File: | llvm/lib/Transforms/Vectorize/LoopVectorize.cpp |
Warning: | line 7270, column 35 Potential leak of memory pointed to by 'BlockMask' |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | //===- LoopVectorize.cpp - A Loop Vectorizer ------------------------------===// | ||||||||
2 | // | ||||||||
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. | ||||||||
4 | // See https://llvm.org/LICENSE.txt for license information. | ||||||||
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception | ||||||||
6 | // | ||||||||
7 | //===----------------------------------------------------------------------===// | ||||||||
8 | // | ||||||||
9 | // This is the LLVM loop vectorizer. This pass modifies 'vectorizable' loops | ||||||||
10 | // and generates target-independent LLVM-IR. | ||||||||
11 | // The vectorizer uses the TargetTransformInfo analysis to estimate the costs | ||||||||
12 | // of instructions in order to estimate the profitability of vectorization. | ||||||||
13 | // | ||||||||
14 | // The loop vectorizer combines consecutive loop iterations into a single | ||||||||
15 | // 'wide' iteration. After this transformation the index is incremented | ||||||||
16 | // by the SIMD vector width, and not by one. | ||||||||
17 | // | ||||||||
18 | // This pass has three parts: | ||||||||
19 | // 1. The main loop pass that drives the different parts. | ||||||||
20 | // 2. LoopVectorizationLegality - A unit that checks for the legality | ||||||||
21 | // of the vectorization. | ||||||||
22 | // 3. InnerLoopVectorizer - A unit that performs the actual | ||||||||
23 | // widening of instructions. | ||||||||
24 | // 4. LoopVectorizationCostModel - A unit that checks for the profitability | ||||||||
25 | // of vectorization. It decides on the optimal vector width, which | ||||||||
26 | // can be one, if vectorization is not profitable. | ||||||||
27 | // | ||||||||
28 | // There is a development effort going on to migrate loop vectorizer to the | ||||||||
29 | // VPlan infrastructure and to introduce outer loop vectorization support (see | ||||||||
30 | // docs/Proposal/VectorizationPlan.rst and | ||||||||
31 | // http://lists.llvm.org/pipermail/llvm-dev/2017-December/119523.html). For this | ||||||||
32 | // purpose, we temporarily introduced the VPlan-native vectorization path: an | ||||||||
33 | // alternative vectorization path that is natively implemented on top of the | ||||||||
34 | // VPlan infrastructure. See EnableVPlanNativePath for enabling. | ||||||||
35 | // | ||||||||
36 | //===----------------------------------------------------------------------===// | ||||||||
37 | // | ||||||||
38 | // The reduction-variable vectorization is based on the paper: | ||||||||
39 | // D. Nuzman and R. Henderson. Multi-platform Auto-vectorization. | ||||||||
40 | // | ||||||||
41 | // Variable uniformity checks are inspired by: | ||||||||
42 | // Karrenberg, R. and Hack, S. Whole Function Vectorization. | ||||||||
43 | // | ||||||||
44 | // The interleaved access vectorization is based on the paper: | ||||||||
45 | // Dorit Nuzman, Ira Rosen and Ayal Zaks. Auto-Vectorization of Interleaved | ||||||||
46 | // Data for SIMD | ||||||||
47 | // | ||||||||
48 | // Other ideas/concepts are from: | ||||||||
49 | // A. Zaks and D. Nuzman. Autovectorization in GCC-two years later. | ||||||||
50 | // | ||||||||
51 | // S. Maleki, Y. Gao, M. Garzaran, T. Wong and D. Padua. An Evaluation of | ||||||||
52 | // Vectorizing Compilers. | ||||||||
53 | // | ||||||||
54 | //===----------------------------------------------------------------------===// | ||||||||
55 | |||||||||
56 | #include "llvm/Transforms/Vectorize/LoopVectorize.h" | ||||||||
57 | #include "LoopVectorizationPlanner.h" | ||||||||
58 | #include "VPRecipeBuilder.h" | ||||||||
59 | #include "VPlan.h" | ||||||||
60 | #include "VPlanHCFGBuilder.h" | ||||||||
61 | #include "VPlanPredicator.h" | ||||||||
62 | #include "VPlanTransforms.h" | ||||||||
63 | #include "llvm/ADT/APInt.h" | ||||||||
64 | #include "llvm/ADT/ArrayRef.h" | ||||||||
65 | #include "llvm/ADT/DenseMap.h" | ||||||||
66 | #include "llvm/ADT/DenseMapInfo.h" | ||||||||
67 | #include "llvm/ADT/Hashing.h" | ||||||||
68 | #include "llvm/ADT/MapVector.h" | ||||||||
69 | #include "llvm/ADT/None.h" | ||||||||
70 | #include "llvm/ADT/Optional.h" | ||||||||
71 | #include "llvm/ADT/STLExtras.h" | ||||||||
72 | #include "llvm/ADT/SetVector.h" | ||||||||
73 | #include "llvm/ADT/SmallPtrSet.h" | ||||||||
74 | #include "llvm/ADT/SmallVector.h" | ||||||||
75 | #include "llvm/ADT/Statistic.h" | ||||||||
76 | #include "llvm/ADT/StringRef.h" | ||||||||
77 | #include "llvm/ADT/Twine.h" | ||||||||
78 | #include "llvm/ADT/iterator_range.h" | ||||||||
79 | #include "llvm/Analysis/AssumptionCache.h" | ||||||||
80 | #include "llvm/Analysis/BasicAliasAnalysis.h" | ||||||||
81 | #include "llvm/Analysis/BlockFrequencyInfo.h" | ||||||||
82 | #include "llvm/Analysis/CFG.h" | ||||||||
83 | #include "llvm/Analysis/CodeMetrics.h" | ||||||||
84 | #include "llvm/Analysis/DemandedBits.h" | ||||||||
85 | #include "llvm/Analysis/GlobalsModRef.h" | ||||||||
86 | #include "llvm/Analysis/LoopAccessAnalysis.h" | ||||||||
87 | #include "llvm/Analysis/LoopAnalysisManager.h" | ||||||||
88 | #include "llvm/Analysis/LoopInfo.h" | ||||||||
89 | #include "llvm/Analysis/LoopIterator.h" | ||||||||
90 | #include "llvm/Analysis/MemorySSA.h" | ||||||||
91 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" | ||||||||
92 | #include "llvm/Analysis/ProfileSummaryInfo.h" | ||||||||
93 | #include "llvm/Analysis/ScalarEvolution.h" | ||||||||
94 | #include "llvm/Analysis/ScalarEvolutionExpressions.h" | ||||||||
95 | #include "llvm/Analysis/TargetLibraryInfo.h" | ||||||||
96 | #include "llvm/Analysis/TargetTransformInfo.h" | ||||||||
97 | #include "llvm/Analysis/VectorUtils.h" | ||||||||
98 | #include "llvm/IR/Attributes.h" | ||||||||
99 | #include "llvm/IR/BasicBlock.h" | ||||||||
100 | #include "llvm/IR/CFG.h" | ||||||||
101 | #include "llvm/IR/Constant.h" | ||||||||
102 | #include "llvm/IR/Constants.h" | ||||||||
103 | #include "llvm/IR/DataLayout.h" | ||||||||
104 | #include "llvm/IR/DebugInfoMetadata.h" | ||||||||
105 | #include "llvm/IR/DebugLoc.h" | ||||||||
106 | #include "llvm/IR/DerivedTypes.h" | ||||||||
107 | #include "llvm/IR/DiagnosticInfo.h" | ||||||||
108 | #include "llvm/IR/Dominators.h" | ||||||||
109 | #include "llvm/IR/Function.h" | ||||||||
110 | #include "llvm/IR/IRBuilder.h" | ||||||||
111 | #include "llvm/IR/InstrTypes.h" | ||||||||
112 | #include "llvm/IR/Instruction.h" | ||||||||
113 | #include "llvm/IR/Instructions.h" | ||||||||
114 | #include "llvm/IR/IntrinsicInst.h" | ||||||||
115 | #include "llvm/IR/Intrinsics.h" | ||||||||
116 | #include "llvm/IR/LLVMContext.h" | ||||||||
117 | #include "llvm/IR/Metadata.h" | ||||||||
118 | #include "llvm/IR/Module.h" | ||||||||
119 | #include "llvm/IR/Operator.h" | ||||||||
120 | #include "llvm/IR/Type.h" | ||||||||
121 | #include "llvm/IR/Use.h" | ||||||||
122 | #include "llvm/IR/User.h" | ||||||||
123 | #include "llvm/IR/Value.h" | ||||||||
124 | #include "llvm/IR/ValueHandle.h" | ||||||||
125 | #include "llvm/IR/Verifier.h" | ||||||||
126 | #include "llvm/InitializePasses.h" | ||||||||
127 | #include "llvm/Pass.h" | ||||||||
128 | #include "llvm/Support/Casting.h" | ||||||||
129 | #include "llvm/Support/CommandLine.h" | ||||||||
130 | #include "llvm/Support/Compiler.h" | ||||||||
131 | #include "llvm/Support/Debug.h" | ||||||||
132 | #include "llvm/Support/ErrorHandling.h" | ||||||||
133 | #include "llvm/Support/MathExtras.h" | ||||||||
134 | #include "llvm/Support/raw_ostream.h" | ||||||||
135 | #include "llvm/Transforms/Utils/BasicBlockUtils.h" | ||||||||
136 | #include "llvm/Transforms/Utils/InjectTLIMappings.h" | ||||||||
137 | #include "llvm/Transforms/Utils/LoopSimplify.h" | ||||||||
138 | #include "llvm/Transforms/Utils/LoopUtils.h" | ||||||||
139 | #include "llvm/Transforms/Utils/LoopVersioning.h" | ||||||||
140 | #include "llvm/Transforms/Utils/ScalarEvolutionExpander.h" | ||||||||
141 | #include "llvm/Transforms/Utils/SizeOpts.h" | ||||||||
142 | #include "llvm/Transforms/Vectorize/LoopVectorizationLegality.h" | ||||||||
143 | #include <algorithm> | ||||||||
144 | #include <cassert> | ||||||||
145 | #include <cstdint> | ||||||||
146 | #include <cstdlib> | ||||||||
147 | #include <functional> | ||||||||
148 | #include <iterator> | ||||||||
149 | #include <limits> | ||||||||
150 | #include <memory> | ||||||||
151 | #include <string> | ||||||||
152 | #include <tuple> | ||||||||
153 | #include <utility> | ||||||||
154 | |||||||||
155 | using namespace llvm; | ||||||||
156 | |||||||||
157 | #define LV_NAME"loop-vectorize" "loop-vectorize" | ||||||||
158 | #define DEBUG_TYPE"loop-vectorize" LV_NAME"loop-vectorize" | ||||||||
159 | |||||||||
160 | /// @{ | ||||||||
161 | /// Metadata attribute names | ||||||||
162 | static const char *const LLVMLoopVectorizeFollowupAll = | ||||||||
163 | "llvm.loop.vectorize.followup_all"; | ||||||||
164 | static const char *const LLVMLoopVectorizeFollowupVectorized = | ||||||||
165 | "llvm.loop.vectorize.followup_vectorized"; | ||||||||
166 | static const char *const LLVMLoopVectorizeFollowupEpilogue = | ||||||||
167 | "llvm.loop.vectorize.followup_epilogue"; | ||||||||
168 | /// @} | ||||||||
169 | |||||||||
170 | STATISTIC(LoopsVectorized, "Number of loops vectorized")static llvm::Statistic LoopsVectorized = {"loop-vectorize", "LoopsVectorized" , "Number of loops vectorized"}; | ||||||||
171 | STATISTIC(LoopsAnalyzed, "Number of loops analyzed for vectorization")static llvm::Statistic LoopsAnalyzed = {"loop-vectorize", "LoopsAnalyzed" , "Number of loops analyzed for vectorization"}; | ||||||||
172 | |||||||||
173 | /// Loops with a known constant trip count below this number are vectorized only | ||||||||
174 | /// if no scalar iteration overheads are incurred. | ||||||||
175 | static cl::opt<unsigned> TinyTripCountVectorThreshold( | ||||||||
176 | "vectorizer-min-trip-count", cl::init(16), cl::Hidden, | ||||||||
177 | cl::desc("Loops with a constant trip count that is smaller than this " | ||||||||
178 | "value are vectorized only if no scalar iteration overheads " | ||||||||
179 | "are incurred.")); | ||||||||
180 | |||||||||
181 | // Option prefer-predicate-over-epilogue indicates that an epilogue is undesired, | ||||||||
182 | // that predication is preferred, and this lists all options. I.e., the | ||||||||
183 | // vectorizer will try to fold the tail-loop (epilogue) into the vector body | ||||||||
184 | // and predicate the instructions accordingly. If tail-folding fails, there are | ||||||||
185 | // different fallback strategies depending on these values: | ||||||||
186 | namespace PreferPredicateTy { | ||||||||
187 | enum Option { | ||||||||
188 | ScalarEpilogue = 0, | ||||||||
189 | PredicateElseScalarEpilogue, | ||||||||
190 | PredicateOrDontVectorize | ||||||||
191 | }; | ||||||||
192 | } // namespace PreferPredicateTy | ||||||||
193 | |||||||||
194 | static cl::opt<PreferPredicateTy::Option> PreferPredicateOverEpilogue( | ||||||||
195 | "prefer-predicate-over-epilogue", | ||||||||
196 | cl::init(PreferPredicateTy::ScalarEpilogue), | ||||||||
197 | cl::Hidden, | ||||||||
198 | cl::desc("Tail-folding and predication preferences over creating a scalar " | ||||||||
199 | "epilogue loop."), | ||||||||
200 | cl::values(clEnumValN(PreferPredicateTy::ScalarEpilogue,llvm::cl::OptionEnumValue { "scalar-epilogue", int(PreferPredicateTy ::ScalarEpilogue), "Don't tail-predicate loops, create scalar epilogue" } | ||||||||
201 | "scalar-epilogue",llvm::cl::OptionEnumValue { "scalar-epilogue", int(PreferPredicateTy ::ScalarEpilogue), "Don't tail-predicate loops, create scalar epilogue" } | ||||||||
202 | "Don't tail-predicate loops, create scalar epilogue")llvm::cl::OptionEnumValue { "scalar-epilogue", int(PreferPredicateTy ::ScalarEpilogue), "Don't tail-predicate loops, create scalar epilogue" }, | ||||||||
203 | clEnumValN(PreferPredicateTy::PredicateElseScalarEpilogue,llvm::cl::OptionEnumValue { "predicate-else-scalar-epilogue", int(PreferPredicateTy::PredicateElseScalarEpilogue), "prefer tail-folding, create scalar epilogue if tail " "folding fails." } | ||||||||
204 | "predicate-else-scalar-epilogue",llvm::cl::OptionEnumValue { "predicate-else-scalar-epilogue", int(PreferPredicateTy::PredicateElseScalarEpilogue), "prefer tail-folding, create scalar epilogue if tail " "folding fails." } | ||||||||
