File: | build/source/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp |
Warning: | line 8133, column 35 Potential leak of memory pointed to by 'BlockMask' |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | //===- LoopVectorize.cpp - A Loop Vectorizer ------------------------------===// | ||||||||
2 | // | ||||||||
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. | ||||||||
4 | // See https://llvm.org/LICENSE.txt for license information. | ||||||||
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception | ||||||||
6 | // | ||||||||
7 | //===----------------------------------------------------------------------===// | ||||||||
8 | // | ||||||||
9 | // This is the LLVM loop vectorizer. This pass modifies 'vectorizable' loops | ||||||||
10 | // and generates target-independent LLVM-IR. | ||||||||
11 | // The vectorizer uses the TargetTransformInfo analysis to estimate the costs | ||||||||
12 | // of instructions in order to estimate the profitability of vectorization. | ||||||||
13 | // | ||||||||
14 | // The loop vectorizer combines consecutive loop iterations into a single | ||||||||
15 | // 'wide' iteration. After this transformation the index is incremented | ||||||||
16 | // by the SIMD vector width, and not by one. | ||||||||
17 | // | ||||||||
18 | // This pass has three parts: | ||||||||
19 | // 1. The main loop pass that drives the different parts. | ||||||||
20 | // 2. LoopVectorizationLegality - A unit that checks for the legality | ||||||||
21 | // of the vectorization. | ||||||||
22 | // 3. InnerLoopVectorizer - A unit that performs the actual | ||||||||
23 | // widening of instructions. | ||||||||
24 | // 4. LoopVectorizationCostModel - A unit that checks for the profitability | ||||||||
25 | // of vectorization. It decides on the optimal vector width, which | ||||||||
26 | // can be one, if vectorization is not profitable. | ||||||||
27 | // | ||||||||
28 | // There is a development effort going on to migrate loop vectorizer to the | ||||||||
29 | // VPlan infrastructure and to introduce outer loop vectorization support (see | ||||||||
30 | // docs/Proposal/VectorizationPlan.rst and | ||||||||
31 | // http://lists.llvm.org/pipermail/llvm-dev/2017-December/119523.html). For this | ||||||||
32 | // purpose, we temporarily introduced the VPlan-native vectorization path: an | ||||||||
33 | // alternative vectorization path that is natively implemented on top of the | ||||||||
34 | // VPlan infrastructure. See EnableVPlanNativePath for enabling. | ||||||||
35 | // | ||||||||
36 | //===----------------------------------------------------------------------===// | ||||||||
37 | // | ||||||||
38 | // The reduction-variable vectorization is based on the paper: | ||||||||
39 | // D. Nuzman and R. Henderson. Multi-platform Auto-vectorization. | ||||||||
40 | // | ||||||||
41 | // Variable uniformity checks are inspired by: | ||||||||
42 | // Karrenberg, R. and Hack, S. Whole Function Vectorization. | ||||||||
43 | // | ||||||||
44 | // The interleaved access vectorization is based on the paper: | ||||||||
45 | // Dorit Nuzman, Ira Rosen and Ayal Zaks. Auto-Vectorization of Interleaved | ||||||||
46 | // Data for SIMD | ||||||||
47 | // | ||||||||
48 | // Other ideas/concepts are from: | ||||||||
49 | // A. Zaks and D. Nuzman. Autovectorization in GCC-two years later. | ||||||||
50 | // | ||||||||
51 | // S. Maleki, Y. Gao, M. Garzaran, T. Wong and D. Padua. An Evaluation of | ||||||||
52 | // Vectorizing Compilers. | ||||||||
53 | // | ||||||||
54 | //===----------------------------------------------------------------------===// | ||||||||
55 | |||||||||
56 | #include "llvm/Transforms/Vectorize/LoopVectorize.h" | ||||||||
57 | #include "LoopVectorizationPlanner.h" | ||||||||
58 | #include "VPRecipeBuilder.h" | ||||||||
59 | #include "VPlan.h" | ||||||||
60 | #include "VPlanHCFGBuilder.h" | ||||||||
61 | #include "VPlanTransforms.h" | ||||||||
62 | #include "llvm/ADT/APInt.h" | ||||||||
63 | #include "llvm/ADT/ArrayRef.h" | ||||||||
64 | #include "llvm/ADT/DenseMap.h" | ||||||||
65 | #include "llvm/ADT/DenseMapInfo.h" | ||||||||
66 | #include "llvm/ADT/Hashing.h" | ||||||||
67 | #include "llvm/ADT/MapVector.h" | ||||||||
68 | #include "llvm/ADT/STLExtras.h" | ||||||||
69 | #include "llvm/ADT/SmallPtrSet.h" | ||||||||
70 | #include "llvm/ADT/SmallSet.h" | ||||||||
71 | #include "llvm/ADT/SmallVector.h" | ||||||||
72 | #include "llvm/ADT/Statistic.h" | ||||||||
73 | #include "llvm/ADT/StringRef.h" | ||||||||
74 | #include "llvm/ADT/Twine.h" | ||||||||
75 | #include "llvm/ADT/iterator_range.h" | ||||||||
76 | #include "llvm/Analysis/AssumptionCache.h" | ||||||||
77 | #include "llvm/Analysis/BasicAliasAnalysis.h" | ||||||||
78 | #include "llvm/Analysis/BlockFrequencyInfo.h" | ||||||||
79 | #include "llvm/Analysis/CFG.h" | ||||||||
80 | #include "llvm/Analysis/CodeMetrics.h" | ||||||||
81 | #include "llvm/Analysis/DemandedBits.h" | ||||||||
82 | #include "llvm/Analysis/GlobalsModRef.h" | ||||||||
83 | #include "llvm/Analysis/LoopAccessAnalysis.h" | ||||||||
84 | #include "llvm/Analysis/LoopAnalysisManager.h" | ||||||||
85 | #include "llvm/Analysis/LoopInfo.h" | ||||||||
86 | #include "llvm/Analysis/LoopIterator.h" | ||||||||
87 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" | ||||||||
88 | #include "llvm/Analysis/ProfileSummaryInfo.h" | ||||||||
89 | #include "llvm/Analysis/ScalarEvolution.h" | ||||||||
90 | #include "llvm/Analysis/ScalarEvolutionExpressions.h" | ||||||||
91 | #include "llvm/Analysis/TargetLibraryInfo.h" | ||||||||
92 | #include "llvm/Analysis/TargetTransformInfo.h" | ||||||||
93 | #include "llvm/Analysis/ValueTracking.h" | ||||||||
94 | #include "llvm/Analysis/VectorUtils.h" | ||||||||
95 | #include "llvm/IR/Attributes.h" | ||||||||
96 | #include "llvm/IR/BasicBlock.h" | ||||||||
97 | #include "llvm/IR/CFG.h" | ||||||||
98 | #include "llvm/IR/Constant.h" | ||||||||
99 | #include "llvm/IR/Constants.h" | ||||||||
100 | #include "llvm/IR/DataLayout.h" | ||||||||
101 | #include "llvm/IR/DebugInfoMetadata.h" | ||||||||
102 | #include "llvm/IR/DebugLoc.h" | ||||||||
103 | #include "llvm/IR/DerivedTypes.h" | ||||||||
104 | #include "llvm/IR/DiagnosticInfo.h" | ||||||||
105 | #include "llvm/IR/Dominators.h" | ||||||||
106 | #include "llvm/IR/Function.h" | ||||||||
107 | #include "llvm/IR/IRBuilder.h" | ||||||||
108 | #include "llvm/IR/InstrTypes.h" | ||||||||
109 | #include "llvm/IR/Instruction.h" | ||||||||
110 | #include "llvm/IR/Instructions.h" | ||||||||
111 | #include "llvm/IR/IntrinsicInst.h" | ||||||||
112 | #include "llvm/IR/Intrinsics.h" | ||||||||
113 | #include "llvm/IR/Metadata.h" | ||||||||
114 | #include "llvm/IR/Module.h" | ||||||||
115 | #include "llvm/IR/Operator.h" | ||||||||
116 | #include "llvm/IR/PatternMatch.h" | ||||||||
117 | #include "llvm/IR/Type.h" | ||||||||
118 | #include "llvm/IR/Use.h" | ||||||||
119 | #include "llvm/IR/User.h" | ||||||||
120 | #include "llvm/IR/Value.h" | ||||||||
121 | #include "llvm/IR/ValueHandle.h" | ||||||||
122 | #include "llvm/IR/Verifier.h" | ||||||||
123 | #include "llvm/InitializePasses.h" | ||||||||
124 | #include "llvm/Pass.h" | ||||||||
125 | #include "llvm/Support/Casting.h" | ||||||||
126 | #include "llvm/Support/CommandLine.h" | ||||||||
127 | #include "llvm/Support/Compiler.h" | ||||||||
128 | #include "llvm/Support/Debug.h" | ||||||||
129 | #include "llvm/Support/ErrorHandling.h" | ||||||||
130 | #include "llvm/Support/InstructionCost.h" | ||||||||
131 | #include "llvm/Support/MathExtras.h" | ||||||||
132 | #include "llvm/Support/raw_ostream.h" | ||||||||
133 | #include "llvm/Transforms/Utils/BasicBlockUtils.h" | ||||||||
134 | #include "llvm/Transforms/Utils/InjectTLIMappings.h" | ||||||||
135 | #include "llvm/Transforms/Utils/LoopSimplify.h" | ||||||||
136 | #include "llvm/Transforms/Utils/LoopUtils.h" | ||||||||
137 | #include "llvm/Transforms/Utils/LoopVersioning.h" | ||||||||
138 | #include "llvm/Transforms/Utils/ScalarEvolutionExpander.h" | ||||||||
139 | #include "llvm/Transforms/Utils/SizeOpts.h" | ||||||||
140 | #include "llvm/Transforms/Vectorize/LoopVectorizationLegality.h" | ||||||||
141 | #include <algorithm> | ||||||||
142 | #include <cassert> | ||||||||
143 | #include <cmath> | ||||||||
144 | #include <cstdint> | ||||||||
145 | #include <functional> | ||||||||
146 | #include <iterator> | ||||||||
147 | #include <limits> | ||||||||
148 | #include <map> | ||||||||
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> VectorizeMemoryCheckThreshold( | ||||||||
200 | "vectorize-memory-check-threshold", cl::init(128), cl::Hidden, | ||||||||
201 | cl::desc("The maximum allowed number of runtime memory checks")); | ||||||||
202 | |||||||||
203 | // Option prefer-predicate-over-epilogue indicates that an epilogue is undesired, | ||||||||
204 | // that predication is preferred, and this lists all options. I.e., the | ||||||||
205 | // vectorizer will try to fold the tail-loop (epilogue) into the vector body | ||||||||
206 | // and predicate the instructions accordingly. If tail-folding fails, there are | ||||||||
207 | // different fallback strategies depending on these values: | ||||||||
208 | namespace PreferPredicateTy { | ||||||||
209 | enum Option { | ||||||||
210 | ScalarEpilogue = 0, | ||||||||
211 | PredicateElseScalarEpilogue, | ||||||||
212 | PredicateOrDontVectorize | ||||||||
213 | }; | ||||||||
214 | } // namespace PreferPredicateTy | ||||||||
215 | |||||||||
216 | static cl::opt<PreferPredicateTy::Option> PreferPredicateOverEpilogue( | ||||||||
217 | "prefer-predicate-over-epilogue", | ||||||||
218 | cl::init(PreferPredicateTy::ScalarEpilogue), | ||||||||
219 | cl::Hidden, | ||||||||
220 | cl::desc("Tail-folding and predication preferences over creating a scalar " | ||||||||
221 | "epilogue loop."), | ||||||||
222 | cl::values(clEnumValN(PreferPredicateTy::ScalarEpilogue,llvm::cl::OptionEnumValue { "scalar-epilogue", int(PreferPredicateTy ::ScalarEpilogue), "Don't tail-predicate loops, create scalar epilogue" } | ||||||||
223 | "scalar-epilogue",llvm::cl::OptionEnumValue { "scalar-epilogue", int(PreferPredicateTy ::ScalarEpilogue), "Don't tail-predicate loops, create scalar epilogue" } | ||||||||
224 | "Don't tail-predicate loops, create scalar epilogue")llvm::cl::OptionEnumValue { "scalar-epilogue", int(PreferPredicateTy ::ScalarEpilogue), "Don't tail-predicate loops, create scalar epilogue" }, | ||||||||
225 | clEnumValN(PreferPredicateTy::PredicateElseScalarEpilogue,llvm::cl::OptionEnumValue { "predicate-else-scalar-epilogue", int(PreferPredicateTy::PredicateElseScalarEpilogue), "prefer tail-folding, create scalar epilogue if tail " "folding fails." } | ||||||||
226 | "predicate-else-scalar-epilogue",llvm::cl::OptionEnumValue { "predicate-else-scalar-epilogue", int(PreferPredicateTy::PredicateElseScalarEpilogue), "prefer tail-folding, create scalar epilogue if tail " "folding fails." } | ||||||||
227 | "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." } | ||||||||
228 | "folding fails.")llvm::cl::OptionEnumValue { "predicate-else-scalar-epilogue", int(PreferPredicateTy::PredicateElseScalarEpilogue), "prefer tail-folding, create scalar epilogue if tail " "folding fails." }, | ||||||||
229 | clEnumValN(PreferPredicateTy::PredicateOrDontVectorize,llvm::cl::OptionEnumValue { "predicate-dont-vectorize", int(PreferPredicateTy ::PredicateOrDontVectorize), "prefers tail-folding, don't attempt vectorization if " "tail-folding fails." } | ||||||||
230 | "predicate-dont-vectorize",llvm::cl::OptionEnumValue { "predicate-dont-vectorize", int(PreferPredicateTy ::PredicateOrDontVectorize), "prefers tail-folding, don't attempt vectorization if " "tail-folding fails." } | ||||||||
231 | "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." } | ||||||||
232 | "tail-folding fails.")llvm::cl::OptionEnumValue { "predicate-dont-vectorize", int(PreferPredicateTy ::PredicateOrDontVectorize), "prefers tail-folding, don't attempt vectorization if " "tail-folding fails." })); | ||||||||
233 | |||||||||
234 | static cl::opt<bool> MaximizeBandwidth( | ||||||||
235 | "vectorizer-maximize-bandwidth", cl::init(false), cl::Hidden, | ||||||||
236 | cl::desc("Maximize bandwidth when selecting vectorization factor which " | ||||||||
237 | "will be determined by the smallest type in loop.")); | ||||||||
238 | |||||||||
239 | static cl::opt<bool> EnableInterleavedMemAccesses( | ||||||||
240 | "enable-interleaved-mem-accesses", cl::init(false), cl::Hidden, | ||||||||
241 | cl::desc("Enable vectorization on interleaved memory accesses in a loop")); | ||||||||
242 | |||||||||
243 | /// An interleave-group may need masking if it resides in a block that needs | ||||||||
244 | /// predication, or in order to mask away gaps. | ||||||||
245 | static cl::opt<bool> EnableMaskedInterleavedMemAccesses( | ||||||||
246 | "enable-masked-interleaved-mem-accesses", cl::init(false), cl::Hidden, | ||||||||
247 | cl::desc("Enable vectorization on masked interleaved memory accesses in a loop")); | ||||||||
248 | |||||||||
249 | static cl::opt<unsigned> TinyTripCountInterleaveThreshold( | ||||||||
250 | "tiny-trip-count-interleave-threshold", cl::init(128), cl::Hidden, | ||||||||
251 | cl::desc("We don't interleave loops with a estimated constant trip count " | ||||||||
252 | "below this number")); | ||||||||
253 | |||||||||
254 | static cl::opt<unsigned> ForceTargetNumScalarRegs( | ||||||||
255 | "force-target-num-scalar-regs", cl::init(0), cl::Hidden, | ||||||||
256 | cl::desc("A flag that overrides the target's number of scalar registers.")); | ||||||||
257 | |||||||||
258 | static cl::opt<unsigned> ForceTargetNumVectorRegs( | ||||||||
259 | "force-target-num-vector-regs", cl::init(0), cl::Hidden, | ||||||||
260 | cl::desc("A flag that overrides the target's number of vector registers.")); | ||||||||
261 | |||||||||
262 | static cl::opt<unsigned> ForceTargetMaxScalarInterleaveFactor( | ||||||||
263 | "force-target-max-scalar-interleave", cl::init(0), cl::Hidden, | ||||||||
264 | cl::desc("A flag that overrides the target's max interleave factor for " | ||||||||
265 | "scalar loops.")); | ||||||||
266 | |||||||||
267 | static cl::opt<unsigned> ForceTargetMaxVectorInterleaveFactor( | ||||||||
268 | "force-target-max-vector-interleave", cl::init(0), cl::Hidden, | ||||||||
269 | cl::desc("A flag that overrides the target's max interleave factor for " | ||||||||
270 | "vectorized loops.")); | ||||||||
271 | |||||||||
272 | static cl::opt<unsigned> ForceTargetInstructionCost( | ||||||||
273 | "force-target-instruction-cost", cl::init(0), cl::Hidden, | ||||||||
274 | cl::desc("A flag that overrides the target's expected cost for " | ||||||||
275 | "an instruction to a single constant value. Mostly " | ||||||||
276 | "useful for getting consistent testing.")); | ||||||||
277 | |||||||||
278 | static cl::opt<bool> ForceTargetSupportsScalableVectors( | ||||||||
279 | "force-target-supports-scalable-vectors", cl::init(false), cl::Hidden, | ||||||||
280 | cl::desc( | ||||||||
281 | "Pretend that scalable vectors are supported, even if the target does " | ||||||||
282 | "not support them. This flag should only be used for testing.")); | ||||||||
283 | |||||||||
284 | static cl::opt<unsigned> SmallLoopCost( | ||||||||
285 | "small-loop-cost", cl::init(20), cl::Hidden, | ||||||||
286 | cl::desc( | ||||||||
287 | "The cost of a loop that is considered 'small' by the interleaver.")); | ||||||||
288 | |||||||||
289 | static cl::opt<bool> LoopVectorizeWithBlockFrequency( | ||||||||
290 | "loop-vectorize-with-block-frequency", cl::init(true), cl::Hidden, | ||||||||
291 | cl::desc("Enable the use of the block frequency analysis to access PGO " | ||||||||
292 | "heuristics minimizing code growth in cold regions and being more " | ||||||||
293 | "aggressive in hot regions.")); | ||||||||
294 | |||||||||
295 | // Runtime interleave loops for load/store throughput. | ||||||||
296 | static cl::opt<bool> EnableLoadStoreRuntimeInterleave( | ||||||||
297 | "enable-loadstore-runtime-interleave", cl::init(true), cl::Hidden, | ||||||||
298 | cl::desc( | ||||||||
299 | "Enable runtime interleaving until load/store ports are saturated")); | ||||||||
300 | |||||||||
301 | /// Interleave small loops with scalar reductions. | ||||||||
302 | static cl::opt<bool> InterleaveSmallLoopScalarReduction( | ||||||||
303 | "interleave-small-loop-scalar-reduction", cl::init(false), cl::Hidden, | ||||||||
304 | cl::desc("Enable interleaving for loops with small iteration counts that " | ||||||||
305 | "contain scalar reductions to expose ILP.")); | ||||||||
306 | |||||||||
307 | /// The number of stores in a loop that are allowed to need predication. | ||||||||
308 | static cl::opt<unsigned> NumberOfStoresToPredicate( | ||||||||
309 | "vectorize-num-stores-pred", cl::init(1), cl::Hidden, | ||||||||
310 | cl::desc("Max number of stores to be predicated behind an if.")); | ||||||||
311 | |||||||||
312 | static cl::opt<bool> EnableIndVarRegisterHeur( | ||||||||
313 | "enable-ind-var-reg-heur", cl::init(true), cl::Hidden, | ||||||||
314 | cl::desc("Count the induction variable only once when interleaving")); | ||||||||
315 | |||||||||
316 | static cl::opt<bool> EnableCondStoresVectorization( | ||||||||
317 | "enable-cond-stores-vec", cl::init(true), cl::Hidden, | ||||||||
318 | cl::desc("Enable if predication of stores during vectorization.")); | ||||||||
319 | |||||||||
320 | static cl::opt<unsigned> MaxNestedScalarReductionIC( | ||||||||
321 | "max-nested-scalar-reduction-interleave", cl::init(2), cl::Hidden, | ||||||||
322 | cl::desc("The maximum interleave count to use when interleaving a scalar " | ||||||||
323 | "reduction in a nested loop.")); | ||||||||
324 | |||||||||
325 | static cl::opt<bool> | ||||||||
326 | PreferInLoopReductions("prefer-inloop-reductions", cl::init(false), | ||||||||
327 | cl::Hidden, | ||||||||
328 | cl::desc("Prefer in-loop vector reductions, " | ||||||||
329 | "overriding the targets preference.")); | ||||||||
330 | |||||||||
331 | static cl::opt<bool> ForceOrderedReductions( | ||||||||
332 | "force-ordered-reductions", cl::init(false), cl::Hidden, | ||||||||
333 | cl::desc("Enable the vectorisation of loops with in-order (strict) " | ||||||||
334 | "FP reductions")); | ||||||||
335 | |||||||||
336 | static cl::opt<bool> PreferPredicatedReductionSelect( | ||||||||
337 | "prefer-predicated-reduction-select", cl::init(false), cl::Hidden, | ||||||||
338 | cl::desc( | ||||||||
339 | "Prefer predicating a reduction operation over an after loop select.")); | ||||||||
340 | |||||||||
341 | cl::opt<bool> EnableVPlanNativePath( | ||||||||
342 | "enable-vplan-native-path", cl::init(false), cl::Hidden, | ||||||||
343 | cl::desc("Enable VPlan-native vectorization path with " | ||||||||
344 | "support for outer loop vectorization.")); | ||||||||
345 | |||||||||
346 | // This flag enables the stress testing of the VPlan H-CFG construction in the | ||||||||
347 | // VPlan-native vectorization path. It must be used in conjuction with | ||||||||
348 | // -enable-vplan-native-path. -vplan-verify-hcfg can also be used to enable the | ||||||||
349 | // verification of the H-CFGs built. | ||||||||
350 | static cl::opt<bool> VPlanBuildStressTest( | ||||||||
351 | "vplan-build-stress-test", cl::init(false), cl::Hidden, | ||||||||
352 | cl::desc( | ||||||||
353 | "Build VPlan for every supported loop nest in the function and bail " | ||||||||
354 | "out right after the build (stress test the VPlan H-CFG construction " | ||||||||
355 | "in the VPlan-native vectorization path).")); | ||||||||
356 | |||||||||
357 | cl::opt<bool> llvm::EnableLoopInterleaving( | ||||||||
358 | "interleave-loops", cl::init(true), cl::Hidden, | ||||||||
359 | cl::desc("Enable loop interleaving in Loop vectorization passes")); | ||||||||
360 | cl::opt<bool> llvm::EnableLoopVectorization( | ||||||||
361 | "vectorize-loops", cl::init(true), cl::Hidden, | ||||||||
362 | cl::desc("Run the Loop vectorization passes")); | ||||||||
363 | |||||||||
364 | static cl::opt<bool> PrintVPlansInDotFormat( | ||||||||
365 | "vplan-print-in-dot-format", cl::Hidden, | ||||||||
366 | cl::desc("Use dot format instead of plain text when dumping VPlans")); | ||||||||
367 | |||||||||
368 | static cl::opt<cl::boolOrDefault> ForceSafeDivisor( | ||||||||
369 | "force-widen-divrem-via-safe-divisor", cl::Hidden, | ||||||||
370 | cl::desc( | ||||||||
371 | "Override cost based safe divisor widening for div/rem instructions")); | ||||||||
372 | |||||||||
373 | /// A helper function that returns true if the given type is irregular. The | ||||||||
374 | /// type is irregular if its allocated size doesn't equal the store size of an | ||||||||
375 | /// element of the corresponding vector type. | ||||||||
376 | static bool hasIrregularType(Type *Ty, const DataLayout &DL) { | ||||||||
377 | // Determine if an array of N elements of type Ty is "bitcast compatible" | ||||||||
378 | // with a <N x Ty> vector. | ||||||||
379 | // This is only true if there is no padding between the array elements. | ||||||||
380 | return DL.getTypeAllocSizeInBits(Ty) != DL.getTypeSizeInBits(Ty); | ||||||||
381 | } | ||||||||
382 | |||||||||
383 | /// A helper function that returns the reciprocal of the block probability of | ||||||||
384 | /// predicated blocks. If we return X, we are assuming the predicated block | ||||||||
385 | /// will execute once for every X iterations of the loop header. | ||||||||
386 | /// | ||||||||
387 | /// TODO: We should use actual block probability here, if available. Currently, | ||||||||
388 | /// we always assume predicated blocks have a 50% chance of executing. | ||||||||
389 | static unsigned getReciprocalPredBlockProb() { return 2; } | ||||||||
390 | |||||||||
391 | /// A helper function that returns an integer or floating-point constant with | ||||||||
392 | /// value C. | ||||||||
393 | static Constant *getSignedIntOrFpConstant(Type *Ty, int64_t C) { | ||||||||
394 | return Ty->isIntegerTy() ? ConstantInt::getSigned(Ty, C) | ||||||||
395 | : ConstantFP::get(Ty, C); | ||||||||
396 | } | ||||||||
397 | |||||||||
398 | /// Returns "best known" trip count for the specified loop \p L as defined by | ||||||||
399 | /// the following procedure: | ||||||||
400 | /// 1) Returns exact trip count if it is known. | ||||||||
401 | /// 2) Returns expected trip count according to profile data if any. | ||||||||
402 | /// 3) Returns upper bound estimate if it is known. | ||||||||
403 | /// 4) Returns std::nullopt if all of the above failed. | ||||||||
404 | static std::optional<unsigned> getSmallBestKnownTC(ScalarEvolution &SE, | ||||||||
405 | Loop *L) { | ||||||||
406 | // Check if exact trip count is known. | ||||||||
407 | if (unsigned ExpectedTC = SE.getSmallConstantTripCount(L)) | ||||||||
408 | return ExpectedTC; | ||||||||
409 | |||||||||
410 | // Check if there is an expected trip count available from profile data. | ||||||||
411 | if (LoopVectorizeWithBlockFrequency) | ||||||||
412 | if (auto EstimatedTC = getLoopEstimatedTripCount(L)) | ||||||||
413 | return *EstimatedTC; | ||||||||
414 | |||||||||
415 | // Check if upper bound estimate is known. | ||||||||
416 | if (unsigned ExpectedTC = SE.getSmallConstantMaxTripCount(L)) | ||||||||
417 | return ExpectedTC; | ||||||||
418 | |||||||||
419 | return std::nullopt; | ||||||||
420 | } | ||||||||
421 | |||||||||
422 | namespace { | ||||||||
423 | // Forward declare GeneratedRTChecks. | ||||||||
424 | class GeneratedRTChecks; | ||||||||
425 | } // namespace | ||||||||
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 | ElementCount MinProfitableTripCount, | ||||||||
453 | unsigned UnrollFactor, LoopVectorizationLegality *LVL, | ||||||||
454 | LoopVectorizationCostModel *CM, BlockFrequencyInfo *BFI, | ||||||||
455 | ProfileSummaryInfo *PSI, GeneratedRTChecks &RTChecks) | ||||||||
456 | : OrigLoop(OrigLoop), PSE(PSE), LI(LI), DT(DT), TLI(TLI), TTI(TTI), | ||||||||
457 | AC(AC), ORE(ORE), VF(VecWidth), UF(UnrollFactor), | ||||||||
458 | Builder(PSE.getSE()->getContext()), Legal(LVL), Cost(CM), BFI(BFI), | ||||||||
459 | PSI(PSI), RTChecks(RTChecks) { | ||||||||
460 | // Query this against the original loop and save it here because the profile | ||||||||
461 | // of the original loop header may change as the transformation happens. | ||||||||
462 | OptForSizeBasedOnProfile = llvm::shouldOptimizeForSize( | ||||||||
463 | OrigLoop->getHeader(), PSI, BFI, PGSOQueryType::IRPass); | ||||||||
464 | |||||||||
465 | if (MinProfitableTripCount.isZero()) | ||||||||
466 | this->MinProfitableTripCount = VecWidth; | ||||||||
467 | else | ||||||||
468 | this->MinProfitableTripCount = MinProfitableTripCount; | ||||||||
469 | } | ||||||||
470 | |||||||||
471 | virtual ~InnerLoopVectorizer() = default; | ||||||||
472 | |||||||||
473 | /// Create a new empty loop that will contain vectorized instructions later | ||||||||
474 | /// on, while the old loop will be used as the scalar remainder. Control flow | ||||||||
475 | /// is generated around the vectorized (and scalar epilogue) loops consisting | ||||||||
476 | /// of various checks and bypasses. Return the pre-header block of the new | ||||||||
477 | /// loop and the start value for the canonical induction, if it is != 0. The | ||||||||
478 | /// latter is the case when vectorizing the epilogue loop. In the case of | ||||||||
479 | /// epilogue vectorization, this function is overriden to handle the more | ||||||||
480 | /// complex control flow around the loops. | ||||||||
481 | virtual std::pair<BasicBlock *, Value *> createVectorizedLoopSkeleton(); | ||||||||
482 | |||||||||
483 | /// Fix the vectorized code, taking care of header phi's, live-outs, and more. | ||||||||
484 | void fixVectorizedLoop(VPTransformState &State, VPlan &Plan); | ||||||||
485 | |||||||||
486 | // Return true if any runtime check is added. | ||||||||
487 | bool areSafetyChecksAdded() { return AddedSafetyChecks; } | ||||||||
488 | |||||||||
489 | /// A type for vectorized values in the new loop. Each value from the | ||||||||
490 | /// original loop, when vectorized, is represented by UF vector values in the | ||||||||
491 | /// new unrolled loop, where UF is the unroll factor. | ||||||||
492 | using VectorParts = SmallVector<Value *, 2>; | ||||||||
493 | |||||||||
494 | /// A helper function to scalarize a single Instruction in the innermost loop. | ||||||||
495 | /// Generates a sequence of scalar instances for each lane between \p MinLane | ||||||||
496 | /// and \p MaxLane, times each part between \p MinPart and \p MaxPart, | ||||||||
497 | /// inclusive. Uses the VPValue operands from \p RepRecipe instead of \p | ||||||||
498 | /// Instr's operands. | ||||||||
499 | void scalarizeInstruction(const Instruction *Instr, | ||||||||
500 | VPReplicateRecipe *RepRecipe, | ||||||||
501 | const VPIteration &Instance, bool IfPredicateInstr, | ||||||||
502 | VPTransformState &State); | ||||||||
503 | |||||||||
504 | /// Construct the vector value of a scalarized value \p V one lane at a time. | ||||||||
505 | void packScalarIntoVectorValue(VPValue *Def, const VPIteration &Instance, | ||||||||
506 | VPTransformState &State); | ||||||||
507 | |||||||||
508 | /// Try to vectorize interleaved access group \p Group with the base address | ||||||||
509 | /// given in \p Addr, optionally masking the vector operations if \p | ||||||||
510 | /// BlockInMask is non-null. Use \p State to translate given VPValues to IR | ||||||||
511 | /// values in the vectorized loop. | ||||||||
512 | void vectorizeInterleaveGroup(const InterleaveGroup<Instruction> *Group, | ||||||||
513 | ArrayRef<VPValue *> VPDefs, | ||||||||
514 | VPTransformState &State, VPValue *Addr, | ||||||||
515 | ArrayRef<VPValue *> StoredValues, | ||||||||
516 | VPValue *BlockInMask = nullptr); | ||||||||
517 | |||||||||
518 | /// Fix the non-induction PHIs in \p Plan. | ||||||||
519 | void fixNonInductionPHIs(VPlan &Plan, VPTransformState &State); | ||||||||
520 | |||||||||
521 | /// Returns true if the reordering of FP operations is not allowed, but we are | ||||||||
522 | /// able to vectorize with strict in-order reductions for the given RdxDesc. | ||||||||
523 | bool useOrderedReductions(const RecurrenceDescriptor &RdxDesc); | ||||||||
524 | |||||||||
525 | /// Create a broadcast instruction. This method generates a broadcast | ||||||||
526 | /// instruction (shuffle) for loop invariant values and for the induction | ||||||||
527 | /// value. If this is the induction variable then we extend it to N, N+1, ... | ||||||||
528 | /// this is needed because each iteration in the loop corresponds to a SIMD | ||||||||
529 | /// element. | ||||||||
530 | virtual Value *getBroadcastInstrs(Value *V); | ||||||||
531 | |||||||||
532 | // Returns the resume value (bc.merge.rdx) for a reduction as | ||||||||
533 | // generated by fixReduction. | ||||||||
534 | PHINode *getReductionResumeValue(const RecurrenceDescriptor &RdxDesc); | ||||||||
535 | |||||||||
536 | /// Create a new phi node for the induction variable \p OrigPhi to resume | ||||||||
537 | /// iteration count in the scalar epilogue, from where the vectorized loop | ||||||||
538 | /// left off. In cases where the loop skeleton is more complicated (eg. | ||||||||
539 | /// epilogue vectorization) and the resume values can come from an additional | ||||||||
540 | /// bypass block, the \p AdditionalBypass pair provides information about the | ||||||||
541 | /// bypass block and the end value on the edge from bypass to this loop. | ||||||||
542 | PHINode *createInductionResumeValue( | ||||||||
543 | PHINode *OrigPhi, const InductionDescriptor &ID, | ||||||||
544 | ArrayRef<BasicBlock *> BypassBlocks, | ||||||||
545 | std::pair<BasicBlock *, Value *> AdditionalBypass = {nullptr, nullptr}); | ||||||||
546 | |||||||||
547 | protected: | ||||||||
548 | friend class LoopVectorizationPlanner; | ||||||||
549 | |||||||||
550 | /// A small list of PHINodes. | ||||||||
551 | using PhiVector = SmallVector<PHINode *, 4>; | ||||||||
552 | |||||||||
553 | /// A type for scalarized values in the new loop. Each value from the | ||||||||
554 | /// original loop, when scalarized, is represented by UF x VF scalar values | ||||||||
555 | /// in the new unrolled loop, where UF is the unroll factor and VF is the | ||||||||
556 | /// vectorization factor. | ||||||||
557 | using ScalarParts = SmallVector<SmallVector<Value *, 4>, 2>; | ||||||||
558 | |||||||||
559 | /// Set up the values of the IVs correctly when exiting the vector loop. | ||||||||
560 | void fixupIVUsers(PHINode *OrigPhi, const InductionDescriptor &II, | ||||||||
561 | Value *VectorTripCount, Value *EndValue, | ||||||||
562 | BasicBlock *MiddleBlock, BasicBlock *VectorHeader, | ||||||||
563 | VPlan &Plan); | ||||||||
564 | |||||||||
565 | /// Handle all cross-iteration phis in the header. | ||||||||
566 | void fixCrossIterationPHIs(VPTransformState &State); | ||||||||
567 | |||||||||
568 | /// Create the exit value of first order recurrences in the middle block and | ||||||||
569 | /// update their users. | ||||||||
570 | void fixFixedOrderRecurrence(VPFirstOrderRecurrencePHIRecipe *PhiR, | ||||||||
571 | VPTransformState &State); | ||||||||
572 | |||||||||
573 | /// Create code for the loop exit value of the reduction. | ||||||||
574 | void fixReduction(VPReductionPHIRecipe *Phi, VPTransformState &State); | ||||||||
575 | |||||||||
576 | /// Clear NSW/NUW flags from reduction instructions if necessary. | ||||||||
577 | void clearReductionWrapFlags(VPReductionPHIRecipe *PhiR, | ||||||||
578 | VPTransformState &State); | ||||||||
579 | |||||||||
580 | /// Iteratively sink the scalarized operands of a predicated instruction into | ||||||||
581 | /// the block that was created for it. | ||||||||
582 | void sinkScalarOperands(Instruction *PredInst); | ||||||||
583 | |||||||||
584 | /// Shrinks vector element sizes to the smallest bitwidth they can be legally | ||||||||
585 | /// represented as. | ||||||||
586 | void truncateToMinimalBitwidths(VPTransformState &State); | ||||||||
587 | |||||||||
588 | /// Returns (and creates if needed) the original loop trip count. | ||||||||
589 | Value *getOrCreateTripCount(BasicBlock *InsertBlock); | ||||||||
590 | |||||||||
591 | /// Returns (and creates if needed) the trip count of the widened loop. | ||||||||
592 | Value *getOrCreateVectorTripCount(BasicBlock *InsertBlock); | ||||||||
593 | |||||||||
594 | /// Returns a bitcasted value to the requested vector type. | ||||||||
595 | /// Also handles bitcasts of vector<float> <-> vector<pointer> types. | ||||||||
596 | Value *createBitOrPointerCast(Value *V, VectorType *DstVTy, | ||||||||
597 | const DataLayout &DL); | ||||||||
598 | |||||||||
599 | /// Emit a bypass check to see if the vector trip count is zero, including if | ||||||||
600 | /// it overflows. | ||||||||
601 | void emitIterationCountCheck(BasicBlock *Bypass); | ||||||||
602 | |||||||||
603 | /// Emit a bypass check to see if all of the SCEV assumptions we've | ||||||||
604 | /// had to make are correct. Returns the block containing the checks or | ||||||||
605 | /// nullptr if no checks have been added. | ||||||||
606 | BasicBlock *emitSCEVChecks(BasicBlock *Bypass); | ||||||||
607 | |||||||||
608 | /// Emit bypass checks to check any memory assumptions we may have made. | ||||||||
609 | /// Returns the block containing the checks or nullptr if no checks have been | ||||||||
610 | /// added. | ||||||||
611 | BasicBlock *emitMemRuntimeChecks(BasicBlock *Bypass); | ||||||||
612 | |||||||||
613 | /// Emit basic blocks (prefixed with \p Prefix) for the iteration check, | ||||||||
614 | /// vector loop preheader, middle block and scalar preheader. | ||||||||
615 | void createVectorLoopSkeleton(StringRef Prefix); | ||||||||
616 | |||||||||
617 | /// Create new phi nodes for the induction variables to resume iteration count | ||||||||
618 | /// in the scalar epilogue, from where the vectorized loop left off. | ||||||||
619 | /// In cases where the loop skeleton is more complicated (eg. epilogue | ||||||||
620 | /// vectorization) and the resume values can come from an additional bypass | ||||||||
621 | /// block, the \p AdditionalBypass pair provides information about the bypass | ||||||||
622 | /// block and the end value on the edge from bypass to this loop. | ||||||||
623 | void createInductionResumeValues( | ||||||||
624 | std::pair<BasicBlock *, Value *> AdditionalBypass = {nullptr, nullptr}); | ||||||||
625 | |||||||||
626 | /// Complete the loop skeleton by adding debug MDs, creating appropriate | ||||||||
627 | /// conditional branches in the middle block, preparing the builder and | ||||||||
628 | /// running the verifier. Return the preheader of the completed vector loop. | ||||||||
629 | BasicBlock *completeLoopSkeleton(); | ||||||||
630 | |||||||||
631 | /// Collect poison-generating recipes that may generate a poison value that is | ||||||||
632 | /// used after vectorization, even when their operands are not poison. Those | ||||||||
633 | /// recipes meet the following conditions: | ||||||||
634 | /// * Contribute to the address computation of a recipe generating a widen | ||||||||
635 | /// memory load/store (VPWidenMemoryInstructionRecipe or | ||||||||
636 | /// VPInterleaveRecipe). | ||||||||
637 | /// * Such a widen memory load/store has at least one underlying Instruction | ||||||||
638 | /// that is in a basic block that needs predication and after vectorization | ||||||||
639 | /// the generated instruction won't be predicated. | ||||||||
640 | void collectPoisonGeneratingRecipes(VPTransformState &State); | ||||||||
641 | |||||||||
642 | /// Allow subclasses to override and print debug traces before/after vplan | ||||||||
643 | /// execution, when trace information is requested. | ||||||||
644 | virtual void printDebugTracesAtStart(){}; | ||||||||
645 | virtual void printDebugTracesAtEnd(){}; | ||||||||
646 | |||||||||
647 | /// The original loop. | ||||||||
648 | Loop *OrigLoop; | ||||||||
649 | |||||||||
650 | /// A wrapper around ScalarEvolution used to add runtime SCEV checks. Applies | ||||||||
651 | /// dynamic knowledge to simplify SCEV expressions and converts them to a | ||||||||
652 | /// more usable form. | ||||||||
653 | PredicatedScalarEvolution &PSE; | ||||||||
654 | |||||||||
655 | /// Loop Info. | ||||||||
656 | LoopInfo *LI; | ||||||||
657 | |||||||||
658 | /// Dominator Tree. | ||||||||
659 | DominatorTree *DT; | ||||||||
660 | |||||||||
661 | /// Target Library Info. | ||||||||
662 | const TargetLibraryInfo *TLI; | ||||||||
663 | |||||||||
664 | /// Target Transform Info. | ||||||||
665 | const TargetTransformInfo *TTI; | ||||||||
666 | |||||||||
667 | /// Assumption Cache. | ||||||||
668 | AssumptionCache *AC; | ||||||||
669 | |||||||||
670 | /// Interface to emit optimization remarks. | ||||||||
671 | OptimizationRemarkEmitter *ORE; | ||||||||
672 | |||||||||
673 | /// The vectorization SIMD factor to use. Each vector will have this many | ||||||||
674 | /// vector elements. | ||||||||
675 | ElementCount VF; | ||||||||
676 | |||||||||
677 | ElementCount MinProfitableTripCount; | ||||||||
678 | |||||||||
679 | /// The vectorization unroll factor to use. Each scalar is vectorized to this | ||||||||
680 | /// many different vector instructions. | ||||||||
681 | unsigned UF; | ||||||||
682 | |||||||||
683 | /// The builder that we use | ||||||||
684 | IRBuilder<> Builder; | ||||||||
685 | |||||||||
686 | // --- Vectorization state --- | ||||||||
687 | |||||||||
688 | /// The vector-loop preheader. | ||||||||
689 | BasicBlock *LoopVectorPreHeader; | ||||||||
690 | |||||||||
691 | /// The scalar-loop preheader. | ||||||||
692 | BasicBlock *LoopScalarPreHeader; | ||||||||
693 | |||||||||
694 | /// Middle Block between the vector and the scalar. | ||||||||
695 | BasicBlock *LoopMiddleBlock; | ||||||||
696 | |||||||||
697 | /// The unique ExitBlock of the scalar loop if one exists. Note that | ||||||||
698 | /// there can be multiple exiting edges reaching this block. | ||||||||
699 | BasicBlock *LoopExitBlock; | ||||||||
700 | |||||||||
701 | /// The scalar loop body. | ||||||||
702 | BasicBlock *LoopScalarBody; | ||||||||
703 | |||||||||
704 | /// A list of all bypass blocks. The first block is the entry of the loop. | ||||||||
705 | SmallVector<BasicBlock *, 4> LoopBypassBlocks; | ||||||||
706 | |||||||||
707 | /// Store instructions that were predicated. | ||||||||
708 | SmallVector<Instruction *, 4> PredicatedInstructions; | ||||||||
709 | |||||||||
710 | /// Trip count of the original loop. | ||||||||
711 | Value *TripCount = nullptr; | ||||||||
712 | |||||||||
713 | /// Trip count of the widened loop (TripCount - TripCount % (VF*UF)) | ||||||||
714 | Value *VectorTripCount = nullptr; | ||||||||
715 | |||||||||
716 | /// The legality analysis. | ||||||||
717 | LoopVectorizationLegality *Legal; | ||||||||
718 | |||||||||
719 | /// The profitablity analysis. | ||||||||
720 | LoopVectorizationCostModel *Cost; | ||||||||
721 | |||||||||
722 | // Record whether runtime checks are added. | ||||||||
723 | bool AddedSafetyChecks = false; | ||||||||
724 | |||||||||
725 | // Holds the end values for each induction variable. We save the end values | ||||||||
726 | // so we can later fix-up the external users of the induction variables. | ||||||||
727 | DenseMap<PHINode *, Value *> IVEndValues; | ||||||||
728 | |||||||||
729 | /// BFI and PSI are used to check for profile guided size optimizations. | ||||||||
730 | BlockFrequencyInfo *BFI; | ||||||||
731 | ProfileSummaryInfo *PSI; | ||||||||
732 | |||||||||
733 | // Whether this loop should be optimized for size based on profile guided size | ||||||||
734 | // optimizatios. | ||||||||
735 | bool OptForSizeBasedOnProfile; | ||||||||
736 | |||||||||
737 | /// Structure to hold information about generated runtime checks, responsible | ||||||||
738 | /// for cleaning the checks, if vectorization turns out unprofitable. | ||||||||
739 | GeneratedRTChecks &RTChecks; | ||||||||
740 | |||||||||
741 | // Holds the resume values for reductions in the loops, used to set the | ||||||||
742 | // correct start value of reduction PHIs when vectorizing the epilogue. | ||||||||
743 | SmallMapVector<const RecurrenceDescriptor *, PHINode *, 4> | ||||||||
744 | ReductionResumeValues; | ||||||||
745 | }; | ||||||||
746 | |||||||||
747 | class InnerLoopUnroller : public InnerLoopVectorizer { | ||||||||
748 | public: | ||||||||
749 | InnerLoopUnroller(Loop *OrigLoop, PredicatedScalarEvolution &PSE, | ||||||||
750 | LoopInfo *LI, DominatorTree *DT, | ||||||||
751 | const TargetLibraryInfo *TLI, | ||||||||
752 | const TargetTransformInfo *TTI, AssumptionCache *AC, | ||||||||
753 | OptimizationRemarkEmitter *ORE, unsigned UnrollFactor, | ||||||||
754 | LoopVectorizationLegality *LVL, | ||||||||
755 | LoopVectorizationCostModel *CM, BlockFrequencyInfo *BFI, | ||||||||
756 | ProfileSummaryInfo *PSI, GeneratedRTChecks &Check) | ||||||||
757 | : InnerLoopVectorizer(OrigLoop, PSE, LI, DT, TLI, TTI, AC, ORE, | ||||||||
758 | ElementCount::getFixed(1), | ||||||||
759 | ElementCount::getFixed(1), UnrollFactor, LVL, CM, | ||||||||
760 | BFI, PSI, Check) {} | ||||||||
761 | |||||||||
762 | private: | ||||||||
763 | Value *getBroadcastInstrs(Value *V) override; | ||||||||
764 | }; | ||||||||
765 | |||||||||
766 | /// Encapsulate information regarding vectorization of a loop and its epilogue. | ||||||||
767 | /// This information is meant to be updated and used across two stages of | ||||||||
768 | /// epilogue vectorization. | ||||||||
769 | struct EpilogueLoopVectorizationInfo { | ||||||||
770 | ElementCount MainLoopVF = ElementCount::getFixed(0); | ||||||||
771 | unsigned MainLoopUF = 0; | ||||||||
772 | ElementCount EpilogueVF = ElementCount::getFixed(0); | ||||||||
773 | unsigned EpilogueUF = 0; | ||||||||
774 | BasicBlock *MainLoopIterationCountCheck = nullptr; | ||||||||
775 | BasicBlock *EpilogueIterationCountCheck = nullptr; | ||||||||
776 | BasicBlock *SCEVSafetyCheck = nullptr; | ||||||||
777 | BasicBlock *MemSafetyCheck = nullptr; | ||||||||
778 | Value *TripCount = nullptr; | ||||||||
779 | Value *VectorTripCount = nullptr; | ||||||||
780 | |||||||||
781 | EpilogueLoopVectorizationInfo(ElementCount MVF, unsigned MUF, | ||||||||
782 | ElementCount EVF, unsigned EUF) | ||||||||
783 | : MainLoopVF(MVF), MainLoopUF(MUF), EpilogueVF(EVF), EpilogueUF(EUF) { | ||||||||
784 | 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", 785, __extension__ __PRETTY_FUNCTION__)) | ||||||||
785 | "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", 785, __extension__ __PRETTY_FUNCTION__)); | ||||||||
786 | } | ||||||||
787 | }; | ||||||||
788 | |||||||||
789 | /// An extension of the inner loop vectorizer that creates a skeleton for a | ||||||||
790 | /// vectorized loop that has its epilogue (residual) also vectorized. | ||||||||
791 | /// The idea is to run the vplan on a given loop twice, firstly to setup the | ||||||||
792 | /// skeleton and vectorize the main loop, and secondly to complete the skeleton | ||||||||
793 | /// from the first step and vectorize the epilogue. This is achieved by | ||||||||
794 | /// deriving two concrete strategy classes from this base class and invoking | ||||||||
795 | /// them in succession from the loop vectorizer planner. | ||||||||
796 | class InnerLoopAndEpilogueVectorizer : public InnerLoopVectorizer { | ||||||||
797 | public: | ||||||||
798 | InnerLoopAndEpilogueVectorizer( | ||||||||
799 | Loop *OrigLoop, PredicatedScalarEvolution &PSE, LoopInfo *LI, | ||||||||
800 | DominatorTree *DT, const TargetLibraryInfo *TLI, | ||||||||
801 | const TargetTransformInfo *TTI, AssumptionCache *AC, | ||||||||
802 | OptimizationRemarkEmitter *ORE, EpilogueLoopVectorizationInfo &EPI, | ||||||||
803 | LoopVectorizationLegality *LVL, llvm::LoopVectorizationCostModel *CM, | ||||||||
804 | BlockFrequencyInfo *BFI, ProfileSummaryInfo *PSI, | ||||||||
805 | GeneratedRTChecks &Checks) | ||||||||
806 | : InnerLoopVectorizer(OrigLoop, PSE, LI, DT, TLI, TTI, AC, ORE, | ||||||||
807 | EPI.MainLoopVF, EPI.MainLoopVF, EPI.MainLoopUF, LVL, | ||||||||
808 | CM, BFI, PSI, Checks), | ||||||||
809 | EPI(EPI) {} | ||||||||
810 | |||||||||
811 | // Override this function to handle the more complex control flow around the | ||||||||
812 | // three loops. | ||||||||
813 | std::pair<BasicBlock *, Value *> createVectorizedLoopSkeleton() final { | ||||||||
814 | return createEpilogueVectorizedLoopSkeleton(); | ||||||||
815 | } | ||||||||
816 | |||||||||
817 | /// The interface for creating a vectorized skeleton using one of two | ||||||||
818 | /// different strategies, each corresponding to one execution of the vplan | ||||||||
819 | /// as described above. | ||||||||
820 | virtual std::pair<BasicBlock *, Value *> | ||||||||
821 | createEpilogueVectorizedLoopSkeleton() = 0; | ||||||||
822 | |||||||||
823 | /// Holds and updates state information required to vectorize the main loop | ||||||||
824 | /// and its epilogue in two separate passes. This setup helps us avoid | ||||||||
825 | /// regenerating and recomputing runtime safety checks. It also helps us to | ||||||||
826 | /// shorten the iteration-count-check path length for the cases where the | ||||||||
827 | /// iteration count of the loop is so small that the main vector loop is | ||||||||
828 | /// completely skipped. | ||||||||
829 | EpilogueLoopVectorizationInfo &EPI; | ||||||||
830 | }; | ||||||||
831 | |||||||||
832 | /// A specialized derived class of inner loop vectorizer that performs | ||||||||
833 | /// vectorization of *main* loops in the process of vectorizing loops and their | ||||||||
834 | /// epilogues. | ||||||||
835 | class EpilogueVectorizerMainLoop : public InnerLoopAndEpilogueVectorizer { | ||||||||
836 | public: | ||||||||
837 | EpilogueVectorizerMainLoop( | ||||||||
838 | Loop *OrigLoop, PredicatedScalarEvolution &PSE, LoopInfo *LI, | ||||||||
839 | DominatorTree *DT, const TargetLibraryInfo *TLI, | ||||||||
840 | const TargetTransformInfo *TTI, AssumptionCache *AC, | ||||||||
841 | OptimizationRemarkEmitter *ORE, EpilogueLoopVectorizationInfo &EPI, | ||||||||
842 | LoopVectorizationLegality *LVL, llvm::LoopVectorizationCostModel *CM, | ||||||||
843 | BlockFrequencyInfo *BFI, ProfileSummaryInfo *PSI, | ||||||||
844 | GeneratedRTChecks &Check) | ||||||||
845 | : InnerLoopAndEpilogueVectorizer(OrigLoop, PSE, LI, DT, TLI, TTI, AC, ORE, | ||||||||
846 | EPI, LVL, CM, BFI, PSI, Check) {} | ||||||||
847 | /// Implements the interface for creating a vectorized skeleton using the | ||||||||
848 | /// *main loop* strategy (ie the first pass of vplan execution). | ||||||||
849 | std::pair<BasicBlock *, Value *> createEpilogueVectorizedLoopSkeleton() final; | ||||||||
850 | |||||||||
851 | protected: | ||||||||
852 | /// Emits an iteration count bypass check once for the main loop (when \p | ||||||||
853 | /// ForEpilogue is false) and once for the epilogue loop (when \p | ||||||||
854 | /// ForEpilogue is true). | ||||||||
855 | BasicBlock *emitIterationCountCheck(BasicBlock *Bypass, bool ForEpilogue); | ||||||||
856 | void printDebugTracesAtStart() override; | ||||||||
857 | void printDebugTracesAtEnd() override; | ||||||||
858 | }; | ||||||||
859 | |||||||||
860 | // A specialized derived class of inner loop vectorizer that performs | ||||||||
861 | // vectorization of *epilogue* loops in the process of vectorizing loops and | ||||||||
862 | // their epilogues. | ||||||||
863 | class EpilogueVectorizerEpilogueLoop : public InnerLoopAndEpilogueVectorizer { | ||||||||
864 | public: | ||||||||
865 | EpilogueVectorizerEpilogueLoop( | ||||||||
866 | Loop *OrigLoop, PredicatedScalarEvolution &PSE, LoopInfo *LI, | ||||||||
867 | DominatorTree *DT, const TargetLibraryInfo *TLI, | ||||||||
868 | const TargetTransformInfo *TTI, AssumptionCache *AC, | ||||||||
869 | OptimizationRemarkEmitter *ORE, EpilogueLoopVectorizationInfo &EPI, | ||||||||
870 | LoopVectorizationLegality *LVL, llvm::LoopVectorizationCostModel *CM, | ||||||||
871 | BlockFrequencyInfo *BFI, ProfileSummaryInfo *PSI, | ||||||||
872 | GeneratedRTChecks &Checks) | ||||||||
873 | : InnerLoopAndEpilogueVectorizer(OrigLoop, PSE, LI, DT, TLI, TTI, AC, ORE, | ||||||||
874 | EPI, LVL, CM, BFI, PSI, Checks) { | ||||||||
875 | TripCount = EPI.TripCount; | ||||||||
876 | } | ||||||||
877 | /// Implements the interface for creating a vectorized skeleton using the | ||||||||
878 | /// *epilogue loop* strategy (ie the second pass of vplan execution). | ||||||||
879 | std::pair<BasicBlock *, Value *> createEpilogueVectorizedLoopSkeleton() final; | ||||||||
880 | |||||||||
881 | protected: | ||||||||
882 | /// Emits an iteration count bypass check after the main vector loop has | ||||||||
883 | /// finished to see if there are any iterations left to execute by either | ||||||||
884 | /// the vector epilogue or the scalar epilogue. | ||||||||
885 | BasicBlock *emitMinimumVectorEpilogueIterCountCheck( | ||||||||
886 | BasicBlock *Bypass, | ||||||||
887 | BasicBlock *Insert); | ||||||||
888 | void printDebugTracesAtStart() override; | ||||||||
889 | void printDebugTracesAtEnd() override; | ||||||||
890 | }; | ||||||||
891 | } // end namespace llvm | ||||||||
892 | |||||||||
893 | /// Look for a meaningful debug location on the instruction or it's | ||||||||
894 | /// operands. | ||||||||
895 | static Instruction *getDebugLocFromInstOrOperands(Instruction *I) { | ||||||||
896 | if (!I) | ||||||||
897 | return I; | ||||||||
898 | |||||||||
899 | DebugLoc Empty; | ||||||||
900 | if (I->getDebugLoc() != Empty) | ||||||||
901 | return I; | ||||||||
902 | |||||||||
903 | for (Use &Op : I->operands()) { | ||||||||
904 | if (Instruction *OpInst = dyn_cast<Instruction>(Op)) | ||||||||
905 | if (OpInst->getDebugLoc() != Empty) | ||||||||
906 | return OpInst; | ||||||||
907 | } | ||||||||
908 | |||||||||
909 | return I; | ||||||||
910 | } | ||||||||
911 | |||||||||
912 | /// Write a \p DebugMsg about vectorization to the debug output stream. If \p I | ||||||||
913 | /// is passed, the message relates to that particular instruction. | ||||||||
914 | #ifndef NDEBUG | ||||||||
915 | static void debugVectorizationMessage(const StringRef Prefix, | ||||||||
916 | const StringRef DebugMsg, | ||||||||
917 | Instruction *I) { | ||||||||
918 | dbgs() << "LV: " << Prefix << DebugMsg; | ||||||||
919 | if (I != nullptr) | ||||||||
920 | dbgs() << " " << *I; | ||||||||
921 | else | ||||||||
922 | dbgs() << '.'; | ||||||||
923 | dbgs() << '\n'; | ||||||||
924 | } | ||||||||
925 | #endif | ||||||||
926 | |||||||||
927 | /// Create an analysis remark that explains why vectorization failed | ||||||||
928 | /// | ||||||||
929 | /// \p PassName is the name of the pass (e.g. can be AlwaysPrint). \p | ||||||||
930 | /// RemarkName is the identifier for the remark. If \p I is passed it is an | ||||||||
931 | /// instruction that prevents vectorization. Otherwise \p TheLoop is used for | ||||||||
932 | /// the location of the remark. \return the remark object that can be | ||||||||
933 | /// streamed to. | ||||||||
934 | static OptimizationRemarkAnalysis createLVAnalysis(const char *PassName, | ||||||||
935 | StringRef RemarkName, Loop *TheLoop, Instruction *I) { | ||||||||
936 | Value *CodeRegion = TheLoop->getHeader(); | ||||||||
937 | DebugLoc DL = TheLoop->getStartLoc(); | ||||||||
938 | |||||||||
939 | if (I) { | ||||||||
940 | CodeRegion = I->getParent(); | ||||||||
941 | // If there is no debug location attached to the instruction, revert back to | ||||||||
942 | // using the loop's. | ||||||||
943 | if (I->getDebugLoc()) | ||||||||
944 | DL = I->getDebugLoc(); | ||||||||
945 | } | ||||||||
946 | |||||||||
947 | return OptimizationRemarkAnalysis(PassName, RemarkName, DL, CodeRegion); | ||||||||
948 | } | ||||||||
949 | |||||||||
950 | namespace llvm { | ||||||||
951 | |||||||||
952 | /// Return a value for Step multiplied by VF. | ||||||||
953 | Value *createStepForVF(IRBuilderBase &B, Type *Ty, ElementCount VF, | ||||||||
954 | int64_t Step) { | ||||||||
955 | 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", 955, __extension__ __PRETTY_FUNCTION__)); | ||||||||
956 | Constant *StepVal = ConstantInt::get(Ty, Step * VF.getKnownMinValue()); | ||||||||
957 | return VF.isScalable() ? B.CreateVScale(StepVal) : StepVal; | ||||||||
958 | } | ||||||||
959 | |||||||||
960 | /// Return the runtime value for VF. | ||||||||
961 | Value *getRuntimeVF(IRBuilderBase &B, Type *Ty, ElementCount VF) { | ||||||||
962 | Constant *EC = ConstantInt::get(Ty, VF.getKnownMinValue()); | ||||||||
963 | return VF.isScalable() ? B.CreateVScale(EC) : EC; | ||||||||
964 | } | ||||||||
965 | |||||||||
966 | const SCEV *createTripCountSCEV(Type *IdxTy, PredicatedScalarEvolution &PSE) { | ||||||||
967 | const SCEV *BackedgeTakenCount = PSE.getBackedgeTakenCount(); | ||||||||
968 | assert(!isa<SCEVCouldNotCompute>(BackedgeTakenCount) && "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", 968, __extension__ __PRETTY_FUNCTION__)); | ||||||||
969 | |||||||||
970 | ScalarEvolution &SE = *PSE.getSE(); | ||||||||
971 | |||||||||
972 | // The exit count might have the type of i64 while the phi is i32. This can | ||||||||
973 | // happen if we have an induction variable that is sign extended before the | ||||||||
974 | // compare. The only way that we get a backedge taken count is that the | ||||||||
975 | // induction variable was signed and as such will not overflow. In such a case | ||||||||
976 | // truncation is legal. | ||||||||
977 | if (SE.getTypeSizeInBits(BackedgeTakenCount->getType()) > | ||||||||
978 | IdxTy->getPrimitiveSizeInBits()) | ||||||||
979 | BackedgeTakenCount = SE.getTruncateOrNoop(BackedgeTakenCount, IdxTy); | ||||||||
980 | BackedgeTakenCount = SE.getNoopOrZeroExtend(BackedgeTakenCount, IdxTy); | ||||||||
981 | |||||||||
982 | // Get the total trip count from the count by adding 1. | ||||||||
983 | return SE.getAddExpr(BackedgeTakenCount, | ||||||||
984 | SE.getOne(BackedgeTakenCount->getType())); | ||||||||
985 | } | ||||||||
986 | |||||||||
987 | static Value *getRuntimeVFAsFloat(IRBuilderBase &B, Type *FTy, | ||||||||
988 | ElementCount VF) { | ||||||||
989 | 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", 989, __extension__ __PRETTY_FUNCTION__)); | ||||||||
990 | Type *IntTy = IntegerType::get(FTy->getContext(), FTy->getScalarSizeInBits()); | ||||||||
991 | Value *RuntimeVF = getRuntimeVF(B, IntTy, VF); | ||||||||
992 | return B.CreateUIToFP(RuntimeVF, FTy); | ||||||||
993 | } | ||||||||
994 | |||||||||
995 | void reportVectorizationFailure(const StringRef DebugMsg, | ||||||||
996 | const StringRef OREMsg, const StringRef ORETag, | ||||||||
997 | OptimizationRemarkEmitter *ORE, Loop *TheLoop, | ||||||||
998 | Instruction *I) { | ||||||||
999 | LLVM_DEBUG(debugVectorizationMessage("Not vectorizing: ", DebugMsg, I))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { debugVectorizationMessage("Not vectorizing: " , DebugMsg, I); } } while (false); | ||||||||
1000 | LoopVectorizeHints Hints(TheLoop, true /* doesn't matter */, *ORE); | ||||||||
1001 | ORE->emit( | ||||||||
1002 | createLVAnalysis(Hints.vectorizeAnalysisPassName(), ORETag, TheLoop, I) | ||||||||
1003 | << "loop not vectorized: " << OREMsg); | ||||||||
1004 | } | ||||||||
1005 | |||||||||
1006 | void reportVectorizationInfo(const StringRef Msg, const StringRef ORETag, | ||||||||
1007 | OptimizationRemarkEmitter *ORE, Loop *TheLoop, | ||||||||
1008 | Instruction *I) { | ||||||||
1009 | LLVM_DEBUG(debugVectorizationMessage("", Msg, I))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { debugVectorizationMessage("", Msg, I); } } while (false); | ||||||||
1010 | LoopVectorizeHints Hints(TheLoop, true /* doesn't matter */, *ORE); | ||||||||
1011 | ORE->emit( | ||||||||
1012 | createLVAnalysis(Hints.vectorizeAnalysisPassName(), ORETag, TheLoop, I) | ||||||||
1013 | << Msg); | ||||||||
1014 | } | ||||||||
1015 | |||||||||
1016 | } // end namespace llvm | ||||||||
1017 | |||||||||
1018 | #ifndef NDEBUG | ||||||||
1019 | /// \return string containing a file name and a line # for the given loop. | ||||||||
1020 | static std::string getDebugLocString(const Loop *L) { | ||||||||
1021 | std::string Result; | ||||||||
1022 | if (L) { | ||||||||
1023 | raw_string_ostream OS(Result); | ||||||||
1024 | if (const DebugLoc LoopDbgLoc = L->getStartLoc()) | ||||||||
1025 | LoopDbgLoc.print(OS); | ||||||||
1026 | else | ||||||||
1027 | // Just print the module name. | ||||||||
1028 | OS << L->getHeader()->getParent()->getParent()->getModuleIdentifier(); | ||||||||
1029 | OS.flush(); | ||||||||
1030 | } | ||||||||
1031 | return Result; | ||||||||
1032 | } | ||||||||
1033 | #endif | ||||||||
1034 | |||||||||
1035 | void InnerLoopVectorizer::collectPoisonGeneratingRecipes( | ||||||||
1036 | VPTransformState &State) { | ||||||||
1037 | |||||||||
1038 | // Collect recipes in the backward slice of `Root` that may generate a poison | ||||||||
1039 | // value that is used after vectorization. | ||||||||
1040 | SmallPtrSet<VPRecipeBase *, 16> Visited; | ||||||||
1041 | auto collectPoisonGeneratingInstrsInBackwardSlice([&](VPRecipeBase *Root) { | ||||||||
1042 | SmallVector<VPRecipeBase *, 16> Worklist; | ||||||||
1043 | Worklist.push_back(Root); | ||||||||
1044 | |||||||||
1045 | // Traverse the backward slice of Root through its use-def chain. | ||||||||
1046 | while (!Worklist.empty()) { | ||||||||
1047 | VPRecipeBase *CurRec = Worklist.back(); | ||||||||
1048 | Worklist.pop_back(); | ||||||||
1049 | |||||||||
1050 | if (!Visited.insert(CurRec).second) | ||||||||
1051 | continue; | ||||||||
1052 | |||||||||
1053 | // Prune search if we find another recipe generating a widen memory | ||||||||
1054 | // instruction. Widen memory instructions involved in address computation | ||||||||
1055 | // will lead to gather/scatter instructions, which don't need to be | ||||||||
1056 | // handled. | ||||||||
1057 | if (isa<VPWidenMemoryInstructionRecipe>(CurRec) || | ||||||||
1058 | isa<VPInterleaveRecipe>(CurRec) || | ||||||||
1059 | isa<VPScalarIVStepsRecipe>(CurRec) || | ||||||||
1060 | isa<VPCanonicalIVPHIRecipe>(CurRec) || | ||||||||
1061 | isa<VPActiveLaneMaskPHIRecipe>(CurRec)) | ||||||||
1062 | continue; | ||||||||
1063 | |||||||||
1064 | // This recipe contributes to the address computation of a widen | ||||||||
1065 | // load/store. Collect recipe if its underlying instruction has | ||||||||
1066 | // poison-generating flags. | ||||||||
1067 | Instruction *Instr = CurRec->getUnderlyingInstr(); | ||||||||
1068 | if (Instr && Instr->hasPoisonGeneratingFlags()) | ||||||||
1069 | State.MayGeneratePoisonRecipes.insert(CurRec); | ||||||||
1070 | |||||||||
1071 | // Add new definitions to the worklist. | ||||||||
1072 | for (VPValue *operand : CurRec->operands()) | ||||||||
1073 | if (VPRecipeBase *OpDef = operand->getDefiningRecipe()) | ||||||||
1074 | Worklist.push_back(OpDef); | ||||||||
1075 | } | ||||||||
1076 | }); | ||||||||
1077 | |||||||||
1078 | // Traverse all the recipes in the VPlan and collect the poison-generating | ||||||||
1079 | // recipes in the backward slice starting at the address of a VPWidenRecipe or | ||||||||
1080 | // VPInterleaveRecipe. | ||||||||
1081 | auto Iter = vp_depth_first_deep(State.Plan->getEntry()); | ||||||||
1082 | for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(Iter)) { | ||||||||
1083 | for (VPRecipeBase &Recipe : *VPBB) { | ||||||||
1084 | if (auto *WidenRec = dyn_cast<VPWidenMemoryInstructionRecipe>(&Recipe)) { | ||||||||
1085 | Instruction &UnderlyingInstr = WidenRec->getIngredient(); | ||||||||
1086 | VPRecipeBase *AddrDef = WidenRec->getAddr()->getDefiningRecipe(); | ||||||||
1087 | if (AddrDef && WidenRec->isConsecutive() && | ||||||||
1088 | Legal->blockNeedsPredication(UnderlyingInstr.getParent())) | ||||||||
1089 | collectPoisonGeneratingInstrsInBackwardSlice(AddrDef); | ||||||||
1090 | } else if (auto *InterleaveRec = dyn_cast<VPInterleaveRecipe>(&Recipe)) { | ||||||||
1091 | VPRecipeBase *AddrDef = InterleaveRec->getAddr()->getDefiningRecipe(); | ||||||||
1092 | if (AddrDef) { | ||||||||
1093 | // Check if any member of the interleave group needs predication. | ||||||||
1094 | const InterleaveGroup<Instruction> *InterGroup = | ||||||||
1095 | InterleaveRec->getInterleaveGroup(); | ||||||||
1096 | bool NeedPredication = false; | ||||||||
1097 | for (int I = 0, NumMembers = InterGroup->getNumMembers(); | ||||||||
1098 | I < NumMembers; ++I) { | ||||||||
1099 | Instruction *Member = InterGroup->getMember(I); | ||||||||
1100 | if (Member) | ||||||||
1101 | NeedPredication |= | ||||||||
1102 | Legal->blockNeedsPredication(Member->getParent()); | ||||||||
1103 | } | ||||||||
1104 | |||||||||
1105 | if (NeedPredication) | ||||||||
1106 | collectPoisonGeneratingInstrsInBackwardSlice(AddrDef); | ||||||||
1107 | } | ||||||||
1108 | } | ||||||||
1109 | } | ||||||||
1110 | } | ||||||||
1111 | } | ||||||||
1112 | |||||||||
1113 | PHINode *InnerLoopVectorizer::getReductionResumeValue( | ||||||||
1114 | const RecurrenceDescriptor &RdxDesc) { | ||||||||
1115 | auto It = ReductionResumeValues.find(&RdxDesc); | ||||||||
1116 | 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", 1117, __extension__ __PRETTY_FUNCTION__)) | ||||||||
1117 | "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", 1117, __extension__ __PRETTY_FUNCTION__)); | ||||||||
1118 | return It->second; | ||||||||
1119 | } | ||||||||
1120 | |||||||||
1121 | namespace llvm { | ||||||||
1122 | |||||||||
1123 | // Loop vectorization cost-model hints how the scalar epilogue loop should be | ||||||||
1124 | // lowered. | ||||||||
1125 | enum ScalarEpilogueLowering { | ||||||||
1126 | |||||||||
1127 | // The default: allowing scalar epilogues. | ||||||||
1128 | CM_ScalarEpilogueAllowed, | ||||||||
1129 | |||||||||
1130 | // Vectorization with OptForSize: don't allow epilogues. | ||||||||
1131 | CM_ScalarEpilogueNotAllowedOptSize, | ||||||||
1132 | |||||||||
1133 | // A special case of vectorisation with OptForSize: loops with a very small | ||||||||
1134 | // trip count are considered for vectorization under OptForSize, thereby | ||||||||
1135 | // making sure the cost of their loop body is dominant, free of runtime | ||||||||
1136 | // guards and scalar iteration overheads. | ||||||||
1137 | CM_ScalarEpilogueNotAllowedLowTripLoop, | ||||||||
1138 | |||||||||
1139 | // Loop hint predicate indicating an epilogue is undesired. | ||||||||
1140 | CM_ScalarEpilogueNotNeededUsePredicate, | ||||||||
1141 | |||||||||
1142 | // Directive indicating we must either tail fold or not vectorize | ||||||||
1143 | CM_ScalarEpilogueNotAllowedUsePredicate | ||||||||
1144 | }; | ||||||||
1145 | |||||||||
1146 | /// ElementCountComparator creates a total ordering for ElementCount | ||||||||
1147 | /// for the purposes of using it in a set structure. | ||||||||
1148 | struct ElementCountComparator { | ||||||||
1149 | bool operator()(const ElementCount &LHS, const ElementCount &RHS) const { | ||||||||
1150 | return std::make_tuple(LHS.isScalable(), LHS.getKnownMinValue()) < | ||||||||
1151 | std::make_tuple(RHS.isScalable(), RHS.getKnownMinValue()); | ||||||||
1152 | } | ||||||||
1153 | }; | ||||||||
1154 | using ElementCountSet = SmallSet<ElementCount, 16, ElementCountComparator>; | ||||||||
1155 | |||||||||
1156 | /// LoopVectorizationCostModel - estimates the expected speedups due to | ||||||||
1157 | /// vectorization. | ||||||||
1158 | /// In many cases vectorization is not profitable. This can happen because of | ||||||||
1159 | /// a number of reasons. In this class we mainly attempt to predict the | ||||||||
1160 | /// expected speedup/slowdowns due to the supported instruction set. We use the | ||||||||
1161 | /// TargetTransformInfo to query the different backends for the cost of | ||||||||
1162 | /// different operations. | ||||||||
1163 | class LoopVectorizationCostModel { | ||||||||
1164 | public: | ||||||||
1165 | LoopVectorizationCostModel(ScalarEpilogueLowering SEL, Loop *L, | ||||||||
1166 | PredicatedScalarEvolution &PSE, LoopInfo *LI, | ||||||||
1167 | LoopVectorizationLegality *Legal, | ||||||||
1168 | const TargetTransformInfo &TTI, | ||||||||
1169 | const TargetLibraryInfo *TLI, DemandedBits *DB, | ||||||||
1170 | AssumptionCache *AC, | ||||||||
1171 | OptimizationRemarkEmitter *ORE, const Function *F, | ||||||||
1172 | const LoopVectorizeHints *Hints, | ||||||||
1173 | InterleavedAccessInfo &IAI) | ||||||||
1174 | : ScalarEpilogueStatus(SEL), TheLoop(L), PSE(PSE), LI(LI), Legal(Legal), | ||||||||
1175 | TTI(TTI), TLI(TLI), DB(DB), AC(AC), ORE(ORE), TheFunction(F), | ||||||||
1176 | Hints(Hints), InterleaveInfo(IAI) {} | ||||||||
1177 | |||||||||
1178 | /// \return An upper bound for the vectorization factors (both fixed and | ||||||||
1179 | /// scalable). If the factors are 0, vectorization and interleaving should be | ||||||||
1180 | /// avoided up front. | ||||||||
1181 | FixedScalableVFPair computeMaxVF(ElementCount UserVF, unsigned UserIC); | ||||||||
1182 | |||||||||
1183 | /// \return True if runtime checks are required for vectorization, and false | ||||||||
1184 | /// otherwise. | ||||||||
1185 | bool runtimeChecksRequired(); | ||||||||
1186 | |||||||||
1187 | /// \return The most profitable vectorization factor and the cost of that VF. | ||||||||
1188 | /// This method checks every VF in \p CandidateVFs. If UserVF is not ZERO | ||||||||
1189 | /// then this vectorization factor will be selected if vectorization is | ||||||||
1190 | /// possible. | ||||||||
1191 | VectorizationFactor | ||||||||
1192 | selectVectorizationFactor(const ElementCountSet &CandidateVFs); | ||||||||
1193 | |||||||||
1194 | VectorizationFactor | ||||||||
1195 | selectEpilogueVectorizationFactor(const ElementCount MaxVF, | ||||||||
1196 | const LoopVectorizationPlanner &LVP); | ||||||||
1197 | |||||||||
1198 | /// Setup cost-based decisions for user vectorization factor. | ||||||||
1199 | /// \return true if the UserVF is a feasible VF to be chosen. | ||||||||
1200 | bool selectUserVectorizationFactor(ElementCount UserVF) { | ||||||||
1201 | collectUniformsAndScalars(UserVF); | ||||||||
1202 | collectInstsToScalarize(UserVF); | ||||||||
1203 | return expectedCost(UserVF).first.isValid(); | ||||||||
1204 | } | ||||||||
1205 | |||||||||
1206 | /// \return The size (in bits) of the smallest and widest types in the code | ||||||||
1207 | /// that needs to be vectorized. We ignore values that remain scalar such as | ||||||||
1208 | /// 64 bit loop indices. | ||||||||
1209 | std::pair<unsigned, unsigned> getSmallestAndWidestTypes(); | ||||||||
1210 | |||||||||
1211 | /// \return The desired interleave count. | ||||||||
1212 | /// If interleave count has been specified by metadata it will be returned. | ||||||||
1213 | /// Otherwise, the interleave count is computed and returned. VF and LoopCost | ||||||||
1214 | /// are the selected vectorization factor and the cost of the selected VF. | ||||||||
1215 | unsigned selectInterleaveCount(ElementCount VF, InstructionCost LoopCost); | ||||||||
1216 | |||||||||
1217 | /// Memory access instruction may be vectorized in more than one way. | ||||||||
1218 | /// Form of instruction after vectorization depends on cost. | ||||||||
1219 | /// This function takes cost-based decisions for Load/Store instructions | ||||||||
1220 | /// and collects them in a map. This decisions map is used for building | ||||||||
1221 | /// the lists of loop-uniform and loop-scalar instructions. | ||||||||
1222 | /// The calculated cost is saved with widening decision in order to | ||||||||
1223 | /// avoid redundant calculations. | ||||||||
1224 | void setCostBasedWideningDecision(ElementCount VF); | ||||||||
1225 | |||||||||
1226 | /// A struct that represents some properties of the register usage | ||||||||
1227 | /// of a loop. | ||||||||
1228 | struct RegisterUsage { | ||||||||
1229 | /// Holds the number of loop invariant values that are used in the loop. | ||||||||
1230 | /// The key is ClassID of target-provided register class. | ||||||||
1231 | SmallMapVector<unsigned, unsigned, 4> LoopInvariantRegs; | ||||||||
1232 | /// Holds the maximum number of concurrent live intervals in the loop. | ||||||||
1233 | /// The key is ClassID of target-provided register class. | ||||||||
1234 | SmallMapVector<unsigned, unsigned, 4> MaxLocalUsers; | ||||||||
1235 | }; | ||||||||
1236 | |||||||||
1237 | /// \return Returns information about the register usages of the loop for the | ||||||||
1238 | /// given vectorization factors. | ||||||||
1239 | SmallVector<RegisterUsage, 8> | ||||||||
1240 | calculateRegisterUsage(ArrayRef<ElementCount> VFs); | ||||||||
1241 | |||||||||
1242 | /// Collect values we want to ignore in the cost model. | ||||||||
1243 | void collectValuesToIgnore(); | ||||||||
1244 | |||||||||
1245 | /// Collect all element types in the loop for which widening is needed. | ||||||||
1246 | void collectElementTypesForWidening(); | ||||||||
1247 | |||||||||
1248 | /// Split reductions into those that happen in the loop, and those that happen | ||||||||
1249 | /// outside. In loop reductions are collected into InLoopReductionChains. | ||||||||
1250 | void collectInLoopReductions(); | ||||||||
1251 | |||||||||
1252 | /// Returns true if we should use strict in-order reductions for the given | ||||||||
1253 | /// RdxDesc. This is true if the -enable-strict-reductions flag is passed, | ||||||||
1254 | /// the IsOrdered flag of RdxDesc is set and we do not allow reordering | ||||||||
1255 | /// of FP operations. | ||||||||
1256 | bool useOrderedReductions(const RecurrenceDescriptor &RdxDesc) const { | ||||||||
1257 | return !Hints->allowReordering() && RdxDesc.isOrdered(); | ||||||||
1258 | } | ||||||||
1259 | |||||||||
1260 | /// \returns The smallest bitwidth each instruction can be represented with. | ||||||||
1261 | /// The vector equivalents of these instructions should be truncated to this | ||||||||
1262 | /// type. | ||||||||
1263 | const MapVector<Instruction *, uint64_t> &getMinimalBitwidths() const { | ||||||||
1264 | return MinBWs; | ||||||||
1265 | } | ||||||||
1266 | |||||||||
1267 | /// \returns True if it is more profitable to scalarize instruction \p I for | ||||||||
1268 | /// vectorization factor \p VF. | ||||||||
1269 | bool isProfitableToScalarize(Instruction *I, ElementCount VF) const { | ||||||||
1270 | 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", 1271, __extension__ __PRETTY_FUNCTION__)) | ||||||||
1271 | "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", 1271, __extension__ __PRETTY_FUNCTION__)); | ||||||||
1272 | |||||||||
1273 | // Cost model is not run in the VPlan-native path - return conservative | ||||||||
1274 | // result until this changes. | ||||||||
1275 | if (EnableVPlanNativePath) | ||||||||
1276 | return false; | ||||||||
1277 | |||||||||
1278 | auto Scalars = InstsToScalarize.find(VF); | ||||||||
1279 | 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", 1280, __extension__ __PRETTY_FUNCTION__)) | ||||||||
1280 | "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", 1280, __extension__ __PRETTY_FUNCTION__)); | ||||||||
1281 | return Scalars->second.find(I) != Scalars->second.end(); | ||||||||
1282 | } | ||||||||
1283 | |||||||||
1284 | /// Returns true if \p I is known to be uniform after vectorization. | ||||||||
1285 | bool isUniformAfterVectorization(Instruction *I, ElementCount VF) const { | ||||||||
1286 | if (VF.isScalar()) | ||||||||
1287 | return true; | ||||||||
1288 | |||||||||
1289 | // Cost model is not run in the VPlan-native path - return conservative | ||||||||
1290 | // result until this changes. | ||||||||
1291 | if (EnableVPlanNativePath) | ||||||||
1292 | return false; | ||||||||
1293 | |||||||||
1294 | auto UniformsPerVF = Uniforms.find(VF); | ||||||||
1295 | 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", 1296, __extension__ __PRETTY_FUNCTION__)) | ||||||||
1296 | "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", 1296, __extension__ __PRETTY_FUNCTION__)); | ||||||||
1297 | return UniformsPerVF->second.count(I); | ||||||||
1298 | } | ||||||||
1299 | |||||||||
1300 | /// Returns true if \p I is known to be scalar after vectorization. | ||||||||
1301 | bool isScalarAfterVectorization(Instruction *I, ElementCount VF) const { | ||||||||
1302 | if (VF.isScalar()) | ||||||||
1303 | return true; | ||||||||
1304 | |||||||||
1305 | // Cost model is not run in the VPlan-native path - return conservative | ||||||||
1306 | // result until this changes. | ||||||||
1307 | if (EnableVPlanNativePath) | ||||||||
1308 | return false; | ||||||||
1309 | |||||||||
1310 | auto ScalarsPerVF = Scalars.find(VF); | ||||||||
1311 | 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", 1312, __extension__ __PRETTY_FUNCTION__)) | ||||||||
1312 | "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", 1312, __extension__ __PRETTY_FUNCTION__)); | ||||||||
1313 | return ScalarsPerVF->second.count(I); | ||||||||
1314 | } | ||||||||
1315 | |||||||||
1316 | /// \returns True if instruction \p I can be truncated to a smaller bitwidth | ||||||||
1317 | /// for vectorization factor \p VF. | ||||||||
1318 | bool canTruncateToMinimalBitwidth(Instruction *I, ElementCount VF) const { | ||||||||
1319 | return VF.isVector() && MinBWs.find(I) != MinBWs.end() && | ||||||||
1320 | !isProfitableToScalarize(I, VF) && | ||||||||
1321 | !isScalarAfterVectorization(I, VF); | ||||||||
1322 | } | ||||||||
1323 | |||||||||
1324 | /// Decision that was taken during cost calculation for memory instruction. | ||||||||
1325 | enum InstWidening { | ||||||||
1326 | CM_Unknown, | ||||||||
1327 | CM_Widen, // For consecutive accesses with stride +1. | ||||||||
1328 | CM_Widen_Reverse, // For consecutive accesses with stride -1. | ||||||||
1329 | CM_Interleave, | ||||||||
1330 | CM_GatherScatter, | ||||||||
1331 | CM_Scalarize | ||||||||
1332 | }; | ||||||||
1333 | |||||||||
1334 | /// Save vectorization decision \p W and \p Cost taken by the cost model for | ||||||||
1335 | /// instruction \p I and vector width \p VF. | ||||||||
1336 | void setWideningDecision(Instruction *I, ElementCount VF, InstWidening W, | ||||||||
1337 | InstructionCost Cost) { | ||||||||
1338 | 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", 1338, __extension__ __PRETTY_FUNCTION__)); | ||||||||
1339 | WideningDecisions[std::make_pair(I, VF)] = std::make_pair(W, Cost); | ||||||||
1340 | } | ||||||||
1341 | |||||||||
1342 | /// Save vectorization decision \p W and \p Cost taken by the cost model for | ||||||||
1343 | /// interleaving group \p Grp and vector width \p VF. | ||||||||
1344 | void setWideningDecision(const InterleaveGroup<Instruction> *Grp, | ||||||||
1345 | ElementCount VF, InstWidening W, | ||||||||
1346 | InstructionCost Cost) { | ||||||||
1347 | 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", 1347, __extension__ __PRETTY_FUNCTION__)); | ||||||||
1348 | /// Broadcast this decicion to all instructions inside the group. | ||||||||
1349 | /// But the cost will be assigned to one instruction only. | ||||||||
1350 | for (unsigned i = 0; i < Grp->getFactor(); ++i) { | ||||||||
1351 | if (auto *I = Grp->getMember(i)) { | ||||||||
1352 | if (Grp->getInsertPos() == I) | ||||||||
1353 | WideningDecisions[std::make_pair(I, VF)] = std::make_pair(W, Cost); | ||||||||
1354 | else | ||||||||
1355 | WideningDecisions[std::make_pair(I, VF)] = std::make_pair(W, 0); | ||||||||
1356 | } | ||||||||
1357 | } | ||||||||
1358 | } | ||||||||
1359 | |||||||||
1360 | /// Return the cost model decision for the given instruction \p I and vector | ||||||||
1361 | /// width \p VF. Return CM_Unknown if this instruction did not pass | ||||||||
1362 | /// through the cost modeling. | ||||||||
1363 | InstWidening getWideningDecision(Instruction *I, ElementCount VF) const { | ||||||||
1364 | 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", 1364, __extension__ __PRETTY_FUNCTION__)); | ||||||||
1365 | // Cost model is not run in the VPlan-native path - return conservative | ||||||||
1366 | // result until this changes. | ||||||||
1367 | if (EnableVPlanNativePath) | ||||||||
1368 | return CM_GatherScatter; | ||||||||
1369 | |||||||||
1370 | std::pair<Instruction *, ElementCount> InstOnVF = std::make_pair(I, VF); | ||||||||
1371 | auto Itr = WideningDecisions.find(InstOnVF); | ||||||||
1372 | if (Itr == WideningDecisions.end()) | ||||||||
1373 | return CM_Unknown; | ||||||||
1374 | return Itr->second.first; | ||||||||
1375 | } | ||||||||
1376 | |||||||||
1377 | /// Return the vectorization cost for the given instruction \p I and vector | ||||||||
1378 | /// width \p VF. | ||||||||
1379 | InstructionCost getWideningCost(Instruction *I, ElementCount VF) { | ||||||||
1380 | 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", 1380, __extension__ __PRETTY_FUNCTION__)); | ||||||||
1381 | std::pair<Instruction *, ElementCount> InstOnVF = std::make_pair(I, VF); | ||||||||
1382 | 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", 1383, __extension__ __PRETTY_FUNCTION__)) | ||||||||
1383 | "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", 1383, __extension__ __PRETTY_FUNCTION__)); | ||||||||
1384 | return WideningDecisions[InstOnVF].second; | ||||||||
1385 | } | ||||||||
1386 | |||||||||
1387 | /// Return True if instruction \p I is an optimizable truncate whose operand | ||||||||
1388 | /// is an induction variable. Such a truncate will be removed by adding a new | ||||||||
1389 | /// induction variable with the destination type. | ||||||||
1390 | bool isOptimizableIVTruncate(Instruction *I, ElementCount VF) { | ||||||||
1391 | // If the instruction is not a truncate, return false. | ||||||||
1392 | auto *Trunc = dyn_cast<TruncInst>(I); | ||||||||
1393 | if (!Trunc) | ||||||||
1394 | return false; | ||||||||
1395 | |||||||||
1396 | // Get the source and destination types of the truncate. | ||||||||
1397 | Type *SrcTy = ToVectorTy(cast<CastInst>(I)->getSrcTy(), VF); | ||||||||
1398 | Type *DestTy = ToVectorTy(cast<CastInst>(I)->getDestTy(), VF); | ||||||||
1399 | |||||||||
1400 | // If the truncate is free for the given types, return false. Replacing a | ||||||||
1401 | // free truncate with an induction variable would add an induction variable | ||||||||
1402 | // update instruction to each iteration of the loop. We exclude from this | ||||||||
1403 | // check the primary induction variable since it will need an update | ||||||||
1404 | // instruction regardless. | ||||||||
1405 | Value *Op = Trunc->getOperand(0); | ||||||||
1406 | if (Op != Legal->getPrimaryInduction() && TTI.isTruncateFree(SrcTy, DestTy)) | ||||||||
1407 | return false; | ||||||||
1408 | |||||||||
1409 | // If the truncated value is not an induction variable, return false. | ||||||||
1410 | return Legal->isInductionPhi(Op); | ||||||||
1411 | } | ||||||||
1412 | |||||||||
1413 | /// Collects the instructions to scalarize for each predicated instruction in | ||||||||
1414 | /// the loop. | ||||||||
1415 | void collectInstsToScalarize(ElementCount VF); | ||||||||
1416 | |||||||||
1417 | /// Collect Uniform and Scalar values for the given \p VF. | ||||||||
1418 | /// The sets depend on CM decision for Load/Store instructions | ||||||||
1419 | /// that may be vectorized as interleave, gather-scatter or scalarized. | ||||||||
1420 | void collectUniformsAndScalars(ElementCount VF) { | ||||||||
1421 | // Do the analysis once. | ||||||||
1422 | if (VF.isScalar() || Uniforms.find(VF) != Uniforms.end()) | ||||||||
1423 | return; | ||||||||
1424 | setCostBasedWideningDecision(VF); | ||||||||
1425 | collectLoopUniforms(VF); | ||||||||
1426 | collectLoopScalars(VF); | ||||||||
1427 | } | ||||||||
1428 | |||||||||
1429 | /// Returns true if the target machine supports masked store operation | ||||||||
1430 | /// for the given \p DataType and kind of access to \p Ptr. | ||||||||
1431 | bool isLegalMaskedStore(Type *DataType, Value *Ptr, Align Alignment) const { | ||||||||
1432 | return Legal->isConsecutivePtr(DataType, Ptr) && | ||||||||
1433 | TTI.isLegalMaskedStore(DataType, Alignment); | ||||||||
1434 | } | ||||||||
1435 | |||||||||
1436 | /// Returns true if the target machine supports masked load operation | ||||||||
1437 | /// for the given \p DataType and kind of access to \p Ptr. | ||||||||
1438 | bool isLegalMaskedLoad(Type *DataType, Value *Ptr, Align Alignment) const { | ||||||||
1439 | return Legal->isConsecutivePtr(DataType, Ptr) && | ||||||||
1440 | TTI.isLegalMaskedLoad(DataType, Alignment); | ||||||||
1441 | } | ||||||||
1442 | |||||||||
1443 | /// Returns true if the target machine can represent \p V as a masked gather | ||||||||
1444 | /// or scatter operation. | ||||||||
1445 | bool isLegalGatherOrScatter(Value *V, | ||||||||
1446 | ElementCount VF = ElementCount::getFixed(1)) { | ||||||||
1447 | bool LI = isa<LoadInst>(V); | ||||||||
1448 | bool SI = isa<StoreInst>(V); | ||||||||