205 | "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." } | ||||||||
206 | "folding fails.")llvm::cl::OptionEnumValue { "predicate-else-scalar-epilogue", int(PreferPredicateTy::PredicateElseScalarEpilogue), "prefer tail-folding, create scalar epilogue if tail " "folding fails." }, | ||||||||
207 | clEnumValN(PreferPredicateTy::PredicateOrDontVectorize,llvm::cl::OptionEnumValue { "predicate-dont-vectorize", int(PreferPredicateTy ::PredicateOrDontVectorize), "prefers tail-folding, don't attempt vectorization if " "tail-folding fails." } | ||||||||
208 | "predicate-dont-vectorize",llvm::cl::OptionEnumValue { "predicate-dont-vectorize", int(PreferPredicateTy ::PredicateOrDontVectorize), "prefers tail-folding, don't attempt vectorization if " "tail-folding fails." } | ||||||||
209 | "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." } | ||||||||
210 | "tail-folding fails.")llvm::cl::OptionEnumValue { "predicate-dont-vectorize", int(PreferPredicateTy ::PredicateOrDontVectorize), "prefers tail-folding, don't attempt vectorization if " "tail-folding fails." })); | ||||||||
211 | |||||||||
212 | static cl::opt<bool> MaximizeBandwidth( | ||||||||
213 | "vectorizer-maximize-bandwidth", cl::init(false), cl::Hidden, | ||||||||
214 | cl::desc("Maximize bandwidth when selecting vectorization factor which " | ||||||||
215 | "will be determined by the smallest type in loop.")); | ||||||||
216 | |||||||||
217 | static cl::opt<bool> EnableInterleavedMemAccesses( | ||||||||
218 | "enable-interleaved-mem-accesses", cl::init(false), cl::Hidden, | ||||||||
219 | cl::desc("Enable vectorization on interleaved memory accesses in a loop")); | ||||||||
220 | |||||||||
221 | /// An interleave-group may need masking if it resides in a block that needs | ||||||||
222 | /// predication, or in order to mask away gaps. | ||||||||
223 | static cl::opt<bool> EnableMaskedInterleavedMemAccesses( | ||||||||
224 | "enable-masked-interleaved-mem-accesses", cl::init(false), cl::Hidden, | ||||||||
225 | cl::desc("Enable vectorization on masked interleaved memory accesses in a loop")); | ||||||||
226 | |||||||||
227 | static cl::opt<unsigned> TinyTripCountInterleaveThreshold( | ||||||||
228 | "tiny-trip-count-interleave-threshold", cl::init(128), cl::Hidden, | ||||||||
229 | cl::desc("We don't interleave loops with a estimated constant trip count " | ||||||||
230 | "below this number")); | ||||||||
231 | |||||||||
232 | static cl::opt<unsigned> ForceTargetNumScalarRegs( | ||||||||
233 | "force-target-num-scalar-regs", cl::init(0), cl::Hidden, | ||||||||
234 | cl::desc("A flag that overrides the target's number of scalar registers.")); | ||||||||
235 | |||||||||
236 | static cl::opt<unsigned> ForceTargetNumVectorRegs( | ||||||||
237 | "force-target-num-vector-regs", cl::init(0), cl::Hidden, | ||||||||
238 | cl::desc("A flag that overrides the target's number of vector registers.")); | ||||||||
239 | |||||||||
240 | static cl::opt<unsigned> ForceTargetMaxScalarInterleaveFactor( | ||||||||
241 | "force-target-max-scalar-interleave", cl::init(0), cl::Hidden, | ||||||||
242 | cl::desc("A flag that overrides the target's max interleave factor for " | ||||||||
243 | "scalar loops.")); | ||||||||
244 | |||||||||
245 | static cl::opt<unsigned> ForceTargetMaxVectorInterleaveFactor( | ||||||||
246 | "force-target-max-vector-interleave", cl::init(0), cl::Hidden, | ||||||||
247 | cl::desc("A flag that overrides the target's max interleave factor for " | ||||||||
248 | "vectorized loops.")); | ||||||||
249 | |||||||||
250 | static cl::opt<unsigned> ForceTargetInstructionCost( | ||||||||
251 | "force-target-instruction-cost", cl::init(0), cl::Hidden, | ||||||||
252 | cl::desc("A flag that overrides the target's expected cost for " | ||||||||
253 | "an instruction to a single constant value. Mostly " | ||||||||
254 | "useful for getting consistent testing.")); | ||||||||
255 | |||||||||
256 | static cl::opt<unsigned> SmallLoopCost( | ||||||||
257 | "small-loop-cost", cl::init(20), cl::Hidden, | ||||||||
258 | cl::desc( | ||||||||
259 | "The cost of a loop that is considered 'small' by the interleaver.")); | ||||||||
260 | |||||||||
261 | static cl::opt<bool> LoopVectorizeWithBlockFrequency( | ||||||||
262 | "loop-vectorize-with-block-frequency", cl::init(true), cl::Hidden, | ||||||||
263 | cl::desc("Enable the use of the block frequency analysis to access PGO " | ||||||||
264 | "heuristics minimizing code growth in cold regions and being more " | ||||||||
265 | "aggressive in hot regions.")); | ||||||||
266 | |||||||||
267 | // Runtime interleave loops for load/store throughput. | ||||||||
268 | static cl::opt<bool> EnableLoadStoreRuntimeInterleave( | ||||||||
269 | "enable-loadstore-runtime-interleave", cl::init(true), cl::Hidden, | ||||||||
270 | cl::desc( | ||||||||
271 | "Enable runtime interleaving until load/store ports are saturated")); | ||||||||
272 | |||||||||
273 | /// Interleave small loops with scalar reductions. | ||||||||
274 | static cl::opt<bool> InterleaveSmallLoopScalarReduction( | ||||||||
275 | "interleave-small-loop-scalar-reduction", cl::init(false), cl::Hidden, | ||||||||
276 | cl::desc("Enable interleaving for loops with small iteration counts that " | ||||||||
277 | "contain scalar reductions to expose ILP.")); | ||||||||
278 | |||||||||
279 | /// The number of stores in a loop that are allowed to need predication. | ||||||||
280 | static cl::opt<unsigned> NumberOfStoresToPredicate( | ||||||||
281 | "vectorize-num-stores-pred", cl::init(1), cl::Hidden, | ||||||||
282 | cl::desc("Max number of stores to be predicated behind an if.")); | ||||||||
283 | |||||||||
284 | static cl::opt<bool> EnableIndVarRegisterHeur( | ||||||||
285 | "enable-ind-var-reg-heur", cl::init(true), cl::Hidden, | ||||||||
286 | cl::desc("Count the induction variable only once when interleaving")); | ||||||||
287 | |||||||||
288 | static cl::opt<bool> EnableCondStoresVectorization( | ||||||||
289 | "enable-cond-stores-vec", cl::init(true), cl::Hidden, | ||||||||
290 | cl::desc("Enable if predication of stores during vectorization.")); | ||||||||
291 | |||||||||
292 | static cl::opt<unsigned> MaxNestedScalarReductionIC( | ||||||||
293 | "max-nested-scalar-reduction-interleave", cl::init(2), cl::Hidden, | ||||||||
294 | cl::desc("The maximum interleave count to use when interleaving a scalar " | ||||||||
295 | "reduction in a nested loop.")); | ||||||||
296 | |||||||||
297 | static cl::opt<bool> | ||||||||
298 | PreferInLoopReductions("prefer-inloop-reductions", cl::init(false), | ||||||||
299 | cl::Hidden, | ||||||||
300 | cl::desc("Prefer in-loop vector reductions, " | ||||||||
301 | "overriding the targets preference.")); | ||||||||
302 | |||||||||
303 | static cl::opt<bool> PreferPredicatedReductionSelect( | ||||||||
304 | "prefer-predicated-reduction-select", cl::init(false), cl::Hidden, | ||||||||
305 | cl::desc( | ||||||||
306 | "Prefer predicating a reduction operation over an after loop select.")); | ||||||||
307 | |||||||||
308 | cl::opt<bool> EnableVPlanNativePath( | ||||||||
309 | "enable-vplan-native-path", cl::init(false), cl::Hidden, | ||||||||
310 | cl::desc("Enable VPlan-native vectorization path with " | ||||||||
311 | "support for outer loop vectorization.")); | ||||||||
312 | |||||||||
313 | // FIXME: Remove this switch once we have divergence analysis. Currently we | ||||||||
314 | // assume divergent non-backedge branches when this switch is true. | ||||||||
315 | cl::opt<bool> EnableVPlanPredication( | ||||||||
316 | "enable-vplan-predication", cl::init(false), cl::Hidden, | ||||||||
317 | cl::desc("Enable VPlan-native vectorization path predicator with " | ||||||||
318 | "support for outer loop vectorization.")); | ||||||||
319 | |||||||||
320 | // This flag enables the stress testing of the VPlan H-CFG construction in the | ||||||||
321 | // VPlan-native vectorization path. It must be used in conjuction with | ||||||||
322 | // -enable-vplan-native-path. -vplan-verify-hcfg can also be used to enable the | ||||||||
323 | // verification of the H-CFGs built. | ||||||||
324 | static cl::opt<bool> VPlanBuildStressTest( | ||||||||
325 | "vplan-build-stress-test", cl::init(false), cl::Hidden, | ||||||||
326 | cl::desc( | ||||||||
327 | "Build VPlan for every supported loop nest in the function and bail " | ||||||||
328 | "out right after the build (stress test the VPlan H-CFG construction " | ||||||||
329 | "in the VPlan-native vectorization path).")); | ||||||||
330 | |||||||||
331 | cl::opt<bool> llvm::EnableLoopInterleaving( | ||||||||
332 | "interleave-loops", cl::init(true), cl::Hidden, | ||||||||
333 | cl::desc("Enable loop interleaving in Loop vectorization passes")); | ||||||||
334 | cl::opt<bool> llvm::EnableLoopVectorization( | ||||||||
335 | "vectorize-loops", cl::init(true), cl::Hidden, | ||||||||
336 | cl::desc("Run the Loop vectorization passes")); | ||||||||
337 | |||||||||
338 | /// A helper function that returns the type of loaded or stored value. | ||||||||
339 | static Type *getMemInstValueType(Value *I) { | ||||||||
340 | assert((isa<LoadInst>(I) || isa<StoreInst>(I)) &&(((isa<LoadInst>(I) || isa<StoreInst>(I)) && "Expected Load or Store instruction") ? static_cast<void> (0) : __assert_fail ("(isa<LoadInst>(I) || isa<StoreInst>(I)) && \"Expected Load or Store instruction\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 341, __PRETTY_FUNCTION__)) | ||||||||
341 | "Expected Load or Store instruction")(((isa<LoadInst>(I) || isa<StoreInst>(I)) && "Expected Load or Store instruction") ? static_cast<void> (0) : __assert_fail ("(isa<LoadInst>(I) || isa<StoreInst>(I)) && \"Expected Load or Store instruction\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 341, __PRETTY_FUNCTION__)); | ||||||||
342 | if (auto *LI = dyn_cast<LoadInst>(I)) | ||||||||
343 | return LI->getType(); | ||||||||
344 | return cast<StoreInst>(I)->getValueOperand()->getType(); | ||||||||
345 | } | ||||||||
346 | |||||||||
347 | /// A helper function that returns true if the given type is irregular. The | ||||||||
348 | /// type is irregular if its allocated size doesn't equal the store size of an | ||||||||
349 | /// element of the corresponding vector type at the given vectorization factor. | ||||||||
350 | static bool hasIrregularType(Type *Ty, const DataLayout &DL, ElementCount VF) { | ||||||||
351 | assert(!VF.isScalable() && "scalable vectors not yet supported.")((!VF.isScalable() && "scalable vectors not yet supported." ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"scalable vectors not yet supported.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 351, __PRETTY_FUNCTION__)); | ||||||||
352 | // Determine if an array of VF elements of type Ty is "bitcast compatible" | ||||||||
353 | // with a <VF x Ty> vector. | ||||||||
354 | if (VF.isVector()) { | ||||||||
355 | auto *VectorTy = VectorType::get(Ty, VF); | ||||||||
356 | return VF * DL.getTypeAllocSize(Ty) != DL.getTypeStoreSize(VectorTy); | ||||||||
357 | } | ||||||||
358 | |||||||||
359 | // If the vectorization factor is one, we just check if an array of type Ty | ||||||||
360 | // requires padding between elements. | ||||||||
361 | return DL.getTypeAllocSizeInBits(Ty) != DL.getTypeSizeInBits(Ty); | ||||||||
362 | } | ||||||||
363 | |||||||||
364 | /// A helper function that returns the reciprocal of the block probability of | ||||||||
365 | /// predicated blocks. If we return X, we are assuming the predicated block | ||||||||
366 | /// will execute once for every X iterations of the loop header. | ||||||||
367 | /// | ||||||||
368 | /// TODO: We should use actual block probability here, if available. Currently, | ||||||||
369 | /// we always assume predicated blocks have a 50% chance of executing. | ||||||||
370 | static unsigned getReciprocalPredBlockProb() { return 2; } | ||||||||
371 | |||||||||
372 | /// A helper function that adds a 'fast' flag to floating-point operations. | ||||||||
373 | static Value *addFastMathFlag(Value *V) { | ||||||||
374 | if (isa<FPMathOperator>(V)) | ||||||||
375 | cast<Instruction>(V)->setFastMathFlags(FastMathFlags::getFast()); | ||||||||
376 | return V; | ||||||||
377 | } | ||||||||
378 | |||||||||
379 | static Value *addFastMathFlag(Value *V, FastMathFlags FMF) { | ||||||||
380 | if (isa<FPMathOperator>(V)) | ||||||||
381 | cast<Instruction>(V)->setFastMathFlags(FMF); | ||||||||
382 | return V; | ||||||||
383 | } | ||||||||
384 | |||||||||
385 | /// A helper function that returns an integer or floating-point constant with | ||||||||
386 | /// value C. | ||||||||
387 | static Constant *getSignedIntOrFpConstant(Type *Ty, int64_t C) { | ||||||||
388 | return Ty->isIntegerTy() ? ConstantInt::getSigned(Ty, C) | ||||||||
389 | : ConstantFP::get(Ty, C); | ||||||||
390 | } | ||||||||
391 | |||||||||
392 | /// Returns "best known" trip count for the specified loop \p L as defined by | ||||||||
393 | /// the following procedure: | ||||||||
394 | /// 1) Returns exact trip count if it is known. | ||||||||
395 | /// 2) Returns expected trip count according to profile data if any. | ||||||||
396 | /// 3) Returns upper bound estimate if it is known. | ||||||||
397 | /// 4) Returns None if all of the above failed. | ||||||||
398 | static Optional<unsigned> getSmallBestKnownTC(ScalarEvolution &SE, Loop *L) { | ||||||||
399 | // Check if exact trip count is known. | ||||||||
400 | if (unsigned ExpectedTC = SE.getSmallConstantTripCount(L)) | ||||||||
401 | return ExpectedTC; | ||||||||
402 | |||||||||
403 | // Check if there is an expected trip count available from profile data. | ||||||||
404 | if (LoopVectorizeWithBlockFrequency) | ||||||||
405 | if (auto EstimatedTC = getLoopEstimatedTripCount(L)) | ||||||||
406 | return EstimatedTC; | ||||||||
407 | |||||||||
408 | // Check if upper bound estimate is known. | ||||||||
409 | if (unsigned ExpectedTC = SE.getSmallConstantMaxTripCount(L)) | ||||||||
410 | return ExpectedTC; | ||||||||
411 | |||||||||
412 | return None; | ||||||||
413 | } | ||||||||
414 | |||||||||
415 | namespace llvm { | ||||||||
416 | |||||||||
417 | /// InnerLoopVectorizer vectorizes loops which contain only one basic | ||||||||
418 | /// block to a specified vectorization factor (VF). | ||||||||
419 | /// This class performs the widening of scalars into vectors, or multiple | ||||||||
420 | /// scalars. This class also implements the following features: | ||||||||
421 | /// * It inserts an epilogue loop for handling loops that don't have iteration | ||||||||
422 | /// counts that are known to be a multiple of the vectorization factor. | ||||||||
423 | /// * It handles the code generation for reduction variables. | ||||||||
424 | /// * Scalarization (implementation using scalars) of un-vectorizable | ||||||||
425 | /// instructions. | ||||||||
426 | /// InnerLoopVectorizer does not perform any vectorization-legality | ||||||||
427 | /// checks, and relies on the caller to check for the different legality | ||||||||
428 | /// aspects. The InnerLoopVectorizer relies on the | ||||||||
429 | /// LoopVectorizationLegality class to provide information about the induction | ||||||||
430 | /// and reduction variables that were found to a given vectorization factor. | ||||||||
431 | class InnerLoopVectorizer { | ||||||||
432 | public: | ||||||||
433 | InnerLoopVectorizer(Loop *OrigLoop, PredicatedScalarEvolution &PSE, | ||||||||
434 | LoopInfo *LI, DominatorTree *DT, | ||||||||
435 | const TargetLibraryInfo *TLI, | ||||||||
436 | const TargetTransformInfo *TTI, AssumptionCache *AC, | ||||||||
437 | OptimizationRemarkEmitter *ORE, ElementCount VecWidth, | ||||||||
438 | unsigned UnrollFactor, LoopVectorizationLegality *LVL, | ||||||||
439 | LoopVectorizationCostModel *CM, BlockFrequencyInfo *BFI, | ||||||||
440 | ProfileSummaryInfo *PSI) | ||||||||
441 | : OrigLoop(OrigLoop), PSE(PSE), LI(LI), DT(DT), TLI(TLI), TTI(TTI), | ||||||||
442 | AC(AC), ORE(ORE), VF(VecWidth), UF(UnrollFactor), | ||||||||
443 | Builder(PSE.getSE()->getContext()), | ||||||||
444 | VectorLoopValueMap(UnrollFactor, VecWidth), Legal(LVL), Cost(CM), | ||||||||
445 | BFI(BFI), PSI(PSI) { | ||||||||
446 | // Query this against the original loop and save it here because the profile | ||||||||
447 | // of the original loop header may change as the transformation happens. | ||||||||
448 | OptForSizeBasedOnProfile = llvm::shouldOptimizeForSize( | ||||||||
449 | OrigLoop->getHeader(), PSI, BFI, PGSOQueryType::IRPass); | ||||||||
450 | } | ||||||||
451 | |||||||||
452 | virtual ~InnerLoopVectorizer() = default; | ||||||||
453 | |||||||||
454 | /// Create a new empty loop that will contain vectorized instructions later | ||||||||
455 | /// on, while the old loop will be used as the scalar remainder. Control flow | ||||||||
456 | /// is generated around the vectorized (and scalar epilogue) loops consisting | ||||||||
457 | /// of various checks and bypasses. Return the pre-header block of the new | ||||||||
458 | /// loop. | ||||||||
459 | BasicBlock *createVectorizedLoopSkeleton(); | ||||||||
460 | |||||||||
461 | /// Widen a single instruction within the innermost loop. | ||||||||
462 | void widenInstruction(Instruction &I, VPUser &Operands, | ||||||||
463 | VPTransformState &State); | ||||||||
464 | |||||||||
465 | /// Widen a single call instruction within the innermost loop. | ||||||||
466 | void widenCallInstruction(CallInst &I, VPUser &ArgOperands, | ||||||||
467 | VPTransformState &State); | ||||||||
468 | |||||||||
469 | /// Widen a single select instruction within the innermost loop. | ||||||||
470 | void widenSelectInstruction(SelectInst &I, VPUser &Operands, | ||||||||
471 | bool InvariantCond, VPTransformState &State); | ||||||||
472 | |||||||||
473 | /// Fix the vectorized code, taking care of header phi's, live-outs, and more. | ||||||||
474 | void fixVectorizedLoop(); | ||||||||
475 | |||||||||
476 | // Return true if any runtime check is added. | ||||||||
477 | bool areSafetyChecksAdded() { return AddedSafetyChecks; } | ||||||||
478 | |||||||||
479 | /// A type for vectorized values in the new loop. Each value from the | ||||||||
480 | /// original loop, when vectorized, is represented by UF vector values in the | ||||||||
481 | /// new unrolled loop, where UF is the unroll factor. | ||||||||
482 | using VectorParts = SmallVector<Value *, 2>; | ||||||||
483 | |||||||||
484 | /// Vectorize a single GetElementPtrInst based on information gathered and | ||||||||
485 | /// decisions taken during planning. | ||||||||
486 | void widenGEP(GetElementPtrInst *GEP, VPUser &Indices, unsigned UF, | ||||||||
487 | ElementCount VF, bool IsPtrLoopInvariant, | ||||||||
488 | SmallBitVector &IsIndexLoopInvariant, VPTransformState &State); | ||||||||
489 | |||||||||
490 | /// Vectorize a single PHINode in a block. This method handles the induction | ||||||||
491 | /// variable canonicalization. It supports both VF = 1 for unrolled loops and | ||||||||
492 | /// arbitrary length vectors. | ||||||||
493 | void widenPHIInstruction(Instruction *PN, unsigned UF, ElementCount VF); | ||||||||
494 | |||||||||
495 | /// A helper function to scalarize a single Instruction in the innermost loop. | ||||||||
496 | /// Generates a sequence of scalar instances for each lane between \p MinLane | ||||||||
497 | /// and \p MaxLane, times each part between \p MinPart and \p MaxPart, | ||||||||
498 | /// inclusive. Uses the VPValue operands from \p Operands instead of \p | ||||||||
499 | /// Instr's operands. | ||||||||
500 | void scalarizeInstruction(Instruction *Instr, VPUser &Operands, | ||||||||
501 | const VPIteration &Instance, bool IfPredicateInstr, | ||||||||
502 | VPTransformState &State); | ||||||||
503 | |||||||||
504 | /// Widen an integer or floating-point induction variable \p IV. If \p Trunc | ||||||||
505 | /// is provided, the integer induction variable will first be truncated to | ||||||||
506 | /// the corresponding type. | ||||||||
507 | void widenIntOrFpInduction(PHINode *IV, TruncInst *Trunc = nullptr); | ||||||||
508 | |||||||||
509 | /// getOrCreateVectorValue and getOrCreateScalarValue coordinate to generate a | ||||||||
510 | /// vector or scalar value on-demand if one is not yet available. When | ||||||||
511 | /// vectorizing a loop, we visit the definition of an instruction before its | ||||||||
512 | /// uses. When visiting the definition, we either vectorize or scalarize the | ||||||||
513 | /// instruction, creating an entry for it in the corresponding map. (In some | ||||||||
514 | /// cases, such as induction variables, we will create both vector and scalar | ||||||||
515 | /// entries.) Then, as we encounter uses of the definition, we derive values | ||||||||
516 | /// for each scalar or vector use unless such a value is already available. | ||||||||
517 | /// For example, if we scalarize a definition and one of its uses is vector, | ||||||||
518 | /// we build the required vector on-demand with an insertelement sequence | ||||||||
519 | /// when visiting the use. Otherwise, if the use is scalar, we can use the | ||||||||
520 | /// existing scalar definition. | ||||||||
521 | /// | ||||||||
522 | /// Return a value in the new loop corresponding to \p V from the original | ||||||||
523 | /// loop at unroll index \p Part. If the value has already been vectorized, | ||||||||
524 | /// the corresponding vector entry in VectorLoopValueMap is returned. If, | ||||||||
525 | /// however, the value has a scalar entry in VectorLoopValueMap, we construct | ||||||||
526 | /// a new vector value on-demand by inserting the scalar values into a vector | ||||||||
527 | /// with an insertelement sequence. If the value has been neither vectorized | ||||||||
528 | /// nor scalarized, it must be loop invariant, so we simply broadcast the | ||||||||
529 | /// value into a vector. | ||||||||
530 | Value *getOrCreateVectorValue(Value *V, unsigned Part); | ||||||||
531 | |||||||||
532 | /// Return a value in the new loop corresponding to \p V from the original | ||||||||
533 | /// loop at unroll and vector indices \p Instance. If the value has been | ||||||||
534 | /// vectorized but not scalarized, the necessary extractelement instruction | ||||||||
535 | /// will be generated. | ||||||||
536 | Value *getOrCreateScalarValue(Value *V, const VPIteration &Instance); | ||||||||
537 | |||||||||
538 | /// Construct the vector value of a scalarized value \p V one lane at a time. | ||||||||
539 | void packScalarIntoVectorValue(Value *V, const VPIteration &Instance); | ||||||||
540 | |||||||||
541 | /// Try to vectorize interleaved access group \p Group with the base address | ||||||||
542 | /// given in \p Addr, optionally masking the vector operations if \p | ||||||||
543 | /// BlockInMask is non-null. Use \p State to translate given VPValues to IR | ||||||||
544 | /// values in the vectorized loop. | ||||||||
545 | void vectorizeInterleaveGroup(const InterleaveGroup<Instruction> *Group, | ||||||||
546 | VPTransformState &State, VPValue *Addr, | ||||||||
547 | VPValue *BlockInMask = nullptr); | ||||||||
548 | |||||||||
549 | /// Vectorize Load and Store instructions with the base address given in \p | ||||||||
550 | /// Addr, optionally masking the vector operations if \p BlockInMask is | ||||||||
551 | /// non-null. Use \p State to translate given VPValues to IR values in the | ||||||||
552 | /// vectorized loop. | ||||||||
553 | void vectorizeMemoryInstruction(Instruction *Instr, VPTransformState &State, | ||||||||
554 | VPValue *Addr, VPValue *StoredValue, | ||||||||
555 | VPValue *BlockInMask); | ||||||||
556 | |||||||||
557 | /// Set the debug location in the builder using the debug location in | ||||||||
558 | /// the instruction. | ||||||||
559 | void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr); | ||||||||
560 | |||||||||
561 | /// Fix the non-induction PHIs in the OrigPHIsToFix vector. | ||||||||
562 | void fixNonInductionPHIs(void); | ||||||||
563 | |||||||||
564 | protected: | ||||||||
565 | friend class LoopVectorizationPlanner; | ||||||||
566 | |||||||||
567 | /// A small list of PHINodes. | ||||||||
568 | using PhiVector = SmallVector<PHINode *, 4>; | ||||||||
569 | |||||||||
570 | /// A type for scalarized values in the new loop. Each value from the | ||||||||
571 | /// original loop, when scalarized, is represented by UF x VF scalar values | ||||||||
572 | /// in the new unrolled loop, where UF is the unroll factor and VF is the | ||||||||
573 | /// vectorization factor. | ||||||||
574 | using ScalarParts = SmallVector<SmallVector<Value *, 4>, 2>; | ||||||||
575 | |||||||||
576 | /// Set up the values of the IVs correctly when exiting the vector loop. | ||||||||
577 | void fixupIVUsers(PHINode *OrigPhi, const InductionDescriptor &II, | ||||||||
578 | Value *CountRoundDown, Value *EndValue, | ||||||||
579 | BasicBlock *MiddleBlock); | ||||||||
580 | |||||||||
581 | /// Create a new induction variable inside L. | ||||||||
582 | PHINode *createInductionVariable(Loop *L, Value *Start, Value *End, | ||||||||