1449 | if (!LI && !SI) | ||||||||
1450 | return false; | ||||||||
1451 | auto *Ty = getLoadStoreType(V); | ||||||||
1452 | Align Align = getLoadStoreAlignment(V); | ||||||||
1453 | if (VF.isVector()) | ||||||||
1454 | Ty = VectorType::get(Ty, VF); | ||||||||
1455 | return (LI && TTI.isLegalMaskedGather(Ty, Align)) || | ||||||||
1456 | (SI && TTI.isLegalMaskedScatter(Ty, Align)); | ||||||||
1457 | } | ||||||||
1458 | |||||||||
1459 | /// Returns true if the target machine supports all of the reduction | ||||||||
1460 | /// variables found for the given VF. | ||||||||
1461 | bool canVectorizeReductions(ElementCount VF) const { | ||||||||
1462 | return (all_of(Legal->getReductionVars(), [&](auto &Reduction) -> bool { | ||||||||
1463 | const RecurrenceDescriptor &RdxDesc = Reduction.second; | ||||||||
1464 | return TTI.isLegalToVectorizeReduction(RdxDesc, VF); | ||||||||
1465 | })); | ||||||||
1466 | } | ||||||||
1467 | |||||||||
1468 | /// Given costs for both strategies, return true if the scalar predication | ||||||||
1469 | /// lowering should be used for div/rem. This incorporates an override | ||||||||
1470 | /// option so it is not simply a cost comparison. | ||||||||
1471 | bool isDivRemScalarWithPredication(InstructionCost ScalarCost, | ||||||||
1472 | InstructionCost SafeDivisorCost) const { | ||||||||
1473 | switch (ForceSafeDivisor) { | ||||||||
1474 | case cl::BOU_UNSET: | ||||||||
1475 | return ScalarCost < SafeDivisorCost; | ||||||||
1476 | case cl::BOU_TRUE: | ||||||||
1477 | return false; | ||||||||
1478 | case cl::BOU_FALSE: | ||||||||
1479 | return true; | ||||||||
1480 | }; | ||||||||
1481 | llvm_unreachable("impossible case value")::llvm::llvm_unreachable_internal("impossible case value", "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 1481); | ||||||||
1482 | } | ||||||||
1483 | |||||||||
1484 | /// Returns true if \p I is an instruction which requires predication and | ||||||||
1485 | /// for which our chosen predication strategy is scalarization (i.e. we | ||||||||
1486 | /// don't have an alternate strategy such as masking available). | ||||||||
1487 | /// \p VF is the vectorization factor that will be used to vectorize \p I. | ||||||||
1488 | bool isScalarWithPredication(Instruction *I, ElementCount VF) const; | ||||||||
1489 | |||||||||
1490 | /// Returns true if \p I is an instruction that needs to be predicated | ||||||||
1491 | /// at runtime. The result is independent of the predication mechanism. | ||||||||
1492 | /// Superset of instructions that return true for isScalarWithPredication. | ||||||||
1493 | bool isPredicatedInst(Instruction *I) const; | ||||||||
1494 | |||||||||
1495 | /// Return the costs for our two available strategies for lowering a | ||||||||
1496 | /// div/rem operation which requires speculating at least one lane. | ||||||||
1497 | /// First result is for scalarization (will be invalid for scalable | ||||||||
1498 | /// vectors); second is for the safe-divisor strategy. | ||||||||
1499 | std::pair<InstructionCost, InstructionCost> | ||||||||
1500 | getDivRemSpeculationCost(Instruction *I, | ||||||||
1501 | ElementCount VF) const; | ||||||||
1502 | |||||||||
1503 | /// Returns true if \p I is a memory instruction with consecutive memory | ||||||||
1504 | /// access that can be widened. | ||||||||
1505 | bool memoryInstructionCanBeWidened(Instruction *I, ElementCount VF); | ||||||||
1506 | |||||||||
1507 | /// Returns true if \p I is a memory instruction in an interleaved-group | ||||||||
1508 | /// of memory accesses that can be vectorized with wide vector loads/stores | ||||||||
1509 | /// and shuffles. | ||||||||
1510 | bool interleavedAccessCanBeWidened(Instruction *I, ElementCount VF); | ||||||||
1511 | |||||||||
1512 | /// Check if \p Instr belongs to any interleaved access group. | ||||||||
1513 | bool isAccessInterleaved(Instruction *Instr) { | ||||||||
1514 | return InterleaveInfo.isInterleaved(Instr); | ||||||||
1515 | } | ||||||||
1516 | |||||||||
1517 | /// Get the interleaved access group that \p Instr belongs to. | ||||||||
1518 | const InterleaveGroup<Instruction> * | ||||||||
1519 | getInterleavedAccessGroup(Instruction *Instr) { | ||||||||
1520 | return InterleaveInfo.getInterleaveGroup(Instr); | ||||||||
1521 | } | ||||||||
1522 | |||||||||
1523 | /// Returns true if we're required to use a scalar epilogue for at least | ||||||||
1524 | /// the final iteration of the original loop. | ||||||||
1525 | bool requiresScalarEpilogue(ElementCount VF) const { | ||||||||
1526 | if (!isScalarEpilogueAllowed()) | ||||||||
1527 | return false; | ||||||||
1528 | // If we might exit from anywhere but the latch, must run the exiting | ||||||||
1529 | // iteration in scalar form. | ||||||||
1530 | if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) | ||||||||
1531 | return true; | ||||||||
1532 | return VF.isVector() && InterleaveInfo.requiresScalarEpilogue(); | ||||||||
1533 | } | ||||||||
1534 | |||||||||
1535 | /// Returns true if a scalar epilogue is not allowed due to optsize or a | ||||||||
1536 | /// loop hint annotation. | ||||||||
1537 | bool isScalarEpilogueAllowed() const { | ||||||||
1538 | return ScalarEpilogueStatus == CM_ScalarEpilogueAllowed; | ||||||||
1539 | } | ||||||||
1540 | |||||||||
1541 | /// Returns true if all loop blocks should be masked to fold tail loop. | ||||||||
1542 | bool foldTailByMasking() const { return FoldTailByMasking; } | ||||||||
1543 | |||||||||
1544 | /// Returns true if were tail-folding and want to use the active lane mask | ||||||||
1545 | /// for vector loop control flow. | ||||||||
1546 | bool useActiveLaneMaskForControlFlow() const { | ||||||||
1547 | return FoldTailByMasking && | ||||||||
1548 | TTI.emitGetActiveLaneMask() == PredicationStyle::DataAndControlFlow; | ||||||||
1549 | } | ||||||||
1550 | |||||||||
1551 | /// Returns true if the instructions in this block requires predication | ||||||||
1552 | /// for any reason, e.g. because tail folding now requires a predicate | ||||||||
1553 | /// or because the block in the original loop was predicated. | ||||||||
1554 | bool blockNeedsPredicationForAnyReason(BasicBlock *BB) const { | ||||||||
1555 | return foldTailByMasking() || Legal->blockNeedsPredication(BB); | ||||||||
1556 | } | ||||||||
1557 | |||||||||
1558 | /// A SmallMapVector to store the InLoop reduction op chains, mapping phi | ||||||||
1559 | /// nodes to the chain of instructions representing the reductions. Uses a | ||||||||
1560 | /// MapVector to ensure deterministic iteration order. | ||||||||
1561 | using ReductionChainMap = | ||||||||
1562 | SmallMapVector<PHINode *, SmallVector<Instruction *, 4>, 4>; | ||||||||
1563 | |||||||||
1564 | /// Return the chain of instructions representing an inloop reduction. | ||||||||
1565 | const ReductionChainMap &getInLoopReductionChains() const { | ||||||||
1566 | return InLoopReductionChains; | ||||||||
1567 | } | ||||||||
1568 | |||||||||
1569 | /// Returns true if the Phi is part of an inloop reduction. | ||||||||
1570 | bool isInLoopReduction(PHINode *Phi) const { | ||||||||
1571 | return InLoopReductionChains.count(Phi); | ||||||||
1572 | } | ||||||||
1573 | |||||||||
1574 | /// Estimate cost of an intrinsic call instruction CI if it were vectorized | ||||||||
1575 | /// with factor VF. Return the cost of the instruction, including | ||||||||
1576 | /// scalarization overhead if it's needed. | ||||||||
1577 | InstructionCost getVectorIntrinsicCost(CallInst *CI, ElementCount VF) const; | ||||||||
1578 | |||||||||
1579 | /// Estimate cost of a call instruction CI if it were vectorized with factor | ||||||||
1580 | /// VF. Return the cost of the instruction, including scalarization overhead | ||||||||
1581 | /// if it's needed. The flag NeedToScalarize shows if the call needs to be | ||||||||
1582 | /// scalarized - | ||||||||
1583 | /// i.e. either vector version isn't available, or is too expensive. | ||||||||
1584 | InstructionCost getVectorCallCost(CallInst *CI, ElementCount VF, | ||||||||
1585 | bool &NeedToScalarize) const; | ||||||||
1586 | |||||||||
1587 | /// Returns true if the per-lane cost of VectorizationFactor A is lower than | ||||||||
1588 | /// that of B. | ||||||||
1589 | bool isMoreProfitable(const VectorizationFactor &A, | ||||||||
1590 | const VectorizationFactor &B) const; | ||||||||
1591 | |||||||||
1592 | /// Invalidates decisions already taken by the cost model. | ||||||||
1593 | void invalidateCostModelingDecisions() { | ||||||||
1594 | WideningDecisions.clear(); | ||||||||
1595 | Uniforms.clear(); | ||||||||
1596 | Scalars.clear(); | ||||||||
1597 | } | ||||||||
1598 | |||||||||
1599 | /// Convenience function that returns the value of vscale_range iff | ||||||||
1600 | /// vscale_range.min == vscale_range.max or otherwise returns the value | ||||||||
1601 | /// returned by the corresponding TLI method. | ||||||||
1602 | std::optional<unsigned> getVScaleForTuning() const; | ||||||||
1603 | |||||||||
1604 | private: | ||||||||
1605 | unsigned NumPredStores = 0; | ||||||||
1606 | |||||||||
1607 | /// \return An upper bound for the vectorization factors for both | ||||||||
1608 | /// fixed and scalable vectorization, where the minimum-known number of | ||||||||
1609 | /// elements is a power-of-2 larger than zero. If scalable vectorization is | ||||||||
1610 | /// disabled or unsupported, then the scalable part will be equal to | ||||||||
1611 | /// ElementCount::getScalable(0). | ||||||||
1612 | FixedScalableVFPair computeFeasibleMaxVF(unsigned ConstTripCount, | ||||||||
1613 | ElementCount UserVF, | ||||||||
1614 | bool FoldTailByMasking); | ||||||||
1615 | |||||||||
1616 | /// \return the maximized element count based on the targets vector | ||||||||
1617 | /// registers and the loop trip-count, but limited to a maximum safe VF. | ||||||||
1618 | /// This is a helper function of computeFeasibleMaxVF. | ||||||||
1619 | ElementCount getMaximizedVFForTarget(unsigned ConstTripCount, | ||||||||
1620 | unsigned SmallestType, | ||||||||
1621 | unsigned WidestType, | ||||||||
1622 | ElementCount MaxSafeVF, | ||||||||
1623 | bool FoldTailByMasking); | ||||||||
1624 | |||||||||
1625 | /// \return the maximum legal scalable VF, based on the safe max number | ||||||||
1626 | /// of elements. | ||||||||
1627 | ElementCount getMaxLegalScalableVF(unsigned MaxSafeElements); | ||||||||
1628 | |||||||||
1629 | /// The vectorization cost is a combination of the cost itself and a boolean | ||||||||
1630 | /// indicating whether any of the contributing operations will actually | ||||||||
1631 | /// operate on vector values after type legalization in the backend. If this | ||||||||
1632 | /// latter value is false, then all operations will be scalarized (i.e. no | ||||||||
1633 | /// vectorization has actually taken place). | ||||||||
1634 | using VectorizationCostTy = std::pair<InstructionCost, bool>; | ||||||||
1635 | |||||||||
1636 | /// Returns the expected execution cost. The unit of the cost does | ||||||||
1637 | /// not matter because we use the 'cost' units to compare different | ||||||||
1638 | /// vector widths. The cost that is returned is *not* normalized by | ||||||||
1639 | /// the factor width. If \p Invalid is not nullptr, this function | ||||||||
1640 | /// will add a pair(Instruction*, ElementCount) to \p Invalid for | ||||||||
1641 | /// each instruction that has an Invalid cost for the given VF. | ||||||||
1642 | using InstructionVFPair = std::pair<Instruction *, ElementCount>; | ||||||||
1643 | VectorizationCostTy | ||||||||
1644 | expectedCost(ElementCount VF, | ||||||||
1645 | SmallVectorImpl<InstructionVFPair> *Invalid = nullptr); | ||||||||
1646 | |||||||||
1647 | /// Returns the execution time cost of an instruction for a given vector | ||||||||
1648 | /// width. Vector width of one means scalar. | ||||||||
1649 | VectorizationCostTy getInstructionCost(Instruction *I, ElementCount VF); | ||||||||
1650 | |||||||||
1651 | /// The cost-computation logic from getInstructionCost which provides | ||||||||
1652 | /// the vector type as an output parameter. | ||||||||
1653 | InstructionCost getInstructionCost(Instruction *I, ElementCount VF, | ||||||||
1654 | Type *&VectorTy); | ||||||||
1655 | |||||||||
1656 | /// Return the cost of instructions in an inloop reduction pattern, if I is | ||||||||
1657 | /// part of that pattern. | ||||||||
1658 | std::optional<InstructionCost> | ||||||||
1659 | getReductionPatternCost(Instruction *I, ElementCount VF, Type *VectorTy, | ||||||||
1660 | TTI::TargetCostKind CostKind); | ||||||||
1661 | |||||||||
1662 | /// Calculate vectorization cost of memory instruction \p I. | ||||||||
1663 | InstructionCost getMemoryInstructionCost(Instruction *I, ElementCount VF); | ||||||||
1664 | |||||||||
1665 | /// The cost computation for scalarized memory instruction. | ||||||||
1666 | InstructionCost getMemInstScalarizationCost(Instruction *I, ElementCount VF); | ||||||||
1667 | |||||||||
1668 | /// The cost computation for interleaving group of memory instructions. | ||||||||
1669 | InstructionCost getInterleaveGroupCost(Instruction *I, ElementCount VF); | ||||||||
1670 | |||||||||
1671 | /// The cost computation for Gather/Scatter instruction. | ||||||||
1672 | InstructionCost getGatherScatterCost(Instruction *I, ElementCount VF); | ||||||||
1673 | |||||||||
1674 | /// The cost computation for widening instruction \p I with consecutive | ||||||||
1675 | /// memory access. | ||||||||
1676 | InstructionCost getConsecutiveMemOpCost(Instruction *I, ElementCount VF); | ||||||||
1677 | |||||||||
1678 | /// The cost calculation for Load/Store instruction \p I with uniform pointer - | ||||||||
1679 | /// Load: scalar load + broadcast. | ||||||||
1680 | /// Store: scalar store + (loop invariant value stored? 0 : extract of last | ||||||||
1681 | /// element) | ||||||||
1682 | InstructionCost getUniformMemOpCost(Instruction *I, ElementCount VF); | ||||||||
1683 | |||||||||
1684 | /// Estimate the overhead of scalarizing an instruction. This is a | ||||||||
1685 | /// convenience wrapper for the type-based getScalarizationOverhead API. | ||||||||
1686 | InstructionCost getScalarizationOverhead(Instruction *I, ElementCount VF, | ||||||||
1687 | TTI::TargetCostKind CostKind) const; | ||||||||
1688 | |||||||||
1689 | /// Returns true if an artificially high cost for emulated masked memrefs | ||||||||
1690 | /// should be used. | ||||||||
1691 | bool useEmulatedMaskMemRefHack(Instruction *I, ElementCount VF); | ||||||||
1692 | |||||||||
1693 | /// Map of scalar integer values to the smallest bitwidth they can be legally | ||||||||
1694 | /// represented as. The vector equivalents of these values should be truncated | ||||||||
1695 | /// to this type. | ||||||||
1696 | MapVector<Instruction *, uint64_t> MinBWs; | ||||||||
1697 | |||||||||
1698 | /// A type representing the costs for instructions if they were to be | ||||||||
1699 | /// scalarized rather than vectorized. The entries are Instruction-Cost | ||||||||
1700 | /// pairs. | ||||||||
1701 | using ScalarCostsTy = DenseMap<Instruction *, InstructionCost>; | ||||||||
1702 | |||||||||
1703 | /// A set containing all BasicBlocks that are known to present after | ||||||||
1704 | /// vectorization as a predicated block. | ||||||||
1705 | DenseMap<ElementCount, SmallPtrSet<BasicBlock *, 4>> | ||||||||
1706 | PredicatedBBsAfterVectorization; | ||||||||
1707 | |||||||||
1708 | /// Records whether it is allowed to have the original scalar loop execute at | ||||||||
1709 | /// least once. This may be needed as a fallback loop in case runtime | ||||||||
1710 | /// aliasing/dependence checks fail, or to handle the tail/remainder | ||||||||
1711 | /// iterations when the trip count is unknown or doesn't divide by the VF, | ||||||||
1712 | /// or as a peel-loop to handle gaps in interleave-groups. | ||||||||
1713 | /// Under optsize and when the trip count is very small we don't allow any | ||||||||
1714 | /// iterations to execute in the scalar loop. | ||||||||
1715 | ScalarEpilogueLowering ScalarEpilogueStatus = CM_ScalarEpilogueAllowed; | ||||||||
1716 | |||||||||
1717 | /// All blocks of loop are to be masked to fold tail of scalar iterations. | ||||||||
1718 | bool FoldTailByMasking = false; | ||||||||
1719 | |||||||||
1720 | /// A map holding scalar costs for different vectorization factors. The | ||||||||
1721 | /// presence of a cost for an instruction in the mapping indicates that the | ||||||||
1722 | /// instruction will be scalarized when vectorizing with the associated | ||||||||
1723 | /// vectorization factor. The entries are VF-ScalarCostTy pairs. | ||||||||
1724 | DenseMap<ElementCount, ScalarCostsTy> InstsToScalarize; | ||||||||
1725 | |||||||||
1726 | /// Holds the instructions known to be uniform after vectorization. | ||||||||
1727 | /// The data is collected per VF. | ||||||||
1728 | DenseMap<ElementCount, SmallPtrSet<Instruction *, 4>> Uniforms; | ||||||||
1729 | |||||||||
1730 | /// Holds the instructions known to be scalar after vectorization. | ||||||||
1731 | /// The data is collected per VF. | ||||||||
1732 | DenseMap<ElementCount, SmallPtrSet<Instruction *, 4>> Scalars; | ||||||||
1733 | |||||||||
1734 | /// Holds the instructions (address computations) that are forced to be | ||||||||
1735 | /// scalarized. | ||||||||
1736 | DenseMap<ElementCount, SmallPtrSet<Instruction *, 4>> ForcedScalars; | ||||||||
1737 | |||||||||
1738 | /// PHINodes of the reductions that should be expanded in-loop along with | ||||||||
1739 | /// their associated chains of reduction operations, in program order from top | ||||||||
1740 | /// (PHI) to bottom | ||||||||
1741 | ReductionChainMap InLoopReductionChains; | ||||||||
1742 | |||||||||
1743 | /// A Map of inloop reduction operations and their immediate chain operand. | ||||||||
1744 | /// FIXME: This can be removed once reductions can be costed correctly in | ||||||||
1745 | /// vplan. This was added to allow quick lookup to the inloop operations, | ||||||||
1746 | /// without having to loop through InLoopReductionChains. | ||||||||
1747 | DenseMap<Instruction *, Instruction *> InLoopReductionImmediateChains; | ||||||||
1748 | |||||||||
1749 | /// Returns the expected difference in cost from scalarizing the expression | ||||||||
1750 | /// feeding a predicated instruction \p PredInst. The instructions to | ||||||||
1751 | /// scalarize and their scalar costs are collected in \p ScalarCosts. A | ||||||||
1752 | /// non-negative return value implies the expression will be scalarized. | ||||||||
1753 | /// Currently, only single-use chains are considered for scalarization. | ||||||||
1754 | InstructionCost computePredInstDiscount(Instruction *PredInst, | ||||||||
1755 | ScalarCostsTy &ScalarCosts, | ||||||||
1756 | ElementCount VF); | ||||||||
1757 | |||||||||
1758 | /// Collect the instructions that are uniform after vectorization. An | ||||||||
1759 | /// instruction is uniform if we represent it with a single scalar value in | ||||||||
1760 | /// the vectorized loop corresponding to each vector iteration. Examples of | ||||||||
1761 | /// uniform instructions include pointer operands of consecutive or | ||||||||
1762 | /// interleaved memory accesses. Note that although uniformity implies an | ||||||||
1763 | /// instruction will be scalar, the reverse is not true. In general, a | ||||||||
1764 | /// scalarized instruction will be represented by VF scalar values in the | ||||||||
1765 | /// vectorized loop, each corresponding to an iteration of the original | ||||||||
1766 | /// scalar loop. | ||||||||
1767 | void collectLoopUniforms(ElementCount VF); | ||||||||
1768 | |||||||||
1769 | /// Collect the instructions that are scalar after vectorization. An | ||||||||
1770 | /// instruction is scalar if it is known to be uniform or will be scalarized | ||||||||
1771 | /// during vectorization. collectLoopScalars should only add non-uniform nodes | ||||||||
1772 | /// to the list if they are used by a load/store instruction that is marked as | ||||||||
1773 | /// CM_Scalarize. Non-uniform scalarized instructions will be represented by | ||||||||
1774 | /// VF values in the vectorized loop, each corresponding to an iteration of | ||||||||
1775 | /// the original scalar loop. | ||||||||
1776 | void collectLoopScalars(ElementCount VF); | ||||||||
1777 | |||||||||
1778 | /// Keeps cost model vectorization decision and cost for instructions. | ||||||||
1779 | /// Right now it is used for memory instructions only. | ||||||||
1780 | using DecisionList = DenseMap<std::pair<Instruction *, ElementCount>, | ||||||||
1781 | std::pair<InstWidening, InstructionCost>>; | ||||||||
1782 | |||||||||
1783 | DecisionList WideningDecisions; | ||||||||
1784 | |||||||||
1785 | /// Returns true if \p V is expected to be vectorized and it needs to be | ||||||||
1786 | /// extracted. | ||||||||
1787 | bool needsExtract(Value *V, ElementCount VF) const { | ||||||||
1788 | Instruction *I = dyn_cast<Instruction>(V); | ||||||||
1789 | if (VF.isScalar() || !I || !TheLoop->contains(I) || | ||||||||
1790 | TheLoop->isLoopInvariant(I)) | ||||||||
1791 | return false; | ||||||||
1792 | |||||||||
1793 | // Assume we can vectorize V (and hence we need extraction) if the | ||||||||
1794 | // scalars are not computed yet. This can happen, because it is called | ||||||||
1795 | // via getScalarizationOverhead from setCostBasedWideningDecision, before | ||||||||
1796 | // the scalars are collected. That should be a safe assumption in most | ||||||||
1797 | // cases, because we check if the operands have vectorizable types | ||||||||
1798 | // beforehand in LoopVectorizationLegality. | ||||||||
1799 | return Scalars.find(VF) == Scalars.end() || | ||||||||
1800 | !isScalarAfterVectorization(I, VF); | ||||||||
1801 | }; | ||||||||
1802 | |||||||||
1803 | /// Returns a range containing only operands needing to be extracted. | ||||||||
1804 | SmallVector<Value *, 4> filterExtractingOperands(Instruction::op_range Ops, | ||||||||
1805 | ElementCount VF) const { | ||||||||
1806 | return SmallVector<Value *, 4>(make_filter_range( | ||||||||
1807 | Ops, [this, VF](Value *V) { return this->needsExtract(V, VF); })); | ||||||||
1808 | } | ||||||||
1809 | |||||||||
1810 | /// Determines if we have the infrastructure to vectorize loop \p L and its | ||||||||
1811 | /// epilogue, assuming the main loop is vectorized by \p VF. | ||||||||
1812 | bool isCandidateForEpilogueVectorization(const Loop &L, | ||||||||
1813 | const ElementCount VF) const; | ||||||||
1814 | |||||||||
1815 | /// Returns true if epilogue vectorization is considered profitable, and | ||||||||
1816 | /// false otherwise. | ||||||||
1817 | /// \p VF is the vectorization factor chosen for the original loop. | ||||||||
1818 | bool isEpilogueVectorizationProfitable(const ElementCount VF) const; | ||||||||
1819 | |||||||||
1820 | public: | ||||||||
1821 | /// The loop that we evaluate. | ||||||||
1822 | Loop *TheLoop; | ||||||||
1823 | |||||||||
1824 | /// Predicated scalar evolution analysis. | ||||||||
1825 | PredicatedScalarEvolution &PSE; | ||||||||
1826 | |||||||||
1827 | /// Loop Info analysis. | ||||||||
1828 | LoopInfo *LI; | ||||||||
1829 | |||||||||
1830 | /// Vectorization legality. | ||||||||
1831 | LoopVectorizationLegality *Legal; | ||||||||
1832 | |||||||||
1833 | /// Vector target information. | ||||||||
1834 | const TargetTransformInfo &TTI; | ||||||||
1835 | |||||||||
1836 | /// Target Library Info. | ||||||||
1837 | const TargetLibraryInfo *TLI; | ||||||||
1838 | |||||||||
1839 | /// Demanded bits analysis. | ||||||||
1840 | DemandedBits *DB; | ||||||||
1841 | |||||||||
1842 | /// Assumption cache. | ||||||||
1843 | AssumptionCache *AC; | ||||||||
1844 | |||||||||
1845 | /// Interface to emit optimization remarks. | ||||||||
1846 | OptimizationRemarkEmitter *ORE; | ||||||||
1847 | |||||||||
1848 | const Function *TheFunction; | ||||||||
1849 | |||||||||
1850 | /// Loop Vectorize Hint. | ||||||||
1851 | const LoopVectorizeHints *Hints; | ||||||||
1852 | |||||||||
1853 | /// The interleave access information contains groups of interleaved accesses | ||||||||
1854 | /// with the same stride and close to each other. | ||||||||
1855 | InterleavedAccessInfo &InterleaveInfo; | ||||||||
1856 | |||||||||
1857 | /// Values to ignore in the cost model. | ||||||||
1858 | SmallPtrSet<const Value *, 16> ValuesToIgnore; | ||||||||
1859 | |||||||||
1860 | /// Values to ignore in the cost model when VF > 1. | ||||||||
1861 | SmallPtrSet<const Value *, 16> VecValuesToIgnore; | ||||||||
1862 | |||||||||
1863 | /// All element types found in the loop. | ||||||||
1864 | SmallPtrSet<Type *, 16> ElementTypesInLoop; | ||||||||
1865 | |||||||||
1866 | /// Profitable vector factors. | ||||||||
1867 | SmallVector<VectorizationFactor, 8> ProfitableVFs; | ||||||||
1868 | }; | ||||||||
1869 | } // end namespace llvm | ||||||||
1870 | |||||||||
1871 | namespace { | ||||||||
1872 | /// Helper struct to manage generating runtime checks for vectorization. | ||||||||
1873 | /// | ||||||||
1874 | /// The runtime checks are created up-front in temporary blocks to allow better | ||||||||
1875 | /// estimating the cost and un-linked from the existing IR. After deciding to | ||||||||
1876 | /// vectorize, the checks are moved back. If deciding not to vectorize, the | ||||||||
1877 | /// temporary blocks are completely removed. | ||||||||
1878 | class GeneratedRTChecks { | ||||||||
1879 | /// Basic block which contains the generated SCEV checks, if any. | ||||||||
1880 | BasicBlock *SCEVCheckBlock = nullptr; | ||||||||
1881 | |||||||||
1882 | /// The value representing the result of the generated SCEV checks. If it is | ||||||||
1883 | /// nullptr, either no SCEV checks have been generated or they have been used. | ||||||||
1884 | Value *SCEVCheckCond = nullptr; | ||||||||
1885 | |||||||||
1886 | /// Basic block which contains the generated memory runtime checks, if any. | ||||||||
1887 | BasicBlock *MemCheckBlock = nullptr; | ||||||||
1888 | |||||||||
1889 | /// The value representing the result of the generated memory runtime checks. | ||||||||
1890 | /// If it is nullptr, either no memory runtime checks have been generated or | ||||||||
1891 | /// they have been used. | ||||||||
1892 | Value *MemRuntimeCheckCond = nullptr; | ||||||||
1893 | |||||||||
1894 | DominatorTree *DT; | ||||||||
1895 | LoopInfo *LI; | ||||||||
1896 | TargetTransformInfo *TTI; | ||||||||
1897 | |||||||||
1898 | SCEVExpander SCEVExp; | ||||||||
1899 | SCEVExpander MemCheckExp; | ||||||||
1900 | |||||||||
1901 | bool CostTooHigh = false; | ||||||||
1902 | |||||||||
1903 | public: | ||||||||
1904 | GeneratedRTChecks(ScalarEvolution &SE, DominatorTree *DT, LoopInfo *LI, | ||||||||
1905 | TargetTransformInfo *TTI, const DataLayout &DL) | ||||||||
1906 | : DT(DT), LI(LI), TTI(TTI), SCEVExp(SE, DL, "scev.check"), | ||||||||
1907 | MemCheckExp(SE, DL, "scev.check") {} | ||||||||
1908 | |||||||||
1909 | /// Generate runtime checks in SCEVCheckBlock and MemCheckBlock, so we can | ||||||||
1910 | /// accurately estimate the cost of the runtime checks. The blocks are | ||||||||
1911 | /// un-linked from the IR and is added back during vector code generation. If | ||||||||
1912 | /// there is no vector code generation, the check blocks are removed | ||||||||
1913 | /// completely. | ||||||||
1914 | void Create(Loop *L, const LoopAccessInfo &LAI, | ||||||||
1915 | const SCEVPredicate &UnionPred, ElementCount VF, unsigned IC) { | ||||||||
1916 | |||||||||
1917 | // Hard cutoff to limit compile-time increase in case a very large number of | ||||||||
1918 | // runtime checks needs to be generated. | ||||||||
1919 | // TODO: Skip cutoff if the loop is guaranteed to execute, e.g. due to | ||||||||
1920 | // profile info. | ||||||||
1921 | CostTooHigh = | ||||||||
1922 | LAI.getNumRuntimePointerChecks() > VectorizeMemoryCheckThreshold; | ||||||||
1923 | if (CostTooHigh) | ||||||||
1924 | return; | ||||||||
1925 | |||||||||
1926 | BasicBlock *LoopHeader = L->getHeader(); | ||||||||
1927 | BasicBlock *Preheader = L->getLoopPreheader(); | ||||||||
1928 | |||||||||
1929 | // Use SplitBlock to create blocks for SCEV & memory runtime checks to | ||||||||
1930 | // ensure the blocks are properly added to LoopInfo & DominatorTree. Those | ||||||||
1931 | // may be used by SCEVExpander. The blocks will be un-linked from their | ||||||||
1932 | // predecessors and removed from LI & DT at the end of the function. | ||||||||
1933 | if (!UnionPred.isAlwaysTrue()) { | ||||||||
1934 | SCEVCheckBlock = SplitBlock(Preheader, Preheader->getTerminator(), DT, LI, | ||||||||
1935 | nullptr, "vector.scevcheck"); | ||||||||
1936 | |||||||||
1937 | SCEVCheckCond = SCEVExp.expandCodeForPredicate( | ||||||||
1938 | &UnionPred, SCEVCheckBlock->getTerminator()); | ||||||||
1939 | } | ||||||||
1940 | |||||||||
1941 | const auto &RtPtrChecking = *LAI.getRuntimePointerChecking(); | ||||||||
1942 | if (RtPtrChecking.Need) { | ||||||||
1943 | auto *Pred = SCEVCheckBlock ? SCEVCheckBlock : Preheader; | ||||||||
1944 | MemCheckBlock = SplitBlock(Pred, Pred->getTerminator(), DT, LI, nullptr, | ||||||||
1945 | "vector.memcheck"); | ||||||||
1946 | |||||||||
1947 | auto DiffChecks = RtPtrChecking.getDiffChecks(); | ||||||||
1948 | if (DiffChecks) { | ||||||||
1949 | Value *RuntimeVF = nullptr; | ||||||||
1950 | MemRuntimeCheckCond = addDiffRuntimeChecks( | ||||||||
1951 | MemCheckBlock->getTerminator(), *DiffChecks, MemCheckExp, | ||||||||
1952 | [VF, &RuntimeVF](IRBuilderBase &B, unsigned Bits) { | ||||||||
1953 | if (!RuntimeVF) | ||||||||
1954 | RuntimeVF = getRuntimeVF(B, B.getIntNTy(Bits), VF); | ||||||||
1955 | return RuntimeVF; | ||||||||
1956 | }, | ||||||||
1957 | IC); | ||||||||
1958 | } else { | ||||||||
1959 | MemRuntimeCheckCond = | ||||||||
1960 | addRuntimeChecks(MemCheckBlock->getTerminator(), L, | ||||||||
1961 | RtPtrChecking.getChecks(), MemCheckExp); | ||||||||
1962 | } | ||||||||
1963 | 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", 1965, __extension__ __PRETTY_FUNCTION__)) | ||||||||
1964 | "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", 1965, __extension__ __PRETTY_FUNCTION__)) | ||||||||
1965 | "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", 1965, __extension__ __PRETTY_FUNCTION__)); | ||||||||
1966 | } | ||||||||
1967 | |||||||||
1968 | if (!MemCheckBlock && !SCEVCheckBlock) | ||||||||
1969 | return; | ||||||||
1970 | |||||||||
1971 | // Unhook the temporary block with the checks, update various places | ||||||||
1972 | // accordingly. | ||||||||
1973 | if (SCEVCheckBlock) | ||||||||
1974 | SCEVCheckBlock->replaceAllUsesWith(Preheader); | ||||||||
1975 | if (MemCheckBlock) | ||||||||
1976 | MemCheckBlock->replaceAllUsesWith(Preheader); | ||||||||
1977 | |||||||||
1978 | if (SCEVCheckBlock) { | ||||||||
1979 | SCEVCheckBlock->getTerminator()->moveBefore(Preheader->getTerminator()); | ||||||||
1980 | new UnreachableInst(Preheader->getContext(), SCEVCheckBlock); | ||||||||
1981 | Preheader->getTerminator()->eraseFromParent(); | ||||||||
1982 | } | ||||||||
1983 | if (MemCheckBlock) { | ||||||||
1984 | MemCheckBlock->getTerminator()->moveBefore(Preheader->getTerminator()); | ||||||||
1985 | new UnreachableInst(Preheader->getContext(), MemCheckBlock); | ||||||||
1986 | Preheader->getTerminator()->eraseFromParent(); | ||||||||
1987 | } | ||||||||
1988 | |||||||||
1989 | DT->changeImmediateDominator(LoopHeader, Preheader); | ||||||||
1990 | if (MemCheckBlock) { | ||||||||
1991 | DT->eraseNode(MemCheckBlock); | ||||||||
1992 | LI->removeBlock(MemCheckBlock); | ||||||||
1993 | } | ||||||||
1994 | if (SCEVCheckBlock) { | ||||||||
1995 | DT->eraseNode(SCEVCheckBlock); | ||||||||
1996 | LI->removeBlock(SCEVCheckBlock); | ||||||||
1997 | } | ||||||||
1998 | } | ||||||||
1999 | |||||||||
2000 | InstructionCost getCost() { | ||||||||
2001 | if (SCEVCheckBlock || MemCheckBlock) | ||||||||
2002 | LLVM_DEBUG(dbgs() << "Calculating cost of runtime checks:\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "Calculating cost of runtime checks:\n" ; } } while (false); | ||||||||
2003 | |||||||||
2004 | if (CostTooHigh) { | ||||||||
2005 | InstructionCost Cost; | ||||||||
2006 | Cost.setInvalid(); | ||||||||
2007 | LLVM_DEBUG(dbgs() << " number of checks exceeded threshold\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << " number of checks exceeded threshold\n" ; } } while (false); | ||||||||
2008 | return Cost; | ||||||||
2009 | } | ||||||||
2010 | |||||||||
2011 | InstructionCost RTCheckCost = 0; | ||||||||
2012 | if (SCEVCheckBlock) | ||||||||
2013 | for (Instruction &I : *SCEVCheckBlock) { | ||||||||
2014 | if (SCEVCheckBlock->getTerminator() == &I) | ||||||||
2015 | continue; | ||||||||
2016 | InstructionCost C = | ||||||||
2017 | TTI->getInstructionCost(&I, TTI::TCK_RecipThroughput); | ||||||||
2018 | LLVM_DEBUG(dbgs() << " " << C << " for " << I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << " " << C << " for " << I << "\n"; } } while (false); | ||||||||
2019 | RTCheckCost += C; | ||||||||
2020 | } | ||||||||
2021 | if (MemCheckBlock) | ||||||||
2022 | for (Instruction &I : *MemCheckBlock) { | ||||||||
2023 | if (MemCheckBlock->getTerminator() == &I) | ||||||||
2024 | continue; | ||||||||
2025 | InstructionCost C = | ||||||||
2026 | TTI->getInstructionCost(&I, TTI::TCK_RecipThroughput); | ||||||||
2027 | LLVM_DEBUG(dbgs() << " " << C << " for " << I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << " " << C << " for " << I << "\n"; } } while (false); | ||||||||
2028 | RTCheckCost += C; | ||||||||
2029 | } | ||||||||
2030 | |||||||||
2031 | if (SCEVCheckBlock || MemCheckBlock) | ||||||||
2032 | LLVM_DEBUG(dbgs() << "Total cost of runtime checks: " << RTCheckCostdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "Total cost of runtime checks: " << RTCheckCost << "\n"; } } while (false) | ||||||||
2033 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "Total cost of runtime checks: " << RTCheckCost << "\n"; } } while (false); | ||||||||
2034 | |||||||||
2035 | return RTCheckCost; | ||||||||
2036 | } | ||||||||
2037 | |||||||||
2038 | /// Remove the created SCEV & memory runtime check blocks & instructions, if | ||||||||
2039 | /// unused. | ||||||||
2040 | ~GeneratedRTChecks() { | ||||||||
2041 | SCEVExpanderCleaner SCEVCleaner(SCEVExp); | ||||||||
2042 | SCEVExpanderCleaner MemCheckCleaner(MemCheckExp); | ||||||||
2043 | if (!SCEVCheckCond) | ||||||||
2044 | SCEVCleaner.markResultUsed(); | ||||||||
2045 | |||||||||
2046 | if (!MemRuntimeCheckCond) | ||||||||
2047 | MemCheckCleaner.markResultUsed(); | ||||||||
2048 | |||||||||
2049 | if (MemRuntimeCheckCond) { | ||||||||
2050 | auto &SE = *MemCheckExp.getSE(); | ||||||||
2051 | // Memory runtime check generation creates compares that use expanded | ||||||||
2052 | // values. Remove them before running the SCEVExpanderCleaners. | ||||||||
2053 | for (auto &I : make_early_inc_range(reverse(*MemCheckBlock))) { | ||||||||
2054 | if (MemCheckExp.isInsertedInstruction(&I)) | ||||||||
2055 | continue; | ||||||||
2056 | SE.forgetValue(&I); | ||||||||
2057 | I.eraseFromParent(); | ||||||||
2058 | } | ||||||||
2059 | } | ||||||||
2060 | MemCheckCleaner.cleanup(); | ||||||||
2061 | SCEVCleaner.cleanup(); | ||||||||
2062 | |||||||||
2063 | if (SCEVCheckCond) | ||||||||
2064 | SCEVCheckBlock->eraseFromParent(); | ||||||||
2065 | if (MemRuntimeCheckCond) | ||||||||
2066 | MemCheckBlock->eraseFromParent(); | ||||||||
2067 | } | ||||||||
2068 | |||||||||
2069 | /// Adds the generated SCEVCheckBlock before \p LoopVectorPreHeader and | ||||||||
2070 | /// adjusts the branches to branch to the vector preheader or \p Bypass, | ||||||||
2071 | /// depending on the generated condition. | ||||||||
2072 | BasicBlock *emitSCEVChecks(BasicBlock *Bypass, | ||||||||
2073 | BasicBlock *LoopVectorPreHeader, | ||||||||
2074 | BasicBlock *LoopExitBlock) { | ||||||||
2075 | if (!SCEVCheckCond) | ||||||||
2076 | return nullptr; | ||||||||
2077 | |||||||||
2078 | Value *Cond = SCEVCheckCond; | ||||||||
2079 | // Mark the check as used, to prevent it from being removed during cleanup. | ||||||||
2080 | SCEVCheckCond = nullptr; | ||||||||
2081 | if (auto *C = dyn_cast<ConstantInt>(Cond)) | ||||||||
2082 | if (C->isZero()) | ||||||||
2083 | return nullptr; | ||||||||
2084 | |||||||||
2085 | auto *Pred = LoopVectorPreHeader->getSinglePredecessor(); | ||||||||
2086 | |||||||||
2087 | BranchInst::Create(LoopVectorPreHeader, SCEVCheckBlock); | ||||||||
2088 | // Create new preheader for vector loop. | ||||||||
2089 | if (auto *PL = LI->getLoopFor(LoopVectorPreHeader)) | ||||||||
2090 | PL->addBasicBlockToLoop(SCEVCheckBlock, *LI); | ||||||||
2091 | |||||||||
2092 | SCEVCheckBlock->getTerminator()->eraseFromParent(); | ||||||||
2093 | SCEVCheckBlock->moveBefore(LoopVectorPreHeader); | ||||||||
2094 | Pred->getTerminator()->replaceSuccessorWith(LoopVectorPreHeader, | ||||||||
2095 | SCEVCheckBlock); | ||||||||
2096 | |||||||||
2097 | DT->addNewBlock(SCEVCheckBlock, Pred); | ||||||||
2098 | DT->changeImmediateDominator(LoopVectorPreHeader, SCEVCheckBlock); | ||||||||
2099 | |||||||||
2100 | ReplaceInstWithInst(SCEVCheckBlock->getTerminator(), | ||||||||
2101 | BranchInst::Create(Bypass, LoopVectorPreHeader, Cond)); | ||||||||
2102 | return SCEVCheckBlock; | ||||||||
2103 | } | ||||||||
2104 | |||||||||
2105 | /// Adds the generated MemCheckBlock before \p LoopVectorPreHeader and adjusts | ||||||||
2106 | /// the branches to branch to the vector preheader or \p Bypass, depending on | ||||||||
2107 | /// the generated condition. | ||||||||
2108 | BasicBlock *emitMemRuntimeChecks(BasicBlock *Bypass, | ||||||||
2109 | BasicBlock *LoopVectorPreHeader) { | ||||||||
2110 | // Check if we generated code that checks in runtime if arrays overlap. | ||||||||
2111 | if (!MemRuntimeCheckCond) | ||||||||
2112 | return nullptr; | ||||||||
2113 | |||||||||
2114 | auto *Pred = LoopVectorPreHeader->getSinglePredecessor(); | ||||||||
2115 | Pred->getTerminator()->replaceSuccessorWith(LoopVectorPreHeader, | ||||||||
2116 | MemCheckBlock); | ||||||||
2117 | |||||||||
2118 | DT->addNewBlock(MemCheckBlock, Pred); | ||||||||
2119 | DT->changeImmediateDominator(LoopVectorPreHeader, MemCheckBlock); | ||||||||
2120 | MemCheckBlock->moveBefore(LoopVectorPreHeader); | ||||||||
2121 | |||||||||
2122 | if (auto *PL = LI->getLoopFor(LoopVectorPreHeader)) | ||||||||
2123 | PL->addBasicBlockToLoop(MemCheckBlock, *LI); | ||||||||
2124 | |||||||||
2125 | ReplaceInstWithInst( | ||||||||
2126 | MemCheckBlock->getTerminator(), | ||||||||
2127 | BranchInst::Create(Bypass, LoopVectorPreHeader, MemRuntimeCheckCond)); | ||||||||
2128 | MemCheckBlock->getTerminator()->setDebugLoc( | ||||||||
2129 | Pred->getTerminator()->getDebugLoc()); | ||||||||
2130 | |||||||||
2131 | // Mark the check as used, to prevent it from being removed during cleanup. | ||||||||
2132 | MemRuntimeCheckCond = nullptr; | ||||||||
2133 | return MemCheckBlock; | ||||||||
2134 | } | ||||||||
2135 | }; | ||||||||
2136 | } // namespace | ||||||||
2137 | |||||||||
2138 | // Return true if \p OuterLp is an outer loop annotated with hints for explicit | ||||||||
2139 | // vectorization. The loop needs to be annotated with #pragma omp simd | ||||||||
2140 | // simdlen(#) or #pragma clang vectorize(enable) vectorize_width(#). If the | ||||||||
2141 | // vector length information is not provided, vectorization is not considered | ||||||||
2142 | // explicit. Interleave hints are not allowed either. These limitations will be | ||||||||
2143 | // relaxed in the future. | ||||||||
2144 | // Please, note that we are currently forced to abuse the pragma 'clang | ||||||||
2145 | // vectorize' semantics. This pragma provides *auto-vectorization hints* | ||||||||
2146 | // (i.e., LV must check that vectorization is legal) whereas pragma 'omp simd' | ||||||||
2147 | // provides *explicit vectorization hints* (LV can bypass legal checks and | ||||||||
2148 | // assume that vectorization is legal). However, both hints are implemented | ||||||||
2149 | // using the same metadata (llvm.loop.vectorize, processed by | ||||||||
2150 | // LoopVectorizeHints). This will be fixed in the future when the native IR | ||||||||
2151 | // representation for pragma 'omp simd' is introduced. | ||||||||
2152 | static bool isExplicitVecOuterLoop(Loop *OuterLp, | ||||||||
2153 | OptimizationRemarkEmitter *ORE) { | ||||||||
2154 | 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", 2154, __extension__ __PRETTY_FUNCTION__)); | ||||||||
2155 | LoopVectorizeHints Hints(OuterLp, true /*DisableInterleaving*/, *ORE); | ||||||||
2156 | |||||||||
2157 | // Only outer loops with an explicit vectorization hint are supported. | ||||||||
2158 | // Unannotated outer loops are ignored. | ||||||||
2159 | if (Hints.getForce() == LoopVectorizeHints::FK_Undefined) | ||||||||
2160 | return false; | ||||||||
2161 | |||||||||
2162 | Function *Fn = OuterLp->getHeader()->getParent(); | ||||||||
2163 | if (!Hints.allowVectorization(Fn, OuterLp, | ||||||||
2164 | true /*VectorizeOnlyWhenForced*/)) { | ||||||||
2165 | 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); | ||||||||
2166 | return false; | ||||||||
2167 | } | ||||||||
2168 | |||||||||
2169 | if (Hints.getInterleave() > 1) { | ||||||||
2170 | // TODO: Interleave support is future work. | ||||||||
2171 | 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) | ||||||||
2172 | "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); | ||||||||
2173 | Hints.emitRemarkWithHints(); | ||||||||
2174 | return false; | ||||||||
2175 | } | ||||||||
2176 | |||||||||
2177 | return true; | ||||||||
2178 | } | ||||||||
2179 | |||||||||
2180 | static void collectSupportedLoops(Loop &L, LoopInfo *LI, | ||||||||
2181 | OptimizationRemarkEmitter *ORE, | ||||||||
2182 | SmallVectorImpl<Loop *> &V) { | ||||||||
2183 | // Collect inner loops and outer loops without irreducible control flow. For | ||||||||
2184 | // now, only collect outer loops that have explicit vectorization hints. If we | ||||||||
2185 | // are stress testing the VPlan H-CFG construction, we collect the outermost | ||||||||
2186 | // loop of every loop nest. | ||||||||
2187 | if (L.isInnermost() || VPlanBuildStressTest || | ||||||||
2188 | (EnableVPlanNativePath && isExplicitVecOuterLoop(&L, ORE))) { | ||||||||
2189 | LoopBlocksRPO RPOT(&L); | ||||||||
2190 | RPOT.perform(LI); | ||||||||
2191 | if (!containsIrreducibleCFG<const BasicBlock *>(RPOT, *LI)) { | ||||||||
2192 | V.push_back(&L); | ||||||||
2193 | // TODO: Collect inner loops inside marked outer loops in case | ||||||||
2194 | // vectorization fails for the outer loop. Do not invoke | ||||||||
2195 | // 'containsIrreducibleCFG' again for inner loops when the outer loop is | ||||||||
2196 | // already known to be reducible. We can use an inherited attribute for | ||||||||
2197 | // that. | ||||||||
2198 | return; | ||||||||
2199 | } | ||||||||
2200 | } | ||||||||
2201 | for (Loop *InnerL : L) | ||||||||
2202 | collectSupportedLoops(*InnerL, LI, ORE, V); | ||||||||
2203 | } | ||||||||
2204 | |||||||||
2205 | namespace { | ||||||||
2206 | |||||||||
2207 | /// The LoopVectorize Pass. | ||||||||
2208 | struct LoopVectorize : public FunctionPass { | ||||||||
2209 | /// Pass identification, replacement for typeid | ||||||||
2210 | static char ID; | ||||||||
2211 | |||||||||
2212 | LoopVectorizePass Impl; | ||||||||
2213 | |||||||||
2214 | explicit LoopVectorize(bool InterleaveOnlyWhenForced = false, | ||||||||
2215 | bool VectorizeOnlyWhenForced = false) | ||||||||
2216 | : FunctionPass(ID), | ||||||||
2217 | Impl({InterleaveOnlyWhenForced, VectorizeOnlyWhenForced}) { | ||||||||
2218 | initializeLoopVectorizePass(*PassRegistry::getPassRegistry()); | ||||||||
2219 | } | ||||||||
2220 | |||||||||
2221 | bool runOnFunction(Function &F) override { | ||||||||
2222 | if (skipFunction(F)) | ||||||||
2223 | return false; | ||||||||
2224 | |||||||||
2225 | auto *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE(); | ||||||||
2226 | auto *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); | ||||||||
2227 | auto *TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F); | ||||||||
2228 | auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); | ||||||||
2229 | auto *BFI = &getAnalysis<BlockFrequencyInfoWrapperPass>().getBFI(); | ||||||||
2230 | auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>(); | ||||||||
2231 | auto *TLI = TLIP ? &TLIP->getTLI(F) : nullptr; | ||||||||
2232 | auto *AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F); | ||||||||
2233 | auto &LAIs = getAnalysis<LoopAccessLegacyAnalysis>().getLAIs(); | ||||||||
2234 | auto *DB = &getAnalysis<DemandedBitsWrapperPass>().getDemandedBits(); | ||||||||
2235 | auto *ORE = &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE(); | ||||||||
2236 | auto *PSI = &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI(); | ||||||||
2237 | |||||||||
2238 | return Impl | ||||||||
2239 | .runImpl(F, *SE, *LI, *TTI, *DT, *BFI, TLI, *DB, *AC, LAIs, *ORE, PSI) | ||||||||
2240 | .MadeAnyChange; | ||||||||
2241 | } | ||||||||
2242 | |||||||||
2243 | void getAnalysisUsage(AnalysisUsage &AU) const override { | ||||||||
2244 | AU.addRequired<AssumptionCacheTracker>(); | ||||||||
2245 | AU.addRequired<BlockFrequencyInfoWrapperPass>(); | ||||||||
2246 | AU.addRequired<DominatorTreeWrapperPass>(); | ||||||||
2247 | AU.addRequired<LoopInfoWrapperPass>(); | ||||||||
2248 | AU.addRequired<ScalarEvolutionWrapperPass>(); | ||||||||
2249 | AU.addRequired<TargetTransformInfoWrapperPass>(); | ||||||||
2250 | AU.addRequired<LoopAccessLegacyAnalysis>(); | ||||||||
2251 | AU.addRequired<DemandedBitsWrapperPass>(); | ||||||||
2252 | AU.addRequired<OptimizationRemarkEmitterWrapperPass>(); | ||||||||
2253 | AU.addRequired<InjectTLIMappingsLegacy>(); | ||||||||
2254 | |||||||||
2255 | // We currently do not preserve loopinfo/dominator analyses with outer loop | ||||||||
2256 | // vectorization. Until this is addressed, mark these analyses as preserved | ||||||||
2257 | // only for non-VPlan-native path. | ||||||||
2258 | // TODO: Preserve Loop and Dominator analyses for VPlan-native path. | ||||||||
2259 | if (!EnableVPlanNativePath) { | ||||||||
2260 | AU.addPreserved<LoopInfoWrapperPass>(); | ||||||||
2261 | AU.addPreserved<DominatorTreeWrapperPass>(); | ||||||||
2262 | } | ||||||||
2263 | |||||||||
2264 | AU.addPreserved<BasicAAWrapperPass>(); | ||||||||
2265 | AU.addPreserved<GlobalsAAWrapperPass>(); | ||||||||
2266 | AU.addRequired<ProfileSummaryInfoWrapperPass>(); | ||||||||
2267 | } | ||||||||
2268 | }; | ||||||||
2269 | |||||||||
2270 | } // end anonymous namespace | ||||||||
2271 | |||||||||
2272 | //===----------------------------------------------------------------------===// | ||||||||
2273 | // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and | ||||||||
2274 | // LoopVectorizationCostModel and LoopVectorizationPlanner. | ||||||||
2275 | //===----------------------------------------------------------------------===// | ||||||||
2276 | |||||||||
2277 | Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) { | ||||||||
2278 | // We need to place the broadcast of invariant variables outside the loop, | ||||||||
2279 | // but only if it's proven safe to do so. Else, broadcast will be inside | ||||||||
2280 | // vector loop body. | ||||||||
2281 | Instruction *Instr = dyn_cast<Instruction>(V); | ||||||||
2282 | bool SafeToHoist = OrigLoop->isLoopInvariant(V) && | ||||||||
2283 | (!Instr || | ||||||||
2284 | DT->dominates(Instr->getParent(), LoopVectorPreHeader)); | ||||||||
2285 | // Place the code for broadcasting invariant variables in the new preheader. | ||||||||
2286 | IRBuilder<>::InsertPointGuard Guard(Builder); | ||||||||
2287 | if (SafeToHoist) | ||||||||
2288 | Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator()); | ||||||||
2289 | |||||||||
2290 | // Broadcast the scalar into all locations in the vector. | ||||||||
2291 | Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast"); | ||||||||
2292 | |||||||||
2293 | return Shuf; | ||||||||
2294 | } | ||||||||
2295 | |||||||||
2296 | /// This function adds | ||||||||
2297 | /// (StartIdx * Step, (StartIdx + 1) * Step, (StartIdx + 2) * Step, ...) | ||||||||
2298 | /// to each vector element of Val. The sequence starts at StartIndex. | ||||||||
2299 | /// \p Opcode is relevant for FP induction variable. | ||||||||
2300 | static Value *getStepVector(Value *Val, Value *StartIdx, Value *Step, | ||||||||
2301 | Instruction::BinaryOps BinOp, ElementCount VF, | ||||||||
2302 | IRBuilderBase &Builder) { | ||||||||
2303 | 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", 2303, __extension__ __PRETTY_FUNCTION__)); | ||||||||
2304 | |||||||||
2305 | // Create and check the types. | ||||||||
2306 | auto *ValVTy = cast<VectorType>(Val->getType()); | ||||||||
2307 | ElementCount VLen = ValVTy->getElementCount(); | ||||||||
2308 | |||||||||
2309 | Type *STy = Val->getType()->getScalarType(); | ||||||||
2310 | 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", 2311, __extension__ __PRETTY_FUNCTION__)) | ||||||||
2311 | "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", 2311, __extension__ __PRETTY_FUNCTION__)); | ||||||||
2312 | 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", 2312, __extension__ __PRETTY_FUNCTION__)); | ||||||||
2313 | |||||||||
2314 | SmallVector<Constant *, 8> Indices; | ||||||||
2315 | |||||||||
2316 | // Create a vector of consecutive numbers from zero to VF. | ||||||||
2317 | VectorType *InitVecValVTy = ValVTy; | ||||||||
2318 | if (STy->isFloatingPointTy()) { | ||||||||
2319 | Type *InitVecValSTy = | ||||||||
2320 | IntegerType::get(STy->getContext(), STy->getScalarSizeInBits()); | ||||||||
2321 | InitVecValVTy = VectorType::get(InitVecValSTy, VLen); | ||||||||
2322 | } | ||||||||
2323 | Value *InitVec = Builder.CreateStepVector(InitVecValVTy); | ||||||||
2324 | |||||||||
2325 | // Splat the StartIdx | ||||||||
2326 | Value *StartIdxSplat = Builder.CreateVectorSplat(VLen, StartIdx); | ||||||||
2327 | |||||||||
2328 | if (STy->isIntegerTy()) { | ||||||||
2329 | InitVec = Builder.CreateAdd(InitVec, StartIdxSplat); | ||||||||
2330 | Step = Builder.CreateVectorSplat(VLen, Step); | ||||||||
2331 | 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", 2331, __extension__ __PRETTY_FUNCTION__)); | ||||||||
2332 | // FIXME: The newly created binary instructions should contain nsw/nuw | ||||||||
2333 | // flags, which can be found from the original scalar operations. | ||||||||
2334 | Step = Builder.CreateMul(InitVec, Step); | ||||||||
2335 | return Builder.CreateAdd(Val, Step, "induction"); | ||||||||
2336 | } | ||||||||
2337 | |||||||||
2338 | // Floating point induction. | ||||||||
2339 | 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", 2340, __extension__ __PRETTY_FUNCTION__)) | ||||||||
2340 | "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", 2340, __extension__ __PRETTY_FUNCTION__)); | ||||||||
2341 | InitVec = Builder.CreateUIToFP(InitVec, ValVTy); | ||||||||
2342 | InitVec = Builder.CreateFAdd(InitVec, StartIdxSplat); | ||||||||
2343 | |||||||||
2344 | Step = Builder.CreateVectorSplat(VLen, Step); | ||||||||
2345 | Value *MulOp = Builder.CreateFMul(InitVec, Step); | ||||||||
2346 | return Builder.CreateBinOp(BinOp, Val, MulOp, "induction"); | ||||||||
2347 | } | ||||||||
2348 | |||||||||
2349 | /// Compute scalar induction steps. \p ScalarIV is the scalar induction | ||||||||
2350 | /// variable on which to base the steps, \p Step is the size of the step. | ||||||||
2351 | static void buildScalarSteps(Value *ScalarIV, Value *Step, | ||||||||
2352 | const InductionDescriptor &ID, VPValue *Def, | ||||||||
2353 | VPTransformState &State) { | ||||||||
2354 | IRBuilderBase &Builder = State.Builder; | ||||||||
2355 | |||||||||
2356 | // Ensure step has the same type as that of scalar IV. | ||||||||
2357 | Type *ScalarIVTy = ScalarIV->getType()->getScalarType(); | ||||||||
2358 | if (ScalarIVTy != Step->getType()) { | ||||||||
2359 | // TODO: Also use VPDerivedIVRecipe when only the step needs truncating, to | ||||||||
2360 | // avoid separate truncate here. | ||||||||
2361 | 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", 2362, __extension__ __PRETTY_FUNCTION__)) | ||||||||
2362 | "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", 2362, __extension__ __PRETTY_FUNCTION__)); | ||||||||
2363 | Step = State.Builder.CreateTrunc(Step, ScalarIVTy); | ||||||||
2364 | } | ||||||||
2365 | |||||||||
2366 | // We build scalar steps for both integer and floating-point induction | ||||||||
2367 | // variables. Here, we determine the kind of arithmetic we will perform. | ||||||||
2368 | Instruction::BinaryOps AddOp; | ||||||||
2369 | Instruction::BinaryOps MulOp; | ||||||||
2370 | if (ScalarIVTy->isIntegerTy()) { | ||||||||
2371 | AddOp = Instruction::Add; | ||||||||
2372 | MulOp = Instruction::Mul; | ||||||||
2373 | } else { | ||||||||
2374 | AddOp = ID.getInductionOpcode(); | ||||||||
2375 | MulOp = Instruction::FMul; | ||||||||
2376 | } | ||||||||
2377 | |||||||||
2378 | // Determine the number of scalars we need to generate for each unroll | ||||||||
2379 | // iteration. | ||||||||
2380 | bool FirstLaneOnly = vputils::onlyFirstLaneUsed(Def); | ||||||||
2381 | // Compute the scalar steps and save the results in State. | ||||||||
2382 | Type *IntStepTy = IntegerType::get(ScalarIVTy->getContext(), | ||||||||
2383 | ScalarIVTy->getScalarSizeInBits()); | ||||||||
2384 | Type *VecIVTy = nullptr; | ||||||||
2385 | Value *UnitStepVec = nullptr, *SplatStep = nullptr, *SplatIV = nullptr; | ||||||||
2386 | if (!FirstLaneOnly && State.VF.isScalable()) { | ||||||||
2387 | VecIVTy = VectorType::get(ScalarIVTy, State.VF); | ||||||||
2388 | UnitStepVec = | ||||||||
2389 | Builder.CreateStepVector(VectorType::get(IntStepTy, State.VF)); | ||||||||
2390 | SplatStep = Builder.CreateVectorSplat(State.VF, Step); | ||||||||
2391 | SplatIV = Builder.CreateVectorSplat(State.VF, ScalarIV); | ||||||||
2392 | } | ||||||||
2393 | |||||||||
2394 | unsigned StartPart = 0; | ||||||||
2395 | unsigned EndPart = State.UF; | ||||||||
2396 | unsigned StartLane = 0; | ||||||||
2397 | unsigned EndLane = FirstLaneOnly ? 1 : State.VF.getKnownMinValue(); | ||||||||
2398 | if (State.Instance) { | ||||||||
2399 | StartPart = State.Instance->Part; | ||||||||
2400 | EndPart = StartPart + 1; | ||||||||
2401 | StartLane = State.Instance->Lane.getKnownLane(); | ||||||||
2402 | EndLane = StartLane + 1; | ||||||||
2403 | } | ||||||||
2404 | for (unsigned Part = StartPart; Part < EndPart; ++Part) { | ||||||||
2405 | Value *StartIdx0 = createStepForVF(Builder, IntStepTy, State.VF, Part); | ||||||||
2406 | |||||||||
2407 | if (!FirstLaneOnly && State.VF.isScalable()) { | ||||||||
2408 | auto *SplatStartIdx = Builder.CreateVectorSplat(State.VF, StartIdx0); | ||||||||
2409 | auto *InitVec = Builder.CreateAdd(SplatStartIdx, UnitStepVec); | ||||||||
2410 | if (ScalarIVTy->isFloatingPointTy()) | ||||||||
2411 | InitVec = Builder.CreateSIToFP(InitVec, VecIVTy); | ||||||||
2412 | auto *Mul = Builder.CreateBinOp(MulOp, InitVec, SplatStep); | ||||||||
2413 | auto *Add = Builder.CreateBinOp(AddOp, SplatIV, Mul); | ||||||||
2414 | State.set(Def, Add, Part); | ||||||||
2415 | // It's useful to record the lane values too for the known minimum number | ||||||||
2416 | // of elements so we do those below. This improves the code quality when | ||||||||
2417 | // trying to extract the first element, for example. | ||||||||
2418 | } | ||||||||
2419 | |||||||||
2420 | if (ScalarIVTy->isFloatingPointTy()) | ||||||||
2421 | StartIdx0 = Builder.CreateSIToFP(StartIdx0, ScalarIVTy); | ||||||||
2422 | |||||||||
2423 | for (unsigned Lane = StartLane; Lane < EndLane; ++Lane) { | ||||||||
2424 | Value *StartIdx = Builder.CreateBinOp( | ||||||||
2425 | AddOp, StartIdx0, getSignedIntOrFpConstant(ScalarIVTy, Lane)); | ||||||||
2426 | // The step returned by `createStepForVF` is a runtime-evaluated value | ||||||||
2427 | // when VF is scalable. Otherwise, it should be folded into a Constant. | ||||||||
2428 | 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", 2430, __extension__ __PRETTY_FUNCTION__)) | ||||||||
2429 | "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", 2430, __extension__ __PRETTY_FUNCTION__)) | ||||||||
2430 | "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", 2430, __extension__ __PRETTY_FUNCTION__)); | ||||||||
2431 | auto *Mul = Builder.CreateBinOp(MulOp, StartIdx, Step); | ||||||||
2432 | auto *Add = Builder.CreateBinOp(AddOp, ScalarIV, Mul); | ||||||||
2433 | State.set(Def, Add, VPIteration(Part, Lane)); | ||||||||
2434 | } | ||||||||
2435 | } | ||||||||
2436 | } | ||||||||
2437 | |||||||||
2438 | // Generate code for the induction step. Note that induction steps are | ||||||||
2439 | // required to be loop-invariant | ||||||||
2440 | static Value *CreateStepValue(const SCEV *Step, ScalarEvolution &SE, | ||||||||
2441 | Instruction *InsertBefore, | ||||||||
2442 | Loop *OrigLoop = nullptr) { | ||||||||
2443 | const DataLayout &DL = SE.getDataLayout(); | ||||||||
2444 | 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", 2445, __extension__ __PRETTY_FUNCTION__)) | ||||||||
2445 | "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", 2445, __extension__ __PRETTY_FUNCTION__)); | ||||||||
2446 | if (auto *E = dyn_cast<SCEVUnknown>(Step)) | ||||||||
2447 | return E->getValue(); | ||||||||
2448 | |||||||||
2449 | SCEVExpander Exp(SE, DL, "induction"); | ||||||||
2450 | return Exp.expandCodeFor(Step, Step->getType(), InsertBefore); | ||||||||
2451 | } | ||||||||
2452 | |||||||||
2453 | /// Compute the transformed value of Index at offset StartValue using step | ||||||||
2454 | /// StepValue. | ||||||||
2455 | /// For integer induction, returns StartValue + Index * StepValue. | ||||||||
2456 | /// For pointer induction, returns StartValue[Index * StepValue]. | ||||||||
2457 | /// FIXME: The newly created binary instructions should contain nsw/nuw | ||||||||
2458 | /// flags, which can be found from the original scalar operations. | ||||||||
2459 | static Value *emitTransformedIndex(IRBuilderBase &B, Value *Index, | ||||||||
2460 | Value *StartValue, Value *Step, | ||||||||
2461 | const InductionDescriptor &ID) { | ||||||||
2462 | Type *StepTy = Step->getType(); | ||||||||
2463 | Value *CastedIndex = StepTy->isIntegerTy() | ||||||||
2464 | ? B.CreateSExtOrTrunc(Index, StepTy) | ||||||||
2465 | : B.CreateCast(Instruction::SIToFP, Index, StepTy); | ||||||||
2466 | if (CastedIndex != Index) { | ||||||||
2467 | CastedIndex->setName(CastedIndex->getName() + ".cast"); | ||||||||
2468 | Index = CastedIndex; | ||||||||
2469 | } | ||||||||
2470 | |||||||||
2471 | // Note: the IR at this point is broken. We cannot use SE to create any new | ||||||||
2472 | // SCEV and then expand it, hoping that SCEV's simplification will give us | ||||||||
2473 | // a more optimal code. Unfortunately, attempt of doing so on invalid IR may | ||||||||
2474 | // lead to various SCEV crashes. So all we can do is to use builder and rely | ||||||||
2475 | // on InstCombine for future simplifications. Here we handle some trivial | ||||||||
2476 | // cases only. | ||||||||
2477 | auto CreateAdd = [&B](Value *X, Value *Y) { | ||||||||
2478 | 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", 2478, __extension__ __PRETTY_FUNCTION__)); | ||||||||
2479 | if (auto *CX = dyn_cast<ConstantInt>(X)) | ||||||||
2480 | if (CX->isZero()) | ||||||||
2481 | return Y; | ||||||||
2482 | if (auto *CY = dyn_cast<ConstantInt>(Y)) | ||||||||
2483 | if (CY->isZero()) | ||||||||
2484 | return X; | ||||||||
2485 | return B.CreateAdd(X, Y); | ||||||||
2486 | }; | ||||||||
2487 | |||||||||
2488 | // We allow X to be a vector type, in which case Y will potentially be | ||||||||
2489 | // splatted into a vector with the same element count. | ||||||||
2490 | auto CreateMul = [&B](Value *X, Value *Y) { | ||||||||
2491 | 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", 2492, __extension__ __PRETTY_FUNCTION__)) | ||||||||
2492 | "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", 2492, __extension__ __PRETTY_FUNCTION__)); | ||||||||
2493 | if (auto *CX = dyn_cast<ConstantInt>(X)) | ||||||||
2494 | if (CX->isOne()) | ||||||||
2495 | return Y; | ||||||||
2496 | if (auto *CY = dyn_cast<ConstantInt>(Y)) | ||||||||
2497 | if (CY->isOne()) | ||||||||
2498 | return X; | ||||||||
2499 | VectorType *XVTy = dyn_cast<VectorType>(X->getType()); | ||||||||
2500 | if (XVTy && !isa<VectorType>(Y->getType())) | ||||||||
2501 | Y = B.CreateVectorSplat(XVTy->getElementCount(), Y); | ||||||||
2502 | return B.CreateMul(X, Y); | ||||||||
2503 | }; | ||||||||
2504 | |||||||||
2505 | switch (ID.getKind()) { | ||||||||
2506 | case InductionDescriptor::IK_IntInduction: { | ||||||||
2507 | 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", 2508, __extension__ __PRETTY_FUNCTION__)) | ||||||||
2508 | "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", 2508, __extension__ __PRETTY_FUNCTION__)); | ||||||||
2509 | 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", 2510, __extension__ __PRETTY_FUNCTION__)) | ||||||||
2510 | "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", 2510, __extension__ __PRETTY_FUNCTION__)); | ||||||||
2511 | if (isa<ConstantInt>(Step) && cast<ConstantInt>(Step)->isMinusOne()) | ||||||||
2512 | return B.CreateSub(StartValue, Index); | ||||||||
2513 | auto *Offset = CreateMul(Index, Step); | ||||||||
2514 | return CreateAdd(StartValue, Offset); | ||||||||
2515 | } | ||||||||
2516 | case InductionDescriptor::IK_PtrInduction: { | ||||||||
2517 | 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", 2518, __extension__ __PRETTY_FUNCTION__)) | ||||||||
2518 | "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", 2518, __extension__ __PRETTY_FUNCTION__)); | ||||||||
2519 | return B.CreateGEP(ID.getElementType(), StartValue, CreateMul(Index, Step)); | ||||||||
2520 | } | ||||||||
2521 | case InductionDescriptor::IK_FpInduction: { | ||||||||
2522 | 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", 2523, __extension__ __PRETTY_FUNCTION__)) | ||||||||
2523 | "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", 2523, __extension__ __PRETTY_FUNCTION__)); | ||||||||
2524 | 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", 2524, __extension__ __PRETTY_FUNCTION__)); | ||||||||
2525 | auto InductionBinOp = ID.getInductionBinOp(); | ||||||||
2526 | 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", 2529, __extension__ __PRETTY_FUNCTION__)) | ||||||||
2527 | (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", 2529, __extension__ __PRETTY_FUNCTION__)) | ||||||||
2528 | 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", 2529, __extension__ __PRETTY_FUNCTION__)) | ||||||||
2529 | "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", 2529, __extension__ __PRETTY_FUNCTION__)); | ||||||||
2530 | |||||||||
2531 | Value *MulExp = B.CreateFMul(Step, Index); | ||||||||
2532 | return B.CreateBinOp(InductionBinOp->getOpcode(), StartValue, MulExp, | ||||||||
2533 | "induction"); | ||||||||
2534 | } | ||||||||
2535 | case InductionDescriptor::IK_NoInduction: | ||||||||
2536 | return nullptr; | ||||||||
2537 | } | ||||||||
2538 | llvm_unreachable("invalid enum")::llvm::llvm_unreachable_internal("invalid enum", "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2538); | ||||||||
2539 | } | ||||||||
2540 | |||||||||
2541 | void InnerLoopVectorizer::packScalarIntoVectorValue(VPValue *Def, | ||||||||
2542 | const VPIteration &Instance, | ||||||||
2543 | VPTransformState &State) { | ||||||||
2544 | Value *ScalarInst = State.get(Def, Instance); | ||||||||
2545 | Value *VectorValue = State.get(Def, Instance.Part); | ||||||||
2546 | VectorValue = Builder.CreateInsertElement( | ||||||||
2547 | VectorValue, ScalarInst, | ||||||||
2548 | Instance.Lane.getAsRuntimeExpr(State.Builder, VF)); | ||||||||
2549 | State.set(Def, VectorValue, Instance.Part); | ||||||||
2550 | } | ||||||||
2551 | |||||||||
2552 | // Return whether we allow using masked interleave-groups (for dealing with | ||||||||
2553 | // strided loads/stores that reside in predicated blocks, or for dealing | ||||||||
2554 | // with gaps). | ||||||||
2555 | static bool useMaskedInterleavedAccesses(const TargetTransformInfo &TTI) { | ||||||||
2556 | // If an override option has been passed in for interleaved accesses, use it. | ||||||||
2557 | if (EnableMaskedInterleavedMemAccesses.getNumOccurrences() > 0) | ||||||||
2558 | return EnableMaskedInterleavedMemAccesses; | ||||||||
2559 | |||||||||
2560 | return TTI.enableMaskedInterleavedAccessVectorization(); | ||||||||
2561 | } | ||||||||
2562 | |||||||||
2563 | // Try to vectorize the interleave group that \p Instr belongs to. | ||||||||
2564 | // | ||||||||
2565 | // E.g. Translate following interleaved load group (factor = 3): | ||||||||
2566 | // for (i = 0; i < N; i+=3) { | ||||||||
2567 | // R = Pic[i]; // Member of index 0 | ||||||||
2568 | // G = Pic[i+1]; // Member of index 1 | ||||||||
2569 | // B = Pic[i+2]; // Member of index 2 | ||||||||
2570 | // ... // do something to R, G, B | ||||||||
2571 | // } | ||||||||
2572 | // To: | ||||||||
2573 | // %wide.vec = load <12 x i32> ; Read 4 tuples of R,G,B | ||||||||
2574 | // %R.vec = shuffle %wide.vec, poison, <0, 3, 6, 9> ; R elements | ||||||||
2575 | // %G.vec = shuffle %wide.vec, poison, <1, 4, 7, 10> ; G elements | ||||||||
2576 | // %B.vec = shuffle %wide.vec, poison, <2, 5, 8, 11> ; B elements | ||||||||
2577 | // | ||||||||
2578 | // Or translate following interleaved store group (factor = 3): | ||||||||
2579 | // for (i = 0; i < N; i+=3) { | ||||||||
2580 | // ... do something to R, G, B | ||||||||
2581 | // Pic[i] = R; // Member of index 0 | ||||||||
2582 | // Pic[i+1] = G; // Member of index 1 | ||||||||
2583 | // Pic[i+2] = B; // Member of index 2 | ||||||||
2584 | // } | ||||||||
2585 | // To: | ||||||||
2586 | // %R_G.vec = shuffle %R.vec, %G.vec, <0, 1, 2, ..., 7> | ||||||||
2587 | // %B_U.vec = shuffle %B.vec, poison, <0, 1, 2, 3, u, u, u, u> | ||||||||
2588 | // %interleaved.vec = shuffle %R_G.vec, %B_U.vec, | ||||||||
2589 | // <0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11> ; Interleave R,G,B elements | ||||||||
2590 | // store <12 x i32> %interleaved.vec ; Write 4 tuples of R,G,B | ||||||||
2591 | void InnerLoopVectorizer::vectorizeInterleaveGroup( | ||||||||
2592 | const InterleaveGroup<Instruction> *Group, ArrayRef<VPValue *> VPDefs, | ||||||||
2593 | VPTransformState &State, VPValue *Addr, ArrayRef<VPValue *> StoredValues, | ||||||||
2594 | VPValue *BlockInMask) { | ||||||||
2595 | Instruction *Instr = Group->getInsertPos(); | ||||||||
2596 | const DataLayout &DL = Instr->getModule()->getDataLayout(); | ||||||||
2597 | |||||||||
2598 | // Prepare for the vector type of the interleaved load/store. | ||||||||
2599 | Type *ScalarTy = getLoadStoreType(Instr); | ||||||||
2600 | unsigned InterleaveFactor = Group->getFactor(); | ||||||||
2601 | 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", 2601, __extension__ __PRETTY_FUNCTION__)); | ||||||||
2602 | auto *VecTy = VectorType::get(ScalarTy, VF * InterleaveFactor); | ||||||||
2603 | |||||||||
2604 | // Prepare for the new pointers. | ||||||||
2605 | SmallVector<Value *, 2> AddrParts; | ||||||||
2606 | unsigned Index = Group->getIndex(Instr); | ||||||||
2607 | |||||||||
2608 | // TODO: extend the masked interleaved-group support to reversed access. | ||||||||
2609 | 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", 2610, __extension__ __PRETTY_FUNCTION__)) | ||||||||
2610 | "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", 2610, __extension__ __PRETTY_FUNCTION__)); | ||||||||
2611 | |||||||||
2612 | // If the group is reverse, adjust the index to refer to the last vector lane | ||||||||
2613 | // instead of the first. We adjust the index from the first vector lane, | ||||||||
2614 | // rather than directly getting the pointer for lane VF - 1, because the | ||||||||
2615 | // pointer operand of the interleaved access is supposed to be uniform. For | ||||||||
2616 | // uniform instructions, we're only required to generate a value for the | ||||||||
2617 | // first vector lane in each unroll iteration. | ||||||||
2618 | if (Group->isReverse()) | ||||||||
2619 | Index += (VF.getKnownMinValue() - 1) * Group->getFactor(); | ||||||||
2620 | |||||||||
2621 | for (unsigned Part = 0; Part < UF; Part++) { | ||||||||
2622 | Value *AddrPart = State.get(Addr, VPIteration(Part, 0)); | ||||||||
2623 | State.setDebugLocFromInst(AddrPart); | ||||||||
2624 | |||||||||
2625 | // Notice current instruction could be any index. Need to adjust the address | ||||||||
2626 | // to the member of index 0. | ||||||||
2627 | // | ||||||||
2628 | // E.g. a = A[i+1]; // Member of index 1 (Current instruction) | ||||||||
2629 | // b = A[i]; // Member of index 0 | ||||||||
2630 | // Current pointer is pointed to A[i+1], adjust it to A[i]. | ||||||||
2631 | // | ||||||||
2632 | // E.g. A[i+1] = a; // Member of index 1 | ||||||||
2633 | // A[i] = b; // Member of index 0 | ||||||||
2634 | // A[i+2] = c; // Member of index 2 (Current instruction) | ||||||||
2635 | // Current pointer is pointed to A[i+2], adjust it to A[i]. | ||||||||
2636 | |||||||||
2637 | bool InBounds = false; | ||||||||
2638 | if (auto *gep = dyn_cast<GetElementPtrInst>(AddrPart->stripPointerCasts())) | ||||||||
2639 | InBounds = gep->isInBounds(); | ||||||||
2640 | AddrPart = Builder.CreateGEP(ScalarTy, AddrPart, Builder.getInt32(-Index)); | ||||||||
2641 | cast<GetElementPtrInst>(AddrPart)->setIsInBounds(InBounds); | ||||||||
2642 | |||||||||
2643 | // Cast to the vector pointer type. | ||||||||
2644 | unsigned AddressSpace = AddrPart->getType()->getPointerAddressSpace(); | ||||||||
2645 | Type *PtrTy = VecTy->getPointerTo(AddressSpace); | ||||||||
2646 | AddrParts.push_back(Builder.CreateBitCast(AddrPart, PtrTy)); | ||||||||
2647 | } | ||||||||
2648 | |||||||||
2649 | State.setDebugLocFromInst(Instr); | ||||||||
2650 | Value *PoisonVec = PoisonValue::get(VecTy); | ||||||||
2651 | |||||||||
2652 | Value *MaskForGaps = nullptr; | ||||||||
2653 | if (Group->requiresScalarEpilogue() && !Cost->isScalarEpilogueAllowed()) { | ||||||||
2654 | MaskForGaps = createBitMaskForGaps(Builder, VF.getKnownMinValue(), *Group); | ||||||||
2655 | 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", 2655, __extension__ __PRETTY_FUNCTION__)); | ||||||||
2656 | } | ||||||||
2657 | |||||||||
2658 | // Vectorize the interleaved load group. | ||||||||
2659 | if (isa<LoadInst>(Instr)) { | ||||||||
2660 | // For each unroll part, create a wide load for the group. | ||||||||
2661 | SmallVector<Value *, 2> NewLoads; | ||||||||
2662 | for (unsigned Part = 0; Part < UF; Part++) { | ||||||||
2663 | Instruction *NewLoad; | ||||||||
2664 | if (BlockInMask || MaskForGaps) { | ||||||||
2665 | 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", 2666, __extension__ __PRETTY_FUNCTION__)) | ||||||||
2666 | "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", 2666, __extension__ __PRETTY_FUNCTION__)); | ||||||||
2667 | Value *GroupMask = MaskForGaps; | ||||||||
2668 | if (BlockInMask) { | ||||||||
2669 | Value *BlockInMaskPart = State.get(BlockInMask, Part); | ||||||||
2670 | Value *ShuffledMask = Builder.CreateShuffleVector( | ||||||||
2671 | BlockInMaskPart, | ||||||||
2672 | createReplicatedMask(InterleaveFactor, VF.getKnownMinValue()), | ||||||||
2673 | "interleaved.mask"); | ||||||||
2674 | GroupMask = MaskForGaps | ||||||||
2675 | ? Builder.CreateBinOp(Instruction::And, ShuffledMask, | ||||||||
2676 | MaskForGaps) | ||||||||
2677 | : ShuffledMask; | ||||||||
2678 | } | ||||||||
2679 | NewLoad = | ||||||||
2680 | Builder.CreateMaskedLoad(VecTy, AddrParts[Part], Group->getAlign(), | ||||||||
2681 | GroupMask, PoisonVec, "wide.masked.vec"); | ||||||||
2682 | } | ||||||||
2683 | else | ||||||||
2684 | NewLoad = Builder.CreateAlignedLoad(VecTy, AddrParts[Part], | ||||||||
2685 | Group->getAlign(), "wide.vec"); | ||||||||
2686 | Group->addMetadata(NewLoad); | ||||||||
2687 | NewLoads.push_back(NewLoad); | ||||||||
2688 | } | ||||||||
2689 | |||||||||
2690 | // For each member in the group, shuffle out the appropriate data from the | ||||||||
2691 | // wide loads. | ||||||||
2692 | unsigned J = 0; | ||||||||
2693 | for (unsigned I = 0; I < InterleaveFactor; ++I) { | ||||||||
2694 | Instruction *Member = Group->getMember(I); | ||||||||
2695 | |||||||||
2696 | // Skip the gaps in the group. | ||||||||
2697 | if (!Member) | ||||||||
2698 | continue; | ||||||||
2699 | |||||||||
2700 | auto StrideMask = | ||||||||
2701 | createStrideMask(I, InterleaveFactor, VF.getKnownMinValue()); | ||||||||
2702 | for (unsigned Part = 0; Part < UF; Part++) { | ||||||||
2703 | Value *StridedVec = Builder.CreateShuffleVector( | ||||||||
2704 | NewLoads[Part], StrideMask, "strided.vec"); | ||||||||
2705 | |||||||||
2706 | // If this member has different type, cast the result type. | ||||||||
2707 | if (Member->getType() != ScalarTy) { | ||||||||
2708 | 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", 2708, __extension__ __PRETTY_FUNCTION__)); | ||||||||
2709 | VectorType *OtherVTy = VectorType::get(Member->getType(), VF); | ||||||||
2710 | StridedVec = createBitOrPointerCast(StridedVec, OtherVTy, DL); | ||||||||
2711 | } | ||||||||
2712 | |||||||||
2713 | if (Group->isReverse()) | ||||||||
2714 | StridedVec = Builder.CreateVectorReverse(StridedVec, "reverse"); | ||||||||
2715 | |||||||||
2716 | State.set(VPDefs[J], StridedVec, Part); | ||||||||
2717 | } | ||||||||
2718 | ++J; | ||||||||
2719 | } | ||||||||
2720 | return; | ||||||||
2721 | } | ||||||||
2722 | |||||||||
2723 | // The sub vector type for current instruction. | ||||||||
2724 | auto *SubVT = VectorType::get(ScalarTy, VF); | ||||||||
2725 | |||||||||
2726 | // Vectorize the interleaved store group. | ||||||||
2727 | MaskForGaps = createBitMaskForGaps(Builder, VF.getKnownMinValue(), *Group); | ||||||||
2728 | 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", 2729, __extension__ __PRETTY_FUNCTION__)) | ||||||||
2729 | "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", 2729, __extension__ __PRETTY_FUNCTION__)); | ||||||||
2730 | 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", 2731, __extension__ __PRETTY_FUNCTION__)) | ||||||||
2731 | "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", 2731, __extension__ __PRETTY_FUNCTION__)); | ||||||||
2732 | for (unsigned Part = 0; Part < UF; Part++) { | ||||||||
2733 | // Collect the stored vector from each member. | ||||||||
2734 | SmallVector<Value *, 4> StoredVecs; | ||||||||
2735 | unsigned StoredIdx = 0; | ||||||||
2736 | for (unsigned i = 0; i < InterleaveFactor; i++) { | ||||||||
2737 | 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", 2738, __extension__ __PRETTY_FUNCTION__)) | ||||||||
2738 | "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", 2738, __extension__ __PRETTY_FUNCTION__)); | ||||||||
2739 | Instruction *Member = Group->getMember(i); | ||||||||
2740 | |||||||||
2741 | // Skip the gaps in the group. | ||||||||
2742 | if (!Member) { | ||||||||
2743 | Value *Undef = PoisonValue::get(SubVT); | ||||||||
2744 | StoredVecs.push_back(Undef); | ||||||||
2745 | continue; | ||||||||
2746 | } | ||||||||
2747 | |||||||||
2748 | Value *StoredVec = State.get(StoredValues[StoredIdx], Part); | ||||||||
2749 | ++StoredIdx; | ||||||||
2750 | |||||||||
2751 | if (Group->isReverse()) | ||||||||
2752 | StoredVec = Builder.CreateVectorReverse(StoredVec, "reverse"); | ||||||||
2753 | |||||||||
2754 | // If this member has different type, cast it to a unified type. | ||||||||
2755 | |||||||||
2756 | if (StoredVec->getType() != SubVT) | ||||||||
2757 | StoredVec = createBitOrPointerCast(StoredVec, SubVT, DL); | ||||||||
2758 | |||||||||
2759 | StoredVecs.push_back(StoredVec); | ||||||||
2760 | } | ||||||||
2761 | |||||||||
2762 | // Concatenate all vectors into a wide vector. | ||||||||
2763 | Value *WideVec = concatenateVectors(Builder, StoredVecs); | ||||||||
2764 | |||||||||
2765 | // Interleave the elements in the wide vector. | ||||||||
2766 | Value *IVec = Builder.CreateShuffleVector( | ||||||||
2767 | WideVec, createInterleaveMask(VF.getKnownMinValue(), InterleaveFactor), | ||||||||
2768 | "interleaved.vec"); | ||||||||
2769 | |||||||||
2770 | Instruction *NewStoreInstr; | ||||||||
2771 | if (BlockInMask || MaskForGaps) { | ||||||||
2772 | Value *GroupMask = MaskForGaps; | ||||||||
2773 | if (BlockInMask) { | ||||||||
2774 | Value *BlockInMaskPart = State.get(BlockInMask, Part); | ||||||||
2775 | Value *ShuffledMask = Builder.CreateShuffleVector( | ||||||||
2776 | BlockInMaskPart, | ||||||||
2777 | createReplicatedMask(InterleaveFactor, VF.getKnownMinValue()), | ||||||||
2778 | "interleaved.mask"); | ||||||||
2779 | GroupMask = MaskForGaps ? Builder.CreateBinOp(Instruction::And, | ||||||||
2780 | ShuffledMask, MaskForGaps) | ||||||||
2781 | : ShuffledMask; | ||||||||
2782 | } | ||||||||
2783 | NewStoreInstr = Builder.CreateMaskedStore(IVec, AddrParts[Part], | ||||||||
2784 | Group->getAlign(), GroupMask); | ||||||||
2785 | } else | ||||||||
2786 | NewStoreInstr = | ||||||||
2787 | Builder.CreateAlignedStore(IVec, AddrParts[Part], Group->getAlign()); | ||||||||
2788 | |||||||||
2789 | Group->addMetadata(NewStoreInstr); | ||||||||
2790 | } | ||||||||
2791 | } | ||||||||
2792 | |||||||||
2793 | void InnerLoopVectorizer::scalarizeInstruction(const Instruction *Instr, | ||||||||
2794 | VPReplicateRecipe *RepRecipe, | ||||||||
2795 | const VPIteration &Instance, | ||||||||
2796 | bool IfPredicateInstr, | ||||||||
2797 | VPTransformState &State) { | ||||||||
2798 | 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", 2798, __extension__ __PRETTY_FUNCTION__)); | ||||||||
2799 | |||||||||
2800 | // llvm.experimental.noalias.scope.decl intrinsics must only be duplicated for | ||||||||
2801 | // the first lane and part. | ||||||||
2802 | if (isa<NoAliasScopeDeclInst>(Instr)) | ||||||||
2803 | if (!Instance.isFirstIteration()) | ||||||||
2804 | return; | ||||||||
2805 | |||||||||
2806 | // Does this instruction return a value ? | ||||||||
2807 | bool IsVoidRetTy = Instr->getType()->isVoidTy(); | ||||||||
2808 | |||||||||
2809 | Instruction *Cloned = Instr->clone(); | ||||||||
2810 | if (!IsVoidRetTy) | ||||||||
2811 | Cloned->setName(Instr->getName() + ".cloned"); | ||||||||
2812 | |||||||||
2813 | // If the scalarized instruction contributes to the address computation of a | ||||||||
2814 | // widen masked load/store which was in a basic block that needed predication | ||||||||
2815 | // and is not predicated after vectorization, we can't propagate | ||||||||
2816 | // poison-generating flags (nuw/nsw, exact, inbounds, etc.). The scalarized | ||||||||
2817 | // instruction could feed a poison value to the base address of the widen | ||||||||
2818 | // load/store. | ||||||||
2819 | if (State.MayGeneratePoisonRecipes.contains(RepRecipe)) | ||||||||
2820 | Cloned->dropPoisonGeneratingFlags(); | ||||||||
2821 | |||||||||
2822 | if (Instr->getDebugLoc()) | ||||||||
2823 | State.setDebugLocFromInst(Instr); | ||||||||
2824 | |||||||||
2825 | // Replace the operands of the cloned instructions with their scalar | ||||||||
2826 | // equivalents in the new loop. | ||||||||
2827 | for (const auto &I : enumerate(RepRecipe->operands())) { | ||||||||
2828 | auto InputInstance = Instance; | ||||||||
2829 | VPValue *Operand = I.value(); | ||||||||
2830 | if (vputils::isUniformAfterVectorization(Operand)) | ||||||||
2831 | InputInstance.Lane = VPLane::getFirstLane(); | ||||||||
2832 | Cloned->setOperand(I.index(), State.get(Operand, InputInstance)); | ||||||||
2833 | } | ||||||||
2834 | State.addNewMetadata(Cloned, Instr); | ||||||||
2835 | |||||||||
2836 | // Place the cloned scalar in the new loop. | ||||||||
2837 | State.Builder.Insert(Cloned); | ||||||||
2838 | |||||||||
2839 | State.set(RepRecipe, Cloned, Instance); | ||||||||
2840 | |||||||||
2841 | // If we just cloned a new assumption, add it the assumption cache. | ||||||||
2842 | if (auto *II = dyn_cast<AssumeInst>(Cloned)) | ||||||||
2843 | AC->registerAssumption(II); | ||||||||
2844 | |||||||||
2845 | // End if-block. | ||||||||
2846 | if (IfPredicateInstr) | ||||||||
2847 | PredicatedInstructions.push_back(Cloned); | ||||||||
2848 | } | ||||||||
2849 | |||||||||
2850 | Value *InnerLoopVectorizer::getOrCreateTripCount(BasicBlock *InsertBlock) { | ||||||||
2851 | if (TripCount) | ||||||||
2852 | return TripCount; | ||||||||
2853 | |||||||||
2854 | assert(InsertBlock)(static_cast <bool> (InsertBlock) ? void (0) : __assert_fail ("InsertBlock", "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2854, __extension__ __PRETTY_FUNCTION__)); | ||||||||
2855 | IRBuilder<> Builder(InsertBlock->getTerminator()); | ||||||||
2856 | // Find the loop boundaries. | ||||||||