583 | Value *Step, Instruction *DL); | ||||||||
584 | |||||||||
585 | /// Handle all cross-iteration phis in the header. | ||||||||
586 | void fixCrossIterationPHIs(); | ||||||||
587 | |||||||||
588 | /// Fix a first-order recurrence. This is the second phase of vectorizing | ||||||||
589 | /// this phi node. | ||||||||
590 | void fixFirstOrderRecurrence(PHINode *Phi); | ||||||||
591 | |||||||||
592 | /// Fix a reduction cross-iteration phi. This is the second phase of | ||||||||
593 | /// vectorizing this phi node. | ||||||||
594 | void fixReduction(PHINode *Phi); | ||||||||
595 | |||||||||
596 | /// Clear NSW/NUW flags from reduction instructions if necessary. | ||||||||
597 | void clearReductionWrapFlags(RecurrenceDescriptor &RdxDesc); | ||||||||
598 | |||||||||
599 | /// The Loop exit block may have single value PHI nodes with some | ||||||||
600 | /// incoming value. While vectorizing we only handled real values | ||||||||
601 | /// that were defined inside the loop and we should have one value for | ||||||||
602 | /// each predecessor of its parent basic block. See PR14725. | ||||||||
603 | void fixLCSSAPHIs(); | ||||||||
604 | |||||||||
605 | /// Iteratively sink the scalarized operands of a predicated instruction into | ||||||||
606 | /// the block that was created for it. | ||||||||
607 | void sinkScalarOperands(Instruction *PredInst); | ||||||||
608 | |||||||||
609 | /// Shrinks vector element sizes to the smallest bitwidth they can be legally | ||||||||
610 | /// represented as. | ||||||||
611 | void truncateToMinimalBitwidths(); | ||||||||
612 | |||||||||
613 | /// Create a broadcast instruction. This method generates a broadcast | ||||||||
614 | /// instruction (shuffle) for loop invariant values and for the induction | ||||||||
615 | /// value. If this is the induction variable then we extend it to N, N+1, ... | ||||||||
616 | /// this is needed because each iteration in the loop corresponds to a SIMD | ||||||||
617 | /// element. | ||||||||
618 | virtual Value *getBroadcastInstrs(Value *V); | ||||||||
619 | |||||||||
620 | /// This function adds (StartIdx, StartIdx + Step, StartIdx + 2*Step, ...) | ||||||||
621 | /// to each vector element of Val. The sequence starts at StartIndex. | ||||||||
622 | /// \p Opcode is relevant for FP induction variable. | ||||||||
623 | virtual Value *getStepVector(Value *Val, int StartIdx, Value *Step, | ||||||||
624 | Instruction::BinaryOps Opcode = | ||||||||
625 | Instruction::BinaryOpsEnd); | ||||||||
626 | |||||||||
627 | /// Compute scalar induction steps. \p ScalarIV is the scalar induction | ||||||||
628 | /// variable on which to base the steps, \p Step is the size of the step, and | ||||||||
629 | /// \p EntryVal is the value from the original loop that maps to the steps. | ||||||||
630 | /// Note that \p EntryVal doesn't have to be an induction variable - it | ||||||||
631 | /// can also be a truncate instruction. | ||||||||
632 | void buildScalarSteps(Value *ScalarIV, Value *Step, Instruction *EntryVal, | ||||||||
633 | const InductionDescriptor &ID); | ||||||||
634 | |||||||||
635 | /// Create a vector induction phi node based on an existing scalar one. \p | ||||||||
636 | /// EntryVal is the value from the original loop that maps to the vector phi | ||||||||
637 | /// node, and \p Step is the loop-invariant step. If \p EntryVal is a | ||||||||
638 | /// truncate instruction, instead of widening the original IV, we widen a | ||||||||
639 | /// version of the IV truncated to \p EntryVal's type. | ||||||||
640 | void createVectorIntOrFpInductionPHI(const InductionDescriptor &II, | ||||||||
641 | Value *Step, Instruction *EntryVal); | ||||||||
642 | |||||||||
643 | /// Returns true if an instruction \p I should be scalarized instead of | ||||||||
644 | /// vectorized for the chosen vectorization factor. | ||||||||
645 | bool shouldScalarizeInstruction(Instruction *I) const; | ||||||||
646 | |||||||||
647 | /// Returns true if we should generate a scalar version of \p IV. | ||||||||
648 | bool needsScalarInduction(Instruction *IV) const; | ||||||||
649 | |||||||||
650 | /// If there is a cast involved in the induction variable \p ID, which should | ||||||||
651 | /// be ignored in the vectorized loop body, this function records the | ||||||||
652 | /// VectorLoopValue of the respective Phi also as the VectorLoopValue of the | ||||||||
653 | /// cast. We had already proved that the casted Phi is equal to the uncasted | ||||||||
654 | /// Phi in the vectorized loop (under a runtime guard), and therefore | ||||||||
655 | /// there is no need to vectorize the cast - the same value can be used in the | ||||||||
656 | /// vector loop for both the Phi and the cast. | ||||||||
657 | /// If \p VectorLoopValue is a scalarized value, \p Lane is also specified, | ||||||||
658 | /// Otherwise, \p VectorLoopValue is a widened/vectorized value. | ||||||||
659 | /// | ||||||||
660 | /// \p EntryVal is the value from the original loop that maps to the vector | ||||||||
661 | /// phi node and is used to distinguish what is the IV currently being | ||||||||
662 | /// processed - original one (if \p EntryVal is a phi corresponding to the | ||||||||
663 | /// original IV) or the "newly-created" one based on the proof mentioned above | ||||||||
664 | /// (see also buildScalarSteps() and createVectorIntOrFPInductionPHI()). In the | ||||||||
665 | /// latter case \p EntryVal is a TruncInst and we must not record anything for | ||||||||
666 | /// that IV, but it's error-prone to expect callers of this routine to care | ||||||||
667 | /// about that, hence this explicit parameter. | ||||||||
668 | void recordVectorLoopValueForInductionCast(const InductionDescriptor &ID, | ||||||||
669 | const Instruction *EntryVal, | ||||||||
670 | Value *VectorLoopValue, | ||||||||
671 | unsigned Part, | ||||||||
672 | unsigned Lane = UINT_MAX(2147483647 *2U +1U)); | ||||||||
673 | |||||||||
674 | /// Generate a shuffle sequence that will reverse the vector Vec. | ||||||||
675 | virtual Value *reverseVector(Value *Vec); | ||||||||
676 | |||||||||
677 | /// Returns (and creates if needed) the original loop trip count. | ||||||||
678 | Value *getOrCreateTripCount(Loop *NewLoop); | ||||||||
679 | |||||||||
680 | /// Returns (and creates if needed) the trip count of the widened loop. | ||||||||
681 | Value *getOrCreateVectorTripCount(Loop *NewLoop); | ||||||||
682 | |||||||||
683 | /// Returns a bitcasted value to the requested vector type. | ||||||||
684 | /// Also handles bitcasts of vector<float> <-> vector<pointer> types. | ||||||||
685 | Value *createBitOrPointerCast(Value *V, VectorType *DstVTy, | ||||||||
686 | const DataLayout &DL); | ||||||||
687 | |||||||||
688 | /// Emit a bypass check to see if the vector trip count is zero, including if | ||||||||
689 | /// it overflows. | ||||||||
690 | void emitMinimumIterationCountCheck(Loop *L, BasicBlock *Bypass); | ||||||||
691 | |||||||||
692 | /// Emit a bypass check to see if all of the SCEV assumptions we've | ||||||||
693 | /// had to make are correct. | ||||||||
694 | void emitSCEVChecks(Loop *L, BasicBlock *Bypass); | ||||||||
695 | |||||||||
696 | /// Emit bypass checks to check any memory assumptions we may have made. | ||||||||
697 | void emitMemRuntimeChecks(Loop *L, BasicBlock *Bypass); | ||||||||
698 | |||||||||
699 | /// Compute the transformed value of Index at offset StartValue using step | ||||||||
700 | /// StepValue. | ||||||||
701 | /// For integer induction, returns StartValue + Index * StepValue. | ||||||||
702 | /// For pointer induction, returns StartValue[Index * StepValue]. | ||||||||
703 | /// FIXME: The newly created binary instructions should contain nsw/nuw | ||||||||
704 | /// flags, which can be found from the original scalar operations. | ||||||||
705 | Value *emitTransformedIndex(IRBuilder<> &B, Value *Index, ScalarEvolution *SE, | ||||||||
706 | const DataLayout &DL, | ||||||||
707 | const InductionDescriptor &ID) const; | ||||||||
708 | |||||||||
709 | /// Emit basic blocks (prefixed with \p Prefix) for the iteration check, | ||||||||
710 | /// vector loop preheader, middle block and scalar preheader. Also | ||||||||
711 | /// allocate a loop object for the new vector loop and return it. | ||||||||
712 | Loop *createVectorLoopSkeleton(StringRef Prefix); | ||||||||
713 | |||||||||
714 | /// Create new phi nodes for the induction variables to resume iteration count | ||||||||
715 | /// in the scalar epilogue, from where the vectorized loop left off (given by | ||||||||
716 | /// \p VectorTripCount). | ||||||||
717 | void createInductionResumeValues(Loop *L, Value *VectorTripCount); | ||||||||
718 | |||||||||
719 | /// Complete the loop skeleton by adding debug MDs, creating appropriate | ||||||||
720 | /// conditional branches in the middle block, preparing the builder and | ||||||||
721 | /// running the verifier. Take in the vector loop \p L as argument, and return | ||||||||
722 | /// the preheader of the completed vector loop. | ||||||||
723 | BasicBlock *completeLoopSkeleton(Loop *L, MDNode *OrigLoopID); | ||||||||
724 | |||||||||
725 | /// Add additional metadata to \p To that was not present on \p Orig. | ||||||||
726 | /// | ||||||||
727 | /// Currently this is used to add the noalias annotations based on the | ||||||||
728 | /// inserted memchecks. Use this for instructions that are *cloned* into the | ||||||||
729 | /// vector loop. | ||||||||
730 | void addNewMetadata(Instruction *To, const Instruction *Orig); | ||||||||
731 | |||||||||
732 | /// Add metadata from one instruction to another. | ||||||||
733 | /// | ||||||||
734 | /// This includes both the original MDs from \p From and additional ones (\see | ||||||||
735 | /// addNewMetadata). Use this for *newly created* instructions in the vector | ||||||||
736 | /// loop. | ||||||||
737 | void addMetadata(Instruction *To, Instruction *From); | ||||||||
738 | |||||||||
739 | /// Similar to the previous function but it adds the metadata to a | ||||||||
740 | /// vector of instructions. | ||||||||
741 | void addMetadata(ArrayRef<Value *> To, Instruction *From); | ||||||||
742 | |||||||||
743 | /// The original loop. | ||||||||
744 | Loop *OrigLoop; | ||||||||
745 | |||||||||
746 | /// A wrapper around ScalarEvolution used to add runtime SCEV checks. Applies | ||||||||
747 | /// dynamic knowledge to simplify SCEV expressions and converts them to a | ||||||||
748 | /// more usable form. | ||||||||
749 | PredicatedScalarEvolution &PSE; | ||||||||
750 | |||||||||
751 | /// Loop Info. | ||||||||
752 | LoopInfo *LI; | ||||||||
753 | |||||||||
754 | /// Dominator Tree. | ||||||||
755 | DominatorTree *DT; | ||||||||
756 | |||||||||
757 | /// Alias Analysis. | ||||||||
758 | AAResults *AA; | ||||||||
759 | |||||||||
760 | /// Target Library Info. | ||||||||
761 | const TargetLibraryInfo *TLI; | ||||||||
762 | |||||||||
763 | /// Target Transform Info. | ||||||||
764 | const TargetTransformInfo *TTI; | ||||||||
765 | |||||||||
766 | /// Assumption Cache. | ||||||||
767 | AssumptionCache *AC; | ||||||||
768 | |||||||||
769 | /// Interface to emit optimization remarks. | ||||||||
770 | OptimizationRemarkEmitter *ORE; | ||||||||
771 | |||||||||
772 | /// LoopVersioning. It's only set up (non-null) if memchecks were | ||||||||
773 | /// used. | ||||||||
774 | /// | ||||||||
775 | /// This is currently only used to add no-alias metadata based on the | ||||||||
776 | /// memchecks. The actually versioning is performed manually. | ||||||||
777 | std::unique_ptr<LoopVersioning> LVer; | ||||||||
778 | |||||||||
779 | /// The vectorization SIMD factor to use. Each vector will have this many | ||||||||
780 | /// vector elements. | ||||||||
781 | ElementCount VF; | ||||||||
782 | |||||||||
783 | /// The vectorization unroll factor to use. Each scalar is vectorized to this | ||||||||
784 | /// many different vector instructions. | ||||||||
785 | unsigned UF; | ||||||||
786 | |||||||||
787 | /// The builder that we use | ||||||||
788 | IRBuilder<> Builder; | ||||||||
789 | |||||||||
790 | // --- Vectorization state --- | ||||||||
791 | |||||||||
792 | /// The vector-loop preheader. | ||||||||
793 | BasicBlock *LoopVectorPreHeader; | ||||||||
794 | |||||||||
795 | /// The scalar-loop preheader. | ||||||||
796 | BasicBlock *LoopScalarPreHeader; | ||||||||
797 | |||||||||
798 | /// Middle Block between the vector and the scalar. | ||||||||
799 | BasicBlock *LoopMiddleBlock; | ||||||||
800 | |||||||||
801 | /// The ExitBlock of the scalar loop. | ||||||||
802 | BasicBlock *LoopExitBlock; | ||||||||
803 | |||||||||
804 | /// The vector loop body. | ||||||||
805 | BasicBlock *LoopVectorBody; | ||||||||
806 | |||||||||
807 | /// The scalar loop body. | ||||||||
808 | BasicBlock *LoopScalarBody; | ||||||||
809 | |||||||||
810 | /// A list of all bypass blocks. The first block is the entry of the loop. | ||||||||
811 | SmallVector<BasicBlock *, 4> LoopBypassBlocks; | ||||||||
812 | |||||||||
813 | /// The new Induction variable which was added to the new block. | ||||||||
814 | PHINode *Induction = nullptr; | ||||||||
815 | |||||||||
816 | /// The induction variable of the old basic block. | ||||||||
817 | PHINode *OldInduction = nullptr; | ||||||||
818 | |||||||||
819 | /// Maps values from the original loop to their corresponding values in the | ||||||||
820 | /// vectorized loop. A key value can map to either vector values, scalar | ||||||||
821 | /// values or both kinds of values, depending on whether the key was | ||||||||
822 | /// vectorized and scalarized. | ||||||||
823 | VectorizerValueMap VectorLoopValueMap; | ||||||||
824 | |||||||||
825 | /// Store instructions that were predicated. | ||||||||
826 | SmallVector<Instruction *, 4> PredicatedInstructions; | ||||||||
827 | |||||||||
828 | /// Trip count of the original loop. | ||||||||
829 | Value *TripCount = nullptr; | ||||||||
830 | |||||||||
831 | /// Trip count of the widened loop (TripCount - TripCount % (VF*UF)) | ||||||||
832 | Value *VectorTripCount = nullptr; | ||||||||
833 | |||||||||
834 | /// The legality analysis. | ||||||||
835 | LoopVectorizationLegality *Legal; | ||||||||
836 | |||||||||
837 | /// The profitablity analysis. | ||||||||
838 | LoopVectorizationCostModel *Cost; | ||||||||
839 | |||||||||
840 | // Record whether runtime checks are added. | ||||||||
841 | bool AddedSafetyChecks = false; | ||||||||
842 | |||||||||
843 | // Holds the end values for each induction variable. We save the end values | ||||||||
844 | // so we can later fix-up the external users of the induction variables. | ||||||||
845 | DenseMap<PHINode *, Value *> IVEndValues; | ||||||||
846 | |||||||||
847 | // Vector of original scalar PHIs whose corresponding widened PHIs need to be | ||||||||
848 | // fixed up at the end of vector code generation. | ||||||||
849 | SmallVector<PHINode *, 8> OrigPHIsToFix; | ||||||||
850 | |||||||||
851 | /// BFI and PSI are used to check for profile guided size optimizations. | ||||||||
852 | BlockFrequencyInfo *BFI; | ||||||||
853 | ProfileSummaryInfo *PSI; | ||||||||
854 | |||||||||
855 | // Whether this loop should be optimized for size based on profile guided size | ||||||||
856 | // optimizatios. | ||||||||
857 | bool OptForSizeBasedOnProfile; | ||||||||
858 | }; | ||||||||
859 | |||||||||
860 | class InnerLoopUnroller : public InnerLoopVectorizer { | ||||||||
861 | public: | ||||||||
862 | InnerLoopUnroller(Loop *OrigLoop, PredicatedScalarEvolution &PSE, | ||||||||
863 | LoopInfo *LI, DominatorTree *DT, | ||||||||
864 | const TargetLibraryInfo *TLI, | ||||||||
865 | const TargetTransformInfo *TTI, AssumptionCache *AC, | ||||||||
866 | OptimizationRemarkEmitter *ORE, unsigned UnrollFactor, | ||||||||
867 | LoopVectorizationLegality *LVL, | ||||||||
868 | LoopVectorizationCostModel *CM, BlockFrequencyInfo *BFI, | ||||||||
869 | ProfileSummaryInfo *PSI) | ||||||||
870 | : InnerLoopVectorizer(OrigLoop, PSE, LI, DT, TLI, TTI, AC, ORE, | ||||||||
871 | ElementCount::getFixed(1), UnrollFactor, LVL, CM, | ||||||||
872 | BFI, PSI) {} | ||||||||
873 | |||||||||
874 | private: | ||||||||
875 | Value *getBroadcastInstrs(Value *V) override; | ||||||||
876 | Value *getStepVector(Value *Val, int StartIdx, Value *Step, | ||||||||
877 | Instruction::BinaryOps Opcode = | ||||||||
878 | Instruction::BinaryOpsEnd) override; | ||||||||
879 | Value *reverseVector(Value *Vec) override; | ||||||||
880 | }; | ||||||||
881 | |||||||||
882 | } // end namespace llvm | ||||||||
883 | |||||||||
884 | /// Look for a meaningful debug location on the instruction or it's | ||||||||
885 | /// operands. | ||||||||
886 | static Instruction *getDebugLocFromInstOrOperands(Instruction *I) { | ||||||||
887 | if (!I) | ||||||||
888 | return I; | ||||||||
889 | |||||||||
890 | DebugLoc Empty; | ||||||||
891 | if (I->getDebugLoc() != Empty) | ||||||||
892 | return I; | ||||||||
893 | |||||||||
894 | for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) { | ||||||||
895 | if (Instruction *OpInst = dyn_cast<Instruction>(*OI)) | ||||||||
896 | if (OpInst->getDebugLoc() != Empty) | ||||||||
897 | return OpInst; | ||||||||
898 | } | ||||||||
899 | |||||||||
900 | return I; | ||||||||
901 | } | ||||||||
902 | |||||||||
903 | void InnerLoopVectorizer::setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) { | ||||||||
904 | if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr)) { | ||||||||
905 | const DILocation *DIL = Inst->getDebugLoc(); | ||||||||
906 | if (DIL && Inst->getFunction()->isDebugInfoForProfiling() && | ||||||||
907 | !isa<DbgInfoIntrinsic>(Inst)) { | ||||||||
908 | assert(!VF.isScalable() && "scalable vectors not yet supported.")((!VF.isScalable() && "scalable vectors not yet supported." ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"scalable vectors not yet supported.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 908, __PRETTY_FUNCTION__)); | ||||||||
909 | auto NewDIL = | ||||||||
910 | DIL->cloneByMultiplyingDuplicationFactor(UF * VF.getKnownMinValue()); | ||||||||
911 | if (NewDIL) | ||||||||
912 | B.SetCurrentDebugLocation(NewDIL.getValue()); | ||||||||
913 | else | ||||||||
914 | LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "Failed to create new discriminator: " << DIL->getFilename() << " Line: " << DIL ->getLine(); } } while (false) | ||||||||
915 | << "Failed to create new discriminator: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "Failed to create new discriminator: " << DIL->getFilename() << " Line: " << DIL ->getLine(); } } while (false) | ||||||||
916 | << DIL->getFilename() << " Line: " << DIL->getLine())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "Failed to create new discriminator: " << DIL->getFilename() << " Line: " << DIL ->getLine(); } } while (false); | ||||||||
917 | } | ||||||||
918 | else | ||||||||
919 | B.SetCurrentDebugLocation(DIL); | ||||||||
920 | } else | ||||||||
921 | B.SetCurrentDebugLocation(DebugLoc()); | ||||||||
922 | } | ||||||||
923 | |||||||||
924 | /// Write a record \p DebugMsg about vectorization failure to the debug | ||||||||
925 | /// output stream. If \p I is passed, it is an instruction that prevents | ||||||||
926 | /// vectorization. | ||||||||
927 | #ifndef NDEBUG | ||||||||
928 | static void debugVectorizationFailure(const StringRef DebugMsg, | ||||||||
929 | Instruction *I) { | ||||||||
930 | dbgs() << "LV: Not vectorizing: " << DebugMsg; | ||||||||
931 | if (I != nullptr) | ||||||||
932 | dbgs() << " " << *I; | ||||||||
933 | else | ||||||||
934 | dbgs() << '.'; | ||||||||
935 | dbgs() << '\n'; | ||||||||
936 | } | ||||||||
937 | #endif | ||||||||
938 | |||||||||
939 | /// Create an analysis remark that explains why vectorization failed | ||||||||
940 | /// | ||||||||
941 | /// \p PassName is the name of the pass (e.g. can be AlwaysPrint). \p | ||||||||
942 | /// RemarkName is the identifier for the remark. If \p I is passed it is an | ||||||||
943 | /// instruction that prevents vectorization. Otherwise \p TheLoop is used for | ||||||||
944 | /// the location of the remark. \return the remark object that can be | ||||||||
945 | /// streamed to. | ||||||||
946 | static OptimizationRemarkAnalysis createLVAnalysis(const char *PassName, | ||||||||
947 | StringRef RemarkName, Loop *TheLoop, Instruction *I) { | ||||||||
948 | Value *CodeRegion = TheLoop->getHeader(); | ||||||||
949 | DebugLoc DL = TheLoop->getStartLoc(); | ||||||||
950 | |||||||||
951 | if (I) { | ||||||||
952 | CodeRegion = I->getParent(); | ||||||||
953 | // If there is no debug location attached to the instruction, revert back to | ||||||||
954 | // using the loop's. | ||||||||
955 | if (I->getDebugLoc()) | ||||||||
956 | DL = I->getDebugLoc(); | ||||||||
957 | } | ||||||||
958 | |||||||||
959 | OptimizationRemarkAnalysis R(PassName, RemarkName, DL, CodeRegion); | ||||||||
960 | R << "loop not vectorized: "; | ||||||||
961 | return R; | ||||||||
962 | } | ||||||||
963 | |||||||||
964 | namespace llvm { | ||||||||
965 | |||||||||
966 | void reportVectorizationFailure(const StringRef DebugMsg, | ||||||||
967 | const StringRef OREMsg, const StringRef ORETag, | ||||||||
968 | OptimizationRemarkEmitter *ORE, Loop *TheLoop, Instruction *I) { | ||||||||
969 | LLVM_DEBUG(debugVectorizationFailure(DebugMsg, I))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { debugVectorizationFailure(DebugMsg, I); } } while (false); | ||||||||
970 | LoopVectorizeHints Hints(TheLoop, true /* doesn't matter */, *ORE); | ||||||||
971 | ORE->emit(createLVAnalysis(Hints.vectorizeAnalysisPassName(), | ||||||||
972 | ORETag, TheLoop, I) << OREMsg); | ||||||||
973 | } | ||||||||
974 | |||||||||
975 | } // end namespace llvm | ||||||||
976 | |||||||||
977 | #ifndef NDEBUG | ||||||||
978 | /// \return string containing a file name and a line # for the given loop. | ||||||||
979 | static std::string getDebugLocString(const Loop *L) { | ||||||||
980 | std::string Result; | ||||||||
981 | if (L) { | ||||||||
982 | raw_string_ostream OS(Result); | ||||||||
983 | if (const DebugLoc LoopDbgLoc = L->getStartLoc()) | ||||||||
984 | LoopDbgLoc.print(OS); | ||||||||
985 | else | ||||||||
986 | // Just print the module name. | ||||||||
987 | OS << L->getHeader()->getParent()->getParent()->getModuleIdentifier(); | ||||||||
988 | OS.flush(); | ||||||||
989 | } | ||||||||
990 | return Result; | ||||||||
991 | } | ||||||||
992 | #endif | ||||||||
993 | |||||||||
994 | void InnerLoopVectorizer::addNewMetadata(Instruction *To, | ||||||||
995 | const Instruction *Orig) { | ||||||||
996 | // If the loop was versioned with memchecks, add the corresponding no-alias | ||||||||
997 | // metadata. | ||||||||
998 | if (LVer && (isa<LoadInst>(Orig) || isa<StoreInst>(Orig))) | ||||||||
999 | LVer->annotateInstWithNoAlias(To, Orig); | ||||||||
1000 | } | ||||||||
1001 | |||||||||
1002 | void InnerLoopVectorizer::addMetadata(Instruction *To, | ||||||||
1003 | Instruction *From) { | ||||||||
1004 | propagateMetadata(To, From); | ||||||||
1005 | addNewMetadata(To, From); | ||||||||
1006 | } | ||||||||
1007 | |||||||||
1008 | void InnerLoopVectorizer::addMetadata(ArrayRef<Value *> To, | ||||||||
1009 | Instruction *From) { | ||||||||
1010 | for (Value *V : To) { | ||||||||
1011 | if (Instruction *I = dyn_cast<Instruction>(V)) | ||||||||
1012 | addMetadata(I, From); | ||||||||
1013 | } | ||||||||
1014 | } | ||||||||
1015 | |||||||||
1016 | namespace llvm { | ||||||||
1017 | |||||||||
1018 | // Loop vectorization cost-model hints how the scalar epilogue loop should be | ||||||||
1019 | // lowered. | ||||||||
1020 | enum ScalarEpilogueLowering { | ||||||||
1021 | |||||||||
1022 | // The default: allowing scalar epilogues. | ||||||||
1023 | CM_ScalarEpilogueAllowed, | ||||||||
1024 | |||||||||
1025 | // Vectorization with OptForSize: don't allow epilogues. | ||||||||
1026 | CM_ScalarEpilogueNotAllowedOptSize, | ||||||||
1027 | |||||||||
1028 | // A special case of vectorisation with OptForSize: loops with a very small | ||||||||
1029 | // trip count are considered for vectorization under OptForSize, thereby | ||||||||
1030 | // making sure the cost of their loop body is dominant, free of runtime | ||||||||
1031 | // guards and scalar iteration overheads. | ||||||||
1032 | CM_ScalarEpilogueNotAllowedLowTripLoop, | ||||||||
1033 | |||||||||
1034 | // Loop hint predicate indicating an epilogue is undesired. | ||||||||
1035 | CM_ScalarEpilogueNotNeededUsePredicate | ||||||||
1036 | }; | ||||||||
1037 | |||||||||
1038 | /// LoopVectorizationCostModel - estimates the expected speedups due to | ||||||||
1039 | /// vectorization. | ||||||||
1040 | /// In many cases vectorization is not profitable. This can happen because of | ||||||||
1041 | /// a number of reasons. In this class we mainly attempt to predict the | ||||||||
1042 | /// expected speedup/slowdowns due to the supported instruction set. We use the | ||||||||
1043 | /// TargetTransformInfo to query the different backends for the cost of | ||||||||