2857 | Type *IdxTy = Legal->getWidestInductionType(); | ||||||||
2858 | 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", 2858, __extension__ __PRETTY_FUNCTION__)); | ||||||||
2859 | const SCEV *ExitCount = createTripCountSCEV(IdxTy, PSE); | ||||||||
2860 | |||||||||
2861 | const DataLayout &DL = InsertBlock->getModule()->getDataLayout(); | ||||||||
2862 | |||||||||
2863 | // Expand the trip count and place the new instructions in the preheader. | ||||||||
2864 | // Notice that the pre-header does not change, only the loop body. | ||||||||
2865 | SCEVExpander Exp(*PSE.getSE(), DL, "induction"); | ||||||||
2866 | |||||||||
2867 | // Count holds the overall loop count (N). | ||||||||
2868 | TripCount = Exp.expandCodeFor(ExitCount, ExitCount->getType(), | ||||||||
2869 | InsertBlock->getTerminator()); | ||||||||
2870 | |||||||||
2871 | if (TripCount->getType()->isPointerTy()) | ||||||||
2872 | TripCount = | ||||||||
2873 | CastInst::CreatePointerCast(TripCount, IdxTy, "exitcount.ptrcnt.to.int", | ||||||||
2874 | InsertBlock->getTerminator()); | ||||||||
2875 | |||||||||
2876 | return TripCount; | ||||||||
2877 | } | ||||||||
2878 | |||||||||
2879 | Value * | ||||||||
2880 | InnerLoopVectorizer::getOrCreateVectorTripCount(BasicBlock *InsertBlock) { | ||||||||
2881 | if (VectorTripCount) | ||||||||
2882 | return VectorTripCount; | ||||||||
2883 | |||||||||
2884 | Value *TC = getOrCreateTripCount(InsertBlock); | ||||||||
2885 | IRBuilder<> Builder(InsertBlock->getTerminator()); | ||||||||
2886 | |||||||||
2887 | Type *Ty = TC->getType(); | ||||||||
2888 | // This is where we can make the step a runtime constant. | ||||||||
2889 | Value *Step = createStepForVF(Builder, Ty, VF, UF); | ||||||||
2890 | |||||||||
2891 | // If the tail is to be folded by masking, round the number of iterations N | ||||||||
2892 | // up to a multiple of Step instead of rounding down. This is done by first | ||||||||
2893 | // adding Step-1 and then rounding down. Note that it's ok if this addition | ||||||||
2894 | // overflows: the vector induction variable will eventually wrap to zero given | ||||||||
2895 | // that it starts at zero and its Step is a power of two; the loop will then | ||||||||
2896 | // exit, with the last early-exit vector comparison also producing all-true. | ||||||||
2897 | // For scalable vectors the VF is not guaranteed to be a power of 2, but this | ||||||||
2898 | // is accounted for in emitIterationCountCheck that adds an overflow check. | ||||||||
2899 | if (Cost->foldTailByMasking()) { | ||||||||
2900 | 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", 2901, __extension__ __PRETTY_FUNCTION__)) | ||||||||
2901 | "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", 2901, __extension__ __PRETTY_FUNCTION__)); | ||||||||
2902 | Value *NumLanes = getRuntimeVF(Builder, Ty, VF * UF); | ||||||||
2903 | TC = Builder.CreateAdd( | ||||||||
2904 | TC, Builder.CreateSub(NumLanes, ConstantInt::get(Ty, 1)), "n.rnd.up"); | ||||||||
2905 | } | ||||||||
2906 | |||||||||
2907 | // Now we need to generate the expression for the part of the loop that the | ||||||||
2908 | // vectorized body will execute. This is equal to N - (N % Step) if scalar | ||||||||
2909 | // iterations are not required for correctness, or N - Step, otherwise. Step | ||||||||
2910 | // is equal to the vectorization factor (number of SIMD elements) times the | ||||||||
2911 | // unroll factor (number of SIMD instructions). | ||||||||
2912 | Value *R = Builder.CreateURem(TC, Step, "n.mod.vf"); | ||||||||
2913 | |||||||||
2914 | // There are cases where we *must* run at least one iteration in the remainder | ||||||||
2915 | // loop. See the cost model for when this can happen. If the step evenly | ||||||||
2916 | // divides the trip count, we set the remainder to be equal to the step. If | ||||||||
2917 | // the step does not evenly divide the trip count, no adjustment is necessary | ||||||||
2918 | // since there will already be scalar iterations. Note that the minimum | ||||||||
2919 | // iterations check ensures that N >= Step. | ||||||||
2920 | if (Cost->requiresScalarEpilogue(VF)) { | ||||||||
2921 | auto *IsZero = Builder.CreateICmpEQ(R, ConstantInt::get(R->getType(), 0)); | ||||||||
2922 | R = Builder.CreateSelect(IsZero, Step, R); | ||||||||
2923 | } | ||||||||
2924 | |||||||||
2925 | VectorTripCount = Builder.CreateSub(TC, R, "n.vec"); | ||||||||
2926 | |||||||||
2927 | return VectorTripCount; | ||||||||
2928 | } | ||||||||
2929 | |||||||||
2930 | Value *InnerLoopVectorizer::createBitOrPointerCast(Value *V, VectorType *DstVTy, | ||||||||
2931 | const DataLayout &DL) { | ||||||||
2932 | // Verify that V is a vector type with same number of elements as DstVTy. | ||||||||
2933 | auto *DstFVTy = cast<FixedVectorType>(DstVTy); | ||||||||
2934 | unsigned VF = DstFVTy->getNumElements(); | ||||||||
2935 | auto *SrcVecTy = cast<FixedVectorType>(V->getType()); | ||||||||
2936 | 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", 2936, __extension__ __PRETTY_FUNCTION__)); | ||||||||
2937 | Type *SrcElemTy = SrcVecTy->getElementType(); | ||||||||
2938 | Type *DstElemTy = DstFVTy->getElementType(); | ||||||||
2939 | 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", 2940, __extension__ __PRETTY_FUNCTION__)) | ||||||||
2940 | "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", 2940, __extension__ __PRETTY_FUNCTION__)); | ||||||||
2941 | |||||||||
2942 | // Do a direct cast if element types are castable. | ||||||||
2943 | if (CastInst::isBitOrNoopPointerCastable(SrcElemTy, DstElemTy, DL)) { | ||||||||
2944 | return Builder.CreateBitOrPointerCast(V, DstFVTy); | ||||||||
2945 | } | ||||||||
2946 | // V cannot be directly casted to desired vector type. | ||||||||
2947 | // May happen when V is a floating point vector but DstVTy is a vector of | ||||||||
2948 | // pointers or vice-versa. Handle this using a two-step bitcast using an | ||||||||
2949 | // intermediate Integer type for the bitcast i.e. Ptr <-> Int <-> Float. | ||||||||
2950 | 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", 2951, __extension__ __PRETTY_FUNCTION__)) | ||||||||
2951 | "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", 2951, __extension__ __PRETTY_FUNCTION__)); | ||||||||
2952 | 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", 2953, __extension__ __PRETTY_FUNCTION__)) | ||||||||
2953 | "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", 2953, __extension__ __PRETTY_FUNCTION__)); | ||||||||
2954 | Type *IntTy = | ||||||||
2955 | IntegerType::getIntNTy(V->getContext(), DL.getTypeSizeInBits(SrcElemTy)); | ||||||||
2956 | auto *VecIntTy = FixedVectorType::get(IntTy, VF); | ||||||||
2957 | Value *CastVal = Builder.CreateBitOrPointerCast(V, VecIntTy); | ||||||||
2958 | return Builder.CreateBitOrPointerCast(CastVal, DstFVTy); | ||||||||
2959 | } | ||||||||
2960 | |||||||||
2961 | void InnerLoopVectorizer::emitIterationCountCheck(BasicBlock *Bypass) { | ||||||||
2962 | Value *Count = getOrCreateTripCount(LoopVectorPreHeader); | ||||||||
2963 | // Reuse existing vector loop preheader for TC checks. | ||||||||
2964 | // Note that new preheader block is generated for vector loop. | ||||||||
2965 | BasicBlock *const TCCheckBlock = LoopVectorPreHeader; | ||||||||
2966 | IRBuilder<> Builder(TCCheckBlock->getTerminator()); | ||||||||
2967 | |||||||||
2968 | // Generate code to check if the loop's trip count is less than VF * UF, or | ||||||||
2969 | // equal to it in case a scalar epilogue is required; this implies that the | ||||||||
2970 | // vector trip count is zero. This check also covers the case where adding one | ||||||||
2971 | // to the backedge-taken count overflowed leading to an incorrect trip count | ||||||||
2972 | // of zero. In this case we will also jump to the scalar loop. | ||||||||
2973 | auto P = Cost->requiresScalarEpilogue(VF) ? ICmpInst::ICMP_ULE | ||||||||
2974 | : ICmpInst::ICMP_ULT; | ||||||||
2975 | |||||||||
2976 | // If tail is to be folded, vector loop takes care of all iterations. | ||||||||
2977 | Type *CountTy = Count->getType(); | ||||||||
2978 | Value *CheckMinIters = Builder.getFalse(); | ||||||||
2979 | auto CreateStep = [&]() -> Value * { | ||||||||
2980 | // Create step with max(MinProTripCount, UF * VF). | ||||||||
2981 | if (UF * VF.getKnownMinValue() >= MinProfitableTripCount.getKnownMinValue()) | ||||||||
2982 | return createStepForVF(Builder, CountTy, VF, UF); | ||||||||
2983 | |||||||||
2984 | Value *MinProfTC = | ||||||||
2985 | createStepForVF(Builder, CountTy, MinProfitableTripCount, 1); | ||||||||
2986 | if (!VF.isScalable()) | ||||||||
2987 | return MinProfTC; | ||||||||
2988 | return Builder.CreateBinaryIntrinsic( | ||||||||
2989 | Intrinsic::umax, MinProfTC, createStepForVF(Builder, CountTy, VF, UF)); | ||||||||
2990 | }; | ||||||||
2991 | |||||||||
2992 | if (!Cost->foldTailByMasking()) | ||||||||
2993 | CheckMinIters = | ||||||||
2994 | Builder.CreateICmp(P, Count, CreateStep(), "min.iters.check"); | ||||||||
2995 | else if (VF.isScalable()) { | ||||||||
2996 | // vscale is not necessarily a power-of-2, which means we cannot guarantee | ||||||||
2997 | // an overflow to zero when updating induction variables and so an | ||||||||
2998 | // additional overflow check is required before entering the vector loop. | ||||||||
2999 | |||||||||
3000 | // Get the maximum unsigned value for the type. | ||||||||
3001 | Value *MaxUIntTripCount = | ||||||||
3002 | ConstantInt::get(CountTy, cast<IntegerType>(CountTy)->getMask()); | ||||||||
3003 | Value *LHS = Builder.CreateSub(MaxUIntTripCount, Count); | ||||||||
3004 | |||||||||
3005 | // Don't execute the vector loop if (UMax - n) < (VF * UF). | ||||||||
3006 | CheckMinIters = Builder.CreateICmp(ICmpInst::ICMP_ULT, LHS, CreateStep()); | ||||||||
3007 | } | ||||||||
3008 | |||||||||
3009 | // Create new preheader for vector loop. | ||||||||
3010 | LoopVectorPreHeader = | ||||||||
3011 | SplitBlock(TCCheckBlock, TCCheckBlock->getTerminator(), DT, LI, nullptr, | ||||||||
3012 | "vector.ph"); | ||||||||
3013 | |||||||||
3014 | 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", 3016, __extension__ __PRETTY_FUNCTION__)) | ||||||||
3015 | 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", 3016, __extension__ __PRETTY_FUNCTION__)) | ||||||||
3016 | "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", 3016, __extension__ __PRETTY_FUNCTION__)); | ||||||||
3017 | |||||||||
3018 | // Update dominator for Bypass & LoopExit (if needed). | ||||||||
3019 | DT->changeImmediateDominator(Bypass, TCCheckBlock); | ||||||||
3020 | if (!Cost->requiresScalarEpilogue(VF)) | ||||||||
3021 | // If there is an epilogue which must run, there's no edge from the | ||||||||
3022 | // middle block to exit blocks and thus no need to update the immediate | ||||||||
3023 | // dominator of the exit blocks. | ||||||||
3024 | DT->changeImmediateDominator(LoopExitBlock, TCCheckBlock); | ||||||||
3025 | |||||||||
3026 | ReplaceInstWithInst( | ||||||||
3027 | TCCheckBlock->getTerminator(), | ||||||||
3028 | BranchInst::Create(Bypass, LoopVectorPreHeader, CheckMinIters)); | ||||||||
3029 | LoopBypassBlocks.push_back(TCCheckBlock); | ||||||||
3030 | } | ||||||||
3031 | |||||||||
3032 | BasicBlock *InnerLoopVectorizer::emitSCEVChecks(BasicBlock *Bypass) { | ||||||||
3033 | BasicBlock *const SCEVCheckBlock = | ||||||||
3034 | RTChecks.emitSCEVChecks(Bypass, LoopVectorPreHeader, LoopExitBlock); | ||||||||
3035 | if (!SCEVCheckBlock) | ||||||||
3036 | return nullptr; | ||||||||
3037 | |||||||||
3038 | 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", 3041, __extension__ __PRETTY_FUNCTION__)) | ||||||||
3039 | (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", 3041, __extension__ __PRETTY_FUNCTION__)) | ||||||||
3040 | 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", 3041, __extension__ __PRETTY_FUNCTION__)) | ||||||||
3041 | "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", 3041, __extension__ __PRETTY_FUNCTION__)); | ||||||||
3042 | |||||||||
3043 | |||||||||
3044 | // Update dominator only if this is first RT check. | ||||||||
3045 | if (LoopBypassBlocks.empty()) { | ||||||||
3046 | DT->changeImmediateDominator(Bypass, SCEVCheckBlock); | ||||||||
3047 | if (!Cost->requiresScalarEpilogue(VF)) | ||||||||
3048 | // If there is an epilogue which must run, there's no edge from the | ||||||||
3049 | // middle block to exit blocks and thus no need to update the immediate | ||||||||
3050 | // dominator of the exit blocks. | ||||||||
3051 | DT->changeImmediateDominator(LoopExitBlock, SCEVCheckBlock); | ||||||||
3052 | } | ||||||||
3053 | |||||||||
3054 | LoopBypassBlocks.push_back(SCEVCheckBlock); | ||||||||
3055 | AddedSafetyChecks = true; | ||||||||
3056 | return SCEVCheckBlock; | ||||||||
3057 | } | ||||||||
3058 | |||||||||
3059 | BasicBlock *InnerLoopVectorizer::emitMemRuntimeChecks(BasicBlock *Bypass) { | ||||||||
3060 | // VPlan-native path does not do any analysis for runtime checks currently. | ||||||||
3061 | if (EnableVPlanNativePath) | ||||||||
3062 | return nullptr; | ||||||||
3063 | |||||||||
3064 | BasicBlock *const MemCheckBlock = | ||||||||
3065 | RTChecks.emitMemRuntimeChecks(Bypass, LoopVectorPreHeader); | ||||||||
3066 | |||||||||
3067 | // Check if we generated code that checks in runtime if arrays overlap. We put | ||||||||
3068 | // the checks into a separate block to make the more common case of few | ||||||||
3069 | // elements faster. | ||||||||
3070 | if (!MemCheckBlock) | ||||||||
3071 | return nullptr; | ||||||||
3072 | |||||||||
3073 | if (MemCheckBlock->getParent()->hasOptSize() || OptForSizeBasedOnProfile) { | ||||||||
3074 | 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", 3076, __extension__ __PRETTY_FUNCTION__)) | ||||||||
3075 | "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", 3076, __extension__ __PRETTY_FUNCTION__)) | ||||||||
3076 | "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", 3076, __extension__ __PRETTY_FUNCTION__)); | ||||||||
3077 | ORE->emit([&]() { | ||||||||
3078 | return OptimizationRemarkAnalysis(DEBUG_TYPE"loop-vectorize", "VectorizationCodeSize", | ||||||||
3079 | OrigLoop->getStartLoc(), | ||||||||
3080 | OrigLoop->getHeader()) | ||||||||
3081 | << "Code-size may be reduced by not forcing " | ||||||||
3082 | "vectorization, or by source-code modifications " | ||||||||
3083 | "eliminating the need for runtime checks " | ||||||||
3084 | "(e.g., adding 'restrict')."; | ||||||||
3085 | }); | ||||||||
3086 | } | ||||||||
3087 | |||||||||
3088 | LoopBypassBlocks.push_back(MemCheckBlock); | ||||||||
3089 | |||||||||
3090 | AddedSafetyChecks = true; | ||||||||
3091 | |||||||||
3092 | return MemCheckBlock; | ||||||||
3093 | } | ||||||||
3094 | |||||||||
3095 | void InnerLoopVectorizer::createVectorLoopSkeleton(StringRef Prefix) { | ||||||||
3096 | LoopScalarBody = OrigLoop->getHeader(); | ||||||||
3097 | LoopVectorPreHeader = OrigLoop->getLoopPreheader(); | ||||||||
3098 | 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", 3098, __extension__ __PRETTY_FUNCTION__)); | ||||||||
3099 | LoopExitBlock = OrigLoop->getUniqueExitBlock(); // may be nullptr | ||||||||
3100 | 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", 3101, __extension__ __PRETTY_FUNCTION__)) | ||||||||
3101 | "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", 3101, __extension__ __PRETTY_FUNCTION__)); | ||||||||
3102 | |||||||||
3103 | LoopMiddleBlock = | ||||||||
3104 | SplitBlock(LoopVectorPreHeader, LoopVectorPreHeader->getTerminator(), DT, | ||||||||
3105 | LI, nullptr, Twine(Prefix) + "middle.block"); | ||||||||
3106 | LoopScalarPreHeader = | ||||||||
3107 | SplitBlock(LoopMiddleBlock, LoopMiddleBlock->getTerminator(), DT, LI, | ||||||||
3108 | nullptr, Twine(Prefix) + "scalar.ph"); | ||||||||
3109 | |||||||||
3110 | auto *ScalarLatchTerm = OrigLoop->getLoopLatch()->getTerminator(); | ||||||||
3111 | |||||||||
3112 | // Set up the middle block terminator. Two cases: | ||||||||
3113 | // 1) If we know that we must execute the scalar epilogue, emit an | ||||||||
3114 | // unconditional branch. | ||||||||
3115 | // 2) Otherwise, we must have a single unique exit block (due to how we | ||||||||
3116 | // implement the multiple exit case). In this case, set up a conditional | ||||||||
3117 | // branch from the middle block to the loop scalar preheader, and the | ||||||||
3118 | // exit block. completeLoopSkeleton will update the condition to use an | ||||||||
3119 | // iteration check, if required to decide whether to execute the remainder. | ||||||||
3120 | BranchInst *BrInst = Cost->requiresScalarEpilogue(VF) ? | ||||||||
3121 | BranchInst::Create(LoopScalarPreHeader) : | ||||||||
3122 | BranchInst::Create(LoopExitBlock, LoopScalarPreHeader, | ||||||||
3123 | Builder.getTrue()); | ||||||||
3124 | BrInst->setDebugLoc(ScalarLatchTerm->getDebugLoc()); | ||||||||
3125 | ReplaceInstWithInst(LoopMiddleBlock->getTerminator(), BrInst); | ||||||||
3126 | |||||||||
3127 | // Update dominator for loop exit. During skeleton creation, only the vector | ||||||||
3128 | // pre-header and the middle block are created. The vector loop is entirely | ||||||||
3129 | // created during VPlan exection. | ||||||||
3130 | if (!Cost->requiresScalarEpilogue(VF)) | ||||||||
3131 | // If there is an epilogue which must run, there's no edge from the | ||||||||
3132 | // middle block to exit blocks and thus no need to update the immediate | ||||||||
3133 | // dominator of the exit blocks. | ||||||||
3134 | DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock); | ||||||||
3135 | } | ||||||||
3136 | |||||||||
3137 | PHINode *InnerLoopVectorizer::createInductionResumeValue( | ||||||||
3138 | PHINode *OrigPhi, const InductionDescriptor &II, | ||||||||
3139 | ArrayRef<BasicBlock *> BypassBlocks, | ||||||||
3140 | std::pair<BasicBlock *, Value *> AdditionalBypass) { | ||||||||
3141 | Value *VectorTripCount = getOrCreateVectorTripCount(LoopVectorPreHeader); | ||||||||
3142 | assert(VectorTripCount && "Expected valid arguments")(static_cast <bool> (VectorTripCount && "Expected valid arguments" ) ? void (0) : __assert_fail ("VectorTripCount && \"Expected valid arguments\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3142, __extension__ __PRETTY_FUNCTION__)); | ||||||||
3143 | |||||||||
3144 | Instruction *OldInduction = Legal->getPrimaryInduction(); | ||||||||
3145 | Value *&EndValue = IVEndValues[OrigPhi]; | ||||||||
3146 | Value *EndValueFromAdditionalBypass = AdditionalBypass.second; | ||||||||
3147 | if (OrigPhi == OldInduction) { | ||||||||
3148 | // We know what the end value is. | ||||||||
3149 | EndValue = VectorTripCount; | ||||||||
3150 | } else { | ||||||||
3151 | IRBuilder<> B(LoopVectorPreHeader->getTerminator()); | ||||||||
3152 | |||||||||
3153 | // Fast-math-flags propagate from the original induction instruction. | ||||||||
3154 | if (II.getInductionBinOp() && isa<FPMathOperator>(II.getInductionBinOp())) | ||||||||
3155 | B.setFastMathFlags(II.getInductionBinOp()->getFastMathFlags()); | ||||||||
3156 | |||||||||
3157 | Value *Step = | ||||||||
3158 | CreateStepValue(II.getStep(), *PSE.getSE(), &*B.GetInsertPoint()); | ||||||||
3159 | EndValue = | ||||||||
3160 | emitTransformedIndex(B, VectorTripCount, II.getStartValue(), Step, II); | ||||||||
3161 | EndValue->setName("ind.end"); | ||||||||
3162 | |||||||||
3163 | // Compute the end value for the additional bypass (if applicable). | ||||||||
3164 | if (AdditionalBypass.first) { | ||||||||
3165 | B.SetInsertPoint(&(*AdditionalBypass.first->getFirstInsertionPt())); | ||||||||
3166 | Value *Step = | ||||||||
3167 | CreateStepValue(II.getStep(), *PSE.getSE(), &*B.GetInsertPoint()); | ||||||||
3168 | EndValueFromAdditionalBypass = emitTransformedIndex( | ||||||||
3169 | B, AdditionalBypass.second, II.getStartValue(), Step, II); | ||||||||
3170 | EndValueFromAdditionalBypass->setName("ind.end"); | ||||||||
3171 | } | ||||||||
3172 | } | ||||||||
3173 | |||||||||
3174 | // Create phi nodes to merge from the backedge-taken check block. | ||||||||
3175 | PHINode *BCResumeVal = PHINode::Create(OrigPhi->getType(), 3, "bc.resume.val", | ||||||||
3176 | LoopScalarPreHeader->getTerminator()); | ||||||||
3177 | // Copy original phi DL over to the new one. | ||||||||
3178 | BCResumeVal->setDebugLoc(OrigPhi->getDebugLoc()); | ||||||||
3179 | |||||||||
3180 | // The new PHI merges the original incoming value, in case of a bypass, | ||||||||
3181 | // or the value at the end of the vectorized loop. | ||||||||
3182 | BCResumeVal->addIncoming(EndValue, LoopMiddleBlock); | ||||||||
3183 | |||||||||
3184 | // Fix the scalar body counter (PHI node). | ||||||||
3185 | // The old induction's phi node in the scalar body needs the truncated | ||||||||
3186 | // value. | ||||||||
3187 | for (BasicBlock *BB : BypassBlocks) | ||||||||
3188 | BCResumeVal->addIncoming(II.getStartValue(), BB); | ||||||||
3189 | |||||||||
3190 | if (AdditionalBypass.first) | ||||||||
3191 | BCResumeVal->setIncomingValueForBlock(AdditionalBypass.first, | ||||||||
3192 | EndValueFromAdditionalBypass); | ||||||||
3193 | return BCResumeVal; | ||||||||
3194 | } | ||||||||
3195 | |||||||||
3196 | void InnerLoopVectorizer::createInductionResumeValues( | ||||||||
3197 | std::pair<BasicBlock *, Value *> AdditionalBypass) { | ||||||||
3198 | 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", 3200, __extension__ __PRETTY_FUNCTION__)) | ||||||||
3199 | (!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", 3200, __extension__ __PRETTY_FUNCTION__)) | ||||||||
3200 | "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", 3200, __extension__ __PRETTY_FUNCTION__)); | ||||||||
3201 | // We are going to resume the execution of the scalar loop. | ||||||||
3202 | // Go over all of the induction variables that we found and fix the | ||||||||
3203 | // PHIs that are left in the scalar version of the loop. | ||||||||
3204 | // The starting values of PHI nodes depend on the counter of the last | ||||||||
3205 | // iteration in the vectorized loop. | ||||||||
3206 | // If we come from a bypass edge then we need to start from the original | ||||||||
3207 | // start value. | ||||||||
3208 | for (const auto &InductionEntry : Legal->getInductionVars()) { | ||||||||
3209 | PHINode *OrigPhi = InductionEntry.first; | ||||||||
3210 | const InductionDescriptor &II = InductionEntry.second; | ||||||||
3211 | PHINode *BCResumeVal = createInductionResumeValue( | ||||||||
3212 | OrigPhi, II, LoopBypassBlocks, AdditionalBypass); | ||||||||
3213 | OrigPhi->setIncomingValueForBlock(LoopScalarPreHeader, BCResumeVal); | ||||||||
3214 | } | ||||||||
3215 | } | ||||||||
3216 | |||||||||
3217 | BasicBlock *InnerLoopVectorizer::completeLoopSkeleton() { | ||||||||
3218 | // The trip counts should be cached by now. | ||||||||
3219 | Value *Count = getOrCreateTripCount(LoopVectorPreHeader); | ||||||||
3220 | Value *VectorTripCount = getOrCreateVectorTripCount(LoopVectorPreHeader); | ||||||||
3221 | |||||||||
3222 | auto *ScalarLatchTerm = OrigLoop->getLoopLatch()->getTerminator(); | ||||||||
3223 | |||||||||
3224 | // Add a check in the middle block to see if we have completed | ||||||||
3225 | // all of the iterations in the first vector loop. Three cases: | ||||||||
3226 | // 1) If we require a scalar epilogue, there is no conditional branch as | ||||||||
3227 | // we unconditionally branch to the scalar preheader. Do nothing. | ||||||||
3228 | // 2) If (N - N%VF) == N, then we *don't* need to run the remainder. | ||||||||
3229 | // Thus if tail is to be folded, we know we don't need to run the | ||||||||
3230 | // remainder and we can use the previous value for the condition (true). | ||||||||
3231 | // 3) Otherwise, construct a runtime check. | ||||||||
3232 | if (!Cost->requiresScalarEpilogue(VF) && !Cost->foldTailByMasking()) { | ||||||||
3233 | Instruction *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, | ||||||||
3234 | Count, VectorTripCount, "cmp.n", | ||||||||
3235 | LoopMiddleBlock->getTerminator()); | ||||||||
3236 | |||||||||
3237 | // Here we use the same DebugLoc as the scalar loop latch terminator instead | ||||||||
3238 | // of the corresponding compare because they may have ended up with | ||||||||
3239 | // different line numbers and we want to avoid awkward line stepping while | ||||||||
3240 | // debugging. Eg. if the compare has got a line number inside the loop. | ||||||||
3241 | CmpN->setDebugLoc(ScalarLatchTerm->getDebugLoc()); | ||||||||
3242 | cast<BranchInst>(LoopMiddleBlock->getTerminator())->setCondition(CmpN); | ||||||||
3243 | } | ||||||||
3244 | |||||||||
3245 | #ifdef EXPENSIVE_CHECKS | ||||||||
3246 | 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", 3246, __extension__ __PRETTY_FUNCTION__)); | ||||||||
3247 | #endif | ||||||||
3248 | |||||||||
3249 | return LoopVectorPreHeader; | ||||||||
3250 | } | ||||||||
3251 | |||||||||
3252 | std::pair<BasicBlock *, Value *> | ||||||||
3253 | InnerLoopVectorizer::createVectorizedLoopSkeleton() { | ||||||||
3254 | /* | ||||||||
3255 | In this function we generate a new loop. The new loop will contain | ||||||||
3256 | the vectorized instructions while the old loop will continue to run the | ||||||||
3257 | scalar remainder. | ||||||||
3258 | |||||||||
3259 | [ ] <-- loop iteration number check. | ||||||||
3260 | / | | ||||||||
3261 | / v | ||||||||
3262 | | [ ] <-- vector loop bypass (may consist of multiple blocks). | ||||||||
3263 | | / | | ||||||||
3264 | | / v | ||||||||
3265 | || [ ] <-- vector pre header. | ||||||||
3266 | |/ | | ||||||||
3267 | | v | ||||||||
3268 | | [ ] \ | ||||||||
3269 | | [ ]_| <-- vector loop (created during VPlan execution). | ||||||||
3270 | | | | ||||||||
3271 | | v | ||||||||
3272 | \ -[ ] <--- middle-block. | ||||||||
3273 | \/ | | ||||||||
3274 | /\ v | ||||||||
3275 | | ->[ ] <--- new preheader. | ||||||||
3276 | | | | ||||||||
3277 | (opt) v <-- edge from middle to exit iff epilogue is not required. | ||||||||
3278 | | [ ] \ | ||||||||
3279 | | [ ]_| <-- old scalar loop to handle remainder (scalar epilogue). | ||||||||
3280 | \ | | ||||||||
3281 | \ v | ||||||||
3282 | >[ ] <-- exit block(s). | ||||||||
3283 | ... | ||||||||
3284 | */ | ||||||||
3285 | |||||||||
3286 | // Create an empty vector loop, and prepare basic blocks for the runtime | ||||||||
3287 | // checks. | ||||||||
3288 | createVectorLoopSkeleton(""); | ||||||||
3289 | |||||||||
3290 | // Now, compare the new count to zero. If it is zero skip the vector loop and | ||||||||
3291 | // jump to the scalar loop. This check also covers the case where the | ||||||||
3292 | // backedge-taken count is uint##_max: adding one to it will overflow leading | ||||||||
3293 | // to an incorrect trip count of zero. In this (rare) case we will also jump | ||||||||
3294 | // to the scalar loop. | ||||||||
3295 | emitIterationCountCheck(LoopScalarPreHeader); | ||||||||
3296 | |||||||||
3297 | // Generate the code to check any assumptions that we've made for SCEV | ||||||||
3298 | // expressions. | ||||||||
3299 | emitSCEVChecks(LoopScalarPreHeader); | ||||||||
3300 | |||||||||
3301 | // Generate the code that checks in runtime if arrays overlap. We put the | ||||||||
3302 | // checks into a separate block to make the more common case of few elements | ||||||||
3303 | // faster. | ||||||||
3304 | emitMemRuntimeChecks(LoopScalarPreHeader); | ||||||||
3305 | |||||||||
3306 | // Emit phis for the new starting index of the scalar loop. | ||||||||
3307 | createInductionResumeValues(); | ||||||||
3308 | |||||||||
3309 | return {completeLoopSkeleton(), nullptr}; | ||||||||
3310 | } | ||||||||
3311 | |||||||||
3312 | // Fix up external users of the induction variable. At this point, we are | ||||||||
3313 | // in LCSSA form, with all external PHIs that use the IV having one input value, | ||||||||
3314 | // coming from the remainder loop. We need those PHIs to also have a correct | ||||||||
3315 | // value for the IV when arriving directly from the middle block. | ||||||||
3316 | void InnerLoopVectorizer::fixupIVUsers(PHINode *OrigPhi, | ||||||||
3317 | const InductionDescriptor &II, | ||||||||
3318 | Value *VectorTripCount, Value *EndValue, | ||||||||
3319 | BasicBlock *MiddleBlock, | ||||||||
3320 | BasicBlock *VectorHeader, VPlan &Plan) { | ||||||||
3321 | // There are two kinds of external IV usages - those that use the value | ||||||||
3322 | // computed in the last iteration (the PHI) and those that use the penultimate | ||||||||
3323 | // value (the value that feeds into the phi from the loop latch). | ||||||||
3324 | // We allow both, but they, obviously, have different values. | ||||||||
3325 | |||||||||
3326 | 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", 3326, __extension__ __PRETTY_FUNCTION__)); | ||||||||
3327 | |||||||||
3328 | DenseMap<Value *, Value *> MissingVals; | ||||||||
3329 | |||||||||
3330 | // An external user of the last iteration's value should see the value that | ||||||||
3331 | // the remainder loop uses to initialize its own IV. | ||||||||
3332 | Value *PostInc = OrigPhi->getIncomingValueForBlock(OrigLoop->getLoopLatch()); | ||||||||
3333 | for (User *U : PostInc->users()) { | ||||||||
3334 | Instruction *UI = cast<Instruction>(U); | ||||||||
3335 | if (!OrigLoop->contains(UI)) { | ||||||||
3336 | 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", 3336, __extension__ __PRETTY_FUNCTION__)); | ||||||||
3337 | MissingVals[UI] = EndValue; | ||||||||
3338 | } | ||||||||
3339 | } | ||||||||
3340 | |||||||||
3341 | // An external user of the penultimate value need to see EndValue - Step. | ||||||||
3342 | // The simplest way to get this is to recompute it from the constituent SCEVs, | ||||||||
3343 | // that is Start + (Step * (CRD - 1)). | ||||||||
3344 | for (User *U : OrigPhi->users()) { | ||||||||
3345 | auto *UI = cast<Instruction>(U); | ||||||||
3346 | if (!OrigLoop->contains(UI)) { | ||||||||
3347 | 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", 3347, __extension__ __PRETTY_FUNCTION__)); | ||||||||
3348 | |||||||||
3349 | IRBuilder<> B(MiddleBlock->getTerminator()); | ||||||||
3350 | |||||||||
3351 | // Fast-math-flags propagate from the original induction instruction. | ||||||||
3352 | if (II.getInductionBinOp() && isa<FPMathOperator>(II.getInductionBinOp())) | ||||||||
3353 | B.setFastMathFlags(II.getInductionBinOp()->getFastMathFlags()); | ||||||||
3354 | |||||||||
3355 | Value *CountMinusOne = B.CreateSub( | ||||||||
3356 | VectorTripCount, ConstantInt::get(VectorTripCount->getType(), 1)); | ||||||||
3357 | CountMinusOne->setName("cmo"); | ||||||||
3358 | Value *Step = CreateStepValue(II.getStep(), *PSE.getSE(), | ||||||||
3359 | VectorHeader->getTerminator()); | ||||||||
3360 | Value *Escape = | ||||||||
3361 | emitTransformedIndex(B, CountMinusOne, II.getStartValue(), Step, II); | ||||||||
3362 | Escape->setName("ind.escape"); | ||||||||
3363 | MissingVals[UI] = Escape; | ||||||||
3364 | } | ||||||||
3365 | } | ||||||||
3366 | |||||||||
3367 | for (auto &I : MissingVals) { | ||||||||
3368 | PHINode *PHI = cast<PHINode>(I.first); | ||||||||
3369 | // One corner case we have to handle is two IVs "chasing" each-other, | ||||||||
3370 | // that is %IV2 = phi [...], [ %IV1, %latch ] | ||||||||
3371 | // In this case, if IV1 has an external use, we need to avoid adding both | ||||||||
3372 | // "last value of IV1" and "penultimate value of IV2". So, verify that we | ||||||||
3373 | // don't already have an incoming value for the middle block. | ||||||||
3374 | if (PHI->getBasicBlockIndex(MiddleBlock) == -1) { | ||||||||
3375 | PHI->addIncoming(I.second, MiddleBlock); | ||||||||
3376 | Plan.removeLiveOut(PHI); | ||||||||
3377 | } | ||||||||
3378 | } | ||||||||
3379 | } | ||||||||
3380 | |||||||||
3381 | namespace { | ||||||||
3382 | |||||||||
3383 | struct CSEDenseMapInfo { | ||||||||
3384 | static bool canHandle(const Instruction *I) { | ||||||||
3385 | return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) || | ||||||||
3386 | isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I); | ||||||||
3387 | } | ||||||||
3388 | |||||||||
3389 | static inline Instruction *getEmptyKey() { | ||||||||
3390 | return DenseMapInfo<Instruction *>::getEmptyKey(); | ||||||||
3391 | } | ||||||||
3392 | |||||||||
3393 | static inline Instruction *getTombstoneKey() { | ||||||||
3394 | return DenseMapInfo<Instruction *>::getTombstoneKey(); | ||||||||
3395 | } | ||||||||
3396 | |||||||||
3397 | static unsigned getHashValue(const Instruction *I) { | ||||||||
3398 | 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", 3398, __extension__ __PRETTY_FUNCTION__)); | ||||||||
3399 | return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(), | ||||||||
3400 | I->value_op_end())); | ||||||||
3401 | } | ||||||||
3402 | |||||||||
3403 | static bool isEqual(const Instruction *LHS, const Instruction *RHS) { | ||||||||
3404 | if (LHS == getEmptyKey() || RHS == getEmptyKey() || | ||||||||
3405 | LHS == getTombstoneKey() || RHS == getTombstoneKey()) | ||||||||
3406 | return LHS == RHS; | ||||||||
3407 | return LHS->isIdenticalTo(RHS); | ||||||||
3408 | } | ||||||||
3409 | }; | ||||||||
3410 | |||||||||
3411 | } // end anonymous namespace | ||||||||
3412 | |||||||||
3413 | ///Perform cse of induction variable instructions. | ||||||||
3414 | static void cse(BasicBlock *BB) { | ||||||||
3415 | // Perform simple cse. | ||||||||
3416 | SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap; | ||||||||
3417 | for (Instruction &In : llvm::make_early_inc_range(*BB)) { | ||||||||
3418 | if (!CSEDenseMapInfo::canHandle(&In)) | ||||||||
3419 | continue; | ||||||||
3420 | |||||||||
3421 | // Check if we can replace this instruction with any of the | ||||||||
3422 | // visited instructions. | ||||||||
3423 | if (Instruction *V = CSEMap.lookup(&In)) { | ||||||||
3424 | In.replaceAllUsesWith(V); | ||||||||
3425 | In.eraseFromParent(); | ||||||||
3426 | continue; | ||||||||
3427 | } | ||||||||
3428 | |||||||||
3429 | CSEMap[&In] = &In; | ||||||||
3430 | } | ||||||||
3431 | } | ||||||||
3432 | |||||||||
3433 | InstructionCost | ||||||||
3434 | LoopVectorizationCostModel::getVectorCallCost(CallInst *CI, ElementCount VF, | ||||||||
3435 | bool &NeedToScalarize) const { | ||||||||
3436 | Function *F = CI->getCalledFunction(); | ||||||||
3437 | Type *ScalarRetTy = CI->getType(); | ||||||||
3438 | SmallVector<Type *, 4> Tys, ScalarTys; | ||||||||
3439 | for (auto &ArgOp : CI->args()) | ||||||||
3440 | ScalarTys.push_back(ArgOp->getType()); | ||||||||
3441 | |||||||||
3442 | // Estimate cost of scalarized vector call. The source operands are assumed | ||||||||
3443 | // to be vectors, so we need to extract individual elements from there, | ||||||||
3444 | // execute VF scalar calls, and then gather the result into the vector return | ||||||||
3445 | // value. | ||||||||
3446 | TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput; | ||||||||
3447 | InstructionCost ScalarCallCost = | ||||||||
3448 | TTI.getCallInstrCost(F, ScalarRetTy, ScalarTys, CostKind); | ||||||||
3449 | if (VF.isScalar()) | ||||||||
3450 | return ScalarCallCost; | ||||||||
3451 | |||||||||
3452 | // Compute corresponding vector type for return value and arguments. | ||||||||
3453 | Type *RetTy = ToVectorTy(ScalarRetTy, VF); | ||||||||
3454 | for (Type *ScalarTy : ScalarTys) | ||||||||
3455 | Tys.push_back(ToVectorTy(ScalarTy, VF)); | ||||||||
3456 | |||||||||
3457 | // Compute costs of unpacking argument values for the scalar calls and | ||||||||
3458 | // packing the return values to a vector. | ||||||||
3459 | InstructionCost ScalarizationCost = | ||||||||
3460 | getScalarizationOverhead(CI, VF, CostKind); | ||||||||
3461 | |||||||||
3462 | InstructionCost Cost = | ||||||||
3463 | ScalarCallCost * VF.getKnownMinValue() + ScalarizationCost; | ||||||||
3464 | |||||||||
3465 | // If we can't emit a vector call for this function, then the currently found | ||||||||
3466 | // cost is the cost we need to return. | ||||||||
3467 | NeedToScalarize = true; | ||||||||
3468 | VFShape Shape = VFShape::get(*CI, VF, false /*HasGlobalPred*/); | ||||||||
3469 | Function *VecFunc = VFDatabase(*CI).getVectorizedFunction(Shape); | ||||||||
3470 | |||||||||
3471 | if (!TLI || CI->isNoBuiltin() || !VecFunc) | ||||||||
3472 | return Cost; | ||||||||
3473 | |||||||||
3474 | // If the corresponding vector cost is cheaper, return its cost. | ||||||||
3475 | InstructionCost VectorCallCost = | ||||||||
3476 | TTI.getCallInstrCost(nullptr, RetTy, Tys, CostKind); | ||||||||
3477 | if (VectorCallCost < Cost) { | ||||||||
3478 | NeedToScalarize = false; | ||||||||
3479 | Cost = VectorCallCost; | ||||||||
3480 | } | ||||||||
3481 | return Cost; | ||||||||
3482 | } | ||||||||
3483 | |||||||||
3484 | static Type *MaybeVectorizeType(Type *Elt, ElementCount VF) { | ||||||||
3485 | if (VF.isScalar() || (!Elt->isIntOrPtrTy() && !Elt->isFloatingPointTy())) | ||||||||
3486 | return Elt; | ||||||||
3487 | return VectorType::get(Elt, VF); | ||||||||
3488 | } | ||||||||
3489 | |||||||||
3490 | InstructionCost | ||||||||
3491 | LoopVectorizationCostModel::getVectorIntrinsicCost(CallInst *CI, | ||||||||
3492 | ElementCount VF) const { | ||||||||
3493 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); | ||||||||
3494 | 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", 3494, __extension__ __PRETTY_FUNCTION__)); | ||||||||
3495 | Type *RetTy = MaybeVectorizeType(CI->getType(), VF); | ||||||||
3496 | FastMathFlags FMF; | ||||||||
3497 | if (auto *FPMO = dyn_cast<FPMathOperator>(CI)) | ||||||||
3498 | FMF = FPMO->getFastMathFlags(); | ||||||||
3499 | |||||||||
3500 | SmallVector<const Value *> Arguments(CI->args()); | ||||||||
3501 | FunctionType *FTy = CI->getCalledFunction()->getFunctionType(); | ||||||||
3502 | SmallVector<Type *> ParamTys; | ||||||||
3503 | std::transform(FTy->param_begin(), FTy->param_end(), | ||||||||
3504 | std::back_inserter(ParamTys), | ||||||||
3505 | [&](Type *Ty) { return MaybeVectorizeType(Ty, VF); }); | ||||||||
3506 | |||||||||
3507 | IntrinsicCostAttributes CostAttrs(ID, RetTy, Arguments, ParamTys, FMF, | ||||||||
3508 | dyn_cast<IntrinsicInst>(CI)); | ||||||||
3509 | return TTI.getIntrinsicInstrCost(CostAttrs, | ||||||||
3510 | TargetTransformInfo::TCK_RecipThroughput); | ||||||||
3511 | } | ||||||||
3512 | |||||||||
3513 | static Type *smallestIntegerVectorType(Type *T1, Type *T2) { | ||||||||
3514 | auto *I1 = cast<IntegerType>(cast<VectorType>(T1)->getElementType()); | ||||||||
3515 | auto *I2 = cast<IntegerType>(cast<VectorType>(T2)->getElementType()); | ||||||||
3516 | return I1->getBitWidth() < I2->getBitWidth() ? T1 : T2; | ||||||||
3517 | } | ||||||||
3518 | |||||||||
3519 | static Type *largestIntegerVectorType(Type *T1, Type *T2) { | ||||||||
3520 | auto *I1 = cast<IntegerType>(cast<VectorType>(T1)->getElementType()); | ||||||||
3521 | auto *I2 = cast<IntegerType>(cast<VectorType>(T2)->getElementType()); | ||||||||
3522 | return I1->getBitWidth() > I2->getBitWidth() ? T1 : T2; | ||||||||
3523 | } | ||||||||
3524 | |||||||||
3525 | void InnerLoopVectorizer::truncateToMinimalBitwidths(VPTransformState &State) { | ||||||||
3526 | // For every instruction `I` in MinBWs, truncate the operands, create a | ||||||||
3527 | // truncated version of `I` and reextend its result. InstCombine runs | ||||||||
3528 | // later and will remove any ext/trunc pairs. | ||||||||
3529 | SmallPtrSet<Value *, 4> Erased; | ||||||||
3530 | for (const auto &KV : Cost->getMinimalBitwidths()) { | ||||||||
3531 | // If the value wasn't vectorized, we must maintain the original scalar | ||||||||
3532 | // type. The absence of the value from State indicates that it | ||||||||
3533 | // wasn't vectorized. | ||||||||
3534 | // FIXME: Should not rely on getVPValue at this point. | ||||||||
3535 | VPValue *Def = State.Plan->getVPValue(KV.first, true); | ||||||||
3536 | if (!State.hasAnyVectorValue(Def)) | ||||||||
3537 | continue; | ||||||||
3538 | for (unsigned Part = 0; Part < UF; ++Part) { | ||||||||
3539 | Value *I = State.get(Def, Part); | ||||||||
3540 | if (Erased.count(I) || I->use_empty() || !isa<Instruction>(I)) | ||||||||
3541 | continue; | ||||||||
3542 | Type *OriginalTy = I->getType(); | ||||||||
3543 | Type *ScalarTruncatedTy = | ||||||||
3544 | IntegerType::get(OriginalTy->getContext(), KV.second); | ||||||||
3545 | auto *TruncatedTy = VectorType::get( | ||||||||
3546 | ScalarTruncatedTy, cast<VectorType>(OriginalTy)->getElementCount()); | ||||||||
3547 | if (TruncatedTy == OriginalTy) | ||||||||
3548 | continue; | ||||||||
3549 | |||||||||
3550 | IRBuilder<> B(cast<Instruction>(I)); | ||||||||
3551 | auto ShrinkOperand = [&](Value *V) -> Value * { | ||||||||
3552 | if (auto *ZI = dyn_cast<ZExtInst>(V)) | ||||||||
3553 | if (ZI->getSrcTy() == TruncatedTy) | ||||||||
3554 | return ZI->getOperand(0); | ||||||||
3555 | return B.CreateZExtOrTrunc(V, TruncatedTy); | ||||||||
3556 | }; | ||||||||
3557 | |||||||||
3558 | // The actual instruction modification depends on the instruction type, | ||||||||
3559 | // unfortunately. | ||||||||
3560 | Value *NewI = nullptr; | ||||||||
3561 | if (auto *BO = dyn_cast<BinaryOperator>(I)) { | ||||||||
3562 | NewI = B.CreateBinOp(BO->getOpcode(), ShrinkOperand(BO->getOperand(0)), | ||||||||
3563 | ShrinkOperand(BO->getOperand(1))); | ||||||||
3564 | |||||||||
3565 | // Any wrapping introduced by shrinking this operation shouldn't be | ||||||||
3566 | // considered undefined behavior. So, we can't unconditionally copy | ||||||||
3567 | // arithmetic wrapping flags to NewI. | ||||||||
3568 | cast<BinaryOperator>(NewI)->copyIRFlags(I, /*IncludeWrapFlags=*/false); | ||||||||
3569 | } else if (auto *CI = dyn_cast<ICmpInst>(I)) { | ||||||||
3570 | NewI = | ||||||||
3571 | B.CreateICmp(CI->getPredicate(), ShrinkOperand(CI->getOperand(0)), | ||||||||
3572 | ShrinkOperand(CI->getOperand(1))); | ||||||||
3573 | } else if (auto *SI = dyn_cast<SelectInst>(I)) { | ||||||||
3574 | NewI = B.CreateSelect(SI->getCondition(), | ||||||||
3575 | ShrinkOperand(SI->getTrueValue()), | ||||||||
3576 | ShrinkOperand(SI->getFalseValue())); | ||||||||
3577 | } else if (auto *CI = dyn_cast<CastInst>(I)) { | ||||||||
3578 | switch (CI->getOpcode()) { | ||||||||
3579 | default: | ||||||||
3580 | llvm_unreachable("Unhandled cast!")::llvm::llvm_unreachable_internal("Unhandled cast!", "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3580); | ||||||||
3581 | case Instruction::Trunc: | ||||||||
3582 | NewI = ShrinkOperand(CI->getOperand(0)); | ||||||||
3583 | break; | ||||||||
3584 | case Instruction::SExt: | ||||||||
3585 | NewI = B.CreateSExtOrTrunc( | ||||||||
3586 | CI->getOperand(0), | ||||||||
3587 | smallestIntegerVectorType(OriginalTy, TruncatedTy)); | ||||||||
3588 | break; | ||||||||
3589 | case Instruction::ZExt: | ||||||||
3590 | NewI = B.CreateZExtOrTrunc( | ||||||||
3591 | CI->getOperand(0), | ||||||||
3592 | smallestIntegerVectorType(OriginalTy, TruncatedTy)); | ||||||||
3593 | break; | ||||||||
3594 | } | ||||||||
3595 | } else if (auto *SI = dyn_cast<ShuffleVectorInst>(I)) { | ||||||||
3596 | auto Elements0 = | ||||||||
3597 | cast<VectorType>(SI->getOperand(0)->getType())->getElementCount(); | ||||||||
3598 | auto *O0 = B.CreateZExtOrTrunc( | ||||||||
3599 | SI->getOperand(0), VectorType::get(ScalarTruncatedTy, Elements0)); | ||||||||
3600 | auto Elements1 = | ||||||||
3601 | cast<VectorType>(SI->getOperand(1)->getType())->getElementCount(); | ||||||||
3602 | auto *O1 = B.CreateZExtOrTrunc( | ||||||||
3603 | SI->getOperand(1), VectorType::get(ScalarTruncatedTy, Elements1)); | ||||||||
3604 | |||||||||
3605 | NewI = B.CreateShuffleVector(O0, O1, SI->getShuffleMask()); | ||||||||
3606 | } else if (isa<LoadInst>(I) || isa<PHINode>(I)) { | ||||||||
3607 | // Don't do anything with the operands, just extend the result. | ||||||||
3608 | continue; | ||||||||
3609 | } else if (auto *IE = dyn_cast<InsertElementInst>(I)) { | ||||||||
3610 | auto Elements = | ||||||||
3611 | cast<VectorType>(IE->getOperand(0)->getType())->getElementCount(); | ||||||||
3612 | auto *O0 = B.CreateZExtOrTrunc( | ||||||||
3613 | IE->getOperand(0), VectorType::get(ScalarTruncatedTy, Elements)); | ||||||||
3614 | auto *O1 = B.CreateZExtOrTrunc(IE->getOperand(1), ScalarTruncatedTy); | ||||||||
3615 | NewI = B.CreateInsertElement(O0, O1, IE->getOperand(2)); | ||||||||
3616 | } else if (auto *EE = dyn_cast<ExtractElementInst>(I)) { | ||||||||
3617 | auto Elements = | ||||||||
3618 | cast<VectorType>(EE->getOperand(0)->getType())->getElementCount(); | ||||||||
3619 | auto *O0 = B.CreateZExtOrTrunc( | ||||||||
3620 | EE->getOperand(0), VectorType::get(ScalarTruncatedTy, Elements)); | ||||||||
3621 | NewI = B.CreateExtractElement(O0, EE->getOperand(2)); | ||||||||
3622 | } else { | ||||||||
3623 | // If we don't know what to do, be conservative and don't do anything. | ||||||||
3624 | continue; | ||||||||
3625 | } | ||||||||
3626 | |||||||||
3627 | // Lastly, extend the result. | ||||||||
3628 | NewI->takeName(cast<Instruction>(I)); | ||||||||
3629 | Value *Res = B.CreateZExtOrTrunc(NewI, OriginalTy); | ||||||||
3630 | I->replaceAllUsesWith(Res); | ||||||||
3631 | cast<Instruction>(I)->eraseFromParent(); | ||||||||
3632 | Erased.insert(I); | ||||||||
3633 | State.reset(Def, Res, Part); | ||||||||
3634 | } | ||||||||
3635 | } | ||||||||
3636 | |||||||||
3637 | // We'll have created a bunch of ZExts that are now parentless. Clean up. | ||||||||
3638 | for (const auto &KV : Cost->getMinimalBitwidths()) { | ||||||||
3639 | // If the value wasn't vectorized, we must maintain the original scalar | ||||||||
3640 | // type. The absence of the value from State indicates that it | ||||||||
3641 | // wasn't vectorized. | ||||||||
3642 | // FIXME: Should not rely on getVPValue at this point. | ||||||||
3643 | VPValue *Def = State.Plan->getVPValue(KV.first, true); | ||||||||
3644 | if (!State.hasAnyVectorValue(Def)) | ||||||||
3645 | continue; | ||||||||
3646 | for (unsigned Part = 0; Part < UF; ++Part) { | ||||||||
3647 | Value *I = State.get(Def, Part); | ||||||||
3648 | ZExtInst *Inst = dyn_cast<ZExtInst>(I); | ||||||||
3649 | if (Inst && Inst->use_empty()) { | ||||||||
3650 | Value *NewI = Inst->getOperand(0); | ||||||||
3651 | Inst->eraseFromParent(); | ||||||||
3652 | State.reset(Def, NewI, Part); | ||||||||
3653 | } | ||||||||
3654 | } | ||||||||
3655 | } | ||||||||
3656 | } | ||||||||
3657 | |||||||||
3658 | void InnerLoopVectorizer::fixVectorizedLoop(VPTransformState &State, | ||||||||
3659 | VPlan &Plan) { | ||||||||
3660 | // Insert truncates and extends for any truncated instructions as hints to | ||||||||
3661 | // InstCombine. | ||||||||
3662 | if (VF.isVector()) | ||||||||
3663 | truncateToMinimalBitwidths(State); | ||||||||
3664 | |||||||||
3665 | // Fix widened non-induction PHIs by setting up the PHI operands. | ||||||||
3666 | if (EnableVPlanNativePath) | ||||||||
3667 | fixNonInductionPHIs(Plan, State); | ||||||||
3668 | |||||||||
3669 | // At this point every instruction in the original loop is widened to a | ||||||||
3670 | // vector form. Now we need to fix the recurrences in the loop. These PHI | ||||||||
3671 | // nodes are currently empty because we did not want to introduce cycles. | ||||||||
3672 | // This is the second stage of vectorizing recurrences. | ||||||||
3673 | fixCrossIterationPHIs(State); | ||||||||
3674 | |||||||||
3675 | // Forget the original basic block. | ||||||||
3676 | PSE.getSE()->forgetLoop(OrigLoop); | ||||||||
3677 | |||||||||
3678 | VPBasicBlock *LatchVPBB = Plan.getVectorLoopRegion()->getExitingBasicBlock(); | ||||||||
3679 | Loop *VectorLoop = LI->getLoopFor(State.CFG.VPBB2IRBB[LatchVPBB]); | ||||||||
3680 | if (Cost->requiresScalarEpilogue(VF)) { | ||||||||
3681 | // No edge from the middle block to the unique exit block has been inserted | ||||||||
3682 | // and there is nothing to fix from vector loop; phis should have incoming | ||||||||
3683 | // from scalar loop only. | ||||||||
3684 | Plan.clearLiveOuts(); | ||||||||
3685 | } else { | ||||||||
3686 | // If we inserted an edge from the middle block to the unique exit block, | ||||||||
3687 | // update uses outside the loop (phis) to account for the newly inserted | ||||||||
3688 | // edge. | ||||||||
3689 | |||||||||
3690 | // Fix-up external users of the induction variables. | ||||||||
3691 | for (const auto &Entry : Legal->getInductionVars()) | ||||||||
3692 | fixupIVUsers(Entry.first, Entry.second, | ||||||||
3693 | getOrCreateVectorTripCount(VectorLoop->getLoopPreheader()), | ||||||||
3694 | IVEndValues[Entry.first], LoopMiddleBlock, | ||||||||
3695 | VectorLoop->getHeader(), Plan); | ||||||||
3696 | } | ||||||||
3697 | |||||||||
3698 | // Fix LCSSA phis not already fixed earlier. Extracts may need to be generated | ||||||||
3699 | // in the exit block, so update the builder. | ||||||||
3700 | State.Builder.SetInsertPoint(State.CFG.ExitBB->getFirstNonPHI()); | ||||||||
3701 | for (const auto &KV : Plan.getLiveOuts()) | ||||||||
3702 | KV.second->fixPhi(Plan, State); | ||||||||
3703 | |||||||||
3704 | for (Instruction *PI : PredicatedInstructions) | ||||||||
3705 | sinkScalarOperands(&*PI); | ||||||||
3706 | |||||||||
3707 | // Remove redundant induction instructions. | ||||||||
3708 | cse(VectorLoop->getHeader()); | ||||||||
3709 | |||||||||
3710 | // Set/update profile weights for the vector and remainder loops as original | ||||||||
3711 | // loop iterations are now distributed among them. Note that original loop | ||||||||
3712 | // represented by LoopScalarBody becomes remainder loop after vectorization. | ||||||||
3713 | // | ||||||||
3714 | // For cases like foldTailByMasking() and requiresScalarEpiloque() we may | ||||||||
3715 | // end up getting slightly roughened result but that should be OK since | ||||||||
3716 | // profile is not inherently precise anyway. Note also possible bypass of | ||||||||
3717 | // vector code caused by legality checks is ignored, assigning all the weight | ||||||||
3718 | // to the vector loop, optimistically. | ||||||||
3719 | // | ||||||||
3720 | // For scalable vectorization we can't know at compile time how many iterations | ||||||||
3721 | // of the loop are handled in one vector iteration, so instead assume a pessimistic | ||||||||
3722 | // vscale of '1'. | ||||||||
3723 | setProfileInfoAfterUnrolling(LI->getLoopFor(LoopScalarBody), VectorLoop, | ||||||||
3724 | LI->getLoopFor(LoopScalarBody), | ||||||||
3725 | VF.getKnownMinValue() * UF); | ||||||||
3726 | } | ||||||||
3727 | |||||||||
3728 | void InnerLoopVectorizer::fixCrossIterationPHIs(VPTransformState &State) { | ||||||||
3729 | // In order to support recurrences we need to be able to vectorize Phi nodes. | ||||||||
3730 | // Phi nodes have cycles, so we need to vectorize them in two stages. This is | ||||||||
3731 | // stage #2: We now need to fix the recurrences by adding incoming edges to | ||||||||
3732 | // the currently empty PHI nodes. At this point every instruction in the | ||||||||
3733 | // original loop is widened to a vector form so we can use them to construct | ||||||||
3734 | // the incoming edges. | ||||||||
3735 | VPBasicBlock *Header = | ||||||||
3736 | State.Plan->getVectorLoopRegion()->getEntryBasicBlock(); | ||||||||
3737 | for (VPRecipeBase &R : Header->phis()) { | ||||||||
3738 | if (auto *ReductionPhi = dyn_cast<VPReductionPHIRecipe>(&R)) | ||||||||
3739 | fixReduction(ReductionPhi, State); | ||||||||
3740 | else if (auto *FOR = dyn_cast<VPFirstOrderRecurrencePHIRecipe>(&R)) | ||||||||
3741 | fixFixedOrderRecurrence(FOR, State); | ||||||||
3742 | } | ||||||||
3743 | } | ||||||||
3744 | |||||||||
3745 | void InnerLoopVectorizer::fixFixedOrderRecurrence( | ||||||||
3746 | VPFirstOrderRecurrencePHIRecipe *PhiR, VPTransformState &State) { | ||||||||
3747 | // This is the second phase of vectorizing first-order recurrences. An | ||||||||
3748 | // overview of the transformation is described below. Suppose we have the | ||||||||
3749 | // following loop. | ||||||||
3750 | // | ||||||||
3751 | // for (int i = 0; i < n; ++i) | ||||||||
3752 | // b[i] = a[i] - a[i - 1]; | ||||||||
3753 | // | ||||||||
3754 | // There is a first-order recurrence on "a". For this loop, the shorthand | ||||||||
3755 | // scalar IR looks like: | ||||||||
3756 | // | ||||||||
3757 | // scalar.ph: | ||||||||
3758 | // s_init = a[-1] | ||||||||
3759 | // br scalar.body | ||||||||
3760 | // | ||||||||
3761 | // scalar.body: | ||||||||
3762 | // i = phi [0, scalar.ph], [i+1, scalar.body] | ||||||||
3763 | // s1 = phi [s_init, scalar.ph], [s2, scalar.body] | ||||||||
3764 | // s2 = a[i] | ||||||||
3765 | // b[i] = s2 - s1 | ||||||||
3766 | // br cond, scalar.body, ... | ||||||||
3767 | // | ||||||||
3768 | // In this example, s1 is a recurrence because it's value depends on the | ||||||||
3769 | // previous iteration. In the first phase of vectorization, we created a | ||||||||
3770 | // vector phi v1 for s1. We now complete the vectorization and produce the | ||||||||
3771 | // shorthand vector IR shown below (for VF = 4, UF = 1). | ||||||||
3772 | // | ||||||||
3773 | // vector.ph: | ||||||||
3774 | // v_init = vector(..., ..., ..., a[-1]) | ||||||||
3775 | // br vector.body | ||||||||
3776 | // | ||||||||
3777 | // vector.body | ||||||||
3778 | // i = phi [0, vector.ph], [i+4, vector.body] | ||||||||
3779 | // v1 = phi [v_init, vector.ph], [v2, vector.body] | ||||||||
3780 | // v2 = a[i, i+1, i+2, i+3]; | ||||||||
3781 | // v3 = vector(v1(3), v2(0, 1, 2)) | ||||||||
3782 | // b[i, i+1, i+2, i+3] = v2 - v3 | ||||||||
3783 | // br cond, vector.body, middle.block | ||||||||
3784 | // | ||||||||
3785 | // middle.block: | ||||||||
3786 | // x = v2(3) | ||||||||
3787 | // br scalar.ph | ||||||||
3788 | // | ||||||||
3789 | // scalar.ph: | ||||||||
3790 | // s_init = phi [x, middle.block], [a[-1], otherwise] | ||||||||
3791 | // br scalar.body | ||||||||
3792 | // | ||||||||
3793 | // After execution completes the vector loop, we extract the next value of | ||||||||
3794 | // the recurrence (x) to use as the initial value in the scalar loop. | ||||||||
3795 | |||||||||
3796 | // Extract the last vector element in the middle block. This will be the | ||||||||
3797 | // initial value for the recurrence when jumping to the scalar loop. | ||||||||
3798 | VPValue *PreviousDef = PhiR->getBackedgeValue(); | ||||||||
3799 | Value *Incoming = State.get(PreviousDef, UF - 1); | ||||||||
3800 | auto *ExtractForScalar = Incoming; | ||||||||
3801 | auto *IdxTy = Builder.getInt32Ty(); | ||||||||
3802 | if (VF.isVector()) { | ||||||||
3803 | auto *One = ConstantInt::get(IdxTy, 1); | ||||||||
3804 | Builder.SetInsertPoint(LoopMiddleBlock->getTerminator()); | ||||||||
3805 | auto *RuntimeVF = getRuntimeVF(Builder, IdxTy, VF); | ||||||||
3806 | auto *LastIdx = Builder.CreateSub(RuntimeVF, One); | ||||||||
3807 | ExtractForScalar = Builder.CreateExtractElement(ExtractForScalar, LastIdx, | ||||||||
3808 | "vector.recur.extract"); | ||||||||
3809 | } | ||||||||
3810 | // Extract the second last element in the middle block if the | ||||||||
3811 | // Phi is used outside the loop. We need to extract the phi itself | ||||||||
3812 | // and not the last element (the phi update in the current iteration). This | ||||||||
3813 | // will be the value when jumping to the exit block from the LoopMiddleBlock, | ||||||||
3814 | // when the scalar loop is not run at all. | ||||||||
3815 | Value *ExtractForPhiUsedOutsideLoop = nullptr; | ||||||||
3816 | if (VF.isVector()) { | ||||||||
3817 | auto *RuntimeVF = getRuntimeVF(Builder, IdxTy, VF); | ||||||||
3818 | auto *Idx = Builder.CreateSub(RuntimeVF, ConstantInt::get(IdxTy, 2)); | ||||||||
3819 | ExtractForPhiUsedOutsideLoop = Builder.CreateExtractElement( | ||||||||
3820 | Incoming, Idx, "vector.recur.extract.for.phi"); | ||||||||
3821 | } else if (UF > 1) | ||||||||
3822 | // When loop is unrolled without vectorizing, initialize | ||||||||
3823 | // ExtractForPhiUsedOutsideLoop with the value just prior to unrolled value | ||||||||
3824 | // of `Incoming`. This is analogous to the vectorized case above: extracting | ||||||||
3825 | // the second last element when VF > 1. | ||||||||
3826 | ExtractForPhiUsedOutsideLoop = State.get(PreviousDef, UF - 2); | ||||||||
3827 | |||||||||
3828 | // Fix the initial value of the original recurrence in the scalar loop. | ||||||||
3829 | Builder.SetInsertPoint(&*LoopScalarPreHeader->begin()); | ||||||||
3830 | PHINode *Phi = cast<PHINode>(PhiR->getUnderlyingValue()); | ||||||||
3831 | auto *Start = Builder.CreatePHI(Phi->getType(), 2, "scalar.recur.init"); | ||||||||
3832 | auto *ScalarInit = PhiR->getStartValue()->getLiveInIRValue(); | ||||||||
3833 | for (auto *BB : predecessors(LoopScalarPreHeader)) { | ||||||||
3834 | auto *Incoming = BB == LoopMiddleBlock ? ExtractForScalar : ScalarInit; | ||||||||
3835 | Start->addIncoming(Incoming, BB); | ||||||||
3836 | } | ||||||||
3837 | |||||||||
3838 | Phi->setIncomingValueForBlock(LoopScalarPreHeader, Start); | ||||||||
3839 | Phi->setName("scalar.recur"); | ||||||||
3840 | |||||||||
3841 | // Finally, fix users of the recurrence outside the loop. The users will need | ||||||||
3842 | // either the last value of the scalar recurrence or the last value of the | ||||||||
3843 | // vector recurrence we extracted in the middle block. Since the loop is in | ||||||||
3844 | // LCSSA form, we just need to find all the phi nodes for the original scalar | ||||||||
3845 | // recurrence in the exit block, and then add an edge for the middle block. | ||||||||
3846 | // Note that LCSSA does not imply single entry when the original scalar loop | ||||||||
3847 | // had multiple exiting edges (as we always run the last iteration in the | ||||||||
3848 | // scalar epilogue); in that case, there is no edge from middle to exit and | ||||||||
3849 | // and thus no phis which needed updated. | ||||||||
3850 | if (!Cost->requiresScalarEpilogue(VF)) | ||||||||
3851 | for (PHINode &LCSSAPhi : LoopExitBlock->phis()) | ||||||||
3852 | if (llvm::is_contained(LCSSAPhi.incoming_values(), Phi)) { | ||||||||
3853 | LCSSAPhi.addIncoming(ExtractForPhiUsedOutsideLoop, LoopMiddleBlock); | ||||||||
3854 | State.Plan->removeLiveOut(&LCSSAPhi); | ||||||||
3855 | } | ||||||||
3856 | } | ||||||||
3857 | |||||||||
3858 | void InnerLoopVectorizer::fixReduction(VPReductionPHIRecipe *PhiR, | ||||||||
3859 | VPTransformState &State) { | ||||||||
3860 | PHINode *OrigPhi = cast<PHINode>(PhiR->getUnderlyingValue()); | ||||||||
3861 | // Get it's reduction variable descriptor. | ||||||||
3862 | 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", 3863, __extension__ __PRETTY_FUNCTION__)) | ||||||||
3863 | "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", 3863, __extension__ __PRETTY_FUNCTION__)); | ||||||||
3864 | const RecurrenceDescriptor &RdxDesc = PhiR->getRecurrenceDescriptor(); | ||||||||
3865 | |||||||||
3866 | RecurKind RK = RdxDesc.getRecurrenceKind(); | ||||||||
3867 | TrackingVH<Value> ReductionStartValue = RdxDesc.getRecurrenceStartValue(); | ||||||||
3868 | Instruction *LoopExitInst = RdxDesc.getLoopExitInstr(); | ||||||||
3869 | State.setDebugLocFromInst(ReductionStartValue); | ||||||||
3870 | |||||||||
3871 | VPValue *LoopExitInstDef = PhiR->getBackedgeValue(); | ||||||||
3872 | // This is the vector-clone of the value that leaves the loop. | ||||||||
3873 | Type *VecTy = State.get(LoopExitInstDef, 0)->getType(); | ||||||||
3874 | |||||||||
3875 | // Wrap flags are in general invalid after vectorization, clear them. | ||||||||
3876 | clearReductionWrapFlags(PhiR, State); | ||||||||
3877 | |||||||||
3878 | // Before each round, move the insertion point right between | ||||||||
3879 | // the PHIs and the values we are going to write. | ||||||||
3880 | // This allows us to write both PHINodes and the extractelement | ||||||||
3881 | // instructions. | ||||||||
3882 | Builder.SetInsertPoint(&*LoopMiddleBlock->getFirstInsertionPt()); | ||||||||
3883 | |||||||||
3884 | State.setDebugLocFromInst(LoopExitInst); | ||||||||
3885 | |||||||||
3886 | Type *PhiTy = OrigPhi->getType(); | ||||||||
3887 | |||||||||
3888 | VPBasicBlock *LatchVPBB = | ||||||||
3889 | PhiR->getParent()->getEnclosingLoopRegion()->getExitingBasicBlock(); | ||||||||
3890 | BasicBlock *VectorLoopLatch = State.CFG.VPBB2IRBB[LatchVPBB]; | ||||||||
3891 | // If tail is folded by masking, the vector value to leave the loop should be | ||||||||
3892 | // a Select choosing between the vectorized LoopExitInst and vectorized Phi, | ||||||||
3893 | // instead of the former. For an inloop reduction the reduction will already | ||||||||
3894 | // be predicated, and does not need to be handled here. | ||||||||
3895 | if (Cost->foldTailByMasking() && !PhiR->isInLoop()) { | ||||||||
3896 | for (unsigned Part = 0; Part < UF; ++Part) { | ||||||||
3897 | Value *VecLoopExitInst = State.get(LoopExitInstDef, Part); | ||||||||
3898 | SelectInst *Sel = nullptr; | ||||||||
3899 | for (User *U : VecLoopExitInst->users()) { | ||||||||
3900 | if (isa<SelectInst>(U)) { | ||||||||
3901 | 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", 3901, __extension__ __PRETTY_FUNCTION__)); | ||||||||
3902 | Sel = cast<SelectInst>(U); | ||||||||
3903 | } else | ||||||||
3904 | 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", 3904, __extension__ __PRETTY_FUNCTION__)); | ||||||||
3905 | } | ||||||||
3906 | 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", 3906, __extension__ __PRETTY_FUNCTION__)); | ||||||||
3907 | State.reset(LoopExitInstDef, Sel, Part); | ||||||||
3908 | |||||||||
3909 | if (isa<FPMathOperator>(Sel)) | ||||||||
3910 | Sel->setFastMathFlags(RdxDesc.getFastMathFlags()); | ||||||||
3911 | |||||||||
3912 | // If the target can create a predicated operator for the reduction at no | ||||||||
3913 | // extra cost in the loop (for example a predicated vadd), it can be | ||||||||
3914 | // cheaper for the select to remain in the loop than be sunk out of it, | ||||||||
3915 | // and so use the select value for the phi instead of the old | ||||||||
3916 | // LoopExitValue. | ||||||||
3917 | if (PreferPredicatedReductionSelect || | ||||||||
3918 | TTI->preferPredicatedReductionSelect( | ||||||||
3919 | RdxDesc.getOpcode(), PhiTy, | ||||||||
3920 | TargetTransformInfo::ReductionFlags())) { | ||||||||
3921 | auto *VecRdxPhi = | ||||||||
3922 | cast<PHINode>(State.get(PhiR, Part)); | ||||||||
3923 | VecRdxPhi->setIncomingValueForBlock(VectorLoopLatch, Sel); | ||||||||
3924 | } | ||||||||
3925 | } | ||||||||
3926 | } | ||||||||
3927 | |||||||||
3928 | // If the vector reduction can be performed in a smaller type, we truncate | ||||||||
3929 | // then extend the loop exit value to enable InstCombine to evaluate the | ||||||||
3930 | // entire expression in the smaller type. | ||||||||
3931 | if (VF.isVector() && PhiTy != RdxDesc.getRecurrenceType()) { | ||||||||
3932 | 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", 3932, __extension__ __PRETTY_FUNCTION__)); | ||||||||
3933 | Type *RdxVecTy = VectorType::get(RdxDesc.getRecurrenceType(), VF); | ||||||||
3934 | Builder.SetInsertPoint(VectorLoopLatch->getTerminator()); | ||||||||
3935 | VectorParts RdxParts(UF); | ||||||||
3936 | for (unsigned Part = 0; Part < UF; ++Part) { | ||||||||
3937 | RdxParts[Part] = State.get(LoopExitInstDef, Part); | ||||||||
3938 | Value *Trunc = Builder.CreateTrunc(RdxParts[Part], RdxVecTy); | ||||||||
3939 | Value *Extnd = RdxDesc.isSigned() ? Builder.CreateSExt(Trunc, VecTy) | ||||||||
3940 | : Builder.CreateZExt(Trunc, VecTy); | ||||||||
3941 | for (User *U : llvm::make_early_inc_range(RdxParts[Part]->users())) | ||||||||
3942 | if (U != Trunc) { | ||||||||
3943 | U->replaceUsesOfWith(RdxParts[Part], Extnd); | ||||||||
3944 | RdxParts[Part] = Extnd; | ||||||||
3945 | } | ||||||||
3946 | } | ||||||||
3947 | Builder.SetInsertPoint(&*LoopMiddleBlock->getFirstInsertionPt()); | ||||||||
3948 | for (unsigned Part = 0; Part < UF; ++Part) { | ||||||||
3949 | RdxParts[Part] = Builder.CreateTrunc(RdxParts[Part], RdxVecTy); | ||||||||
3950 | State.reset(LoopExitInstDef, RdxParts[Part], Part); | ||||||||
3951 | } | ||||||||
3952 | } | ||||||||
3953 | |||||||||
3954 | // Reduce all of the unrolled parts into a single vector. | ||||||||
3955 | Value *ReducedPartRdx = State.get(LoopExitInstDef, 0); | ||||||||
3956 | unsigned Op = RecurrenceDescriptor::getOpcode(RK); | ||||||||
3957 | |||||||||
3958 | // The middle block terminator has already been assigned a DebugLoc here (the | ||||||||
3959 | // OrigLoop's single latch terminator). We want the whole middle block to | ||||||||
3960 | // appear to execute on this line because: (a) it is all compiler generated, | ||||||||
3961 | // (b) these instructions are always executed after evaluating the latch | ||||||||
3962 | // conditional branch, and (c) other passes may add new predecessors which | ||||||||
3963 | // terminate on this line. This is the easiest way to ensure we don't | ||||||||
3964 | // accidentally cause an extra step back into the loop while debugging. | ||||||||
3965 | State.setDebugLocFromInst(LoopMiddleBlock->getTerminator()); | ||||||||
3966 | if (PhiR->isOrdered()) | ||||||||
3967 | ReducedPartRdx = State.get(LoopExitInstDef, UF - 1); | ||||||||
3968 | else { | ||||||||
3969 | // Floating-point operations should have some FMF to enable the reduction. | ||||||||
3970 | IRBuilderBase::FastMathFlagGuard FMFG(Builder); | ||||||||
3971 | Builder.setFastMathFlags(RdxDesc.getFastMathFlags()); | ||||||||
3972 | for (unsigned Part = 1; Part < UF; ++Part) { | ||||||||
3973 | Value *RdxPart = State.get(LoopExitInstDef, Part); | ||||||||
3974 | if (Op != Instruction::ICmp && Op != Instruction::FCmp) { | ||||||||
3975 | ReducedPartRdx = Builder.CreateBinOp( | ||||||||
3976 | (Instruction::BinaryOps)Op, RdxPart, ReducedPartRdx, "bin.rdx"); | ||||||||
3977 | } else if (RecurrenceDescriptor::isSelectCmpRecurrenceKind(RK)) | ||||||||
3978 | ReducedPartRdx = createSelectCmpOp(Builder, ReductionStartValue, RK, | ||||||||
3979 | ReducedPartRdx, RdxPart); | ||||||||
3980 | else | ||||||||
3981 | ReducedPartRdx = createMinMaxOp(Builder, RK, ReducedPartRdx, RdxPart); | ||||||||
3982 | } | ||||||||
3983 | } | ||||||||
3984 | |||||||||
3985 | // Create the reduction after the loop. Note that inloop reductions create the | ||||||||
3986 | // target reduction in the loop using a Reduction recipe. | ||||||||
3987 | if (VF.isVector() && !PhiR->isInLoop()) { | ||||||||
3988 | ReducedPartRdx = | ||||||||
3989 | createTargetReduction(Builder, TTI, RdxDesc, ReducedPartRdx, OrigPhi); | ||||||||
3990 | // If the reduction can be performed in a smaller type, we need to extend | ||||||||
3991 | // the reduction to the wider type before we branch to the original loop. | ||||||||
3992 | if (PhiTy != RdxDesc.getRecurrenceType()) | ||||||||
3993 | ReducedPartRdx = RdxDesc.isSigned() | ||||||||
3994 | ? Builder.CreateSExt(ReducedPartRdx, PhiTy) | ||||||||
3995 | : Builder.CreateZExt(ReducedPartRdx, PhiTy); | ||||||||
3996 | } | ||||||||
3997 | |||||||||
3998 | PHINode *ResumePhi = | ||||||||
3999 | dyn_cast<PHINode>(PhiR->getStartValue()->getUnderlyingValue()); | ||||||||
4000 | |||||||||
4001 | // Create a phi node that merges control-flow from the backedge-taken check | ||||||||
4002 | // block and the middle block. | ||||||||
4003 | PHINode *BCBlockPhi = PHINode::Create(PhiTy, 2, "bc.merge.rdx", | ||||||||
4004 | LoopScalarPreHeader->getTerminator()); | ||||||||
4005 | |||||||||
4006 | // If we are fixing reductions in the epilogue loop then we should already | ||||||||
4007 | // have created a bc.merge.rdx Phi after the main vector body. Ensure that | ||||||||
4008 | // we carry over the incoming values correctly. | ||||||||
4009 | for (auto *Incoming : predecessors(LoopScalarPreHeader)) { | ||||||||
4010 | if (Incoming == LoopMiddleBlock) | ||||||||
4011 | BCBlockPhi->addIncoming(ReducedPartRdx, Incoming); | ||||||||
4012 | else if (ResumePhi && llvm::is_contained(ResumePhi->blocks(), Incoming)) | ||||||||
4013 | BCBlockPhi->addIncoming(ResumePhi->getIncomingValueForBlock(Incoming), | ||||||||
4014 | Incoming); | ||||||||
4015 | else | ||||||||
4016 | BCBlockPhi->addIncoming(ReductionStartValue, Incoming); | ||||||||
4017 | } | ||||||||
4018 | |||||||||
4019 | // Set the resume value for this reduction | ||||||||
4020 | ReductionResumeValues.insert({&RdxDesc, BCBlockPhi}); | ||||||||
4021 | |||||||||
4022 | // If there were stores of the reduction value to a uniform memory address | ||||||||
4023 | // inside the loop, create the final store here. | ||||||||
4024 | if (StoreInst *SI = RdxDesc.IntermediateStore) { | ||||||||
4025 | StoreInst *NewSI = | ||||||||
4026 | Builder.CreateStore(ReducedPartRdx, SI->getPointerOperand()); | ||||||||
4027 | propagateMetadata(NewSI, SI); | ||||||||
4028 | |||||||||
4029 | // If the reduction value is used in other places, | ||||||||
4030 | // then let the code below create PHI's for that. | ||||||||
4031 | } | ||||||||
4032 | |||||||||
4033 | // Now, we need to fix the users of the reduction variable | ||||||||
4034 | // inside and outside of the scalar remainder loop. | ||||||||
4035 | |||||||||
4036 | // We know that the loop is in LCSSA form. We need to update the PHI nodes | ||||||||
4037 | // in the exit blocks. See comment on analogous loop in | ||||||||
4038 | // fixFixedOrderRecurrence for a more complete explaination of the logic. | ||||||||
4039 | if (!Cost->requiresScalarEpilogue(VF)) | ||||||||
4040 | for (PHINode &LCSSAPhi : LoopExitBlock->phis()) | ||||||||
4041 | if (llvm::is_contained(LCSSAPhi.incoming_values(), LoopExitInst)) { | ||||||||
4042 | LCSSAPhi.addIncoming(ReducedPartRdx, LoopMiddleBlock); | ||||||||
4043 | State.Plan->removeLiveOut(&LCSSAPhi); | ||||||||
4044 | } | ||||||||
4045 | |||||||||
4046 | // Fix the scalar loop reduction variable with the incoming reduction sum | ||||||||
4047 | // from the vector body and from the backedge value. | ||||||||
4048 | int IncomingEdgeBlockIdx = | ||||||||
4049 | OrigPhi->getBasicBlockIndex(OrigLoop->getLoopLatch()); | ||||||||
4050 | 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", 4050, __extension__ __PRETTY_FUNCTION__)); | ||||||||
4051 | // Pick the other block. | ||||||||
4052 | int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1); | ||||||||
4053 | OrigPhi->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi); | ||||||||
4054 | OrigPhi->setIncomingValue(IncomingEdgeBlockIdx, LoopExitInst); | ||||||||
4055 | } | ||||||||
4056 | |||||||||
4057 | void InnerLoopVectorizer::clearReductionWrapFlags(VPReductionPHIRecipe *PhiR, | ||||||||
4058 | VPTransformState &State) { | ||||||||
4059 | const RecurrenceDescriptor &RdxDesc = PhiR->getRecurrenceDescriptor(); | ||||||||
4060 | RecurKind RK = RdxDesc.getRecurrenceKind(); | ||||||||
4061 | if (RK != RecurKind::Add && RK != RecurKind::Mul) | ||||||||
4062 | return; | ||||||||
4063 | |||||||||
4064 | SmallVector<VPValue *, 8> Worklist; | ||||||||
4065 | SmallPtrSet<VPValue *, 8> Visited; | ||||||||
4066 | Worklist.push_back(PhiR); | ||||||||
4067 | Visited.insert(PhiR); | ||||||||
4068 | |||||||||
4069 | while (!Worklist.empty()) { | ||||||||
4070 | VPValue *Cur = Worklist.pop_back_val(); | ||||||||
4071 | for (unsigned Part = 0; Part < UF; ++Part) { | ||||||||
4072 | Value *V = State.get(Cur, Part); | ||||||||
4073 | if (!isa<OverflowingBinaryOperator>(V)) | ||||||||
4074 | break; | ||||||||
4075 | cast<Instruction>(V)->dropPoisonGeneratingFlags(); | ||||||||
4076 | } | ||||||||
4077 | |||||||||
4078 | for (VPUser *U : Cur->users()) { | ||||||||
4079 | auto *UserRecipe = dyn_cast<VPRecipeBase>(U); | ||||||||
4080 | if (!UserRecipe) | ||||||||
4081 | continue; | ||||||||
4082 | for (VPValue *V : UserRecipe->definedValues()) | ||||||||
4083 | if (Visited.insert(V).second) | ||||||||
4084 | Worklist.push_back(V); | ||||||||
4085 | } | ||||||||
4086 | } | ||||||||
4087 | } | ||||||||
4088 | |||||||||
4089 | void InnerLoopVectorizer::sinkScalarOperands(Instruction *PredInst) { | ||||||||
4090 | // The basic block and loop containing the predicated instruction. | ||||||||
4091 | auto *PredBB = PredInst->getParent(); | ||||||||
4092 | auto *VectorLoop = LI->getLoopFor(PredBB); | ||||||||
4093 | |||||||||
4094 | // Initialize a worklist with the operands of the predicated instruction. | ||||||||
4095 | SetVector<Value *> Worklist(PredInst->op_begin(), PredInst->op_end()); | ||||||||
4096 | |||||||||
4097 | // Holds instructions that we need to analyze again. An instruction may be | ||||||||
4098 | // reanalyzed if we don't yet know if we can sink it or not. | ||||||||
4099 | SmallVector<Instruction *, 8> InstsToReanalyze; | ||||||||
4100 | |||||||||
4101 | // Returns true if a given use occurs in the predicated block. Phi nodes use | ||||||||
4102 | // their operands in their corresponding predecessor blocks. | ||||||||
4103 | auto isBlockOfUsePredicated = [&](Use &U) -> bool { | ||||||||
4104 | auto *I = cast<Instruction>(U.getUser()); | ||||||||
4105 | BasicBlock *BB = I->getParent(); | ||||||||
4106 | if (auto *Phi = dyn_cast<PHINode>(I)) | ||||||||
4107 | BB = Phi->getIncomingBlock( | ||||||||
4108 | PHINode::getIncomingValueNumForOperand(U.getOperandNo())); | ||||||||
4109 | return BB == PredBB; | ||||||||
4110 | }; | ||||||||
4111 | |||||||||
4112 | // Iteratively sink the scalarized operands of the predicated instruction | ||||||||
4113 | // into the block we created for it. When an instruction is sunk, it's | ||||||||
4114 | // operands are then added to the worklist. The algorithm ends after one pass | ||||||||
4115 | // through the worklist doesn't sink a single instruction. | ||||||||
4116 | bool Changed; | ||||||||
4117 | do { | ||||||||
4118 | // Add the instructions that need to be reanalyzed to the worklist, and | ||||||||
4119 | // reset the changed indicator. | ||||||||
4120 | Worklist.insert(InstsToReanalyze.begin(), InstsToReanalyze.end()); | ||||||||
4121 | InstsToReanalyze.clear(); | ||||||||
4122 | Changed = false; | ||||||||
4123 | |||||||||
4124 | while (!Worklist.empty()) { | ||||||||
4125 | auto *I = dyn_cast<Instruction>(Worklist.pop_back_val()); | ||||||||
4126 | |||||||||
4127 | // We can't sink an instruction if it is a phi node, is not in the loop, | ||||||||
4128 | // or may have side effects. | ||||||||
4129 | if (!I || isa<PHINode>(I) || !VectorLoop->contains(I) || | ||||||||
4130 | I->mayHaveSideEffects()) | ||||||||
4131 | continue; | ||||||||
4132 | |||||||||
4133 | // If the instruction is already in PredBB, check if we can sink its | ||||||||
4134 | // operands. In that case, VPlan's sinkScalarOperands() succeeded in | ||||||||
4135 | // sinking the scalar instruction I, hence it appears in PredBB; but it | ||||||||
4136 | // may have failed to sink I's operands (recursively), which we try | ||||||||
4137 | // (again) here. | ||||||||
4138 | if (I->getParent() == PredBB) { | ||||||||
4139 | Worklist.insert(I->op_begin(), I->op_end()); | ||||||||
4140 | continue; | ||||||||
4141 | } | ||||||||
4142 | |||||||||
4143 | // It's legal to sink the instruction if all its uses occur in the | ||||||||
4144 | // predicated block. Otherwise, there's nothing to do yet, and we may | ||||||||
4145 | // need to reanalyze the instruction. | ||||||||
4146 | if (!llvm::all_of(I->uses(), isBlockOfUsePredicated)) { | ||||||||
4147 | InstsToReanalyze.push_back(I); | ||||||||
4148 | continue; | ||||||||
4149 | } | ||||||||
4150 | |||||||||
4151 | // Move the instruction to the beginning of the predicated block, and add | ||||||||
4152 | // it's operands to the worklist. | ||||||||
4153 | I->moveBefore(&*PredBB->getFirstInsertionPt()); | ||||||||
4154 | Worklist.insert(I->op_begin(), I->op_end()); | ||||||||
4155 | |||||||||
4156 | // The sinking may have enabled other instructions to be sunk, so we will | ||||||||
4157 | // need to iterate. | ||||||||
4158 | Changed = true; | ||||||||
4159 | } | ||||||||
4160 | } while (Changed); | ||||||||
4161 | } | ||||||||
4162 | |||||||||
4163 | void InnerLoopVectorizer::fixNonInductionPHIs(VPlan &Plan, | ||||||||
4164 | VPTransformState &State) { | ||||||||
4165 | auto Iter = vp_depth_first_deep(Plan.getEntry()); | ||||||||
4166 | for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(Iter)) { | ||||||||
4167 | for (VPRecipeBase &P : VPBB->phis()) { | ||||||||
4168 | VPWidenPHIRecipe *VPPhi = dyn_cast<VPWidenPHIRecipe>(&P); | ||||||||
4169 | if (!VPPhi) | ||||||||
4170 | continue; | ||||||||
4171 | PHINode *NewPhi = cast<PHINode>(State.get(VPPhi, 0)); | ||||||||
4172 | // Make sure the builder has a valid insert point. | ||||||||
4173 | Builder.SetInsertPoint(NewPhi); | ||||||||
4174 | for (unsigned i = 0; i < VPPhi->getNumOperands(); ++i) { | ||||||||
4175 | VPValue *Inc = VPPhi->getIncomingValue(i); | ||||||||
4176 | VPBasicBlock *VPBB = VPPhi->getIncomingBlock(i); | ||||||||
4177 | NewPhi->addIncoming(State.get(Inc, 0), State.CFG.VPBB2IRBB[VPBB]); | ||||||||
4178 | } | ||||||||
4179 | } | ||||||||
4180 | } | ||||||||
4181 | } | ||||||||
4182 | |||||||||
4183 | bool InnerLoopVectorizer::useOrderedReductions( | ||||||||
4184 | const RecurrenceDescriptor &RdxDesc) { | ||||||||
4185 | return Cost->useOrderedReductions(RdxDesc); | ||||||||
4186 | } | ||||||||
4187 | |||||||||
4188 | void LoopVectorizationCostModel::collectLoopScalars(ElementCount VF) { | ||||||||
4189 | // We should not collect Scalars more than once per VF. Right now, this | ||||||||
4190 | // function is called from collectUniformsAndScalars(), which already does | ||||||||
4191 | // this check. Collecting Scalars for VF=1 does not make any sense. | ||||||||
4192 | 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", 4193, __extension__ __PRETTY_FUNCTION__)) | ||||||||
4193 | "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", 4193, __extension__ __PRETTY_FUNCTION__)); | ||||||||
4194 | |||||||||
4195 | // This avoids any chances of creating a REPLICATE recipe during planning | ||||||||
4196 | // since that would result in generation of scalarized code during execution, | ||||||||
4197 | // which is not supported for scalable vectors. | ||||||||
4198 | if (VF.isScalable()) { | ||||||||
4199 | Scalars[VF].insert(Uniforms[VF].begin(), Uniforms[VF].end()); | ||||||||
4200 | return; | ||||||||
4201 | } | ||||||||
4202 | |||||||||
4203 | SmallSetVector<Instruction *, 8> Worklist; | ||||||||
4204 | |||||||||
4205 | // These sets are used to seed the analysis with pointers used by memory | ||||||||
4206 | // accesses that will remain scalar. | ||||||||
4207 | SmallSetVector<Instruction *, 8> ScalarPtrs; | ||||||||
4208 | SmallPtrSet<Instruction *, 8> PossibleNonScalarPtrs; | ||||||||
4209 | auto *Latch = TheLoop->getLoopLatch(); | ||||||||
4210 | |||||||||
4211 | // A helper that returns true if the use of Ptr by MemAccess will be scalar. | ||||||||
4212 | // The pointer operands of loads and stores will be scalar as long as the | ||||||||
4213 | // memory access is not a gather or scatter operation. The value operand of a | ||||||||
4214 | // store will remain scalar if the store is scalarized. | ||||||||
4215 | auto isScalarUse = [&](Instruction *MemAccess, Value *Ptr) { | ||||||||
4216 | InstWidening WideningDecision = getWideningDecision(MemAccess, VF); | ||||||||
4217 | 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", 4218, __extension__ __PRETTY_FUNCTION__)) | ||||||||
4218 | "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", 4218, __extension__ __PRETTY_FUNCTION__)); | ||||||||
4219 | if (auto *Store = dyn_cast<StoreInst>(MemAccess)) | ||||||||
4220 | if (Ptr == Store->getValueOperand()) | ||||||||
4221 | return WideningDecision == CM_Scalarize; | ||||||||
4222 | 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", 4223, __extension__ __PRETTY_FUNCTION__)) | ||||||||
4223 | "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", 4223, __extension__ __PRETTY_FUNCTION__)); | ||||||||
4224 | return WideningDecision != CM_GatherScatter; | ||||||||
4225 | }; | ||||||||
4226 | |||||||||
4227 | // A helper that returns true if the given value is a bitcast or | ||||||||
4228 | // getelementptr instruction contained in the loop. | ||||||||
4229 | auto isLoopVaryingBitCastOrGEP = [&](Value *V) { | ||||||||
4230 | return ((isa<BitCastInst>(V) && V->getType()->isPointerTy()) || | ||||||||
4231 | isa<GetElementPtrInst>(V)) && | ||||||||
4232 | !TheLoop->isLoopInvariant(V); | ||||||||
4233 | }; | ||||||||
4234 | |||||||||
4235 | // A helper that evaluates a memory access's use of a pointer. If the use will | ||||||||
4236 | // be a scalar use and the pointer is only used by memory accesses, we place | ||||||||
4237 | // the pointer in ScalarPtrs. Otherwise, the pointer is placed in | ||||||||
4238 | // PossibleNonScalarPtrs. | ||||||||
4239 | auto evaluatePtrUse = [&](Instruction *MemAccess, Value *Ptr) { | ||||||||
4240 | // We only care about bitcast and getelementptr instructions contained in | ||||||||
4241 | // the loop. | ||||||||
4242 | if (!isLoopVaryingBitCastOrGEP(Ptr)) | ||||||||
4243 | return; | ||||||||
4244 | |||||||||
4245 | // If the pointer has already been identified as scalar (e.g., if it was | ||||||||
4246 | // also identified as uniform), there's nothing to do. | ||||||||
4247 | auto *I = cast<Instruction>(Ptr); | ||||||||
4248 | if (Worklist.count(I)) | ||||||||
4249 | return; | ||||||||
4250 | |||||||||
4251 | // If the use of the pointer will be a scalar use, and all users of the | ||||||||
4252 | // pointer are memory accesses, place the pointer in ScalarPtrs. Otherwise, | ||||||||
4253 | // place the pointer in PossibleNonScalarPtrs. | ||||||||
4254 | if (isScalarUse(MemAccess, Ptr) && llvm::all_of(I->users(), [&](User *U) { | ||||||||
4255 | return isa<LoadInst>(U) || isa<StoreInst>(U); | ||||||||
4256 | })) | ||||||||
4257 | ScalarPtrs.insert(I); | ||||||||
4258 | else | ||||||||
4259 | PossibleNonScalarPtrs.insert(I); | ||||||||
4260 | }; | ||||||||
4261 | |||||||||
4262 | // We seed the scalars analysis with three classes of instructions: (1) | ||||||||
4263 | // instructions marked uniform-after-vectorization and (2) bitcast, | ||||||||
4264 | // getelementptr and (pointer) phi instructions used by memory accesses | ||||||||
4265 | // requiring a scalar use. | ||||||||
4266 | // | ||||||||
4267 | // (1) Add to the worklist all instructions that have been identified as | ||||||||
4268 | // uniform-after-vectorization. | ||||||||
4269 | Worklist.insert(Uniforms[VF].begin(), Uniforms[VF].end()); | ||||||||
4270 | |||||||||
4271 | // (2) Add to the worklist all bitcast and getelementptr instructions used by | ||||||||
4272 | // memory accesses requiring a scalar use. The pointer operands of loads and | ||||||||
4273 | // stores will be scalar as long as the memory accesses is not a gather or | ||||||||
4274 | // scatter operation. The value operand of a store will remain scalar if the | ||||||||
4275 | // store is scalarized. | ||||||||
4276 | for (auto *BB : TheLoop->blocks()) | ||||||||
4277 | for (auto &I : *BB) { | ||||||||
4278 | if (auto *Load = dyn_cast<LoadInst>(&I)) { | ||||||||
4279 | evaluatePtrUse(Load, Load->getPointerOperand()); | ||||||||
4280 | } else if (auto *Store = dyn_cast<StoreInst>(&I)) { | ||||||||
4281 | evaluatePtrUse(Store, Store->getPointerOperand()); | ||||||||
4282 | evaluatePtrUse(Store, Store->getValueOperand()); | ||||||||
4283 | } | ||||||||
4284 | } | ||||||||
4285 | for (auto *I : ScalarPtrs) | ||||||||
4286 | if (!PossibleNonScalarPtrs.count(I)) { | ||||||||
4287 | 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); | ||||||||
4288 | Worklist.insert(I); | ||||||||
4289 | } | ||||||||
4290 | |||||||||
4291 | // Insert the forced scalars. | ||||||||
4292 | // FIXME: Currently VPWidenPHIRecipe() often creates a dead vector | ||||||||
4293 | // induction variable when the PHI user is scalarized. | ||||||||
4294 | auto ForcedScalar = ForcedScalars.find(VF); | ||||||||
4295 | if (ForcedScalar != ForcedScalars.end()) | ||||||||
4296 | for (auto *I : ForcedScalar->second) { | ||||||||
4297 | LLVM_DEBUG(dbgs() << "LV: Found (forced) scalar instruction: " << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found (forced) scalar instruction: " << *I << "\n"; } } while (false); | ||||||||
4298 | Worklist.insert(I); | ||||||||
4299 | } | ||||||||
4300 | |||||||||
4301 | // Expand the worklist by looking through any bitcasts and getelementptr | ||||||||
4302 | // instructions we've already identified as scalar. This is similar to the | ||||||||
4303 | // expansion step in collectLoopUniforms(); however, here we're only | ||||||||
4304 | // expanding to include additional bitcasts and getelementptr instructions. | ||||||||
4305 | unsigned Idx = 0; | ||||||||
4306 | while (Idx != Worklist.size()) { | ||||||||
4307 | Instruction *Dst = Worklist[Idx++]; | ||||||||
4308 | if (!isLoopVaryingBitCastOrGEP(Dst->getOperand(0))) | ||||||||
4309 | continue; | ||||||||
4310 | auto *Src = cast<Instruction>(Dst->getOperand(0)); | ||||||||
4311 | if (llvm::all_of(Src->users(), [&](User *U) -> bool { | ||||||||
4312 | auto *J = cast<Instruction>(U); | ||||||||
4313 | return !TheLoop->contains(J) || Worklist.count(J) || | ||||||||
4314 | ((isa<LoadInst>(J) || isa<StoreInst>(J)) && | ||||||||
4315 | isScalarUse(J, Src)); | ||||||||
4316 | })) { | ||||||||
4317 | Worklist.insert(Src); | ||||||||
4318 | 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); | ||||||||
4319 | } | ||||||||
4320 | } | ||||||||
4321 | |||||||||
4322 | // An induction variable will remain scalar if all users of the induction | ||||||||
4323 | // variable and induction variable update remain scalar. | ||||||||
4324 | for (const auto &Induction : Legal->getInductionVars()) { | ||||||||
4325 | auto *Ind = Induction.first; | ||||||||
4326 | auto *IndUpdate = cast<Instruction>(Ind->getIncomingValueForBlock(Latch)); | ||||||||
4327 | |||||||||
4328 | // If tail-folding is applied, the primary induction variable will be used | ||||||||
4329 | // to feed a vector compare. | ||||||||
4330 | if (Ind == Legal->getPrimaryInduction() && foldTailByMasking()) | ||||||||
4331 | continue; | ||||||||
4332 | |||||||||
4333 | // Returns true if \p Indvar is a pointer induction that is used directly by | ||||||||
4334 | // load/store instruction \p I. | ||||||||
4335 | auto IsDirectLoadStoreFromPtrIndvar = [&](Instruction *Indvar, | ||||||||
4336 | Instruction *I) { | ||||||||
4337 | return Induction.second.getKind() == | ||||||||
4338 | InductionDescriptor::IK_PtrInduction && | ||||||||
4339 | (isa<LoadInst>(I) || isa<StoreInst>(I)) && | ||||||||
4340 | Indvar == getLoadStorePointerOperand(I) && isScalarUse(I, Indvar); | ||||||||
4341 | }; | ||||||||
4342 | |||||||||
4343 | // Determine if all users of the induction variable are scalar after | ||||||||
4344 | // vectorization. | ||||||||
4345 | auto ScalarInd = llvm::all_of(Ind->users(), [&](User *U) -> bool { | ||||||||
4346 | auto *I = cast<Instruction>(U); | ||||||||
4347 | return I == IndUpdate || !TheLoop->contains(I) || Worklist.count(I) || | ||||||||
4348 | IsDirectLoadStoreFromPtrIndvar(Ind, I); | ||||||||
4349 | }); | ||||||||
4350 | if (!ScalarInd) | ||||||||
4351 | continue; | ||||||||
4352 | |||||||||
4353 | // Determine if all users of the induction variable update instruction are | ||||||||
4354 | // scalar after vectorization. | ||||||||
4355 | auto ScalarIndUpdate = | ||||||||
4356 | llvm::all_of(IndUpdate->users(), [&](User *U) -> bool { | ||||||||
4357 | auto *I = cast<Instruction>(U); | ||||||||
4358 | return I == Ind || !TheLoop->contains(I) || Worklist.count(I) || | ||||||||
4359 | IsDirectLoadStoreFromPtrIndvar(IndUpdate, I); | ||||||||
4360 | }); | ||||||||
4361 | if (!ScalarIndUpdate) | ||||||||
4362 | continue; | ||||||||
4363 | |||||||||
4364 | // The induction variable and its update instruction will remain scalar. | ||||||||
4365 | Worklist.insert(Ind); | ||||||||
4366 | Worklist.insert(IndUpdate); | ||||||||
4367 | 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); | ||||||||
4368 | LLVM_DEBUG(dbgs() << "LV: Found scalar instruction: " << *IndUpdatedo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found scalar instruction: " << *IndUpdate << "\n"; } } while (false) | ||||||||
4369 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found scalar instruction: " << *IndUpdate << "\n"; } } while (false); | ||||||||
4370 | } | ||||||||
4371 | |||||||||
4372 | Scalars[VF].insert(Worklist.begin(), Worklist.end()); | ||||||||
4373 | } | ||||||||
4374 | |||||||||
4375 | bool LoopVectorizationCostModel::isScalarWithPredication( | ||||||||
4376 | Instruction *I, ElementCount VF) const { | ||||||||
4377 | if (!isPredicatedInst(I)) | ||||||||
4378 | return false; | ||||||||
4379 | |||||||||
4380 | // Do we have a non-scalar lowering for this predicated | ||||||||
4381 | // instruction? No - it is scalar with predication. | ||||||||
4382 | switch(I->getOpcode()) { | ||||||||
4383 | default: | ||||||||
4384 | return true; | ||||||||
4385 | case Instruction::Load: | ||||||||
4386 | case Instruction::Store: { | ||||||||
4387 | auto *Ptr = getLoadStorePointerOperand(I); | ||||||||
4388 | auto *Ty = getLoadStoreType(I); | ||||||||
4389 | Type *VTy = Ty; | ||||||||
4390 | if (VF.isVector()) | ||||||||
4391 | VTy = VectorType::get(Ty, VF); | ||||||||
4392 | const Align Alignment = getLoadStoreAlignment(I); | ||||||||
4393 | return isa<LoadInst>(I) ? !(isLegalMaskedLoad(Ty, Ptr, Alignment) || | ||||||||
4394 | TTI.isLegalMaskedGather(VTy, Alignment)) | ||||||||
4395 | : !(isLegalMaskedStore(Ty, Ptr, Alignment) || | ||||||||
4396 | TTI.isLegalMaskedScatter(VTy, Alignment)); | ||||||||
4397 | } | ||||||||
4398 | case Instruction::UDiv: | ||||||||
4399 | case Instruction::SDiv: | ||||||||
4400 | case Instruction::SRem: | ||||||||
4401 | case Instruction::URem: { | ||||||||
4402 | // We have the option to use the safe-divisor idiom to avoid predication. | ||||||||
4403 | // The cost based decision here will always select safe-divisor for | ||||||||
4404 | // scalable vectors as scalarization isn't legal. | ||||||||
4405 | const auto [ScalarCost, SafeDivisorCost] = getDivRemSpeculationCost(I, VF); | ||||||||
4406 | return isDivRemScalarWithPredication(ScalarCost, SafeDivisorCost); | ||||||||
4407 | } | ||||||||
4408 | } | ||||||||
4409 | } | ||||||||
4410 | |||||||||
4411 | bool LoopVectorizationCostModel::isPredicatedInst(Instruction *I) const { | ||||||||
4412 | if (!blockNeedsPredicationForAnyReason(I->getParent())) | ||||||||
4413 | return false; | ||||||||
4414 | |||||||||
4415 | // Can we prove this instruction is safe to unconditionally execute? | ||||||||
4416 | // If not, we must use some form of predication. | ||||||||
4417 | switch(I->getOpcode()) { | ||||||||
4418 | default: | ||||||||
4419 | return false; | ||||||||
4420 | case Instruction::Load: | ||||||||
4421 | case Instruction::Store: { | ||||||||
4422 | if (!Legal->isMaskRequired(I)) | ||||||||
4423 | return false; | ||||||||
4424 | // When we know the load's address is loop invariant and the instruction | ||||||||
4425 | // in the original scalar loop was unconditionally executed then we | ||||||||
4426 | // don't need to mark it as a predicated instruction. Tail folding may | ||||||||
4427 | // introduce additional predication, but we're guaranteed to always have | ||||||||
4428 | // at least one active lane. We call Legal->blockNeedsPredication here | ||||||||
4429 | // because it doesn't query tail-folding. For stores, we need to prove | ||||||||
4430 | // both speculation safety (which follows from the same argument as loads), | ||||||||
4431 | // but also must prove the value being stored is correct. The easiest | ||||||||
4432 | // form of the later is to require that all values stored are the same. | ||||||||
4433 | if (Legal->isUniformMemOp(*I) && | ||||||||
4434 | (isa<LoadInst>(I) || | ||||||||
4435 | (isa<StoreInst>(I) && | ||||||||
4436 | TheLoop->isLoopInvariant(cast<StoreInst>(I)->getValueOperand()))) && | ||||||||
4437 | !Legal->blockNeedsPredication(I->getParent())) | ||||||||
4438 | return false; | ||||||||
4439 | return true; | ||||||||
4440 | } | ||||||||
4441 | case Instruction::UDiv: | ||||||||
4442 | case Instruction::SDiv: | ||||||||
4443 | case Instruction::SRem: | ||||||||
4444 | case Instruction::URem: | ||||||||
4445 | // TODO: We can use the loop-preheader as context point here and get | ||||||||
4446 | // context sensitive reasoning | ||||||||
4447 | return !isSafeToSpeculativelyExecute(I); | ||||||||
4448 | } | ||||||||
4449 | } | ||||||||
4450 | |||||||||
4451 | std::pair<InstructionCost, InstructionCost> | ||||||||
4452 | LoopVectorizationCostModel::getDivRemSpeculationCost(Instruction *I, | ||||||||
4453 | ElementCount VF) const { | ||||||||
4454 | assert(I->getOpcode() == Instruction::UDiv ||(static_cast <bool> (I->getOpcode() == Instruction:: UDiv || I->getOpcode() == Instruction::SDiv || I->getOpcode () == Instruction::SRem || I->getOpcode() == Instruction:: URem) ? void (0) : __assert_fail ("I->getOpcode() == Instruction::UDiv || I->getOpcode() == Instruction::SDiv || I->getOpcode() == Instruction::SRem || I->getOpcode() == Instruction::URem" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4457, __extension__ __PRETTY_FUNCTION__)) | ||||||||
4455 | I->getOpcode() == Instruction::SDiv ||(static_cast <bool> (I->getOpcode() == Instruction:: UDiv || I->getOpcode() == Instruction::SDiv || I->getOpcode () == Instruction::SRem || I->getOpcode() == Instruction:: URem) ? void (0) : __assert_fail ("I->getOpcode() == Instruction::UDiv || I->getOpcode() == Instruction::SDiv || I->getOpcode() == Instruction::SRem || I->getOpcode() == Instruction::URem" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4457, __extension__ __PRETTY_FUNCTION__)) | ||||||||
4456 | I->getOpcode() == Instruction::SRem ||(static_cast <bool> (I->getOpcode() == Instruction:: UDiv || I->getOpcode() == Instruction::SDiv || I->getOpcode () == Instruction::SRem || I->getOpcode() == Instruction:: URem) ? void (0) : __assert_fail ("I->getOpcode() == Instruction::UDiv || I->getOpcode() == Instruction::SDiv || I->getOpcode() == Instruction::SRem || I->getOpcode() == Instruction::URem" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4457, __extension__ __PRETTY_FUNCTION__)) | ||||||||
4457 | I->getOpcode() == Instruction::URem)(static_cast <bool> (I->getOpcode() == Instruction:: UDiv || I->getOpcode() == Instruction::SDiv || I->getOpcode () == Instruction::SRem || I->getOpcode() == Instruction:: URem) ? void (0) : __assert_fail ("I->getOpcode() == Instruction::UDiv || I->getOpcode() == Instruction::SDiv || I->getOpcode() == Instruction::SRem || I->getOpcode() == Instruction::URem" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4457, __extension__ __PRETTY_FUNCTION__)); | ||||||||
4458 | assert(!isSafeToSpeculativelyExecute(I))(static_cast <bool> (!isSafeToSpeculativelyExecute(I)) ? void (0) : __assert_fail ("!isSafeToSpeculativelyExecute(I)" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4458, __extension__ __PRETTY_FUNCTION__)); | ||||||||
4459 | |||||||||
4460 | const TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput; | ||||||||
4461 | |||||||||
4462 | // Scalarization isn't legal for scalable vector types | ||||||||
4463 | InstructionCost ScalarizationCost = InstructionCost::getInvalid(); | ||||||||
4464 | if (!VF.isScalable()) { | ||||||||
4465 | // Get the scalarization cost and scale this amount by the probability of | ||||||||
4466 | // executing the predicated block. If the instruction is not predicated, | ||||||||
4467 | // we fall through to the next case. | ||||||||
4468 | ScalarizationCost = 0; | ||||||||
4469 | |||||||||
4470 | // These instructions have a non-void type, so account for the phi nodes | ||||||||
4471 | // that we will create. This cost is likely to be zero. The phi node | ||||||||
4472 | // cost, if any, should be scaled by the block probability because it | ||||||||
4473 | // models a copy at the end of each predicated block. | ||||||||
4474 | ScalarizationCost += VF.getKnownMinValue() * | ||||||||
4475 | TTI.getCFInstrCost(Instruction::PHI, CostKind); | ||||||||
4476 | |||||||||
4477 | // The cost of the non-predicated instruction. | ||||||||
4478 | ScalarizationCost += VF.getKnownMinValue() * | ||||||||
4479 | TTI.getArithmeticInstrCost(I->getOpcode(), I->getType(), CostKind); | ||||||||
4480 | |||||||||
4481 | // The cost of insertelement and extractelement instructions needed for | ||||||||
4482 | // scalarization. | ||||||||
4483 | ScalarizationCost += getScalarizationOverhead(I, VF, CostKind); | ||||||||
4484 | |||||||||
4485 | // Scale the cost by the probability of executing the predicated blocks. | ||||||||
4486 | // This assumes the predicated block for each vector lane is equally | ||||||||
4487 | // likely. | ||||||||
4488 | ScalarizationCost = ScalarizationCost / getReciprocalPredBlockProb(); | ||||||||
4489 | } | ||||||||
4490 | InstructionCost SafeDivisorCost = 0; | ||||||||
4491 | |||||||||
4492 | auto *VecTy = ToVectorTy(I->getType(), VF); | ||||||||
4493 | |||||||||
4494 | // The cost of the select guard to ensure all lanes are well defined | ||||||||
4495 | // after we speculate above any internal control flow. | ||||||||
4496 | SafeDivisorCost += TTI.getCmpSelInstrCost( | ||||||||
4497 | Instruction::Select, VecTy, | ||||||||
4498 | ToVectorTy(Type::getInt1Ty(I->getContext()), VF), | ||||||||
4499 | CmpInst::BAD_ICMP_PREDICATE, CostKind); | ||||||||
4500 | |||||||||
4501 | // Certain instructions can be cheaper to vectorize if they have a constant | ||||||||
4502 | // second vector operand. One example of this are shifts on x86. | ||||||||
4503 | Value *Op2 = I->getOperand(1); | ||||||||
4504 | auto Op2Info = TTI.getOperandInfo(Op2); | ||||||||
4505 | if (Op2Info.Kind == TargetTransformInfo::OK_AnyValue && Legal->isUniform(Op2)) | ||||||||
4506 | Op2Info.Kind = TargetTransformInfo::OK_UniformValue; | ||||||||
4507 | |||||||||
4508 | SmallVector<const Value *, 4> Operands(I->operand_values()); | ||||||||
4509 | SafeDivisorCost += TTI.getArithmeticInstrCost( | ||||||||
4510 | I->getOpcode(), VecTy, CostKind, | ||||||||
4511 | {TargetTransformInfo::OK_AnyValue, TargetTransformInfo::OP_None}, | ||||||||
4512 | Op2Info, Operands, I); | ||||||||
4513 | return {ScalarizationCost, SafeDivisorCost}; | ||||||||
4514 | } | ||||||||
4515 | |||||||||
4516 | bool LoopVectorizationCostModel::interleavedAccessCanBeWidened( | ||||||||
4517 | Instruction *I, ElementCount VF) { | ||||||||
4518 | 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", 4518, __extension__ __PRETTY_FUNCTION__)); | ||||||||
4519 | 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", 4520, __extension__ __PRETTY_FUNCTION__)) | ||||||||
4520 | "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", 4520, __extension__ __PRETTY_FUNCTION__)); | ||||||||
4521 | auto *Group = getInterleavedAccessGroup(I); | ||||||||
4522 | 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", 4522, __extension__ __PRETTY_FUNCTION__)); | ||||||||
4523 | |||||||||
4524 | // If the instruction's allocated size doesn't equal it's type size, it | ||||||||
4525 | // requires padding and will be scalarized. | ||||||||
4526 | auto &DL = I->getModule()->getDataLayout(); | ||||||||
4527 | auto *ScalarTy = getLoadStoreType(I); | ||||||||
4528 | if (hasIrregularType(ScalarTy, DL)) | ||||||||
4529 | return false; | ||||||||
4530 | |||||||||
4531 | // If the group involves a non-integral pointer, we may not be able to | ||||||||
4532 | // losslessly cast all values to a common type. | ||||||||
4533 | unsigned InterleaveFactor = Group->getFactor(); | ||||||||
4534 | bool ScalarNI = DL.isNonIntegralPointerType(ScalarTy); | ||||||||
4535 | for (unsigned i = 0; i < InterleaveFactor; i++) { | ||||||||
4536 | Instruction *Member = Group->getMember(i); | ||||||||
4537 | if (!Member) | ||||||||
4538 | continue; | ||||||||
4539 | auto *MemberTy = getLoadStoreType(Member); | ||||||||
4540 | bool MemberNI = DL.isNonIntegralPointerType(MemberTy); | ||||||||
4541 | // Don't coerce non-integral pointers to integers or vice versa. | ||||||||
4542 | if (MemberNI != ScalarNI) { | ||||||||
4543 | // TODO: Consider adding special nullptr value case here | ||||||||
4544 | return false; | ||||||||
4545 | } else if (MemberNI && ScalarNI && | ||||||||
4546 | ScalarTy->getPointerAddressSpace() != | ||||||||
4547 | MemberTy->getPointerAddressSpace()) { | ||||||||
4548 | return false; | ||||||||
4549 | } | ||||||||
4550 | } | ||||||||
4551 | |||||||||
4552 | // Check if masking is required. | ||||||||
4553 | // A Group may need masking for one of two reasons: it resides in a block that | ||||||||
4554 | // needs predication, or it was decided to use masking to deal with gaps | ||||||||
4555 | // (either a gap at the end of a load-access that may result in a speculative | ||||||||
4556 | // load, or any gaps in a store-access). | ||||||||
4557 | bool PredicatedAccessRequiresMasking = | ||||||||
4558 | blockNeedsPredicationForAnyReason(I->getParent()) && | ||||||||
4559 | Legal->isMaskRequired(I); | ||||||||
4560 | bool LoadAccessWithGapsRequiresEpilogMasking = | ||||||||
4561 | isa<LoadInst>(I) && Group->requiresScalarEpilogue() && | ||||||||
4562 | !isScalarEpilogueAllowed(); | ||||||||
4563 | bool StoreAccessWithGapsRequiresMasking = | ||||||||
4564 | isa<StoreInst>(I) && (Group->getNumMembers() < Group->getFactor()); | ||||||||
4565 | if (!PredicatedAccessRequiresMasking && | ||||||||
4566 | !LoadAccessWithGapsRequiresEpilogMasking && | ||||||||
4567 | !StoreAccessWithGapsRequiresMasking) | ||||||||
4568 | return true; | ||||||||
4569 | |||||||||
4570 | // If masked interleaving is required, we expect that the user/target had | ||||||||
4571 | // enabled it, because otherwise it either wouldn't have been created or | ||||||||
4572 | // it should have been invalidated by the CostModel. | ||||||||
4573 | 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", 4574, __extension__ __PRETTY_FUNCTION__)) | ||||||||
4574 | "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", 4574, __extension__ __PRETTY_FUNCTION__)); | ||||||||
4575 | |||||||||
4576 | if (Group->isReverse()) | ||||||||
4577 | return false; | ||||||||
4578 | |||||||||
4579 | auto *Ty = getLoadStoreType(I); | ||||||||
4580 | const Align Alignment = getLoadStoreAlignment(I); | ||||||||
4581 | return isa<LoadInst>(I) ? TTI.isLegalMaskedLoad(Ty, Alignment) | ||||||||
4582 | : TTI.isLegalMaskedStore(Ty, Alignment); | ||||||||
4583 | } | ||||||||
4584 | |||||||||
4585 | bool LoopVectorizationCostModel::memoryInstructionCanBeWidened( | ||||||||
4586 | Instruction *I, ElementCount VF) { | ||||||||
4587 | // Get and ensure we have a valid memory instruction. | ||||||||
4588 | 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", 4588, __extension__ __PRETTY_FUNCTION__)); | ||||||||
4589 | |||||||||
4590 | auto *Ptr = getLoadStorePointerOperand(I); | ||||||||
4591 | auto *ScalarTy = getLoadStoreType(I); | ||||||||
4592 | |||||||||
4593 | // In order to be widened, the pointer should be consecutive, first of all. | ||||||||
4594 | if (!Legal->isConsecutivePtr(ScalarTy, Ptr)) | ||||||||
4595 | return false; | ||||||||
4596 | |||||||||
4597 | // If the instruction is a store located in a predicated block, it will be | ||||||||
4598 | // scalarized. | ||||||||
4599 | if (isScalarWithPredication(I, VF)) | ||||||||
4600 | return false; | ||||||||
4601 | |||||||||
4602 | // If the instruction's allocated size doesn't equal it's type size, it | ||||||||
4603 | // requires padding and will be scalarized. | ||||||||
4604 | auto &DL = I->getModule()->getDataLayout(); | ||||||||
4605 | if (hasIrregularType(ScalarTy, DL)) | ||||||||
4606 | return false; | ||||||||
4607 | |||||||||
4608 | return true; | ||||||||
4609 | } | ||||||||
4610 | |||||||||
4611 | void LoopVectorizationCostModel::collectLoopUniforms(ElementCount VF) { | ||||||||
4612 | // We should not collect Uniforms more than once per VF. Right now, | ||||||||
4613 | // this function is called from collectUniformsAndScalars(), which | ||||||||
4614 | // already does this check. Collecting Uniforms for VF=1 does not make any | ||||||||
4615 | // sense. | ||||||||
4616 | |||||||||
4617 | 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", 4618, __extension__ __PRETTY_FUNCTION__)) | ||||||||
4618 | "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", 4618, __extension__ __PRETTY_FUNCTION__)); | ||||||||
4619 | |||||||||
4620 | // Visit the list of Uniforms. If we'll not find any uniform value, we'll | ||||||||
4621 | // not analyze again. Uniforms.count(VF) will return 1. | ||||||||
4622 | Uniforms[VF].clear(); | ||||||||
4623 | |||||||||
4624 | // We now know that the loop is vectorizable! | ||||||||
4625 | // Collect instructions inside the loop that will remain uniform after | ||||||||
4626 | // vectorization. | ||||||||
4627 | |||||||||
4628 | // Global values, params and instructions outside of current loop are out of | ||||||||
4629 | // scope. | ||||||||
4630 | auto isOutOfScope = [&](Value *V) -> bool { | ||||||||
4631 | Instruction *I = dyn_cast<Instruction>(V); | ||||||||
4632 | return (!I || !TheLoop->contains(I)); | ||||||||
4633 | }; | ||||||||
4634 | |||||||||
4635 | // Worklist containing uniform instructions demanding lane 0. | ||||||||
4636 | SetVector<Instruction *> Worklist; | ||||||||
4637 | BasicBlock *Latch = TheLoop->getLoopLatch(); | ||||||||
4638 | |||||||||
4639 | // Add uniform instructions demanding lane 0 to the worklist. Instructions | ||||||||
4640 | // that are scalar with predication must not be considered uniform after | ||||||||
4641 | // vectorization, because that would create an erroneous replicating region | ||||||||
4642 | // where only a single instance out of VF should be formed. | ||||||||
4643 | // TODO: optimize such seldom cases if found important, see PR40816. | ||||||||
4644 | auto addToWorklistIfAllowed = [&](Instruction *I) -> void { | ||||||||
4645 | if (isOutOfScope(I)) { | ||||||||
4646 | 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) | ||||||||
4647 | << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found not uniform due to scope: " << *I << "\n"; } } while (false); | ||||||||
4648 | return; | ||||||||
4649 | } | ||||||||
4650 | if (isScalarWithPredication(I, VF)) { | ||||||||
4651 | 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) | ||||||||
4652 | << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found not uniform being ScalarWithPredication: " << *I << "\n"; } } while (false); | ||||||||
4653 | return; | ||||||||
4654 | } | ||||||||
4655 | 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); | ||||||||
4656 | Worklist.insert(I); | ||||||||
4657 | }; | ||||||||
4658 | |||||||||
4659 | // Start with the conditional branch. If the branch condition is an | ||||||||
4660 | // instruction contained in the loop that is only used by the branch, it is | ||||||||
4661 | // uniform. | ||||||||
4662 | auto *Cmp = dyn_cast<Instruction>(Latch->getTerminator()->getOperand(0)); | ||||||||
4663 | if (Cmp && TheLoop->contains(Cmp) && Cmp->hasOneUse()) | ||||||||
4664 | addToWorklistIfAllowed(Cmp); | ||||||||
4665 | |||||||||
4666 | // Return true if all lanes perform the same memory operation, and we can | ||||||||
4667 | // thus chose to execute only one. | ||||||||
4668 | auto isUniformMemOpUse = [&](Instruction *I) { | ||||||||
4669 | if (!Legal->isUniformMemOp(*I)) | ||||||||
4670 | return false; | ||||||||
4671 | if (isa<LoadInst>(I)) | ||||||||
4672 | // Loading the same address always produces the same result - at least | ||||||||
4673 | // assuming aliasing and ordering which have already been checked. | ||||||||
4674 | return true; | ||||||||
4675 | // Storing the same value on every iteration. | ||||||||
4676 | return TheLoop->isLoopInvariant(cast<StoreInst>(I)->getValueOperand()); | ||||||||
4677 | }; | ||||||||
4678 | |||||||||
4679 | auto isUniformDecision = [&](Instruction *I, ElementCount VF) { | ||||||||
4680 | InstWidening WideningDecision = getWideningDecision(I, VF); | ||||||||
4681 | 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", 4682, __extension__ __PRETTY_FUNCTION__)) | ||||||||
4682 | "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", 4682, __extension__ __PRETTY_FUNCTION__)); | ||||||||
4683 | |||||||||
4684 | if (isUniformMemOpUse(I)) | ||||||||
4685 | return true; | ||||||||
4686 | |||||||||
4687 | return (WideningDecision == CM_Widen || | ||||||||
4688 | WideningDecision == CM_Widen_Reverse || | ||||||||
4689 | WideningDecision == CM_Interleave); | ||||||||
4690 | }; | ||||||||
4691 | |||||||||
4692 | |||||||||
4693 | // Returns true if Ptr is the pointer operand of a memory access instruction | ||||||||
4694 | // I, and I is known to not require scalarization. | ||||||||
4695 | auto isVectorizedMemAccessUse = [&](Instruction *I, Value *Ptr) -> bool { | ||||||||
4696 | return getLoadStorePointerOperand(I) == Ptr && isUniformDecision(I, VF); | ||||||||
4697 | }; | ||||||||
4698 | |||||||||
4699 | // Holds a list of values which are known to have at least one uniform use. | ||||||||
4700 | // Note that there may be other uses which aren't uniform. A "uniform use" | ||||||||
4701 | // here is something which only demands lane 0 of the unrolled iterations; | ||||||||
4702 | // it does not imply that all lanes produce the same value (e.g. this is not | ||||||||
4703 | // the usual meaning of uniform) | ||||||||
4704 | SetVector<Value *> HasUniformUse; | ||||||||
4705 | |||||||||
4706 | // Scan the loop for instructions which are either a) known to have only | ||||||||
4707 | // lane 0 demanded or b) are uses which demand only lane 0 of their operand. | ||||||||
4708 | for (auto *BB : TheLoop->blocks()) | ||||||||
4709 | for (auto &I : *BB) { | ||||||||
4710 | if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(&I)) { | ||||||||
4711 | switch (II->getIntrinsicID()) { | ||||||||
4712 | case Intrinsic::sideeffect: | ||||||||
4713 | case Intrinsic::experimental_noalias_scope_decl: | ||||||||
4714 | case Intrinsic::assume: | ||||||||
4715 | case Intrinsic::lifetime_start: | ||||||||
4716 | case Intrinsic::lifetime_end: | ||||||||
4717 | if (TheLoop->hasLoopInvariantOperands(&I)) | ||||||||
4718 | addToWorklistIfAllowed(&I); | ||||||||
4719 | break; | ||||||||
4720 | default: | ||||||||
4721 | break; | ||||||||
4722 | } | ||||||||
4723 | } | ||||||||
4724 | |||||||||
4725 | // ExtractValue instructions must be uniform, because the operands are | ||||||||
4726 | // known to be loop-invariant. | ||||||||
4727 | if (auto *EVI = dyn_cast<ExtractValueInst>(&I)) { | ||||||||
4728 | 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", 4729, __extension__ __PRETTY_FUNCTION__)) | ||||||||
4729 | "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", 4729, __extension__ __PRETTY_FUNCTION__)); | ||||||||
4730 | addToWorklistIfAllowed(EVI); | ||||||||
4731 | continue; | ||||||||
4732 | } | ||||||||
4733 | |||||||||
4734 | // If there's no pointer operand, there's nothing to do. | ||||||||
4735 | auto *Ptr = getLoadStorePointerOperand(&I); | ||||||||
4736 | if (!Ptr) | ||||||||
4737 | continue; | ||||||||
4738 | |||||||||
4739 | if (isUniformMemOpUse(&I)) | ||||||||
4740 | addToWorklistIfAllowed(&I); | ||||||||
4741 | |||||||||
4742 | if (isUniformDecision(&I, VF)) { | ||||||||
4743 | 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", 4743, __extension__ __PRETTY_FUNCTION__)); | ||||||||
4744 | HasUniformUse.insert(Ptr); | ||||||||
4745 | } | ||||||||
4746 | } | ||||||||
4747 | |||||||||
4748 | // Add to the worklist any operands which have *only* uniform (e.g. lane 0 | ||||||||
4749 | // demanding) users. Since loops are assumed to be in LCSSA form, this | ||||||||
4750 | // disallows uses outside the loop as well. | ||||||||
4751 | for (auto *V : HasUniformUse) { | ||||||||
4752 | if (isOutOfScope(V)) | ||||||||
4753 | continue; | ||||||||
4754 | auto *I = cast<Instruction>(V); | ||||||||
4755 | auto UsersAreMemAccesses = | ||||||||
4756 | llvm::all_of(I->users(), [&](User *U) -> bool { | ||||||||
4757 | return isVectorizedMemAccessUse(cast<Instruction>(U), V); | ||||||||
4758 | }); | ||||||||
4759 | if (UsersAreMemAccesses) | ||||||||
4760 | addToWorklistIfAllowed(I); | ||||||||
4761 | } | ||||||||
4762 | |||||||||
4763 | // Expand Worklist in topological order: whenever a new instruction | ||||||||
4764 | // is added , its users should be already inside Worklist. It ensures | ||||||||
4765 | // a uniform instruction will only be used by uniform instructions. | ||||||||
4766 | unsigned idx = 0; | ||||||||
4767 | while (idx != Worklist.size()) { | ||||||||
4768 | Instruction *I = Worklist[idx++]; | ||||||||
4769 | |||||||||
4770 | for (auto *OV : I->operand_values()) { | ||||||||
4771 | // isOutOfScope operands cannot be uniform instructions. | ||||||||
4772 | if (isOutOfScope(OV)) | ||||||||
4773 | continue; | ||||||||
4774 | // First order recurrence Phi's should typically be considered | ||||||||
4775 | // non-uniform. | ||||||||
4776 | auto *OP = dyn_cast<PHINode>(OV); | ||||||||
4777 | if (OP && Legal->isFixedOrderRecurrence(OP)) | ||||||||
4778 | continue; | ||||||||
4779 | // If all the users of the operand are uniform, then add the | ||||||||
4780 | // operand into the uniform worklist. | ||||||||
4781 | auto *OI = cast<Instruction>(OV); | ||||||||
4782 | if (llvm::all_of(OI->users(), [&](User *U) -> bool { | ||||||||
4783 | auto *J = cast<Instruction>(U); | ||||||||
4784 | return Worklist.count(J) || isVectorizedMemAccessUse(J, OI); | ||||||||
4785 | })) | ||||||||
4786 | addToWorklistIfAllowed(OI); | ||||||||
4787 | } | ||||||||
4788 | } | ||||||||
4789 | |||||||||
4790 | // For an instruction to be added into Worklist above, all its users inside | ||||||||
4791 | // the loop should also be in Worklist. However, this condition cannot be | ||||||||
4792 | // true for phi nodes that form a cyclic dependence. We must process phi | ||||||||
4793 | // nodes separately. An induction variable will remain uniform if all users | ||||||||
4794 | // of the induction variable and induction variable update remain uniform. | ||||||||
4795 | // The code below handles both pointer and non-pointer induction variables. | ||||||||
4796 | for (const auto &Induction : Legal->getInductionVars()) { | ||||||||
4797 | auto *Ind = Induction.first; | ||||||||
4798 | auto *IndUpdate = cast<Instruction>(Ind->getIncomingValueForBlock(Latch)); | ||||||||
4799 | |||||||||
4800 | // Determine if all users of the induction variable are uniform after | ||||||||
4801 | // vectorization. | ||||||||
4802 | auto UniformInd = llvm::all_of(Ind->users(), [&](User *U) -> bool { | ||||||||
4803 | auto *I = cast<Instruction>(U); | ||||||||
4804 | return I == IndUpdate || !TheLoop->contains(I) || Worklist.count(I) || | ||||||||
4805 | isVectorizedMemAccessUse(I, Ind); | ||||||||
4806 | }); | ||||||||
4807 | if (!UniformInd) | ||||||||
4808 | continue; | ||||||||
4809 | |||||||||
4810 | // Determine if all users of the induction variable update instruction are | ||||||||
4811 | // uniform after vectorization. | ||||||||
4812 | auto UniformIndUpdate = | ||||||||
4813 | llvm::all_of(IndUpdate->users(), [&](User *U) -> bool { | ||||||||
4814 | auto *I = cast<Instruction>(U); | ||||||||
4815 | return I == Ind || !TheLoop->contains(I) || Worklist.count(I) || | ||||||||
4816 | isVectorizedMemAccessUse(I, IndUpdate); | ||||||||
4817 | }); | ||||||||
4818 | if (!UniformIndUpdate) | ||||||||
4819 | continue; | ||||||||
4820 | |||||||||
4821 | // The induction variable and its update instruction will remain uniform. | ||||||||
4822 | addToWorklistIfAllowed(Ind); | ||||||||
4823 | addToWorklistIfAllowed(IndUpdate); | ||||||||
4824 | } | ||||||||
4825 | |||||||||
4826 | Uniforms[VF].insert(Worklist.begin(), Worklist.end()); | ||||||||
4827 | } | ||||||||
4828 | |||||||||
4829 | bool LoopVectorizationCostModel::runtimeChecksRequired() { | ||||||||
4830 | 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); | ||||||||
4831 | |||||||||
4832 | if (Legal->getRuntimePointerChecking()->Need) { | ||||||||
4833 | reportVectorizationFailure("Runtime ptr check is required with -Os/-Oz", | ||||||||
4834 | "runtime pointer checks needed. Enable vectorization of this " | ||||||||
4835 | "loop with '#pragma clang loop vectorize(enable)' when " | ||||||||
4836 | "compiling with -Os/-Oz", | ||||||||
4837 | "CantVersionLoopWithOptForSize", ORE, TheLoop); | ||||||||
4838 | return true; | ||||||||
4839 | } | ||||||||
4840 | |||||||||
4841 | if (!PSE.getPredicate().isAlwaysTrue()) { | ||||||||
4842 | reportVectorizationFailure("Runtime SCEV check is required with -Os/-Oz", | ||||||||
4843 | "runtime SCEV checks needed. Enable vectorization of this " | ||||||||
4844 | "loop with '#pragma clang loop vectorize(enable)' when " | ||||||||
4845 | "compiling with -Os/-Oz", | ||||||||
4846 | "CantVersionLoopWithOptForSize", ORE, TheLoop); | ||||||||
4847 | return true; | ||||||||
4848 | } | ||||||||
4849 | |||||||||
4850 | // FIXME: Avoid specializing for stride==1 instead of bailing out. | ||||||||
4851 | if (!Legal->getLAI()->getSymbolicStrides().empty()) { | ||||||||
4852 | reportVectorizationFailure("Runtime stride check for small trip count", | ||||||||
4853 | "runtime stride == 1 checks needed. Enable vectorization of " | ||||||||
4854 | "this loop without such check by compiling with -Os/-Oz", | ||||||||
4855 | "CantVersionLoopWithOptForSize", ORE, TheLoop); | ||||||||
4856 | return true; | ||||||||
4857 | } | ||||||||
4858 | |||||||||
4859 | return false; | ||||||||
4860 | } | ||||||||
4861 | |||||||||
4862 | ElementCount | ||||||||
4863 | LoopVectorizationCostModel::getMaxLegalScalableVF(unsigned MaxSafeElements) { | ||||||||
4864 | if (!TTI.supportsScalableVectors() && !ForceTargetSupportsScalableVectors) | ||||||||
4865 | return ElementCount::getScalable(0); | ||||||||
4866 | |||||||||
4867 | if (Hints->isScalableVectorizationDisabled()) { | ||||||||
4868 | reportVectorizationInfo("Scalable vectorization is explicitly disabled", | ||||||||
4869 | "ScalableVectorizationDisabled", ORE, TheLoop); | ||||||||
4870 | return ElementCount::getScalable(0); | ||||||||
4871 | } | ||||||||
4872 | |||||||||
4873 | 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); | ||||||||
4874 | |||||||||
4875 | auto MaxScalableVF = ElementCount::getScalable( | ||||||||
4876 | std::numeric_limits<ElementCount::ScalarTy>::max()); | ||||||||
4877 | |||||||||
4878 | // Test that the loop-vectorizer can legalize all operations for this MaxVF. | ||||||||
4879 | // FIXME: While for scalable vectors this is currently sufficient, this should | ||||||||
4880 | // be replaced by a more detailed mechanism that filters out specific VFs, | ||||||||
4881 | // instead of invalidating vectorization for a whole set of VFs based on the | ||||||||
4882 | // MaxVF. | ||||||||
4883 | |||||||||
4884 | // Disable scalable vectorization if the loop contains unsupported reductions. | ||||||||
4885 | if (!canVectorizeReductions(MaxScalableVF)) { | ||||||||
4886 | reportVectorizationInfo( | ||||||||
4887 | "Scalable vectorization not supported for the reduction " | ||||||||
4888 | "operations found in this loop.", | ||||||||
4889 | "ScalableVFUnfeasible", ORE, TheLoop); | ||||||||
4890 | return ElementCount::getScalable(0); | ||||||||
4891 | } | ||||||||
4892 | |||||||||
4893 | // Disable scalable vectorization if the loop contains any instructions | ||||||||
4894 | // with element types not supported for scalable vectors. | ||||||||
4895 | if (any_of(ElementTypesInLoop, [&](Type *Ty) { | ||||||||
4896 | return !Ty->isVoidTy() && | ||||||||
4897 | !this->TTI.isElementTypeLegalForScalableVector(Ty); | ||||||||
4898 | })) { | ||||||||
4899 | reportVectorizationInfo("Scalable vectorization is not supported " | ||||||||
4900 | "for all element types found in this loop.", | ||||||||
4901 | "ScalableVFUnfeasible", ORE, TheLoop); | ||||||||
4902 | return ElementCount::getScalable(0); | ||||||||
4903 | } | ||||||||
4904 | |||||||||
4905 | if (Legal->isSafeForAnyVectorWidth()) | ||||||||
4906 | return MaxScalableVF; | ||||||||
4907 | |||||||||
4908 | // Limit MaxScalableVF by the maximum safe dependence distance. | ||||||||
4909 | std::optional<unsigned> MaxVScale = TTI.getMaxVScale(); | ||||||||
4910 | if (!MaxVScale && TheFunction->hasFnAttribute(Attribute::VScaleRange)) | ||||||||
4911 | MaxVScale = | ||||||||
4912 | TheFunction->getFnAttribute(Attribute::VScaleRange).getVScaleRangeMax(); | ||||||||
4913 | MaxScalableVF = | ||||||||
4914 | ElementCount::getScalable(MaxVScale ? (MaxSafeElements / *MaxVScale) : 0); | ||||||||
4915 | if (!MaxScalableVF) | ||||||||
4916 | reportVectorizationInfo( | ||||||||
4917 | "Max legal vector width too small, scalable vectorization " | ||||||||
4918 | "unfeasible.", | ||||||||
4919 | "ScalableVFUnfeasible", ORE, TheLoop); | ||||||||
4920 | |||||||||
4921 | return MaxScalableVF; | ||||||||
4922 | } | ||||||||
4923 | |||||||||
4924 | FixedScalableVFPair LoopVectorizationCostModel::computeFeasibleMaxVF( | ||||||||
4925 | unsigned ConstTripCount, ElementCount UserVF, bool FoldTailByMasking) { | ||||||||
4926 | MinBWs = computeMinimumValueSizes(TheLoop->getBlocks(), *DB, &TTI); | ||||||||
4927 | unsigned SmallestType, WidestType; | ||||||||
4928 | std::tie(SmallestType, WidestType) = getSmallestAndWidestTypes(); | ||||||||
4929 | |||||||||
4930 | // Get the maximum safe dependence distance in bits computed by LAA. | ||||||||
4931 | // It is computed by MaxVF * sizeOf(type) * 8, where type is taken from | ||||||||
4932 | // the memory accesses that is most restrictive (involved in the smallest | ||||||||
4933 | // dependence distance). | ||||||||
4934 | unsigned MaxSafeElements = | ||||||||
4935 | PowerOf2Floor(Legal->getMaxSafeVectorWidthInBits() / WidestType); | ||||||||
4936 | |||||||||
4937 | auto MaxSafeFixedVF = ElementCount::getFixed(MaxSafeElements); | ||||||||
4938 | auto MaxSafeScalableVF = getMaxLegalScalableVF(MaxSafeElements); | ||||||||
4939 | |||||||||
4940 | 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) | ||||||||
4941 | << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: The max safe fixed VF is: " << MaxSafeFixedVF << ".\n"; } } while (false); | ||||||||
4942 | 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) | ||||||||
4943 | << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: The max safe scalable VF is: " << MaxSafeScalableVF << ".\n"; } } while (false); | ||||||||
4944 | |||||||||
4945 | // First analyze the UserVF, fall back if the UserVF should be ignored. | ||||||||
4946 | if (UserVF) { | ||||||||
4947 | auto MaxSafeUserVF = | ||||||||
4948 | UserVF.isScalable() ? MaxSafeScalableVF : MaxSafeFixedVF; | ||||||||
4949 | |||||||||
4950 | if (ElementCount::isKnownLE(UserVF, MaxSafeUserVF)) { | ||||||||
4951 | // If `VF=vscale x N` is safe, then so is `VF=N` | ||||||||
4952 | if (UserVF.isScalable()) | ||||||||
4953 | return FixedScalableVFPair( | ||||||||
4954 | ElementCount::getFixed(UserVF.getKnownMinValue()), UserVF); | ||||||||
4955 | else | ||||||||
4956 | return UserVF; | ||||||||
4957 | } | ||||||||
4958 | |||||||||
4959 | 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", 4959, __extension__ __PRETTY_FUNCTION__)); | ||||||||
4960 | |||||||||
4961 | // Only clamp if the UserVF is not scalable. If the UserVF is scalable, it | ||||||||
4962 | // is better to ignore the hint and let the compiler choose a suitable VF. | ||||||||
4963 | if (!UserVF.isScalable()) { | ||||||||
4964 | 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) | ||||||||
4965 | << " 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) | ||||||||
4966 | << 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); | ||||||||
4967 | ORE->emit([&]() { | ||||||||
4968 | return OptimizationRemarkAnalysis(DEBUG_TYPE"loop-vectorize", "VectorizationFactor", | ||||||||
4969 | TheLoop->getStartLoc(), | ||||||||
4970 | TheLoop->getHeader()) | ||||||||
4971 | << "User-specified vectorization factor " | ||||||||
4972 | << ore::NV("UserVectorizationFactor", UserVF) | ||||||||
4973 | << " is unsafe, clamping to maximum safe vectorization factor " | ||||||||
4974 | << ore::NV("VectorizationFactor", MaxSafeFixedVF); | ||||||||
4975 | }); | ||||||||
4976 | return MaxSafeFixedVF; | ||||||||
4977 | } | ||||||||
4978 | |||||||||
4979 | if (!TTI.supportsScalableVectors() && !ForceTargetSupportsScalableVectors) { | ||||||||
4980 | 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) | ||||||||
4981 | << " 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) | ||||||||
4982 | "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); | ||||||||
4983 | ORE->emit([&]() { | ||||||||
4984 | return OptimizationRemarkAnalysis(DEBUG_TYPE"loop-vectorize", "VectorizationFactor", | ||||||||
4985 | TheLoop->getStartLoc(), | ||||||||
4986 | TheLoop->getHeader()) | ||||||||
4987 | << "User-specified vectorization factor " | ||||||||
4988 | << ore::NV("UserVectorizationFactor", UserVF) | ||||||||
4989 | << " is ignored because the target does not support scalable " | ||||||||
4990 | "vectors. The compiler will pick a more suitable value."; | ||||||||
4991 | }); | ||||||||
4992 | } else { | ||||||||
4993 | 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) | ||||||||
4994 | << " 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); | ||||||||
4995 | ORE->emit([&]() { | ||||||||
4996 | return OptimizationRemarkAnalysis(DEBUG_TYPE"loop-vectorize", "VectorizationFactor", | ||||||||
4997 | TheLoop->getStartLoc(), | ||||||||
4998 | TheLoop->getHeader()) | ||||||||
4999 | << "User-specified vectorization factor " | ||||||||
5000 | << ore::NV("UserVectorizationFactor", UserVF) | ||||||||
5001 | << " is unsafe. Ignoring the hint to let the compiler pick a " | ||||||||
5002 | "more suitable value."; | ||||||||
5003 | }); | ||||||||
5004 | } | ||||||||
5005 | } | ||||||||
5006 | |||||||||
5007 | 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) | ||||||||
5008 | << " / " << WidestType << " bits.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: The Smallest and Widest types: " << SmallestType << " / " << WidestType << " bits.\n"; } } while (false); | ||||||||
5009 | |||||||||
5010 | FixedScalableVFPair Result(ElementCount::getFixed(1), | ||||||||
5011 | ElementCount::getScalable(0)); | ||||||||
5012 | if (auto MaxVF = | ||||||||
5013 | getMaximizedVFForTarget(ConstTripCount, SmallestType, WidestType, | ||||||||
5014 | MaxSafeFixedVF, FoldTailByMasking)) | ||||||||
5015 | Result.FixedVF = MaxVF; | ||||||||
5016 | |||||||||
5017 | if (auto MaxVF = | ||||||||
5018 | getMaximizedVFForTarget(ConstTripCount, SmallestType, WidestType, | ||||||||
5019 | MaxSafeScalableVF, FoldTailByMasking)) | ||||||||
5020 | if (MaxVF.isScalable()) { | ||||||||
5021 | Result.ScalableVF = MaxVF; | ||||||||
5022 | 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) | ||||||||
5023 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found feasible scalable VF = " << MaxVF << "\n"; } } while (false); | ||||||||
5024 | } | ||||||||
5025 | |||||||||
5026 | return Result; | ||||||||
5027 | } | ||||||||
5028 | |||||||||
5029 | FixedScalableVFPair | ||||||||
5030 | LoopVectorizationCostModel::computeMaxVF(ElementCount UserVF, unsigned UserIC) { | ||||||||
5031 | if (Legal->getRuntimePointerChecking()->Need && TTI.hasBranchDivergence()) { | ||||||||
5032 | // TODO: It may by useful to do since it's still likely to be dynamically | ||||||||
5033 | // uniform if the target can skip. | ||||||||
5034 | reportVectorizationFailure( | ||||||||
5035 | "Not inserting runtime ptr check for divergent target", | ||||||||
5036 | "runtime pointer checks needed. Not enabled for divergent target", | ||||||||
5037 | "CantVersionLoopWithDivergentTarget", ORE, TheLoop); | ||||||||
5038 | return FixedScalableVFPair::getNone(); | ||||||||
5039 | } | ||||||||
5040 | |||||||||
5041 | unsigned TC = PSE.getSE()->getSmallConstantTripCount(TheLoop); | ||||||||
5042 | 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); | ||||||||
5043 | if (TC == 1) { | ||||||||
5044 | reportVectorizationFailure("Single iteration (non) loop", | ||||||||
5045 | "loop trip count is one, irrelevant for vectorization", | ||||||||
5046 | "SingleIterationLoop", ORE, TheLoop); | ||||||||
5047 | return FixedScalableVFPair::getNone(); | ||||||||
5048 | } | ||||||||
5049 | |||||||||
5050 | switch (ScalarEpilogueStatus) { | ||||||||
5051 | case CM_ScalarEpilogueAllowed: | ||||||||
5052 | return computeFeasibleMaxVF(TC, UserVF, false); | ||||||||
5053 | case CM_ScalarEpilogueNotAllowedUsePredicate: | ||||||||
5054 | [[fallthrough]]; | ||||||||
5055 | case CM_ScalarEpilogueNotNeededUsePredicate: | ||||||||
5056 | 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) | ||||||||
5057 | 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) | ||||||||
5058 | << "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) | ||||||||
5059 | << "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); | ||||||||
5060 | break; | ||||||||
5061 | case CM_ScalarEpilogueNotAllowedLowTripLoop: | ||||||||
5062 | // fallthrough as a special case of OptForSize | ||||||||
5063 | case CM_ScalarEpilogueNotAllowedOptSize: | ||||||||
5064 | if (ScalarEpilogueStatus == CM_ScalarEpilogueNotAllowedOptSize) | ||||||||
5065 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Not allowing scalar epilogue due to -Os/-Oz.\n" ; } } while (false) | ||||||||
5066 | 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); | ||||||||
5067 | else | ||||||||
5068 | 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) | ||||||||
5069 | << "count.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Not allowing scalar epilogue due to low trip " << "count.\n"; } } while (false); | ||||||||
5070 | |||||||||
5071 | // Bail if runtime checks are required, which are not good when optimising | ||||||||
5072 | // for size. | ||||||||
5073 | if (runtimeChecksRequired()) | ||||||||
5074 | return FixedScalableVFPair::getNone(); | ||||||||
5075 | |||||||||
5076 | break; | ||||||||
5077 | } | ||||||||
5078 | |||||||||
5079 | // The only loops we can vectorize without a scalar epilogue, are loops with | ||||||||
5080 | // a bottom-test and a single exiting block. We'd have to handle the fact | ||||||||
5081 | // that not every instruction executes on the last iteration. This will | ||||||||
5082 | // require a lane mask which varies through the vector loop body. (TODO) | ||||||||
5083 | if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) { | ||||||||
5084 | // If there was a tail-folding hint/switch, but we can't fold the tail by | ||||||||
5085 | // masking, fallback to a vectorization with a scalar epilogue. | ||||||||
5086 | if (ScalarEpilogueStatus == CM_ScalarEpilogueNotNeededUsePredicate) { | ||||||||
5087 | 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) | ||||||||
5088 | "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); | ||||||||
5089 | ScalarEpilogueStatus = CM_ScalarEpilogueAllowed; | ||||||||
5090 | return computeFeasibleMaxVF(TC, UserVF, false); | ||||||||
5091 | } | ||||||||
5092 | return FixedScalableVFPair::getNone(); | ||||||||
5093 | } | ||||||||
5094 | |||||||||
5095 | // Now try the tail folding | ||||||||
5096 | |||||||||
5097 | // Invalidate interleave groups that require an epilogue if we can't mask | ||||||||
5098 | // the interleave-group. | ||||||||
5099 | if (!useMaskedInterleavedAccesses(TTI)) { | ||||||||
5100 | 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", 5101, __extension__ __PRETTY_FUNCTION__)) | ||||||||
5101 | "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", 5101, __extension__ __PRETTY_FUNCTION__)); | ||||||||
5102 | // Note: There is no need to invalidate any cost modeling decisions here, as | ||||||||
5103 | // non where taken so far. | ||||||||
5104 | InterleaveInfo.invalidateGroupsRequiringScalarEpilogue(); | ||||||||
5105 | } | ||||||||
5106 | |||||||||
5107 | FixedScalableVFPair MaxFactors = computeFeasibleMaxVF(TC, UserVF, true); | ||||||||
5108 | // Avoid tail folding if the trip count is known to be a multiple of any VF | ||||||||
5109 | // we chose. | ||||||||
5110 | // FIXME: The condition below pessimises the case for fixed-width vectors, | ||||||||
5111 | // when scalable VFs are also candidates for vectorization. | ||||||||
5112 | if (MaxFactors.FixedVF.isVector() && !MaxFactors.ScalableVF) { | ||||||||
5113 | ElementCount MaxFixedVF = MaxFactors.FixedVF; | ||||||||
5114 | 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", 5115, __extension__ __PRETTY_FUNCTION__)) | ||||||||
5115 | "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", 5115, __extension__ __PRETTY_FUNCTION__)); | ||||||||
5116 | unsigned MaxVFtimesIC = UserIC ? MaxFixedVF.getFixedValue() * UserIC | ||||||||
5117 | : MaxFixedVF.getFixedValue(); | ||||||||
5118 | ScalarEvolution *SE = PSE.getSE(); | ||||||||
5119 | const SCEV *BackedgeTakenCount = PSE.getBackedgeTakenCount(); | ||||||||
5120 | const SCEV *ExitCount = SE->getAddExpr( | ||||||||
5121 | BackedgeTakenCount, SE->getOne(BackedgeTakenCount->getType())); | ||||||||
5122 | const SCEV *Rem = SE->getURemExpr( | ||||||||
5123 | SE->applyLoopGuards(ExitCount, TheLoop), | ||||||||
5124 | SE->getConstant(BackedgeTakenCount->getType(), MaxVFtimesIC)); | ||||||||
5125 | if (Rem->isZero()) { | ||||||||
5126 | // Accept MaxFixedVF if we do not have a tail. | ||||||||
5127 | 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); | ||||||||
5128 | return MaxFactors; | ||||||||
5129 | } | ||||||||
5130 | } | ||||||||
5131 | |||||||||
5132 | // If we don't know the precise trip count, or if the trip count that we | ||||||||
5133 | // found modulo the vectorization factor is not zero, try to fold the tail | ||||||||
5134 | // by masking. | ||||||||
5135 | // FIXME: look for a smaller MaxVF that does divide TC rather than masking. | ||||||||
5136 | if (Legal->prepareToFoldTailByMasking()) { | ||||||||
5137 | FoldTailByMasking = true; | ||||||||
5138 | return MaxFactors; | ||||||||
5139 | } | ||||||||
5140 | |||||||||
5141 | // If there was a tail-folding hint/switch, but we can't fold the tail by | ||||||||
5142 | // masking, fallback to a vectorization with a scalar epilogue. | ||||||||
5143 | if (ScalarEpilogueStatus == CM_ScalarEpilogueNotNeededUsePredicate) { | ||||||||
5144 | 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) | ||||||||
5145 | "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); | ||||||||
5146 | ScalarEpilogueStatus = CM_ScalarEpilogueAllowed; | ||||||||
5147 | return MaxFactors; | ||||||||
5148 | } | ||||||||
5149 | |||||||||
5150 | if (ScalarEpilogueStatus == CM_ScalarEpilogueNotAllowedUsePredicate) { | ||||||||
5151 | 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); | ||||||||
5152 | return FixedScalableVFPair::getNone(); | ||||||||
5153 | } | ||||||||
5154 | |||||||||
5155 | if (TC == 0) { | ||||||||
5156 | reportVectorizationFailure( | ||||||||
5157 | "Unable to calculate the loop count due to complex control flow", | ||||||||
5158 | "unable to calculate the loop count due to complex control flow", | ||||||||
5159 | "UnknownLoopCountComplexCFG", ORE, TheLoop); | ||||||||
5160 | return FixedScalableVFPair::getNone(); | ||||||||
5161 | } | ||||||||
5162 | |||||||||
5163 | reportVectorizationFailure( | ||||||||
5164 | "Cannot optimize for size and vectorize at the same time.", | ||||||||
5165 | "cannot optimize for size and vectorize at the same time. " | ||||||||
5166 | "Enable vectorization of this loop with '#pragma clang loop " | ||||||||
5167 | "vectorize(enable)' when compiling with -Os/-Oz", | ||||||||
5168 | "NoTailLoopWithOptForSize", ORE, TheLoop); | ||||||||
5169 | return FixedScalableVFPair::getNone(); | ||||||||
5170 | } | ||||||||
5171 | |||||||||
5172 | ElementCount LoopVectorizationCostModel::getMaximizedVFForTarget( | ||||||||
5173 | unsigned ConstTripCount, unsigned SmallestType, unsigned WidestType, | ||||||||
5174 | ElementCount MaxSafeVF, bool FoldTailByMasking) { | ||||||||
5175 | bool ComputeScalableMaxVF = MaxSafeVF.isScalable(); | ||||||||
5176 | const TypeSize WidestRegister = TTI.getRegisterBitWidth( | ||||||||
5177 | ComputeScalableMaxVF ? TargetTransformInfo::RGK_ScalableVector | ||||||||
5178 | : TargetTransformInfo::RGK_FixedWidthVector); | ||||||||
5179 | |||||||||
5180 | // Convenience function to return the minimum of two ElementCounts. | ||||||||
5181 | auto MinVF = [](const ElementCount &LHS, const ElementCount &RHS) { | ||||||||
5182 | 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", 5183, __extension__ __PRETTY_FUNCTION__)) | ||||||||
5183 | "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", 5183, __extension__ __PRETTY_FUNCTION__)); | ||||||||
5184 | return ElementCount::isKnownLT(LHS, RHS) ? LHS : RHS; | ||||||||
5185 | }; | ||||||||
5186 | |||||||||
5187 | // Ensure MaxVF is a power of 2; the dependence distance bound may not be. | ||||||||
5188 | // Note that both WidestRegister and WidestType may not be a powers of 2. | ||||||||
5189 | auto MaxVectorElementCount = ElementCount::get( | ||||||||
5190 | PowerOf2Floor(WidestRegister.getKnownMinValue() / WidestType), | ||||||||
5191 | ComputeScalableMaxVF); | ||||||||
5192 | MaxVectorElementCount = MinVF(MaxVectorElementCount, MaxSafeVF); | ||||||||
5193 | 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) | ||||||||
5194 | << (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); | ||||||||
5195 | |||||||||
5196 | if (!MaxVectorElementCount) { | ||||||||
5197 | 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) | ||||||||
5198 | << (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) | ||||||||
5199 | << " 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); | ||||||||
5200 | return ElementCount::getFixed(1); | ||||||||
5201 | } | ||||||||
5202 | |||||||||
5203 | unsigned WidestRegisterMinEC = MaxVectorElementCount.getKnownMinValue(); | ||||||||
5204 | if (MaxVectorElementCount.isScalable() && | ||||||||
5205 | TheFunction->hasFnAttribute(Attribute::VScaleRange)) { | ||||||||
5206 | auto Attr = TheFunction->getFnAttribute(Attribute::VScaleRange); | ||||||||
5207 | auto Min = Attr.getVScaleRangeMin(); | ||||||||
5208 | WidestRegisterMinEC *= Min; | ||||||||
5209 | } | ||||||||
5210 | if (ConstTripCount && ConstTripCount <= WidestRegisterMinEC && | ||||||||
5211 | (!FoldTailByMasking || isPowerOf2_32(ConstTripCount))) { | ||||||||
5212 | // If loop trip count (TC) is known at compile time there is no point in | ||||||||
5213 | // choosing VF greater than TC (as done in the loop below). Select maximum | ||||||||
5214 | // power of two which doesn't exceed TC. | ||||||||
5215 | // If MaxVectorElementCount is scalable, we only fall back on a fixed VF | ||||||||
5216 | // when the TC is less than or equal to the known number of lanes. | ||||||||
5217 | auto ClampedConstTripCount = PowerOf2Floor(ConstTripCount); | ||||||||
5218 | 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) | ||||||||
5219 | "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) | ||||||||
5220 | << 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); | ||||||||
5221 | return ElementCount::getFixed(ClampedConstTripCount); | ||||||||
5222 | } | ||||||||
5223 | |||||||||
5224 | TargetTransformInfo::RegisterKind RegKind = | ||||||||
5225 | ComputeScalableMaxVF ? TargetTransformInfo::RGK_ScalableVector | ||||||||
5226 | : TargetTransformInfo::RGK_FixedWidthVector; | ||||||||
5227 | ElementCount MaxVF = MaxVectorElementCount; | ||||||||
5228 | if (MaximizeBandwidth || (MaximizeBandwidth.getNumOccurrences() == 0 && | ||||||||
5229 | TTI.shouldMaximizeVectorBandwidth(RegKind))) { | ||||||||
5230 | auto MaxVectorElementCountMaxBW = ElementCount::get( | ||||||||
5231 | PowerOf2Floor(WidestRegister.getKnownMinValue() / SmallestType), | ||||||||
5232 | ComputeScalableMaxVF); | ||||||||
5233 | MaxVectorElementCountMaxBW = MinVF(MaxVectorElementCountMaxBW, MaxSafeVF); | ||||||||
5234 | |||||||||
5235 | // Collect all viable vectorization factors larger than the default MaxVF | ||||||||
5236 | // (i.e. MaxVectorElementCount). | ||||||||
5237 | SmallVector<ElementCount, 8> VFs; | ||||||||
5238 | for (ElementCount VS = MaxVectorElementCount * 2; | ||||||||
5239 | ElementCount::isKnownLE(VS, MaxVectorElementCountMaxBW); VS *= 2) | ||||||||
5240 | VFs.push_back(VS); | ||||||||
5241 | |||||||||
5242 | // For each VF calculate its register usage. | ||||||||
5243 | auto RUs = calculateRegisterUsage(VFs); | ||||||||
5244 | |||||||||
5245 | // Select the largest VF which doesn't require more registers than existing | ||||||||
5246 | // ones. | ||||||||
5247 | for (int i = RUs.size() - 1; i >= 0; --i) { | ||||||||
5248 | bool Selected = true; | ||||||||
5249 | for (auto &pair : RUs[i].MaxLocalUsers) { | ||||||||
5250 | unsigned TargetNumRegisters = TTI.getNumberOfRegisters(pair.first); | ||||||||
5251 | if (pair.second > TargetNumRegisters) | ||||||||
5252 | Selected = false; | ||||||||
5253 | } | ||||||||
5254 | if (Selected) { | ||||||||
5255 | MaxVF = VFs[i]; | ||||||||
5256 | break; | ||||||||
5257 | } | ||||||||
5258 | } | ||||||||
5259 | if (ElementCount MinVF = | ||||||||
5260 | TTI.getMinimumVF(SmallestType, ComputeScalableMaxVF)) { | ||||||||
5261 | if (ElementCount::isKnownLT(MaxVF, MinVF)) { | ||||||||
5262 | 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) | ||||||||
5263 | << ") 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); | ||||||||
5264 | MaxVF = MinVF; | ||||||||
5265 | } | ||||||||
5266 | } | ||||||||
5267 | |||||||||
5268 | // Invalidate any widening decisions we might have made, in case the loop | ||||||||
5269 | // requires prediction (decided later), but we have already made some | ||||||||
5270 | // load/store widening decisions. | ||||||||
5271 | invalidateCostModelingDecisions(); | ||||||||
5272 | } | ||||||||
5273 | return MaxVF; | ||||||||
5274 | } | ||||||||
5275 | |||||||||
5276 | std::optional<unsigned> LoopVectorizationCostModel::getVScaleForTuning() const { | ||||||||
5277 | if (TheFunction->hasFnAttribute(Attribute::VScaleRange)) { | ||||||||
5278 | auto Attr = TheFunction->getFnAttribute(Attribute::VScaleRange); | ||||||||
5279 | auto Min = Attr.getVScaleRangeMin(); | ||||||||
5280 | auto Max = Attr.getVScaleRangeMax(); | ||||||||
5281 | if (Max && Min == Max) | ||||||||
5282 | return Max; | ||||||||
5283 | } | ||||||||
5284 | |||||||||
5285 | return TTI.getVScaleForTuning(); | ||||||||
5286 | } | ||||||||
5287 | |||||||||
5288 | bool LoopVectorizationCostModel::isMoreProfitable( | ||||||||
5289 | const VectorizationFactor &A, const VectorizationFactor &B) const { | ||||||||
5290 | InstructionCost CostA = A.Cost; | ||||||||
5291 | InstructionCost CostB = B.Cost; | ||||||||
5292 | |||||||||
5293 | unsigned MaxTripCount = PSE.getSE()->getSmallConstantMaxTripCount(TheLoop); | ||||||||
5294 | |||||||||
5295 | if (!A.Width.isScalable() && !B.Width.isScalable() && FoldTailByMasking && | ||||||||
5296 | MaxTripCount) { | ||||||||
5297 | // If we are folding the tail and the trip count is a known (possibly small) | ||||||||
5298 | // constant, the trip count will be rounded up to an integer number of | ||||||||
5299 | // iterations. The total cost will be PerIterationCost*ceil(TripCount/VF), | ||||||||
5300 | // which we compare directly. When not folding the tail, the total cost will | ||||||||
5301 | // be PerIterationCost*floor(TC/VF) + Scalar remainder cost, and so is | ||||||||
5302 | // approximated with the per-lane cost below instead of using the tripcount | ||||||||
5303 | // as here. | ||||||||
5304 | auto RTCostA = CostA * divideCeil(MaxTripCount, A.Width.getFixedValue()); | ||||||||
5305 | auto RTCostB = CostB * divideCeil(MaxTripCount, B.Width.getFixedValue()); | ||||||||
5306 | return RTCostA < RTCostB; | ||||||||
5307 | } | ||||||||
5308 | |||||||||
5309 | // Improve estimate for the vector width if it is scalable. | ||||||||
5310 | unsigned EstimatedWidthA = A.Width.getKnownMinValue(); | ||||||||
5311 | unsigned EstimatedWidthB = B.Width.getKnownMinValue(); | ||||||||
5312 | if (std::optional<unsigned> VScale = getVScaleForTuning()) { | ||||||||
5313 | if (A.Width.isScalable()) | ||||||||
5314 | EstimatedWidthA *= *VScale; | ||||||||
5315 | if (B.Width.isScalable()) | ||||||||
5316 | EstimatedWidthB *= *VScale; | ||||||||
5317 | } | ||||||||
5318 | |||||||||
5319 | // Assume vscale may be larger than 1 (or the value being tuned for), | ||||||||
5320 | // so that scalable vectorization is slightly favorable over fixed-width | ||||||||
5321 | // vectorization. | ||||||||
5322 | if (A.Width.isScalable() && !B.Width.isScalable()) | ||||||||
5323 | return (CostA * B.Width.getFixedValue()) <= (CostB * EstimatedWidthA); | ||||||||
5324 | |||||||||
5325 | // To avoid the need for FP division: | ||||||||
5326 | // (CostA / A.Width) < (CostB / B.Width) | ||||||||
5327 | // <=> (CostA * B.Width) < (CostB * A.Width) | ||||||||
5328 | return (CostA * EstimatedWidthB) < (CostB * EstimatedWidthA); | ||||||||
5329 | } | ||||||||
5330 | |||||||||
5331 | VectorizationFactor LoopVectorizationCostModel::selectVectorizationFactor( | ||||||||
5332 | const ElementCountSet &VFCandidates) { | ||||||||
5333 | InstructionCost ExpectedCost = expectedCost(ElementCount::getFixed(1)).first; | ||||||||
5334 | 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); | ||||||||
5335 | 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", 5335, __extension__ __PRETTY_FUNCTION__)); | ||||||||
5336 | 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", 5337, __extension__ __PRETTY_FUNCTION__)) | ||||||||
5337 | "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", 5337, __extension__ __PRETTY_FUNCTION__)); | ||||||||
5338 | |||||||||
5339 | const VectorizationFactor ScalarCost(ElementCount::getFixed(1), ExpectedCost, | ||||||||
5340 | ExpectedCost); | ||||||||
5341 | VectorizationFactor ChosenFactor = ScalarCost; | ||||||||
5342 | |||||||||
5343 | bool ForceVectorization = Hints->getForce() == LoopVectorizeHints::FK_Enabled; | ||||||||
5344 | if (ForceVectorization && VFCandidates.size() > 1) { | ||||||||
5345 | // Ignore scalar width, because the user explicitly wants vectorization. | ||||||||
5346 | // Initialize cost to max so that VF = 2 is, at least, chosen during cost | ||||||||
5347 | // evaluation. | ||||||||
5348 | ChosenFactor.Cost = InstructionCost::getMax(); | ||||||||
5349 | } | ||||||||
5350 | |||||||||
5351 | SmallVector<InstructionVFPair> InvalidCosts; | ||||||||
5352 | for (const auto &i : VFCandidates) { | ||||||||
5353 | // The cost for scalar VF=1 is already calculated, so ignore it. | ||||||||
5354 | if (i.isScalar()) | ||||||||
5355 | continue; | ||||||||
5356 | |||||||||
5357 | VectorizationCostTy C = expectedCost(i, &InvalidCosts); | ||||||||
5358 | VectorizationFactor Candidate(i, C.first, ScalarCost.ScalarCost); | ||||||||
5359 | |||||||||
5360 | #ifndef NDEBUG | ||||||||
5361 | unsigned AssumedMinimumVscale = 1; | ||||||||
5362 | if (std::optional<unsigned> VScale = getVScaleForTuning()) | ||||||||
5363 | AssumedMinimumVscale = *VScale; | ||||||||
5364 | unsigned Width = | ||||||||
5365 | Candidate.Width.isScalable() | ||||||||
5366 | ? Candidate.Width.getKnownMinValue() * AssumedMinimumVscale | ||||||||
5367 | : Candidate.Width.getFixedValue(); | ||||||||
5368 | 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) | ||||||||
5369 | << " 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); | ||||||||
5370 | if (i.isScalable()) | ||||||||
5371 | 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 ) | ||||||||
5372 | << AssumedMinimumVscale << ")")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << " (assuming a minimum vscale of " << AssumedMinimumVscale << ")"; } } while (false ); | ||||||||
5373 | LLVM_DEBUG(dbgs() << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << ".\n"; } } while (false ); | ||||||||
5374 | #endif | ||||||||
5375 | |||||||||
5376 | if (!C.second && !ForceVectorization) { | ||||||||
5377 | 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) | ||||||||
5378 | 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) | ||||||||
5379 | << " 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); | ||||||||
5380 | continue; | ||||||||
5381 | } | ||||||||
5382 | |||||||||
5383 | // If profitable add it to ProfitableVF list. | ||||||||
5384 | if (isMoreProfitable(Candidate, ScalarCost)) | ||||||||
5385 | ProfitableVFs.push_back(Candidate); | ||||||||
5386 | |||||||||
5387 | if (isMoreProfitable(Candidate, ChosenFactor)) | ||||||||
5388 | ChosenFactor = Candidate; | ||||||||
5389 | } | ||||||||
5390 | |||||||||
5391 | // Emit a report of VFs with invalid costs in the loop. | ||||||||
5392 | if (!InvalidCosts.empty()) { | ||||||||
5393 | // Group the remarks per instruction, keeping the instruction order from | ||||||||
5394 | // InvalidCosts. | ||||||||
5395 | std::map<Instruction *, unsigned> Numbering; | ||||||||
5396 | unsigned I = 0; | ||||||||
5397 | for (auto &Pair : InvalidCosts) | ||||||||
5398 | if (!Numbering.count(Pair.first)) | ||||||||
5399 | Numbering[Pair.first] = I++; | ||||||||
5400 | |||||||||
5401 | // Sort the list, first on instruction(number) then on VF. | ||||||||
5402 | llvm::sort(InvalidCosts, | ||||||||
5403 | [&Numbering](InstructionVFPair &A, InstructionVFPair &B) { | ||||||||
5404 | if (Numbering[A.first] != Numbering[B.first]) | ||||||||
5405 | return Numbering[A.first] < Numbering[B.first]; | ||||||||
5406 | ElementCountComparator ECC; | ||||||||
5407 | return ECC(A.second, B.second); | ||||||||
5408 | }); | ||||||||
5409 | |||||||||
5410 | // For a list of ordered instruction-vf pairs: | ||||||||
5411 | // [(load, vf1), (load, vf2), (store, vf1)] | ||||||||
5412 | // Group the instructions together to emit separate remarks for: | ||||||||
5413 | // load (vf1, vf2) | ||||||||
5414 | // store (vf1) | ||||||||
5415 | auto Tail = ArrayRef<InstructionVFPair>(InvalidCosts); | ||||||||
5416 | auto Subset = ArrayRef<InstructionVFPair>(); | ||||||||
5417 | do { | ||||||||
5418 | if (Subset.empty()) | ||||||||
5419 | Subset = Tail.take_front(1); | ||||||||
5420 | |||||||||
5421 | Instruction *I = Subset.front().first; | ||||||||
5422 | |||||||||
5423 | // If the next instruction is different, or if there are no other pairs, | ||||||||
5424 | // emit a remark for the collated subset. e.g. | ||||||||
5425 | // [(load, vf1), (load, vf2))] | ||||||||
5426 | // to emit: | ||||||||
5427 | // remark: invalid costs for 'load' at VF=(vf, vf2) | ||||||||
5428 | if (Subset == Tail || Tail[Subset.size()].first != I) { | ||||||||
5429 | std::string OutString; | ||||||||
5430 | raw_string_ostream OS(OutString); | ||||||||
5431 | 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", 5431, __extension__ __PRETTY_FUNCTION__)); | ||||||||
5432 | OS << "Instruction with invalid costs prevented vectorization at VF=("; | ||||||||
5433 | for (const auto &Pair : Subset) | ||||||||
5434 | OS << (Pair.second == Subset.front().second ? "" : ", ") | ||||||||
5435 | << Pair.second; | ||||||||
5436 | OS << "):"; | ||||||||
5437 | if (auto *CI = dyn_cast<CallInst>(I)) | ||||||||
5438 | OS << " call to " << CI->getCalledFunction()->getName(); | ||||||||
5439 | else | ||||||||
5440 | OS << " " << I->getOpcodeName(); | ||||||||
5441 | OS.flush(); | ||||||||
5442 | reportVectorizationInfo(OutString, "InvalidCost", ORE, TheLoop, I); | ||||||||
5443 | Tail = Tail.drop_front(Subset.size()); | ||||||||
5444 | Subset = {}; | ||||||||
5445 | } else | ||||||||
5446 | // Grow the subset by one element | ||||||||
5447 | Subset = Tail.take_front(Subset.size() + 1); | ||||||||
5448 | } while (!Tail.empty()); | ||||||||
5449 | } | ||||||||
5450 | |||||||||
5451 | if (!EnableCondStoresVectorization && NumPredStores) { | ||||||||
5452 | reportVectorizationFailure("There are conditional stores.", | ||||||||
5453 | "store that is conditionally executed prevents vectorization", | ||||||||
5454 | "ConditionalStore", ORE, TheLoop); | ||||||||
5455 | ChosenFactor = ScalarCost; | ||||||||
5456 | } | ||||||||
5457 | |||||||||
5458 | LLVM_DEBUG(if (ForceVectorization && !ChosenFactor.Width.isScalar() &&do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { if (ForceVectorization && !ChosenFactor .Width.isScalar() && !isMoreProfitable(ChosenFactor, ScalarCost )) dbgs() << "LV: Vectorization seems to be not beneficial, " << "but was forced by a user.\n"; } } while (false) | ||||||||
5459 | !isMoreProfitable(ChosenFactor, ScalarCost)) dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { if (ForceVectorization && !ChosenFactor .Width.isScalar() && !isMoreProfitable(ChosenFactor, ScalarCost )) dbgs() << "LV: Vectorization seems to be not beneficial, " << "but was forced by a user.\n"; } } while (false) | ||||||||
5460 | << "LV: Vectorization seems to be not beneficial, "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { if (ForceVectorization && !ChosenFactor .Width.isScalar() && !isMoreProfitable(ChosenFactor, ScalarCost )) dbgs() << "LV: Vectorization seems to be not beneficial, " << "but was forced by a user.\n"; } } while (false) | ||||||||
5461 | << "but was forced by a user.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { if (ForceVectorization && !ChosenFactor .Width.isScalar() && !isMoreProfitable(ChosenFactor, ScalarCost )) dbgs() << "LV: Vectorization seems to be not beneficial, " << "but was forced by a user.\n"; } } while (false); | ||||||||
5462 | 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); | ||||||||
5463 | return ChosenFactor; | ||||||||
5464 | } | ||||||||
5465 | |||||||||
5466 | bool LoopVectorizationCostModel::isCandidateForEpilogueVectorization( | ||||||||
5467 | const Loop &L, ElementCount VF) const { | ||||||||
5468 | // Cross iteration phis such as reductions need special handling and are | ||||||||
5469 | // currently unsupported. | ||||||||
5470 | if (any_of(L.getHeader()->phis(), | ||||||||
5471 | [&](PHINode &Phi) { return Legal->isFixedOrderRecurrence(&Phi); })) | ||||||||
5472 | return false; | ||||||||
5473 | |||||||||
5474 | // Phis with uses outside of the loop require special handling and are | ||||||||
5475 | // currently unsupported. | ||||||||
5476 | for (const auto &Entry : Legal->getInductionVars()) { | ||||||||
5477 | // Look for uses of the value of the induction at the last iteration. | ||||||||
5478 | Value *PostInc = Entry.first->getIncomingValueForBlock(L.getLoopLatch()); | ||||||||
5479 | for (User *U : PostInc->users()) | ||||||||
5480 | if (!L.contains(cast<Instruction>(U))) | ||||||||
5481 | return false; | ||||||||
5482 | // Look for uses of penultimate value of the induction. | ||||||||
5483 | for (User *U : Entry.first->users()) | ||||||||
5484 | if (!L.contains(cast<Instruction>(U))) | ||||||||
5485 | return false; | ||||||||
5486 | } | ||||||||
5487 | |||||||||
5488 | // Epilogue vectorization code has not been auditted to ensure it handles | ||||||||
5489 | // non-latch exits properly. It may be fine, but it needs auditted and | ||||||||
5490 | // tested. | ||||||||
5491 | if (L.getExitingBlock() != L.getLoopLatch()) | ||||||||
5492 | return false; | ||||||||
5493 | |||||||||
5494 | return true; | ||||||||
5495 | } | ||||||||
5496 | |||||||||
5497 | bool LoopVectorizationCostModel::isEpilogueVectorizationProfitable( | ||||||||
5498 | const ElementCount VF) const { | ||||||||
5499 | // FIXME: We need a much better cost-model to take different parameters such | ||||||||
5500 | // as register pressure, code size increase and cost of extra branches into | ||||||||
5501 | // account. For now we apply a very crude heuristic and only consider loops | ||||||||
5502 | // with vectorization factors larger than a certain value. | ||||||||
5503 | |||||||||
5504 | // Allow the target to opt out entirely. | ||||||||
5505 | if (!TTI.preferEpilogueVectorization()) | ||||||||
5506 | return false; | ||||||||
5507 | |||||||||
5508 | // We also consider epilogue vectorization unprofitable for targets that don't | ||||||||
5509 | // consider interleaving beneficial (eg. MVE). | ||||||||
5510 | if (TTI.getMaxInterleaveFactor(VF.getKnownMinValue()) <= 1) | ||||||||
5511 | return false; | ||||||||
5512 | // FIXME: We should consider changing the threshold for scalable | ||||||||
5513 | // vectors to take VScaleForTuning into account. | ||||||||
5514 | if (VF.getKnownMinValue() >= EpilogueVectorizationMinVF) | ||||||||
5515 | return true; | ||||||||
5516 | return false; | ||||||||
5517 | } | ||||||||
5518 | |||||||||
5519 | VectorizationFactor | ||||||||
5520 | LoopVectorizationCostModel::selectEpilogueVectorizationFactor( | ||||||||
5521 | const ElementCount MainLoopVF, const LoopVectorizationPlanner &LVP) { | ||||||||
5522 | VectorizationFactor Result = VectorizationFactor::Disabled(); | ||||||||
5523 | if (!EnableEpilogueVectorization) { | ||||||||
5524 | 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); | ||||||||
5525 | return Result; | ||||||||
5526 | } | ||||||||
5527 | |||||||||
5528 | if (!isScalarEpilogueAllowed()) { | ||||||||
5529 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Unable to vectorize epilogue because no epilogue is " "allowed.\n";; } } while (false) | ||||||||
5530 | 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) | ||||||||
5531 | "allowed.\n";)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Unable to vectorize epilogue because no epilogue is " "allowed.\n";; } } while (false); | ||||||||
5532 | return Result; | ||||||||
5533 | } | ||||||||
5534 | |||||||||
5535 | // Not really a cost consideration, but check for unsupported cases here to | ||||||||
5536 | // simplify the logic. | ||||||||
5537 | if (!isCandidateForEpilogueVectorization(*TheLoop, MainLoopVF)) { | ||||||||
5538 | 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) | ||||||||
5539 | 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) | ||||||||
5540 | "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); | ||||||||
5541 | return Result; | ||||||||
5542 | } | ||||||||
5543 | |||||||||
5544 | if (EpilogueVectorizationForceVF > 1) { | ||||||||
5545 | 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); | ||||||||
5546 | ElementCount ForcedEC = ElementCount::getFixed(EpilogueVectorizationForceVF); | ||||||||
5547 | if (LVP.hasPlanWithVF(ForcedEC)) | ||||||||
5548 | return {ForcedEC, 0, 0}; | ||||||||
5549 | else { | ||||||||
5550 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Epilogue vectorization forced factor is not viable.\n" ;; } } while (false) | ||||||||
5551 | dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Epilogue vectorization forced factor is not viable.\n" ;; } } while (false) | ||||||||
5552 | << "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); | ||||||||
5553 | return Result; | ||||||||
5554 | } | ||||||||
5555 | } | ||||||||
5556 | |||||||||
5557 | if (TheLoop->getHeader()->getParent()->hasOptSize() || | ||||||||
5558 | TheLoop->getHeader()->getParent()->hasMinSize()) { | ||||||||
5559 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Epilogue vectorization skipped due to opt for size.\n" ;; } } while (false) | ||||||||
5560 | dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Epilogue vectorization skipped due to opt for size.\n" ;; } } while (false) | ||||||||
5561 | << "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); | ||||||||
5562 | return Result; | ||||||||
5563 | } | ||||||||
5564 | |||||||||
5565 | if (!isEpilogueVectorizationProfitable(MainLoopVF)) { | ||||||||
5566 | 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) | ||||||||
5567 | "this loop\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Epilogue vectorization is not profitable for " "this loop\n"; } } while (false); | ||||||||
5568 | return Result; | ||||||||
5569 | } | ||||||||
5570 | |||||||||
5571 | // If MainLoopVF = vscale x 2, and vscale is expected to be 4, then we know | ||||||||
5572 | // the main loop handles 8 lanes per iteration. We could still benefit from | ||||||||
5573 | // vectorizing the epilogue loop with VF=4. | ||||||||
5574 | ElementCount EstimatedRuntimeVF = MainLoopVF; | ||||||||
5575 | if (MainLoopVF.isScalable()) { | ||||||||
5576 | EstimatedRuntimeVF = ElementCount::getFixed(MainLoopVF.getKnownMinValue()); | ||||||||
5577 | if (std::optional<unsigned> VScale = getVScaleForTuning()) | ||||||||
5578 | EstimatedRuntimeVF *= *VScale; | ||||||||
5579 | } | ||||||||
5580 | |||||||||
5581 | for (auto &NextVF : ProfitableVFs) | ||||||||
5582 | if (((!NextVF.Width.isScalable() && MainLoopVF.isScalable() && | ||||||||
5583 | ElementCount::isKnownLT(NextVF.Width, EstimatedRuntimeVF)) || | ||||||||
5584 | ElementCount::isKnownLT(NextVF.Width, MainLoopVF)) && | ||||||||
5585 | (Result.Width.isScalar() || isMoreProfitable(NextVF, Result)) && | ||||||||
5586 | LVP.hasPlanWithVF(NextVF.Width)) | ||||||||
5587 | Result = NextVF; | ||||||||
5588 | |||||||||
5589 | if (Result != VectorizationFactor::Disabled()) | ||||||||
5590 | 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) | ||||||||
5591 | << Result.Width << "\n";)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Vectorizing epilogue loop with VF = " << Result.Width << "\n";; } } while (false); | ||||||||
5592 | return Result; | ||||||||
5593 | } | ||||||||
5594 | |||||||||
5595 | std::pair<unsigned, unsigned> | ||||||||
5596 | LoopVectorizationCostModel::getSmallestAndWidestTypes() { | ||||||||
5597 | unsigned MinWidth = -1U; | ||||||||
5598 | unsigned MaxWidth = 8; | ||||||||
5599 | const DataLayout &DL = TheFunction->getParent()->getDataLayout(); | ||||||||
5600 | // For in-loop reductions, no element types are added to ElementTypesInLoop | ||||||||
5601 | // if there are no loads/stores in the loop. In this case, check through the | ||||||||
5602 | // reduction variables to determine the maximum width. | ||||||||