1044 | /// different operations. | ||||||||
1045 | class LoopVectorizationCostModel { | ||||||||
1046 | public: | ||||||||
1047 | LoopVectorizationCostModel(ScalarEpilogueLowering SEL, Loop *L, | ||||||||
1048 | PredicatedScalarEvolution &PSE, LoopInfo *LI, | ||||||||
1049 | LoopVectorizationLegality *Legal, | ||||||||
1050 | const TargetTransformInfo &TTI, | ||||||||
1051 | const TargetLibraryInfo *TLI, DemandedBits *DB, | ||||||||
1052 | AssumptionCache *AC, | ||||||||
1053 | OptimizationRemarkEmitter *ORE, const Function *F, | ||||||||
1054 | const LoopVectorizeHints *Hints, | ||||||||
1055 | InterleavedAccessInfo &IAI) | ||||||||
1056 | : ScalarEpilogueStatus(SEL), TheLoop(L), PSE(PSE), LI(LI), Legal(Legal), | ||||||||
1057 | TTI(TTI), TLI(TLI), DB(DB), AC(AC), ORE(ORE), TheFunction(F), | ||||||||
1058 | Hints(Hints), InterleaveInfo(IAI) {} | ||||||||
1059 | |||||||||
1060 | /// \return An upper bound for the vectorization factor, or None if | ||||||||
1061 | /// vectorization and interleaving should be avoided up front. | ||||||||
1062 | Optional<unsigned> computeMaxVF(unsigned UserVF, unsigned UserIC); | ||||||||
1063 | |||||||||
1064 | /// \return True if runtime checks are required for vectorization, and false | ||||||||
1065 | /// otherwise. | ||||||||
1066 | bool runtimeChecksRequired(); | ||||||||
1067 | |||||||||
1068 | /// \return The most profitable vectorization factor and the cost of that VF. | ||||||||
1069 | /// This method checks every power of two up to MaxVF. If UserVF is not ZERO | ||||||||
1070 | /// then this vectorization factor will be selected if vectorization is | ||||||||
1071 | /// possible. | ||||||||
1072 | VectorizationFactor selectVectorizationFactor(unsigned MaxVF); | ||||||||
1073 | |||||||||
1074 | /// Setup cost-based decisions for user vectorization factor. | ||||||||
1075 | void selectUserVectorizationFactor(ElementCount UserVF) { | ||||||||
1076 | collectUniformsAndScalars(UserVF); | ||||||||
1077 | collectInstsToScalarize(UserVF); | ||||||||
1078 | } | ||||||||
1079 | |||||||||
1080 | /// \return The size (in bits) of the smallest and widest types in the code | ||||||||
1081 | /// that needs to be vectorized. We ignore values that remain scalar such as | ||||||||
1082 | /// 64 bit loop indices. | ||||||||
1083 | std::pair<unsigned, unsigned> getSmallestAndWidestTypes(); | ||||||||
1084 | |||||||||
1085 | /// \return The desired interleave count. | ||||||||
1086 | /// If interleave count has been specified by metadata it will be returned. | ||||||||
1087 | /// Otherwise, the interleave count is computed and returned. VF and LoopCost | ||||||||
1088 | /// are the selected vectorization factor and the cost of the selected VF. | ||||||||
1089 | unsigned selectInterleaveCount(ElementCount VF, unsigned LoopCost); | ||||||||
1090 | |||||||||
1091 | /// Memory access instruction may be vectorized in more than one way. | ||||||||
1092 | /// Form of instruction after vectorization depends on cost. | ||||||||
1093 | /// This function takes cost-based decisions for Load/Store instructions | ||||||||
1094 | /// and collects them in a map. This decisions map is used for building | ||||||||
1095 | /// the lists of loop-uniform and loop-scalar instructions. | ||||||||
1096 | /// The calculated cost is saved with widening decision in order to | ||||||||
1097 | /// avoid redundant calculations. | ||||||||
1098 | void setCostBasedWideningDecision(ElementCount VF); | ||||||||
1099 | |||||||||
1100 | /// A struct that represents some properties of the register usage | ||||||||
1101 | /// of a loop. | ||||||||
1102 | struct RegisterUsage { | ||||||||
1103 | /// Holds the number of loop invariant values that are used in the loop. | ||||||||
1104 | /// The key is ClassID of target-provided register class. | ||||||||
1105 | SmallMapVector<unsigned, unsigned, 4> LoopInvariantRegs; | ||||||||
1106 | /// Holds the maximum number of concurrent live intervals in the loop. | ||||||||
1107 | /// The key is ClassID of target-provided register class. | ||||||||
1108 | SmallMapVector<unsigned, unsigned, 4> MaxLocalUsers; | ||||||||
1109 | }; | ||||||||
1110 | |||||||||
1111 | /// \return Returns information about the register usages of the loop for the | ||||||||
1112 | /// given vectorization factors. | ||||||||
1113 | SmallVector<RegisterUsage, 8> | ||||||||
1114 | calculateRegisterUsage(ArrayRef<ElementCount> VFs); | ||||||||
1115 | |||||||||
1116 | /// Collect values we want to ignore in the cost model. | ||||||||
1117 | void collectValuesToIgnore(); | ||||||||
1118 | |||||||||
1119 | /// Split reductions into those that happen in the loop, and those that happen | ||||||||
1120 | /// outside. In loop reductions are collected into InLoopReductionChains. | ||||||||
1121 | void collectInLoopReductions(); | ||||||||
1122 | |||||||||
1123 | /// \returns The smallest bitwidth each instruction can be represented with. | ||||||||
1124 | /// The vector equivalents of these instructions should be truncated to this | ||||||||
1125 | /// type. | ||||||||
1126 | const MapVector<Instruction *, uint64_t> &getMinimalBitwidths() const { | ||||||||
1127 | return MinBWs; | ||||||||
1128 | } | ||||||||
1129 | |||||||||
1130 | /// \returns True if it is more profitable to scalarize instruction \p I for | ||||||||
1131 | /// vectorization factor \p VF. | ||||||||
1132 | bool isProfitableToScalarize(Instruction *I, ElementCount VF) const { | ||||||||
1133 | assert(VF.isVector() &&((VF.isVector() && "Profitable to scalarize relevant only for VF > 1." ) ? static_cast<void> (0) : __assert_fail ("VF.isVector() && \"Profitable to scalarize relevant only for VF > 1.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 1134, __PRETTY_FUNCTION__)) | ||||||||
1134 | "Profitable to scalarize relevant only for VF > 1.")((VF.isVector() && "Profitable to scalarize relevant only for VF > 1." ) ? static_cast<void> (0) : __assert_fail ("VF.isVector() && \"Profitable to scalarize relevant only for VF > 1.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 1134, __PRETTY_FUNCTION__)); | ||||||||
1135 | |||||||||
1136 | // Cost model is not run in the VPlan-native path - return conservative | ||||||||
1137 | // result until this changes. | ||||||||
1138 | if (EnableVPlanNativePath) | ||||||||
1139 | return false; | ||||||||
1140 | |||||||||
1141 | auto Scalars = InstsToScalarize.find(VF); | ||||||||
1142 | assert(Scalars != InstsToScalarize.end() &&((Scalars != InstsToScalarize.end() && "VF not yet analyzed for scalarization profitability" ) ? static_cast<void> (0) : __assert_fail ("Scalars != InstsToScalarize.end() && \"VF not yet analyzed for scalarization profitability\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 1143, __PRETTY_FUNCTION__)) | ||||||||
1143 | "VF not yet analyzed for scalarization profitability")((Scalars != InstsToScalarize.end() && "VF not yet analyzed for scalarization profitability" ) ? static_cast<void> (0) : __assert_fail ("Scalars != InstsToScalarize.end() && \"VF not yet analyzed for scalarization profitability\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 1143, __PRETTY_FUNCTION__)); | ||||||||
1144 | return Scalars->second.find(I) != Scalars->second.end(); | ||||||||
1145 | } | ||||||||
1146 | |||||||||
1147 | /// Returns true if \p I is known to be uniform after vectorization. | ||||||||
1148 | bool isUniformAfterVectorization(Instruction *I, ElementCount VF) const { | ||||||||
1149 | if (VF.isScalar()) | ||||||||
1150 | return true; | ||||||||
1151 | |||||||||
1152 | // Cost model is not run in the VPlan-native path - return conservative | ||||||||
1153 | // result until this changes. | ||||||||
1154 | if (EnableVPlanNativePath) | ||||||||
1155 | return false; | ||||||||
1156 | |||||||||
1157 | auto UniformsPerVF = Uniforms.find(VF); | ||||||||
1158 | assert(UniformsPerVF != Uniforms.end() &&((UniformsPerVF != Uniforms.end() && "VF not yet analyzed for uniformity" ) ? static_cast<void> (0) : __assert_fail ("UniformsPerVF != Uniforms.end() && \"VF not yet analyzed for uniformity\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 1159, __PRETTY_FUNCTION__)) | ||||||||
1159 | "VF not yet analyzed for uniformity")((UniformsPerVF != Uniforms.end() && "VF not yet analyzed for uniformity" ) ? static_cast<void> (0) : __assert_fail ("UniformsPerVF != Uniforms.end() && \"VF not yet analyzed for uniformity\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 1159, __PRETTY_FUNCTION__)); | ||||||||
1160 | return UniformsPerVF->second.count(I); | ||||||||
1161 | } | ||||||||
1162 | |||||||||
1163 | /// Returns true if \p I is known to be scalar after vectorization. | ||||||||
1164 | bool isScalarAfterVectorization(Instruction *I, ElementCount VF) const { | ||||||||
1165 | if (VF.isScalar()) | ||||||||
1166 | return true; | ||||||||
1167 | |||||||||
1168 | // Cost model is not run in the VPlan-native path - return conservative | ||||||||
1169 | // result until this changes. | ||||||||
1170 | if (EnableVPlanNativePath) | ||||||||
1171 | return false; | ||||||||
1172 | |||||||||
1173 | auto ScalarsPerVF = Scalars.find(VF); | ||||||||
1174 | assert(ScalarsPerVF != Scalars.end() &&((ScalarsPerVF != Scalars.end() && "Scalar values are not calculated for VF" ) ? static_cast<void> (0) : __assert_fail ("ScalarsPerVF != Scalars.end() && \"Scalar values are not calculated for VF\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 1175, __PRETTY_FUNCTION__)) | ||||||||
1175 | "Scalar values are not calculated for VF")((ScalarsPerVF != Scalars.end() && "Scalar values are not calculated for VF" ) ? static_cast<void> (0) : __assert_fail ("ScalarsPerVF != Scalars.end() && \"Scalar values are not calculated for VF\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 1175, __PRETTY_FUNCTION__)); | ||||||||
1176 | return ScalarsPerVF->second.count(I); | ||||||||
1177 | } | ||||||||
1178 | |||||||||
1179 | /// \returns True if instruction \p I can be truncated to a smaller bitwidth | ||||||||
1180 | /// for vectorization factor \p VF. | ||||||||
1181 | bool canTruncateToMinimalBitwidth(Instruction *I, ElementCount VF) const { | ||||||||
1182 | return VF.isVector() && MinBWs.find(I) != MinBWs.end() && | ||||||||
1183 | !isProfitableToScalarize(I, VF) && | ||||||||
1184 | !isScalarAfterVectorization(I, VF); | ||||||||
1185 | } | ||||||||
1186 | |||||||||
1187 | /// Decision that was taken during cost calculation for memory instruction. | ||||||||
1188 | enum InstWidening { | ||||||||
1189 | CM_Unknown, | ||||||||
1190 | CM_Widen, // For consecutive accesses with stride +1. | ||||||||
1191 | CM_Widen_Reverse, // For consecutive accesses with stride -1. | ||||||||
1192 | CM_Interleave, | ||||||||
1193 | CM_GatherScatter, | ||||||||
1194 | CM_Scalarize | ||||||||
1195 | }; | ||||||||
1196 | |||||||||
1197 | /// Save vectorization decision \p W and \p Cost taken by the cost model for | ||||||||
1198 | /// instruction \p I and vector width \p VF. | ||||||||
1199 | void setWideningDecision(Instruction *I, ElementCount VF, InstWidening W, | ||||||||
1200 | unsigned Cost) { | ||||||||
1201 | assert(VF.isVector() && "Expected VF >=2")((VF.isVector() && "Expected VF >=2") ? static_cast <void> (0) : __assert_fail ("VF.isVector() && \"Expected VF >=2\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 1201, __PRETTY_FUNCTION__)); | ||||||||
1202 | WideningDecisions[std::make_pair(I, VF)] = std::make_pair(W, Cost); | ||||||||
1203 | } | ||||||||
1204 | |||||||||
1205 | /// Save vectorization decision \p W and \p Cost taken by the cost model for | ||||||||
1206 | /// interleaving group \p Grp and vector width \p VF. | ||||||||
1207 | void setWideningDecision(const InterleaveGroup<Instruction> *Grp, | ||||||||
1208 | ElementCount VF, InstWidening W, unsigned Cost) { | ||||||||
1209 | assert(VF.isVector() && "Expected VF >=2")((VF.isVector() && "Expected VF >=2") ? static_cast <void> (0) : __assert_fail ("VF.isVector() && \"Expected VF >=2\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 1209, __PRETTY_FUNCTION__)); | ||||||||
1210 | /// Broadcast this decicion to all instructions inside the group. | ||||||||
1211 | /// But the cost will be assigned to one instruction only. | ||||||||
1212 | for (unsigned i = 0; i < Grp->getFactor(); ++i) { | ||||||||
1213 | if (auto *I = Grp->getMember(i)) { | ||||||||
1214 | if (Grp->getInsertPos() == I) | ||||||||
1215 | WideningDecisions[std::make_pair(I, VF)] = std::make_pair(W, Cost); | ||||||||
1216 | else | ||||||||