5603 | if (ElementTypesInLoop.empty() && !Legal->getReductionVars().empty()) { | ||||||||
5604 | // Reset MaxWidth so that we can find the smallest type used by recurrences | ||||||||
5605 | // in the loop. | ||||||||
5606 | MaxWidth = -1U; | ||||||||
5607 | for (const auto &PhiDescriptorPair : Legal->getReductionVars()) { | ||||||||
5608 | const RecurrenceDescriptor &RdxDesc = PhiDescriptorPair.second; | ||||||||
5609 | // When finding the min width used by the recurrence we need to account | ||||||||
5610 | // for casts on the input operands of the recurrence. | ||||||||
5611 | MaxWidth = std::min<unsigned>( | ||||||||
5612 | MaxWidth, std::min<unsigned>( | ||||||||
5613 | RdxDesc.getMinWidthCastToRecurrenceTypeInBits(), | ||||||||
5614 | RdxDesc.getRecurrenceType()->getScalarSizeInBits())); | ||||||||
5615 | } | ||||||||
5616 | } else { | ||||||||
5617 | for (Type *T : ElementTypesInLoop) { | ||||||||
5618 | MinWidth = std::min<unsigned>( | ||||||||
5619 | MinWidth, DL.getTypeSizeInBits(T->getScalarType()).getFixedValue()); | ||||||||
5620 | MaxWidth = std::max<unsigned>( | ||||||||
5621 | MaxWidth, DL.getTypeSizeInBits(T->getScalarType()).getFixedValue()); | ||||||||
5622 | } | ||||||||
5623 | } | ||||||||
5624 | return {MinWidth, MaxWidth}; | ||||||||
5625 | } | ||||||||
5626 | |||||||||
5627 | void LoopVectorizationCostModel::collectElementTypesForWidening() { | ||||||||
5628 | ElementTypesInLoop.clear(); | ||||||||
5629 | // For each block. | ||||||||
5630 | for (BasicBlock *BB : TheLoop->blocks()) { | ||||||||
5631 | // For each instruction in the loop. | ||||||||
5632 | for (Instruction &I : BB->instructionsWithoutDebug()) { | ||||||||
5633 | Type *T = I.getType(); | ||||||||
5634 | |||||||||
5635 | // Skip ignored values. | ||||||||
5636 | if (ValuesToIgnore.count(&I)) | ||||||||
5637 | continue; | ||||||||
5638 | |||||||||
5639 | // Only examine Loads, Stores and PHINodes. | ||||||||
5640 | if (!isa<LoadInst>(I) && !isa<StoreInst>(I) && !isa<PHINode>(I)) | ||||||||
5641 | continue; | ||||||||
5642 | |||||||||
5643 | // Examine PHI nodes that are reduction variables. Update the type to | ||||||||
5644 | // account for the recurrence type. | ||||||||
5645 | if (auto *PN = dyn_cast<PHINode>(&I)) { | ||||||||
5646 | if (!Legal->isReductionVariable(PN)) | ||||||||
5647 | continue; | ||||||||
5648 | const RecurrenceDescriptor &RdxDesc = | ||||||||
5649 | Legal->getReductionVars().find(PN)->second; | ||||||||
5650 | if (PreferInLoopReductions || useOrderedReductions(RdxDesc) || | ||||||||
5651 | TTI.preferInLoopReduction(RdxDesc.getOpcode(), | ||||||||
5652 | RdxDesc.getRecurrenceType(), | ||||||||
5653 | TargetTransformInfo::ReductionFlags())) | ||||||||
5654 | continue; | ||||||||
5655 | T = RdxDesc.getRecurrenceType(); | ||||||||
5656 | } | ||||||||
5657 | |||||||||
5658 | // Examine the stored values. | ||||||||
5659 | if (auto *ST = dyn_cast<StoreInst>(&I)) | ||||||||
5660 | T = ST->getValueOperand()->getType(); | ||||||||
5661 | |||||||||
5662 | 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", 5663, __extension__ __PRETTY_FUNCTION__)) | ||||||||
5663 | "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", 5663, __extension__ __PRETTY_FUNCTION__)); | ||||||||
5664 | |||||||||
5665 | ElementTypesInLoop.insert(T); | ||||||||
5666 | } | ||||||||
5667 | } | ||||||||
5668 | } | ||||||||
5669 | |||||||||
5670 | unsigned | ||||||||
5671 | LoopVectorizationCostModel::selectInterleaveCount(ElementCount VF, | ||||||||
5672 | InstructionCost LoopCost) { | ||||||||
5673 | // -- The interleave heuristics -- | ||||||||
5674 | // We interleave the loop in order to expose ILP and reduce the loop overhead. | ||||||||
5675 | // There are many micro-architectural considerations that we can't predict | ||||||||
5676 | // at this level. For example, frontend pressure (on decode or fetch) due to | ||||||||
5677 | // code size, or the number and capabilities of the execution ports. | ||||||||
5678 | // | ||||||||
5679 | // We use the following heuristics to select the interleave count: | ||||||||
5680 | // 1. If the code has reductions, then we interleave to break the cross | ||||||||
5681 | // iteration dependency. | ||||||||
5682 | // 2. If the loop is really small, then we interleave to reduce the loop | ||||||||
5683 | // overhead. | ||||||||
5684 | // 3. We don't interleave if we think that we will spill registers to memory | ||||||||
5685 | // due to the increased register pressure. | ||||||||
5686 | |||||||||
5687 | if (!isScalarEpilogueAllowed()) | ||||||||
5688 | return 1; | ||||||||
5689 | |||||||||
5690 | // We used the distance for the interleave count. | ||||||||
5691 | if (Legal->getMaxSafeDepDistBytes() != -1U) | ||||||||
5692 | return 1; | ||||||||
5693 | |||||||||
5694 | auto BestKnownTC = getSmallBestKnownTC(*PSE.getSE(), TheLoop); | ||||||||
5695 | const bool HasReductions = !Legal->getReductionVars().empty(); | ||||||||
5696 | // Do not interleave loops with a relatively small known or estimated trip | ||||||||
5697 | // count. But we will interleave when InterleaveSmallLoopScalarReduction is | ||||||||
5698 | // enabled, and the code has scalar reductions(HasReductions && VF = 1), | ||||||||
5699 | // because with the above conditions interleaving can expose ILP and break | ||||||||
5700 | // cross iteration dependences for reductions. | ||||||||
5701 | if (BestKnownTC && (*BestKnownTC < TinyTripCountInterleaveThreshold) && | ||||||||
5702 | !(InterleaveSmallLoopScalarReduction && HasReductions && VF.isScalar())) | ||||||||
5703 | return 1; | ||||||||
5704 | |||||||||
5705 | // If we did not calculate the cost for VF (because the user selected the VF) | ||||||||
5706 | // then we calculate the cost of VF here. | ||||||||
5707 | if (LoopCost == 0) { | ||||||||
5708 | LoopCost = expectedCost(VF).first; | ||||||||
5709 | assert(LoopCost.isValid() && "Expected to have chosen a VF with valid cost")(static_cast <bool> (LoopCost.isValid() && "Expected to have chosen a VF with valid cost" ) ? void (0) : __assert_fail ("LoopCost.isValid() && \"Expected to have chosen a VF with valid cost\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 5709, __extension__ __PRETTY_FUNCTION__)); | ||||||||
5710 | |||||||||
5711 | // Loop body is free and there is no need for interleaving. | ||||||||
5712 | if (LoopCost == 0) | ||||||||
5713 | return 1; | ||||||||
5714 | } | ||||||||
5715 | |||||||||
5716 | RegisterUsage R = calculateRegisterUsage({VF})[0]; | ||||||||
5717 | // We divide by these constants so assume that we have at least one | ||||||||
5718 | // instruction that uses at least one register. | ||||||||
5719 | for (auto& pair : R.MaxLocalUsers) { | ||||||||
5720 | pair.second = std::max(pair.second, 1U); | ||||||||
5721 | } | ||||||||
5722 | |||||||||
5723 | // We calculate the interleave count using the following formula. | ||||||||
5724 | // Subtract the number of loop invariants from the number of available | ||||||||
5725 | // registers. These registers are used by all of the interleaved instances. | ||||||||
5726 | // Next, divide the remaining registers by the number of registers that is | ||||||||
5727 | // required by the loop, in order to estimate how many parallel instances | ||||||||
5728 | // fit without causing spills. All of this is rounded down if necessary to be | ||||||||
5729 | // a power of two. We want power of two interleave count to simplify any | ||||||||
5730 | // addressing operations or alignment considerations. | ||||||||
5731 | // We also want power of two interleave counts to ensure that the induction | ||||||||
5732 | // variable of the vector loop wraps to zero, when tail is folded by masking; | ||||||||
5733 | // this currently happens when OptForSize, in which case IC is set to 1 above. | ||||||||
5734 | unsigned IC = UINT_MAX(2147483647 *2U +1U); | ||||||||
5735 | |||||||||
5736 | for (auto& pair : R.MaxLocalUsers) { | ||||||||
5737 | unsigned TargetNumRegisters = TTI.getNumberOfRegisters(pair.first); | ||||||||
5738 | 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) | ||||||||
5739 | << " 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) | ||||||||
5740 | << 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); | ||||||||
5741 | if (VF.isScalar()) { | ||||||||
5742 | if (ForceTargetNumScalarRegs.getNumOccurrences() > 0) | ||||||||
5743 | TargetNumRegisters = ForceTargetNumScalarRegs; | ||||||||
5744 | } else { | ||||||||
5745 | if (ForceTargetNumVectorRegs.getNumOccurrences() > 0) | ||||||||
5746 | TargetNumRegisters = ForceTargetNumVectorRegs; | ||||||||
5747 | } | ||||||||
5748 | unsigned MaxLocalUsers = pair.second; | ||||||||
5749 | unsigned LoopInvariantRegs = 0; | ||||||||
5750 | if (R.LoopInvariantRegs.find(pair.first) != R.LoopInvariantRegs.end()) | ||||||||
5751 | LoopInvariantRegs = R.LoopInvariantRegs[pair.first]; | ||||||||
5752 | |||||||||
5753 | unsigned TmpIC = PowerOf2Floor((TargetNumRegisters - LoopInvariantRegs) / MaxLocalUsers); | ||||||||
5754 | // Don't count the induction variable as interleaved. | ||||||||
5755 | if (EnableIndVarRegisterHeur) { | ||||||||
5756 | TmpIC = | ||||||||
5757 | PowerOf2Floor((TargetNumRegisters - LoopInvariantRegs - 1) / | ||||||||
5758 | std::max(1U, (MaxLocalUsers - 1))); | ||||||||
5759 | } | ||||||||
5760 | |||||||||
5761 | IC = std::min(IC, TmpIC); | ||||||||
5762 | } | ||||||||
5763 | |||||||||
5764 | // Clamp the interleave ranges to reasonable counts. | ||||||||
5765 | unsigned MaxInterleaveCount = | ||||||||
5766 | TTI.getMaxInterleaveFactor(VF.getKnownMinValue()); | ||||||||
5767 | |||||||||
5768 | // Check if the user has overridden the max. | ||||||||
5769 | if (VF.isScalar()) { | ||||||||
5770 | if (ForceTargetMaxScalarInterleaveFactor.getNumOccurrences() > 0) | ||||||||
5771 | MaxInterleaveCount = ForceTargetMaxScalarInterleaveFactor; | ||||||||
5772 | } else { | ||||||||
5773 | if (ForceTargetMaxVectorInterleaveFactor.getNumOccurrences() > 0) | ||||||||
5774 | MaxInterleaveCount = ForceTargetMaxVectorInterleaveFactor; | ||||||||
5775 | } | ||||||||
5776 | |||||||||
5777 | // If trip count is known or estimated compile time constant, limit the | ||||||||
5778 | // interleave count to be less than the trip count divided by VF, provided it | ||||||||
5779 | // is at least 1. | ||||||||
5780 | // | ||||||||
5781 | // For scalable vectors we can't know if interleaving is beneficial. It may | ||||||||
5782 | // not be beneficial for small loops if none of the lanes in the second vector | ||||||||
5783 | // iterations is enabled. However, for larger loops, there is likely to be a | ||||||||
5784 | // similar benefit as for fixed-width vectors. For now, we choose to leave | ||||||||
5785 | // the InterleaveCount as if vscale is '1', although if some information about | ||||||||
5786 | // the vector is known (e.g. min vector size), we can make a better decision. | ||||||||
5787 | if (BestKnownTC) { | ||||||||
5788 | MaxInterleaveCount = | ||||||||
5789 | std::min(*BestKnownTC / VF.getKnownMinValue(), MaxInterleaveCount); | ||||||||
5790 | // Make sure MaxInterleaveCount is greater than 0. | ||||||||
5791 | MaxInterleaveCount = std::max(1u, MaxInterleaveCount); | ||||||||
5792 | } | ||||||||
5793 | |||||||||
5794 | 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", 5795, __extension__ __PRETTY_FUNCTION__)) | ||||||||
5795 | "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", 5795, __extension__ __PRETTY_FUNCTION__)); | ||||||||
5796 | |||||||||
5797 | // Clamp the calculated IC to be between the 1 and the max interleave count | ||||||||
5798 | // that the target and trip count allows. | ||||||||
5799 | if (IC > MaxInterleaveCount) | ||||||||
5800 | IC = MaxInterleaveCount; | ||||||||
5801 | else | ||||||||
5802 | // Make sure IC is greater than 0. | ||||||||
5803 | IC = std::max(1u, IC); | ||||||||
5804 | |||||||||
5805 | 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", 5805, __extension__ __PRETTY_FUNCTION__)); | ||||||||
5806 | |||||||||
5807 | // Interleave if we vectorized this loop and there is a reduction that could | ||||||||
5808 | // benefit from interleaving. | ||||||||
5809 | if (VF.isVector() && HasReductions) { | ||||||||
5810 | 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); | ||||||||
5811 | return IC; | ||||||||
5812 | } | ||||||||
5813 | |||||||||
5814 | // For any scalar loop that either requires runtime checks or predication we | ||||||||
5815 | // are better off leaving this to the unroller. Note that if we've already | ||||||||
5816 | // vectorized the loop we will have done the runtime check and so interleaving | ||||||||
5817 | // won't require further checks. | ||||||||
5818 | bool ScalarInterleavingRequiresPredication = | ||||||||
5819 | (VF.isScalar() && any_of(TheLoop->blocks(), [this](BasicBlock *BB) { | ||||||||
5820 | return Legal->blockNeedsPredication(BB); | ||||||||
5821 | })); | ||||||||
5822 | bool ScalarInterleavingRequiresRuntimePointerCheck = | ||||||||
5823 | (VF.isScalar() && Legal->getRuntimePointerChecking()->Need); | ||||||||
5824 | |||||||||
5825 | // We want to interleave small loops in order to reduce the loop overhead and | ||||||||
5826 | // potentially expose ILP opportunities. | ||||||||
5827 | 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) | ||||||||
5828 | << "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) | ||||||||
5829 | << "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); | ||||||||
5830 | const bool AggressivelyInterleaveReductions = | ||||||||
5831 | TTI.enableAggressiveInterleaving(HasReductions); | ||||||||
5832 | if (!ScalarInterleavingRequiresRuntimePointerCheck && | ||||||||
5833 | !ScalarInterleavingRequiresPredication && LoopCost < SmallLoopCost) { | ||||||||
5834 | // We assume that the cost overhead is 1 and we use the cost model | ||||||||
5835 | // to estimate the cost of the loop and interleave until the cost of the | ||||||||
5836 | // loop overhead is about 5% of the cost of the loop. | ||||||||
5837 | unsigned SmallIC = std::min( | ||||||||
5838 | IC, (unsigned)PowerOf2Floor(SmallLoopCost / *LoopCost.getValue())); | ||||||||
5839 | |||||||||
5840 | // Interleave until store/load ports (estimated by max interleave count) are | ||||||||
5841 | // saturated. | ||||||||
5842 | unsigned NumStores = Legal->getNumStores(); | ||||||||
5843 | unsigned NumLoads = Legal->getNumLoads(); | ||||||||
5844 | unsigned StoresIC = IC / (NumStores ? NumStores : 1); | ||||||||
5845 | unsigned LoadsIC = IC / (NumLoads ? NumLoads : 1); | ||||||||
5846 | |||||||||
5847 | // There is little point in interleaving for reductions containing selects | ||||||||
5848 | // and compares when VF=1 since it may just create more overhead than it's | ||||||||
5849 | // worth for loops with small trip counts. This is because we still have to | ||||||||
5850 | // do the final reduction after the loop. | ||||||||
5851 | bool HasSelectCmpReductions = | ||||||||
5852 | HasReductions && | ||||||||
5853 | any_of(Legal->getReductionVars(), [&](auto &Reduction) -> bool { | ||||||||
5854 | const RecurrenceDescriptor &RdxDesc = Reduction.second; | ||||||||
5855 | return RecurrenceDescriptor::isSelectCmpRecurrenceKind( | ||||||||
5856 | RdxDesc.getRecurrenceKind()); | ||||||||
5857 | }); | ||||||||
5858 | if (HasSelectCmpReductions) { | ||||||||
5859 | 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); | ||||||||
5860 | return 1; | ||||||||
5861 | } | ||||||||
5862 | |||||||||
5863 | // If we have a scalar reduction (vector reductions are already dealt with | ||||||||
5864 | // by this point), we can increase the critical path length if the loop | ||||||||
5865 | // we're interleaving is inside another loop. For tree-wise reductions | ||||||||
5866 | // set the limit to 2, and for ordered reductions it's best to disable | ||||||||
5867 | // interleaving entirely. | ||||||||
5868 | if (HasReductions && TheLoop->getLoopDepth() > 1) { | ||||||||
5869 | bool HasOrderedReductions = | ||||||||
5870 | any_of(Legal->getReductionVars(), [&](auto &Reduction) -> bool { | ||||||||
5871 | const RecurrenceDescriptor &RdxDesc = Reduction.second; | ||||||||
5872 | return RdxDesc.isOrdered(); | ||||||||
5873 | }); | ||||||||
5874 | if (HasOrderedReductions) { | ||||||||
5875 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Not interleaving scalar ordered reductions.\n" ; } } while (false) | ||||||||
5876 | 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); | ||||||||
5877 | return 1; | ||||||||
5878 | } | ||||||||
5879 | |||||||||
5880 | unsigned F = static_cast<unsigned>(MaxNestedScalarReductionIC); | ||||||||
5881 | SmallIC = std::min(SmallIC, F); | ||||||||
5882 | StoresIC = std::min(StoresIC, F); | ||||||||
5883 | LoadsIC = std::min(LoadsIC, F); | ||||||||
5884 | } | ||||||||
5885 | |||||||||
5886 | if (EnableLoadStoreRuntimeInterleave && | ||||||||
5887 | std::max(StoresIC, LoadsIC) > SmallIC) { | ||||||||
5888 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Interleaving to saturate store or load ports.\n" ; } } while (false) | ||||||||
5889 | 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); | ||||||||
5890 | return std::max(StoresIC, LoadsIC); | ||||||||
5891 | } | ||||||||
5892 | |||||||||
5893 | // If there are scalar reductions and TTI has enabled aggressive | ||||||||
5894 | // interleaving for reductions, we will interleave to expose ILP. | ||||||||
5895 | if (InterleaveSmallLoopScalarReduction && VF.isScalar() && | ||||||||
5896 | AggressivelyInterleaveReductions) { | ||||||||
5897 | 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); | ||||||||
5898 | // Interleave no less than SmallIC but not as aggressive as the normal IC | ||||||||
5899 | // to satisfy the rare situation when resources are too limited. | ||||||||
5900 | return std::max(IC / 2, SmallIC); | ||||||||
5901 | } else { | ||||||||
5902 | 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); | ||||||||
5903 | return SmallIC; | ||||||||
5904 | } | ||||||||
5905 | } | ||||||||
5906 | |||||||||
5907 | // Interleave if this is a large loop (small loops are already dealt with by | ||||||||
5908 | // this point) that could benefit from interleaving. | ||||||||
5909 | if (AggressivelyInterleaveReductions) { | ||||||||
5910 | 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); | ||||||||
5911 | return IC; | ||||||||
5912 | } | ||||||||
5913 | |||||||||
5914 | LLVM_DEBUG(dbgs() << "LV: Not Interleaving.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Not Interleaving.\n" ; } } while (false); | ||||||||
5915 | return 1; | ||||||||
5916 | } | ||||||||
5917 | |||||||||
5918 | SmallVector<LoopVectorizationCostModel::RegisterUsage, 8> | ||||||||
5919 | LoopVectorizationCostModel::calculateRegisterUsage(ArrayRef<ElementCount> VFs) { | ||||||||
5920 | // This function calculates the register usage by measuring the highest number | ||||||||
5921 | // of values that are alive at a single location. Obviously, this is a very | ||||||||
5922 | // rough estimation. We scan the loop in a topological order in order and | ||||||||
5923 | // assign a number to each instruction. We use RPO to ensure that defs are | ||||||||
5924 | // met before their users. We assume that each instruction that has in-loop | ||||||||
5925 | // users starts an interval. We record every time that an in-loop value is | ||||||||
5926 | // used, so we have a list of the first and last occurrences of each | ||||||||
5927 | // instruction. Next, we transpose this data structure into a multi map that | ||||||||
5928 | // holds the list of intervals that *end* at a specific location. This multi | ||||||||
5929 | // map allows us to perform a linear search. We scan the instructions linearly | ||||||||
5930 | // and record each time that a new interval starts, by placing it in a set. | ||||||||
5931 | // If we find this value in the multi-map then we remove it from the set. | ||||||||
5932 | // The max register usage is the maximum size of the set. | ||||||||
5933 | // We also search for instructions that are defined outside the loop, but are | ||||||||
5934 | // used inside the loop. We need this number separately from the max-interval | ||||||||
5935 | // usage number because when we unroll, loop-invariant values do not take | ||||||||
5936 | // more register. | ||||||||
5937 | LoopBlocksDFS DFS(TheLoop); | ||||||||
5938 | DFS.perform(LI); | ||||||||
5939 | |||||||||
5940 | RegisterUsage RU; | ||||||||
5941 | |||||||||
5942 | // Each 'key' in the map opens a new interval. The values | ||||||||
5943 | // of the map are the index of the 'last seen' usage of the | ||||||||
5944 | // instruction that is the key. | ||||||||
5945 | using IntervalMap = DenseMap<Instruction *, unsigned>; | ||||||||
5946 | |||||||||
5947 | // Maps instruction to its index. | ||||||||
5948 | SmallVector<Instruction *, 64> IdxToInstr; | ||||||||
5949 | // Marks the end of each interval. | ||||||||
5950 | IntervalMap EndPoint; | ||||||||
5951 | // Saves the list of instruction indices that are used in the loop. | ||||||||
5952 | SmallPtrSet<Instruction *, 8> Ends; | ||||||||
5953 | // Saves the list of values that are used in the loop but are defined outside | ||||||||
5954 | // the loop (not including non-instruction values such as arguments and | ||||||||
5955 | // constants). | ||||||||
5956 | SmallPtrSet<Value *, 8> LoopInvariants; | ||||||||
5957 | |||||||||
5958 | for (BasicBlock *BB : make_range(DFS.beginRPO(), DFS.endRPO())) { | ||||||||
5959 | for (Instruction &I : BB->instructionsWithoutDebug()) { | ||||||||
5960 | IdxToInstr.push_back(&I); | ||||||||
5961 | |||||||||
5962 | // Save the end location of each USE. | ||||||||
5963 | for (Value *U : I.operands()) { | ||||||||
5964 | auto *Instr = dyn_cast<Instruction>(U); | ||||||||
5965 | |||||||||
5966 | // Ignore non-instruction values such as arguments, constants, etc. | ||||||||
5967 | // FIXME: Might need some motivation why these values are ignored. If | ||||||||
5968 | // for example an argument is used inside the loop it will increase the | ||||||||
5969 | // register pressure (so shouldn't we add it to LoopInvariants). | ||||||||
5970 | if (!Instr) | ||||||||
5971 | continue; | ||||||||
5972 | |||||||||
5973 | // If this instruction is outside the loop then record it and continue. | ||||||||
5974 | if (!TheLoop->contains(Instr)) { | ||||||||
5975 | LoopInvariants.insert(Instr); | ||||||||
5976 | continue; | ||||||||
5977 | } | ||||||||
5978 | |||||||||
5979 | // Overwrite previous end points. | ||||||||
5980 | EndPoint[Instr] = IdxToInstr.size(); | ||||||||
5981 | Ends.insert(Instr); | ||||||||
5982 | } | ||||||||
5983 | } | ||||||||
5984 | } | ||||||||
5985 | |||||||||
5986 | // Saves the list of intervals that end with the index in 'key'. | ||||||||
5987 | using InstrList = SmallVector<Instruction *, 2>; | ||||||||
5988 | DenseMap<unsigned, InstrList> TransposeEnds; | ||||||||
5989 | |||||||||
5990 | // Transpose the EndPoints to a list of values that end at each index. | ||||||||
5991 | for (auto &Interval : EndPoint) | ||||||||
5992 | TransposeEnds[Interval.second].push_back(Interval.first); | ||||||||
5993 | |||||||||
5994 | SmallPtrSet<Instruction *, 8> OpenIntervals; | ||||||||
5995 | SmallVector<RegisterUsage, 8> RUs(VFs.size()); | ||||||||
5996 | SmallVector<SmallMapVector<unsigned, unsigned, 4>, 8> MaxUsages(VFs.size()); | ||||||||
5997 | |||||||||
5998 | 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); | ||||||||
5999 | |||||||||
6000 | const auto &TTICapture = TTI; | ||||||||
6001 | auto GetRegUsage = [&TTICapture](Type *Ty, ElementCount VF) -> unsigned { | ||||||||
6002 | if (Ty->isTokenTy() || !VectorType::isValidElementType(Ty)) | ||||||||
6003 | return 0; | ||||||||
6004 | return TTICapture.getRegUsageForType(VectorType::get(Ty, VF)); | ||||||||
6005 | }; | ||||||||
6006 | |||||||||
6007 | for (unsigned int i = 0, s = IdxToInstr.size(); i < s; ++i) { | ||||||||
6008 | Instruction *I = IdxToInstr[i]; | ||||||||
6009 | |||||||||
6010 | // Remove all of the instructions that end at this location. | ||||||||
6011 | InstrList &List = TransposeEnds[i]; | ||||||||
6012 | for (Instruction *ToRemove : List) | ||||||||
6013 | OpenIntervals.erase(ToRemove); | ||||||||
6014 | |||||||||
6015 | // Ignore instructions that are never used within the loop. | ||||||||
6016 | if (!Ends.count(I)) | ||||||||
6017 | continue; | ||||||||
6018 | |||||||||
6019 | // Skip ignored values. | ||||||||
6020 | if (ValuesToIgnore.count(I)) | ||||||||
6021 | continue; | ||||||||
6022 | |||||||||
6023 | // For each VF find the maximum usage of registers. | ||||||||
6024 | for (unsigned j = 0, e = VFs.size(); j < e; ++j) { | ||||||||
6025 | // Count the number of registers used, per register class, given all open | ||||||||
6026 | // intervals. | ||||||||
6027 | // Note that elements in this SmallMapVector will be default constructed | ||||||||
6028 | // as 0. So we can use "RegUsage[ClassID] += n" in the code below even if | ||||||||
6029 | // there is no previous entry for ClassID. | ||||||||
6030 | SmallMapVector<unsigned, unsigned, 4> RegUsage; | ||||||||
6031 | |||||||||
6032 | if (VFs[j].isScalar()) { | ||||||||
6033 | for (auto *Inst : OpenIntervals) { | ||||||||
6034 | unsigned ClassID = | ||||||||
6035 | TTI.getRegisterClassForType(false, Inst->getType()); | ||||||||
6036 | // FIXME: The target might use more than one register for the type | ||||||||
6037 | // even in the scalar case. | ||||||||
6038 | RegUsage[ClassID] += 1; | ||||||||
6039 | } | ||||||||
6040 | } else { | ||||||||
6041 | collectUniformsAndScalars(VFs[j]); | ||||||||
6042 | for (auto *Inst : OpenIntervals) { | ||||||||
6043 | // Skip ignored values for VF > 1. | ||||||||
6044 | if (VecValuesToIgnore.count(Inst)) | ||||||||
6045 | continue; | ||||||||
6046 | if (isScalarAfterVectorization(Inst, VFs[j])) { | ||||||||
6047 | unsigned ClassID = | ||||||||
6048 | TTI.getRegisterClassForType(false, Inst->getType()); | ||||||||
6049 | // FIXME: The target might use more than one register for the type | ||||||||
6050 | // even in the scalar case. | ||||||||
6051 | RegUsage[ClassID] += 1; | ||||||||
6052 | } else { | ||||||||
6053 | unsigned ClassID = | ||||||||
6054 | TTI.getRegisterClassForType(true, Inst->getType()); | ||||||||
6055 | RegUsage[ClassID] += GetRegUsage(Inst->getType(), VFs[j]); | ||||||||
6056 | } | ||||||||
6057 | } | ||||||||
6058 | } | ||||||||
6059 | |||||||||
6060 | for (auto& pair : RegUsage) { | ||||||||
6061 | auto &Entry = MaxUsages[j][pair.first]; | ||||||||
6062 | Entry = std::max(Entry, pair.second); | ||||||||
6063 | } | ||||||||
6064 | } | ||||||||
6065 | |||||||||
6066 | 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) | ||||||||
6067 | << OpenIntervals.size() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV(REG): At #" << i << " Interval # " << OpenIntervals.size() << '\n'; } } while (false); | ||||||||
6068 | |||||||||
6069 | // Add the current instruction to the list of open intervals. | ||||||||
6070 | OpenIntervals.insert(I); | ||||||||
6071 | } | ||||||||
6072 | |||||||||
6073 | for (unsigned i = 0, e = VFs.size(); i < e; ++i) { | ||||||||
6074 | // Note that elements in this SmallMapVector will be default constructed | ||||||||
6075 | // as 0. So we can use "Invariant[ClassID] += n" in the code below even if | ||||||||
6076 | // there is no previous entry for ClassID. | ||||||||
6077 | SmallMapVector<unsigned, unsigned, 4> Invariant; | ||||||||
6078 | |||||||||
6079 | for (auto *Inst : LoopInvariants) { | ||||||||
6080 | // FIXME: The target might use more than one register for the type | ||||||||
6081 | // even in the scalar case. | ||||||||
6082 | unsigned Usage = | ||||||||
6083 | VFs[i].isScalar() ? 1 : GetRegUsage(Inst->getType(), VFs[i]); | ||||||||
6084 | unsigned ClassID = | ||||||||
6085 | TTI.getRegisterClassForType(VFs[i].isVector(), Inst->getType()); | ||||||||
6086 | Invariant[ClassID] += Usage; | ||||||||
6087 | } | ||||||||
6088 | |||||||||
6089 | 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) | ||||||||
6090 | 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) | ||||||||
6091 | 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) | ||||||||
6092 | << " 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) | ||||||||
6093 | 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) | ||||||||
6094 | 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) | ||||||||
6095 | << 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) | ||||||||
6096 | << " 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) | ||||||||
6097 | }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) | ||||||||
6098 | 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) | ||||||||
6099 | << " 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) | ||||||||
6100 | 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) | ||||||||
6101 | 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) | ||||||||
6102 | << 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) | ||||||||
6103 | << " 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) | ||||||||
6104 | }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) | ||||||||
6105 | })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); | ||||||||
6106 | |||||||||
6107 | RU.LoopInvariantRegs = Invariant; | ||||||||
6108 | RU.MaxLocalUsers = MaxUsages[i]; | ||||||||
6109 | RUs[i] = RU; | ||||||||
6110 | } | ||||||||
6111 | |||||||||
6112 | return RUs; | ||||||||
6113 | } | ||||||||
6114 | |||||||||
6115 | bool LoopVectorizationCostModel::useEmulatedMaskMemRefHack(Instruction *I, | ||||||||
6116 | ElementCount VF) { | ||||||||
6117 | // TODO: Cost model for emulated masked load/store is completely | ||||||||
6118 | // broken. This hack guides the cost model to use an artificially | ||||||||
6119 | // high enough value to practically disable vectorization with such | ||||||||
6120 | // operations, except where previously deployed legality hack allowed | ||||||||
6121 | // using very low cost values. This is to avoid regressions coming simply | ||||||||
6122 | // from moving "masked load/store" check from legality to cost model. | ||||||||
6123 | // Masked Load/Gather emulation was previously never allowed. | ||||||||
6124 | // Limited number of Masked Store/Scatter emulation was allowed. | ||||||||
6125 | assert((isPredicatedInst(I)) &&(static_cast <bool> ((isPredicatedInst(I)) && "Expecting a scalar emulated instruction" ) ? void (0) : __assert_fail ("(isPredicatedInst(I)) && \"Expecting a scalar emulated instruction\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 6126, __extension__ __PRETTY_FUNCTION__)) | ||||||||
6126 | "Expecting a scalar emulated instruction")(static_cast <bool> ((isPredicatedInst(I)) && "Expecting a scalar emulated instruction" ) ? void (0) : __assert_fail ("(isPredicatedInst(I)) && \"Expecting a scalar emulated instruction\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 6126, __extension__ __PRETTY_FUNCTION__)); | ||||||||
6127 | return isa<LoadInst>(I) || | ||||||||
6128 | (isa<StoreInst>(I) && | ||||||||
6129 | NumPredStores > NumberOfStoresToPredicate); | ||||||||
6130 | } | ||||||||
6131 | |||||||||
6132 | void LoopVectorizationCostModel::collectInstsToScalarize(ElementCount VF) { | ||||||||
6133 | // If we aren't vectorizing the loop, or if we've already collected the | ||||||||
6134 | // instructions to scalarize, there's nothing to do. Collection may already | ||||||||
6135 | // have occurred if we have a user-selected VF and are now computing the | ||||||||
6136 | // expected cost for interleaving. | ||||||||
6137 | if (VF.isScalar() || VF.isZero() || | ||||||||
6138 | InstsToScalarize.find(VF) != InstsToScalarize.end()) | ||||||||
6139 | return; | ||||||||
6140 | |||||||||
6141 | // Initialize a mapping for VF in InstsToScalalarize. If we find that it's | ||||||||
6142 | // not profitable to scalarize any instructions, the presence of VF in the | ||||||||
6143 | // map will indicate that we've analyzed it already. | ||||||||
6144 | ScalarCostsTy &ScalarCostsVF = InstsToScalarize[VF]; | ||||||||
6145 | |||||||||
6146 | PredicatedBBsAfterVectorization[VF].clear(); | ||||||||
6147 | |||||||||
6148 | // Find all the instructions that are scalar with predication in the loop and | ||||||||
6149 | // determine if it would be better to not if-convert the blocks they are in. | ||||||||
6150 | // If so, we also record the instructions to scalarize. | ||||||||
6151 | for (BasicBlock *BB : TheLoop->blocks()) { | ||||||||
6152 | if (!blockNeedsPredicationForAnyReason(BB)) | ||||||||
6153 | continue; | ||||||||
6154 | for (Instruction &I : *BB) | ||||||||
6155 | if (isScalarWithPredication(&I, VF)) { | ||||||||
6156 | ScalarCostsTy ScalarCosts; | ||||||||
6157 | // Do not apply discount if scalable, because that would lead to | ||||||||
6158 | // invalid scalarization costs. | ||||||||
6159 | // Do not apply discount logic if hacked cost is needed | ||||||||
6160 | // for emulated masked memrefs. | ||||||||
6161 | if (!VF.isScalable() && !useEmulatedMaskMemRefHack(&I, VF) && | ||||||||
6162 | computePredInstDiscount(&I, ScalarCosts, VF) >= 0) | ||||||||
6163 | ScalarCostsVF.insert(ScalarCosts.begin(), ScalarCosts.end()); | ||||||||
6164 | // Remember that BB will remain after vectorization. | ||||||||
6165 | PredicatedBBsAfterVectorization[VF].insert(BB); | ||||||||
6166 | } | ||||||||
6167 | } | ||||||||
6168 | } | ||||||||
6169 | |||||||||
6170 | InstructionCost LoopVectorizationCostModel::computePredInstDiscount( | ||||||||
6171 | Instruction *PredInst, ScalarCostsTy &ScalarCosts, ElementCount VF) { | ||||||||
6172 | 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", 6173, __extension__ __PRETTY_FUNCTION__)) | ||||||||
6173 | "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", 6173, __extension__ __PRETTY_FUNCTION__)); | ||||||||
6174 | |||||||||
6175 | // Initialize the discount to zero, meaning that the scalar version and the | ||||||||
6176 | // vector version cost the same. | ||||||||
6177 | InstructionCost Discount = 0; | ||||||||
6178 | |||||||||
6179 | // Holds instructions to analyze. The instructions we visit are mapped in | ||||||||
6180 | // ScalarCosts. Those instructions are the ones that would be scalarized if | ||||||||
6181 | // we find that the scalar version costs less. | ||||||||
6182 | SmallVector<Instruction *, 8> Worklist; | ||||||||
6183 | |||||||||
6184 | // Returns true if the given instruction can be scalarized. | ||||||||
6185 | auto canBeScalarized = [&](Instruction *I) -> bool { | ||||||||
6186 | // We only attempt to scalarize instructions forming a single-use chain | ||||||||
6187 | // from the original predicated block that would otherwise be vectorized. | ||||||||
6188 | // Although not strictly necessary, we give up on instructions we know will | ||||||||
6189 | // already be scalar to avoid traversing chains that are unlikely to be | ||||||||
6190 | // beneficial. | ||||||||
6191 | if (!I->hasOneUse() || PredInst->getParent() != I->getParent() || | ||||||||
6192 | isScalarAfterVectorization(I, VF)) | ||||||||
6193 | return false; | ||||||||
6194 | |||||||||
6195 | // If the instruction is scalar with predication, it will be analyzed | ||||||||
6196 | // separately. We ignore it within the context of PredInst. | ||||||||
6197 | if (isScalarWithPredication(I, VF)) | ||||||||
6198 | return false; | ||||||||
6199 | |||||||||
6200 | // If any of the instruction's operands are uniform after vectorization, | ||||||||
6201 | // the instruction cannot be scalarized. This prevents, for example, a | ||||||||
6202 | // masked load from being scalarized. | ||||||||
6203 | // | ||||||||
6204 | // We assume we will only emit a value for lane zero of an instruction | ||||||||
6205 | // marked uniform after vectorization, rather than VF identical values. | ||||||||
6206 | // Thus, if we scalarize an instruction that uses a uniform, we would | ||||||||
6207 | // create uses of values corresponding to the lanes we aren't emitting code | ||||||||
6208 | // for. This behavior can be changed by allowing getScalarValue to clone | ||||||||
6209 | // the lane zero values for uniforms rather than asserting. | ||||||||
6210 | for (Use &U : I->operands()) | ||||||||
6211 | if (auto *J = dyn_cast<Instruction>(U.get())) | ||||||||
6212 | if (isUniformAfterVectorization(J, VF)) | ||||||||
6213 | return false; | ||||||||
6214 | |||||||||
6215 | // Otherwise, we can scalarize the instruction. | ||||||||
6216 | return true; | ||||||||
6217 | }; | ||||||||
6218 | |||||||||
6219 | // Compute the expected cost discount from scalarizing the entire expression | ||||||||
6220 | // feeding the predicated instruction. We currently only consider expressions | ||||||||
6221 | // that are single-use instruction chains. | ||||||||
6222 | Worklist.push_back(PredInst); | ||||||||
6223 | while (!Worklist.empty()) { | ||||||||
6224 | Instruction *I = Worklist.pop_back_val(); | ||||||||
6225 | |||||||||
6226 | // If we've already analyzed the instruction, there's nothing to do. | ||||||||
6227 | if (ScalarCosts.find(I) != ScalarCosts.end()) | ||||||||
6228 | continue; | ||||||||
6229 | |||||||||
6230 | // Compute the cost of the vector instruction. Note that this cost already | ||||||||
6231 | // includes the scalarization overhead of the predicated instruction. | ||||||||
6232 | InstructionCost VectorCost = getInstructionCost(I, VF).first; | ||||||||
6233 | |||||||||
6234 | // Compute the cost of the scalarized instruction. This cost is the cost of | ||||||||
6235 | // the instruction as if it wasn't if-converted and instead remained in the | ||||||||
6236 | // predicated block. We will scale this cost by block probability after | ||||||||
6237 | // computing the scalarization overhead. | ||||||||
6238 | InstructionCost ScalarCost = | ||||||||
6239 | VF.getFixedValue() * | ||||||||
6240 | getInstructionCost(I, ElementCount::getFixed(1)).first; | ||||||||
6241 | |||||||||
6242 | // Compute the scalarization overhead of needed insertelement instructions | ||||||||
6243 | // and phi nodes. | ||||||||
6244 | TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput; | ||||||||
6245 | if (isScalarWithPredication(I, VF) && !I->getType()->isVoidTy()) { | ||||||||
6246 | ScalarCost += TTI.getScalarizationOverhead( | ||||||||
6247 | cast<VectorType>(ToVectorTy(I->getType(), VF)), | ||||||||
6248 | APInt::getAllOnes(VF.getFixedValue()), /*Insert*/ true, | ||||||||
6249 | /*Extract*/ false, CostKind); | ||||||||
6250 | ScalarCost += | ||||||||
6251 | VF.getFixedValue() * TTI.getCFInstrCost(Instruction::PHI, CostKind); | ||||||||
6252 | } | ||||||||
6253 | |||||||||
6254 | // Compute the scalarization overhead of needed extractelement | ||||||||
6255 | // instructions. For each of the instruction's operands, if the operand can | ||||||||
6256 | // be scalarized, add it to the worklist; otherwise, account for the | ||||||||
6257 | // overhead. | ||||||||
6258 | for (Use &U : I->operands()) | ||||||||
6259 | if (auto *J = dyn_cast<Instruction>(U.get())) { | ||||||||
6260 | 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", 6261, __extension__ __PRETTY_FUNCTION__)) | ||||||||
6261 | "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", 6261, __extension__ __PRETTY_FUNCTION__)); | ||||||||
6262 | if (canBeScalarized(J)) | ||||||||
6263 | Worklist.push_back(J); | ||||||||
6264 | else if (needsExtract(J, VF)) { | ||||||||
6265 | ScalarCost += TTI.getScalarizationOverhead( | ||||||||
6266 | cast<VectorType>(ToVectorTy(J->getType(), VF)), | ||||||||
6267 | APInt::getAllOnes(VF.getFixedValue()), /*Insert*/ false, | ||||||||
6268 | /*Extract*/ true, CostKind); | ||||||||
6269 | } | ||||||||
6270 | } | ||||||||
6271 | |||||||||
6272 | // Scale the total scalar cost by block probability. | ||||||||
6273 | ScalarCost /= getReciprocalPredBlockProb(); | ||||||||
6274 | |||||||||
6275 | // Compute the discount. A non-negative discount means the vector version | ||||||||
6276 | // of the instruction costs more, and scalarizing would be beneficial. | ||||||||
6277 | Discount += VectorCost - ScalarCost; | ||||||||
6278 | ScalarCosts[I] = ScalarCost; | ||||||||
6279 | } | ||||||||
6280 | |||||||||
6281 | return Discount; | ||||||||
6282 | } | ||||||||
6283 | |||||||||
6284 | LoopVectorizationCostModel::VectorizationCostTy | ||||||||
6285 | LoopVectorizationCostModel::expectedCost( | ||||||||
6286 | ElementCount VF, SmallVectorImpl<InstructionVFPair> *Invalid) { | ||||||||
6287 | VectorizationCostTy Cost; | ||||||||
6288 | |||||||||
6289 | // For each block. | ||||||||
6290 | for (BasicBlock *BB : TheLoop->blocks()) { | ||||||||
6291 | VectorizationCostTy BlockCost; | ||||||||
6292 | |||||||||
6293 | // For each instruction in the old loop. | ||||||||
6294 | for (Instruction &I : BB->instructionsWithoutDebug()) { | ||||||||
6295 | // Skip ignored values. | ||||||||
6296 | if (ValuesToIgnore.count(&I) || | ||||||||
6297 | (VF.isVector() && VecValuesToIgnore.count(&I))) | ||||||||
6298 | continue; | ||||||||
6299 | |||||||||
6300 | VectorizationCostTy C = getInstructionCost(&I, VF); | ||||||||
6301 | |||||||||
6302 | // Check if we should override the cost. | ||||||||
6303 | if (C.first.isValid() && | ||||||||
6304 | ForceTargetInstructionCost.getNumOccurrences() > 0) | ||||||||
6305 | C.first = InstructionCost(ForceTargetInstructionCost); | ||||||||
6306 | |||||||||
6307 | // Keep a list of instructions with invalid costs. | ||||||||
6308 | if (Invalid && !C.first.isValid()) | ||||||||
6309 | Invalid->emplace_back(&I, VF); | ||||||||
6310 | |||||||||
6311 | BlockCost.first += C.first; | ||||||||
6312 | BlockCost.second |= C.second; | ||||||||
6313 | 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) | ||||||||
6314 | << " 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) | ||||||||
6315 | << '\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); | ||||||||
6316 | } | ||||||||
6317 | |||||||||
6318 | // If we are vectorizing a predicated block, it will have been | ||||||||
6319 | // if-converted. This means that the block's instructions (aside from | ||||||||
6320 | // stores and instructions that may divide by zero) will now be | ||||||||
6321 | // unconditionally executed. For the scalar case, we may not always execute | ||||||||
6322 | // the predicated block, if it is an if-else block. Thus, scale the block's | ||||||||
6323 | // cost by the probability of executing it. blockNeedsPredication from | ||||||||
6324 | // Legal is used so as to not include all blocks in tail folded loops. | ||||||||
6325 | if (VF.isScalar() && Legal->blockNeedsPredication(BB)) | ||||||||
6326 | BlockCost.first /= getReciprocalPredBlockProb(); | ||||||||
6327 | |||||||||
6328 | Cost.first += BlockCost.first; | ||||||||
6329 | Cost.second |= BlockCost.second; | ||||||||
6330 | } | ||||||||
6331 | |||||||||
6332 | return Cost; | ||||||||
6333 | } | ||||||||
6334 | |||||||||
6335 | /// Gets Address Access SCEV after verifying that the access pattern | ||||||||
6336 | /// is loop invariant except the induction variable dependence. | ||||||||
6337 | /// | ||||||||
6338 | /// This SCEV can be sent to the Target in order to estimate the address | ||||||||
6339 | /// calculation cost. | ||||||||
6340 | static const SCEV *getAddressAccessSCEV( | ||||||||
6341 | Value *Ptr, | ||||||||
6342 | LoopVectorizationLegality *Legal, | ||||||||
6343 | PredicatedScalarEvolution &PSE, | ||||||||
6344 | const Loop *TheLoop) { | ||||||||
6345 | |||||||||
6346 | auto *Gep = dyn_cast<GetElementPtrInst>(Ptr); | ||||||||
6347 | if (!Gep) | ||||||||
6348 | return nullptr; | ||||||||
6349 | |||||||||
6350 | // We are looking for a gep with all loop invariant indices except for one | ||||||||
6351 | // which should be an induction variable. | ||||||||
6352 | auto SE = PSE.getSE(); | ||||||||
6353 | unsigned NumOperands = Gep->getNumOperands(); | ||||||||
6354 | for (unsigned i = 1; i < NumOperands; ++i) { | ||||||||
6355 | Value *Opd = Gep->getOperand(i); | ||||||||
6356 | if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) && | ||||||||
6357 | !Legal->isInductionVariable(Opd)) | ||||||||
6358 | return nullptr; | ||||||||
6359 | } | ||||||||
6360 | |||||||||
6361 | // Now we know we have a GEP ptr, %inv, %ind, %inv. return the Ptr SCEV. | ||||||||
6362 | return PSE.getSCEV(Ptr); | ||||||||
6363 | } | ||||||||
6364 | |||||||||
6365 | static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) { | ||||||||
6366 | return Legal->hasStride(I->getOperand(0)) || | ||||||||
6367 | Legal->hasStride(I->getOperand(1)); | ||||||||
6368 | } | ||||||||
6369 | |||||||||
6370 | InstructionCost | ||||||||
6371 | LoopVectorizationCostModel::getMemInstScalarizationCost(Instruction *I, | ||||||||
6372 | ElementCount VF) { | ||||||||
6373 | 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", 6374, __extension__ __PRETTY_FUNCTION__)) | ||||||||
6374 | "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", 6374, __extension__ __PRETTY_FUNCTION__)); | ||||||||
6375 | if (VF.isScalable()) | ||||||||
6376 | return InstructionCost::getInvalid(); | ||||||||
6377 | |||||||||
6378 | Type *ValTy = getLoadStoreType(I); | ||||||||
6379 | auto SE = PSE.getSE(); | ||||||||
6380 | |||||||||
6381 | unsigned AS = getLoadStoreAddressSpace(I); | ||||||||
6382 | Value *Ptr = getLoadStorePointerOperand(I); | ||||||||
6383 | Type *PtrTy = ToVectorTy(Ptr->getType(), VF); | ||||||||
6384 | // NOTE: PtrTy is a vector to signal `TTI::getAddressComputationCost` | ||||||||
6385 | // that it is being called from this specific place. | ||||||||
6386 | |||||||||
6387 | // Figure out whether the access is strided and get the stride value | ||||||||
6388 | // if it's known in compile time | ||||||||
6389 | const SCEV *PtrSCEV = getAddressAccessSCEV(Ptr, Legal, PSE, TheLoop); | ||||||||
6390 | |||||||||
6391 | // Get the cost of the scalar memory instruction and address computation. | ||||||||
6392 | InstructionCost Cost = | ||||||||
6393 | VF.getKnownMinValue() * TTI.getAddressComputationCost(PtrTy, SE, PtrSCEV); | ||||||||
6394 | |||||||||
6395 | // Don't pass *I here, since it is scalar but will actually be part of a | ||||||||
6396 | // vectorized loop where the user of it is a vectorized instruction. | ||||||||
6397 | TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput; | ||||||||
6398 | const Align Alignment = getLoadStoreAlignment(I); | ||||||||
6399 | Cost += VF.getKnownMinValue() * TTI.getMemoryOpCost(I->getOpcode(), | ||||||||
6400 | ValTy->getScalarType(), | ||||||||
6401 | Alignment, AS, CostKind); | ||||||||
6402 | |||||||||
6403 | // Get the overhead of the extractelement and insertelement instructions | ||||||||
6404 | // we might create due to scalarization. | ||||||||
6405 | Cost += getScalarizationOverhead(I, VF, CostKind); | ||||||||
6406 | |||||||||
6407 | // If we have a predicated load/store, it will need extra i1 extracts and | ||||||||
6408 | // conditional branches, but may not be executed for each vector lane. Scale | ||||||||
6409 | // the cost by the probability of executing the predicated block. | ||||||||
6410 | if (isPredicatedInst(I)) { | ||||||||
6411 | Cost /= getReciprocalPredBlockProb(); | ||||||||
6412 | |||||||||
6413 | // Add the cost of an i1 extract and a branch | ||||||||
6414 | auto *Vec_i1Ty = | ||||||||
6415 | VectorType::get(IntegerType::getInt1Ty(ValTy->getContext()), VF); | ||||||||
6416 | Cost += TTI.getScalarizationOverhead( | ||||||||
6417 | Vec_i1Ty, APInt::getAllOnes(VF.getKnownMinValue()), | ||||||||
6418 | /*Insert=*/false, /*Extract=*/true, CostKind); | ||||||||
6419 | Cost += TTI.getCFInstrCost(Instruction::Br, CostKind); | ||||||||
6420 | |||||||||
6421 | if (useEmulatedMaskMemRefHack(I, VF)) | ||||||||
6422 | // Artificially setting to a high enough value to practically disable | ||||||||
6423 | // vectorization with such operations. | ||||||||
6424 | Cost = 3000000; | ||||||||
6425 | } | ||||||||
6426 | |||||||||
6427 | return Cost; | ||||||||
6428 | } | ||||||||
6429 | |||||||||
6430 | InstructionCost | ||||||||
6431 | LoopVectorizationCostModel::getConsecutiveMemOpCost(Instruction *I, | ||||||||
6432 | ElementCount VF) { | ||||||||
6433 | Type *ValTy = getLoadStoreType(I); | ||||||||
6434 | auto *VectorTy = cast<VectorType>(ToVectorTy(ValTy, VF)); | ||||||||
6435 | Value *Ptr = getLoadStorePointerOperand(I); | ||||||||
6436 | unsigned AS = getLoadStoreAddressSpace(I); | ||||||||
6437 | int ConsecutiveStride = Legal->isConsecutivePtr(ValTy, Ptr); | ||||||||
6438 | enum TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput; | ||||||||
6439 | |||||||||
6440 | 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", 6441, __extension__ __PRETTY_FUNCTION__)) | ||||||||
6441 | "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", 6441, __extension__ __PRETTY_FUNCTION__)); | ||||||||
6442 | const Align Alignment = getLoadStoreAlignment(I); | ||||||||
6443 | InstructionCost Cost = 0; | ||||||||
6444 | if (Legal->isMaskRequired(I)) { | ||||||||
6445 | Cost += TTI.getMaskedMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS, | ||||||||
6446 | CostKind); | ||||||||
6447 | } else { | ||||||||
6448 | TTI::OperandValueInfo OpInfo = TTI::getOperandInfo(I->getOperand(0)); | ||||||||
6449 | Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS, | ||||||||
6450 | CostKind, OpInfo, I); | ||||||||
6451 | } | ||||||||
6452 | |||||||||
6453 | bool Reverse = ConsecutiveStride < 0; | ||||||||
6454 | if (Reverse) | ||||||||
6455 | Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, VectorTy, | ||||||||
6456 | std::nullopt, CostKind, 0); | ||||||||
6457 | return Cost; | ||||||||
6458 | } | ||||||||
6459 | |||||||||
6460 | InstructionCost | ||||||||
6461 | LoopVectorizationCostModel::getUniformMemOpCost(Instruction *I, | ||||||||
6462 | ElementCount VF) { | ||||||||
6463 | assert(Legal->isUniformMemOp(*I))(static_cast <bool> (Legal->isUniformMemOp(*I)) ? void (0) : __assert_fail ("Legal->isUniformMemOp(*I)", "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 6463, __extension__ __PRETTY_FUNCTION__)); | ||||||||
6464 | |||||||||
6465 | Type *ValTy = getLoadStoreType(I); | ||||||||
6466 | auto *VectorTy = cast<VectorType>(ToVectorTy(ValTy, VF)); | ||||||||
6467 | const Align Alignment = getLoadStoreAlignment(I); | ||||||||
6468 | unsigned AS = getLoadStoreAddressSpace(I); | ||||||||
6469 | enum TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput; | ||||||||
6470 | if (isa<LoadInst>(I)) { | ||||||||
6471 | return TTI.getAddressComputationCost(ValTy) + | ||||||||
6472 | TTI.getMemoryOpCost(Instruction::Load, ValTy, Alignment, AS, | ||||||||
6473 | CostKind) + | ||||||||
6474 | TTI.getShuffleCost(TargetTransformInfo::SK_Broadcast, VectorTy); | ||||||||
6475 | } | ||||||||
6476 | StoreInst *SI = cast<StoreInst>(I); | ||||||||
6477 | |||||||||
6478 | bool isLoopInvariantStoreValue = Legal->isUniform(SI->getValueOperand()); | ||||||||
6479 | return TTI.getAddressComputationCost(ValTy) + | ||||||||
6480 | TTI.getMemoryOpCost(Instruction::Store, ValTy, Alignment, AS, | ||||||||
6481 | CostKind) + | ||||||||
6482 | (isLoopInvariantStoreValue | ||||||||
6483 | ? 0 | ||||||||
6484 | : TTI.getVectorInstrCost(Instruction::ExtractElement, VectorTy, | ||||||||
6485 | CostKind, VF.getKnownMinValue() - 1)); | ||||||||
6486 | } | ||||||||
6487 | |||||||||
6488 | InstructionCost | ||||||||
6489 | LoopVectorizationCostModel::getGatherScatterCost(Instruction *I, | ||||||||
6490 | ElementCount VF) { | ||||||||
6491 | Type *ValTy = getLoadStoreType(I); | ||||||||
6492 | auto *VectorTy = cast<VectorType>(ToVectorTy(ValTy, VF)); | ||||||||
6493 | const Align Alignment = getLoadStoreAlignment(I); | ||||||||
6494 | const Value *Ptr = getLoadStorePointerOperand(I); | ||||||||
6495 | |||||||||
6496 | return TTI.getAddressComputationCost(VectorTy) + | ||||||||
6497 | TTI.getGatherScatterOpCost( | ||||||||
6498 | I->getOpcode(), VectorTy, Ptr, Legal->isMaskRequired(I), Alignment, | ||||||||
6499 | TargetTransformInfo::TCK_RecipThroughput, I); | ||||||||
6500 | } | ||||||||
6501 | |||||||||
6502 | InstructionCost | ||||||||
6503 | LoopVectorizationCostModel::getInterleaveGroupCost(Instruction *I, | ||||||||
6504 | ElementCount VF) { | ||||||||
6505 | // TODO: Once we have support for interleaving with scalable vectors | ||||||||
6506 | // we can calculate the cost properly here. | ||||||||
6507 | if (VF.isScalable()) | ||||||||
6508 | return InstructionCost::getInvalid(); | ||||||||
6509 | |||||||||
6510 | Type *ValTy = getLoadStoreType(I); | ||||||||
6511 | auto *VectorTy = cast<VectorType>(ToVectorTy(ValTy, VF)); | ||||||||
6512 | unsigned AS = getLoadStoreAddressSpace(I); | ||||||||
6513 | enum TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput; | ||||||||
6514 | |||||||||
6515 | auto Group = getInterleavedAccessGroup(I); | ||||||||
6516 | 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", 6516, __extension__ __PRETTY_FUNCTION__)); | ||||||||
6517 | |||||||||
6518 | unsigned InterleaveFactor = Group->getFactor(); | ||||||||
6519 | auto *WideVecTy = VectorType::get(ValTy, VF * InterleaveFactor); | ||||||||
6520 | |||||||||
6521 | // Holds the indices of existing members in the interleaved group. | ||||||||
6522 | SmallVector<unsigned, 4> Indices; | ||||||||
6523 | for (unsigned IF = 0; IF < InterleaveFactor; IF++) | ||||||||
6524 | if (Group->getMember(IF)) | ||||||||
6525 | Indices.push_back(IF); | ||||||||
6526 | |||||||||
6527 | // Calculate the cost of the whole interleaved group. | ||||||||
6528 | bool UseMaskForGaps = | ||||||||
6529 | (Group->requiresScalarEpilogue() && !isScalarEpilogueAllowed()) || | ||||||||
6530 | (isa<StoreInst>(I) && (Group->getNumMembers() < Group->getFactor())); | ||||||||
6531 | InstructionCost Cost = TTI.getInterleavedMemoryOpCost( | ||||||||
6532 | I->getOpcode(), WideVecTy, Group->getFactor(), Indices, Group->getAlign(), | ||||||||
6533 | AS, CostKind, Legal->isMaskRequired(I), UseMaskForGaps); | ||||||||
6534 | |||||||||
6535 | if (Group->isReverse()) { | ||||||||
6536 | // TODO: Add support for reversed masked interleaved access. | ||||||||
6537 | 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", 6538, __extension__ __PRETTY_FUNCTION__)) | ||||||||
6538 | "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", 6538, __extension__ __PRETTY_FUNCTION__)); | ||||||||
6539 | Cost += Group->getNumMembers() * | ||||||||
6540 | TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, VectorTy, | ||||||||
6541 | std::nullopt, CostKind, 0); | ||||||||
6542 | } | ||||||||
6543 | return Cost; | ||||||||
6544 | } | ||||||||
6545 | |||||||||
6546 | std::optional<InstructionCost> | ||||||||
6547 | LoopVectorizationCostModel::getReductionPatternCost( | ||||||||
6548 | Instruction *I, ElementCount VF, Type *Ty, TTI::TargetCostKind CostKind) { | ||||||||
6549 | using namespace llvm::PatternMatch; | ||||||||
6550 | // Early exit for no inloop reductions | ||||||||
6551 | if (InLoopReductionChains.empty() || VF.isScalar() || !isa<VectorType>(Ty)) | ||||||||
6552 | return std::nullopt; | ||||||||
6553 | auto *VectorTy = cast<VectorType>(Ty); | ||||||||
6554 | |||||||||
6555 | // We are looking for a pattern of, and finding the minimal acceptable cost: | ||||||||
6556 | // reduce(mul(ext(A), ext(B))) or | ||||||||
6557 | // reduce(mul(A, B)) or | ||||||||
6558 | // reduce(ext(A)) or | ||||||||
6559 | // reduce(A). | ||||||||
6560 | // The basic idea is that we walk down the tree to do that, finding the root | ||||||||
6561 | // reduction instruction in InLoopReductionImmediateChains. From there we find | ||||||||
6562 | // the pattern of mul/ext and test the cost of the entire pattern vs the cost | ||||||||
6563 | // of the components. If the reduction cost is lower then we return it for the | ||||||||
6564 | // reduction instruction and 0 for the other instructions in the pattern. If | ||||||||
6565 | // it is not we return an invalid cost specifying the orignal cost method | ||||||||
6566 | // should be used. | ||||||||
6567 | Instruction *RetI = I; | ||||||||
6568 | if (match(RetI, m_ZExtOrSExt(m_Value()))) { | ||||||||
6569 | if (!RetI->hasOneUser()) | ||||||||
6570 | return std::nullopt; | ||||||||
6571 | RetI = RetI->user_back(); | ||||||||
6572 | } | ||||||||
6573 | |||||||||
6574 | if (match(RetI, m_OneUse(m_Mul(m_Value(), m_Value()))) && | ||||||||
6575 | RetI->user_back()->getOpcode() == Instruction::Add) { | ||||||||
6576 | RetI = RetI->user_back(); | ||||||||
6577 | } | ||||||||
6578 | |||||||||
6579 | // Test if the found instruction is a reduction, and if not return an invalid | ||||||||
6580 | // cost specifying the parent to use the original cost modelling. | ||||||||
6581 | if (!InLoopReductionImmediateChains.count(RetI)) | ||||||||
6582 | return std::nullopt; | ||||||||
6583 | |||||||||
6584 | // Find the reduction this chain is a part of and calculate the basic cost of | ||||||||
6585 | // the reduction on its own. | ||||||||
6586 | Instruction *LastChain = InLoopReductionImmediateChains[RetI]; | ||||||||
6587 | Instruction *ReductionPhi = LastChain; | ||||||||
6588 | while (!isa<PHINode>(ReductionPhi)) | ||||||||
6589 | ReductionPhi = InLoopReductionImmediateChains[ReductionPhi]; | ||||||||
6590 | |||||||||
6591 | const RecurrenceDescriptor &RdxDesc = | ||||||||
6592 | Legal->getReductionVars().find(cast<PHINode>(ReductionPhi))->second; | ||||||||
6593 | |||||||||
6594 | InstructionCost BaseCost = TTI.getArithmeticReductionCost( | ||||||||
6595 | RdxDesc.getOpcode(), VectorTy, RdxDesc.getFastMathFlags(), CostKind); | ||||||||
6596 | |||||||||
6597 | // For a call to the llvm.fmuladd intrinsic we need to add the cost of a | ||||||||
6598 | // normal fmul instruction to the cost of the fadd reduction. | ||||||||
6599 | if (RdxDesc.getRecurrenceKind() == RecurKind::FMulAdd) | ||||||||
6600 | BaseCost += | ||||||||
6601 | TTI.getArithmeticInstrCost(Instruction::FMul, VectorTy, CostKind); | ||||||||
6602 | |||||||||
6603 | // If we're using ordered reductions then we can just return the base cost | ||||||||
6604 | // here, since getArithmeticReductionCost calculates the full ordered | ||||||||
6605 | // reduction cost when FP reassociation is not allowed. | ||||||||
6606 | if (useOrderedReductions(RdxDesc)) | ||||||||
6607 | return BaseCost; | ||||||||
6608 | |||||||||
6609 | // Get the operand that was not the reduction chain and match it to one of the | ||||||||
6610 | // patterns, returning the better cost if it is found. | ||||||||
6611 | Instruction *RedOp = RetI->getOperand(1) == LastChain | ||||||||
6612 | ? dyn_cast<Instruction>(RetI->getOperand(0)) | ||||||||
6613 | : dyn_cast<Instruction>(RetI->getOperand(1)); | ||||||||
6614 | |||||||||
6615 | VectorTy = VectorType::get(I->getOperand(0)->getType(), VectorTy); | ||||||||
6616 | |||||||||
6617 | Instruction *Op0, *Op1; | ||||||||
6618 | if (RedOp && RdxDesc.getOpcode() == Instruction::Add && | ||||||||
6619 | match(RedOp, | ||||||||
6620 | m_ZExtOrSExt(m_Mul(m_Instruction(Op0), m_Instruction(Op1)))) && | ||||||||
6621 | match(Op0, m_ZExtOrSExt(m_Value())) && | ||||||||
6622 | Op0->getOpcode() == Op1->getOpcode() && | ||||||||
6623 | Op0->getOperand(0)->getType() == Op1->getOperand(0)->getType() && | ||||||||
6624 | !TheLoop->isLoopInvariant(Op0) && !TheLoop->isLoopInvariant(Op1) && | ||||||||
6625 | (Op0->getOpcode() == RedOp->getOpcode() || Op0 == Op1)) { | ||||||||
6626 | |||||||||
6627 | // Matched reduce.add(ext(mul(ext(A), ext(B))) | ||||||||
6628 | // Note that the extend opcodes need to all match, or if A==B they will have | ||||||||
6629 | // been converted to zext(mul(sext(A), sext(A))) as it is known positive, | ||||||||
6630 | // which is equally fine. | ||||||||
6631 | bool IsUnsigned = isa<ZExtInst>(Op0); | ||||||||
6632 | auto *ExtType = VectorType::get(Op0->getOperand(0)->getType(), VectorTy); | ||||||||
6633 | auto *MulType = VectorType::get(Op0->getType(), VectorTy); | ||||||||
6634 | |||||||||
6635 | InstructionCost ExtCost = | ||||||||
6636 | TTI.getCastInstrCost(Op0->getOpcode(), MulType, ExtType, | ||||||||
6637 | TTI::CastContextHint::None, CostKind, Op0); | ||||||||
6638 | InstructionCost MulCost = | ||||||||
6639 | TTI.getArithmeticInstrCost(Instruction::Mul, MulType, CostKind); | ||||||||
6640 | InstructionCost Ext2Cost = | ||||||||
6641 | TTI.getCastInstrCost(RedOp->getOpcode(), VectorTy, MulType, | ||||||||
6642 | TTI::CastContextHint::None, CostKind, RedOp); | ||||||||
6643 | |||||||||
6644 | InstructionCost RedCost = TTI.getMulAccReductionCost( | ||||||||
6645 | IsUnsigned, RdxDesc.getRecurrenceType(), ExtType, CostKind); | ||||||||
6646 | |||||||||
6647 | if (RedCost.isValid() && | ||||||||
6648 | RedCost < ExtCost * 2 + MulCost + Ext2Cost + BaseCost) | ||||||||
6649 | return I == RetI ? RedCost : 0; | ||||||||
6650 | } else if (RedOp && match(RedOp, m_ZExtOrSExt(m_Value())) && | ||||||||
6651 | !TheLoop->isLoopInvariant(RedOp)) { | ||||||||
6652 | // Matched reduce(ext(A)) | ||||||||
6653 | bool IsUnsigned = isa<ZExtInst>(RedOp); | ||||||||
6654 | auto *ExtType = VectorType::get(RedOp->getOperand(0)->getType(), VectorTy); | ||||||||
6655 | InstructionCost RedCost = TTI.getExtendedReductionCost( | ||||||||
6656 | RdxDesc.getOpcode(), IsUnsigned, RdxDesc.getRecurrenceType(), ExtType, | ||||||||
6657 | RdxDesc.getFastMathFlags(), CostKind); | ||||||||
6658 | |||||||||
6659 | InstructionCost ExtCost = | ||||||||
6660 | TTI.getCastInstrCost(RedOp->getOpcode(), VectorTy, ExtType, | ||||||||
6661 | TTI::CastContextHint::None, CostKind, RedOp); | ||||||||
6662 | if (RedCost.isValid() && RedCost < BaseCost + ExtCost) | ||||||||
6663 | return I == RetI ? RedCost : 0; | ||||||||
6664 | } else if (RedOp && RdxDesc.getOpcode() == Instruction::Add && | ||||||||
6665 | match(RedOp, m_Mul(m_Instruction(Op0), m_Instruction(Op1)))) { | ||||||||
6666 | if (match(Op0, m_ZExtOrSExt(m_Value())) && | ||||||||
6667 | Op0->getOpcode() == Op1->getOpcode() && | ||||||||
6668 | !TheLoop->isLoopInvariant(Op0) && !TheLoop->isLoopInvariant(Op1)) { | ||||||||
6669 | bool IsUnsigned = isa<ZExtInst>(Op0); | ||||||||
6670 | Type *Op0Ty = Op0->getOperand(0)->getType(); | ||||||||
6671 | Type *Op1Ty = Op1->getOperand(0)->getType(); | ||||||||
6672 | Type *LargestOpTy = | ||||||||
6673 | Op0Ty->getIntegerBitWidth() < Op1Ty->getIntegerBitWidth() ? Op1Ty | ||||||||
6674 | : Op0Ty; | ||||||||
6675 | auto *ExtType = VectorType::get(LargestOpTy, VectorTy); | ||||||||
6676 | |||||||||
6677 | // Matched reduce.add(mul(ext(A), ext(B))), where the two ext may be of | ||||||||
6678 | // different sizes. We take the largest type as the ext to reduce, and add | ||||||||
6679 | // the remaining cost as, for example reduce(mul(ext(ext(A)), ext(B))). | ||||||||
6680 | InstructionCost ExtCost0 = TTI.getCastInstrCost( | ||||||||
6681 | Op0->getOpcode(), VectorTy, VectorType::get(Op0Ty, VectorTy), | ||||||||
6682 | TTI::CastContextHint::None, CostKind, Op0); | ||||||||
6683 | InstructionCost ExtCost1 = TTI.getCastInstrCost( | ||||||||
6684 | Op1->getOpcode(), VectorTy, VectorType::get(Op1Ty, VectorTy), | ||||||||
6685 | TTI::CastContextHint::None, CostKind, Op1); | ||||||||
6686 | InstructionCost MulCost = | ||||||||
6687 | TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy, CostKind); | ||||||||
6688 | |||||||||
6689 | InstructionCost RedCost = TTI.getMulAccReductionCost( | ||||||||
6690 | IsUnsigned, RdxDesc.getRecurrenceType(), ExtType, CostKind); | ||||||||
6691 | InstructionCost ExtraExtCost = 0; | ||||||||
6692 | if (Op0Ty != LargestOpTy || Op1Ty != LargestOpTy) { | ||||||||
6693 | Instruction *ExtraExtOp = (Op0Ty != LargestOpTy) ? Op0 : Op1; | ||||||||
6694 | ExtraExtCost = TTI.getCastInstrCost( | ||||||||
6695 | ExtraExtOp->getOpcode(), ExtType, | ||||||||
6696 | VectorType::get(ExtraExtOp->getOperand(0)->getType(), VectorTy), | ||||||||
6697 | TTI::CastContextHint::None, CostKind, ExtraExtOp); | ||||||||
6698 | } | ||||||||
6699 | |||||||||
6700 | if (RedCost.isValid() && | ||||||||
6701 | (RedCost + ExtraExtCost) < (ExtCost0 + ExtCost1 + MulCost + BaseCost)) | ||||||||
6702 | return I == RetI ? RedCost : 0; | ||||||||
6703 | } else if (!match(I, m_ZExtOrSExt(m_Value()))) { | ||||||||
6704 | // Matched reduce.add(mul()) | ||||||||
6705 | InstructionCost MulCost = | ||||||||
6706 | TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy, CostKind); | ||||||||
6707 | |||||||||
6708 | InstructionCost RedCost = TTI.getMulAccReductionCost( | ||||||||
6709 | true, RdxDesc.getRecurrenceType(), VectorTy, CostKind); | ||||||||
6710 | |||||||||
6711 | if (RedCost.isValid() && RedCost < MulCost + BaseCost) | ||||||||
6712 | return I == RetI ? RedCost : 0; | ||||||||
6713 | } | ||||||||
6714 | } | ||||||||
6715 | |||||||||
6716 | return I == RetI ? std::optional<InstructionCost>(BaseCost) : std::nullopt; | ||||||||
6717 | } | ||||||||
6718 | |||||||||
6719 | InstructionCost | ||||||||
6720 | LoopVectorizationCostModel::getMemoryInstructionCost(Instruction *I, | ||||||||
6721 | ElementCount VF) { | ||||||||
6722 | // Calculate scalar cost only. Vectorization cost should be ready at this | ||||||||
6723 | // moment. | ||||||||
6724 | if (VF.isScalar()) { | ||||||||
6725 | Type *ValTy = getLoadStoreType(I); | ||||||||
6726 | const Align Alignment = getLoadStoreAlignment(I); | ||||||||
6727 | unsigned AS = getLoadStoreAddressSpace(I); | ||||||||
6728 | |||||||||
6729 | TTI::OperandValueInfo OpInfo = TTI::getOperandInfo(I->getOperand(0)); | ||||||||
6730 | return TTI.getAddressComputationCost(ValTy) + | ||||||||
6731 | TTI.getMemoryOpCost(I->getOpcode(), ValTy, Alignment, AS, | ||||||||
6732 | TTI::TCK_RecipThroughput, OpInfo, I); | ||||||||
6733 | } | ||||||||
6734 | return getWideningCost(I, VF); | ||||||||
6735 | } | ||||||||
6736 | |||||||||
6737 | LoopVectorizationCostModel::VectorizationCostTy | ||||||||
6738 | LoopVectorizationCostModel::getInstructionCost(Instruction *I, | ||||||||
6739 | ElementCount VF) { | ||||||||
6740 | // If we know that this instruction will remain uniform, check the cost of | ||||||||
6741 | // the scalar version. | ||||||||
6742 | if (isUniformAfterVectorization(I, VF)) | ||||||||
6743 | VF = ElementCount::getFixed(1); | ||||||||
6744 | |||||||||
6745 | if (VF.isVector() && isProfitableToScalarize(I, VF)) | ||||||||
6746 | return VectorizationCostTy(InstsToScalarize[VF][I], false); | ||||||||
6747 | |||||||||
6748 | // Forced scalars do not have any scalarization overhead. | ||||||||
6749 | auto ForcedScalar = ForcedScalars.find(VF); | ||||||||
6750 | if (VF.isVector() && ForcedScalar != ForcedScalars.end()) { | ||||||||
6751 | auto InstSet = ForcedScalar->second; | ||||||||
6752 | if (InstSet.count(I)) | ||||||||
6753 | return VectorizationCostTy( | ||||||||
6754 | (getInstructionCost(I, ElementCount::getFixed(1)).first * | ||||||||
6755 | VF.getKnownMinValue()), | ||||||||
6756 | false); | ||||||||
6757 | } | ||||||||
6758 | |||||||||
6759 | Type *VectorTy; | ||||||||
6760 | InstructionCost C = getInstructionCost(I, VF, VectorTy); | ||||||||
6761 | |||||||||
6762 | bool TypeNotScalarized = false; | ||||||||
6763 | if (VF.isVector() && VectorTy->isVectorTy()) { | ||||||||
6764 | if (unsigned NumParts = TTI.getNumberOfParts(VectorTy)) { | ||||||||
6765 | if (VF.isScalable()) | ||||||||
6766 | // <vscale x 1 x iN> is assumed to be profitable over iN because | ||||||||
6767 | // scalable registers are a distinct register class from scalar ones. | ||||||||
6768 | // If we ever find a target which wants to lower scalable vectors | ||||||||
6769 | // back to scalars, we'll need to update this code to explicitly | ||||||||
6770 | // ask TTI about the register class uses for each part. | ||||||||
6771 | TypeNotScalarized = NumParts <= VF.getKnownMinValue(); | ||||||||
6772 | else | ||||||||
6773 | TypeNotScalarized = NumParts < VF.getKnownMinValue(); | ||||||||
6774 | } else | ||||||||
6775 | C = InstructionCost::getInvalid(); | ||||||||
6776 | } | ||||||||
6777 | return VectorizationCostTy(C, TypeNotScalarized); | ||||||||
6778 | } | ||||||||
6779 | |||||||||
6780 | InstructionCost LoopVectorizationCostModel::getScalarizationOverhead( | ||||||||
6781 | Instruction *I, ElementCount VF, TTI::TargetCostKind CostKind) const { | ||||||||
6782 | |||||||||
6783 | // There is no mechanism yet to create a scalable scalarization loop, | ||||||||
6784 | // so this is currently Invalid. | ||||||||
6785 | if (VF.isScalable()) | ||||||||
6786 | return InstructionCost::getInvalid(); | ||||||||
6787 | |||||||||
6788 | if (VF.isScalar()) | ||||||||
6789 | return 0; | ||||||||
6790 | |||||||||
6791 | InstructionCost Cost = 0; | ||||||||
6792 | Type *RetTy = ToVectorTy(I->getType(), VF); | ||||||||
6793 | if (!RetTy->isVoidTy() && | ||||||||
6794 | (!isa<LoadInst>(I) || !TTI.supportsEfficientVectorElementLoadStore())) | ||||||||
6795 | Cost += TTI.getScalarizationOverhead( | ||||||||
6796 | cast<VectorType>(RetTy), APInt::getAllOnes(VF.getKnownMinValue()), | ||||||||
6797 | /*Insert*/ true, | ||||||||
6798 | /*Extract*/ false, CostKind); | ||||||||
6799 | |||||||||
6800 | // Some targets keep addresses scalar. | ||||||||
6801 | if (isa<LoadInst>(I) && !TTI.prefersVectorizedAddressing()) | ||||||||
6802 | return Cost; | ||||||||
6803 | |||||||||
6804 | // Some targets support efficient element stores. | ||||||||
6805 | if (isa<StoreInst>(I) && TTI.supportsEfficientVectorElementLoadStore()) | ||||||||
6806 | return Cost; | ||||||||
6807 | |||||||||
6808 | // Collect operands to consider. | ||||||||
6809 | CallInst *CI = dyn_cast<CallInst>(I); | ||||||||
6810 | Instruction::op_range Ops = CI ? CI->args() : I->operands(); | ||||||||
6811 | |||||||||
6812 | // Skip operands that do not require extraction/scalarization and do not incur | ||||||||
6813 | // any overhead. | ||||||||
6814 | SmallVector<Type *> Tys; | ||||||||
6815 | for (auto *V : filterExtractingOperands(Ops, VF)) | ||||||||
6816 | Tys.push_back(MaybeVectorizeType(V->getType(), VF)); | ||||||||
6817 | return Cost + TTI.getOperandsScalarizationOverhead( | ||||||||
6818 | filterExtractingOperands(Ops, VF), Tys, CostKind); | ||||||||
6819 | } | ||||||||
6820 | |||||||||
6821 | void LoopVectorizationCostModel::setCostBasedWideningDecision(ElementCount VF) { | ||||||||
6822 | if (VF.isScalar()) | ||||||||
6823 | return; | ||||||||
6824 | NumPredStores = 0; | ||||||||
6825 | for (BasicBlock *BB : TheLoop->blocks()) { | ||||||||
6826 | // For each instruction in the old loop. | ||||||||
6827 | for (Instruction &I : *BB) { | ||||||||
6828 | Value *Ptr = getLoadStorePointerOperand(&I); | ||||||||
6829 | if (!Ptr) | ||||||||
6830 | continue; | ||||||||
6831 | |||||||||
6832 | // TODO: We should generate better code and update the cost model for | ||||||||
6833 | // predicated uniform stores. Today they are treated as any other | ||||||||
6834 | // predicated store (see added test cases in | ||||||||
6835 | // invariant-store-vectorization.ll). | ||||||||
6836 | if (isa<StoreInst>(&I) && isScalarWithPredication(&I, VF)) | ||||||||
6837 | NumPredStores++; | ||||||||
6838 | |||||||||
6839 | if (Legal->isUniformMemOp(I)) { | ||||||||
6840 | auto isLegalToScalarize = [&]() { | ||||||||
6841 | if (!VF.isScalable()) | ||||||||
6842 | // Scalarization of fixed length vectors "just works". | ||||||||
6843 | return true; | ||||||||
6844 | |||||||||
6845 | // We have dedicated lowering for unpredicated uniform loads and | ||||||||
6846 | // stores. Note that even with tail folding we know that at least | ||||||||
6847 | // one lane is active (i.e. generalized predication is not possible | ||||||||
6848 | // here), and the logic below depends on this fact. | ||||||||
6849 | if (!foldTailByMasking()) | ||||||||
6850 | return true; | ||||||||
6851 | |||||||||
6852 | // For scalable vectors, a uniform memop load is always | ||||||||
6853 | // uniform-by-parts and we know how to scalarize that. | ||||||||
6854 | if (isa<LoadInst>(I)) | ||||||||
6855 | return true; | ||||||||
6856 | |||||||||
6857 | // A uniform store isn't neccessarily uniform-by-part | ||||||||
6858 | // and we can't assume scalarization. | ||||||||
6859 | auto &SI = cast<StoreInst>(I); | ||||||||
6860 | return TheLoop->isLoopInvariant(SI.getValueOperand()); | ||||||||
6861 | }; | ||||||||
6862 | |||||||||
6863 | const InstructionCost GatherScatterCost = | ||||||||
6864 | isLegalGatherOrScatter(&I, VF) ? | ||||||||
6865 | getGatherScatterCost(&I, VF) : InstructionCost::getInvalid(); | ||||||||
6866 | |||||||||
6867 | // Load: Scalar load + broadcast | ||||||||
6868 | // Store: Scalar store + isLoopInvariantStoreValue ? 0 : extract | ||||||||
6869 | // FIXME: This cost is a significant under-estimate for tail folded | ||||||||
6870 | // memory ops. | ||||||||
6871 | const InstructionCost ScalarizationCost = isLegalToScalarize() ? | ||||||||
6872 | getUniformMemOpCost(&I, VF) : InstructionCost::getInvalid(); | ||||||||
6873 | |||||||||
6874 | // Choose better solution for the current VF, Note that Invalid | ||||||||
6875 | // costs compare as maximumal large. If both are invalid, we get | ||||||||
6876 | // scalable invalid which signals a failure and a vectorization abort. | ||||||||
6877 | if (GatherScatterCost < ScalarizationCost) | ||||||||
6878 | setWideningDecision(&I, VF, CM_GatherScatter, GatherScatterCost); | ||||||||
6879 | else | ||||||||
6880 | setWideningDecision(&I, VF, CM_Scalarize, ScalarizationCost); | ||||||||
6881 | continue; | ||||||||
6882 | } | ||||||||
6883 | |||||||||
6884 | // We assume that widening is the best solution when possible. | ||||||||
6885 | if (memoryInstructionCanBeWidened(&I, VF)) { | ||||||||
6886 | InstructionCost Cost = getConsecutiveMemOpCost(&I, VF); | ||||||||
6887 | int ConsecutiveStride = Legal->isConsecutivePtr( | ||||||||
6888 | getLoadStoreType(&I), getLoadStorePointerOperand(&I)); | ||||||||
6889 | 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", 6890, __extension__ __PRETTY_FUNCTION__)) | ||||||||
6890 | "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", 6890, __extension__ __PRETTY_FUNCTION__)); | ||||||||
6891 | InstWidening Decision = | ||||||||
6892 | ConsecutiveStride == 1 ? CM_Widen : CM_Widen_Reverse; | ||||||||
6893 | setWideningDecision(&I, VF, Decision, Cost); | ||||||||
6894 | continue; | ||||||||
6895 | } | ||||||||
6896 | |||||||||
6897 | // Choose between Interleaving, Gather/Scatter or Scalarization. | ||||||||
6898 | InstructionCost InterleaveCost = InstructionCost::getInvalid(); | ||||||||
6899 | unsigned NumAccesses = 1; | ||||||||
6900 | if (isAccessInterleaved(&I)) { | ||||||||
6901 | auto Group = getInterleavedAccessGroup(&I); | ||||||||
6902 | 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", 6902, __extension__ __PRETTY_FUNCTION__)); | ||||||||
6903 | |||||||||
6904 | // Make one decision for the whole group. | ||||||||
6905 | if (getWideningDecision(&I, VF) != CM_Unknown) | ||||||||
6906 | continue; | ||||||||
6907 | |||||||||
6908 | NumAccesses = Group->getNumMembers(); | ||||||||
6909 | if (interleavedAccessCanBeWidened(&I, VF)) | ||||||||
6910 | InterleaveCost = getInterleaveGroupCost(&I, VF); | ||||||||
6911 | } | ||||||||
6912 | |||||||||
6913 | InstructionCost GatherScatterCost = | ||||||||
6914 | isLegalGatherOrScatter(&I, VF) | ||||||||
6915 | ? getGatherScatterCost(&I, VF) * NumAccesses | ||||||||
6916 | : InstructionCost::getInvalid(); | ||||||||
6917 | |||||||||
6918 | InstructionCost ScalarizationCost = | ||||||||
6919 | getMemInstScalarizationCost(&I, VF) * NumAccesses; | ||||||||
6920 | |||||||||
6921 | // Choose better solution for the current VF, | ||||||||
6922 | // write down this decision and use it during vectorization. | ||||||||
6923 | InstructionCost Cost; | ||||||||
6924 | InstWidening Decision; | ||||||||
6925 | if (InterleaveCost <= GatherScatterCost && | ||||||||
6926 | InterleaveCost < ScalarizationCost) { | ||||||||
6927 | Decision = CM_Interleave; | ||||||||
6928 | Cost = InterleaveCost; | ||||||||
6929 | } else if (GatherScatterCost < ScalarizationCost) { | ||||||||
6930 | Decision = CM_GatherScatter; | ||||||||
6931 | Cost = GatherScatterCost; | ||||||||
6932 | } else { | ||||||||
6933 | Decision = CM_Scalarize; | ||||||||
6934 | Cost = ScalarizationCost; | ||||||||
6935 | } | ||||||||
6936 | // If the instructions belongs to an interleave group, the whole group | ||||||||
6937 | // receives the same decision. The whole group receives the cost, but | ||||||||
6938 | // the cost will actually be assigned to one instruction. | ||||||||
6939 | if (auto Group = getInterleavedAccessGroup(&I)) | ||||||||
6940 | setWideningDecision(Group, VF, Decision, Cost); | ||||||||
6941 | else | ||||||||
6942 | setWideningDecision(&I, VF, Decision, Cost); | ||||||||
6943 | } | ||||||||
6944 | } | ||||||||
6945 | |||||||||
6946 | // Make sure that any load of address and any other address computation | ||||||||
6947 | // remains scalar unless there is gather/scatter support. This avoids | ||||||||
6948 | // inevitable extracts into address registers, and also has the benefit of | ||||||||
6949 | // activating LSR more, since that pass can't optimize vectorized | ||||||||
6950 | // addresses. | ||||||||
6951 | if (TTI.prefersVectorizedAddressing()) | ||||||||
6952 | return; | ||||||||
6953 | |||||||||
6954 | // Start with all scalar pointer uses. | ||||||||
6955 | SmallPtrSet<Instruction *, 8> AddrDefs; | ||||||||
6956 | for (BasicBlock *BB : TheLoop->blocks()) | ||||||||
6957 | for (Instruction &I : *BB) { | ||||||||
6958 | Instruction *PtrDef = | ||||||||
6959 | dyn_cast_or_null<Instruction>(getLoadStorePointerOperand(&I)); | ||||||||
6960 | if (PtrDef && TheLoop->contains(PtrDef) && | ||||||||
6961 | getWideningDecision(&I, VF) != CM_GatherScatter) | ||||||||
6962 | AddrDefs.insert(PtrDef); | ||||||||
6963 | } | ||||||||
6964 | |||||||||
6965 | // Add all instructions used to generate the addresses. | ||||||||
6966 | SmallVector<Instruction *, 4> Worklist; | ||||||||
6967 | append_range(Worklist, AddrDefs); | ||||||||
6968 | while (!Worklist.empty()) { | ||||||||
6969 | Instruction *I = Worklist.pop_back_val(); | ||||||||
6970 | for (auto &Op : I->operands()) | ||||||||
6971 | if (auto *InstOp = dyn_cast<Instruction>(Op)) | ||||||||
6972 | if ((InstOp->getParent() == I->getParent()) && !isa<PHINode>(InstOp) && | ||||||||
6973 | AddrDefs.insert(InstOp).second) | ||||||||
6974 | Worklist.push_back(InstOp); | ||||||||
6975 | } | ||||||||
6976 | |||||||||
6977 | for (auto *I : AddrDefs) { | ||||||||
6978 | if (isa<LoadInst>(I)) { | ||||||||
6979 | // Setting the desired widening decision should ideally be handled in | ||||||||
6980 | // by cost functions, but since this involves the task of finding out | ||||||||
6981 | // if the loaded register is involved in an address computation, it is | ||||||||
6982 | // instead changed here when we know this is the case. | ||||||||
6983 | InstWidening Decision = getWideningDecision(I, VF); | ||||||||
6984 | if (Decision == CM_Widen || Decision == CM_Widen_Reverse) | ||||||||
6985 | // Scalarize a widened load of address. | ||||||||
6986 | setWideningDecision( | ||||||||
6987 | I, VF, CM_Scalarize, | ||||||||
6988 | (VF.getKnownMinValue() * | ||||||||
6989 | getMemoryInstructionCost(I, ElementCount::getFixed(1)))); | ||||||||
6990 | else if (auto Group = getInterleavedAccessGroup(I)) { | ||||||||
6991 | // Scalarize an interleave group of address loads. | ||||||||
6992 | for (unsigned I = 0; I < Group->getFactor(); ++I) { | ||||||||
6993 | if (Instruction *Member = Group->getMember(I)) | ||||||||
6994 | setWideningDecision( | ||||||||
6995 | Member, VF, CM_Scalarize, | ||||||||
6996 | (VF.getKnownMinValue() * | ||||||||
6997 | getMemoryInstructionCost(Member, ElementCount::getFixed(1)))); | ||||||||
6998 | } | ||||||||
6999 | } | ||||||||
7000 | } else | ||||||||
7001 | // Make sure I gets scalarized and a cost estimate without | ||||||||
7002 | // scalarization overhead. | ||||||||
7003 | ForcedScalars[VF].insert(I); | ||||||||
7004 | } | ||||||||
7005 | } | ||||||||
7006 | |||||||||
7007 | InstructionCost | ||||||||
7008 | LoopVectorizationCostModel::getInstructionCost(Instruction *I, ElementCount VF, | ||||||||
7009 | Type *&VectorTy) { | ||||||||
7010 | Type *RetTy = I->getType(); | ||||||||
7011 | if (canTruncateToMinimalBitwidth(I, VF)) | ||||||||
7012 | RetTy = IntegerType::get(RetTy->getContext(), MinBWs[I]); | ||||||||
7013 | auto SE = PSE.getSE(); | ||||||||
7014 | TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput; | ||||||||
7015 | |||||||||
7016 | auto hasSingleCopyAfterVectorization = [this](Instruction *I, | ||||||||
7017 | ElementCount VF) -> bool { | ||||||||
7018 | if (VF.isScalar()) | ||||||||
7019 | return true; | ||||||||
7020 | |||||||||
7021 | auto Scalarized = InstsToScalarize.find(VF); | ||||||||
7022 | 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", 7023, __extension__ __PRETTY_FUNCTION__)) | ||||||||
7023 | "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", 7023, __extension__ __PRETTY_FUNCTION__)); | ||||||||
7024 | return !Scalarized->second.count(I) && | ||||||||
7025 | llvm::all_of(I->users(), [&](User *U) { | ||||||||
7026 | auto *UI = cast<Instruction>(U); | ||||||||
7027 | return !Scalarized->second.count(UI); | ||||||||
7028 | }); | ||||||||
7029 | }; | ||||||||
7030 | (void) hasSingleCopyAfterVectorization; | ||||||||
7031 | |||||||||
7032 | if (isScalarAfterVectorization(I, VF)) { | ||||||||
7033 | // With the exception of GEPs and PHIs, after scalarization there should | ||||||||
7034 | // only be one copy of the instruction generated in the loop. This is | ||||||||
7035 | // because the VF is either 1, or any instructions that need scalarizing | ||||||||
7036 | // have already been dealt with by the the time we get here. As a result, | ||||||||
7037 | // it means we don't have to multiply the instruction cost by VF. | ||||||||
7038 | 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", 7042, __extension__ __PRETTY_FUNCTION__)) | ||||||||
7039 | 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", 7042, __extension__ __PRETTY_FUNCTION__)) | ||||||||
7040 | (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", 7042, __extension__ __PRETTY_FUNCTION__)) | ||||||||
7041 | 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", 7042, __extension__ __PRETTY_FUNCTION__)) | ||||||||
7042 | 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", 7042, __extension__ __PRETTY_FUNCTION__)); | ||||||||
7043 | VectorTy = RetTy; | ||||||||
7044 | } else | ||||||||
7045 | VectorTy = ToVectorTy(RetTy, VF); | ||||||||
7046 | |||||||||
7047 | // TODO: We need to estimate the cost of intrinsic calls. | ||||||||
7048 | switch (I->getOpcode()) { | ||||||||
7049 | case Instruction::GetElementPtr: | ||||||||
7050 | // We mark this instruction as zero-cost because the cost of GEPs in | ||||||||
7051 | // vectorized code depends on whether the corresponding memory instruction | ||||||||
7052 | // is scalarized or not. Therefore, we handle GEPs with the memory | ||||||||
7053 | // instruction cost. | ||||||||
7054 | return 0; | ||||||||
7055 | case Instruction::Br: { | ||||||||
7056 | // In cases of scalarized and predicated instructions, there will be VF | ||||||||
7057 | // predicated blocks in the vectorized loop. Each branch around these | ||||||||
7058 | // blocks requires also an extract of its vector compare i1 element. | ||||||||
7059 | bool ScalarPredicatedBB = false; | ||||||||
7060 | BranchInst *BI = cast<BranchInst>(I); | ||||||||
7061 | if (VF.isVector() && BI->isConditional() && | ||||||||
7062 | (PredicatedBBsAfterVectorization[VF].count(BI->getSuccessor(0)) || | ||||||||
7063 | PredicatedBBsAfterVectorization[VF].count(BI->getSuccessor(1)))) | ||||||||
7064 | ScalarPredicatedBB = true; | ||||||||
7065 | |||||||||
7066 | if (ScalarPredicatedBB) { | ||||||||
7067 | // Not possible to scalarize scalable vector with predicated instructions. | ||||||||
7068 | if (VF.isScalable()) | ||||||||
7069 | return InstructionCost::getInvalid(); | ||||||||
7070 | // Return cost for branches around scalarized and predicated blocks. | ||||||||
7071 | auto *Vec_i1Ty = | ||||||||
7072 | VectorType::get(IntegerType::getInt1Ty(RetTy->getContext()), VF); | ||||||||
7073 | return ( | ||||||||
7074 | TTI.getScalarizationOverhead( | ||||||||
7075 | Vec_i1Ty, APInt::getAllOnes(VF.getFixedValue()), | ||||||||
7076 | /*Insert*/ false, /*Extract*/ true, CostKind) + | ||||||||
7077 | (TTI.getCFInstrCost(Instruction::Br, CostKind) * VF.getFixedValue())); | ||||||||
7078 | } else if (I->getParent() == TheLoop->getLoopLatch() || VF.isScalar()) | ||||||||
7079 | // The back-edge branch will remain, as will all scalar branches. | ||||||||
7080 | return TTI.getCFInstrCost(Instruction::Br, CostKind); | ||||||||
7081 | else | ||||||||
7082 | // This branch will be eliminated by if-conversion. | ||||||||
7083 | return 0; | ||||||||
7084 | // Note: We currently assume zero cost for an unconditional branch inside | ||||||||
7085 | // a predicated block since it will become a fall-through, although we | ||||||||
7086 | // may decide in the future to call TTI for all branches. | ||||||||
7087 | } | ||||||||
7088 | case Instruction::PHI: { | ||||||||
7089 | auto *Phi = cast<PHINode>(I); | ||||||||
7090 | |||||||||
7091 | // First-order recurrences are replaced by vector shuffles inside the loop. | ||||||||
7092 | if (VF.isVector() && Legal->isFixedOrderRecurrence(Phi)) { | ||||||||
7093 | SmallVector<int> Mask(VF.getKnownMinValue()); | ||||||||
7094 | std::iota(Mask.begin(), Mask.end(), VF.getKnownMinValue() - 1); | ||||||||
7095 | return TTI.getShuffleCost(TargetTransformInfo::SK_Splice, | ||||||||
7096 | cast<VectorType>(VectorTy), Mask, CostKind, | ||||||||
7097 | VF.getKnownMinValue() - 1); | ||||||||
7098 | } | ||||||||
7099 | |||||||||
7100 | // Phi nodes in non-header blocks (not inductions, reductions, etc.) are | ||||||||
7101 | // converted into select instructions. We require N - 1 selects per phi | ||||||||
7102 | // node, where N is the number of incoming values. | ||||||||
7103 | if (VF.isVector() && Phi->getParent() != TheLoop->getHeader()) | ||||||||
7104 | return (Phi->getNumIncomingValues() - 1) * | ||||||||
7105 | TTI.getCmpSelInstrCost( | ||||||||
7106 | Instruction::Select, ToVectorTy(Phi->getType(), VF), | ||||||||
7107 | ToVectorTy(Type::getInt1Ty(Phi->getContext()), VF), | ||||||||
7108 | CmpInst::BAD_ICMP_PREDICATE, CostKind); | ||||||||
7109 | |||||||||
7110 | return TTI.getCFInstrCost(Instruction::PHI, CostKind); | ||||||||
7111 | } | ||||||||
7112 | case Instruction::UDiv: | ||||||||
7113 | case Instruction::SDiv: | ||||||||
7114 | case Instruction::URem: | ||||||||
7115 | case Instruction::SRem: | ||||||||
7116 | if (VF.isVector() && isPredicatedInst(I)) { | ||||||||
7117 | const auto [ScalarCost, SafeDivisorCost] = getDivRemSpeculationCost(I, VF); | ||||||||
7118 | return isDivRemScalarWithPredication(ScalarCost, SafeDivisorCost) ? | ||||||||
7119 | ScalarCost : SafeDivisorCost; | ||||||||
7120 | } | ||||||||
7121 | // We've proven all lanes safe to speculate, fall through. | ||||||||
7122 | [[fallthrough]]; | ||||||||
7123 | case Instruction::Add: | ||||||||
7124 | case Instruction::FAdd: | ||||||||
7125 | case Instruction::Sub: | ||||||||
7126 | case Instruction::FSub: | ||||||||
7127 | case Instruction::Mul: | ||||||||
7128 | case Instruction::FMul: | ||||||||
7129 | case Instruction::FDiv: | ||||||||
7130 | case Instruction::FRem: | ||||||||
7131 | case Instruction::Shl: | ||||||||
7132 | case Instruction::LShr: | ||||||||
7133 | case Instruction::AShr: | ||||||||
7134 | case Instruction::And: | ||||||||
7135 | case Instruction::Or: | ||||||||
7136 | case Instruction::Xor: { | ||||||||
7137 | // Since we will replace the stride by 1 the multiplication should go away. | ||||||||
7138 | if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal)) | ||||||||
7139 | return 0; | ||||||||
7140 | |||||||||
7141 | // Detect reduction patterns | ||||||||
7142 | if (auto RedCost = getReductionPatternCost(I, VF, VectorTy, CostKind)) | ||||||||
7143 | return *RedCost; | ||||||||
7144 | |||||||||
7145 | // Certain instructions can be cheaper to vectorize if they have a constant | ||||||||
7146 | // second vector operand. One example of this are shifts on x86. | ||||||||
7147 | Value *Op2 = I->getOperand(1); | ||||||||
7148 | auto Op2Info = TTI.getOperandInfo(Op2); | ||||||||
7149 | if (Op2Info.Kind == TargetTransformInfo::OK_AnyValue && Legal->isUniform(Op2)) | ||||||||
7150 | Op2Info.Kind = TargetTransformInfo::OK_UniformValue; | ||||||||
7151 | |||||||||
7152 | SmallVector<const Value *, 4> Operands(I->operand_values()); | ||||||||
7153 | return TTI.getArithmeticInstrCost( | ||||||||
7154 | I->getOpcode(), VectorTy, CostKind, | ||||||||
7155 | {TargetTransformInfo::OK_AnyValue, TargetTransformInfo::OP_None}, | ||||||||
7156 | Op2Info, Operands, I); | ||||||||
7157 | } | ||||||||
7158 | case Instruction::FNeg: { | ||||||||
7159 | return TTI.getArithmeticInstrCost( | ||||||||
7160 | I->getOpcode(), VectorTy, CostKind, | ||||||||
7161 | {TargetTransformInfo::OK_AnyValue, TargetTransformInfo::OP_None}, | ||||||||
7162 | {TargetTransformInfo::OK_AnyValue, TargetTransformInfo::OP_None}, | ||||||||
7163 | I->getOperand(0), I); | ||||||||
7164 | } | ||||||||
7165 | case Instruction::Select: { | ||||||||
7166 | SelectInst *SI = cast<SelectInst>(I); | ||||||||
7167 | const SCEV *CondSCEV = SE->getSCEV(SI->getCondition()); | ||||||||
7168 | bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop)); | ||||||||
7169 | |||||||||
7170 | const Value *Op0, *Op1; | ||||||||
7171 | using namespace llvm::PatternMatch; | ||||||||
7172 | if (!ScalarCond && (match(I, m_LogicalAnd(m_Value(Op0), m_Value(Op1))) || | ||||||||
7173 | match(I, m_LogicalOr(m_Value(Op0), m_Value(Op1))))) { | ||||||||
7174 | // select x, y, false --> x & y | ||||||||
7175 | // select x, true, y --> x | y | ||||||||
7176 | const auto [Op1VK, Op1VP] = TTI::getOperandInfo(Op0); | ||||||||
7177 | const auto [Op2VK, Op2VP] = TTI::getOperandInfo(Op1); | ||||||||
7178 | 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", 7179, __extension__ __PRETTY_FUNCTION__)) | ||||||||
7179 | 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", 7179, __extension__ __PRETTY_FUNCTION__)); | ||||||||
7180 | |||||||||
7181 | SmallVector<const Value *, 2> Operands{Op0, Op1}; | ||||||||
7182 | return TTI.getArithmeticInstrCost( | ||||||||
7183 | match(I, m_LogicalOr()) ? Instruction::Or : Instruction::And, VectorTy, | ||||||||
7184 | CostKind, {Op1VK, Op1VP}, {Op2VK, Op2VP}, Operands, I); | ||||||||
7185 | } | ||||||||
7186 | |||||||||
7187 | Type *CondTy = SI->getCondition()->getType(); | ||||||||
7188 | if (!ScalarCond) | ||||||||
7189 | CondTy = VectorType::get(CondTy, VF); | ||||||||
7190 | |||||||||
7191 | CmpInst::Predicate Pred = CmpInst::BAD_ICMP_PREDICATE; | ||||||||
7192 | if (auto *Cmp = dyn_cast<CmpInst>(SI->getCondition())) | ||||||||
7193 | Pred = Cmp->getPredicate(); | ||||||||
7194 | return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy, Pred, | ||||||||
7195 | CostKind, I); | ||||||||
7196 | } | ||||||||
7197 | case Instruction::ICmp: | ||||||||
7198 | case Instruction::FCmp: { | ||||||||
7199 | Type *ValTy = I->getOperand(0)->getType(); | ||||||||
7200 | Instruction *Op0AsInstruction = dyn_cast<Instruction>(I->getOperand(0)); | ||||||||
7201 | if (canTruncateToMinimalBitwidth(Op0AsInstruction, VF)) | ||||||||
7202 | ValTy = IntegerType::get(ValTy->getContext(), MinBWs[Op0AsInstruction]); | ||||||||
7203 | VectorTy = ToVectorTy(ValTy, VF); | ||||||||
7204 | return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, nullptr, | ||||||||
7205 | cast<CmpInst>(I)->getPredicate(), CostKind, | ||||||||
7206 | I); | ||||||||
7207 | } | ||||||||
7208 | case Instruction::Store: | ||||||||
7209 | case Instruction::Load: { | ||||||||
7210 | ElementCount Width = VF; | ||||||||
7211 | if (Width.isVector()) { | ||||||||
7212 | InstWidening Decision = getWideningDecision(I, Width); | ||||||||
7213 | 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", 7214, __extension__ __PRETTY_FUNCTION__)) | ||||||||
7214 | "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", 7214, __extension__ __PRETTY_FUNCTION__)); | ||||||||
7215 | if (getWideningCost(I, VF) == InstructionCost::getInvalid()) | ||||||||
7216 | return InstructionCost::getInvalid(); | ||||||||
7217 | if (Decision == CM_Scalarize) | ||||||||
7218 | Width = ElementCount::getFixed(1); | ||||||||
7219 | } | ||||||||
7220 | VectorTy = ToVectorTy(getLoadStoreType(I), Width); | ||||||||
7221 | return getMemoryInstructionCost(I, VF); | ||||||||
7222 | } | ||||||||
7223 | case Instruction::BitCast: | ||||||||
7224 | if (I->getType()->isPointerTy()) | ||||||||
7225 | return 0; | ||||||||
7226 | [[fallthrough]]; | ||||||||
7227 | case Instruction::ZExt: | ||||||||
7228 | case Instruction::SExt: | ||||||||
7229 | case Instruction::FPToUI: | ||||||||
7230 | case Instruction::FPToSI: | ||||||||
7231 | case Instruction::FPExt: | ||||||||
7232 | case Instruction::PtrToInt: | ||||||||
7233 | case Instruction::IntToPtr: | ||||||||
7234 | case Instruction::SIToFP: | ||||||||
7235 | case Instruction::UIToFP: | ||||||||
7236 | case Instruction::Trunc: | ||||||||
7237 | case Instruction::FPTrunc: { | ||||||||
7238 | // Computes the CastContextHint from a Load/Store instruction. | ||||||||
7239 | auto ComputeCCH = [&](Instruction *I) -> TTI::CastContextHint { | ||||||||
7240 | 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", 7241, __extension__ __PRETTY_FUNCTION__)) | ||||||||
7241 | "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", 7241, __extension__ __PRETTY_FUNCTION__)); | ||||||||
7242 | |||||||||
7243 | if (VF.isScalar() || !TheLoop->contains(I)) | ||||||||
7244 | return TTI::CastContextHint::Normal; | ||||||||
7245 | |||||||||
7246 | switch (getWideningDecision(I, VF)) { | ||||||||
7247 | case LoopVectorizationCostModel::CM_GatherScatter: | ||||||||
7248 | return TTI::CastContextHint::GatherScatter; | ||||||||
7249 | case LoopVectorizationCostModel::CM_Interleave: | ||||||||
7250 | return TTI::CastContextHint::Interleave; | ||||||||
7251 | case LoopVectorizationCostModel::CM_Scalarize: | ||||||||
7252 | case LoopVectorizationCostModel::CM_Widen: | ||||||||
7253 | return Legal->isMaskRequired(I) ? TTI::CastContextHint::Masked | ||||||||
7254 | : TTI::CastContextHint::Normal; | ||||||||
7255 | case LoopVectorizationCostModel::CM_Widen_Reverse: | ||||||||
7256 | return TTI::CastContextHint::Reversed; | ||||||||
7257 | case LoopVectorizationCostModel::CM_Unknown: | ||||||||
7258 | 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", 7258); | ||||||||
7259 | } | ||||||||
7260 | |||||||||
7261 | llvm_unreachable("Unhandled case!")::llvm::llvm_unreachable_internal("Unhandled case!", "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 7261); | ||||||||
7262 | }; | ||||||||
7263 | |||||||||
7264 | unsigned Opcode = I->getOpcode(); | ||||||||
7265 | TTI::CastContextHint CCH = TTI::CastContextHint::None; | ||||||||
7266 | // For Trunc, the context is the only user, which must be a StoreInst. | ||||||||
7267 | if (Opcode == Instruction::Trunc || Opcode == Instruction::FPTrunc) { | ||||||||
7268 | if (I->hasOneUse()) | ||||||||
7269 | if (StoreInst *Store = dyn_cast<StoreInst>(*I->user_begin())) | ||||||||
7270 | CCH = ComputeCCH(Store); | ||||||||
7271 | } | ||||||||
7272 | // For Z/Sext, the context is the operand, which must be a LoadInst. | ||||||||
7273 | else if (Opcode == Instruction::ZExt || Opcode == Instruction::SExt || | ||||||||
7274 | Opcode == Instruction::FPExt) { | ||||||||
7275 | if (LoadInst *Load = dyn_cast<LoadInst>(I->getOperand(0))) | ||||||||
7276 | CCH = ComputeCCH(Load); | ||||||||
7277 | } | ||||||||
7278 | |||||||||
7279 | // We optimize the truncation of induction variables having constant | ||||||||
7280 | // integer steps. The cost of these truncations is the same as the scalar | ||||||||
7281 | // operation. | ||||||||
7282 | if (isOptimizableIVTruncate(I, VF)) { | ||||||||
7283 | auto *Trunc = cast<TruncInst>(I); | ||||||||
7284 | return TTI.getCastInstrCost(Instruction::Trunc, Trunc->getDestTy(), | ||||||||
7285 | Trunc->getSrcTy(), CCH, CostKind, Trunc); | ||||||||
7286 | } | ||||||||
7287 | |||||||||
7288 | // Detect reduction patterns | ||||||||
7289 | if (auto RedCost = getReductionPatternCost(I, VF, VectorTy, CostKind)) | ||||||||
7290 | return *RedCost; | ||||||||
7291 | |||||||||
7292 | Type *SrcScalarTy = I->getOperand(0)->getType(); | ||||||||
7293 | Type *SrcVecTy = | ||||||||
7294 | VectorTy->isVectorTy() ? ToVectorTy(SrcScalarTy, VF) : SrcScalarTy; | ||||||||
7295 | if (canTruncateToMinimalBitwidth(I, VF)) { | ||||||||
7296 | // This cast is going to be shrunk. This may remove the cast or it might | ||||||||
7297 | // turn it into slightly different cast. For example, if MinBW == 16, | ||||||||
7298 | // "zext i8 %1 to i32" becomes "zext i8 %1 to i16". | ||||||||
7299 | // | ||||||||
7300 | // Calculate the modified src and dest types. | ||||||||
7301 | Type *MinVecTy = VectorTy; | ||||||||
7302 | if (Opcode == Instruction::Trunc) { | ||||||||
7303 | SrcVecTy = smallestIntegerVectorType(SrcVecTy, MinVecTy); | ||||||||
7304 | VectorTy = | ||||||||
7305 | largestIntegerVectorType(ToVectorTy(I->getType(), VF), MinVecTy); | ||||||||
7306 | } else if (Opcode == Instruction::ZExt || Opcode == Instruction::SExt) { | ||||||||
7307 | SrcVecTy = largestIntegerVectorType(SrcVecTy, MinVecTy); | ||||||||
7308 | VectorTy = | ||||||||
7309 | smallestIntegerVectorType(ToVectorTy(I->getType(), VF), MinVecTy); | ||||||||
7310 | } | ||||||||
7311 | } | ||||||||
7312 | |||||||||
7313 | return TTI.getCastInstrCost(Opcode, VectorTy, SrcVecTy, CCH, CostKind, I); | ||||||||
7314 | } | ||||||||
7315 | case Instruction::Call: { | ||||||||
7316 | if (RecurrenceDescriptor::isFMulAddIntrinsic(I)) | ||||||||
7317 | if (auto RedCost = getReductionPatternCost(I, VF, VectorTy, CostKind)) | ||||||||
7318 | return *RedCost; | ||||||||
7319 | bool NeedToScalarize; | ||||||||
7320 | CallInst *CI = cast<CallInst>(I); | ||||||||
7321 | InstructionCost CallCost = getVectorCallCost(CI, VF, NeedToScalarize); | ||||||||
7322 | if (getVectorIntrinsicIDForCall(CI, TLI)) { | ||||||||
7323 | InstructionCost IntrinsicCost = getVectorIntrinsicCost(CI, VF); | ||||||||
7324 | return std::min(CallCost, IntrinsicCost); | ||||||||
7325 | } | ||||||||
7326 | return CallCost; | ||||||||
7327 | } | ||||||||
7328 | case Instruction::ExtractValue: | ||||||||
7329 | return TTI.getInstructionCost(I, TTI::TCK_RecipThroughput); | ||||||||
7330 | case Instruction::Alloca: | ||||||||
7331 | // We cannot easily widen alloca to a scalable alloca, as | ||||||||
7332 | // the result would need to be a vector of pointers. | ||||||||
7333 | if (VF.isScalable()) | ||||||||
7334 | return InstructionCost::getInvalid(); | ||||||||
7335 | [[fallthrough]]; | ||||||||
7336 | default: | ||||||||
7337 | // This opcode is unknown. Assume that it is the same as 'mul'. | ||||||||
7338 | return TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy, CostKind); | ||||||||
7339 | } // end of switch. | ||||||||
7340 | } | ||||||||
7341 | |||||||||
7342 | char LoopVectorize::ID = 0; | ||||||||
7343 | |||||||||
7344 | static const char lv_name[] = "Loop Vectorization"; | ||||||||
7345 | |||||||||
7346 | INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)static void *initializeLoopVectorizePassOnce(PassRegistry & Registry) { | ||||||||
7347 | INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)initializeTargetTransformInfoWrapperPassPass(Registry); | ||||||||
7348 | INITIALIZE_PASS_DEPENDENCY(BasicAAWrapperPass)initializeBasicAAWrapperPassPass(Registry); | ||||||||
7349 | INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)initializeGlobalsAAWrapperPassPass(Registry); | ||||||||
7350 | INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)initializeAssumptionCacheTrackerPass(Registry); | ||||||||
7351 | INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass)initializeBlockFrequencyInfoWrapperPassPass(Registry); | ||||||||
7352 | INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)initializeDominatorTreeWrapperPassPass(Registry); | ||||||||
7353 | INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)initializeScalarEvolutionWrapperPassPass(Registry); | ||||||||
7354 | INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)initializeLoopInfoWrapperPassPass(Registry); | ||||||||
7355 | INITIALIZE_PASS_DEPENDENCY(LoopAccessLegacyAnalysis)initializeLoopAccessLegacyAnalysisPass(Registry); | ||||||||
7356 | INITIALIZE_PASS_DEPENDENCY(DemandedBitsWrapperPass)initializeDemandedBitsWrapperPassPass(Registry); | ||||||||
7357 | INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass)initializeOptimizationRemarkEmitterWrapperPassPass(Registry); | ||||||||
7358 | INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass)initializeProfileSummaryInfoWrapperPassPass(Registry); | ||||||||
7359 | INITIALIZE_PASS_DEPENDENCY(InjectTLIMappingsLegacy)initializeInjectTLIMappingsLegacyPass(Registry); | ||||||||
7360 | 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)); } | ||||||||
7361 | |||||||||
7362 | namespace llvm { | ||||||||
7363 | |||||||||
7364 | Pass *createLoopVectorizePass() { return new LoopVectorize(); } | ||||||||
7365 | |||||||||
7366 | Pass *createLoopVectorizePass(bool InterleaveOnlyWhenForced, | ||||||||
7367 | bool VectorizeOnlyWhenForced) { | ||||||||
7368 | return new LoopVectorize(InterleaveOnlyWhenForced, VectorizeOnlyWhenForced); | ||||||||
7369 | } | ||||||||
7370 | |||||||||
7371 | } // end namespace llvm | ||||||||
7372 | |||||||||
7373 | void LoopVectorizationCostModel::collectValuesToIgnore() { | ||||||||
7374 | // Ignore ephemeral values. | ||||||||
7375 | CodeMetrics::collectEphemeralValues(TheLoop, AC, ValuesToIgnore); | ||||||||
7376 | |||||||||
7377 | // Find all stores to invariant variables. Since they are going to sink | ||||||||
7378 | // outside the loop we do not need calculate cost for them. | ||||||||
7379 | for (BasicBlock *BB : TheLoop->blocks()) | ||||||||
7380 | for (Instruction &I : *BB) { | ||||||||
7381 | StoreInst *SI; | ||||||||
7382 | if ((SI = dyn_cast<StoreInst>(&I)) && | ||||||||
7383 | Legal->isInvariantAddressOfReduction(SI->getPointerOperand())) | ||||||||
7384 | ValuesToIgnore.insert(&I); | ||||||||
7385 | } | ||||||||
7386 | |||||||||
7387 | // Ignore type-promoting instructions we identified during reduction | ||||||||
7388 | // detection. | ||||||||
7389 | for (const auto &Reduction : Legal->getReductionVars()) { | ||||||||
7390 | const RecurrenceDescriptor &RedDes = Reduction.second; | ||||||||
7391 | const SmallPtrSetImpl<Instruction *> &Casts = RedDes.getCastInsts(); | ||||||||
7392 | VecValuesToIgnore.insert(Casts.begin(), Casts.end()); | ||||||||
7393 | } | ||||||||
7394 | // Ignore type-casting instructions we identified during induction | ||||||||
7395 | // detection. | ||||||||
7396 | for (const auto &Induction : Legal->getInductionVars()) { | ||||||||
7397 | const InductionDescriptor &IndDes = Induction.second; | ||||||||
7398 | const SmallVectorImpl<Instruction *> &Casts = IndDes.getCastInsts(); | ||||||||
7399 | VecValuesToIgnore.insert(Casts.begin(), Casts.end()); | ||||||||
7400 | } | ||||||||
7401 | } | ||||||||
7402 | |||||||||
7403 | void LoopVectorizationCostModel::collectInLoopReductions() { | ||||||||
7404 | for (const auto &Reduction : Legal->getReductionVars()) { | ||||||||
7405 | PHINode *Phi = Reduction.first; | ||||||||
7406 | const RecurrenceDescriptor &RdxDesc = Reduction.second; | ||||||||
7407 | |||||||||
7408 | // We don't collect reductions that are type promoted (yet). | ||||||||
7409 | if (RdxDesc.getRecurrenceType() != Phi->getType()) | ||||||||
7410 | continue; | ||||||||
7411 | |||||||||
7412 | // If the target would prefer this reduction to happen "in-loop", then we | ||||||||
7413 | // want to record it as such. | ||||||||
7414 | unsigned Opcode = RdxDesc.getOpcode(); | ||||||||
7415 | if (!PreferInLoopReductions && !useOrderedReductions(RdxDesc) && | ||||||||
7416 | !TTI.preferInLoopReduction(Opcode, Phi->getType(), | ||||||||
7417 | TargetTransformInfo::ReductionFlags())) | ||||||||
7418 | continue; | ||||||||
7419 | |||||||||
7420 | // Check that we can correctly put the reductions into the loop, by | ||||||||
7421 | // finding the chain of operations that leads from the phi to the loop | ||||||||
7422 | // exit value. | ||||||||
7423 | SmallVector<Instruction *, 4> ReductionOperations = | ||||||||
7424 | RdxDesc.getReductionOpChain(Phi, TheLoop); | ||||||||
7425 | bool InLoop = !ReductionOperations.empty(); | ||||||||
7426 | if (InLoop) { | ||||||||
7427 | InLoopReductionChains[Phi] = ReductionOperations; | ||||||||
7428 | // Add the elements to InLoopReductionImmediateChains for cost modelling. | ||||||||
7429 | Instruction *LastChain = Phi; | ||||||||
7430 | for (auto *I : ReductionOperations) { | ||||||||
7431 | InLoopReductionImmediateChains[I] = LastChain; | ||||||||
7432 | LastChain = I; | ||||||||
7433 | } | ||||||||
7434 | } | ||||||||
7435 | 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) | ||||||||
7436 | << " 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); | ||||||||
7437 | } | ||||||||
7438 | } | ||||||||
7439 | |||||||||
7440 | // TODO: we could return a pair of values that specify the max VF and | ||||||||
7441 | // min VF, to be used in `buildVPlans(MinVF, MaxVF)` instead of | ||||||||
7442 | // `buildVPlans(VF, VF)`. We cannot do it because VPLAN at the moment | ||||||||
7443 | // doesn't have a cost model that can choose which plan to execute if | ||||||||
7444 | // more than one is generated. | ||||||||
7445 | static unsigned determineVPlanVF(const unsigned WidestVectorRegBits, | ||||||||
7446 | LoopVectorizationCostModel &CM) { | ||||||||
7447 | unsigned WidestType; | ||||||||
7448 | std::tie(std::ignore, WidestType) = CM.getSmallestAndWidestTypes(); | ||||||||
7449 | return WidestVectorRegBits / WidestType; | ||||||||
7450 | } | ||||||||
7451 | |||||||||
7452 | VectorizationFactor | ||||||||
7453 | LoopVectorizationPlanner::planInVPlanNativePath(ElementCount UserVF) { | ||||||||
7454 | 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", 7454, __extension__ __PRETTY_FUNCTION__)); | ||||||||
7455 | ElementCount VF = UserVF; | ||||||||
7456 | // Outer loop handling: They may require CFG and instruction level | ||||||||
7457 | // transformations before even evaluating whether vectorization is profitable. | ||||||||
7458 | // Since we cannot modify the incoming IR, we need to build VPlan upfront in | ||||||||
7459 | // the vectorization pipeline. | ||||||||
7460 | if (!OrigLoop->isInnermost()) { | ||||||||
7461 | // If the user doesn't provide a vectorization factor, determine a | ||||||||
7462 | // reasonable one. | ||||||||
7463 | if (UserVF.isZero()) { | ||||||||
7464 | VF = ElementCount::getFixed(determineVPlanVF( | ||||||||
7465 | TTI->getRegisterBitWidth(TargetTransformInfo::RGK_FixedWidthVector) | ||||||||
7466 | .getFixedValue(), | ||||||||
7467 | CM)); | ||||||||
7468 | 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); | ||||||||
7469 | |||||||||
7470 | // Make sure we have a VF > 1 for stress testing. | ||||||||
7471 | if (VPlanBuildStressTest && (VF.isScalar() || VF.isZero())) { | ||||||||
7472 | 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) | ||||||||
7473 | << "overriding computed VF.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: VPlan stress testing: " << "overriding computed VF.\n"; } } while (false); | ||||||||
7474 | VF = ElementCount::getFixed(4); | ||||||||
7475 | } | ||||||||
7476 | } | ||||||||
7477 | 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", 7477, __extension__ __PRETTY_FUNCTION__)); | ||||||||
7478 | 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", 7479, __extension__ __PRETTY_FUNCTION__)) | ||||||||
7479 | "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", 7479, __extension__ __PRETTY_FUNCTION__)); | ||||||||
7480 | 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) | ||||||||
7481 | << "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); | ||||||||
7482 | buildVPlans(VF, VF); | ||||||||
7483 | |||||||||
7484 | // For VPlan build stress testing, we bail out after VPlan construction. | ||||||||
7485 | if (VPlanBuildStressTest) | ||||||||
7486 | return VectorizationFactor::Disabled(); | ||||||||
7487 | |||||||||
7488 | return {VF, 0 /*Cost*/, 0 /* ScalarCost */}; | ||||||||
7489 | } | ||||||||
7490 | |||||||||
7491 | 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) | ||||||||
7492 | 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) | ||||||||
7493 | "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); | ||||||||
7494 | return VectorizationFactor::Disabled(); | ||||||||
7495 | } | ||||||||
7496 | |||||||||
7497 | std::optional<VectorizationFactor> | ||||||||
7498 | LoopVectorizationPlanner::plan(ElementCount UserVF, unsigned UserIC) { | ||||||||
7499 | 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", 7499, __extension__ __PRETTY_FUNCTION__)); | ||||||||
7500 | FixedScalableVFPair MaxFactors = CM.computeMaxVF(UserVF, UserIC); | ||||||||
7501 | if (!MaxFactors) // Cases that should not to be vectorized nor interleaved. | ||||||||
7502 | return std::nullopt; | ||||||||
7503 | |||||||||
7504 | // Invalidate interleave groups if all blocks of loop will be predicated. | ||||||||
7505 | if (CM.blockNeedsPredicationForAnyReason(OrigLoop->getHeader()) && | ||||||||
7506 | !useMaskedInterleavedAccesses(*TTI)) { | ||||||||
7507 | 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 ) | ||||||||
7508 | 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 ) | ||||||||
7509 | << "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 ) | ||||||||
7510 | "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 ); | ||||||||
7511 | if (CM.InterleaveInfo.invalidateGroups()) | ||||||||
7512 | // Invalidating interleave groups also requires invalidating all decisions | ||||||||
7513 | // based on them, which includes widening decisions and uniform and scalar | ||||||||
7514 | // values. | ||||||||
7515 | CM.invalidateCostModelingDecisions(); | ||||||||
7516 | } | ||||||||
7517 | |||||||||
7518 | ElementCount MaxUserVF = | ||||||||
7519 | UserVF.isScalable() ? MaxFactors.ScalableVF : MaxFactors.FixedVF; | ||||||||
7520 | bool UserVFIsLegal = ElementCount::isKnownLE(UserVF, MaxUserVF); | ||||||||
7521 | if (!UserVF.isZero() && UserVFIsLegal) { | ||||||||
7522 | 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", 7523, __extension__ __PRETTY_FUNCTION__)) | ||||||||
7523 | "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", 7523, __extension__ __PRETTY_FUNCTION__)); | ||||||||
7524 | // Collect the instructions (and their associated costs) that will be more | ||||||||
7525 | // profitable to scalarize. | ||||||||
7526 | if (CM.selectUserVectorizationFactor(UserVF)) { | ||||||||
7527 | 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); | ||||||||
7528 | CM.collectInLoopReductions(); | ||||||||
7529 | buildVPlansWithVPRecipes(UserVF, UserVF); | ||||||||
7530 | LLVM_DEBUG(printPlans(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { printPlans(dbgs()); } } while (false); | ||||||||
7531 | return {{UserVF, 0, 0}}; | ||||||||
7532 | } else | ||||||||
7533 | reportVectorizationInfo("UserVF ignored because of invalid costs.", | ||||||||
7534 | "InvalidCost", ORE, OrigLoop); | ||||||||
7535 | } | ||||||||
7536 | |||||||||
7537 | // Populate the set of Vectorization Factor Candidates. | ||||||||
7538 | ElementCountSet VFCandidates; | ||||||||
7539 | for (auto VF = ElementCount::getFixed(1); | ||||||||
7540 | ElementCount::isKnownLE(VF, MaxFactors.FixedVF); VF *= 2) | ||||||||
7541 | VFCandidates.insert(VF); | ||||||||
7542 | for (auto VF = ElementCount::getScalable(1); | ||||||||
7543 | ElementCount::isKnownLE(VF, MaxFactors.ScalableVF); VF *= 2) | ||||||||
7544 | VFCandidates.insert(VF); | ||||||||
7545 | |||||||||
7546 | for (const auto &VF : VFCandidates) { | ||||||||
7547 | // Collect Uniform and Scalar instructions after vectorization with VF. | ||||||||
7548 | CM.collectUniformsAndScalars(VF); | ||||||||
7549 | |||||||||
7550 | // Collect the instructions (and their associated costs) that will be more | ||||||||
7551 | // profitable to scalarize. | ||||||||
7552 | if (VF.isVector()) | ||||||||
7553 | CM.collectInstsToScalarize(VF); | ||||||||
7554 | } | ||||||||
7555 | |||||||||
7556 | CM.collectInLoopReductions(); | ||||||||
7557 | buildVPlansWithVPRecipes(ElementCount::getFixed(1), MaxFactors.FixedVF); | ||||||||
7558 | buildVPlansWithVPRecipes(ElementCount::getScalable(1), MaxFactors.ScalableVF); | ||||||||
7559 | |||||||||
7560 | LLVM_DEBUG(printPlans(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { printPlans(dbgs()); } } while (false); | ||||||||
7561 | if (!MaxFactors.hasVector()) | ||||||||
7562 | return VectorizationFactor::Disabled(); | ||||||||
7563 | |||||||||
7564 | // Select the optimal vectorization factor. | ||||||||
7565 | VectorizationFactor VF = CM.selectVectorizationFactor(VFCandidates); | ||||||||
7566 | assert((VF.Width.isScalar() || VF.ScalarCost > 0) && "when vectorizing, the scalar cost must be non-zero.")(static_cast <bool> ((VF.Width.isScalar() || VF.ScalarCost > 0) && "when vectorizing, the scalar cost must be non-zero." ) ? void (0) : __assert_fail ("(VF.Width.isScalar() || VF.ScalarCost > 0) && \"when vectorizing, the scalar cost must be non-zero.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7566, __extension__ __PRETTY_FUNCTION__)); | ||||||||
7567 | return VF; | ||||||||
7568 | } | ||||||||
7569 | |||||||||
7570 | VPlan &LoopVectorizationPlanner::getBestPlanFor(ElementCount VF) const { | ||||||||
7571 | 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", 7574, __extension__ __PRETTY_FUNCTION__)) | ||||||||
7572 | [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", 7574, __extension__ __PRETTY_FUNCTION__)) | ||||||||
7573 | 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", 7574, __extension__ __PRETTY_FUNCTION__)) | ||||||||
7574 | "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", 7574, __extension__ __PRETTY_FUNCTION__)); | ||||||||
7575 | |||||||||
7576 | for (const VPlanPtr &Plan : VPlans) { | ||||||||
7577 | if (Plan->hasVF(VF)) | ||||||||
7578 | return *Plan.get(); | ||||||||
7579 | } | ||||||||
7580 | llvm_unreachable("No plan found!")::llvm::llvm_unreachable_internal("No plan found!", "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 7580); | ||||||||
7581 | } | ||||||||
7582 | |||||||||
7583 | static void AddRuntimeUnrollDisableMetaData(Loop *L) { | ||||||||
7584 | SmallVector<Metadata *, 4> MDs; | ||||||||
7585 | // Reserve first location for self reference to the LoopID metadata node. | ||||||||
7586 | MDs.push_back(nullptr); | ||||||||
7587 | bool IsUnrollMetadata = false; | ||||||||
7588 | MDNode *LoopID = L->getLoopID(); | ||||||||
7589 | if (LoopID) { | ||||||||
7590 | // First find existing loop unrolling disable metadata. | ||||||||
7591 | for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) { | ||||||||
7592 | auto *MD = dyn_cast<MDNode>(LoopID->getOperand(i)); | ||||||||
7593 | if (MD) { | ||||||||
7594 | const auto *S = dyn_cast<MDString>(MD->getOperand(0)); | ||||||||
7595 | IsUnrollMetadata = | ||||||||
7596 | S && S->getString().startswith("llvm.loop.unroll.disable"); | ||||||||
7597 | } | ||||||||
7598 | MDs.push_back(LoopID->getOperand(i)); | ||||||||
7599 | } | ||||||||
7600 | } | ||||||||
7601 | |||||||||
7602 | if (!IsUnrollMetadata) { | ||||||||
7603 | // Add runtime unroll disable metadata. | ||||||||
7604 | LLVMContext &Context = L->getHeader()->getContext(); | ||||||||
7605 | SmallVector<Metadata *, 1> DisableOperands; | ||||||||
7606 | DisableOperands.push_back( | ||||||||
7607 | MDString::get(Context, "llvm.loop.unroll.runtime.disable")); | ||||||||
7608 | MDNode *DisableNode = MDNode::get(Context, DisableOperands); | ||||||||
7609 | MDs.push_back(DisableNode); | ||||||||
7610 | MDNode *NewLoopID = MDNode::get(Context, MDs); | ||||||||
7611 | // Set operand 0 to refer to the loop id itself. | ||||||||
7612 | NewLoopID->replaceOperandWith(0, NewLoopID); | ||||||||
7613 | L->setLoopID(NewLoopID); | ||||||||
7614 | } | ||||||||
7615 | } | ||||||||
7616 | |||||||||
7617 | void LoopVectorizationPlanner::executePlan(ElementCount BestVF, unsigned BestUF, | ||||||||
7618 | VPlan &BestVPlan, | ||||||||
7619 | InnerLoopVectorizer &ILV, | ||||||||
7620 | DominatorTree *DT, | ||||||||
7621 | bool IsEpilogueVectorization) { | ||||||||
7622 | assert(BestVPlan.hasVF(BestVF) &&(static_cast <bool> (BestVPlan.hasVF(BestVF) && "Trying to execute plan with unsupported VF") ? void (0) : __assert_fail ("BestVPlan.hasVF(BestVF) && \"Trying to execute plan with unsupported VF\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7623, __extension__ __PRETTY_FUNCTION__)) | ||||||||
7623 | "Trying to execute plan with unsupported VF")(static_cast <bool> (BestVPlan.hasVF(BestVF) && "Trying to execute plan with unsupported VF") ? void (0) : __assert_fail ("BestVPlan.hasVF(BestVF) && \"Trying to execute plan with unsupported VF\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7623, __extension__ __PRETTY_FUNCTION__)); | ||||||||
7624 | assert(BestVPlan.hasUF(BestUF) &&(static_cast <bool> (BestVPlan.hasUF(BestUF) && "Trying to execute plan with unsupported UF") ? void (0) : __assert_fail ("BestVPlan.hasUF(BestUF) && \"Trying to execute plan with unsupported UF\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7625, __extension__ __PRETTY_FUNCTION__)) | ||||||||
7625 | "Trying to execute plan with unsupported UF")(static_cast <bool> (BestVPlan.hasUF(BestUF) && "Trying to execute plan with unsupported UF") ? void (0) : __assert_fail ("BestVPlan.hasUF(BestUF) && \"Trying to execute plan with unsupported UF\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7625, __extension__ __PRETTY_FUNCTION__)); | ||||||||
7626 | |||||||||
7627 | 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) | ||||||||
7628 | << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "Executing best plan with VF=" << BestVF << ", UF=" << BestUF << '\n' ; } } while (false); | ||||||||
7629 | |||||||||
7630 | // Workaround! Compute the trip count of the original loop and cache it | ||||||||
7631 | // before we start modifying the CFG. This code has a systemic problem | ||||||||
7632 | // wherein it tries to run analysis over partially constructed IR; this is | ||||||||
7633 | // wrong, and not simply for SCEV. The trip count of the original loop | ||||||||
7634 | // simply happens to be prone to hitting this in practice. In theory, we | ||||||||
7635 | // can hit the same issue for any SCEV, or ValueTracking query done during | ||||||||
7636 | // mutation. See PR49900. | ||||||||
7637 | ILV.getOrCreateTripCount(OrigLoop->getLoopPreheader()); | ||||||||
7638 | |||||||||
7639 | if (!IsEpilogueVectorization) | ||||||||
7640 | VPlanTransforms::optimizeForVFAndUF(BestVPlan, BestVF, BestUF, PSE); | ||||||||
7641 | |||||||||
7642 | // Perform the actual loop transformation. | ||||||||
7643 | |||||||||
7644 | // 1. Set up the skeleton for vectorization, including vector pre-header and | ||||||||
7645 | // middle block. The vector loop is created during VPlan execution. | ||||||||
7646 | VPTransformState State{BestVF, BestUF, LI, DT, ILV.Builder, &ILV, &BestVPlan}; | ||||||||
7647 | Value *CanonicalIVStartValue; | ||||||||
7648 | std::tie(State.CFG.PrevBB, CanonicalIVStartValue) = | ||||||||
7649 | ILV.createVectorizedLoopSkeleton(); | ||||||||
7650 | |||||||||
7651 | // Only use noalias metadata when using memory checks guaranteeing no overlap | ||||||||
7652 | // across all iterations. | ||||||||
7653 | const LoopAccessInfo *LAI = ILV.Legal->getLAI(); | ||||||||
7654 | if (LAI && !LAI->getRuntimePointerChecking()->getChecks().empty() && | ||||||||
7655 | !LAI->getRuntimePointerChecking()->getDiffChecks()) { | ||||||||
7656 | |||||||||
7657 | // We currently don't use LoopVersioning for the actual loop cloning but we | ||||||||
7658 | // still use it to add the noalias metadata. | ||||||||
7659 | // TODO: Find a better way to re-use LoopVersioning functionality to add | ||||||||
7660 | // metadata. | ||||||||
7661 | State.LVer = std::make_unique<LoopVersioning>( | ||||||||
7662 | *LAI, LAI->getRuntimePointerChecking()->getChecks(), OrigLoop, LI, DT, | ||||||||
7663 | PSE.getSE()); | ||||||||
7664 | State.LVer->prepareNoAliasMetadata(); | ||||||||
7665 | } | ||||||||
7666 | |||||||||
7667 | ILV.collectPoisonGeneratingRecipes(State); | ||||||||
7668 | |||||||||
7669 | ILV.printDebugTracesAtStart(); | ||||||||
7670 | |||||||||
7671 | //===------------------------------------------------===// | ||||||||
7672 | // | ||||||||
7673 | // Notice: any optimization or new instruction that go | ||||||||
7674 | // into the code below should also be implemented in | ||||||||
7675 | // the cost-model. | ||||||||
7676 | // | ||||||||
7677 | //===------------------------------------------------===// | ||||||||
7678 | |||||||||
7679 | // 2. Copy and widen instructions from the old loop into the new loop. | ||||||||
7680 | BestVPlan.prepareToExecute(ILV.getOrCreateTripCount(nullptr), | ||||||||
7681 | ILV.getOrCreateVectorTripCount(nullptr), | ||||||||
7682 | CanonicalIVStartValue, State, | ||||||||
7683 | IsEpilogueVectorization); | ||||||||
7684 | |||||||||
7685 | BestVPlan.execute(&State); | ||||||||
7686 | |||||||||
7687 | // Keep all loop hints from the original loop on the vector loop (we'll | ||||||||
7688 | // replace the vectorizer-specific hints below). | ||||||||
7689 | MDNode *OrigLoopID = OrigLoop->getLoopID(); | ||||||||
7690 | |||||||||
7691 | std::optional<MDNode *> VectorizedLoopID = | ||||||||
7692 | makeFollowupLoopID(OrigLoopID, {LLVMLoopVectorizeFollowupAll, | ||||||||
7693 | LLVMLoopVectorizeFollowupVectorized}); | ||||||||
7694 | |||||||||
7695 | VPBasicBlock *HeaderVPBB = | ||||||||
7696 | BestVPlan.getVectorLoopRegion()->getEntryBasicBlock(); | ||||||||
7697 | Loop *L = LI->getLoopFor(State.CFG.VPBB2IRBB[HeaderVPBB]); | ||||||||
7698 | if (VectorizedLoopID) | ||||||||
7699 | L->setLoopID(*VectorizedLoopID); | ||||||||
7700 | else { | ||||||||
7701 | // Keep all loop hints from the original loop on the vector loop (we'll | ||||||||
7702 | // replace the vectorizer-specific hints below). | ||||||||
7703 | if (MDNode *LID = OrigLoop->getLoopID()) | ||||||||
7704 | L->setLoopID(LID); | ||||||||
7705 | |||||||||
7706 | LoopVectorizeHints Hints(L, true, *ORE); | ||||||||
7707 | Hints.setAlreadyVectorized(); | ||||||||
7708 | } | ||||||||
7709 | AddRuntimeUnrollDisableMetaData(L); | ||||||||
7710 | |||||||||
7711 | // 3. Fix the vectorized code: take care of header phi's, live-outs, | ||||||||
7712 | // predication, updating analyses. | ||||||||
7713 | ILV.fixVectorizedLoop(State, BestVPlan); | ||||||||
7714 | |||||||||
7715 | ILV.printDebugTracesAtEnd(); | ||||||||
7716 | } | ||||||||
7717 | |||||||||
7718 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||
7719 | void LoopVectorizationPlanner::printPlans(raw_ostream &O) { | ||||||||
7720 | for (const auto &Plan : VPlans) | ||||||||
7721 | if (PrintVPlansInDotFormat) | ||||||||
7722 | Plan->printDOT(O); | ||||||||
7723 | else | ||||||||
7724 | Plan->print(O); | ||||||||
7725 | } | ||||||||
7726 | #endif | ||||||||
7727 | |||||||||
7728 | Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) { return V; } | ||||||||
7729 | |||||||||
7730 | //===--------------------------------------------------------------------===// | ||||||||
7731 | // EpilogueVectorizerMainLoop | ||||||||
7732 | //===--------------------------------------------------------------------===// | ||||||||
7733 | |||||||||
7734 | /// This function is partially responsible for generating the control flow | ||||||||
7735 | /// depicted in https://llvm.org/docs/Vectorizers.html#epilogue-vectorization. | ||||||||
7736 | std::pair<BasicBlock *, Value *> | ||||||||
7737 | EpilogueVectorizerMainLoop::createEpilogueVectorizedLoopSkeleton() { | ||||||||
7738 | createVectorLoopSkeleton(""); | ||||||||
7739 | |||||||||
7740 | // Generate the code to check the minimum iteration count of the vector | ||||||||
7741 | // epilogue (see below). | ||||||||
7742 | EPI.EpilogueIterationCountCheck = | ||||||||
7743 | emitIterationCountCheck(LoopScalarPreHeader, true); | ||||||||
7744 | EPI.EpilogueIterationCountCheck->setName("iter.check"); | ||||||||
7745 | |||||||||
7746 | // Generate the code to check any assumptions that we've made for SCEV | ||||||||
7747 | // expressions. | ||||||||
7748 | EPI.SCEVSafetyCheck = emitSCEVChecks(LoopScalarPreHeader); | ||||||||
7749 | |||||||||
7750 | // Generate the code that checks at runtime if arrays overlap. We put the | ||||||||
7751 | // checks into a separate block to make the more common case of few elements | ||||||||
7752 | // faster. | ||||||||
7753 | EPI.MemSafetyCheck = emitMemRuntimeChecks(LoopScalarPreHeader); | ||||||||
7754 | |||||||||
7755 | // Generate the iteration count check for the main loop, *after* the check | ||||||||
7756 | // for the epilogue loop, so that the path-length is shorter for the case | ||||||||
7757 | // that goes directly through the vector epilogue. The longer-path length for | ||||||||
7758 | // the main loop is compensated for, by the gain from vectorizing the larger | ||||||||
7759 | // trip count. Note: the branch will get updated later on when we vectorize | ||||||||
7760 | // the epilogue. | ||||||||
7761 | EPI.MainLoopIterationCountCheck = | ||||||||
7762 | emitIterationCountCheck(LoopScalarPreHeader, false); | ||||||||
7763 | |||||||||
7764 | // Generate the induction variable. | ||||||||
7765 | EPI.VectorTripCount = getOrCreateVectorTripCount(LoopVectorPreHeader); | ||||||||
7766 | |||||||||
7767 | // Skip induction resume value creation here because they will be created in | ||||||||
7768 | // the second pass for the scalar loop. The induction resume values for the | ||||||||
7769 | // inductions in the epilogue loop are created before executing the plan for | ||||||||
7770 | // the epilogue loop. | ||||||||
7771 | |||||||||
7772 | return {completeLoopSkeleton(), nullptr}; | ||||||||
7773 | } | ||||||||
7774 | |||||||||
7775 | void EpilogueVectorizerMainLoop::printDebugTracesAtStart() { | ||||||||
7776 | 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) | ||||||||
7777 | 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) | ||||||||
7778 | << "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) | ||||||||
7779 | << ", 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) | ||||||||
7780 | << ", 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) | ||||||||
7781 | << ", 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) | ||||||||
7782 | })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); | ||||||||
7783 | } | ||||||||
7784 | |||||||||
7785 | void EpilogueVectorizerMainLoop::printDebugTracesAtEnd() { | ||||||||
7786 | DEBUG_WITH_TYPE(VerboseDebug, {do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType (VerboseDebug)) { { dbgs() << "intermediate fn:\n" << *OrigLoop->getHeader()->getParent() << "\n"; }; } } while (false) | ||||||||
7787 | dbgs() << "intermediate fn:\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType (VerboseDebug)) { { dbgs() << "intermediate fn:\n" << *OrigLoop->getHeader()->getParent() << "\n"; }; } } while (false) | ||||||||
7788 | << *OrigLoop->getHeader()->getParent() << "\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType (VerboseDebug)) { { dbgs() << "intermediate fn:\n" << *OrigLoop->getHeader()->getParent() << "\n"; }; } } while (false) | ||||||||
7789 | })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType (VerboseDebug)) { { dbgs() << "intermediate fn:\n" << *OrigLoop->getHeader()->getParent() << "\n"; }; } } while (false); | ||||||||
7790 | } | ||||||||
7791 | |||||||||
7792 | BasicBlock * | ||||||||
7793 | EpilogueVectorizerMainLoop::emitIterationCountCheck(BasicBlock *Bypass, | ||||||||
7794 | bool ForEpilogue) { | ||||||||
7795 | 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", 7795, __extension__ __PRETTY_FUNCTION__)); | ||||||||
7796 | ElementCount VFactor = ForEpilogue ? EPI.EpilogueVF : VF; | ||||||||
7797 | unsigned UFactor = ForEpilogue ? EPI.EpilogueUF : UF; | ||||||||
7798 | Value *Count = getOrCreateTripCount(LoopVectorPreHeader); | ||||||||
7799 | // Reuse existing vector loop preheader for TC checks. | ||||||||
7800 | // Note that new preheader block is generated for vector loop. | ||||||||
7801 | BasicBlock *const TCCheckBlock = LoopVectorPreHeader; | ||||||||
7802 | IRBuilder<> Builder(TCCheckBlock->getTerminator()); | ||||||||
7803 | |||||||||
7804 | // Generate code to check if the loop's trip count is less than VF * UF of the | ||||||||
7805 | // main vector loop. | ||||||||
7806 | auto P = Cost->requiresScalarEpilogue(ForEpilogue ? EPI.EpilogueVF : VF) ? | ||||||||
7807 | ICmpInst::ICMP_ULE : ICmpInst::ICMP_ULT; | ||||||||
7808 | |||||||||
7809 | Value *CheckMinIters = Builder.CreateICmp( | ||||||||
7810 | P, Count, createStepForVF(Builder, Count->getType(), VFactor, UFactor), | ||||||||
7811 | "min.iters.check"); | ||||||||
7812 | |||||||||
7813 | if (!ForEpilogue) | ||||||||
7814 | TCCheckBlock->setName("vector.main.loop.iter.check"); | ||||||||
7815 | |||||||||
7816 | // Create new preheader for vector loop. | ||||||||
7817 | LoopVectorPreHeader = SplitBlock(TCCheckBlock, TCCheckBlock->getTerminator(), | ||||||||
7818 | DT, LI, nullptr, "vector.ph"); | ||||||||
7819 | |||||||||
7820 | if (ForEpilogue) { | ||||||||
7821 | 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", 7823, __extension__ __PRETTY_FUNCTION__)) | ||||||||
7822 | 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", 7823, __extension__ __PRETTY_FUNCTION__)) | ||||||||
7823 | "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", 7823, __extension__ __PRETTY_FUNCTION__)); | ||||||||
7824 | |||||||||
7825 | // Update dominator for Bypass & LoopExit. | ||||||||
7826 | DT->changeImmediateDominator(Bypass, TCCheckBlock); | ||||||||
7827 | if (!Cost->requiresScalarEpilogue(EPI.EpilogueVF)) | ||||||||
7828 | // For loops with multiple exits, there's no edge from the middle block | ||||||||
7829 | // to exit blocks (as the epilogue must run) and thus no need to update | ||||||||
7830 | // the immediate dominator of the exit blocks. | ||||||||
7831 | DT->changeImmediateDominator(LoopExitBlock, TCCheckBlock); | ||||||||
7832 | |||||||||
7833 | LoopBypassBlocks.push_back(TCCheckBlock); | ||||||||
7834 | |||||||||
7835 | // Save the trip count so we don't have to regenerate it in the | ||||||||
7836 | // vec.epilog.iter.check. This is safe to do because the trip count | ||||||||
7837 | // generated here dominates the vector epilog iter check. | ||||||||
7838 | EPI.TripCount = Count; | ||||||||
7839 | } | ||||||||
7840 | |||||||||
7841 | ReplaceInstWithInst( | ||||||||
7842 | TCCheckBlock->getTerminator(), | ||||||||
7843 | BranchInst::Create(Bypass, LoopVectorPreHeader, CheckMinIters)); | ||||||||
7844 | |||||||||
7845 | return TCCheckBlock; | ||||||||
7846 | } | ||||||||
7847 | |||||||||
7848 | //===--------------------------------------------------------------------===// | ||||||||
7849 | // EpilogueVectorizerEpilogueLoop | ||||||||
7850 | //===--------------------------------------------------------------------===// | ||||||||
7851 | |||||||||
7852 | /// This function is partially responsible for generating the control flow | ||||||||
7853 | /// depicted in https://llvm.org/docs/Vectorizers.html#epilogue-vectorization. | ||||||||
7854 | std::pair<BasicBlock *, Value *> | ||||||||
7855 | EpilogueVectorizerEpilogueLoop::createEpilogueVectorizedLoopSkeleton() { | ||||||||
7856 | createVectorLoopSkeleton("vec.epilog."); | ||||||||
7857 | |||||||||
7858 | // Now, compare the remaining count and if there aren't enough iterations to | ||||||||
7859 | // execute the vectorized epilogue skip to the scalar part. | ||||||||
7860 | BasicBlock *VecEpilogueIterationCountCheck = LoopVectorPreHeader; | ||||||||
7861 | VecEpilogueIterationCountCheck->setName("vec.epilog.iter.check"); | ||||||||
7862 | LoopVectorPreHeader = | ||||||||
7863 | SplitBlock(LoopVectorPreHeader, LoopVectorPreHeader->getTerminator(), DT, | ||||||||
7864 | LI, nullptr, "vec.epilog.ph"); | ||||||||
7865 | emitMinimumVectorEpilogueIterCountCheck(LoopScalarPreHeader, | ||||||||
7866 | VecEpilogueIterationCountCheck); | ||||||||
7867 | |||||||||
7868 | // Adjust the control flow taking the state info from the main loop | ||||||||
7869 | // vectorization into account. | ||||||||
7870 | 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", 7871, __extension__ __PRETTY_FUNCTION__)) | ||||||||
7871 | "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", 7871, __extension__ __PRETTY_FUNCTION__)); | ||||||||
7872 | EPI.MainLoopIterationCountCheck->getTerminator()->replaceUsesOfWith( | ||||||||
7873 | VecEpilogueIterationCountCheck, LoopVectorPreHeader); | ||||||||
7874 | |||||||||
7875 | DT->changeImmediateDominator(LoopVectorPreHeader, | ||||||||
7876 | EPI.MainLoopIterationCountCheck); | ||||||||
7877 | |||||||||
7878 | EPI.EpilogueIterationCountCheck->getTerminator()->replaceUsesOfWith( | ||||||||
7879 | VecEpilogueIterationCountCheck, LoopScalarPreHeader); | ||||||||
7880 | |||||||||
7881 | if (EPI.SCEVSafetyCheck) | ||||||||
7882 | EPI.SCEVSafetyCheck->getTerminator()->replaceUsesOfWith( | ||||||||
7883 | VecEpilogueIterationCountCheck, LoopScalarPreHeader); | ||||||||
7884 | if (EPI.MemSafetyCheck) | ||||||||
7885 | EPI.MemSafetyCheck->getTerminator()->replaceUsesOfWith( | ||||||||
7886 | VecEpilogueIterationCountCheck, LoopScalarPreHeader); | ||||||||
7887 | |||||||||
7888 | DT->changeImmediateDominator( | ||||||||
7889 | VecEpilogueIterationCountCheck, | ||||||||
7890 | VecEpilogueIterationCountCheck->getSinglePredecessor()); | ||||||||
7891 | |||||||||
7892 | DT->changeImmediateDominator(LoopScalarPreHeader, | ||||||||
7893 | EPI.EpilogueIterationCountCheck); | ||||||||
7894 | if (!Cost->requiresScalarEpilogue(EPI.EpilogueVF)) | ||||||||
7895 | // If there is an epilogue which must run, there's no edge from the | ||||||||
7896 | // middle block to exit blocks and thus no need to update the immediate | ||||||||
7897 | // dominator of the exit blocks. | ||||||||
7898 | DT->changeImmediateDominator(LoopExitBlock, | ||||||||
7899 | EPI.EpilogueIterationCountCheck); | ||||||||
7900 | |||||||||
7901 | // Keep track of bypass blocks, as they feed start values to the induction and | ||||||||
7902 | // reduction phis in the scalar loop preheader. | ||||||||
7903 | if (EPI.SCEVSafetyCheck) | ||||||||
7904 | LoopBypassBlocks.push_back(EPI.SCEVSafetyCheck); | ||||||||
7905 | if (EPI.MemSafetyCheck) | ||||||||
7906 | LoopBypassBlocks.push_back(EPI.MemSafetyCheck); | ||||||||
7907 | LoopBypassBlocks.push_back(EPI.EpilogueIterationCountCheck); | ||||||||
7908 | |||||||||
7909 | // The vec.epilog.iter.check block may contain Phi nodes from inductions or | ||||||||
7910 | // reductions which merge control-flow from the latch block and the middle | ||||||||
7911 | // block. Update the incoming values here and move the Phi into the preheader. | ||||||||
7912 | SmallVector<PHINode *, 4> PhisInBlock; | ||||||||
7913 | for (PHINode &Phi : VecEpilogueIterationCountCheck->phis()) | ||||||||
7914 | PhisInBlock.push_back(&Phi); | ||||||||
7915 | |||||||||
7916 | for (PHINode *Phi : PhisInBlock) { | ||||||||
7917 | Phi->moveBefore(LoopVectorPreHeader->getFirstNonPHI()); | ||||||||
7918 | Phi->replaceIncomingBlockWith( | ||||||||
7919 | VecEpilogueIterationCountCheck->getSinglePredecessor(), | ||||||||
7920 | VecEpilogueIterationCountCheck); | ||||||||
7921 | |||||||||
7922 | // If the phi doesn't have an incoming value from the | ||||||||
7923 | // EpilogueIterationCountCheck, we are done. Otherwise remove the incoming | ||||||||
7924 | // value and also those from other check blocks. This is needed for | ||||||||
7925 | // reduction phis only. | ||||||||
7926 | if (none_of(Phi->blocks(), [&](BasicBlock *IncB) { | ||||||||
7927 | return EPI.EpilogueIterationCountCheck == IncB; | ||||||||
7928 | })) | ||||||||
7929 | continue; | ||||||||
7930 | Phi->removeIncomingValue(EPI.EpilogueIterationCountCheck); | ||||||||
7931 | if (EPI.SCEVSafetyCheck) | ||||||||
7932 | Phi->removeIncomingValue(EPI.SCEVSafetyCheck); | ||||||||
7933 | if (EPI.MemSafetyCheck) | ||||||||
7934 | Phi->removeIncomingValue(EPI.MemSafetyCheck); | ||||||||
7935 | } | ||||||||
7936 | |||||||||
7937 | // Generate a resume induction for the vector epilogue and put it in the | ||||||||
7938 | // vector epilogue preheader | ||||||||
7939 | Type *IdxTy = Legal->getWidestInductionType(); | ||||||||
7940 | PHINode *EPResumeVal = PHINode::Create(IdxTy, 2, "vec.epilog.resume.val", | ||||||||
7941 | LoopVectorPreHeader->getFirstNonPHI()); | ||||||||
7942 | EPResumeVal->addIncoming(EPI.VectorTripCount, VecEpilogueIterationCountCheck); | ||||||||
7943 | EPResumeVal->addIncoming(ConstantInt::get(IdxTy, 0), | ||||||||
7944 | EPI.MainLoopIterationCountCheck); | ||||||||
7945 | |||||||||
7946 | // Generate induction resume values. These variables save the new starting | ||||||||
7947 | // indexes for the scalar loop. They are used to test if there are any tail | ||||||||
7948 | // iterations left once the vector loop has completed. | ||||||||
7949 | // Note that when the vectorized epilogue is skipped due to iteration count | ||||||||
7950 | // check, then the resume value for the induction variable comes from | ||||||||
7951 | // the trip count of the main vector loop, hence passing the AdditionalBypass | ||||||||
7952 | // argument. | ||||||||
7953 | createInductionResumeValues({VecEpilogueIterationCountCheck, | ||||||||
7954 | EPI.VectorTripCount} /* AdditionalBypass */); | ||||||||
7955 | |||||||||
7956 | return {completeLoopSkeleton(), EPResumeVal}; | ||||||||
7957 | } | ||||||||
7958 | |||||||||
7959 | BasicBlock * | ||||||||
7960 | EpilogueVectorizerEpilogueLoop::emitMinimumVectorEpilogueIterCountCheck( | ||||||||
7961 | BasicBlock *Bypass, BasicBlock *Insert) { | ||||||||
7962 | |||||||||
7963 | 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", 7964, __extension__ __PRETTY_FUNCTION__)) | ||||||||
7964 | "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", 7964, __extension__ __PRETTY_FUNCTION__)); | ||||||||
7965 | 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", 7968, __extension__ __PRETTY_FUNCTION__)) | ||||||||
7966 | (!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", 7968, __extension__ __PRETTY_FUNCTION__)) | ||||||||
7967 | 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", 7968, __extension__ __PRETTY_FUNCTION__)) | ||||||||
7968 | "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", 7968, __extension__ __PRETTY_FUNCTION__)); | ||||||||
7969 | Value *TC = EPI.TripCount; | ||||||||
7970 | IRBuilder<> Builder(Insert->getTerminator()); | ||||||||
7971 | Value *Count = Builder.CreateSub(TC, EPI.VectorTripCount, "n.vec.remaining"); | ||||||||
7972 | |||||||||
7973 | // Generate code to check if the loop's trip count is less than VF * UF of the | ||||||||
7974 | // vector epilogue loop. | ||||||||
7975 | auto P = Cost->requiresScalarEpilogue(EPI.EpilogueVF) ? | ||||||||
7976 | ICmpInst::ICMP_ULE : ICmpInst::ICMP_ULT; | ||||||||
7977 | |||||||||
7978 | Value *CheckMinIters = | ||||||||
7979 | Builder.CreateICmp(P, Count, | ||||||||
7980 | createStepForVF(Builder, Count->getType(), | ||||||||
7981 | EPI.EpilogueVF, EPI.EpilogueUF), | ||||||||
7982 | "min.epilog.iters.check"); | ||||||||
7983 | |||||||||
7984 | ReplaceInstWithInst( | ||||||||
7985 | Insert->getTerminator(), | ||||||||
7986 | BranchInst::Create(Bypass, LoopVectorPreHeader, CheckMinIters)); | ||||||||
7987 | |||||||||
7988 | LoopBypassBlocks.push_back(Insert); | ||||||||
7989 | return Insert; | ||||||||
7990 | } | ||||||||
7991 | |||||||||
7992 | void EpilogueVectorizerEpilogueLoop::printDebugTracesAtStart() { | ||||||||
7993 | 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) | ||||||||
7994 | 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) | ||||||||
7995 | << "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) | ||||||||
7996 | << ", 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) | ||||||||
7997 | })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); | ||||||||
7998 | } | ||||||||
7999 | |||||||||
8000 | void EpilogueVectorizerEpilogueLoop::printDebugTracesAtEnd() { | ||||||||
8001 | DEBUG_WITH_TYPE(VerboseDebug, {do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType (VerboseDebug)) { { dbgs() << "final fn:\n" << *OrigLoop ->getHeader()->getParent() << "\n"; }; } } while ( false) | ||||||||
8002 | 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) | ||||||||
8003 | })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType (VerboseDebug)) { { dbgs() << "final fn:\n" << *OrigLoop ->getHeader()->getParent() << "\n"; }; } } while ( false); | ||||||||
8004 | } | ||||||||
8005 | |||||||||
8006 | bool LoopVectorizationPlanner::getDecisionAndClampRange( | ||||||||
8007 | const std::function<bool(ElementCount)> &Predicate, VFRange &Range) { | ||||||||
8008 | 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", 8008, __extension__ __PRETTY_FUNCTION__)); | ||||||||
8009 | bool PredicateAtRangeStart = Predicate(Range.Start); | ||||||||
8010 | |||||||||
8011 | for (ElementCount TmpVF = Range.Start * 2; | ||||||||
8012 | ElementCount::isKnownLT(TmpVF, Range.End); TmpVF *= 2) | ||||||||
8013 | if (Predicate(TmpVF) != PredicateAtRangeStart) { | ||||||||
8014 | Range.End = TmpVF; | ||||||||
8015 | break; | ||||||||
8016 | } | ||||||||
8017 | |||||||||
8018 | return PredicateAtRangeStart; | ||||||||
8019 | } | ||||||||
8020 | |||||||||
8021 | /// Build VPlans for the full range of feasible VF's = {\p MinVF, 2 * \p MinVF, | ||||||||
8022 | /// 4 * \p MinVF, ..., \p MaxVF} by repeatedly building a VPlan for a sub-range | ||||||||
8023 | /// of VF's starting at a given VF and extending it as much as possible. Each | ||||||||
8024 | /// vectorization decision can potentially shorten this sub-range during | ||||||||
8025 | /// buildVPlan(). | ||||||||
8026 | void LoopVectorizationPlanner::buildVPlans(ElementCount MinVF, | ||||||||
8027 | ElementCount MaxVF) { | ||||||||
8028 | auto MaxVFPlusOne = MaxVF.getWithIncrement(1); | ||||||||
8029 | for (ElementCount VF = MinVF; ElementCount::isKnownLT(VF, MaxVFPlusOne);) { | ||||||||
8030 | VFRange SubRange = {VF, MaxVFPlusOne}; | ||||||||
8031 | VPlans.push_back(buildVPlan(SubRange)); | ||||||||
8032 | VF = SubRange.End; | ||||||||
8033 | } | ||||||||
8034 | } | ||||||||
8035 | |||||||||
8036 | VPValue *VPRecipeBuilder::createEdgeMask(BasicBlock *Src, BasicBlock *Dst, | ||||||||
8037 | VPlanPtr &Plan) { | ||||||||
8038 | 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", 8038, __extension__ __PRETTY_FUNCTION__)); | ||||||||
8039 | |||||||||
8040 | // Look for cached value. | ||||||||
8041 | std::pair<BasicBlock *, BasicBlock *> Edge(Src, Dst); | ||||||||
8042 | EdgeMaskCacheTy::iterator ECEntryIt = EdgeMaskCache.find(Edge); | ||||||||
8043 | if (ECEntryIt != EdgeMaskCache.end()) | ||||||||
8044 | return ECEntryIt->second; | ||||||||
8045 | |||||||||
8046 | VPValue *SrcMask = createBlockInMask(Src, Plan); | ||||||||
8047 | |||||||||
8048 | // The terminator has to be a branch inst! | ||||||||
8049 | BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator()); | ||||||||
8050 | 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", 8050, __extension__ __PRETTY_FUNCTION__)); | ||||||||
8051 | |||||||||
8052 | if (!BI->isConditional() || BI->getSuccessor(0) == BI->getSuccessor(1)) | ||||||||
8053 | return EdgeMaskCache[Edge] = SrcMask; | ||||||||
8054 | |||||||||
8055 | // If source is an exiting block, we know the exit edge is dynamically dead | ||||||||
8056 | // in the vector loop, and thus we don't need to restrict the mask. Avoid | ||||||||
8057 | // adding uses of an otherwise potentially dead instruction. | ||||||||
8058 | if (OrigLoop->isLoopExiting(Src)) | ||||||||
8059 | return EdgeMaskCache[Edge] = SrcMask; | ||||||||
8060 | |||||||||
8061 | VPValue *EdgeMask = Plan->getOrAddVPValue(BI->getCondition()); | ||||||||
8062 | 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", 8062, __extension__ __PRETTY_FUNCTION__)); | ||||||||
8063 | |||||||||
8064 | if (BI->getSuccessor(0) != Dst) | ||||||||
8065 | EdgeMask = Builder.createNot(EdgeMask, BI->getDebugLoc()); | ||||||||
8066 | |||||||||
8067 | if (SrcMask) { // Otherwise block in-mask is all-one, no need to AND. | ||||||||
8068 | // The condition is 'SrcMask && EdgeMask', which is equivalent to | ||||||||
8069 | // 'select i1 SrcMask, i1 EdgeMask, i1 false'. | ||||||||
8070 | // The select version does not introduce new UB if SrcMask is false and | ||||||||
8071 | // EdgeMask is poison. Using 'and' here introduces undefined behavior. | ||||||||
8072 | VPValue *False = Plan->getOrAddVPValue( | ||||||||
8073 | ConstantInt::getFalse(BI->getCondition()->getType())); | ||||||||
8074 | EdgeMask = | ||||||||
8075 | Builder.createSelect(SrcMask, EdgeMask, False, BI->getDebugLoc()); | ||||||||
8076 | } | ||||||||
8077 | |||||||||
8078 | return EdgeMaskCache[Edge] = EdgeMask; | ||||||||
8079 | } | ||||||||
8080 | |||||||||
8081 | VPValue *VPRecipeBuilder::createBlockInMask(BasicBlock *BB, VPlanPtr &Plan) { | ||||||||
8082 | 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", 8082, __extension__ __PRETTY_FUNCTION__)); | ||||||||
8083 | |||||||||
8084 | // Look for cached value. | ||||||||
8085 | BlockMaskCacheTy::iterator BCEntryIt = BlockMaskCache.find(BB); | ||||||||
8086 | if (BCEntryIt != BlockMaskCache.end()) | ||||||||
8087 | return BCEntryIt->second; | ||||||||
8088 | |||||||||
8089 | // All-one mask is modelled as no-mask following the convention for masked | ||||||||
8090 | // load/store/gather/scatter. Initialize BlockMask to no-mask. | ||||||||
8091 | VPValue *BlockMask = nullptr; | ||||||||
8092 | |||||||||
8093 | if (OrigLoop->getHeader() == BB) { | ||||||||
8094 | if (!CM.blockNeedsPredicationForAnyReason(BB)) | ||||||||
8095 | return BlockMaskCache[BB] = BlockMask; // Loop incoming mask is all-one. | ||||||||
8096 | |||||||||
8097 | 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", 8097, __extension__ __PRETTY_FUNCTION__)); | ||||||||
8098 | |||||||||
8099 | // If we're using the active lane mask for control flow, then we get the | ||||||||
8100 | // mask from the active lane mask PHI that is cached in the VPlan. | ||||||||
8101 | PredicationStyle EmitGetActiveLaneMask = CM.TTI.emitGetActiveLaneMask(); | ||||||||
8102 | if (EmitGetActiveLaneMask == PredicationStyle::DataAndControlFlow) | ||||||||
8103 | return BlockMaskCache[BB] = Plan->getActiveLaneMaskPhi(); | ||||||||
8104 | |||||||||
8105 | // Introduce the early-exit compare IV <= BTC to form header block mask. | ||||||||
8106 | // This is used instead of IV < TC because TC may wrap, unlike BTC. Start by | ||||||||
8107 | // constructing the desired canonical IV in the header block as its first | ||||||||
8108 | // non-phi instructions. | ||||||||
8109 | |||||||||
8110 | VPBasicBlock *HeaderVPBB = | ||||||||
8111 | Plan->getVectorLoopRegion()->getEntryBasicBlock(); | ||||||||
8112 | auto NewInsertionPoint = HeaderVPBB->getFirstNonPhi(); | ||||||||
8113 | auto *IV = new VPWidenCanonicalIVRecipe(Plan->getCanonicalIV()); | ||||||||
8114 | HeaderVPBB->insert(IV, HeaderVPBB->getFirstNonPhi()); | ||||||||
8115 | |||||||||
8116 | VPBuilder::InsertPointGuard Guard(Builder); | ||||||||
8117 | Builder.setInsertPoint(HeaderVPBB, NewInsertionPoint); | ||||||||
8118 | if (EmitGetActiveLaneMask != PredicationStyle::None) { | ||||||||
8119 | VPValue *TC = Plan->getOrCreateTripCount(); | ||||||||
8120 | BlockMask = Builder.createNaryOp(VPInstruction::ActiveLaneMask, {IV, TC}, | ||||||||
8121 | nullptr, "active.lane.mask"); | ||||||||
8122 | } else { | ||||||||
8123 | VPValue *BTC = Plan->getOrCreateBackedgeTakenCount(); | ||||||||
8124 | BlockMask = Builder.createNaryOp(VPInstruction::ICmpULE, {IV, BTC}); | ||||||||
8125 | } | ||||||||
8126 | return BlockMaskCache[BB] = BlockMask; | ||||||||
8127 | } | ||||||||
8128 | |||||||||
8129 | // This is the block mask. We OR all incoming edges. | ||||||||
8130 | for (auto *Predecessor : predecessors(BB)) { | ||||||||
8131 | VPValue *EdgeMask = createEdgeMask(Predecessor, BB, Plan); | ||||||||
8132 | if (!EdgeMask) // Mask of predecessor is all-one so mask of block is too. | ||||||||
8133 | return BlockMaskCache[BB] = EdgeMask; | ||||||||
| |||||||||
8134 | |||||||||
8135 | if (!BlockMask
| ||||||||
8136 | BlockMask = EdgeMask; | ||||||||
8137 | continue; | ||||||||
8138 | } | ||||||||
8139 | |||||||||
8140 | BlockMask = Builder.createOr(BlockMask, EdgeMask, {}); | ||||||||
8141 | } | ||||||||
8142 | |||||||||
8143 | return BlockMaskCache[BB] = BlockMask; | ||||||||
8144 | } | ||||||||
8145 | |||||||||
8146 | VPRecipeBase *VPRecipeBuilder::tryToWidenMemory(Instruction *I, | ||||||||
8147 | ArrayRef<VPValue *> Operands, | ||||||||
8148 | VFRange &Range, | ||||||||
8149 | VPlanPtr &Plan) { | ||||||||
8150 | 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", 8151, __extension__ __PRETTY_FUNCTION__)) | ||||||||
8151 | "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", 8151, __extension__ __PRETTY_FUNCTION__)); | ||||||||
8152 | |||||||||
8153 | auto willWiden = [&](ElementCount VF) -> bool { | ||||||||
8154 | LoopVectorizationCostModel::InstWidening Decision = | ||||||||
8155 | CM.getWideningDecision(I, VF); | ||||||||
8156 | 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", 8157, __extension__ __PRETTY_FUNCTION__)) | ||||||||
8157 | "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", 8157, __extension__ __PRETTY_FUNCTION__)); | ||||||||
8158 | if (Decision == LoopVectorizationCostModel::CM_Interleave) | ||||||||
8159 | return true; | ||||||||
8160 | if (CM.isScalarAfterVectorization(I, VF) || | ||||||||
8161 | CM.isProfitableToScalarize(I, VF)) | ||||||||
8162 | return false; | ||||||||
8163 | return Decision != LoopVectorizationCostModel::CM_Scalarize; | ||||||||
8164 | }; | ||||||||
8165 | |||||||||
8166 | if (!LoopVectorizationPlanner::getDecisionAndClampRange(willWiden, Range)) | ||||||||
8167 | return nullptr; | ||||||||
8168 | |||||||||
8169 | VPValue *Mask = nullptr; | ||||||||
8170 | if (Legal->isMaskRequired(I)) | ||||||||
8171 | Mask = createBlockInMask(I->getParent(), Plan); | ||||||||
8172 | |||||||||
8173 | // Determine if the pointer operand of the access is either consecutive or | ||||||||
8174 | // reverse consecutive. | ||||||||
8175 | LoopVectorizationCostModel::InstWidening Decision = | ||||||||
8176 | CM.getWideningDecision(I, Range.Start); | ||||||||
8177 | bool Reverse = Decision == LoopVectorizationCostModel::CM_Widen_Reverse; | ||||||||
8178 | bool Consecutive = | ||||||||
8179 | Reverse || Decision == LoopVectorizationCostModel::CM_Widen; | ||||||||
8180 | |||||||||
8181 | if (LoadInst *Load = dyn_cast<LoadInst>(I)) | ||||||||
8182 | return new VPWidenMemoryInstructionRecipe(*Load, Operands[0], Mask, | ||||||||
8183 | Consecutive, Reverse); | ||||||||
8184 | |||||||||
8185 | StoreInst *Store = cast<StoreInst>(I); | ||||||||
8186 | return new VPWidenMemoryInstructionRecipe(*Store, Operands[1], Operands[0], | ||||||||
8187 | Mask, Consecutive, Reverse); | ||||||||
8188 | } | ||||||||
8189 | |||||||||
8190 | /// Creates a VPWidenIntOrFpInductionRecpipe for \p Phi. If needed, it will also | ||||||||
8191 | /// insert a recipe to expand the step for the induction recipe. | ||||||||
8192 | static VPWidenIntOrFpInductionRecipe *createWidenInductionRecipes( | ||||||||
8193 | PHINode *Phi, Instruction *PhiOrTrunc, VPValue *Start, | ||||||||
8194 | const InductionDescriptor &IndDesc, LoopVectorizationCostModel &CM, | ||||||||
8195 | VPlan &Plan, ScalarEvolution &SE, Loop &OrigLoop, VFRange &Range) { | ||||||||
8196 | // Returns true if an instruction \p I should be scalarized instead of | ||||||||
8197 | // vectorized for the chosen vectorization factor. | ||||||||
8198 | auto ShouldScalarizeInstruction = [&CM](Instruction *I, ElementCount VF) { | ||||||||
8199 | return CM.isScalarAfterVectorization(I, VF) || | ||||||||
8200 | CM.isProfitableToScalarize(I, VF); | ||||||||
8201 | }; | ||||||||
8202 | |||||||||
8203 | bool NeedsScalarIVOnly = LoopVectorizationPlanner::getDecisionAndClampRange( | ||||||||
8204 | [&](ElementCount VF) { | ||||||||
8205 | return ShouldScalarizeInstruction(PhiOrTrunc, VF); | ||||||||
8206 | }, | ||||||||
8207 | Range); | ||||||||
8208 | 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", 8209, __extension__ __PRETTY_FUNCTION__)) | ||||||||
8209 | 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", 8209, __extension__ __PRETTY_FUNCTION__)); | ||||||||
8210 | 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", 8211, __extension__ __PRETTY_FUNCTION__)) | ||||||||
8211 | "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", 8211, __extension__ __PRETTY_FUNCTION__)); | ||||||||
8212 | |||||||||
8213 | VPValue *Step = | ||||||||
8214 | vputils::getOrCreateVPValueForSCEVExpr(Plan, IndDesc.getStep(), SE); | ||||||||
8215 | if (auto *TruncI = dyn_cast<TruncInst>(PhiOrTrunc)) { | ||||||||
8216 | return new VPWidenIntOrFpInductionRecipe(Phi, Start, Step, IndDesc, TruncI, | ||||||||
8217 | !NeedsScalarIVOnly); | ||||||||
8218 | } | ||||||||
8219 | 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", 8219, __extension__ __PRETTY_FUNCTION__)); | ||||||||
8220 | return new VPWidenIntOrFpInductionRecipe(Phi, Start, Step, IndDesc, | ||||||||
8221 | !NeedsScalarIVOnly); | ||||||||
8222 | } | ||||||||
8223 | |||||||||
8224 | VPRecipeBase *VPRecipeBuilder::tryToOptimizeInductionPHI( | ||||||||
8225 | PHINode *Phi, ArrayRef<VPValue *> Operands, VPlan &Plan, VFRange &Range) { | ||||||||
8226 | |||||||||
8227 | // Check if this is an integer or fp induction. If so, build the recipe that | ||||||||
8228 | // produces its scalar and vector values. | ||||||||
8229 | if (auto *II = Legal->getIntOrFpInductionDescriptor(Phi)) | ||||||||
8230 | return createWidenInductionRecipes(Phi, Phi, Operands[0], *II, CM, Plan, | ||||||||
8231 | *PSE.getSE(), *OrigLoop, Range); | ||||||||
8232 | |||||||||
8233 | // Check if this is pointer induction. If so, build the recipe for it. | ||||||||
8234 | if (auto *II = Legal->getPointerInductionDescriptor(Phi)) { | ||||||||
8235 | VPValue *Step = vputils::getOrCreateVPValueForSCEVExpr(Plan, II->getStep(), | ||||||||
8236 | *PSE.getSE()); | ||||||||
8237 | assert(isa<SCEVConstant>(II->getStep()))(static_cast <bool> (isa<SCEVConstant>(II->getStep ())) ? void (0) : __assert_fail ("isa<SCEVConstant>(II->getStep())" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8237, __extension__ __PRETTY_FUNCTION__)); | ||||||||
8238 | return new VPWidenPointerInductionRecipe( | ||||||||
8239 | Phi, Operands[0], Step, *II, | ||||||||
8240 | LoopVectorizationPlanner::getDecisionAndClampRange( | ||||||||
8241 | [&](ElementCount VF) { | ||||||||
8242 | return CM.isScalarAfterVectorization(Phi, VF); | ||||||||
8243 | }, | ||||||||
8244 | Range)); | ||||||||
8245 | } | ||||||||
8246 | return nullptr; | ||||||||
8247 | } | ||||||||
8248 | |||||||||
8249 | VPWidenIntOrFpInductionRecipe *VPRecipeBuilder::tryToOptimizeInductionTruncate( | ||||||||
8250 | TruncInst *I, ArrayRef<VPValue *> Operands, VFRange &Range, VPlan &Plan) { | ||||||||
8251 | // Optimize the special case where the source is a constant integer | ||||||||
8252 | // induction variable. Notice that we can only optimize the 'trunc' case | ||||||||
8253 | // because (a) FP conversions lose precision, (b) sext/zext may wrap, and | ||||||||
8254 | // (c) other casts depend on pointer size. | ||||||||
8255 | |||||||||
8256 | // Determine whether \p K is a truncation based on an induction variable that | ||||||||
8257 | // can be optimized. | ||||||||
8258 | auto isOptimizableIVTruncate = | ||||||||
8259 | [&](Instruction *K) -> std::function<bool(ElementCount)> { | ||||||||
8260 | return [=](ElementCount VF) -> bool { | ||||||||
8261 | return CM.isOptimizableIVTruncate(K, VF); | ||||||||
8262 | }; | ||||||||
8263 | }; | ||||||||
8264 | |||||||||
8265 | if (LoopVectorizationPlanner::getDecisionAndClampRange( | ||||||||
8266 | isOptimizableIVTruncate(I), Range)) { | ||||||||
8267 | |||||||||
8268 | auto *Phi = cast<PHINode>(I->getOperand(0)); | ||||||||
8269 | const InductionDescriptor &II = *Legal->getIntOrFpInductionDescriptor(Phi); | ||||||||
8270 | VPValue *Start = Plan.getOrAddVPValue(II.getStartValue()); | ||||||||
8271 | return createWidenInductionRecipes(Phi, I, Start, II, CM, Plan, | ||||||||
8272 | *PSE.getSE(), *OrigLoop, Range); | ||||||||
8273 | } | ||||||||
8274 | return nullptr; | ||||||||
8275 | } | ||||||||
8276 | |||||||||
8277 | VPRecipeOrVPValueTy VPRecipeBuilder::tryToBlend(PHINode *Phi, | ||||||||
8278 | ArrayRef<VPValue *> Operands, | ||||||||
8279 | VPlanPtr &Plan) { | ||||||||
8280 | // If all incoming values are equal, the incoming VPValue can be used directly | ||||||||
8281 | // instead of creating a new VPBlendRecipe. | ||||||||
8282 | if (llvm::all_equal(Operands)) | ||||||||
8283 | return Operands[0]; | ||||||||
8284 | |||||||||
8285 | unsigned NumIncoming = Phi->getNumIncomingValues(); | ||||||||
8286 | // For in-loop reductions, we do not need to create an additional select. | ||||||||
8287 | VPValue *InLoopVal = nullptr; | ||||||||
8288 | for (unsigned In = 0; In < NumIncoming; In++) { | ||||||||
8289 | PHINode *PhiOp = | ||||||||
8290 | dyn_cast_or_null<PHINode>(Operands[In]->getUnderlyingValue()); | ||||||||
8291 | if (PhiOp && CM.isInLoopReduction(PhiOp)) { | ||||||||
8292 | 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", 8292, __extension__ __PRETTY_FUNCTION__)); | ||||||||
8293 | InLoopVal = Operands[In]; | ||||||||
8294 | } | ||||||||
8295 | } | ||||||||
8296 | |||||||||
8297 | 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", 8299, __extension__ __PRETTY_FUNCTION__)) | ||||||||
8298 | "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", 8299, __extension__ __PRETTY_FUNCTION__)) | ||||||||
8299 | "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", 8299, __extension__ __PRETTY_FUNCTION__)); | ||||||||
8300 | if (InLoopVal) | ||||||||
8301 | return Operands[Operands[0] == InLoopVal ? 1 : 0]; | ||||||||
8302 | |||||||||
8303 | // We know that all PHIs in non-header blocks are converted into selects, so | ||||||||
8304 | // we don't have to worry about the insertion order and we can just use the | ||||||||
8305 | // builder. At this point we generate the predication tree. There may be | ||||||||
8306 | // duplications since this is a simple recursive scan, but future | ||||||||
8307 | // optimizations will clean it up. | ||||||||
8308 | SmallVector<VPValue *, 2> OperandsWithMask; | ||||||||
8309 | |||||||||
8310 | for (unsigned In = 0; In < NumIncoming; In++) { | ||||||||
8311 | VPValue *EdgeMask = | ||||||||
8312 | createEdgeMask(Phi->getIncomingBlock(In), Phi->getParent(), Plan); | ||||||||
8313 | 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", 8314, __extension__ __PRETTY_FUNCTION__)) | ||||||||
8314 | "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", 8314, __extension__ __PRETTY_FUNCTION__)); | ||||||||
8315 | OperandsWithMask.push_back(Operands[In]); | ||||||||
8316 | if (EdgeMask) | ||||||||
8317 | OperandsWithMask.push_back(EdgeMask); | ||||||||
8318 | } | ||||||||
8319 | return toVPRecipeResult(new VPBlendRecipe(Phi, OperandsWithMask)); | ||||||||
8320 | } | ||||||||
8321 | |||||||||
8322 | VPWidenCallRecipe *VPRecipeBuilder::tryToWidenCall(CallInst *CI, | ||||||||
8323 | ArrayRef<VPValue *> Operands, | ||||||||
8324 | VFRange &Range) const { | ||||||||
8325 | |||||||||
8326 | bool IsPredicated = LoopVectorizationPlanner::getDecisionAndClampRange( | ||||||||
8327 | [this, CI](ElementCount VF) { | ||||||||
8328 | return CM.isScalarWithPredication(CI, VF); | ||||||||
8329 | }, | ||||||||
8330 | Range); | ||||||||
8331 | |||||||||
8332 | if (IsPredicated) | ||||||||
8333 | return nullptr; | ||||||||
8334 | |||||||||
8335 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); | ||||||||
8336 | if (ID && (ID == Intrinsic::assume || ID == Intrinsic::lifetime_end || | ||||||||
8337 | ID == Intrinsic::lifetime_start || ID == Intrinsic::sideeffect || | ||||||||
8338 | ID == Intrinsic::pseudoprobe || | ||||||||
8339 | ID == Intrinsic::experimental_noalias_scope_decl)) | ||||||||
8340 | return nullptr; | ||||||||
8341 | |||||||||
8342 | ArrayRef<VPValue *> Ops = Operands.take_front(CI->arg_size()); | ||||||||
8343 | |||||||||
8344 | // Is it beneficial to perform intrinsic call compared to lib call? | ||||||||
8345 | bool ShouldUseVectorIntrinsic = | ||||||||
8346 | ID && LoopVectorizationPlanner::getDecisionAndClampRange( | ||||||||
8347 | [&](ElementCount VF) -> bool { | ||||||||
8348 | bool NeedToScalarize = false; | ||||||||
8349 | // Is it beneficial to perform intrinsic call compared to lib | ||||||||
8350 | // call? | ||||||||
8351 | InstructionCost CallCost = | ||||||||
8352 | CM.getVectorCallCost(CI, VF, NeedToScalarize); | ||||||||
8353 | InstructionCost IntrinsicCost = | ||||||||
8354 | CM.getVectorIntrinsicCost(CI, VF); | ||||||||
8355 | return IntrinsicCost <= CallCost; | ||||||||
8356 | }, | ||||||||
8357 | Range); | ||||||||
8358 | if (ShouldUseVectorIntrinsic) | ||||||||
8359 | return new VPWidenCallRecipe(*CI, make_range(Ops.begin(), Ops.end()), ID); | ||||||||
8360 | |||||||||
8361 | // Is better to call a vectorized version of the function than to to scalarize | ||||||||
8362 | // the call? | ||||||||
8363 | auto ShouldUseVectorCall = LoopVectorizationPlanner::getDecisionAndClampRange( | ||||||||