1217 | WideningDecisions[std::make_pair(I, VF)] = std::make_pair(W, 0); | ||||||||
1218 | } | ||||||||
1219 | } | ||||||||
1220 | } | ||||||||
1221 | |||||||||
1222 | /// Return the cost model decision for the given instruction \p I and vector | ||||||||
1223 | /// width \p VF. Return CM_Unknown if this instruction did not pass | ||||||||
1224 | /// through the cost modeling. | ||||||||
1225 | InstWidening getWideningDecision(Instruction *I, ElementCount VF) { | ||||||||
1226 | assert(!VF.isScalable() && "scalable vectors not yet supported.")((!VF.isScalable() && "scalable vectors not yet supported." ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"scalable vectors not yet supported.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 1226, __PRETTY_FUNCTION__)); | ||||||||
1227 | assert(VF.isVector() && "Expected VF >=2")((VF.isVector() && "Expected VF >=2") ? static_cast <void> (0) : __assert_fail ("VF.isVector() && \"Expected VF >=2\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 1227, __PRETTY_FUNCTION__)); | ||||||||
1228 | |||||||||
1229 | // Cost model is not run in the VPlan-native path - return conservative | ||||||||
1230 | // result until this changes. | ||||||||
1231 | if (EnableVPlanNativePath) | ||||||||
1232 | return CM_GatherScatter; | ||||||||
1233 | |||||||||
1234 | std::pair<Instruction *, ElementCount> InstOnVF = std::make_pair(I, VF); | ||||||||
1235 | auto Itr = WideningDecisions.find(InstOnVF); | ||||||||
1236 | if (Itr == WideningDecisions.end()) | ||||||||
1237 | return CM_Unknown; | ||||||||
1238 | return Itr->second.first; | ||||||||
1239 | } | ||||||||
1240 | |||||||||
1241 | /// Return the vectorization cost for the given instruction \p I and vector | ||||||||
1242 | /// width \p VF. | ||||||||
1243 | unsigned getWideningCost(Instruction *I, ElementCount VF) { | ||||||||
1244 | assert(VF.isVector() && "Expected VF >=2")((VF.isVector() && "Expected VF >=2") ? static_cast <void> (0) : __assert_fail ("VF.isVector() && \"Expected VF >=2\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 1244, __PRETTY_FUNCTION__)); | ||||||||
1245 | std::pair<Instruction *, ElementCount> InstOnVF = std::make_pair(I, VF); | ||||||||
1246 | assert(WideningDecisions.find(InstOnVF) != WideningDecisions.end() &&((WideningDecisions.find(InstOnVF) != WideningDecisions.end() && "The cost is not calculated") ? static_cast<void > (0) : __assert_fail ("WideningDecisions.find(InstOnVF) != WideningDecisions.end() && \"The cost is not calculated\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 1247, __PRETTY_FUNCTION__)) | ||||||||
1247 | "The cost is not calculated")((WideningDecisions.find(InstOnVF) != WideningDecisions.end() && "The cost is not calculated") ? static_cast<void > (0) : __assert_fail ("WideningDecisions.find(InstOnVF) != WideningDecisions.end() && \"The cost is not calculated\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 1247, __PRETTY_FUNCTION__)); | ||||||||
1248 | return WideningDecisions[InstOnVF].second; | ||||||||
1249 | } | ||||||||
1250 | |||||||||
1251 | /// Return True if instruction \p I is an optimizable truncate whose operand | ||||||||
1252 | /// is an induction variable. Such a truncate will be removed by adding a new | ||||||||
1253 | /// induction variable with the destination type. | ||||||||
1254 | bool isOptimizableIVTruncate(Instruction *I, ElementCount VF) { | ||||||||
1255 | // If the instruction is not a truncate, return false. | ||||||||
1256 | auto *Trunc = dyn_cast<TruncInst>(I); | ||||||||
1257 | if (!Trunc) | ||||||||
1258 | return false; | ||||||||
1259 | |||||||||
1260 | // Get the source and destination types of the truncate. | ||||||||
1261 | Type *SrcTy = ToVectorTy(cast<CastInst>(I)->getSrcTy(), VF); | ||||||||
1262 | Type *DestTy = ToVectorTy(cast<CastInst>(I)->getDestTy(), VF); | ||||||||
1263 | |||||||||
1264 | // If the truncate is free for the given types, return false. Replacing a | ||||||||
1265 | // free truncate with an induction variable would add an induction variable | ||||||||
1266 | // update instruction to each iteration of the loop. We exclude from this | ||||||||
1267 | // check the primary induction variable since it will need an update | ||||||||
1268 | // instruction regardless. | ||||||||
1269 | Value *Op = Trunc->getOperand(0); | ||||||||
1270 | if (Op != Legal->getPrimaryInduction() && TTI.isTruncateFree(SrcTy, DestTy)) | ||||||||
1271 | return false; | ||||||||
1272 | |||||||||
1273 | // If the truncated value is not an induction variable, return false. | ||||||||
1274 | return Legal->isInductionPhi(Op); | ||||||||
1275 | } | ||||||||
1276 | |||||||||
1277 | /// Collects the instructions to scalarize for each predicated instruction in | ||||||||
1278 | /// the loop. | ||||||||
1279 | void collectInstsToScalarize(ElementCount VF); | ||||||||
1280 | |||||||||
1281 | /// Collect Uniform and Scalar values for the given \p VF. | ||||||||
1282 | /// The sets depend on CM decision for Load/Store instructions | ||||||||
1283 | /// that may be vectorized as interleave, gather-scatter or scalarized. | ||||||||
1284 | void collectUniformsAndScalars(ElementCount VF) { | ||||||||
1285 | // Do the analysis once. | ||||||||
1286 | if (VF.isScalar() || Uniforms.find(VF) != Uniforms.end()) | ||||||||
1287 | return; | ||||||||
1288 | setCostBasedWideningDecision(VF); | ||||||||
1289 | collectLoopUniforms(VF); | ||||||||
1290 | collectLoopScalars(VF); | ||||||||
1291 | } | ||||||||
1292 | |||||||||
1293 | /// Returns true if the target machine supports masked store operation | ||||||||
1294 | /// for the given \p DataType and kind of access to \p Ptr. | ||||||||
1295 | bool isLegalMaskedStore(Type *DataType, Value *Ptr, Align Alignment) { | ||||||||
1296 | return Legal->isConsecutivePtr(Ptr) && | ||||||||
1297 | TTI.isLegalMaskedStore(DataType, Alignment); | ||||||||
1298 | } | ||||||||
1299 | |||||||||
1300 | /// Returns true if the target machine supports masked load operation | ||||||||
1301 | /// for the given \p DataType and kind of access to \p Ptr. | ||||||||
1302 | bool isLegalMaskedLoad(Type *DataType, Value *Ptr, Align Alignment) { | ||||||||
1303 | return Legal->isConsecutivePtr(Ptr) && | ||||||||
1304 | TTI.isLegalMaskedLoad(DataType, Alignment); | ||||||||
1305 | } | ||||||||
1306 | |||||||||
1307 | /// Returns true if the target machine supports masked scatter operation | ||||||||
1308 | /// for the given \p DataType. | ||||||||
1309 | bool isLegalMaskedScatter(Type *DataType, Align Alignment) { | ||||||||
1310 | return TTI.isLegalMaskedScatter(DataType, Alignment); | ||||||||
1311 | } | ||||||||
1312 | |||||||||
1313 | /// Returns true if the target machine supports masked gather operation | ||||||||
1314 | /// for the given \p DataType. | ||||||||
1315 | bool isLegalMaskedGather(Type *DataType, Align Alignment) { | ||||||||
1316 | return TTI.isLegalMaskedGather(DataType, Alignment); | ||||||||
1317 | } | ||||||||
1318 | |||||||||
1319 | /// Returns true if the target machine can represent \p V as a masked gather | ||||||||
1320 | /// or scatter operation. | ||||||||
1321 | bool isLegalGatherOrScatter(Value *V) { | ||||||||
1322 | bool LI = isa<LoadInst>(V); | ||||||||
1323 | bool SI = isa<StoreInst>(V); | ||||||||
1324 | if (!LI && !SI) | ||||||||
1325 | return false; | ||||||||
1326 | auto *Ty = getMemInstValueType(V); | ||||||||
1327 | Align Align = getLoadStoreAlignment(V); | ||||||||
1328 | return (LI && isLegalMaskedGather(Ty, Align)) || | ||||||||
1329 | (SI && isLegalMaskedScatter(Ty, Align)); | ||||||||
1330 | } | ||||||||
1331 | |||||||||
1332 | /// Returns true if \p I is an instruction that will be scalarized with | ||||||||
1333 | /// predication. Such instructions include conditional stores and | ||||||||
1334 | /// instructions that may divide by zero. | ||||||||
1335 | /// If a non-zero VF has been calculated, we check if I will be scalarized | ||||||||
1336 | /// predication for that VF. | ||||||||
1337 | bool isScalarWithPredication(Instruction *I, | ||||||||
1338 | ElementCount VF = ElementCount::getFixed(1)); | ||||||||
1339 | |||||||||
1340 | // Returns true if \p I is an instruction that will be predicated either | ||||||||
1341 | // through scalar predication or masked load/store or masked gather/scatter. | ||||||||
1342 | // Superset of instructions that return true for isScalarWithPredication. | ||||||||
1343 | bool isPredicatedInst(Instruction *I) { | ||||||||
1344 | if (!blockNeedsPredication(I->getParent())) | ||||||||
1345 | return false; | ||||||||
1346 | // Loads and stores that need some form of masked operation are predicated | ||||||||
1347 | // instructions. | ||||||||
1348 | if (isa<LoadInst>(I) || isa<StoreInst>(I)) | ||||||||
1349 | return Legal->isMaskRequired(I); | ||||||||
1350 | return isScalarWithPredication(I); | ||||||||
1351 | } | ||||||||
1352 | |||||||||
1353 | /// Returns true if \p I is a memory instruction with consecutive memory | ||||||||
1354 | /// access that can be widened. | ||||||||
1355 | bool | ||||||||
1356 | memoryInstructionCanBeWidened(Instruction *I, | ||||||||
1357 | ElementCount VF = ElementCount::getFixed(1)); | ||||||||
1358 | |||||||||
1359 | /// Returns true if \p I is a memory instruction in an interleaved-group | ||||||||
1360 | /// of memory accesses that can be vectorized with wide vector loads/stores | ||||||||
1361 | /// and shuffles. | ||||||||
1362 | bool | ||||||||
1363 | interleavedAccessCanBeWidened(Instruction *I, | ||||||||
1364 | ElementCount VF = ElementCount::getFixed(1)); | ||||||||
1365 | |||||||||
1366 | /// Check if \p Instr belongs to any interleaved access group. | ||||||||
1367 | bool isAccessInterleaved(Instruction *Instr) { | ||||||||
1368 | return InterleaveInfo.isInterleaved(Instr); | ||||||||
1369 | } | ||||||||
1370 | |||||||||
1371 | /// Get the interleaved access group that \p Instr belongs to. | ||||||||
1372 | const InterleaveGroup<Instruction> * | ||||||||
1373 | getInterleavedAccessGroup(Instruction *Instr) { | ||||||||
1374 | return InterleaveInfo.getInterleaveGroup(Instr); | ||||||||
1375 | } | ||||||||
1376 | |||||||||
1377 | /// Returns true if an interleaved group requires a scalar iteration | ||||||||
1378 | /// to handle accesses with gaps, and there is nothing preventing us from | ||||||||
1379 | /// creating a scalar epilogue. | ||||||||
1380 | bool requiresScalarEpilogue() const { | ||||||||
1381 | return isScalarEpilogueAllowed() && InterleaveInfo.requiresScalarEpilogue(); | ||||||||
1382 | } | ||||||||
1383 | |||||||||
1384 | /// Returns true if a scalar epilogue is not allowed due to optsize or a | ||||||||
1385 | /// loop hint annotation. | ||||||||
1386 | bool isScalarEpilogueAllowed() const { | ||||||||
1387 | return ScalarEpilogueStatus == CM_ScalarEpilogueAllowed; | ||||||||
1388 | } | ||||||||
1389 | |||||||||
1390 | /// Returns true if all loop blocks should be masked to fold tail loop. | ||||||||
1391 | bool foldTailByMasking() const { return FoldTailByMasking; } | ||||||||
1392 | |||||||||
1393 | bool blockNeedsPredication(BasicBlock *BB) { | ||||||||
1394 | return foldTailByMasking() || Legal->blockNeedsPredication(BB); | ||||||||
1395 | } | ||||||||
1396 | |||||||||
1397 | /// A SmallMapVector to store the InLoop reduction op chains, mapping phi | ||||||||
1398 | /// nodes to the chain of instructions representing the reductions. Uses a | ||||||||
1399 | /// MapVector to ensure deterministic iteration order. | ||||||||
1400 | using ReductionChainMap = | ||||||||
1401 | SmallMapVector<PHINode *, SmallVector<Instruction *, 4>, 4>; | ||||||||
1402 | |||||||||
1403 | /// Return the chain of instructions representing an inloop reduction. | ||||||||
1404 | const ReductionChainMap &getInLoopReductionChains() const { | ||||||||
1405 | return InLoopReductionChains; | ||||||||
1406 | } | ||||||||
1407 | |||||||||
1408 | /// Returns true if the Phi is part of an inloop reduction. | ||||||||
1409 | bool isInLoopReduction(PHINode *Phi) const { | ||||||||
1410 | return InLoopReductionChains.count(Phi); | ||||||||
1411 | } | ||||||||
1412 | |||||||||
1413 | /// Estimate cost of an intrinsic call instruction CI if it were vectorized | ||||||||
1414 | /// with factor VF. Return the cost of the instruction, including | ||||||||
1415 | /// scalarization overhead if it's needed. | ||||||||
1416 | unsigned getVectorIntrinsicCost(CallInst *CI, ElementCount VF); | ||||||||
1417 | |||||||||
1418 | /// Estimate cost of a call instruction CI if it were vectorized with factor | ||||||||