8364 | [&](ElementCount VF) -> bool { | ||||||||
8365 | // The following case may be scalarized depending on the VF. | ||||||||
8366 | // The flag shows whether we can use a usual Call for vectorized | ||||||||
8367 | // version of the instruction. | ||||||||
8368 | bool NeedToScalarize = false; | ||||||||
8369 | CM.getVectorCallCost(CI, VF, NeedToScalarize); | ||||||||
8370 | return !NeedToScalarize; | ||||||||
8371 | }, | ||||||||
8372 | Range); | ||||||||
8373 | if (ShouldUseVectorCall) | ||||||||
8374 | return new VPWidenCallRecipe(*CI, make_range(Ops.begin(), Ops.end()), | ||||||||
8375 | Intrinsic::not_intrinsic); | ||||||||
8376 | |||||||||
8377 | return nullptr; | ||||||||
8378 | } | ||||||||
8379 | |||||||||
8380 | bool VPRecipeBuilder::shouldWiden(Instruction *I, VFRange &Range) const { | ||||||||
8381 | 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", 8382, __extension__ __PRETTY_FUNCTION__)) | ||||||||
8382 | !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", 8382, __extension__ __PRETTY_FUNCTION__)); | ||||||||
8383 | // Instruction should be widened, unless it is scalar after vectorization, | ||||||||
8384 | // scalarization is profitable or it is predicated. | ||||||||
8385 | auto WillScalarize = [this, I](ElementCount VF) -> bool { | ||||||||
8386 | return CM.isScalarAfterVectorization(I, VF) || | ||||||||
8387 | CM.isProfitableToScalarize(I, VF) || | ||||||||
8388 | CM.isScalarWithPredication(I, VF); | ||||||||
8389 | }; | ||||||||
8390 | return !LoopVectorizationPlanner::getDecisionAndClampRange(WillScalarize, | ||||||||
8391 | Range); | ||||||||
8392 | } | ||||||||
8393 | |||||||||
8394 | VPRecipeBase *VPRecipeBuilder::tryToWiden(Instruction *I, | ||||||||
8395 | ArrayRef<VPValue *> Operands, | ||||||||
8396 | VPBasicBlock *VPBB, VPlanPtr &Plan) { | ||||||||
8397 | switch (I->getOpcode()) { | ||||||||
8398 | default: | ||||||||
8399 | return nullptr; | ||||||||
8400 | case Instruction::SDiv: | ||||||||
8401 | case Instruction::UDiv: | ||||||||
8402 | case Instruction::SRem: | ||||||||
8403 | case Instruction::URem: { | ||||||||
8404 | // If not provably safe, use a select to form a safe divisor before widening the | ||||||||
8405 | // div/rem operation itself. Otherwise fall through to general handling below. | ||||||||
8406 | if (CM.isPredicatedInst(I)) { | ||||||||
8407 | SmallVector<VPValue *> Ops(Operands.begin(), Operands.end()); | ||||||||
8408 | VPValue *Mask = createBlockInMask(I->getParent(), Plan); | ||||||||
8409 | VPValue *One = | ||||||||
8410 | Plan->getOrAddExternalDef(ConstantInt::get(I->getType(), 1u, false)); | ||||||||
8411 | auto *SafeRHS = | ||||||||
8412 | new VPInstruction(Instruction::Select, {Mask, Ops[1], One}, | ||||||||
8413 | I->getDebugLoc()); | ||||||||
8414 | VPBB->appendRecipe(SafeRHS); | ||||||||
8415 | Ops[1] = SafeRHS; | ||||||||
8416 | return new VPWidenRecipe(*I, make_range(Ops.begin(), Ops.end())); | ||||||||
8417 | } | ||||||||
8418 | LLVM_FALLTHROUGH[[fallthrough]]; | ||||||||
8419 | } | ||||||||
8420 | case Instruction::Add: | ||||||||
8421 | case Instruction::And: | ||||||||
8422 | case Instruction::AShr: | ||||||||
8423 | case Instruction::BitCast: | ||||||||
8424 | case Instruction::FAdd: | ||||||||
8425 | case Instruction::FCmp: | ||||||||
8426 | case Instruction::FDiv: | ||||||||
8427 | case Instruction::FMul: | ||||||||
8428 | case Instruction::FNeg: | ||||||||
8429 | case Instruction::FPExt: | ||||||||
8430 | case Instruction::FPToSI: | ||||||||
8431 | case Instruction::FPToUI: | ||||||||
8432 | case Instruction::FPTrunc: | ||||||||
8433 | case Instruction::FRem: | ||||||||
8434 | case Instruction::FSub: | ||||||||
8435 | case Instruction::ICmp: | ||||||||
8436 | case Instruction::IntToPtr: | ||||||||
8437 | case Instruction::LShr: | ||||||||
8438 | case Instruction::Mul: | ||||||||
8439 | case Instruction::Or: | ||||||||
8440 | case Instruction::PtrToInt: | ||||||||
8441 | case Instruction::Select: | ||||||||
8442 | case Instruction::SExt: | ||||||||
8443 | case Instruction::Shl: | ||||||||
8444 | case Instruction::SIToFP: | ||||||||
8445 | case Instruction::Sub: | ||||||||
8446 | case Instruction::Trunc: | ||||||||
8447 | case Instruction::UIToFP: | ||||||||
8448 | case Instruction::Xor: | ||||||||
8449 | case Instruction::ZExt: | ||||||||
8450 | case Instruction::Freeze: | ||||||||
8451 | return new VPWidenRecipe(*I, make_range(Operands.begin(), Operands.end())); | ||||||||
8452 | }; | ||||||||
8453 | } | ||||||||
8454 | |||||||||
8455 | void VPRecipeBuilder::fixHeaderPhis() { | ||||||||
8456 | BasicBlock *OrigLatch = OrigLoop->getLoopLatch(); | ||||||||
8457 | for (VPHeaderPHIRecipe *R : PhisToFix) { | ||||||||
8458 | auto *PN = cast<PHINode>(R->getUnderlyingValue()); | ||||||||
8459 | VPRecipeBase *IncR = | ||||||||
8460 | getRecipe(cast<Instruction>(PN->getIncomingValueForBlock(OrigLatch))); | ||||||||
8461 | R->addOperand(IncR->getVPSingleValue()); | ||||||||
8462 | } | ||||||||
8463 | } | ||||||||
8464 | |||||||||
8465 | VPBasicBlock *VPRecipeBuilder::handleReplication( | ||||||||
8466 | Instruction *I, VFRange &Range, VPBasicBlock *VPBB, | ||||||||
8467 | VPlanPtr &Plan) { | ||||||||
8468 | bool IsUniform = LoopVectorizationPlanner::getDecisionAndClampRange( | ||||||||
8469 | [&](ElementCount VF) { return CM.isUniformAfterVectorization(I, VF); }, | ||||||||
8470 | Range); | ||||||||
8471 | |||||||||
8472 | bool IsPredicated = CM.isPredicatedInst(I); | ||||||||
8473 | |||||||||
8474 | // Even if the instruction is not marked as uniform, there are certain | ||||||||
8475 | // intrinsic calls that can be effectively treated as such, so we check for | ||||||||
8476 | // them here. Conservatively, we only do this for scalable vectors, since | ||||||||
8477 | // for fixed-width VFs we can always fall back on full scalarization. | ||||||||
8478 | if (!IsUniform && Range.Start.isScalable() && isa<IntrinsicInst>(I)) { | ||||||||
8479 | switch (cast<IntrinsicInst>(I)->getIntrinsicID()) { | ||||||||
8480 | case Intrinsic::assume: | ||||||||
8481 | case Intrinsic::lifetime_start: | ||||||||
8482 | case Intrinsic::lifetime_end: | ||||||||
8483 | // For scalable vectors if one of the operands is variant then we still | ||||||||
8484 | // want to mark as uniform, which will generate one instruction for just | ||||||||
8485 | // the first lane of the vector. We can't scalarize the call in the same | ||||||||
8486 | // way as for fixed-width vectors because we don't know how many lanes | ||||||||
8487 | // there are. | ||||||||
8488 | // | ||||||||
8489 | // The reasons for doing it this way for scalable vectors are: | ||||||||
8490 | // 1. For the assume intrinsic generating the instruction for the first | ||||||||
8491 | // lane is still be better than not generating any at all. For | ||||||||
8492 | // example, the input may be a splat across all lanes. | ||||||||
8493 | // 2. For the lifetime start/end intrinsics the pointer operand only | ||||||||
8494 | // does anything useful when the input comes from a stack object, | ||||||||
8495 | // which suggests it should always be uniform. For non-stack objects | ||||||||
8496 | // the effect is to poison the object, which still allows us to | ||||||||
8497 | // remove the call. | ||||||||
8498 | IsUniform = true; | ||||||||
8499 | break; | ||||||||
8500 | default: | ||||||||
8501 | break; | ||||||||
8502 | } | ||||||||
8503 | } | ||||||||
8504 | |||||||||
8505 | auto *Recipe = new VPReplicateRecipe(I, Plan->mapToVPValues(I->operands()), | ||||||||
8506 | IsUniform, IsPredicated); | ||||||||
8507 | |||||||||
8508 | // Find if I uses a predicated instruction. If so, it will use its scalar | ||||||||
8509 | // value. Avoid hoisting the insert-element which packs the scalar value into | ||||||||
8510 | // a vector value, as that happens iff all users use the vector value. | ||||||||
8511 | for (VPValue *Op : Recipe->operands()) { | ||||||||
8512 | auto *PredR = | ||||||||
8513 | dyn_cast_or_null<VPPredInstPHIRecipe>(Op->getDefiningRecipe()); | ||||||||
8514 | if (!PredR) | ||||||||
8515 | continue; | ||||||||
8516 | auto *RepR = cast<VPReplicateRecipe>( | ||||||||
8517 | PredR->getOperand(0)->getDefiningRecipe()); | ||||||||
8518 | 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", 8519, __extension__ __PRETTY_FUNCTION__)) | ||||||||
8519 | "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", 8519, __extension__ __PRETTY_FUNCTION__)); | ||||||||
8520 | RepR->setAlsoPack(false); | ||||||||
8521 | } | ||||||||
8522 | |||||||||
8523 | // Finalize the recipe for Instr, first if it is not predicated. | ||||||||
8524 | if (!IsPredicated) { | ||||||||
8525 | LLVM_DEBUG(dbgs() << "LV: Scalarizing:" << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Scalarizing:" << *I << "\n"; } } while (false); | ||||||||
8526 | setRecipe(I, Recipe); | ||||||||
8527 | Plan->addVPValue(I, Recipe); | ||||||||
8528 | VPBB->appendRecipe(Recipe); | ||||||||
8529 | return VPBB; | ||||||||
8530 | } | ||||||||
8531 | 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); | ||||||||
8532 | |||||||||
8533 | VPBlockBase *SingleSucc = VPBB->getSingleSuccessor(); | ||||||||
8534 | 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", 8535, __extension__ __PRETTY_FUNCTION__)) | ||||||||
8535 | "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", 8535, __extension__ __PRETTY_FUNCTION__)); | ||||||||
8536 | VPBlockUtils::disconnectBlocks(VPBB, SingleSucc); | ||||||||
8537 | // Record predicated instructions for above packing optimizations. | ||||||||
8538 | VPBlockBase *Region = createReplicateRegion(Recipe, Plan); | ||||||||
8539 | VPBlockUtils::insertBlockAfter(Region, VPBB); | ||||||||
8540 | auto *RegSucc = new VPBasicBlock(); | ||||||||
8541 | VPBlockUtils::insertBlockAfter(RegSucc, Region); | ||||||||
8542 | VPBlockUtils::connectBlocks(RegSucc, SingleSucc); | ||||||||
8543 | return RegSucc; | ||||||||
8544 | } | ||||||||
8545 | |||||||||
8546 | VPRegionBlock * | ||||||||
8547 | VPRecipeBuilder::createReplicateRegion(VPReplicateRecipe *PredRecipe, | ||||||||
8548 | VPlanPtr &Plan) { | ||||||||
8549 | Instruction *Instr = PredRecipe->getUnderlyingInstr(); | ||||||||
8550 | // Instructions marked for predication are replicated and placed under an | ||||||||
8551 | // if-then construct to prevent side-effects. | ||||||||
8552 | // Generate recipes to compute the block mask for this region. | ||||||||
8553 | VPValue *BlockInMask = createBlockInMask(Instr->getParent(), Plan); | ||||||||
| |||||||||
8554 | |||||||||
8555 | // Build the triangular if-then region. | ||||||||
8556 | std::string RegionName = (Twine("pred.") + Instr->getOpcodeName()).str(); | ||||||||
8557 | 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", 8557, __extension__ __PRETTY_FUNCTION__)); | ||||||||
8558 | auto *BOMRecipe = new VPBranchOnMaskRecipe(BlockInMask); | ||||||||
8559 | auto *Entry = new VPBasicBlock(Twine(RegionName) + ".entry", BOMRecipe); | ||||||||
8560 | auto *PHIRecipe = Instr->getType()->isVoidTy() | ||||||||
8561 | ? nullptr | ||||||||
8562 | : new VPPredInstPHIRecipe(PredRecipe); | ||||||||
8563 | if (PHIRecipe) { | ||||||||
8564 | setRecipe(Instr, PHIRecipe); | ||||||||
8565 | Plan->addVPValue(Instr, PHIRecipe); | ||||||||
8566 | } else { | ||||||||
8567 | setRecipe(Instr, PredRecipe); | ||||||||
8568 | Plan->addVPValue(Instr, PredRecipe); | ||||||||
8569 | } | ||||||||
8570 | |||||||||
8571 | auto *Exiting = new VPBasicBlock(Twine(RegionName) + ".continue", PHIRecipe); | ||||||||
8572 | auto *Pred = new VPBasicBlock(Twine(RegionName) + ".if", PredRecipe); | ||||||||
8573 | VPRegionBlock *Region = new VPRegionBlock(Entry, Exiting, RegionName, true); | ||||||||
8574 | |||||||||
8575 | // Note: first set Entry as region entry and then connect successors starting | ||||||||
8576 | // from it in order, to propagate the "parent" of each VPBasicBlock. | ||||||||
8577 | VPBlockUtils::insertTwoBlocksAfter(Pred, Exiting, Entry); | ||||||||
8578 | VPBlockUtils::connectBlocks(Pred, Exiting); | ||||||||
8579 | |||||||||
8580 | return Region; | ||||||||
8581 | } | ||||||||
8582 | |||||||||
8583 | VPRecipeOrVPValueTy | ||||||||
8584 | VPRecipeBuilder::tryToCreateWidenRecipe(Instruction *Instr, | ||||||||
8585 | ArrayRef<VPValue *> Operands, | ||||||||
8586 | VFRange &Range, VPBasicBlock *VPBB, | ||||||||
8587 | VPlanPtr &Plan) { | ||||||||
8588 | // First, check for specific widening recipes that deal with inductions, Phi | ||||||||
8589 | // nodes, calls and memory operations. | ||||||||
8590 | VPRecipeBase *Recipe; | ||||||||
8591 | if (auto Phi = dyn_cast<PHINode>(Instr)) { | ||||||||
8592 | if (Phi->getParent() != OrigLoop->getHeader()) | ||||||||
8593 | return tryToBlend(Phi, Operands, Plan); | ||||||||
8594 | |||||||||
8595 | // Always record recipes for header phis. Later first-order recurrence phis | ||||||||
8596 | // can have earlier phis as incoming values. | ||||||||
8597 | recordRecipeOf(Phi); | ||||||||
8598 | |||||||||
8599 | if ((Recipe = tryToOptimizeInductionPHI(Phi, Operands, *Plan, Range))) | ||||||||
8600 | return toVPRecipeResult(Recipe); | ||||||||
8601 | |||||||||
8602 | VPHeaderPHIRecipe *PhiRecipe = nullptr; | ||||||||
8603 | assert((Legal->isReductionVariable(Phi) ||(static_cast <bool> ((Legal->isReductionVariable(Phi ) || Legal->isFixedOrderRecurrence(Phi)) && "can only widen reductions and fixed-order recurrences here" ) ? void (0) : __assert_fail ("(Legal->isReductionVariable(Phi) || Legal->isFixedOrderRecurrence(Phi)) && \"can only widen reductions and fixed-order recurrences here\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8605, __extension__ __PRETTY_FUNCTION__)) | ||||||||
8604 | Legal->isFixedOrderRecurrence(Phi)) &&(static_cast <bool> ((Legal->isReductionVariable(Phi ) || Legal->isFixedOrderRecurrence(Phi)) && "can only widen reductions and fixed-order recurrences here" ) ? void (0) : __assert_fail ("(Legal->isReductionVariable(Phi) || Legal->isFixedOrderRecurrence(Phi)) && \"can only widen reductions and fixed-order recurrences here\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8605, __extension__ __PRETTY_FUNCTION__)) | ||||||||
8605 | "can only widen reductions and fixed-order recurrences here")(static_cast <bool> ((Legal->isReductionVariable(Phi ) || Legal->isFixedOrderRecurrence(Phi)) && "can only widen reductions and fixed-order recurrences here" ) ? void (0) : __assert_fail ("(Legal->isReductionVariable(Phi) || Legal->isFixedOrderRecurrence(Phi)) && \"can only widen reductions and fixed-order recurrences here\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8605, __extension__ __PRETTY_FUNCTION__)); | ||||||||
8606 | VPValue *StartV = Operands[0]; | ||||||||
8607 | if (Legal->isReductionVariable(Phi)) { | ||||||||
8608 | const RecurrenceDescriptor &RdxDesc = | ||||||||
8609 | Legal->getReductionVars().find(Phi)->second; | ||||||||
8610 | 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", 8611, __extension__ __PRETTY_FUNCTION__)) | ||||||||
8611 | 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", 8611, __extension__ __PRETTY_FUNCTION__)); | ||||||||
8612 | PhiRecipe = new VPReductionPHIRecipe(Phi, RdxDesc, *StartV, | ||||||||
8613 | CM.isInLoopReduction(Phi), | ||||||||
8614 | CM.useOrderedReductions(RdxDesc)); | ||||||||
8615 | } else { | ||||||||
8616 | // TODO: Currently fixed-order recurrences are modeled as chains of | ||||||||
8617 | // first-order recurrences. If there are no users of the intermediate | ||||||||
8618 | // recurrences in the chain, the fixed order recurrence should be modeled | ||||||||
8619 | // directly, enabling more efficient codegen. | ||||||||
8620 | PhiRecipe = new VPFirstOrderRecurrencePHIRecipe(Phi, *StartV); | ||||||||
8621 | } | ||||||||
8622 | |||||||||
8623 | // Record the incoming value from the backedge, so we can add the incoming | ||||||||
8624 | // value from the backedge after all recipes have been created. | ||||||||
8625 | auto *Inc = cast<Instruction>( | ||||||||
8626 | Phi->getIncomingValueForBlock(OrigLoop->getLoopLatch())); | ||||||||
8627 | auto RecipeIter = Ingredient2Recipe.find(Inc); | ||||||||
8628 | if (RecipeIter == Ingredient2Recipe.end()) | ||||||||
8629 | recordRecipeOf(Inc); | ||||||||
8630 | |||||||||
8631 | PhisToFix.push_back(PhiRecipe); | ||||||||
8632 | return toVPRecipeResult(PhiRecipe); | ||||||||
8633 | } | ||||||||
8634 | |||||||||
8635 | if (isa<TruncInst>(Instr) && | ||||||||
8636 | (Recipe = tryToOptimizeInductionTruncate(cast<TruncInst>(Instr), Operands, | ||||||||
8637 | Range, *Plan))) | ||||||||
8638 | return toVPRecipeResult(Recipe); | ||||||||
8639 | |||||||||
8640 | // All widen recipes below deal only with VF > 1. | ||||||||
8641 | if (LoopVectorizationPlanner::getDecisionAndClampRange( | ||||||||
8642 | [&](ElementCount VF) { return VF.isScalar(); }, Range)) | ||||||||
8643 | return nullptr; | ||||||||
8644 | |||||||||
8645 | if (auto *CI = dyn_cast<CallInst>(Instr)) | ||||||||
8646 | return toVPRecipeResult(tryToWidenCall(CI, Operands, Range)); | ||||||||
8647 | |||||||||
8648 | if (isa<LoadInst>(Instr) || isa<StoreInst>(Instr)) | ||||||||
8649 | return toVPRecipeResult(tryToWidenMemory(Instr, Operands, Range, Plan)); | ||||||||
8650 | |||||||||
8651 | if (!shouldWiden(Instr, Range)) | ||||||||
8652 | return nullptr; | ||||||||
8653 | |||||||||
8654 | if (auto GEP = dyn_cast<GetElementPtrInst>(Instr)) | ||||||||
8655 | return toVPRecipeResult(new VPWidenGEPRecipe( | ||||||||
8656 | GEP, make_range(Operands.begin(), Operands.end()), OrigLoop)); | ||||||||
8657 | |||||||||
8658 | if (auto *SI = dyn_cast<SelectInst>(Instr)) { | ||||||||
8659 | bool InvariantCond = | ||||||||
8660 | PSE.getSE()->isLoopInvariant(PSE.getSCEV(SI->getOperand(0)), OrigLoop); | ||||||||
8661 | return toVPRecipeResult(new VPWidenSelectRecipe( | ||||||||
8662 | *SI, make_range(Operands.begin(), Operands.end()), InvariantCond)); | ||||||||
8663 | } | ||||||||
8664 | |||||||||
8665 | return toVPRecipeResult(tryToWiden(Instr, Operands, VPBB, Plan)); | ||||||||
8666 | } | ||||||||
8667 | |||||||||
8668 | void LoopVectorizationPlanner::buildVPlansWithVPRecipes(ElementCount MinVF, | ||||||||
8669 | ElementCount MaxVF) { | ||||||||
8670 | 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", 8670, __extension__ __PRETTY_FUNCTION__)); | ||||||||
8671 | |||||||||
8672 | // Add assume instructions we need to drop to DeadInstructions, to prevent | ||||||||
8673 | // them from being added to the VPlan. | ||||||||
8674 | // TODO: We only need to drop assumes in blocks that get flattend. If the | ||||||||
8675 | // control flow is preserved, we should keep them. | ||||||||
8676 | SmallPtrSet<Instruction *, 4> DeadInstructions; | ||||||||
8677 | auto &ConditionalAssumes = Legal->getConditionalAssumes(); | ||||||||
8678 | DeadInstructions.insert(ConditionalAssumes.begin(), ConditionalAssumes.end()); | ||||||||
8679 | |||||||||
8680 | MapVector<Instruction *, Instruction *> &SinkAfter = Legal->getSinkAfter(); | ||||||||
8681 | // Dead instructions do not need sinking. Remove them from SinkAfter. | ||||||||
8682 | for (Instruction *I : DeadInstructions) | ||||||||
8683 | SinkAfter.erase(I); | ||||||||
8684 | |||||||||
8685 | // Cannot sink instructions after dead instructions (there won't be any | ||||||||
8686 | // recipes for them). Instead, find the first non-dead previous instruction. | ||||||||
8687 | for (auto &P : Legal->getSinkAfter()) { | ||||||||
8688 | Instruction *SinkTarget = P.second; | ||||||||
8689 | Instruction *FirstInst = &*SinkTarget->getParent()->begin(); | ||||||||
8690 | (void)FirstInst; | ||||||||
8691 | while (DeadInstructions.contains(SinkTarget)) { | ||||||||
8692 | assert((static_cast <bool> (SinkTarget != FirstInst && "Must find a live instruction (at least the one feeding the " "fixed-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 \" \"fixed-order recurrence PHI) before reaching beginning of the block\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8695, __extension__ __PRETTY_FUNCTION__)) | ||||||||
8693 | SinkTarget != FirstInst &&(static_cast <bool> (SinkTarget != FirstInst && "Must find a live instruction (at least the one feeding the " "fixed-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 \" \"fixed-order recurrence PHI) before reaching beginning of the block\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8695, __extension__ __PRETTY_FUNCTION__)) | ||||||||
8694 | "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 " "fixed-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 \" \"fixed-order recurrence PHI) before reaching beginning of the block\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8695, __extension__ __PRETTY_FUNCTION__)) | ||||||||
8695 | "fixed-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 " "fixed-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 \" \"fixed-order recurrence PHI) before reaching beginning of the block\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8695, __extension__ __PRETTY_FUNCTION__)); | ||||||||
8696 | SinkTarget = SinkTarget->getPrevNode(); | ||||||||
8697 | 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", 8698, __extension__ __PRETTY_FUNCTION__)) | ||||||||
8698 | "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", 8698, __extension__ __PRETTY_FUNCTION__)); | ||||||||
8699 | } | ||||||||
8700 | P.second = SinkTarget; | ||||||||
8701 | } | ||||||||
8702 | |||||||||
8703 | auto MaxVFPlusOne = MaxVF.getWithIncrement(1); | ||||||||
8704 | for (ElementCount VF = MinVF; ElementCount::isKnownLT(VF, MaxVFPlusOne);) { | ||||||||
8705 | VFRange SubRange = {VF, MaxVFPlusOne}; | ||||||||
8706 | VPlans.push_back( | ||||||||
8707 | buildVPlanWithVPRecipes(SubRange, DeadInstructions, SinkAfter)); | ||||||||
8708 | VF = SubRange.End; | ||||||||
8709 | } | ||||||||
8710 | } | ||||||||
8711 | |||||||||
8712 | // Add the necessary canonical IV and branch recipes required to control the | ||||||||
8713 | // loop. | ||||||||
8714 | static void addCanonicalIVRecipes(VPlan &Plan, Type *IdxTy, DebugLoc DL, | ||||||||
8715 | bool HasNUW, | ||||||||
8716 | bool UseLaneMaskForLoopControlFlow) { | ||||||||
8717 | Value *StartIdx = ConstantInt::get(IdxTy, 0); | ||||||||
8718 | auto *StartV = Plan.getOrAddVPValue(StartIdx); | ||||||||
8719 | |||||||||
8720 | // Add a VPCanonicalIVPHIRecipe starting at 0 to the header. | ||||||||
8721 | auto *CanonicalIVPHI = new VPCanonicalIVPHIRecipe(StartV, DL); | ||||||||
8722 | VPRegionBlock *TopRegion = Plan.getVectorLoopRegion(); | ||||||||
8723 | VPBasicBlock *Header = TopRegion->getEntryBasicBlock(); | ||||||||
8724 | Header->insert(CanonicalIVPHI, Header->begin()); | ||||||||
8725 | |||||||||
8726 | // Add a CanonicalIVIncrement{NUW} VPInstruction to increment the scalar | ||||||||
8727 | // IV by VF * UF. | ||||||||
8728 | auto *CanonicalIVIncrement = | ||||||||
8729 | new VPInstruction(HasNUW ? VPInstruction::CanonicalIVIncrementNUW | ||||||||
8730 | : VPInstruction::CanonicalIVIncrement, | ||||||||
8731 | {CanonicalIVPHI}, DL, "index.next"); | ||||||||
8732 | CanonicalIVPHI->addOperand(CanonicalIVIncrement); | ||||||||
8733 | |||||||||
8734 | VPBasicBlock *EB = TopRegion->getExitingBasicBlock(); | ||||||||
8735 | EB->appendRecipe(CanonicalIVIncrement); | ||||||||
8736 | |||||||||
8737 | if (UseLaneMaskForLoopControlFlow) { | ||||||||
8738 | // Create the active lane mask instruction in the vplan preheader. | ||||||||
8739 | VPBasicBlock *Preheader = Plan.getEntry()->getEntryBasicBlock(); | ||||||||
8740 | |||||||||
8741 | // We can't use StartV directly in the ActiveLaneMask VPInstruction, since | ||||||||
8742 | // we have to take unrolling into account. Each part needs to start at | ||||||||
8743 | // Part * VF | ||||||||
8744 | auto *CanonicalIVIncrementParts = | ||||||||
8745 | new VPInstruction(HasNUW ? VPInstruction::CanonicalIVIncrementForPartNUW | ||||||||
8746 | : VPInstruction::CanonicalIVIncrementForPart, | ||||||||
8747 | {StartV}, DL, "index.part.next"); | ||||||||
8748 | Preheader->appendRecipe(CanonicalIVIncrementParts); | ||||||||
8749 | |||||||||
8750 | // Create the ActiveLaneMask instruction using the correct start values. | ||||||||
8751 | VPValue *TC = Plan.getOrCreateTripCount(); | ||||||||
8752 | auto *EntryALM = new VPInstruction(VPInstruction::ActiveLaneMask, | ||||||||
8753 | {CanonicalIVIncrementParts, TC}, DL, | ||||||||
8754 | "active.lane.mask.entry"); | ||||||||
8755 | Preheader->appendRecipe(EntryALM); | ||||||||
8756 | |||||||||
8757 | // Now create the ActiveLaneMaskPhi recipe in the main loop using the | ||||||||
8758 | // preheader ActiveLaneMask instruction. | ||||||||
8759 | auto *LaneMaskPhi = new VPActiveLaneMaskPHIRecipe(EntryALM, DebugLoc()); | ||||||||
8760 | Header->insert(LaneMaskPhi, Header->getFirstNonPhi()); | ||||||||
8761 | |||||||||
8762 | // Create the active lane mask for the next iteration of the loop. | ||||||||
8763 | CanonicalIVIncrementParts = | ||||||||
8764 | new VPInstruction(HasNUW ? VPInstruction::CanonicalIVIncrementForPartNUW | ||||||||
8765 | : VPInstruction::CanonicalIVIncrementForPart, | ||||||||
8766 | {CanonicalIVIncrement}, DL); | ||||||||
8767 | EB->appendRecipe(CanonicalIVIncrementParts); | ||||||||
8768 | |||||||||
8769 | auto *ALM = new VPInstruction(VPInstruction::ActiveLaneMask, | ||||||||
8770 | {CanonicalIVIncrementParts, TC}, DL, | ||||||||
8771 | "active.lane.mask.next"); | ||||||||
8772 | EB->appendRecipe(ALM); | ||||||||
8773 | LaneMaskPhi->addOperand(ALM); | ||||||||
8774 | |||||||||
8775 | // We have to invert the mask here because a true condition means jumping | ||||||||
8776 | // to the exit block. | ||||||||
8777 | auto *NotMask = new VPInstruction(VPInstruction::Not, ALM, DL); | ||||||||
8778 | EB->appendRecipe(NotMask); | ||||||||
8779 | |||||||||
8780 | VPInstruction *BranchBack = | ||||||||
8781 | new VPInstruction(VPInstruction::BranchOnCond, {NotMask}, DL); | ||||||||
8782 | EB->appendRecipe(BranchBack); | ||||||||
8783 | } else { | ||||||||
8784 | // Add the BranchOnCount VPInstruction to the latch. | ||||||||
8785 | VPInstruction *BranchBack = new VPInstruction( | ||||||||
8786 | VPInstruction::BranchOnCount, | ||||||||
8787 | {CanonicalIVIncrement, &Plan.getVectorTripCount()}, DL); | ||||||||
8788 | EB->appendRecipe(BranchBack); | ||||||||
8789 | } | ||||||||
8790 | } | ||||||||
8791 | |||||||||
8792 | // Add exit values to \p Plan. VPLiveOuts are added for each LCSSA phi in the | ||||||||
8793 | // original exit block. | ||||||||
8794 | static void addUsersInExitBlock(VPBasicBlock *HeaderVPBB, | ||||||||
8795 | VPBasicBlock *MiddleVPBB, Loop *OrigLoop, | ||||||||
8796 | VPlan &Plan) { | ||||||||
8797 | BasicBlock *ExitBB = OrigLoop->getUniqueExitBlock(); | ||||||||
8798 | BasicBlock *ExitingBB = OrigLoop->getExitingBlock(); | ||||||||
8799 | // Only handle single-exit loops with unique exit blocks for now. | ||||||||
8800 | if (!ExitBB || !ExitBB->getSinglePredecessor() || !ExitingBB) | ||||||||
8801 | return; | ||||||||
8802 | |||||||||
8803 | // Introduce VPUsers modeling the exit values. | ||||||||
8804 | for (PHINode &ExitPhi : ExitBB->phis()) { | ||||||||
8805 | Value *IncomingValue = | ||||||||
8806 | ExitPhi.getIncomingValueForBlock(ExitingBB); | ||||||||
8807 | VPValue *V = Plan.getOrAddVPValue(IncomingValue, true); | ||||||||
8808 | Plan.addLiveOut(&ExitPhi, V); | ||||||||
8809 | } | ||||||||
8810 | } | ||||||||
8811 | |||||||||
8812 | VPlanPtr LoopVectorizationPlanner::buildVPlanWithVPRecipes( | ||||||||
8813 | VFRange &Range, SmallPtrSetImpl<Instruction *> &DeadInstructions, | ||||||||
8814 | const MapVector<Instruction *, Instruction *> &SinkAfter) { | ||||||||
8815 | |||||||||
8816 | SmallPtrSet<const InterleaveGroup<Instruction> *, 1> InterleaveGroups; | ||||||||
8817 | |||||||||
8818 | VPRecipeBuilder RecipeBuilder(OrigLoop, TLI, Legal, CM, PSE, Builder); | ||||||||
8819 | |||||||||
8820 | // --------------------------------------------------------------------------- | ||||||||
8821 | // Pre-construction: record ingredients whose recipes we'll need to further | ||||||||
8822 | // process after constructing the initial VPlan. | ||||||||
8823 | // --------------------------------------------------------------------------- | ||||||||
8824 | |||||||||
8825 | // Mark instructions we'll need to sink later and their targets as | ||||||||
8826 | // ingredients whose recipe we'll need to record. | ||||||||
8827 | for (const auto &Entry : SinkAfter) { | ||||||||
8828 | RecipeBuilder.recordRecipeOf(Entry.first); | ||||||||
8829 | RecipeBuilder.recordRecipeOf(Entry.second); | ||||||||
8830 | } | ||||||||
8831 | for (const auto &Reduction : CM.getInLoopReductionChains()) { | ||||||||
8832 | PHINode *Phi = Reduction.first; | ||||||||
8833 | RecurKind Kind = | ||||||||
8834 | Legal->getReductionVars().find(Phi)->second.getRecurrenceKind(); | ||||||||
8835 | const SmallVector<Instruction *, 4> &ReductionOperations = Reduction.second; | ||||||||
8836 | |||||||||
8837 | RecipeBuilder.recordRecipeOf(Phi); | ||||||||
8838 | for (const auto &R : ReductionOperations) { | ||||||||
8839 | RecipeBuilder.recordRecipeOf(R); | ||||||||
8840 | // For min/max reductions, where we have a pair of icmp/select, we also | ||||||||
8841 | // need to record the ICmp recipe, so it can be removed later. | ||||||||
8842 | 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", 8843, __extension__ __PRETTY_FUNCTION__)) | ||||||||
8843 | "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", 8843, __extension__ __PRETTY_FUNCTION__)); | ||||||||
8844 | if (RecurrenceDescriptor::isMinMaxRecurrenceKind(Kind)) | ||||||||
8845 | RecipeBuilder.recordRecipeOf(cast<Instruction>(R->getOperand(0))); | ||||||||
8846 | } | ||||||||
8847 | } | ||||||||
8848 | |||||||||
8849 | // For each interleave group which is relevant for this (possibly trimmed) | ||||||||
8850 | // Range, add it to the set of groups to be later applied to the VPlan and add | ||||||||
8851 | // placeholders for its members' Recipes which we'll be replacing with a | ||||||||
8852 | // single VPInterleaveRecipe. | ||||||||
8853 | for (InterleaveGroup<Instruction> *IG : IAI.getInterleaveGroups()) { | ||||||||
8854 | auto applyIG = [IG, this](ElementCount VF) -> bool { | ||||||||
8855 | return (VF.isVector() && // Query is illegal for VF == 1 | ||||||||
8856 | CM.getWideningDecision(IG->getInsertPos(), VF) == | ||||||||
8857 | LoopVectorizationCostModel::CM_Interleave); | ||||||||
8858 | }; | ||||||||
8859 | if (!getDecisionAndClampRange(applyIG, Range)) | ||||||||
8860 | continue; | ||||||||
8861 | InterleaveGroups.insert(IG); | ||||||||
8862 | for (unsigned i = 0; i < IG->getFactor(); i++) | ||||||||
8863 | if (Instruction *Member = IG->getMember(i)) | ||||||||
8864 | RecipeBuilder.recordRecipeOf(Member); | ||||||||
8865 | }; | ||||||||
8866 | |||||||||
8867 | // --------------------------------------------------------------------------- | ||||||||
8868 | // Build initial VPlan: Scan the body of the loop in a topological order to | ||||||||
8869 | // visit each basic block after having visited its predecessor basic blocks. | ||||||||
8870 | // --------------------------------------------------------------------------- | ||||||||
8871 | |||||||||
8872 | // Create initial VPlan skeleton, starting with a block for the pre-header, | ||||||||
8873 | // followed by a region for the vector loop, followed by the middle block. The | ||||||||
8874 | // skeleton vector loop region contains a header and latch block. | ||||||||
8875 | VPBasicBlock *Preheader = new VPBasicBlock("vector.ph"); | ||||||||
8876 | auto Plan = std::make_unique<VPlan>(Preheader); | ||||||||
8877 | |||||||||
8878 | VPBasicBlock *HeaderVPBB = new VPBasicBlock("vector.body"); | ||||||||
8879 | VPBasicBlock *LatchVPBB = new VPBasicBlock("vector.latch"); | ||||||||
8880 | VPBlockUtils::insertBlockAfter(LatchVPBB, HeaderVPBB); | ||||||||
8881 | auto *TopRegion = new VPRegionBlock(HeaderVPBB, LatchVPBB, "vector loop"); | ||||||||
8882 | VPBlockUtils::insertBlockAfter(TopRegion, Preheader); | ||||||||
8883 | VPBasicBlock *MiddleVPBB = new VPBasicBlock("middle.block"); | ||||||||
8884 | VPBlockUtils::insertBlockAfter(MiddleVPBB, TopRegion); | ||||||||
8885 | |||||||||
8886 | Instruction *DLInst = | ||||||||
8887 | getDebugLocFromInstOrOperands(Legal->getPrimaryInduction()); | ||||||||
8888 | addCanonicalIVRecipes(*Plan, Legal->getWidestInductionType(), | ||||||||
8889 | DLInst ? DLInst->getDebugLoc() : DebugLoc(), | ||||||||
8890 | !CM.foldTailByMasking(), | ||||||||
8891 | CM.useActiveLaneMaskForControlFlow()); | ||||||||
8892 | |||||||||
8893 | // Scan the body of the loop in a topological order to visit each basic block | ||||||||
8894 | // after having visited its predecessor basic blocks. | ||||||||
8895 | LoopBlocksDFS DFS(OrigLoop); | ||||||||
8896 | DFS.perform(LI); | ||||||||
8897 | |||||||||
8898 | VPBasicBlock *VPBB = HeaderVPBB; | ||||||||
8899 | SmallVector<VPWidenIntOrFpInductionRecipe *> InductionsToMove; | ||||||||
8900 | for (BasicBlock *BB : make_range(DFS.beginRPO(), DFS.endRPO())) { | ||||||||
8901 | // Relevant instructions from basic block BB will be grouped into VPRecipe | ||||||||
8902 | // ingredients and fill a new VPBasicBlock. | ||||||||
8903 | unsigned VPBBsForBB = 0; | ||||||||
8904 | if (VPBB != HeaderVPBB) | ||||||||
8905 | VPBB->setName(BB->getName()); | ||||||||
8906 | Builder.setInsertPoint(VPBB); | ||||||||
8907 | |||||||||
8908 | // Introduce each ingredient into VPlan. | ||||||||
8909 | // TODO: Model and preserve debug intrinsics in VPlan. | ||||||||
8910 | for (Instruction &I : BB->instructionsWithoutDebug()) { | ||||||||
8911 | Instruction *Instr = &I; | ||||||||
8912 | |||||||||
8913 | // First filter out irrelevant instructions, to ensure no recipes are | ||||||||
8914 | // built for them. | ||||||||
8915 | if (isa<BranchInst>(Instr) || DeadInstructions.count(Instr)) | ||||||||
8916 | continue; | ||||||||
8917 | |||||||||
8918 | SmallVector<VPValue *, 4> Operands; | ||||||||
8919 | auto *Phi = dyn_cast<PHINode>(Instr); | ||||||||
8920 | if (Phi && Phi->getParent() == OrigLoop->getHeader()) { | ||||||||
8921 | Operands.push_back(Plan->getOrAddVPValue( | ||||||||
8922 | Phi->getIncomingValueForBlock(OrigLoop->getLoopPreheader()))); | ||||||||
8923 | } else { | ||||||||
8924 | auto OpRange = Plan->mapToVPValues(Instr->operands()); | ||||||||
8925 | Operands = {OpRange.begin(), OpRange.end()}; | ||||||||
8926 | } | ||||||||
8927 | |||||||||
8928 | // Invariant stores inside loop will be deleted and a single store | ||||||||
8929 | // with the final reduction value will be added to the exit block | ||||||||
8930 | StoreInst *SI; | ||||||||
8931 | if ((SI = dyn_cast<StoreInst>(&I)) && | ||||||||
8932 | Legal->isInvariantAddressOfReduction(SI->getPointerOperand())) | ||||||||
8933 | continue; | ||||||||
8934 | |||||||||
8935 | if (auto RecipeOrValue = RecipeBuilder.tryToCreateWidenRecipe( | ||||||||
8936 | Instr, Operands, Range, VPBB, Plan)) { | ||||||||
8937 | // If Instr can be simplified to an existing VPValue, use it. | ||||||||
8938 | if (RecipeOrValue.is<VPValue *>()) { | ||||||||
8939 | auto *VPV = RecipeOrValue.get<VPValue *>(); | ||||||||
8940 | Plan->addVPValue(Instr, VPV); | ||||||||
8941 | // If the re-used value is a recipe, register the recipe for the | ||||||||
8942 | // instruction, in case the recipe for Instr needs to be recorded. | ||||||||
8943 | if (VPRecipeBase *R = VPV->getDefiningRecipe()) | ||||||||
8944 | RecipeBuilder.setRecipe(Instr, R); | ||||||||
8945 | continue; | ||||||||
8946 | } | ||||||||
8947 | // Otherwise, add the new recipe. | ||||||||
8948 | VPRecipeBase *Recipe = RecipeOrValue.get<VPRecipeBase *>(); | ||||||||
8949 | for (auto *Def : Recipe->definedValues()) { | ||||||||
8950 | auto *UV = Def->getUnderlyingValue(); | ||||||||
8951 | Plan->addVPValue(UV, Def); | ||||||||
8952 | } | ||||||||
8953 | |||||||||
8954 | if (isa<VPWidenIntOrFpInductionRecipe>(Recipe) && | ||||||||
8955 | HeaderVPBB->getFirstNonPhi() != VPBB->end()) { | ||||||||
8956 | // Keep track of VPWidenIntOrFpInductionRecipes not in the phi section | ||||||||
8957 | // of the header block. That can happen for truncates of induction | ||||||||
8958 | // variables. Those recipes are moved to the phi section of the header | ||||||||
8959 | // block after applying SinkAfter, which relies on the original | ||||||||
8960 | // position of the trunc. | ||||||||
8961 | assert(isa<TruncInst>(Instr))(static_cast <bool> (isa<TruncInst>(Instr)) ? void (0) : __assert_fail ("isa<TruncInst>(Instr)", "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 8961, __extension__ __PRETTY_FUNCTION__)); | ||||||||
8962 | InductionsToMove.push_back( | ||||||||
8963 | cast<VPWidenIntOrFpInductionRecipe>(Recipe)); | ||||||||
8964 | } | ||||||||
8965 | RecipeBuilder.setRecipe(Instr, Recipe); | ||||||||
8966 | VPBB->appendRecipe(Recipe); | ||||||||
8967 | continue; | ||||||||
8968 | } | ||||||||
8969 | |||||||||
8970 | // Otherwise, if all widening options failed, Instruction is to be | ||||||||
8971 | // replicated. This may create a successor for VPBB. | ||||||||
8972 | VPBasicBlock *NextVPBB = | ||||||||
8973 | RecipeBuilder.handleReplication(Instr, Range, VPBB, Plan); | ||||||||
8974 | if (NextVPBB != VPBB) { | ||||||||
8975 | VPBB = NextVPBB; | ||||||||
8976 | VPBB->setName(BB->hasName() ? BB->getName() + "." + Twine(VPBBsForBB++) | ||||||||
8977 | : ""); | ||||||||
8978 | } | ||||||||
8979 | } | ||||||||
8980 | |||||||||
8981 | VPBlockUtils::insertBlockAfter(new VPBasicBlock(), VPBB); | ||||||||
8982 | VPBB = cast<VPBasicBlock>(VPBB->getSingleSuccessor()); | ||||||||
8983 | } | ||||||||
8984 | |||||||||
8985 | // After here, VPBB should not be used. | ||||||||
8986 | VPBB = nullptr; | ||||||||
8987 | |||||||||
8988 | addUsersInExitBlock(HeaderVPBB, MiddleVPBB, OrigLoop, *Plan); | ||||||||
8989 | |||||||||
8990 | assert(isa<VPRegionBlock>(Plan->getVectorLoopRegion()) &&(static_cast <bool> (isa<VPRegionBlock>(Plan-> getVectorLoopRegion()) && !Plan->getVectorLoopRegion ()->getEntryBasicBlock()->empty() && "entry block must be set to a VPRegionBlock having a non-empty entry " "VPBasicBlock") ? void (0) : __assert_fail ("isa<VPRegionBlock>(Plan->getVectorLoopRegion()) && !Plan->getVectorLoopRegion()->getEntryBasicBlock()->empty() && \"entry block must be set to a VPRegionBlock having a non-empty entry \" \"VPBasicBlock\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8993, __extension__ __PRETTY_FUNCTION__)) | ||||||||
8991 | !Plan->getVectorLoopRegion()->getEntryBasicBlock()->empty() &&(static_cast <bool> (isa<VPRegionBlock>(Plan-> getVectorLoopRegion()) && !Plan->getVectorLoopRegion ()->getEntryBasicBlock()->empty() && "entry block must be set to a VPRegionBlock having a non-empty entry " "VPBasicBlock") ? void (0) : __assert_fail ("isa<VPRegionBlock>(Plan->getVectorLoopRegion()) && !Plan->getVectorLoopRegion()->getEntryBasicBlock()->empty() && \"entry block must be set to a VPRegionBlock having a non-empty entry \" \"VPBasicBlock\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8993, __extension__ __PRETTY_FUNCTION__)) | ||||||||
8992 | "entry block must be set to a VPRegionBlock having a non-empty entry "(static_cast <bool> (isa<VPRegionBlock>(Plan-> getVectorLoopRegion()) && !Plan->getVectorLoopRegion ()->getEntryBasicBlock()->empty() && "entry block must be set to a VPRegionBlock having a non-empty entry " "VPBasicBlock") ? void (0) : __assert_fail ("isa<VPRegionBlock>(Plan->getVectorLoopRegion()) && !Plan->getVectorLoopRegion()->getEntryBasicBlock()->empty() && \"entry block must be set to a VPRegionBlock having a non-empty entry \" \"VPBasicBlock\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8993, __extension__ __PRETTY_FUNCTION__)) | ||||||||
8993 | "VPBasicBlock")(static_cast <bool> (isa<VPRegionBlock>(Plan-> getVectorLoopRegion()) && !Plan->getVectorLoopRegion ()->getEntryBasicBlock()->empty() && "entry block must be set to a VPRegionBlock having a non-empty entry " "VPBasicBlock") ? void (0) : __assert_fail ("isa<VPRegionBlock>(Plan->getVectorLoopRegion()) && !Plan->getVectorLoopRegion()->getEntryBasicBlock()->empty() && \"entry block must be set to a VPRegionBlock having a non-empty entry \" \"VPBasicBlock\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8993, __extension__ __PRETTY_FUNCTION__)); | ||||||||
8994 | RecipeBuilder.fixHeaderPhis(); | ||||||||
8995 | |||||||||
8996 | // --------------------------------------------------------------------------- | ||||||||
8997 | // Transform initial VPlan: Apply previously taken decisions, in order, to | ||||||||
8998 | // bring the VPlan to its final state. | ||||||||
8999 | // --------------------------------------------------------------------------- | ||||||||
9000 | |||||||||
9001 | // Apply Sink-After legal constraints. | ||||||||
9002 | auto GetReplicateRegion = [](VPRecipeBase *R) -> VPRegionBlock * { | ||||||||
9003 | auto *Region = dyn_cast_or_null<VPRegionBlock>(R->getParent()->getParent()); | ||||||||
9004 | if (Region && Region->isReplicator()) { | ||||||||
9005 | 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", 9006, __extension__ __PRETTY_FUNCTION__)) | ||||||||
9006 | 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", 9006, __extension__ __PRETTY_FUNCTION__)); | ||||||||
9007 | 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", 9009, __extension__ __PRETTY_FUNCTION__)) | ||||||||
9008 | "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", 9009, __extension__ __PRETTY_FUNCTION__)) | ||||||||
9009 | "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", 9009, __extension__ __PRETTY_FUNCTION__)); | ||||||||
9010 | return Region; | ||||||||
9011 | } | ||||||||
9012 | return nullptr; | ||||||||
9013 | }; | ||||||||
9014 | for (const auto &Entry : SinkAfter) { | ||||||||
9015 | VPRecipeBase *Sink = RecipeBuilder.getRecipe(Entry.first); | ||||||||
9016 | VPRecipeBase *Target = RecipeBuilder.getRecipe(Entry.second); | ||||||||
9017 | |||||||||
9018 | auto *TargetRegion = GetReplicateRegion(Target); | ||||||||
9019 | auto *SinkRegion = GetReplicateRegion(Sink); | ||||||||
9020 | if (!SinkRegion) { | ||||||||
9021 | // If the sink source is not a replicate region, sink the recipe directly. | ||||||||
9022 | if (TargetRegion) { | ||||||||
9023 | // The target is in a replication region, make sure to move Sink to | ||||||||
9024 | // the block after it, not into the replication region itself. | ||||||||
9025 | VPBasicBlock *NextBlock = | ||||||||
9026 | cast<VPBasicBlock>(TargetRegion->getSuccessors().front()); | ||||||||
9027 | Sink->moveBefore(*NextBlock, NextBlock->getFirstNonPhi()); | ||||||||
9028 | } else | ||||||||
9029 | Sink->moveAfter(Target); | ||||||||
9030 | continue; | ||||||||
9031 | } | ||||||||
9032 | |||||||||
9033 | // The sink source is in a replicate region. Unhook the region from the CFG. | ||||||||
9034 | auto *SinkPred = SinkRegion->getSinglePredecessor(); | ||||||||
9035 | auto *SinkSucc = SinkRegion->getSingleSuccessor(); | ||||||||
9036 | VPBlockUtils::disconnectBlocks(SinkPred, SinkRegion); | ||||||||
9037 | VPBlockUtils::disconnectBlocks(SinkRegion, SinkSucc); | ||||||||
9038 | VPBlockUtils::connectBlocks(SinkPred, SinkSucc); | ||||||||
9039 | |||||||||
9040 | if (TargetRegion) { | ||||||||
9041 | // The target recipe is also in a replicate region, move the sink region | ||||||||
9042 | // after the target region. | ||||||||
9043 | auto *TargetSucc = TargetRegion->getSingleSuccessor(); | ||||||||
9044 | VPBlockUtils::disconnectBlocks(TargetRegion, TargetSucc); | ||||||||
9045 | VPBlockUtils::connectBlocks(TargetRegion, SinkRegion); | ||||||||
9046 | VPBlockUtils::connectBlocks(SinkRegion, TargetSucc); | ||||||||
9047 | } else { | ||||||||
9048 | // The sink source is in a replicate region, we need to move the whole | ||||||||
9049 | // replicate region, which should only contain a single recipe in the | ||||||||
9050 | // main block. | ||||||||
9051 | auto *SplitBlock = | ||||||||
9052 | Target->getParent()->splitAt(std::next(Target->getIterator())); | ||||||||
9053 | |||||||||
9054 | auto *SplitPred = SplitBlock->getSinglePredecessor(); | ||||||||
9055 | |||||||||
9056 | VPBlockUtils::disconnectBlocks(SplitPred, SplitBlock); | ||||||||
9057 | VPBlockUtils::connectBlocks(SplitPred, SinkRegion); | ||||||||
9058 | VPBlockUtils::connectBlocks(SinkRegion, SplitBlock); | ||||||||
9059 | } | ||||||||
9060 | } | ||||||||
9061 | |||||||||
9062 | VPlanTransforms::removeRedundantCanonicalIVs(*Plan); | ||||||||
9063 | VPlanTransforms::removeRedundantInductionCasts(*Plan); | ||||||||
9064 | |||||||||
9065 | // Now that sink-after is done, move induction recipes for optimized truncates | ||||||||
9066 | // to the phi section of the header block. | ||||||||
9067 | for (VPWidenIntOrFpInductionRecipe *Ind : InductionsToMove) | ||||||||
9068 | Ind->moveBefore(*HeaderVPBB, HeaderVPBB->getFirstNonPhi()); | ||||||||
9069 | |||||||||
9070 | // Adjust the recipes for any inloop reductions. | ||||||||
9071 | adjustRecipesForReductions(cast<VPBasicBlock>(TopRegion->getExiting()), Plan, | ||||||||
9072 | RecipeBuilder, Range.Start); | ||||||||
9073 | |||||||||
9074 | // Introduce a recipe to combine the incoming and previous values of a | ||||||||
9075 | // fixed-order recurrence. | ||||||||
9076 | for (VPRecipeBase &R : | ||||||||
9077 | Plan->getVectorLoopRegion()->getEntryBasicBlock()->phis()) { | ||||||||
9078 | auto *RecurPhi = dyn_cast<VPFirstOrderRecurrencePHIRecipe>(&R); | ||||||||
9079 | if (!RecurPhi) | ||||||||
9080 | continue; | ||||||||
9081 | |||||||||
9082 | VPRecipeBase *PrevRecipe = &RecurPhi->getBackedgeRecipe(); | ||||||||
9083 | // Fixed-order recurrences do not contain cycles, so this loop is guaranteed | ||||||||
9084 | // to terminate. | ||||||||
9085 | while (auto *PrevPhi = | ||||||||
9086 | dyn_cast<VPFirstOrderRecurrencePHIRecipe>(PrevRecipe)) | ||||||||
9087 | PrevRecipe = &PrevPhi->getBackedgeRecipe(); | ||||||||
9088 | VPBasicBlock *InsertBlock = PrevRecipe->getParent(); | ||||||||
9089 | auto *Region = GetReplicateRegion(PrevRecipe); | ||||||||
9090 | if (Region) | ||||||||
9091 | InsertBlock = dyn_cast<VPBasicBlock>(Region->getSingleSuccessor()); | ||||||||
9092 | if (!InsertBlock) { | ||||||||
9093 | InsertBlock = new VPBasicBlock(Region->getName() + ".succ"); | ||||||||
9094 | VPBlockUtils::insertBlockAfter(InsertBlock, Region); | ||||||||
9095 | } | ||||||||
9096 | if (Region || PrevRecipe->isPhi()) | ||||||||
9097 | Builder.setInsertPoint(InsertBlock, InsertBlock->getFirstNonPhi()); | ||||||||
9098 | else | ||||||||
9099 | Builder.setInsertPoint(InsertBlock, std::next(PrevRecipe->getIterator())); | ||||||||
9100 | |||||||||
9101 | auto *RecurSplice = cast<VPInstruction>( | ||||||||
9102 | Builder.createNaryOp(VPInstruction::FirstOrderRecurrenceSplice, | ||||||||
9103 | {RecurPhi, RecurPhi->getBackedgeValue()})); | ||||||||
9104 | |||||||||
9105 | RecurPhi->replaceAllUsesWith(RecurSplice); | ||||||||
9106 | // Set the first operand of RecurSplice to RecurPhi again, after replacing | ||||||||
9107 | // all users. | ||||||||
9108 | RecurSplice->setOperand(0, RecurPhi); | ||||||||
9109 | } | ||||||||
9110 | |||||||||
9111 | // Interleave memory: for each Interleave Group we marked earlier as relevant | ||||||||
9112 | // for this VPlan, replace the Recipes widening its memory instructions with a | ||||||||
9113 | // single VPInterleaveRecipe at its insertion point. | ||||||||
9114 | for (const auto *IG : InterleaveGroups) { | ||||||||
9115 | auto *Recipe = cast<VPWidenMemoryInstructionRecipe>( | ||||||||
9116 | RecipeBuilder.getRecipe(IG->getInsertPos())); | ||||||||
9117 | SmallVector<VPValue *, 4> StoredValues; | ||||||||
9118 | for (unsigned i = 0; i < IG->getFactor(); ++i) | ||||||||
9119 | if (auto *SI = dyn_cast_or_null<StoreInst>(IG->getMember(i))) { | ||||||||
9120 | auto *StoreR = | ||||||||
9121 | cast<VPWidenMemoryInstructionRecipe>(RecipeBuilder.getRecipe(SI)); | ||||||||
9122 | StoredValues.push_back(StoreR->getStoredValue()); | ||||||||
9123 | } | ||||||||
9124 | |||||||||
9125 | auto *VPIG = new VPInterleaveRecipe(IG, Recipe->getAddr(), StoredValues, | ||||||||
9126 | Recipe->getMask()); | ||||||||
9127 | VPIG->insertBefore(Recipe); | ||||||||
9128 | unsigned J = 0; | ||||||||
9129 | for (unsigned i = 0; i < IG->getFactor(); ++i) | ||||||||
9130 | if (Instruction *Member = IG->getMember(i)) { | ||||||||
9131 | if (!Member->getType()->isVoidTy()) { | ||||||||
9132 | VPValue *OriginalV = Plan->getVPValue(Member); | ||||||||
9133 | Plan->removeVPValueFor(Member); | ||||||||
9134 | Plan->addVPValue(Member, VPIG->getVPValue(J)); | ||||||||
9135 | OriginalV->replaceAllUsesWith(VPIG->getVPValue(J)); | ||||||||
9136 | J++; | ||||||||
9137 | } | ||||||||
9138 | RecipeBuilder.getRecipe(Member)->eraseFromParent(); | ||||||||
9139 | } | ||||||||
9140 | } | ||||||||
9141 | |||||||||
9142 | for (ElementCount VF = Range.Start; ElementCount::isKnownLT(VF, Range.End); | ||||||||
9143 | VF *= 2) | ||||||||
9144 | Plan->addVF(VF); | ||||||||
9145 | Plan->setName("Initial VPlan"); | ||||||||
9146 | |||||||||
9147 | // From this point onwards, VPlan-to-VPlan transformations may change the plan | ||||||||
9148 | // in ways that accessing values using original IR values is incorrect. | ||||||||
9149 | Plan->disableValue2VPValue(); | ||||||||
9150 | |||||||||
9151 | VPlanTransforms::optimizeInductions(*Plan, *PSE.getSE()); | ||||||||
9152 | VPlanTransforms::removeDeadRecipes(*Plan); | ||||||||
9153 | |||||||||
9154 | bool ShouldSimplify = true; | ||||||||
9155 | while (ShouldSimplify) { | ||||||||
9156 | ShouldSimplify = VPlanTransforms::sinkScalarOperands(*Plan); | ||||||||
9157 | ShouldSimplify |= | ||||||||
9158 | VPlanTransforms::mergeReplicateRegionsIntoSuccessors(*Plan); | ||||||||
9159 | ShouldSimplify |= VPlanTransforms::mergeBlocksIntoPredecessors(*Plan); | ||||||||
9160 | } | ||||||||
9161 | |||||||||
9162 | VPlanTransforms::removeRedundantExpandSCEVRecipes(*Plan); | ||||||||
9163 | VPlanTransforms::mergeBlocksIntoPredecessors(*Plan); | ||||||||
9164 | |||||||||
9165 | 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", 9165, __extension__ __PRETTY_FUNCTION__)); | ||||||||
9166 | return Plan; | ||||||||
9167 | } | ||||||||
9168 | |||||||||
9169 | VPlanPtr LoopVectorizationPlanner::buildVPlan(VFRange &Range) { | ||||||||
9170 | // Outer loop handling: They may require CFG and instruction level | ||||||||
9171 | // transformations before even evaluating whether vectorization is profitable. | ||||||||
9172 | // Since we cannot modify the incoming IR, we need to build VPlan upfront in | ||||||||
9173 | // the vectorization pipeline. | ||||||||
9174 | assert(!OrigLoop->isInnermost())(static_cast <bool> (!OrigLoop->isInnermost()) ? void (0) : __assert_fail ("!OrigLoop->isInnermost()", "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 9174, __extension__ __PRETTY_FUNCTION__)); | ||||||||
9175 | 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", 9175, __extension__ __PRETTY_FUNCTION__)); | ||||||||
9176 | |||||||||
9177 | // Create new empty VPlan | ||||||||
9178 | auto Plan = std::make_unique<VPlan>(); | ||||||||
9179 | |||||||||
9180 | // Build hierarchical CFG | ||||||||
9181 | VPlanHCFGBuilder HCFGBuilder(OrigLoop, LI, *Plan); | ||||||||
9182 | HCFGBuilder.buildHierarchicalCFG(); | ||||||||
9183 | |||||||||
9184 | for (ElementCount VF = Range.Start; ElementCount::isKnownLT(VF, Range.End); | ||||||||
9185 | VF *= 2) | ||||||||
9186 | Plan->addVF(VF); | ||||||||
9187 | |||||||||
9188 | SmallPtrSet<Instruction *, 1> DeadInstructions; | ||||||||
9189 | VPlanTransforms::VPInstructionsToVPRecipes( | ||||||||
9190 | OrigLoop, Plan, | ||||||||
9191 | [this](PHINode *P) { return Legal->getIntOrFpInductionDescriptor(P); }, | ||||||||
9192 | DeadInstructions, *PSE.getSE(), *TLI); | ||||||||
9193 | |||||||||
9194 | // Remove the existing terminator of the exiting block of the top-most region. | ||||||||
9195 | // A BranchOnCount will be added instead when adding the canonical IV recipes. | ||||||||
9196 | auto *Term = | ||||||||
9197 | Plan->getVectorLoopRegion()->getExitingBasicBlock()->getTerminator(); | ||||||||
9198 | Term->eraseFromParent(); | ||||||||
9199 | |||||||||
9200 | addCanonicalIVRecipes(*Plan, Legal->getWidestInductionType(), DebugLoc(), | ||||||||
9201 | true, CM.useActiveLaneMaskForControlFlow()); | ||||||||
9202 | return Plan; | ||||||||
9203 | } | ||||||||
9204 | |||||||||
9205 | // Adjust the recipes for reductions. For in-loop reductions the chain of | ||||||||
9206 | // instructions leading from the loop exit instr to the phi need to be converted | ||||||||
9207 | // to reductions, with one operand being vector and the other being the scalar | ||||||||
9208 | // reduction chain. For other reductions, a select is introduced between the phi | ||||||||
9209 | // and live-out recipes when folding the tail. | ||||||||
9210 | void LoopVectorizationPlanner::adjustRecipesForReductions( | ||||||||
9211 | VPBasicBlock *LatchVPBB, VPlanPtr &Plan, VPRecipeBuilder &RecipeBuilder, | ||||||||
9212 | ElementCount MinVF) { | ||||||||
9213 | for (const auto &Reduction : CM.getInLoopReductionChains()) { | ||||||||
9214 | PHINode *Phi = Reduction.first; | ||||||||
9215 | const RecurrenceDescriptor &RdxDesc = | ||||||||
9216 | Legal->getReductionVars().find(Phi)->second; | ||||||||
9217 | const SmallVector<Instruction *, 4> &ReductionOperations = Reduction.second; | ||||||||
9218 | |||||||||
9219 | if (MinVF.isScalar() && !CM.useOrderedReductions(RdxDesc)) | ||||||||
9220 | continue; | ||||||||
9221 | |||||||||
9222 | // ReductionOperations are orders top-down from the phi's use to the | ||||||||
9223 | // LoopExitValue. We keep a track of the previous item (the Chain) to tell | ||||||||
9224 | // which of the two operands will remain scalar and which will be reduced. | ||||||||
9225 | // For minmax the chain will be the select instructions. | ||||||||
9226 | Instruction *Chain = Phi; | ||||||||
9227 | for (Instruction *R : ReductionOperations) { | ||||||||
9228 | VPRecipeBase *WidenRecipe = RecipeBuilder.getRecipe(R); | ||||||||
9229 | RecurKind Kind = RdxDesc.getRecurrenceKind(); | ||||||||
9230 | |||||||||
9231 | VPValue *ChainOp = Plan->getVPValue(Chain); | ||||||||
9232 | unsigned FirstOpId; | ||||||||
9233 | 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", 9234, __extension__ __PRETTY_FUNCTION__)) | ||||||||
9234 | "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", 9234, __extension__ __PRETTY_FUNCTION__)); | ||||||||
9235 | // Recognize a call to the llvm.fmuladd intrinsic. | ||||||||
9236 | bool IsFMulAdd = (Kind == RecurKind::FMulAdd); | ||||||||
9237 | 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", 9238, __extension__ __PRETTY_FUNCTION__)) | ||||||||
9238 | "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", 9238, __extension__ __PRETTY_FUNCTION__)); | ||||||||
9239 | if (RecurrenceDescriptor::isMinMaxRecurrenceKind(Kind)) { | ||||||||
9240 | 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", 9241, __extension__ __PRETTY_FUNCTION__)) | ||||||||
9241 | "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", 9241, __extension__ __PRETTY_FUNCTION__)); | ||||||||
9242 | FirstOpId = 1; | ||||||||
9243 | } else { | ||||||||
9244 | 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", 9246, __extension__ __PRETTY_FUNCTION__)) | ||||||||
9245 | (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", 9246, __extension__ __PRETTY_FUNCTION__)) | ||||||||
9246 | "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", 9246, __extension__ __PRETTY_FUNCTION__)); | ||||||||
9247 | FirstOpId = 0; | ||||||||
9248 | } | ||||||||
9249 | unsigned VecOpId = | ||||||||
9250 | R->getOperand(FirstOpId) == Chain ? FirstOpId + 1 : FirstOpId; | ||||||||
9251 | VPValue *VecOp = Plan->getVPValue(R->getOperand(VecOpId)); | ||||||||
9252 | |||||||||
9253 | VPValue *CondOp = nullptr; | ||||||||
9254 | if (CM.blockNeedsPredicationForAnyReason(R->getParent())) { | ||||||||
9255 | VPBuilder::InsertPointGuard Guard(Builder); | ||||||||
9256 | Builder.setInsertPoint(WidenRecipe->getParent(), | ||||||||
9257 | WidenRecipe->getIterator()); | ||||||||
9258 | CondOp = RecipeBuilder.createBlockInMask(R->getParent(), Plan); | ||||||||
9259 | } | ||||||||
9260 | |||||||||
9261 | if (IsFMulAdd) { | ||||||||
9262 | // If the instruction is a call to the llvm.fmuladd intrinsic then we | ||||||||
9263 | // need to create an fmul recipe to use as the vector operand for the | ||||||||
9264 | // fadd reduction. | ||||||||
9265 | VPInstruction *FMulRecipe = new VPInstruction( | ||||||||
9266 | Instruction::FMul, {VecOp, Plan->getVPValue(R->getOperand(1))}); | ||||||||
9267 | FMulRecipe->setFastMathFlags(R->getFastMathFlags()); | ||||||||
9268 | WidenRecipe->getParent()->insert(FMulRecipe, | ||||||||
9269 | WidenRecipe->getIterator()); | ||||||||
9270 | VecOp = FMulRecipe; | ||||||||
9271 | } | ||||||||
9272 | VPReductionRecipe *RedRecipe = | ||||||||
9273 | new VPReductionRecipe(&RdxDesc, R, ChainOp, VecOp, CondOp, TTI); | ||||||||
9274 | WidenRecipe->getVPSingleValue()->replaceAllUsesWith(RedRecipe); | ||||||||
9275 | Plan->removeVPValueFor(R); | ||||||||
9276 | Plan->addVPValue(R, RedRecipe); | ||||||||
9277 | // Append the recipe to the end of the VPBasicBlock because we need to | ||||||||
9278 | // ensure that it comes after all of it's inputs, including CondOp. | ||||||||
9279 | WidenRecipe->getParent()->appendRecipe(RedRecipe); | ||||||||
9280 | WidenRecipe->getVPSingleValue()->replaceAllUsesWith(RedRecipe); | ||||||||
9281 | WidenRecipe->eraseFromParent(); | ||||||||
9282 | |||||||||
9283 | if (RecurrenceDescriptor::isMinMaxRecurrenceKind(Kind)) { | ||||||||
9284 | VPRecipeBase *CompareRecipe = | ||||||||
9285 | RecipeBuilder.getRecipe(cast<Instruction>(R->getOperand(0))); | ||||||||
9286 | 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", 9287, __extension__ __PRETTY_FUNCTION__)) | ||||||||
9287 | "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", 9287, __extension__ __PRETTY_FUNCTION__)); | ||||||||
9288 | 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", 9289, __extension__ __PRETTY_FUNCTION__)) | ||||||||
9289 | "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", 9289, __extension__ __PRETTY_FUNCTION__)); | ||||||||
9290 | CompareRecipe->eraseFromParent(); | ||||||||
9291 | } | ||||||||
9292 | Chain = R; | ||||||||
9293 | } | ||||||||
9294 | } | ||||||||
9295 | |||||||||
9296 | // If tail is folded by masking, introduce selects between the phi | ||||||||
9297 | // and the live-out instruction of each reduction, at the beginning of the | ||||||||
9298 | // dedicated latch block. | ||||||||
9299 | if (CM.foldTailByMasking()) { | ||||||||
9300 | Builder.setInsertPoint(LatchVPBB, LatchVPBB->begin()); | ||||||||
9301 | for (VPRecipeBase &R : | ||||||||
9302 | Plan->getVectorLoopRegion()->getEntryBasicBlock()->phis()) { | ||||||||
9303 | VPReductionPHIRecipe *PhiR = dyn_cast<VPReductionPHIRecipe>(&R); | ||||||||
9304 | if (!PhiR || PhiR->isInLoop()) | ||||||||
9305 | continue; | ||||||||
9306 | VPValue *Cond = | ||||||||
9307 | RecipeBuilder.createBlockInMask(OrigLoop->getHeader(), Plan); | ||||||||
9308 | VPValue *Red = PhiR->getBackedgeValue(); | ||||||||
9309 | assert(Red->getDefiningRecipe()->getParent() != LatchVPBB &&(static_cast <bool> (Red->getDefiningRecipe()->getParent () != LatchVPBB && "reduction recipe must be defined before latch" ) ? void (0) : __assert_fail ("Red->getDefiningRecipe()->getParent() != LatchVPBB && \"reduction recipe must be defined before latch\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9310, __extension__ __PRETTY_FUNCTION__)) | ||||||||
9310 | "reduction recipe must be defined before latch")(static_cast <bool> (Red->getDefiningRecipe()->getParent () != LatchVPBB && "reduction recipe must be defined before latch" ) ? void (0) : __assert_fail ("Red->getDefiningRecipe()->getParent() != LatchVPBB && \"reduction recipe must be defined before latch\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9310, __extension__ __PRETTY_FUNCTION__)); | ||||||||
9311 | Builder.createNaryOp(Instruction::Select, {Cond, Red, PhiR}); | ||||||||
9312 | } | ||||||||
9313 | } | ||||||||
9314 | } | ||||||||
9315 | |||||||||
9316 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||
9317 | void VPInterleaveRecipe::print(raw_ostream &O, const Twine &Indent, | ||||||||
9318 | VPSlotTracker &SlotTracker) const { | ||||||||
9319 | O << Indent << "INTERLEAVE-GROUP with factor " << IG->getFactor() << " at "; | ||||||||
9320 | IG->getInsertPos()->printAsOperand(O, false); | ||||||||
9321 | O << ", "; | ||||||||
9322 | getAddr()->printAsOperand(O, SlotTracker); | ||||||||
9323 | VPValue *Mask = getMask(); | ||||||||
9324 | if (Mask) { | ||||||||
9325 | O << ", "; | ||||||||
9326 | Mask->printAsOperand(O, SlotTracker); | ||||||||
9327 | } | ||||||||
9328 | |||||||||
9329 | unsigned OpIdx = 0; | ||||||||
9330 | for (unsigned i = 0; i < IG->getFactor(); ++i) { | ||||||||
9331 | if (!IG->getMember(i)) | ||||||||
9332 | continue; | ||||||||
9333 | if (getNumStoreOperands() > 0) { | ||||||||
9334 | O << "\n" << Indent << " store "; | ||||||||
9335 | getOperand(1 + OpIdx)->printAsOperand(O, SlotTracker); | ||||||||
9336 | O << " to index " << i; | ||||||||
9337 | } else { | ||||||||
9338 | O << "\n" << Indent << " "; | ||||||||
9339 | getVPValue(OpIdx)->printAsOperand(O, SlotTracker); | ||||||||
9340 | O << " = load from index " << i; | ||||||||
9341 | } | ||||||||
9342 | ++OpIdx; | ||||||||
9343 | } | ||||||||
9344 | } | ||||||||
9345 | #endif | ||||||||
9346 | |||||||||
9347 | void VPWidenIntOrFpInductionRecipe::execute(VPTransformState &State) { | ||||||||
9348 | 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", 9348, __extension__ __PRETTY_FUNCTION__)); | ||||||||
9349 | |||||||||
9350 | Value *Start = getStartValue()->getLiveInIRValue(); | ||||||||
9351 | const InductionDescriptor &ID = getInductionDescriptor(); | ||||||||
9352 | TruncInst *Trunc = getTruncInst(); | ||||||||
9353 | IRBuilderBase &Builder = State.Builder; | ||||||||
9354 | 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", 9354, __extension__ __PRETTY_FUNCTION__)); | ||||||||
9355 | 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", 9355, __extension__ __PRETTY_FUNCTION__)); | ||||||||
9356 | |||||||||
9357 | // The value from the original loop to which we are mapping the new induction | ||||||||
9358 | // variable. | ||||||||
9359 | Instruction *EntryVal = Trunc ? cast<Instruction>(Trunc) : IV; | ||||||||
9360 | |||||||||
9361 | // Fast-math-flags propagate from the original induction instruction. | ||||||||
9362 | IRBuilder<>::FastMathFlagGuard FMFG(Builder); | ||||||||
9363 | if (ID.getInductionBinOp() && isa<FPMathOperator>(ID.getInductionBinOp())) | ||||||||
9364 | Builder.setFastMathFlags(ID.getInductionBinOp()->getFastMathFlags()); | ||||||||
9365 | |||||||||
9366 | // Now do the actual transformations, and start with fetching the step value. | ||||||||
9367 | Value *Step = State.get(getStepValue(), VPIteration(0, 0)); | ||||||||
9368 | |||||||||
9369 | 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", 9370, __extension__ __PRETTY_FUNCTION__)) | ||||||||
9370 | "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", 9370, __extension__ __PRETTY_FUNCTION__)); | ||||||||
9371 | |||||||||
9372 | // Construct the initial value of the vector IV in the vector loop preheader | ||||||||
9373 | auto CurrIP = Builder.saveIP(); | ||||||||
9374 | BasicBlock *VectorPH = State.CFG.getPreheaderBBFor(this); | ||||||||
9375 | Builder.SetInsertPoint(VectorPH->getTerminator()); | ||||||||
9376 | if (isa<TruncInst>(EntryVal)) { | ||||||||
9377 | 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", 9378, __extension__ __PRETTY_FUNCTION__)) | ||||||||
9378 | "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", 9378, __extension__ __PRETTY_FUNCTION__)); | ||||||||
9379 | auto *TruncType = cast<IntegerType>(EntryVal->getType()); | ||||||||
9380 | Step = Builder.CreateTrunc(Step, TruncType); | ||||||||
9381 | Start = Builder.CreateCast(Instruction::Trunc, Start, TruncType); | ||||||||
9382 | } | ||||||||
9383 | |||||||||
9384 | Value *Zero = getSignedIntOrFpConstant(Start->getType(), 0); | ||||||||
9385 | Value *SplatStart = Builder.CreateVectorSplat(State.VF, Start); | ||||||||
9386 | Value *SteppedStart = getStepVector( | ||||||||
9387 | SplatStart, Zero, Step, ID.getInductionOpcode(), State.VF, State.Builder); | ||||||||
9388 | |||||||||
9389 | // We create vector phi nodes for both integer and floating-point induction | ||||||||
9390 | // variables. Here, we determine the kind of arithmetic we will perform. | ||||||||
9391 | Instruction::BinaryOps AddOp; | ||||||||
9392 | Instruction::BinaryOps MulOp; | ||||||||
9393 | if (Step->getType()->isIntegerTy()) { | ||||||||
9394 | AddOp = Instruction::Add; | ||||||||
9395 | MulOp = Instruction::Mul; | ||||||||
9396 | } else { | ||||||||
9397 | AddOp = ID.getInductionOpcode(); | ||||||||
9398 | MulOp = Instruction::FMul; | ||||||||
9399 | } | ||||||||
9400 | |||||||||
9401 | // Multiply the vectorization factor by the step using integer or | ||||||||
9402 | // floating-point arithmetic as appropriate. | ||||||||
9403 | Type *StepType = Step->getType(); | ||||||||
9404 | Value *RuntimeVF; | ||||||||
9405 | if (Step->getType()->isFloatingPointTy()) | ||||||||
9406 | RuntimeVF = getRuntimeVFAsFloat(Builder, StepType, State.VF); | ||||||||
9407 | else | ||||||||
9408 | RuntimeVF = getRuntimeVF(Builder, StepType, State.VF); | ||||||||
9409 | Value *Mul = Builder.CreateBinOp(MulOp, Step, RuntimeVF); | ||||||||
9410 | |||||||||
9411 | // Create a vector splat to use in the induction update. | ||||||||
9412 | // | ||||||||
9413 | // FIXME: If the step is non-constant, we create the vector splat with | ||||||||
9414 | // IRBuilder. IRBuilder can constant-fold the multiply, but it doesn't | ||||||||
9415 | // handle a constant vector splat. | ||||||||
9416 | Value *SplatVF = isa<Constant>(Mul) | ||||||||
9417 | ? ConstantVector::getSplat(State.VF, cast<Constant>(Mul)) | ||||||||
9418 | : Builder.CreateVectorSplat(State.VF, Mul); | ||||||||
9419 | Builder.restoreIP(CurrIP); | ||||||||
9420 | |||||||||
9421 | // We may need to add the step a number of times, depending on the unroll | ||||||||
9422 | // factor. The last of those goes into the PHI. | ||||||||
9423 | PHINode *VecInd = PHINode::Create(SteppedStart->getType(), 2, "vec.ind", | ||||||||
9424 | &*State.CFG.PrevBB->getFirstInsertionPt()); | ||||||||
9425 | VecInd->setDebugLoc(EntryVal->getDebugLoc()); | ||||||||
9426 | Instruction *LastInduction = VecInd; | ||||||||
9427 | for (unsigned Part = 0; Part < State.UF; ++Part) { | ||||||||
9428 | State.set(this, LastInduction, Part); | ||||||||
9429 | |||||||||
9430 | if (isa<TruncInst>(EntryVal)) | ||||||||
9431 | State.addMetadata(LastInduction, EntryVal); | ||||||||
9432 | |||||||||
9433 | LastInduction = cast<Instruction>( | ||||||||
9434 | Builder.CreateBinOp(AddOp, LastInduction, SplatVF, "step.add")); | ||||||||
9435 | LastInduction->setDebugLoc(EntryVal->getDebugLoc()); | ||||||||
9436 | } | ||||||||
9437 | |||||||||
9438 | LastInduction->setName("vec.ind.next"); | ||||||||
9439 | VecInd->addIncoming(SteppedStart, VectorPH); | ||||||||
9440 | // Add induction update using an incorrect block temporarily. The phi node | ||||||||
9441 | // will be fixed after VPlan execution. Note that at this point the latch | ||||||||
9442 | // block cannot be used, as it does not exist yet. | ||||||||
9443 | // TODO: Model increment value in VPlan, by turning the recipe into a | ||||||||
9444 | // multi-def and a subclass of VPHeaderPHIRecipe. | ||||||||
9445 | VecInd->addIncoming(LastInduction, VectorPH); | ||||||||
9446 | } | ||||||||
9447 | |||||||||
9448 | void VPWidenPointerInductionRecipe::execute(VPTransformState &State) { | ||||||||
9449 | assert(IndDesc.getKind() == InductionDescriptor::IK_PtrInduction &&(static_cast <bool> (IndDesc.getKind() == InductionDescriptor ::IK_PtrInduction && "Not a pointer induction according to InductionDescriptor!" ) ? void (0) : __assert_fail ("IndDesc.getKind() == InductionDescriptor::IK_PtrInduction && \"Not a pointer induction according to InductionDescriptor!\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9450, __extension__ __PRETTY_FUNCTION__)) | ||||||||
9450 | "Not a pointer induction according to InductionDescriptor!")(static_cast <bool> (IndDesc.getKind() == InductionDescriptor ::IK_PtrInduction && "Not a pointer induction according to InductionDescriptor!" ) ? void (0) : __assert_fail ("IndDesc.getKind() == InductionDescriptor::IK_PtrInduction && \"Not a pointer induction according to InductionDescriptor!\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9450, __extension__ __PRETTY_FUNCTION__)); | ||||||||
9451 | assert(cast<PHINode>(getUnderlyingInstr())->getType()->isPointerTy() &&(static_cast <bool> (cast<PHINode>(getUnderlyingInstr ())->getType()->isPointerTy() && "Unexpected type." ) ? void (0) : __assert_fail ("cast<PHINode>(getUnderlyingInstr())->getType()->isPointerTy() && \"Unexpected type.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9452, __extension__ __PRETTY_FUNCTION__)) | ||||||||
9452 | "Unexpected type.")(static_cast <bool> (cast<PHINode>(getUnderlyingInstr ())->getType()->isPointerTy() && "Unexpected type." ) ? void (0) : __assert_fail ("cast<PHINode>(getUnderlyingInstr())->getType()->isPointerTy() && \"Unexpected type.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9452, __extension__ __PRETTY_FUNCTION__)); | ||||||||
9453 | |||||||||
9454 | auto *IVR = getParent()->getPlan()->getCanonicalIV(); | ||||||||
9455 | PHINode *CanonicalIV = cast<PHINode>(State.get(IVR, 0)); | ||||||||
9456 | |||||||||
9457 | if (onlyScalarsGenerated(State.VF)) { | ||||||||
9458 | // This is the normalized GEP that starts counting at zero. | ||||||||
9459 | Value *PtrInd = State.Builder.CreateSExtOrTrunc( | ||||||||
9460 | CanonicalIV, IndDesc.getStep()->getType()); | ||||||||
9461 | // Determine the number of scalars we need to generate for each unroll | ||||||||
9462 | // iteration. If the instruction is uniform, we only need to generate the | ||||||||
9463 | // first lane. Otherwise, we generate all VF values. | ||||||||
9464 | bool IsUniform = vputils::onlyFirstLaneUsed(this); | ||||||||
9465 | 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", 9466, __extension__ __PRETTY_FUNCTION__)) | ||||||||
9466 | "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", 9466, __extension__ __PRETTY_FUNCTION__)); | ||||||||
9467 | unsigned Lanes = IsUniform ? 1 : State.VF.getFixedValue(); | ||||||||
9468 | |||||||||
9469 | for (unsigned Part = 0; Part < State.UF; ++Part) { | ||||||||
9470 | Value *PartStart = | ||||||||
9471 | createStepForVF(State.Builder, PtrInd->getType(), State.VF, Part); | ||||||||
9472 | |||||||||
9473 | for (unsigned Lane = 0; Lane < Lanes; ++Lane) { | ||||||||
9474 | Value *Idx = State.Builder.CreateAdd( | ||||||||
9475 | PartStart, ConstantInt::get(PtrInd->getType(), Lane)); | ||||||||
9476 | Value *GlobalIdx = State.Builder.CreateAdd(PtrInd, Idx); | ||||||||
9477 | |||||||||
9478 | Value *Step = State.get(getOperand(1), VPIteration(0, Part)); | ||||||||
9479 | Value *SclrGep = emitTransformedIndex( | ||||||||
9480 | State.Builder, GlobalIdx, IndDesc.getStartValue(), Step, IndDesc); | ||||||||
9481 | SclrGep->setName("next.gep"); | ||||||||
9482 | State.set(this, SclrGep, VPIteration(Part, Lane)); | ||||||||
9483 | } | ||||||||
9484 | } | ||||||||
9485 | return; | ||||||||
9486 | } | ||||||||
9487 | |||||||||
9488 | assert(isa<SCEVConstant>(IndDesc.getStep()) &&(static_cast <bool> (isa<SCEVConstant>(IndDesc.getStep ()) && "Induction step not a SCEV constant!") ? void ( 0) : __assert_fail ("isa<SCEVConstant>(IndDesc.getStep()) && \"Induction step not a SCEV constant!\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9489, __extension__ __PRETTY_FUNCTION__)) | ||||||||
9489 | "Induction step not a SCEV constant!")(static_cast <bool> (isa<SCEVConstant>(IndDesc.getStep ()) && "Induction step not a SCEV constant!") ? void ( 0) : __assert_fail ("isa<SCEVConstant>(IndDesc.getStep()) && \"Induction step not a SCEV constant!\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9489, __extension__ __PRETTY_FUNCTION__)); | ||||||||
9490 | Type *PhiType = IndDesc.getStep()->getType(); | ||||||||
9491 | |||||||||
9492 | // Build a pointer phi | ||||||||
9493 | Value *ScalarStartValue = getStartValue()->getLiveInIRValue(); | ||||||||
9494 | Type *ScStValueType = ScalarStartValue->getType(); | ||||||||
9495 | PHINode *NewPointerPhi = | ||||||||
9496 | PHINode::Create(ScStValueType, 2, "pointer.phi", CanonicalIV); | ||||||||
9497 | |||||||||
9498 | BasicBlock *VectorPH = State.CFG.getPreheaderBBFor(this); | ||||||||
9499 | NewPointerPhi->addIncoming(ScalarStartValue, VectorPH); | ||||||||
9500 | |||||||||
9501 | // A pointer induction, performed by using a gep | ||||||||
9502 | Instruction *InductionLoc = &*State.Builder.GetInsertPoint(); | ||||||||
9503 | |||||||||
9504 | Value *ScalarStepValue = State.get(getOperand(1), VPIteration(0, 0)); | ||||||||
9505 | Value *RuntimeVF = getRuntimeVF(State.Builder, PhiType, State.VF); | ||||||||
9506 | Value *NumUnrolledElems = | ||||||||
9507 | State.Builder.CreateMul(RuntimeVF, ConstantInt::get(PhiType, State.UF)); | ||||||||
9508 | Value *InductionGEP = GetElementPtrInst::Create( | ||||||||
9509 | IndDesc.getElementType(), NewPointerPhi, | ||||||||
9510 | State.Builder.CreateMul(ScalarStepValue, NumUnrolledElems), "ptr.ind", | ||||||||
9511 | InductionLoc); | ||||||||
9512 | // Add induction update using an incorrect block temporarily. The phi node | ||||||||
9513 | // will be fixed after VPlan execution. Note that at this point the latch | ||||||||
9514 | // block cannot be used, as it does not exist yet. | ||||||||
9515 | // TODO: Model increment value in VPlan, by turning the recipe into a | ||||||||
9516 | // multi-def and a subclass of VPHeaderPHIRecipe. | ||||||||
9517 | NewPointerPhi->addIncoming(InductionGEP, VectorPH); | ||||||||
9518 | |||||||||
9519 | // Create UF many actual address geps that use the pointer | ||||||||
9520 | // phi as base and a vectorized version of the step value | ||||||||
9521 | // (<step*0, ..., step*N>) as offset. | ||||||||
9522 | for (unsigned Part = 0; Part < State.UF; ++Part) { | ||||||||
9523 | Type *VecPhiType = VectorType::get(PhiType, State.VF); | ||||||||
9524 | Value *StartOffsetScalar = | ||||||||
9525 | State.Builder.CreateMul(RuntimeVF, ConstantInt::get(PhiType, Part)); | ||||||||
9526 | Value *StartOffset = | ||||||||
9527 | State.Builder.CreateVectorSplat(State.VF, StartOffsetScalar); | ||||||||
9528 | // Create a vector of consecutive numbers from zero to VF. | ||||||||
9529 | StartOffset = State.Builder.CreateAdd( | ||||||||
9530 | StartOffset, State.Builder.CreateStepVector(VecPhiType)); | ||||||||
9531 | |||||||||
9532 | assert(ScalarStepValue == State.get(getOperand(1), VPIteration(0, Part)) &&(static_cast <bool> (ScalarStepValue == State.get(getOperand (1), VPIteration(0, Part)) && "scalar step must be the same across all parts" ) ? void (0) : __assert_fail ("ScalarStepValue == State.get(getOperand(1), VPIteration(0, Part)) && \"scalar step must be the same across all parts\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9533, __extension__ __PRETTY_FUNCTION__)) | ||||||||
9533 | "scalar step must be the same across all parts")(static_cast <bool> (ScalarStepValue == State.get(getOperand (1), VPIteration(0, Part)) && "scalar step must be the same across all parts" ) ? void (0) : __assert_fail ("ScalarStepValue == State.get(getOperand(1), VPIteration(0, Part)) && \"scalar step must be the same across all parts\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9533, __extension__ __PRETTY_FUNCTION__)); | ||||||||
9534 | Value *GEP = State.Builder.CreateGEP( | ||||||||
9535 | IndDesc.getElementType(), NewPointerPhi, | ||||||||
9536 | State.Builder.CreateMul( | ||||||||
9537 | StartOffset, | ||||||||
9538 | State.Builder.CreateVectorSplat(State.VF, ScalarStepValue), | ||||||||
9539 | "vector.gep")); | ||||||||
9540 | State.set(this, GEP, Part); | ||||||||
9541 | } | ||||||||
9542 | } | ||||||||
9543 | |||||||||
9544 | void VPDerivedIVRecipe::execute(VPTransformState &State) { | ||||||||
9545 | assert(!State.Instance && "VPDerivedIVRecipe being replicated.")(static_cast <bool> (!State.Instance && "VPDerivedIVRecipe being replicated." ) ? void (0) : __assert_fail ("!State.Instance && \"VPDerivedIVRecipe being replicated.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9545, __extension__ __PRETTY_FUNCTION__)); | ||||||||
9546 | |||||||||
9547 | // Fast-math-flags propagate from the original induction instruction. | ||||||||
9548 | IRBuilder<>::FastMathFlagGuard FMFG(State.Builder); | ||||||||
9549 | if (IndDesc.getInductionBinOp() && | ||||||||
9550 | isa<FPMathOperator>(IndDesc.getInductionBinOp())) | ||||||||
9551 | State.Builder.setFastMathFlags( | ||||||||
9552 | IndDesc.getInductionBinOp()->getFastMathFlags()); | ||||||||
9553 | |||||||||
9554 | Value *Step = State.get(getStepValue(), VPIteration(0, 0)); | ||||||||
9555 | Value *CanonicalIV = State.get(getCanonicalIV(), VPIteration(0, 0)); | ||||||||
9556 | Value *DerivedIV = | ||||||||
9557 | emitTransformedIndex(State.Builder, CanonicalIV, | ||||||||
9558 | getStartValue()->getLiveInIRValue(), Step, IndDesc); | ||||||||
9559 | DerivedIV->setName("offset.idx"); | ||||||||
9560 | if (ResultTy != DerivedIV->getType()) { | ||||||||
9561 | 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", 9562, __extension__ __PRETTY_FUNCTION__)) | ||||||||
9562 | "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", 9562, __extension__ __PRETTY_FUNCTION__)); | ||||||||
9563 | DerivedIV = State.Builder.CreateTrunc(DerivedIV, ResultTy); | ||||||||
9564 | } | ||||||||
9565 | assert(DerivedIV != CanonicalIV && "IV didn't need transforming?")(static_cast <bool> (DerivedIV != CanonicalIV && "IV didn't need transforming?") ? void (0) : __assert_fail ( "DerivedIV != CanonicalIV && \"IV didn't need transforming?\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9565, __extension__ __PRETTY_FUNCTION__)); | ||||||||
9566 | |||||||||
9567 | State.set(this, DerivedIV, VPIteration(0, 0)); | ||||||||
9568 | } | ||||||||
9569 | |||||||||
9570 | void VPScalarIVStepsRecipe::execute(VPTransformState &State) { | ||||||||
9571 | // Fast-math-flags propagate from the original induction instruction. | ||||||||
9572 | IRBuilder<>::FastMathFlagGuard FMFG(State.Builder); | ||||||||
9573 | if (IndDesc.getInductionBinOp() && | ||||||||
9574 | isa<FPMathOperator>(IndDesc.getInductionBinOp())) | ||||||||
9575 | State.Builder.setFastMathFlags( | ||||||||
9576 | IndDesc.getInductionBinOp()->getFastMathFlags()); | ||||||||
9577 | |||||||||
9578 | Value *BaseIV = State.get(getOperand(0), VPIteration(0, 0)); | ||||||||
9579 | Value *Step = State.get(getStepValue(), VPIteration(0, 0)); | ||||||||
9580 | |||||||||
9581 | buildScalarSteps(BaseIV, Step, IndDesc, this, State); | ||||||||
9582 | } | ||||||||
9583 | |||||||||
9584 | void VPInterleaveRecipe::execute(VPTransformState &State) { | ||||||||
9585 | 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", 9585, __extension__ __PRETTY_FUNCTION__)); | ||||||||
9586 | State.ILV->vectorizeInterleaveGroup(IG, definedValues(), State, getAddr(), | ||||||||
9587 | getStoredValues(), getMask()); | ||||||||
9588 | } | ||||||||
9589 | |||||||||
9590 | void VPReductionRecipe::execute(VPTransformState &State) { | ||||||||
9591 | 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", 9591, __extension__ __PRETTY_FUNCTION__)); | ||||||||
9592 | Value *PrevInChain = State.get(getChainOp(), 0); | ||||||||
9593 | RecurKind Kind = RdxDesc->getRecurrenceKind(); | ||||||||
9594 | bool IsOrdered = State.ILV->useOrderedReductions(*RdxDesc); | ||||||||
9595 | // Propagate the fast-math flags carried by the underlying instruction. | ||||||||
9596 | IRBuilderBase::FastMathFlagGuard FMFGuard(State.Builder); | ||||||||
9597 | State.Builder.setFastMathFlags(RdxDesc->getFastMathFlags()); | ||||||||
9598 | for (unsigned Part = 0; Part < State.UF; ++Part) { | ||||||||
9599 | Value *NewVecOp = State.get(getVecOp(), Part); | ||||||||
9600 | if (VPValue *Cond = getCondOp()) { | ||||||||
9601 | Value *NewCond = State.get(Cond, Part); | ||||||||
9602 | VectorType *VecTy = cast<VectorType>(NewVecOp->getType()); | ||||||||
9603 | Value *Iden = RdxDesc->getRecurrenceIdentity( | ||||||||
9604 | Kind, VecTy->getElementType(), RdxDesc->getFastMathFlags()); | ||||||||
9605 | Value *IdenVec = | ||||||||
9606 | State.Builder.CreateVectorSplat(VecTy->getElementCount(), Iden); | ||||||||
9607 | Value *Select = State.Builder.CreateSelect(NewCond, NewVecOp, IdenVec); | ||||||||
9608 | NewVecOp = Select; | ||||||||
9609 | } | ||||||||
9610 | Value *NewRed; | ||||||||
9611 | Value *NextInChain; | ||||||||
9612 | if (IsOrdered) { | ||||||||
9613 | if (State.VF.isVector()) | ||||||||
9614 | NewRed = createOrderedReduction(State.Builder, *RdxDesc, NewVecOp, | ||||||||
9615 | PrevInChain); | ||||||||
9616 | else | ||||||||
9617 | NewRed = State.Builder.CreateBinOp( | ||||||||
9618 | (Instruction::BinaryOps)RdxDesc->getOpcode(Kind), PrevInChain, | ||||||||
9619 | NewVecOp); | ||||||||
9620 | PrevInChain = NewRed; | ||||||||
9621 | } else { | ||||||||
9622 | PrevInChain = State.get(getChainOp(), Part); | ||||||||
9623 | NewRed = createTargetReduction(State.Builder, TTI, *RdxDesc, NewVecOp); | ||||||||
9624 | } | ||||||||
9625 | if (RecurrenceDescriptor::isMinMaxRecurrenceKind(Kind)) { | ||||||||
9626 | NextInChain = | ||||||||
9627 | createMinMaxOp(State.Builder, RdxDesc->getRecurrenceKind(), | ||||||||
9628 | NewRed, PrevInChain); | ||||||||
9629 | } else if (IsOrdered) | ||||||||
9630 | NextInChain = NewRed; | ||||||||
9631 | else | ||||||||
9632 | NextInChain = State.Builder.CreateBinOp( | ||||||||
9633 | (Instruction::BinaryOps)RdxDesc->getOpcode(Kind), NewRed, | ||||||||
9634 | PrevInChain); | ||||||||
9635 | State.set(this, NextInChain, Part); | ||||||||
9636 | } | ||||||||
9637 | } | ||||||||
9638 | |||||||||
9639 | void VPReplicateRecipe::execute(VPTransformState &State) { | ||||||||
9640 | Instruction *UI = getUnderlyingInstr(); | ||||||||
9641 | if (State.Instance) { // Generate a single instance. | ||||||||
9642 | 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", 9642, __extension__ __PRETTY_FUNCTION__)); | ||||||||
9643 | State.ILV->scalarizeInstruction(UI, this, *State.Instance, | ||||||||
9644 | IsPredicated, State); | ||||||||
9645 | // Insert scalar instance packing it into a vector. | ||||||||
9646 | if (AlsoPack && State.VF.isVector()) { | ||||||||
9647 | // If we're constructing lane 0, initialize to start from poison. | ||||||||
9648 | if (State.Instance->Lane.isFirstLane()) { | ||||||||
9649 | 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", 9649, __extension__ __PRETTY_FUNCTION__)); | ||||||||
9650 | Value *Poison = PoisonValue::get( | ||||||||
9651 | VectorType::get(UI->getType(), State.VF)); | ||||||||
9652 | State.set(this, Poison, State.Instance->Part); | ||||||||
9653 | } | ||||||||
9654 | State.ILV->packScalarIntoVectorValue(this, *State.Instance, State); | ||||||||
9655 | } | ||||||||
9656 | return; | ||||||||
9657 | } | ||||||||
9658 | |||||||||
9659 | if (IsUniform) { | ||||||||
9660 | // If the recipe is uniform across all parts (instead of just per VF), only | ||||||||
9661 | // generate a single instance. | ||||||||
9662 | if ((isa<LoadInst>(UI) || isa<StoreInst>(UI)) && | ||||||||
9663 | all_of(operands(), [](VPValue *Op) { | ||||||||
9664 | return Op->isDefinedOutsideVectorRegions(); | ||||||||
9665 | })) { | ||||||||
9666 | State.ILV->scalarizeInstruction(UI, this, VPIteration(0, 0), IsPredicated, | ||||||||
9667 | State); | ||||||||
9668 | if (user_begin() != user_end()) { | ||||||||
9669 | for (unsigned Part = 1; Part < State.UF; ++Part) | ||||||||
9670 | State.set(this, State.get(this, VPIteration(0, 0)), | ||||||||
9671 | VPIteration(Part, 0)); | ||||||||
9672 | } | ||||||||
9673 | return; | ||||||||
9674 | } | ||||||||
9675 | |||||||||
9676 | // Uniform within VL means we need to generate lane 0 only for each | ||||||||
9677 | // unrolled copy. | ||||||||
9678 | for (unsigned Part = 0; Part < State.UF; ++Part) | ||||||||
9679 | State.ILV->scalarizeInstruction(UI, this, VPIteration(Part, 0), | ||||||||
9680 | IsPredicated, State); | ||||||||
9681 | return; | ||||||||
9682 | } | ||||||||
9683 | |||||||||
9684 | // A store of a loop varying value to a loop invariant address only | ||||||||
9685 | // needs only the last copy of the store. | ||||||||
9686 | if (isa<StoreInst>(UI) && !getOperand(1)->hasDefiningRecipe()) { | ||||||||
9687 | auto Lane = VPLane::getLastLaneForVF(State.VF); | ||||||||
9688 | State.ILV->scalarizeInstruction(UI, this, VPIteration(State.UF - 1, Lane), IsPredicated, | ||||||||
9689 | State); | ||||||||
9690 | return; | ||||||||
9691 | } | ||||||||
9692 | |||||||||
9693 | // Generate scalar instances for all VF lanes of all UF parts. | ||||||||
9694 | 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", 9694, __extension__ __PRETTY_FUNCTION__)); | ||||||||
9695 | const unsigned EndLane = State.VF.getKnownMinValue(); | ||||||||
9696 | for (unsigned Part = 0; Part < State.UF; ++Part) | ||||||||
9697 | for (unsigned Lane = 0; Lane < EndLane; ++Lane) | ||||||||
9698 | State.ILV->scalarizeInstruction(UI, this, VPIteration(Part, Lane), | ||||||||
9699 | IsPredicated, State); | ||||||||
9700 | } | ||||||||
9701 | |||||||||
9702 | void VPWidenMemoryInstructionRecipe::execute(VPTransformState &State) { | ||||||||
9703 | VPValue *StoredValue = isStore() ? getStoredValue() : nullptr; | ||||||||
9704 | |||||||||
9705 | // Attempt to issue a wide load. | ||||||||
9706 | LoadInst *LI = dyn_cast<LoadInst>(&Ingredient); | ||||||||
9707 | StoreInst *SI = dyn_cast<StoreInst>(&Ingredient); | ||||||||
9708 | |||||||||
9709 | 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", 9709, __extension__ __PRETTY_FUNCTION__)); | ||||||||
9710 | 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", 9710, __extension__ __PRETTY_FUNCTION__)); | ||||||||
9711 | 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", 9711, __extension__ __PRETTY_FUNCTION__)); | ||||||||
9712 | |||||||||
9713 | Type *ScalarDataTy = getLoadStoreType(&Ingredient); | ||||||||
9714 | |||||||||
9715 | auto *DataTy = VectorType::get(ScalarDataTy, State.VF); | ||||||||
9716 | const Align Alignment = getLoadStoreAlignment(&Ingredient); | ||||||||
9717 | bool CreateGatherScatter = !Consecutive; | ||||||||
9718 | |||||||||
9719 | auto &Builder = State.Builder; | ||||||||
9720 | InnerLoopVectorizer::VectorParts BlockInMaskParts(State.UF); | ||||||||
9721 | bool isMaskRequired = getMask(); | ||||||||
9722 | if (isMaskRequired) | ||||||||
9723 | for (unsigned Part = 0; Part < State.UF; ++Part) | ||||||||
9724 | BlockInMaskParts[Part] = State.get(getMask(), Part); | ||||||||
9725 | |||||||||
9726 | const auto CreateVecPtr = [&](unsigned Part, Value *Ptr) -> Value * { | ||||||||
9727 | // Calculate the pointer for the specific unroll-part. | ||||||||
9728 | GetElementPtrInst *PartPtr = nullptr; | ||||||||
9729 | |||||||||
9730 | bool InBounds = false; | ||||||||
9731 | if (auto *gep = dyn_cast<GetElementPtrInst>(Ptr->stripPointerCasts())) | ||||||||
9732 | InBounds = gep->isInBounds(); | ||||||||
9733 | if (Reverse) { | ||||||||
9734 | // If the address is consecutive but reversed, then the | ||||||||
9735 | // wide store needs to start at the last vector element. | ||||||||
9736 | // RunTimeVF = VScale * VF.getKnownMinValue() | ||||||||
9737 | // For fixed-width VScale is 1, then RunTimeVF = VF.getKnownMinValue() | ||||||||
9738 | Value *RunTimeVF = getRuntimeVF(Builder, Builder.getInt32Ty(), State.VF); | ||||||||
9739 | // NumElt = -Part * RunTimeVF | ||||||||
9740 | Value *NumElt = Builder.CreateMul(Builder.getInt32(-Part), RunTimeVF); | ||||||||
9741 | // LastLane = 1 - RunTimeVF | ||||||||
9742 | Value *LastLane = Builder.CreateSub(Builder.getInt32(1), RunTimeVF); | ||||||||
9743 | PartPtr = | ||||||||
9744 | cast<GetElementPtrInst>(Builder.CreateGEP(ScalarDataTy, Ptr, NumElt)); | ||||||||
9745 | PartPtr->setIsInBounds(InBounds); | ||||||||
9746 | PartPtr = cast<GetElementPtrInst>( | ||||||||
9747 | Builder.CreateGEP(ScalarDataTy, PartPtr, LastLane)); | ||||||||
9748 | PartPtr->setIsInBounds(InBounds); | ||||||||
9749 | if (isMaskRequired) // Reverse of a null all-one mask is a null mask. | ||||||||
9750 | BlockInMaskParts[Part] = | ||||||||
9751 | Builder.CreateVectorReverse(BlockInMaskParts[Part], "reverse"); | ||||||||
9752 | } else { | ||||||||
9753 | Value *Increment = | ||||||||
9754 | createStepForVF(Builder, Builder.getInt32Ty(), State.VF, Part); | ||||||||
9755 | PartPtr = cast<GetElementPtrInst>( | ||||||||
9756 | Builder.CreateGEP(ScalarDataTy, Ptr, Increment)); | ||||||||
9757 | PartPtr->setIsInBounds(InBounds); | ||||||||
9758 | } | ||||||||
9759 | |||||||||
9760 | unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace(); | ||||||||
9761 | return Builder.CreateBitCast(PartPtr, DataTy->getPointerTo(AddressSpace)); | ||||||||
9762 | }; | ||||||||
9763 | |||||||||
9764 | // Handle Stores: | ||||||||
9765 | if (SI) { | ||||||||
9766 | State.setDebugLocFromInst(SI); | ||||||||
9767 | |||||||||
9768 | for (unsigned Part = 0; Part < State.UF; ++Part) { | ||||||||
9769 | Instruction *NewSI = nullptr; | ||||||||
9770 | Value *StoredVal = State.get(StoredValue, Part); | ||||||||
9771 | if (CreateGatherScatter) { | ||||||||
9772 | Value *MaskPart = isMaskRequired ? BlockInMaskParts[Part] : nullptr; | ||||||||
9773 | Value *VectorGep = State.get(getAddr(), Part); | ||||||||
9774 | NewSI = Builder.CreateMaskedScatter(StoredVal, VectorGep, Alignment, | ||||||||
9775 | MaskPart); | ||||||||
9776 | } else { | ||||||||
9777 | if (Reverse) { | ||||||||
9778 | // If we store to reverse consecutive memory locations, then we need | ||||||||
9779 | // to reverse the order of elements in the stored value. | ||||||||
9780 | StoredVal = Builder.CreateVectorReverse(StoredVal, "reverse"); | ||||||||
9781 | // We don't want to update the value in the map as it might be used in | ||||||||
9782 | // another expression. So don't call resetVectorValue(StoredVal). | ||||||||
9783 | } | ||||||||
9784 | auto *VecPtr = | ||||||||
9785 | CreateVecPtr(Part, State.get(getAddr(), VPIteration(0, 0))); | ||||||||
9786 | if (isMaskRequired) | ||||||||
9787 | NewSI = Builder.CreateMaskedStore(StoredVal, VecPtr, Alignment, | ||||||||
9788 | BlockInMaskParts[Part]); | ||||||||
9789 | else | ||||||||
9790 | NewSI = Builder.CreateAlignedStore(StoredVal, VecPtr, Alignment); | ||||||||
9791 | } | ||||||||
9792 | State.addMetadata(NewSI, SI); | ||||||||
9793 | } | ||||||||
9794 | return; | ||||||||
9795 | } | ||||||||
9796 | |||||||||
9797 | // Handle loads. | ||||||||
9798 | 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", 9798, __extension__ __PRETTY_FUNCTION__)); | ||||||||
9799 | State.setDebugLocFromInst(LI); | ||||||||
9800 | for (unsigned Part = 0; Part < State.UF; ++Part) { | ||||||||
9801 | Value *NewLI; | ||||||||
9802 | if (CreateGatherScatter) { | ||||||||
9803 | Value *MaskPart = isMaskRequired ? BlockInMaskParts[Part] : nullptr; | ||||||||
9804 | Value *VectorGep = State.get(getAddr(), Part); | ||||||||
9805 | NewLI = Builder.CreateMaskedGather(DataTy, VectorGep, Alignment, MaskPart, | ||||||||
9806 | nullptr, "wide.masked.gather"); | ||||||||
9807 | State.addMetadata(NewLI, LI); | ||||||||
9808 | } else { | ||||||||
9809 | auto *VecPtr = | ||||||||
9810 | CreateVecPtr(Part, State.get(getAddr(), VPIteration(0, 0))); | ||||||||
9811 | if (isMaskRequired) | ||||||||
9812 | NewLI = Builder.CreateMaskedLoad( | ||||||||
9813 | DataTy, VecPtr, Alignment, BlockInMaskParts[Part], | ||||||||
9814 | PoisonValue::get(DataTy), "wide.masked.load"); | ||||||||
9815 | else | ||||||||
9816 | NewLI = | ||||||||
9817 | Builder.CreateAlignedLoad(DataTy, VecPtr, Alignment, "wide.load"); | ||||||||
9818 | |||||||||
9819 | // Add metadata to the load, but setVectorValue to the reverse shuffle. | ||||||||
9820 | State.addMetadata(NewLI, LI); | ||||||||
9821 | if (Reverse) | ||||||||
9822 | NewLI = Builder.CreateVectorReverse(NewLI, "reverse"); | ||||||||
9823 | } | ||||||||
9824 | |||||||||
9825 | State.set(getVPSingleValue(), NewLI, Part); | ||||||||
9826 | } | ||||||||
9827 | } | ||||||||
9828 | |||||||||
9829 | // Determine how to lower the scalar epilogue, which depends on 1) optimising | ||||||||
9830 | // for minimum code-size, 2) predicate compiler options, 3) loop hints forcing | ||||||||
9831 | // predication, and 4) a TTI hook that analyses whether the loop is suitable | ||||||||
9832 | // for predication. | ||||||||
9833 | static ScalarEpilogueLowering getScalarEpilogueLowering( | ||||||||
9834 | Function *F, Loop *L, LoopVectorizeHints &Hints, ProfileSummaryInfo *PSI, | ||||||||
9835 | BlockFrequencyInfo *BFI, TargetTransformInfo *TTI, TargetLibraryInfo *TLI, | ||||||||
9836 | AssumptionCache *AC, LoopInfo *LI, ScalarEvolution *SE, DominatorTree *DT, | ||||||||
9837 | LoopVectorizationLegality &LVL, InterleavedAccessInfo *IAI) { | ||||||||
9838 | // 1) OptSize takes precedence over all other options, i.e. if this is set, | ||||||||
9839 | // don't look at hints or options, and don't request a scalar epilogue. | ||||||||
9840 | // (For PGSO, as shouldOptimizeForSize isn't currently accessible from | ||||||||
9841 | // LoopAccessInfo (due to code dependency and not being able to reliably get | ||||||||
9842 | // PSI/BFI from a loop analysis under NPM), we cannot suppress the collection | ||||||||
9843 | // of strides in LoopAccessInfo::analyzeLoop() and vectorize without | ||||||||
9844 | // versioning when the vectorization is forced, unlike hasOptSize. So revert | ||||||||
9845 | // back to the old way and vectorize with versioning when forced. See D81345.) | ||||||||
9846 | if (F->hasOptSize() || (llvm::shouldOptimizeForSize(L->getHeader(), PSI, BFI, | ||||||||
9847 | PGSOQueryType::IRPass) && | ||||||||
9848 | Hints.getForce() != LoopVectorizeHints::FK_Enabled)) | ||||||||
9849 | return CM_ScalarEpilogueNotAllowedOptSize; | ||||||||
9850 | |||||||||
9851 | // 2) If set, obey the directives | ||||||||
9852 | if (PreferPredicateOverEpilogue.getNumOccurrences()) { | ||||||||
9853 | switch (PreferPredicateOverEpilogue) { | ||||||||
9854 | case PreferPredicateTy::ScalarEpilogue: | ||||||||
9855 | return CM_ScalarEpilogueAllowed; | ||||||||
9856 | case PreferPredicateTy::PredicateElseScalarEpilogue: | ||||||||
9857 | return CM_ScalarEpilogueNotNeededUsePredicate; | ||||||||
9858 | case PreferPredicateTy::PredicateOrDontVectorize: | ||||||||
9859 | return CM_ScalarEpilogueNotAllowedUsePredicate; | ||||||||
9860 | }; | ||||||||
9861 | } | ||||||||
9862 | |||||||||
9863 | // 3) If set, obey the hints | ||||||||
9864 | switch (Hints.getPredicate()) { | ||||||||
9865 | case LoopVectorizeHints::FK_Enabled: | ||||||||
9866 | return CM_ScalarEpilogueNotNeededUsePredicate; | ||||||||
9867 | case LoopVectorizeHints::FK_Disabled: | ||||||||
9868 | return CM_ScalarEpilogueAllowed; | ||||||||
9869 | }; | ||||||||
9870 | |||||||||
9871 | // 4) if the TTI hook indicates this is profitable, request predication. | ||||||||
9872 | if (TTI->preferPredicateOverEpilogue(L, LI, *SE, *AC, TLI, DT, &LVL, IAI)) | ||||||||
9873 | return CM_ScalarEpilogueNotNeededUsePredicate; | ||||||||
9874 | |||||||||
9875 | return CM_ScalarEpilogueAllowed; | ||||||||
9876 | } | ||||||||
9877 | |||||||||
9878 | Value *VPTransformState::get(VPValue *Def, unsigned Part) { | ||||||||
9879 | // If Values have been set for this Def return the one relevant for \p Part. | ||||||||
9880 | if (hasVectorValue(Def, Part)) | ||||||||
9881 | return Data.PerPartOutput[Def][Part]; | ||||||||
9882 | |||||||||
9883 | if (!hasScalarValue(Def, {Part, 0})) { | ||||||||
9884 | Value *IRV = Def->getLiveInIRValue(); | ||||||||
9885 | Value *B = ILV->getBroadcastInstrs(IRV); | ||||||||
9886 | set(Def, B, Part); | ||||||||
9887 | return B; | ||||||||
9888 | } | ||||||||
9889 | |||||||||
9890 | Value *ScalarValue = get(Def, {Part, 0}); | ||||||||
9891 | // If we aren't vectorizing, we can just copy the scalar map values over | ||||||||
9892 | // to the vector map. | ||||||||
9893 | if (VF.isScalar()) { | ||||||||
9894 | set(Def, ScalarValue, Part); | ||||||||
9895 | return ScalarValue; | ||||||||
9896 | } | ||||||||
9897 | |||||||||
9898 | bool IsUniform = vputils::isUniformAfterVectorization(Def); | ||||||||
9899 | |||||||||
9900 | unsigned LastLane = IsUniform ? 0 : VF.getKnownMinValue() - 1; | ||||||||
9901 | // Check if there is a scalar value for the selected lane. | ||||||||
9902 | if (!hasScalarValue(Def, {Part, LastLane})) { | ||||||||
9903 | // At the moment, VPWidenIntOrFpInductionRecipes and VPScalarIVStepsRecipes can also be uniform. | ||||||||
9904 | assert((isa<VPWidenIntOrFpInductionRecipe>(Def->getDefiningRecipe()) ||(static_cast <bool> ((isa<VPWidenIntOrFpInductionRecipe >(Def->getDefiningRecipe()) || isa<VPScalarIVStepsRecipe >(Def->getDefiningRecipe())) && "unexpected recipe found to be invariant" ) ? void (0) : __assert_fail ("(isa<VPWidenIntOrFpInductionRecipe>(Def->getDefiningRecipe()) || isa<VPScalarIVStepsRecipe>(Def->getDefiningRecipe())) && \"unexpected recipe found to be invariant\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9906, __extension__ __PRETTY_FUNCTION__)) | ||||||||
9905 | isa<VPScalarIVStepsRecipe>(Def->getDefiningRecipe())) &&(static_cast <bool> ((isa<VPWidenIntOrFpInductionRecipe >(Def->getDefiningRecipe()) || isa<VPScalarIVStepsRecipe >(Def->getDefiningRecipe())) && "unexpected recipe found to be invariant" ) ? void (0) : __assert_fail ("(isa<VPWidenIntOrFpInductionRecipe>(Def->getDefiningRecipe()) || isa<VPScalarIVStepsRecipe>(Def->getDefiningRecipe())) && \"unexpected recipe found to be invariant\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9906, __extension__ __PRETTY_FUNCTION__)) | ||||||||
9906 | "unexpected recipe found to be invariant")(static_cast <bool> ((isa<VPWidenIntOrFpInductionRecipe >(Def->getDefiningRecipe()) || isa<VPScalarIVStepsRecipe >(Def->getDefiningRecipe())) && "unexpected recipe found to be invariant" ) ? void (0) : __assert_fail ("(isa<VPWidenIntOrFpInductionRecipe>(Def->getDefiningRecipe()) || isa<VPScalarIVStepsRecipe>(Def->getDefiningRecipe())) && \"unexpected recipe found to be invariant\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9906, __extension__ __PRETTY_FUNCTION__)); | ||||||||
9907 | IsUniform = true; | ||||||||
9908 | LastLane = 0; | ||||||||
9909 | } | ||||||||
9910 | |||||||||
9911 | auto *LastInst = cast<Instruction>(get(Def, {Part, LastLane})); | ||||||||
9912 | // Set the insert point after the last scalarized instruction or after the | ||||||||
9913 | // last PHI, if LastInst is a PHI. This ensures the insertelement sequence | ||||||||
9914 | // will directly follow the scalar definitions. | ||||||||
9915 | auto OldIP = Builder.saveIP(); | ||||||||
9916 | auto NewIP = | ||||||||
9917 | isa<PHINode>(LastInst) | ||||||||
9918 | ? BasicBlock::iterator(LastInst->getParent()->getFirstNonPHI()) | ||||||||
9919 | : std::next(BasicBlock::iterator(LastInst)); | ||||||||
9920 | Builder.SetInsertPoint(&*NewIP); | ||||||||
9921 | |||||||||
9922 | // However, if we are vectorizing, we need to construct the vector values. | ||||||||
9923 | // If the value is known to be uniform after vectorization, we can just | ||||||||
9924 | // broadcast the scalar value corresponding to lane zero for each unroll | ||||||||
9925 | // iteration. Otherwise, we construct the vector values using | ||||||||
9926 | // insertelement instructions. Since the resulting vectors are stored in | ||||||||
9927 | // State, we will only generate the insertelements once. | ||||||||
9928 | Value *VectorValue = nullptr; | ||||||||
9929 | if (IsUniform) { | ||||||||
9930 | VectorValue = ILV->getBroadcastInstrs(ScalarValue); | ||||||||
9931 | set(Def, VectorValue, Part); | ||||||||
9932 | } else { | ||||||||
9933 | // Initialize packing with insertelements to start from undef. | ||||||||
9934 | 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", 9934, __extension__ __PRETTY_FUNCTION__)); | ||||||||
9935 | Value *Undef = PoisonValue::get(VectorType::get(LastInst->getType(), VF)); | ||||||||
9936 | set(Def, Undef, Part); | ||||||||
9937 | for (unsigned Lane = 0; Lane < VF.getKnownMinValue(); ++Lane) | ||||||||
9938 | ILV->packScalarIntoVectorValue(Def, {Part, Lane}, *this); | ||||||||
9939 | VectorValue = get(Def, Part); | ||||||||
9940 | } | ||||||||
9941 | Builder.restoreIP(OldIP); | ||||||||
9942 | return VectorValue; | ||||||||
9943 | } | ||||||||
9944 | |||||||||
9945 | // Process the loop in the VPlan-native vectorization path. This path builds | ||||||||
9946 | // VPlan upfront in the vectorization pipeline, which allows to apply | ||||||||
9947 | // VPlan-to-VPlan transformations from the very beginning without modifying the | ||||||||
9948 | // input LLVM IR. | ||||||||
9949 | static bool processLoopInVPlanNativePath( | ||||||||
9950 | Loop *L, PredicatedScalarEvolution &PSE, LoopInfo *LI, DominatorTree *DT, | ||||||||
9951 | LoopVectorizationLegality *LVL, TargetTransformInfo *TTI, | ||||||||
9952 | TargetLibraryInfo *TLI, DemandedBits *DB, AssumptionCache *AC, | ||||||||
9953 | OptimizationRemarkEmitter *ORE, BlockFrequencyInfo *BFI, | ||||||||
9954 | ProfileSummaryInfo *PSI, LoopVectorizeHints &Hints, | ||||||||
9955 | LoopVectorizationRequirements &Requirements) { | ||||||||
9956 | |||||||||
9957 | if (isa<SCEVCouldNotCompute>(PSE.getBackedgeTakenCount())) { | ||||||||
9958 | 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); | ||||||||
9959 | return false; | ||||||||
9960 | } | ||||||||
9961 | 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", 9961, __extension__ __PRETTY_FUNCTION__)); | ||||||||
9962 | Function *F = L->getHeader()->getParent(); | ||||||||
9963 | InterleavedAccessInfo IAI(PSE, L, DT, LI, LVL->getLAI()); | ||||||||
9964 | |||||||||
9965 | ScalarEpilogueLowering SEL = getScalarEpilogueLowering( | ||||||||
9966 | F, L, Hints, PSI, BFI, TTI, TLI, AC, LI, PSE.getSE(), DT, *LVL, &IAI); | ||||||||
9967 | |||||||||
9968 | LoopVectorizationCostModel CM(SEL, L, PSE, LI, LVL, *TTI, TLI, DB, AC, ORE, F, | ||||||||
9969 | &Hints, IAI); | ||||||||
9970 | // Use the planner for outer loop vectorization. | ||||||||
9971 | // TODO: CM is not used at this point inside the planner. Turn CM into an | ||||||||
9972 | // optional argument if we don't need it in the future. | ||||||||
9973 | LoopVectorizationPlanner LVP(L, LI, TLI, TTI, LVL, CM, IAI, PSE, Hints, ORE); | ||||||||
9974 | |||||||||
9975 | // Get user vectorization factor. | ||||||||
9976 | ElementCount UserVF = Hints.getWidth(); | ||||||||
9977 | |||||||||
9978 | CM.collectElementTypesForWidening(); | ||||||||
9979 | |||||||||
9980 | // Plan how to best vectorize, return the best VF and its cost. | ||||||||
9981 | const VectorizationFactor VF = LVP.planInVPlanNativePath(UserVF); | ||||||||
9982 | |||||||||
9983 | // If we are stress testing VPlan builds, do not attempt to generate vector | ||||||||
9984 | // code. Masked vector code generation support will follow soon. | ||||||||
9985 | // Also, do not attempt to vectorize if no vector code will be produced. | ||||||||
9986 | if (VPlanBuildStressTest || VectorizationFactor::Disabled() == VF) | ||||||||
9987 | return false; | ||||||||
9988 | |||||||||
9989 | VPlan &BestPlan = LVP.getBestPlanFor(VF.Width); | ||||||||
9990 | |||||||||
9991 | { | ||||||||
9992 | GeneratedRTChecks Checks(*PSE.getSE(), DT, LI, TTI, | ||||||||
9993 | F->getParent()->getDataLayout()); | ||||||||
9994 | InnerLoopVectorizer LB(L, PSE, LI, DT, TLI, TTI, AC, ORE, VF.Width, | ||||||||
9995 | VF.Width, 1, LVL, &CM, BFI, PSI, Checks); | ||||||||
9996 | 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) | ||||||||
9997 | << 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); | ||||||||
9998 | LVP.executePlan(VF.Width, 1, BestPlan, LB, DT, false); | ||||||||
9999 | } | ||||||||
10000 | |||||||||
10001 | // Mark the loop as already vectorized to avoid vectorizing again. | ||||||||
10002 | Hints.setAlreadyVectorized(); | ||||||||
10003 | 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", 10003, __extension__ __PRETTY_FUNCTION__)); | ||||||||
10004 | return true; | ||||||||
10005 | } | ||||||||
10006 | |||||||||
10007 | // Emit a remark if there are stores to floats that required a floating point | ||||||||
10008 | // extension. If the vectorized loop was generated with floating point there | ||||||||
10009 | // will be a performance penalty from the conversion overhead and the change in | ||||||||
10010 | // the vector width. | ||||||||
10011 | static void checkMixedPrecision(Loop *L, OptimizationRemarkEmitter *ORE) { | ||||||||
10012 | SmallVector<Instruction *, 4> Worklist; | ||||||||
10013 | for (BasicBlock *BB : L->getBlocks()) { | ||||||||
10014 | for (Instruction &Inst : *BB) { | ||||||||
10015 | if (auto *S = dyn_cast<StoreInst>(&Inst)) { | ||||||||
10016 | if (S->getValueOperand()->getType()->isFloatTy()) | ||||||||
10017 | Worklist.push_back(S); | ||||||||
10018 | } | ||||||||
10019 | } | ||||||||
10020 | } | ||||||||
10021 | |||||||||
10022 | // Traverse the floating point stores upwards searching, for floating point | ||||||||
10023 | // conversions. | ||||||||
10024 | SmallPtrSet<const Instruction *, 4> Visited; | ||||||||
10025 | SmallPtrSet<const Instruction *, 4> EmittedRemark; | ||||||||
10026 | while (!Worklist.empty()) { | ||||||||
10027 | auto *I = Worklist.pop_back_val(); | ||||||||
10028 | if (!L->contains(I)) | ||||||||
10029 | continue; | ||||||||
10030 | if (!Visited.insert(I).second) | ||||||||
10031 | continue; | ||||||||
10032 | |||||||||
10033 | // Emit a remark if the floating point store required a floating | ||||||||
10034 | // point conversion. | ||||||||
10035 | // TODO: More work could be done to identify the root cause such as a | ||||||||
10036 | // constant or a function return type and point the user to it. | ||||||||
10037 | if (isa<FPExtInst>(I) && EmittedRemark.insert(I).second) | ||||||||
10038 | ORE->emit([&]() { | ||||||||
10039 | return OptimizationRemarkAnalysis(LV_NAME"loop-vectorize", "VectorMixedPrecision", | ||||||||
10040 | I->getDebugLoc(), L->getHeader()) | ||||||||
10041 | << "floating point conversion changes vector width. " | ||||||||
10042 | << "Mixed floating point precision requires an up/down " | ||||||||
10043 | << "cast that will negatively impact performance."; | ||||||||
10044 | }); | ||||||||
10045 | |||||||||
10046 | for (Use &Op : I->operands()) | ||||||||
10047 | if (auto *OpI = dyn_cast<Instruction>(Op)) | ||||||||
10048 | Worklist.push_back(OpI); | ||||||||
10049 | } | ||||||||
10050 | } | ||||||||
10051 | |||||||||
10052 | static bool areRuntimeChecksProfitable(GeneratedRTChecks &Checks, | ||||||||
10053 | VectorizationFactor &VF, | ||||||||
10054 | std::optional<unsigned> VScale, Loop *L, | ||||||||
10055 | ScalarEvolution &SE) { | ||||||||
10056 | InstructionCost CheckCost = Checks.getCost(); | ||||||||
10057 | if (!CheckCost.isValid()) | ||||||||
10058 | return false; | ||||||||
10059 | |||||||||
10060 | // When interleaving only scalar and vector cost will be equal, which in turn | ||||||||
10061 | // would lead to a divide by 0. Fall back to hard threshold. | ||||||||
10062 | if (VF.Width.isScalar()) { | ||||||||
10063 | if (CheckCost > VectorizeMemoryCheckThreshold) { | ||||||||
10064 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Interleaving only is not profitable due to runtime checks\n" ; } } while (false) | ||||||||
10065 | dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Interleaving only is not profitable due to runtime checks\n" ; } } while (false) | ||||||||
10066 | << "LV: Interleaving only is not profitable due to runtime checks\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Interleaving only is not profitable due to runtime checks\n" ; } } while (false); | ||||||||
10067 | return false; | ||||||||
10068 | } | ||||||||
10069 | return true; | ||||||||
10070 | } | ||||||||
10071 | |||||||||
10072 | // The scalar cost should only be 0 when vectorizing with a user specified VF/IC. In those cases, runtime checks should always be generated. | ||||||||
10073 | double ScalarC = *VF.ScalarCost.getValue(); | ||||||||
10074 | if (ScalarC == 0) | ||||||||
10075 | return true; | ||||||||
10076 | |||||||||
10077 | // First, compute the minimum iteration count required so that the vector | ||||||||
10078 | // loop outperforms the scalar loop. | ||||||||
10079 | // The total cost of the scalar loop is | ||||||||
10080 | // ScalarC * TC | ||||||||
10081 | // where | ||||||||
10082 | // * TC is the actual trip count of the loop. | ||||||||
10083 | // * ScalarC is the cost of a single scalar iteration. | ||||||||
10084 | // | ||||||||
10085 | // The total cost of the vector loop is | ||||||||
10086 | // RtC + VecC * (TC / VF) + EpiC | ||||||||
10087 | // where | ||||||||
10088 | // * RtC is the cost of the generated runtime checks | ||||||||
10089 | // * VecC is the cost of a single vector iteration. | ||||||||
10090 | // * TC is the actual trip count of the loop | ||||||||
10091 | // * VF is the vectorization factor | ||||||||
10092 | // * EpiCost is the cost of the generated epilogue, including the cost | ||||||||
10093 | // of the remaining scalar operations. | ||||||||
10094 | // | ||||||||
10095 | // Vectorization is profitable once the total vector cost is less than the | ||||||||
10096 | // total scalar cost: | ||||||||
10097 | // RtC + VecC * (TC / VF) + EpiC < ScalarC * TC | ||||||||
10098 | // | ||||||||
10099 | // Now we can compute the minimum required trip count TC as | ||||||||
10100 | // (RtC + EpiC) / (ScalarC - (VecC / VF)) < TC | ||||||||
10101 | // | ||||||||
10102 | // For now we assume the epilogue cost EpiC = 0 for simplicity. Note that | ||||||||
10103 | // the computations are performed on doubles, not integers and the result | ||||||||
10104 | // is rounded up, hence we get an upper estimate of the TC. | ||||||||
10105 | unsigned IntVF = VF.Width.getKnownMinValue(); | ||||||||
10106 | if (VF.Width.isScalable()) { | ||||||||
10107 | unsigned AssumedMinimumVscale = 1; | ||||||||
10108 | if (VScale) | ||||||||
10109 | AssumedMinimumVscale = *VScale; | ||||||||
10110 | IntVF *= AssumedMinimumVscale; | ||||||||
10111 | } | ||||||||
10112 | double VecCOverVF = double(*VF.Cost.getValue()) / IntVF; | ||||||||
10113 | double RtC = *CheckCost.getValue(); | ||||||||
10114 | double MinTC1 = RtC / (ScalarC - VecCOverVF); | ||||||||
10115 | |||||||||
10116 | // Second, compute a minimum iteration count so that the cost of the | ||||||||
10117 | // runtime checks is only a fraction of the total scalar loop cost. This | ||||||||
10118 | // adds a loop-dependent bound on the overhead incurred if the runtime | ||||||||
10119 | // checks fail. In case the runtime checks fail, the cost is RtC + ScalarC | ||||||||
10120 | // * TC. To bound the runtime check to be a fraction 1/X of the scalar | ||||||||
10121 | // cost, compute | ||||||||
10122 | // RtC < ScalarC * TC * (1 / X) ==> RtC * X / ScalarC < TC | ||||||||
10123 | double MinTC2 = RtC * 10 / ScalarC; | ||||||||
10124 | |||||||||
10125 | // Now pick the larger minimum. If it is not a multiple of VF, choose the | ||||||||
10126 | // next closest multiple of VF. This should partly compensate for ignoring | ||||||||
10127 | // the epilogue cost. | ||||||||
10128 | uint64_t MinTC = std::ceil(std::max(MinTC1, MinTC2)); | ||||||||
10129 | VF.MinProfitableTripCount = ElementCount::getFixed(alignTo(MinTC, IntVF)); | ||||||||
10130 | |||||||||
10131 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Minimum required TC for runtime checks to be profitable:" << VF.MinProfitableTripCount << "\n"; } } while ( false) | ||||||||
10132 | dbgs() << "LV: Minimum required TC for runtime checks to be profitable:"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Minimum required TC for runtime checks to be profitable:" << VF.MinProfitableTripCount << "\n"; } } while ( false) | ||||||||
10133 | << VF.MinProfitableTripCount << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Minimum required TC for runtime checks to be profitable:" << VF.MinProfitableTripCount << "\n"; } } while ( false); | ||||||||
10134 | |||||||||
10135 | // Skip vectorization if the expected trip count is less than the minimum | ||||||||
10136 | // required trip count. | ||||||||
10137 | if (auto ExpectedTC = getSmallBestKnownTC(SE, L)) { | ||||||||
10138 | if (ElementCount::isKnownLT(ElementCount::getFixed(*ExpectedTC), | ||||||||
10139 | VF.MinProfitableTripCount)) { | ||||||||
10140 | LLVM_DEBUG(dbgs() << "LV: Vectorization is not beneficial: expected "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Vectorization is not beneficial: expected " "trip count < minimum profitable VF (" << *ExpectedTC << " < " << VF.MinProfitableTripCount << ")\n"; } } while (false) | ||||||||
10141 | "trip count < minimum profitable VF ("do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Vectorization is not beneficial: expected " "trip count < minimum profitable VF (" << *ExpectedTC << " < " << VF.MinProfitableTripCount << ")\n"; } } while (false) | ||||||||
10142 | << *ExpectedTC << " < " << VF.MinProfitableTripCountdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Vectorization is not beneficial: expected " "trip count < minimum profitable VF (" << *ExpectedTC << " < " << VF.MinProfitableTripCount << ")\n"; } } while (false) | ||||||||
10143 | << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Vectorization is not beneficial: expected " "trip count < minimum profitable VF (" << *ExpectedTC << " < " << VF.MinProfitableTripCount << ")\n"; } } while (false); | ||||||||
10144 | |||||||||
10145 | return false; | ||||||||
10146 | } | ||||||||
10147 | } | ||||||||
10148 | return true; | ||||||||
10149 | } | ||||||||
10150 | |||||||||
10151 | LoopVectorizePass::LoopVectorizePass(LoopVectorizeOptions Opts) | ||||||||
10152 | : InterleaveOnlyWhenForced(Opts.InterleaveOnlyWhenForced || | ||||||||
10153 | !EnableLoopInterleaving), | ||||||||
10154 | VectorizeOnlyWhenForced(Opts.VectorizeOnlyWhenForced || | ||||||||
10155 | !EnableLoopVectorization) {} | ||||||||
10156 | |||||||||
10157 | bool LoopVectorizePass::processLoop(Loop *L) { | ||||||||
10158 | 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", 10159, __extension__ __PRETTY_FUNCTION__)) | ||||||||
10159 | "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", 10159, __extension__ __PRETTY_FUNCTION__)); | ||||||||
10160 | |||||||||
10161 | #ifndef NDEBUG | ||||||||
10162 | const std::string DebugLocStr = getDebugLocString(L); | ||||||||
10163 | #endif /* NDEBUG */ | ||||||||
10164 | |||||||||
10165 | 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 ) | ||||||||
10166 | << 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 ) | ||||||||
10167 | << 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 ); | ||||||||
10168 | |||||||||
10169 | LoopVectorizeHints Hints(L, InterleaveOnlyWhenForced, *ORE, TTI); | ||||||||
10170 | |||||||||
10171 | 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) | ||||||||
10172 | 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) | ||||||||
10173 | << " 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) | ||||||||
10174 | << (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) | ||||||||
10175 | ? "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) | ||||||||
10176 | : (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) | ||||||||
10177 | ? "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) | ||||||||
10178 | : "?"))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) | ||||||||
10179 | << " 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) | ||||||||
10180 | << " 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); | ||||||||
10181 | |||||||||
10182 | // Function containing loop | ||||||||
10183 | Function *F = L->getHeader()->getParent(); | ||||||||
10184 | |||||||||
10185 | // Looking at the diagnostic output is the only way to determine if a loop | ||||||||
10186 | // was vectorized (other than looking at the IR or machine code), so it | ||||||||
10187 | // is important to generate an optimization remark for each loop. Most of | ||||||||
10188 | // these messages are generated as OptimizationRemarkAnalysis. Remarks | ||||||||
10189 | // generated as OptimizationRemark and OptimizationRemarkMissed are | ||||||||
10190 | // less verbose reporting vectorized loops and unvectorized loops that may | ||||||||
10191 | // benefit from vectorization, respectively. | ||||||||
10192 | |||||||||
10193 | if (!Hints.allowVectorization(F, L, VectorizeOnlyWhenForced)) { | ||||||||
10194 | 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); | ||||||||
10195 | return false; | ||||||||
10196 | } | ||||||||
10197 | |||||||||
10198 | PredicatedScalarEvolution PSE(*SE, *L); | ||||||||
10199 | |||||||||
10200 | // Check if it is legal to vectorize the loop. | ||||||||
10201 | LoopVectorizationRequirements Requirements; | ||||||||
10202 | LoopVectorizationLegality LVL(L, PSE, DT, TTI, TLI, F, *LAIs, LI, ORE, | ||||||||
10203 | &Requirements, &Hints, DB, AC, BFI, PSI); | ||||||||
10204 | if (!LVL.canVectorize(EnableVPlanNativePath)) { | ||||||||
10205 | 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); | ||||||||
10206 | Hints.emitRemarkWithHints(); | ||||||||
10207 | return false; | ||||||||
10208 | } | ||||||||
10209 | |||||||||
10210 | // Entrance to the VPlan-native vectorization path. Outer loops are processed | ||||||||
10211 | // here. They may require CFG and instruction level transformations before | ||||||||
10212 | // even evaluating whether vectorization is profitable. Since we cannot modify | ||||||||
10213 | // the incoming IR, we need to build VPlan upfront in the vectorization | ||||||||
10214 | // pipeline. | ||||||||
10215 | if (!L->isInnermost()) | ||||||||
10216 | return processLoopInVPlanNativePath(L, PSE, LI, DT, &LVL, TTI, TLI, DB, AC, | ||||||||
10217 | ORE, BFI, PSI, Hints, Requirements); | ||||||||
10218 | |||||||||
10219 | 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", 10219, __extension__ __PRETTY_FUNCTION__)); | ||||||||
10220 | |||||||||
10221 | InterleavedAccessInfo IAI(PSE, L, DT, LI, LVL.getLAI()); | ||||||||
10222 | bool UseInterleaved = TTI->enableInterleavedAccessVectorization(); | ||||||||
10223 | |||||||||
10224 | // If an override option has been passed in for interleaved accesses, use it. | ||||||||
10225 | if (EnableInterleavedMemAccesses.getNumOccurrences() > 0) | ||||||||
10226 | UseInterleaved = EnableInterleavedMemAccesses; | ||||||||
10227 | |||||||||
10228 | // Analyze interleaved memory accesses. | ||||||||
10229 | if (UseInterleaved) | ||||||||
10230 | IAI.analyzeInterleaving(useMaskedInterleavedAccesses(*TTI)); | ||||||||
10231 | |||||||||
10232 | // Check the function attributes and profiles to find out if this function | ||||||||
10233 | // should be optimized for size. | ||||||||
10234 | ScalarEpilogueLowering SEL = getScalarEpilogueLowering( | ||||||||
10235 | F, L, Hints, PSI, BFI, TTI, TLI, AC, LI, PSE.getSE(), DT, LVL, &IAI); | ||||||||
10236 | |||||||||
10237 | // Check the loop for a trip count threshold: vectorize loops with a tiny trip | ||||||||
10238 | // count by optimizing for size, to minimize overheads. | ||||||||
10239 | auto ExpectedTC = getSmallBestKnownTC(*SE, L); | ||||||||
10240 | if (ExpectedTC && *ExpectedTC < TinyTripCountVectorThreshold) { | ||||||||
10241 | 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 ) | ||||||||
10242 | << "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 ) | ||||||||
10243 | << "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 ); | ||||||||
10244 | if (Hints.getForce() == LoopVectorizeHints::FK_Enabled) | ||||||||
10245 | 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); | ||||||||
10246 | else { | ||||||||
10247 | if (*ExpectedTC > TTI->getMinTripCountTailFoldingThreshold()) { | ||||||||
10248 | LLVM_DEBUG(dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "\n"; } } while (false); | ||||||||
10249 | SEL = CM_ScalarEpilogueNotAllowedLowTripLoop; | ||||||||
10250 | } else { | ||||||||
10251 | LLVM_DEBUG(dbgs() << " But the target considers the trip count too "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << " But the target considers the trip count too " "small to consider vectorizing.\n"; } } while (false) | ||||||||
10252 | "small to consider vectorizing.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << " But the target considers the trip count too " "small to consider vectorizing.\n"; } } while (false); | ||||||||
10253 | reportVectorizationFailure( | ||||||||
10254 | "The trip count is below the minial threshold value.", | ||||||||
10255 | "loop trip count is too low, avoiding vectorization", | ||||||||
10256 | "LowTripCount", ORE, L); | ||||||||
10257 | Hints.emitRemarkWithHints(); | ||||||||
10258 | return false; | ||||||||
10259 | } | ||||||||
10260 | } | ||||||||
10261 | } | ||||||||
10262 | |||||||||
10263 | // Check the function attributes to see if implicit floats or vectors are | ||||||||
10264 | // allowed. | ||||||||
10265 | if (F->hasFnAttribute(Attribute::NoImplicitFloat)) { | ||||||||
10266 | reportVectorizationFailure( | ||||||||
10267 | "Can't vectorize when the NoImplicitFloat attribute is used", | ||||||||
10268 | "loop not vectorized due to NoImplicitFloat attribute", | ||||||||
10269 | "NoImplicitFloat", ORE, L); | ||||||||
10270 | Hints.emitRemarkWithHints(); | ||||||||
10271 | return false; | ||||||||
10272 | } | ||||||||
10273 | |||||||||
10274 | // Check if the target supports potentially unsafe FP vectorization. | ||||||||
10275 | // FIXME: Add a check for the type of safety issue (denormal, signaling) | ||||||||
10276 | // for the target we're vectorizing for, to make sure none of the | ||||||||
10277 | // additional fp-math flags can help. | ||||||||
10278 | if (Hints.isPotentiallyUnsafe() && | ||||||||
10279 | TTI->isFPVectorizationPotentiallyUnsafe()) { | ||||||||
10280 | reportVectorizationFailure( | ||||||||
10281 | "Potentially unsafe FP op prevents vectorization", | ||||||||
10282 | "loop not vectorized due to unsafe FP support.", | ||||||||
10283 | "UnsafeFP", ORE, L); | ||||||||
10284 | Hints.emitRemarkWithHints(); | ||||||||
10285 | return false; | ||||||||
10286 | } | ||||||||
10287 | |||||||||
10288 | bool AllowOrderedReductions; | ||||||||
10289 | // If the flag is set, use that instead and override the TTI behaviour. | ||||||||
10290 | if (ForceOrderedReductions.getNumOccurrences() > 0) | ||||||||
10291 | AllowOrderedReductions = ForceOrderedReductions; | ||||||||
10292 | else | ||||||||
10293 | AllowOrderedReductions = TTI->enableOrderedReductions(); | ||||||||
10294 | if (!LVL.canVectorizeFPMath(AllowOrderedReductions)) { | ||||||||
10295 | ORE->emit([&]() { | ||||||||
10296 | auto *ExactFPMathInst = Requirements.getExactFPInst(); | ||||||||
10297 | return OptimizationRemarkAnalysisFPCommute(DEBUG_TYPE"loop-vectorize", "CantReorderFPOps", | ||||||||
10298 | ExactFPMathInst->getDebugLoc(), | ||||||||
10299 | ExactFPMathInst->getParent()) | ||||||||
10300 | << "loop not vectorized: cannot prove it is safe to reorder " | ||||||||
10301 | "floating-point operations"; | ||||||||
10302 | }); | ||||||||
10303 | 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) | ||||||||
10304 | "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); | ||||||||
10305 | Hints.emitRemarkWithHints(); | ||||||||
10306 | return false; | ||||||||
10307 | } | ||||||||
10308 | |||||||||
10309 | // Use the cost model. | ||||||||
10310 | LoopVectorizationCostModel CM(SEL, L, PSE, LI, &LVL, *TTI, TLI, DB, AC, ORE, | ||||||||
10311 | F, &Hints, IAI); | ||||||||
10312 | CM.collectValuesToIgnore(); | ||||||||
10313 | CM.collectElementTypesForWidening(); | ||||||||
10314 | |||||||||
10315 | // Use the planner for vectorization. | ||||||||
10316 | LoopVectorizationPlanner LVP(L, LI, TLI, TTI, &LVL, CM, IAI, PSE, Hints, ORE); | ||||||||
10317 | |||||||||
10318 | // Get user vectorization factor and interleave count. | ||||||||
10319 | ElementCount UserVF = Hints.getWidth(); | ||||||||
10320 | unsigned UserIC = Hints.getInterleave(); | ||||||||
10321 | |||||||||
10322 | // Plan how to best vectorize, return the best VF and its cost. | ||||||||
10323 | std::optional<VectorizationFactor> MaybeVF = LVP.plan(UserVF, UserIC); | ||||||||
10324 | |||||||||
10325 | VectorizationFactor VF = VectorizationFactor::Disabled(); | ||||||||
10326 | unsigned IC = 1; | ||||||||
10327 | |||||||||
10328 | GeneratedRTChecks Checks(*PSE.getSE(), DT, LI, TTI, | ||||||||
10329 | F->getParent()->getDataLayout()); | ||||||||
10330 | if (MaybeVF) { | ||||||||
10331 | VF = *MaybeVF; | ||||||||
10332 | // Select the interleave count. | ||||||||
10333 | IC = CM.selectInterleaveCount(VF.Width, VF.Cost); | ||||||||
10334 | |||||||||
10335 | unsigned SelectedIC = std::max(IC, UserIC); | ||||||||
10336 | // Optimistically generate runtime checks if they are needed. Drop them if | ||||||||
10337 | // they turn out to not be profitable. | ||||||||
10338 | if (VF.Width.isVector() || SelectedIC > 1) | ||||||||
10339 | Checks.Create(L, *LVL.getLAI(), PSE.getPredicate(), VF.Width, SelectedIC); | ||||||||
10340 | |||||||||
10341 | // Check if it is profitable to vectorize with runtime checks. | ||||||||
10342 | bool ForceVectorization = | ||||||||
10343 | Hints.getForce() == LoopVectorizeHints::FK_Enabled; | ||||||||
10344 | if (!ForceVectorization && | ||||||||
10345 | !areRuntimeChecksProfitable(Checks, VF, CM.getVScaleForTuning(), L, | ||||||||
10346 | *PSE.getSE())) { | ||||||||
10347 | ORE->emit([&]() { | ||||||||
10348 | return OptimizationRemarkAnalysisAliasing( | ||||||||
10349 | DEBUG_TYPE"loop-vectorize", "CantReorderMemOps", L->getStartLoc(), | ||||||||
10350 | L->getHeader()) | ||||||||
10351 | << "loop not vectorized: cannot prove it is safe to reorder " | ||||||||
10352 | "memory operations"; | ||||||||
10353 | }); | ||||||||
10354 | 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); | ||||||||
10355 | Hints.emitRemarkWithHints(); | ||||||||
10356 | return false; | ||||||||
10357 | } | ||||||||
10358 | } | ||||||||
10359 | |||||||||
10360 | // Identify the diagnostic messages that should be produced. | ||||||||
10361 | std::pair<StringRef, std::string> VecDiagMsg, IntDiagMsg; | ||||||||
10362 | bool VectorizeLoop = true, InterleaveLoop = true; | ||||||||
10363 | if (VF.Width.isScalar()) { | ||||||||
10364 | 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); | ||||||||
10365 | VecDiagMsg = std::make_pair( | ||||||||
10366 | "VectorizationNotBeneficial", | ||||||||
10367 | "the cost-model indicates that vectorization is not beneficial"); | ||||||||
10368 | VectorizeLoop = false; | ||||||||
10369 | } | ||||||||
10370 | |||||||||
10371 | if (!MaybeVF && UserIC > 1) { | ||||||||
10372 | // Tell the user interleaving was avoided up-front, despite being explicitly | ||||||||
10373 | // requested. | ||||||||
10374 | 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 ) | ||||||||
10375 | "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 ); | ||||||||
10376 | IntDiagMsg = std::make_pair( | ||||||||
10377 | "InterleavingAvoided", | ||||||||
10378 | "Ignoring UserIC, because interleaving was avoided up front"); | ||||||||
10379 | InterleaveLoop = false; | ||||||||
10380 | } else if (IC == 1 && UserIC <= 1) { | ||||||||
10381 | // Tell the user interleaving is not beneficial. | ||||||||
10382 | 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); | ||||||||
10383 | IntDiagMsg = std::make_pair( | ||||||||
10384 | "InterleavingNotBeneficial", | ||||||||
10385 | "the cost-model indicates that interleaving is not beneficial"); | ||||||||
10386 | InterleaveLoop = false; | ||||||||
10387 | if (UserIC == 1) { | ||||||||
10388 | IntDiagMsg.first = "InterleavingNotBeneficialAndDisabled"; | ||||||||
10389 | IntDiagMsg.second += | ||||||||
10390 | " and is explicitly disabled or interleave count is set to 1"; | ||||||||
10391 | } | ||||||||
10392 | } else if (IC > 1 && UserIC == 1) { | ||||||||
10393 | // Tell the user interleaving is beneficial, but it explicitly disabled. | ||||||||
10394 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Interleaving is beneficial but is explicitly disabled." ; } } while (false) | ||||||||
10395 | 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); | ||||||||
10396 | IntDiagMsg = std::make_pair( | ||||||||
10397 | "InterleavingBeneficialButDisabled", | ||||||||
10398 | "the cost-model indicates that interleaving is beneficial " | ||||||||
10399 | "but is explicitly disabled or interleave count is set to 1"); | ||||||||
10400 | InterleaveLoop = false; | ||||||||
10401 | } | ||||||||
10402 | |||||||||
10403 | // Override IC if user provided an interleave count. | ||||||||
10404 | IC = UserIC > 0 ? UserIC : IC; | ||||||||
10405 | |||||||||
10406 | // Emit diagnostic messages, if any. | ||||||||
10407 | const char *VAPassName = Hints.vectorizeAnalysisPassName(); | ||||||||
10408 | if (!VectorizeLoop && !InterleaveLoop) { | ||||||||
10409 | // Do not vectorize or interleaving the loop. | ||||||||
10410 | ORE->emit([&]() { | ||||||||
10411 | return OptimizationRemarkMissed(VAPassName, VecDiagMsg.first, | ||||||||
10412 | L->getStartLoc(), L->getHeader()) | ||||||||
10413 | << VecDiagMsg.second; | ||||||||
10414 | }); | ||||||||
10415 | ORE->emit([&]() { | ||||||||
10416 | return OptimizationRemarkMissed(LV_NAME"loop-vectorize", IntDiagMsg.first, | ||||||||
10417 | L->getStartLoc(), L->getHeader()) | ||||||||
10418 | << IntDiagMsg.second; | ||||||||
10419 | }); | ||||||||
10420 | return false; | ||||||||
10421 | } else if (!VectorizeLoop && InterleaveLoop) { | ||||||||
10422 | 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); | ||||||||
10423 | ORE->emit([&]() { | ||||||||
10424 | return OptimizationRemarkAnalysis(VAPassName, VecDiagMsg.first, | ||||||||
10425 | L->getStartLoc(), L->getHeader()) | ||||||||
10426 | << VecDiagMsg.second; | ||||||||
10427 | }); | ||||||||
10428 | } else if (VectorizeLoop && !InterleaveLoop) { | ||||||||
10429 | 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) | ||||||||
10430 | << ") in " << DebugLocStr << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in " << DebugLocStr << '\n'; } } while (false); | ||||||||
10431 | ORE->emit([&]() { | ||||||||
10432 | return OptimizationRemarkAnalysis(LV_NAME"loop-vectorize", IntDiagMsg.first, | ||||||||
10433 | L->getStartLoc(), L->getHeader()) | ||||||||
10434 | << IntDiagMsg.second; | ||||||||
10435 | }); | ||||||||
10436 | } else if (VectorizeLoop && InterleaveLoop) { | ||||||||
10437 | 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) | ||||||||
10438 | << ") in " << DebugLocStr << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in " << DebugLocStr << '\n'; } } while (false); | ||||||||
10439 | 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); | ||||||||
10440 | } | ||||||||
10441 | |||||||||
10442 | bool DisableRuntimeUnroll = false; | ||||||||
10443 | MDNode *OrigLoopID = L->getLoopID(); | ||||||||
10444 | { | ||||||||
10445 | using namespace ore; | ||||||||
10446 | if (!VectorizeLoop) { | ||||||||
10447 | 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", 10447, __extension__ __PRETTY_FUNCTION__)); | ||||||||
10448 | // If we decided that it is not legal to vectorize the loop, then | ||||||||
10449 | // interleave it. | ||||||||
10450 | InnerLoopUnroller Unroller(L, PSE, LI, DT, TLI, TTI, AC, ORE, IC, &LVL, | ||||||||
10451 | &CM, BFI, PSI, Checks); | ||||||||
10452 | |||||||||
10453 | VPlan &BestPlan = LVP.getBestPlanFor(VF.Width); | ||||||||
10454 | LVP.executePlan(VF.Width, IC, BestPlan, Unroller, DT, false); | ||||||||
10455 | |||||||||
10456 | ORE->emit([&]() { | ||||||||
10457 | return OptimizationRemark(LV_NAME"loop-vectorize", "Interleaved", L->getStartLoc(), | ||||||||
10458 | L->getHeader()) | ||||||||
10459 | << "interleaved loop (interleaved count: " | ||||||||
10460 | << NV("InterleaveCount", IC) << ")"; | ||||||||
10461 | }); | ||||||||
10462 | } else { | ||||||||
10463 | // If we decided that it is *legal* to vectorize the loop, then do it. | ||||||||
10464 | |||||||||
10465 | // Consider vectorizing the epilogue too if it's profitable. | ||||||||
10466 | VectorizationFactor EpilogueVF = | ||||||||
10467 | CM.selectEpilogueVectorizationFactor(VF.Width, LVP); | ||||||||
10468 | if (EpilogueVF.Width.isVector()) { | ||||||||
10469 | |||||||||
10470 | // The first pass vectorizes the main loop and creates a scalar epilogue | ||||||||
10471 | // to be vectorized by executing the plan (potentially with a different | ||||||||
10472 | // factor) again shortly afterwards. | ||||||||
10473 | EpilogueLoopVectorizationInfo EPI(VF.Width, IC, EpilogueVF.Width, 1); | ||||||||
10474 | EpilogueVectorizerMainLoop MainILV(L, PSE, LI, DT, TLI, TTI, AC, ORE, | ||||||||
10475 | EPI, &LVL, &CM, BFI, PSI, Checks); | ||||||||
10476 | |||||||||
10477 | VPlan &BestMainPlan = LVP.getBestPlanFor(EPI.MainLoopVF); | ||||||||
10478 | LVP.executePlan(EPI.MainLoopVF, EPI.MainLoopUF, BestMainPlan, MainILV, | ||||||||
10479 | DT, true); | ||||||||
10480 | ++LoopsVectorized; | ||||||||
10481 | |||||||||
10482 | // Second pass vectorizes the epilogue and adjusts the control flow | ||||||||
10483 | // edges from the first pass. | ||||||||
10484 | EPI.MainLoopVF = EPI.EpilogueVF; | ||||||||
10485 | EPI.MainLoopUF = EPI.EpilogueUF; | ||||||||
10486 | EpilogueVectorizerEpilogueLoop EpilogILV(L, PSE, LI, DT, TLI, TTI, AC, | ||||||||
10487 | ORE, EPI, &LVL, &CM, BFI, PSI, | ||||||||
10488 | Checks); | ||||||||
10489 | |||||||||
10490 | VPlan &BestEpiPlan = LVP.getBestPlanFor(EPI.EpilogueVF); | ||||||||
10491 | VPRegionBlock *VectorLoop = BestEpiPlan.getVectorLoopRegion(); | ||||||||
10492 | VPBasicBlock *Header = VectorLoop->getEntryBasicBlock(); | ||||||||
10493 | Header->setName("vec.epilog.vector.body"); | ||||||||
10494 | |||||||||
10495 | // Ensure that the start values for any VPWidenIntOrFpInductionRecipe, | ||||||||
10496 | // VPWidenPointerInductionRecipe and VPReductionPHIRecipes are updated | ||||||||
10497 | // before vectorizing the epilogue loop. | ||||||||
10498 | for (VPRecipeBase &R : Header->phis()) { | ||||||||
10499 | if (isa<VPCanonicalIVPHIRecipe>(&R)) | ||||||||
10500 | continue; | ||||||||
10501 | |||||||||
10502 | Value *ResumeV = nullptr; | ||||||||
10503 | // TODO: Move setting of resume values to prepareToExecute. | ||||||||
10504 | if (auto *ReductionPhi = dyn_cast<VPReductionPHIRecipe>(&R)) { | ||||||||
10505 | ResumeV = MainILV.getReductionResumeValue( | ||||||||
10506 | ReductionPhi->getRecurrenceDescriptor()); | ||||||||
10507 | } else { | ||||||||
10508 | // Create induction resume values for both widened pointer and | ||||||||
10509 | // integer/fp inductions and update the start value of the induction | ||||||||
10510 | // recipes to use the resume value. | ||||||||
10511 | PHINode *IndPhi = nullptr; | ||||||||
10512 | const InductionDescriptor *ID; | ||||||||
10513 | if (auto *Ind = dyn_cast<VPWidenPointerInductionRecipe>(&R)) { | ||||||||
10514 | IndPhi = cast<PHINode>(Ind->getUnderlyingValue()); | ||||||||
10515 | ID = &Ind->getInductionDescriptor(); | ||||||||
10516 | } else { | ||||||||
10517 | auto *WidenInd = cast<VPWidenIntOrFpInductionRecipe>(&R); | ||||||||
10518 | IndPhi = WidenInd->getPHINode(); | ||||||||
10519 | ID = &WidenInd->getInductionDescriptor(); | ||||||||
10520 | } | ||||||||
10521 | |||||||||
10522 | ResumeV = MainILV.createInductionResumeValue( | ||||||||
10523 | IndPhi, *ID, {EPI.MainLoopIterationCountCheck}); | ||||||||
10524 | } | ||||||||
10525 | assert(ResumeV && "Must have a resume value")(static_cast <bool> (ResumeV && "Must have a resume value" ) ? void (0) : __assert_fail ("ResumeV && \"Must have a resume value\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 10525, __extension__ __PRETTY_FUNCTION__)); | ||||||||
10526 | VPValue *StartVal = BestEpiPlan.getOrAddExternalDef(ResumeV); | ||||||||
10527 | cast<VPHeaderPHIRecipe>(&R)->setStartValue(StartVal); | ||||||||
10528 | } | ||||||||
10529 | |||||||||
10530 | LVP.executePlan(EPI.EpilogueVF, EPI.EpilogueUF, BestEpiPlan, EpilogILV, | ||||||||
10531 | DT, true); | ||||||||
10532 | ++LoopsEpilogueVectorized; | ||||||||
10533 | |||||||||
10534 | if (!MainILV.areSafetyChecksAdded()) | ||||||||
10535 | DisableRuntimeUnroll = true; | ||||||||
10536 | } else { | ||||||||
10537 | InnerLoopVectorizer LB(L, PSE, LI, DT, TLI, TTI, AC, ORE, VF.Width, | ||||||||
10538 | VF.MinProfitableTripCount, IC, &LVL, &CM, BFI, | ||||||||
10539 | PSI, Checks); | ||||||||
10540 | |||||||||
10541 | VPlan &BestPlan = LVP.getBestPlanFor(VF.Width); | ||||||||
10542 | LVP.executePlan(VF.Width, IC, BestPlan, LB, DT, false); | ||||||||
10543 | ++LoopsVectorized; | ||||||||
10544 | |||||||||
10545 | // Add metadata to disable runtime unrolling a scalar loop when there | ||||||||
10546 | // are no runtime checks about strides and memory. A scalar loop that is | ||||||||
10547 | // rarely used is not worth unrolling. | ||||||||
10548 | if (!LB.areSafetyChecksAdded()) | ||||||||
10549 | DisableRuntimeUnroll = true; | ||||||||
10550 | } | ||||||||
10551 | // Report the vectorization decision. | ||||||||
10552 | ORE->emit([&]() { | ||||||||
10553 | return OptimizationRemark(LV_NAME"loop-vectorize", "Vectorized", L->getStartLoc(), | ||||||||
10554 | L->getHeader()) | ||||||||
10555 | << "vectorized loop (vectorization width: " | ||||||||
10556 | << NV("VectorizationFactor", VF.Width) | ||||||||
10557 | << ", interleaved count: " << NV("InterleaveCount", IC) << ")"; | ||||||||
10558 | }); | ||||||||
10559 | } | ||||||||
10560 | |||||||||
10561 | if (ORE->allowExtraAnalysis(LV_NAME"loop-vectorize")) | ||||||||
10562 | checkMixedPrecision(L, ORE); | ||||||||
10563 | } | ||||||||
10564 | |||||||||
10565 | std::optional<MDNode *> RemainderLoopID = | ||||||||
10566 | makeFollowupLoopID(OrigLoopID, {LLVMLoopVectorizeFollowupAll, | ||||||||
10567 | LLVMLoopVectorizeFollowupEpilogue}); | ||||||||
10568 | if (RemainderLoopID) { | ||||||||
10569 | L->setLoopID(*RemainderLoopID); | ||||||||
10570 | } else { | ||||||||
10571 | if (DisableRuntimeUnroll) | ||||||||
10572 | AddRuntimeUnrollDisableMetaData(L); | ||||||||
10573 | |||||||||
10574 | // Mark the loop as already vectorized to avoid vectorizing again. | ||||||||
10575 | Hints.setAlreadyVectorized(); | ||||||||
10576 | } | ||||||||
10577 | |||||||||
10578 | 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", 10578, __extension__ __PRETTY_FUNCTION__)); | ||||||||
10579 | return true; | ||||||||
10580 | } | ||||||||
10581 | |||||||||
10582 | LoopVectorizeResult LoopVectorizePass::runImpl( | ||||||||
10583 | Function &F, ScalarEvolution &SE_, LoopInfo &LI_, TargetTransformInfo &TTI_, | ||||||||
10584 | DominatorTree &DT_, BlockFrequencyInfo &BFI_, TargetLibraryInfo *TLI_, | ||||||||
10585 | DemandedBits &DB_, AssumptionCache &AC_, LoopAccessInfoManager &LAIs_, | ||||||||
10586 | OptimizationRemarkEmitter &ORE_, ProfileSummaryInfo *PSI_) { | ||||||||
10587 | SE = &SE_; | ||||||||
10588 | LI = &LI_; | ||||||||
10589 | TTI = &TTI_; | ||||||||
10590 | DT = &DT_; | ||||||||
10591 | BFI = &BFI_; | ||||||||
10592 | TLI = TLI_; | ||||||||
10593 | AC = &AC_; | ||||||||
10594 | LAIs = &LAIs_; | ||||||||
10595 | DB = &DB_; | ||||||||
10596 | ORE = &ORE_; | ||||||||
10597 | PSI = PSI_; | ||||||||
10598 | |||||||||
10599 | // Don't attempt if | ||||||||
10600 | // 1. the target claims to have no vector registers, and | ||||||||
10601 | // 2. interleaving won't help ILP. | ||||||||
10602 | // | ||||||||
10603 | // The second condition is necessary because, even if the target has no | ||||||||
10604 | // vector registers, loop vectorization may still enable scalar | ||||||||
10605 | // interleaving. | ||||||||
10606 | if (!TTI->getNumberOfRegisters(TTI->getRegisterClassForType(true)) && | ||||||||
10607 | TTI->getMaxInterleaveFactor(1) < 2) | ||||||||
10608 | return LoopVectorizeResult(false, false); | ||||||||
10609 | |||||||||
10610 | bool Changed = false, CFGChanged = false; | ||||||||
10611 | |||||||||
10612 | // The vectorizer requires loops to be in simplified form. | ||||||||
10613 | // Since simplification may add new inner loops, it has to run before the | ||||||||
10614 | // legality and profitability checks. This means running the loop vectorizer | ||||||||
10615 | // will simplify all loops, regardless of whether anything end up being | ||||||||
10616 | // vectorized. | ||||||||
10617 | for (const auto &L : *LI) | ||||||||
10618 | Changed |= CFGChanged |= | ||||||||
10619 | simplifyLoop(L, DT, LI, SE, AC, nullptr, false /* PreserveLCSSA */); | ||||||||
10620 | |||||||||
10621 | // Build up a worklist of inner-loops to vectorize. This is necessary as | ||||||||
10622 | // the act of vectorizing or partially unrolling a loop creates new loops | ||||||||
10623 | // and can invalidate iterators across the loops. | ||||||||
10624 | SmallVector<Loop *, 8> Worklist; | ||||||||
10625 | |||||||||
10626 | for (Loop *L : *LI) | ||||||||
10627 | collectSupportedLoops(*L, LI, ORE, Worklist); | ||||||||
10628 | |||||||||
10629 | LoopsAnalyzed += Worklist.size(); | ||||||||
10630 | |||||||||
10631 | // Now walk the identified inner loops. | ||||||||
10632 | while (!Worklist.empty()) { | ||||||||
10633 | Loop *L = Worklist.pop_back_val(); | ||||||||
10634 | |||||||||
10635 | // For the inner loops we actually process, form LCSSA to simplify the | ||||||||
10636 | // transform. | ||||||||
10637 | Changed |= formLCSSARecursively(*L, *DT, LI, SE); | ||||||||
10638 | |||||||||
10639 | Changed |= CFGChanged |= processLoop(L); | ||||||||
10640 | |||||||||
10641 | if (Changed) | ||||||||
10642 | LAIs->clear(); | ||||||||
10643 | } | ||||||||
10644 | |||||||||
10645 | // Process each loop nest in the function. | ||||||||
10646 | return LoopVectorizeResult(Changed, CFGChanged); | ||||||||
10647 | } | ||||||||
10648 | |||||||||
10649 | PreservedAnalyses LoopVectorizePass::run(Function &F, | ||||||||
10650 | FunctionAnalysisManager &AM) { | ||||||||
10651 | auto &LI = AM.getResult<LoopAnalysis>(F); | ||||||||
10652 | // There are no loops in the function. Return before computing other expensive | ||||||||
10653 | // analyses. | ||||||||
10654 | if (LI.empty()) | ||||||||
10655 | return PreservedAnalyses::all(); | ||||||||
10656 | auto &SE = AM.getResult<ScalarEvolutionAnalysis>(F); | ||||||||
10657 | auto &TTI = AM.getResult<TargetIRAnalysis>(F); | ||||||||
10658 | auto &DT = AM.getResult<DominatorTreeAnalysis>(F); | ||||||||
10659 | auto &BFI = AM.getResult<BlockFrequencyAnalysis>(F); | ||||||||
10660 | auto &TLI = AM.getResult<TargetLibraryAnalysis>(F); | ||||||||
10661 | auto &AC = AM.getResult<AssumptionAnalysis>(F); | ||||||||
10662 | auto &DB = AM.getResult<DemandedBitsAnalysis>(F); | ||||||||
10663 | auto &ORE = AM.getResult<OptimizationRemarkEmitterAnalysis>(F); | ||||||||
10664 | |||||||||
10665 | LoopAccessInfoManager &LAIs = AM.getResult<LoopAccessAnalysis>(F); | ||||||||
10666 | auto &MAMProxy = AM.getResult<ModuleAnalysisManagerFunctionProxy>(F); | ||||||||
10667 | ProfileSummaryInfo *PSI = | ||||||||
10668 | MAMProxy.getCachedResult<ProfileSummaryAnalysis>(*F.getParent()); | ||||||||
10669 | LoopVectorizeResult Result = | ||||||||
10670 | runImpl(F, SE, LI, TTI, DT, BFI, &TLI, DB, AC, LAIs, ORE, PSI); | ||||||||
10671 | if (!Result.MadeAnyChange) | ||||||||
10672 | return PreservedAnalyses::all(); | ||||||||
10673 | PreservedAnalyses PA; | ||||||||
10674 | |||||||||
10675 | // We currently do not preserve loopinfo/dominator analyses with outer loop | ||||||||
10676 | // vectorization. Until this is addressed, mark these analyses as preserved | ||||||||
10677 | // only for non-VPlan-native path. | ||||||||
10678 | // TODO: Preserve Loop and Dominator analyses for VPlan-native path. | ||||||||
10679 | if (!EnableVPlanNativePath) { | ||||||||
10680 | PA.preserve<LoopAnalysis>(); | ||||||||
10681 | PA.preserve<DominatorTreeAnalysis>(); | ||||||||
10682 | } | ||||||||
10683 | |||||||||
10684 | if (Result.MadeCFGChange) { | ||||||||
10685 | // Making CFG changes likely means a loop got vectorized. Indicate that | ||||||||
10686 | // extra simplification passes should be run. | ||||||||
10687 | // TODO: MadeCFGChanges is not a prefect proxy. Extra passes should only | ||||||||
10688 | // be run if runtime checks have been added. | ||||||||
10689 | AM.getResult<ShouldRunExtraVectorPasses>(F); | ||||||||
10690 | PA.preserve<ShouldRunExtraVectorPasses>(); | ||||||||
10691 | } else { | ||||||||
10692 | PA.preserveSet<CFGAnalyses>(); | ||||||||
10693 | } | ||||||||
10694 | return PA; | ||||||||
10695 | } | ||||||||
10696 | |||||||||
10697 | void LoopVectorizePass::printPipeline( | ||||||||
10698 | raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) { | ||||||||
10699 | static_cast<PassInfoMixin<LoopVectorizePass> *>(this)->printPipeline( | ||||||||
10700 | OS, MapClassName2PassName); | ||||||||
10701 | |||||||||
10702 | OS << "<"; | ||||||||
10703 | OS << (InterleaveOnlyWhenForced ? "" : "no-") << "interleave-forced-only;"; | ||||||||
10704 | OS << (VectorizeOnlyWhenForced ? "" : "no-") << "vectorize-forced-only;"; | ||||||||
10705 | OS << ">"; | ||||||||
10706 | } |
1 | //===- LoopVectorizationPlanner.h - Planner for LoopVectorization ---------===// |
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 | /// \file |
10 | /// This file provides a LoopVectorizationPlanner class. |
11 | /// InnerLoopVectorizer vectorizes loops which contain only one basic |
12 | /// LoopVectorizationPlanner - drives the vectorization process after having |
13 | /// passed Legality checks. |
14 | /// The planner builds and optimizes the Vectorization Plans which record the |
15 | /// decisions how to vectorize the given loop. In particular, represent the |
16 | /// control-flow of the vectorized version, the replication of instructions that |
17 | /// are to be scalarized, and interleave access groups. |
18 | /// |
19 | /// Also provides a VPlan-based builder utility analogous to IRBuilder. |
20 | /// It provides an instruction-level API for generating VPInstructions while |
21 | /// abstracting away the Recipe manipulation details. |
22 | //===----------------------------------------------------------------------===// |
23 | |
24 | #ifndef LLVM_TRANSFORMS_VECTORIZE_LOOPVECTORIZATIONPLANNER_H |
25 | #define LLVM_TRANSFORMS_VECTORIZE_LOOPVECTORIZATIONPLANNER_H |
26 | |
27 | #include "VPlan.h" |
28 | #include "llvm/Support/InstructionCost.h" |
29 | |
30 | namespace llvm { |
31 | |
32 | class LoopInfo; |
33 | class LoopVectorizationLegality; |
34 | class LoopVectorizationCostModel; |
35 | class PredicatedScalarEvolution; |
36 | class LoopVectorizeHints; |
37 | class OptimizationRemarkEmitter; |
38 | class TargetTransformInfo; |
39 | class TargetLibraryInfo; |
40 | class VPRecipeBuilder; |
41 | |
42 | /// VPlan-based builder utility analogous to IRBuilder. |
43 | class VPBuilder { |
44 | VPBasicBlock *BB = nullptr; |
45 | VPBasicBlock::iterator InsertPt = VPBasicBlock::iterator(); |
46 | |
47 | VPInstruction *createInstruction(unsigned Opcode, |
48 | ArrayRef<VPValue *> Operands, DebugLoc DL, |
49 | const Twine &Name = "") { |
50 | VPInstruction *Instr = new VPInstruction(Opcode, Operands, DL, Name); |
51 | if (BB) |
52 | BB->insert(Instr, InsertPt); |
53 | return Instr; |
54 | } |
55 | |
56 | VPInstruction *createInstruction(unsigned Opcode, |
57 | std::initializer_list<VPValue *> Operands, |
58 | DebugLoc DL, const Twine &Name = "") { |
59 | return createInstruction(Opcode, ArrayRef<VPValue *>(Operands), DL, Name); |
60 | } |
61 | |
62 | public: |
63 | VPBuilder() = default; |
64 | |
65 | /// Clear the insertion point: created instructions will not be inserted into |
66 | /// a block. |
67 | void clearInsertionPoint() { |
68 | BB = nullptr; |
69 | InsertPt = VPBasicBlock::iterator(); |
70 | } |
71 | |
72 | VPBasicBlock *getInsertBlock() const { return BB; } |
73 | VPBasicBlock::iterator getInsertPoint() const { return InsertPt; } |
74 | |
75 | /// InsertPoint - A saved insertion point. |
76 | class VPInsertPoint { |
77 | VPBasicBlock *Block = nullptr; |
78 | VPBasicBlock::iterator Point; |
79 | |
80 | public: |
81 | /// Creates a new insertion point which doesn't point to anything. |
82 | VPInsertPoint() = default; |
83 | |
84 | /// Creates a new insertion point at the given location. |
85 | VPInsertPoint(VPBasicBlock *InsertBlock, VPBasicBlock::iterator InsertPoint) |
86 | : Block(InsertBlock), Point(InsertPoint) {} |
87 | |
88 | /// Returns true if this insert point is set. |
89 | bool isSet() const { return Block != nullptr; } |
90 | |
91 | VPBasicBlock *getBlock() const { return Block; } |
92 | VPBasicBlock::iterator getPoint() const { return Point; } |
93 | }; |
94 | |
95 | /// Sets the current insert point to a previously-saved location. |
96 | void restoreIP(VPInsertPoint IP) { |
97 | if (IP.isSet()) |
98 | setInsertPoint(IP.getBlock(), IP.getPoint()); |
99 | else |
100 | clearInsertionPoint(); |
101 | } |
102 | |
103 | /// This specifies that created VPInstructions should be appended to the end |
104 | /// of the specified block. |
105 | void setInsertPoint(VPBasicBlock *TheBB) { |
106 | assert(TheBB && "Attempting to set a null insert point")(static_cast <bool> (TheBB && "Attempting to set a null insert point" ) ? void (0) : __assert_fail ("TheBB && \"Attempting to set a null insert point\"" , "llvm/lib/Transforms/Vectorize/LoopVectorizationPlanner.h", 106, __extension__ __PRETTY_FUNCTION__)); |
107 | BB = TheBB; |
108 | InsertPt = BB->end(); |
109 | } |
110 | |
111 | /// This specifies that created instructions should be inserted at the |
112 | /// specified point. |
113 | void setInsertPoint(VPBasicBlock *TheBB, VPBasicBlock::iterator IP) { |
114 | BB = TheBB; |
115 | InsertPt = IP; |
116 | } |
117 | |
118 | /// Insert and return the specified instruction. |
119 | VPInstruction *insert(VPInstruction *I) const { |
120 | BB->insert(I, InsertPt); |
121 | return I; |
122 | } |
123 | |
124 | /// Create an N-ary operation with \p Opcode, \p Operands and set \p Inst as |
125 | /// its underlying Instruction. |
126 | VPValue *createNaryOp(unsigned Opcode, ArrayRef<VPValue *> Operands, |
127 | Instruction *Inst = nullptr, const Twine &Name = "") { |
128 | DebugLoc DL; |
129 | if (Inst) |
130 | DL = Inst->getDebugLoc(); |
131 | VPInstruction *NewVPInst = createInstruction(Opcode, Operands, DL, Name); |
132 | NewVPInst->setUnderlyingValue(Inst); |
133 | return NewVPInst; |
134 | } |
135 | VPValue *createNaryOp(unsigned Opcode, ArrayRef<VPValue *> Operands, |
136 | DebugLoc DL, const Twine &Name = "") { |
137 | return createInstruction(Opcode, Operands, DL, Name); |
138 | } |
139 | |
140 | VPValue *createNot(VPValue *Operand, DebugLoc DL, const Twine &Name = "") { |
141 | return createInstruction(VPInstruction::Not, {Operand}, DL, Name); |
142 | } |
143 | |
144 | VPValue *createAnd(VPValue *LHS, VPValue *RHS, DebugLoc DL, |
145 | const Twine &Name = "") { |
146 | return createInstruction(Instruction::BinaryOps::And, {LHS, RHS}, DL, Name); |
147 | } |
148 | |
149 | VPValue *createOr(VPValue *LHS, VPValue *RHS, DebugLoc DL, |
150 | const Twine &Name = "") { |
151 | return createInstruction(Instruction::BinaryOps::Or, {LHS, RHS}, DL, Name); |
152 | } |
153 | |
154 | VPValue *createSelect(VPValue *Cond, VPValue *TrueVal, VPValue *FalseVal, |
155 | DebugLoc DL, const Twine &Name = "") { |
156 | return createNaryOp(Instruction::Select, {Cond, TrueVal, FalseVal}, DL, |
157 | Name); |
158 | } |
159 | |
160 | //===--------------------------------------------------------------------===// |
161 | // RAII helpers. |
162 | //===--------------------------------------------------------------------===// |
163 | |
164 | /// RAII object that stores the current insertion point and restores it when |
165 | /// the object is destroyed. |
166 | class InsertPointGuard { |
167 | VPBuilder &Builder; |
168 | VPBasicBlock *Block; |
169 | VPBasicBlock::iterator Point; |
170 | |
171 | public: |
172 | InsertPointGuard(VPBuilder &B) |
173 | : Builder(B), Block(B.getInsertBlock()), Point(B.getInsertPoint()) {} |
174 | |
175 | InsertPointGuard(const InsertPointGuard &) = delete; |
176 | InsertPointGuard &operator=(const InsertPointGuard &) = delete; |
177 | |
178 | ~InsertPointGuard() { Builder.restoreIP(VPInsertPoint(Block, Point)); } |
179 | }; |
180 | }; |
181 | |
182 | /// TODO: The following VectorizationFactor was pulled out of |
183 | /// LoopVectorizationCostModel class. LV also deals with |
184 | /// VectorizerParams::VectorizationFactor and VectorizationCostTy. |
185 | /// We need to streamline them. |
186 | |
187 | /// Information about vectorization costs. |
188 | struct VectorizationFactor { |
189 | /// Vector width with best cost. |
190 | ElementCount Width; |
191 | |
192 | /// Cost of the loop with that width. |
193 | InstructionCost Cost; |
194 | |
195 | /// Cost of the scalar loop. |
196 | InstructionCost ScalarCost; |
197 | |
198 | /// The minimum trip count required to make vectorization profitable, e.g. due |
199 | /// to runtime checks. |
200 | ElementCount MinProfitableTripCount; |
201 | |
202 | VectorizationFactor(ElementCount Width, InstructionCost Cost, |
203 | InstructionCost ScalarCost) |
204 | : Width(Width), Cost(Cost), ScalarCost(ScalarCost) {} |
205 | |
206 | /// Width 1 means no vectorization, cost 0 means uncomputed cost. |
207 | static VectorizationFactor Disabled() { |
208 | return {ElementCount::getFixed(1), 0, 0}; |
209 | } |
210 | |
211 | bool operator==(const VectorizationFactor &rhs) const { |
212 | return Width == rhs.Width && Cost == rhs.Cost; |
213 | } |
214 | |
215 | bool operator!=(const VectorizationFactor &rhs) const { |
216 | return !(*this == rhs); |
217 | } |
218 | }; |
219 | |
220 | /// A class that represents two vectorization factors (initialized with 0 by |
221 | /// default). One for fixed-width vectorization and one for scalable |
222 | /// vectorization. This can be used by the vectorizer to choose from a range of |
223 | /// fixed and/or scalable VFs in order to find the most cost-effective VF to |
224 | /// vectorize with. |
225 | struct FixedScalableVFPair { |
226 | ElementCount FixedVF; |
227 | ElementCount ScalableVF; |
228 | |
229 | FixedScalableVFPair() |
230 | : FixedVF(ElementCount::getFixed(0)), |
231 | ScalableVF(ElementCount::getScalable(0)) {} |
232 | FixedScalableVFPair(const ElementCount &Max) : FixedScalableVFPair() { |
233 | *(Max.isScalable() ? &ScalableVF : &FixedVF) = Max; |
234 | } |
235 | FixedScalableVFPair(const ElementCount &FixedVF, |
236 | const ElementCount &ScalableVF) |
237 | : FixedVF(FixedVF), ScalableVF(ScalableVF) { |
238 | assert(!FixedVF.isScalable() && ScalableVF.isScalable() &&(static_cast <bool> (!FixedVF.isScalable() && ScalableVF .isScalable() && "Invalid scalable properties") ? void (0) : __assert_fail ("!FixedVF.isScalable() && ScalableVF.isScalable() && \"Invalid scalable properties\"" , "llvm/lib/Transforms/Vectorize/LoopVectorizationPlanner.h", 239, __extension__ __PRETTY_FUNCTION__)) |
239 | "Invalid scalable properties")(static_cast <bool> (!FixedVF.isScalable() && ScalableVF .isScalable() && "Invalid scalable properties") ? void (0) : __assert_fail ("!FixedVF.isScalable() && ScalableVF.isScalable() && \"Invalid scalable properties\"" , "llvm/lib/Transforms/Vectorize/LoopVectorizationPlanner.h", 239, __extension__ __PRETTY_FUNCTION__)); |
240 | } |
241 | |
242 | static FixedScalableVFPair getNone() { return FixedScalableVFPair(); } |
243 | |
244 | /// \return true if either fixed- or scalable VF is non-zero. |
245 | explicit operator bool() const { return FixedVF || ScalableVF; } |
246 | |
247 | /// \return true if either fixed- or scalable VF is a valid vector VF. |
248 | bool hasVector() const { return FixedVF.isVector() || ScalableVF.isVector(); } |
249 | }; |
250 | |
251 | /// Planner drives the vectorization process after having passed |
252 | /// Legality checks. |
253 | class LoopVectorizationPlanner { |
254 | /// The loop that we evaluate. |
255 | Loop *OrigLoop; |
256 | |
257 | /// Loop Info analysis. |
258 | LoopInfo *LI; |
259 | |
260 | /// Target Library Info. |
261 | const TargetLibraryInfo *TLI; |
262 | |
263 | /// Target Transform Info. |
264 | const TargetTransformInfo *TTI; |
265 | |
266 | /// The legality analysis. |
267 | LoopVectorizationLegality *Legal; |
268 | |
269 | /// The profitability analysis. |
270 | LoopVectorizationCostModel &CM; |
271 | |
272 | /// The interleaved access analysis. |
273 | InterleavedAccessInfo &IAI; |
274 | |
275 | PredicatedScalarEvolution &PSE; |
276 | |
277 | const LoopVectorizeHints &Hints; |
278 | |
279 | OptimizationRemarkEmitter *ORE; |
280 | |
281 | SmallVector<VPlanPtr, 4> VPlans; |
282 | |
283 | /// A builder used to construct the current plan. |
284 | VPBuilder Builder; |
285 | |
286 | public: |
287 | LoopVectorizationPlanner(Loop *L, LoopInfo *LI, const TargetLibraryInfo *TLI, |
288 | const TargetTransformInfo *TTI, |
289 | LoopVectorizationLegality *Legal, |
290 | LoopVectorizationCostModel &CM, |
291 | InterleavedAccessInfo &IAI, |
292 | PredicatedScalarEvolution &PSE, |
293 | const LoopVectorizeHints &Hints, |
294 | OptimizationRemarkEmitter *ORE) |
295 | : OrigLoop(L), LI(LI), TLI(TLI), TTI(TTI), Legal(Legal), CM(CM), IAI(IAI), |
296 | PSE(PSE), Hints(Hints), ORE(ORE) {} |
297 | |
298 | /// Plan how to best vectorize, return the best VF and its cost, or |
299 | /// std::nullopt if vectorization and interleaving should be avoided up front. |
300 | std::optional<VectorizationFactor> plan(ElementCount UserVF, unsigned UserIC); |
301 | |
302 | /// Use the VPlan-native path to plan how to best vectorize, return the best |
303 | /// VF and its cost. |
304 | VectorizationFactor planInVPlanNativePath(ElementCount UserVF); |
305 | |
306 | /// Return the best VPlan for \p VF. |
307 | VPlan &getBestPlanFor(ElementCount VF) const; |
308 | |
309 | /// Generate the IR code for the body of the vectorized loop according to the |
310 | /// best selected \p VF, \p UF and VPlan \p BestPlan. |
311 | /// TODO: \p IsEpilogueVectorization is needed to avoid issues due to epilogue |
312 | /// vectorization re-using plans for both the main and epilogue vector loops. |
313 | /// It should be removed once the re-use issue has been fixed. |
314 | void executePlan(ElementCount VF, unsigned UF, VPlan &BestPlan, |
315 | InnerLoopVectorizer &LB, DominatorTree *DT, |
316 | bool IsEpilogueVectorization); |
317 | |
318 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) |
319 | void printPlans(raw_ostream &O); |
320 | #endif |
321 | |
322 | /// Look through the existing plans and return true if we have one with all |
323 | /// the vectorization factors in question. |
324 | bool hasPlanWithVF(ElementCount VF) const { |
325 | return any_of(VPlans, |
326 | [&](const VPlanPtr &Plan) { return Plan->hasVF(VF); }); |
327 | } |
328 | |
329 | /// Test a \p Predicate on a \p Range of VF's. Return the value of applying |
330 | /// \p Predicate on Range.Start, possibly decreasing Range.End such that the |
331 | /// returned value holds for the entire \p Range. |
332 | static bool |
333 | getDecisionAndClampRange(const std::function<bool(ElementCount)> &Predicate, |
334 | VFRange &Range); |
335 | |
336 | /// Check if the number of runtime checks exceeds the threshold. |
337 | bool requiresTooManyRuntimeChecks() const; |
338 | |
339 | protected: |
340 | /// Build VPlans for power-of-2 VF's between \p MinVF and \p MaxVF inclusive, |
341 | /// according to the information gathered by Legal when it checked if it is |
342 | /// legal to vectorize the loop. |
343 | void buildVPlans(ElementCount MinVF, ElementCount MaxVF); |
344 | |
345 | private: |
346 | /// Build a VPlan according to the information gathered by Legal. \return a |
347 | /// VPlan for vectorization factors \p Range.Start and up to \p Range.End |
348 | /// exclusive, possibly decreasing \p Range.End. |
349 | VPlanPtr buildVPlan(VFRange &Range); |
350 | |
351 | /// Build a VPlan using VPRecipes according to the information gather by |
352 | /// Legal. This method is only used for the legacy inner loop vectorizer. |
353 | VPlanPtr buildVPlanWithVPRecipes( |
354 | VFRange &Range, SmallPtrSetImpl<Instruction *> &DeadInstructions, |
355 | const MapVector<Instruction *, Instruction *> &SinkAfter); |
356 | |
357 | /// Build VPlans for power-of-2 VF's between \p MinVF and \p MaxVF inclusive, |
358 | /// according to the information gathered by Legal when it checked if it is |
359 | /// legal to vectorize the loop. This method creates VPlans using VPRecipes. |
360 | void buildVPlansWithVPRecipes(ElementCount MinVF, ElementCount MaxVF); |
361 | |
362 | // Adjust the recipes for reductions. For in-loop reductions the chain of |
363 | // instructions leading from the loop exit instr to the phi need to be |
364 | // converted to reductions, with one operand being vector and the other being |
365 | // the scalar reduction chain. For other reductions, a select is introduced |
366 | // between the phi and live-out recipes when folding the tail. |
367 | void adjustRecipesForReductions(VPBasicBlock *LatchVPBB, VPlanPtr &Plan, |
368 | VPRecipeBuilder &RecipeBuilder, |
369 | ElementCount MinVF); |
370 | }; |
371 | |
372 | } // namespace llvm |
373 | |
374 | #endif // LLVM_TRANSFORMS_VECTORIZE_LOOPVECTORIZATIONPLANNER_H |