1419 | /// VF. Return the cost of the instruction, including scalarization overhead | ||||||||
1420 | /// if it's needed. The flag NeedToScalarize shows if the call needs to be | ||||||||
1421 | /// scalarized - | ||||||||
1422 | /// i.e. either vector version isn't available, or is too expensive. | ||||||||
1423 | unsigned getVectorCallCost(CallInst *CI, ElementCount VF, | ||||||||
1424 | bool &NeedToScalarize); | ||||||||
1425 | |||||||||
1426 | /// Invalidates decisions already taken by the cost model. | ||||||||
1427 | void invalidateCostModelingDecisions() { | ||||||||
1428 | WideningDecisions.clear(); | ||||||||
1429 | Uniforms.clear(); | ||||||||
1430 | Scalars.clear(); | ||||||||
1431 | } | ||||||||
1432 | |||||||||
1433 | private: | ||||||||
1434 | unsigned NumPredStores = 0; | ||||||||
1435 | |||||||||
1436 | /// \return An upper bound for the vectorization factor, a power-of-2 larger | ||||||||
1437 | /// than zero. One is returned if vectorization should best be avoided due | ||||||||
1438 | /// to cost. | ||||||||
1439 | unsigned computeFeasibleMaxVF(unsigned ConstTripCount); | ||||||||
1440 | |||||||||
1441 | /// The vectorization cost is a combination of the cost itself and a boolean | ||||||||
1442 | /// indicating whether any of the contributing operations will actually | ||||||||
1443 | /// operate on | ||||||||
1444 | /// vector values after type legalization in the backend. If this latter value | ||||||||
1445 | /// is | ||||||||
1446 | /// false, then all operations will be scalarized (i.e. no vectorization has | ||||||||
1447 | /// actually taken place). | ||||||||
1448 | using VectorizationCostTy = std::pair<unsigned, bool>; | ||||||||
1449 | |||||||||
1450 | /// Returns the expected execution cost. The unit of the cost does | ||||||||
1451 | /// not matter because we use the 'cost' units to compare different | ||||||||
1452 | /// vector widths. The cost that is returned is *not* normalized by | ||||||||
1453 | /// the factor width. | ||||||||
1454 | VectorizationCostTy expectedCost(ElementCount VF); | ||||||||
1455 | |||||||||
1456 | /// Returns the execution time cost of an instruction for a given vector | ||||||||
1457 | /// width. Vector width of one means scalar. | ||||||||
1458 | VectorizationCostTy getInstructionCost(Instruction *I, ElementCount VF); | ||||||||
1459 | |||||||||
1460 | /// The cost-computation logic from getInstructionCost which provides | ||||||||
1461 | /// the vector type as an output parameter. | ||||||||
1462 | unsigned getInstructionCost(Instruction *I, ElementCount VF, Type *&VectorTy); | ||||||||
1463 | |||||||||
1464 | /// Calculate vectorization cost of memory instruction \p I. | ||||||||
1465 | unsigned getMemoryInstructionCost(Instruction *I, ElementCount VF); | ||||||||
1466 | |||||||||
1467 | /// The cost computation for scalarized memory instruction. | ||||||||
1468 | unsigned getMemInstScalarizationCost(Instruction *I, ElementCount VF); | ||||||||
1469 | |||||||||
1470 | /// The cost computation for interleaving group of memory instructions. | ||||||||
1471 | unsigned getInterleaveGroupCost(Instruction *I, ElementCount VF); | ||||||||
1472 | |||||||||
1473 | /// The cost computation for Gather/Scatter instruction. | ||||||||
1474 | unsigned getGatherScatterCost(Instruction *I, ElementCount VF); | ||||||||
1475 | |||||||||
1476 | /// The cost computation for widening instruction \p I with consecutive | ||||||||
1477 | /// memory access. | ||||||||
1478 | unsigned getConsecutiveMemOpCost(Instruction *I, ElementCount VF); | ||||||||
1479 | |||||||||
1480 | /// The cost calculation for Load/Store instruction \p I with uniform pointer - | ||||||||
1481 | /// Load: scalar load + broadcast. | ||||||||
1482 | /// Store: scalar store + (loop invariant value stored? 0 : extract of last | ||||||||
1483 | /// element) | ||||||||
1484 | unsigned getUniformMemOpCost(Instruction *I, ElementCount VF); | ||||||||
1485 | |||||||||
1486 | /// Estimate the overhead of scalarizing an instruction. This is a | ||||||||
1487 | /// convenience wrapper for the type-based getScalarizationOverhead API. | ||||||||
1488 | unsigned getScalarizationOverhead(Instruction *I, ElementCount VF); | ||||||||
1489 | |||||||||
1490 | /// Returns whether the instruction is a load or store and will be a emitted | ||||||||
1491 | /// as a vector operation. | ||||||||
1492 | bool isConsecutiveLoadOrStore(Instruction *I); | ||||||||
1493 | |||||||||
1494 | /// Returns true if an artificially high cost for emulated masked memrefs | ||||||||
1495 | /// should be used. | ||||||||
1496 | bool useEmulatedMaskMemRefHack(Instruction *I); | ||||||||
1497 | |||||||||
1498 | /// Map of scalar integer values to the smallest bitwidth they can be legally | ||||||||
1499 | /// represented as. The vector equivalents of these values should be truncated | ||||||||
1500 | /// to this type. | ||||||||
1501 | MapVector<Instruction *, uint64_t> MinBWs; | ||||||||
1502 | |||||||||
1503 | /// A type representing the costs for instructions if they were to be | ||||||||
1504 | /// scalarized rather than vectorized. The entries are Instruction-Cost | ||||||||
1505 | /// pairs. | ||||||||
1506 | using ScalarCostsTy = DenseMap<Instruction *, unsigned>; | ||||||||
1507 | |||||||||
1508 | /// A set containing all BasicBlocks that are known to present after | ||||||||
1509 | /// vectorization as a predicated block. | ||||||||
1510 | SmallPtrSet<BasicBlock *, 4> PredicatedBBsAfterVectorization; | ||||||||
1511 | |||||||||
1512 | /// Records whether it is allowed to have the original scalar loop execute at | ||||||||
1513 | /// least once. This may be needed as a fallback loop in case runtime | ||||||||
1514 | /// aliasing/dependence checks fail, or to handle the tail/remainder | ||||||||
1515 | /// iterations when the trip count is unknown or doesn't divide by the VF, | ||||||||
1516 | /// or as a peel-loop to handle gaps in interleave-groups. | ||||||||
1517 | /// Under optsize and when the trip count is very small we don't allow any | ||||||||
1518 | /// iterations to execute in the scalar loop. | ||||||||
1519 | ScalarEpilogueLowering ScalarEpilogueStatus = CM_ScalarEpilogueAllowed; | ||||||||
1520 | |||||||||
1521 | /// All blocks of loop are to be masked to fold tail of scalar iterations. | ||||||||
1522 | bool FoldTailByMasking = false; | ||||||||
1523 | |||||||||
1524 | /// A map holding scalar costs for different vectorization factors. The | ||||||||
1525 | /// presence of a cost for an instruction in the mapping indicates that the | ||||||||
1526 | /// instruction will be scalarized when vectorizing with the associated | ||||||||
1527 | /// vectorization factor. The entries are VF-ScalarCostTy pairs. | ||||||||
1528 | DenseMap<ElementCount, ScalarCostsTy> InstsToScalarize; | ||||||||
1529 | |||||||||
1530 | /// Holds the instructions known to be uniform after vectorization. | ||||||||
1531 | /// The data is collected per VF. | ||||||||
1532 | DenseMap<ElementCount, SmallPtrSet<Instruction *, 4>> Uniforms; | ||||||||
1533 | |||||||||
1534 | /// Holds the instructions known to be scalar after vectorization. | ||||||||
1535 | /// The data is collected per VF. | ||||||||
1536 | DenseMap<ElementCount, SmallPtrSet<Instruction *, 4>> Scalars; | ||||||||
1537 | |||||||||
1538 | /// Holds the instructions (address computations) that are forced to be | ||||||||
1539 | /// scalarized. | ||||||||
1540 | DenseMap<ElementCount, SmallPtrSet<Instruction *, 4>> ForcedScalars; | ||||||||
1541 | |||||||||
1542 | /// PHINodes of the reductions that should be expanded in-loop along with | ||||||||
1543 | /// their associated chains of reduction operations, in program order from top | ||||||||
1544 | /// (PHI) to bottom | ||||||||
1545 | ReductionChainMap InLoopReductionChains; | ||||||||
1546 | |||||||||
1547 | /// Returns the expected difference in cost from scalarizing the expression | ||||||||
1548 | /// feeding a predicated instruction \p PredInst. The instructions to | ||||||||
1549 | /// scalarize and their scalar costs are collected in \p ScalarCosts. A | ||||||||
1550 | /// non-negative return value implies the expression will be scalarized. | ||||||||
1551 | /// Currently, only single-use chains are considered for scalarization. | ||||||||
1552 | int computePredInstDiscount(Instruction *PredInst, ScalarCostsTy &ScalarCosts, | ||||||||
1553 | ElementCount VF); | ||||||||
1554 | |||||||||
1555 | /// Collect the instructions that are uniform after vectorization. An | ||||||||
1556 | /// instruction is uniform if we represent it with a single scalar value in | ||||||||
1557 | /// the vectorized loop corresponding to each vector iteration. Examples of | ||||||||
1558 | /// uniform instructions include pointer operands of consecutive or | ||||||||
1559 | /// interleaved memory accesses. Note that although uniformity implies an | ||||||||
1560 | /// instruction will be scalar, the reverse is not true. In general, a | ||||||||
1561 | /// scalarized instruction will be represented by VF scalar values in the | ||||||||
1562 | /// vectorized loop, each corresponding to an iteration of the original | ||||||||
1563 | /// scalar loop. | ||||||||
1564 | void collectLoopUniforms(ElementCount VF); | ||||||||
1565 | |||||||||
1566 | /// Collect the instructions that are scalar after vectorization. An | ||||||||
1567 | /// instruction is scalar if it is known to be uniform or will be scalarized | ||||||||
1568 | /// during vectorization. Non-uniform scalarized instructions will be | ||||||||
1569 | /// represented by VF values in the vectorized loop, each corresponding to an | ||||||||
1570 | /// iteration of the original scalar loop. | ||||||||
1571 | void collectLoopScalars(ElementCount VF); | ||||||||
1572 | |||||||||
1573 | /// Keeps cost model vectorization decision and cost for instructions. | ||||||||
1574 | /// Right now it is used for memory instructions only. | ||||||||
1575 | using DecisionList = DenseMap<std::pair<Instruction *, ElementCount>, | ||||||||
1576 | std::pair<InstWidening, unsigned>>; | ||||||||
1577 | |||||||||
1578 | DecisionList WideningDecisions; | ||||||||
1579 | |||||||||
1580 | /// Returns true if \p V is expected to be vectorized and it needs to be | ||||||||
1581 | /// extracted. | ||||||||
1582 | bool needsExtract(Value *V, ElementCount VF) const { | ||||||||
1583 | Instruction *I = dyn_cast<Instruction>(V); | ||||||||
1584 | if (VF.isScalar() || !I || !TheLoop->contains(I) || | ||||||||
1585 | TheLoop->isLoopInvariant(I)) | ||||||||
1586 | return false; | ||||||||
1587 | |||||||||
1588 | // Assume we can vectorize V (and hence we need extraction) if the | ||||||||
1589 | // scalars are not computed yet. This can happen, because it is called | ||||||||
1590 | // via getScalarizationOverhead from setCostBasedWideningDecision, before | ||||||||
1591 | // the scalars are collected. That should be a safe assumption in most | ||||||||
1592 | // cases, because we check if the operands have vectorizable types | ||||||||
1593 | // beforehand in LoopVectorizationLegality. | ||||||||
1594 | return Scalars.find(VF) == Scalars.end() || | ||||||||
1595 | !isScalarAfterVectorization(I, VF); | ||||||||
1596 | }; | ||||||||
1597 | |||||||||
1598 | /// Returns a range containing only operands needing to be extracted. | ||||||||
1599 | SmallVector<Value *, 4> filterExtractingOperands(Instruction::op_range Ops, | ||||||||
1600 | ElementCount VF) { | ||||||||
1601 | return SmallVector<Value *, 4>(make_filter_range( | ||||||||
1602 | Ops, [this, VF](Value *V) { return this->needsExtract(V, VF); })); | ||||||||
1603 | } | ||||||||
1604 | |||||||||
1605 | public: | ||||||||
1606 | /// The loop that we evaluate. | ||||||||
1607 | Loop *TheLoop; | ||||||||
1608 | |||||||||
1609 | /// Predicated scalar evolution analysis. | ||||||||
1610 | PredicatedScalarEvolution &PSE; | ||||||||
1611 | |||||||||
1612 | /// Loop Info analysis. | ||||||||
1613 | LoopInfo *LI; | ||||||||
1614 | |||||||||
1615 | /// Vectorization legality. | ||||||||
1616 | LoopVectorizationLegality *Legal; | ||||||||
1617 | |||||||||
1618 | /// Vector target information. | ||||||||
1619 | const TargetTransformInfo &TTI; | ||||||||
1620 | |||||||||
1621 | /// Target Library Info. | ||||||||
1622 | const TargetLibraryInfo *TLI; | ||||||||
1623 | |||||||||
1624 | /// Demanded bits analysis. | ||||||||
1625 | DemandedBits *DB; | ||||||||
1626 | |||||||||
1627 | /// Assumption cache. | ||||||||
1628 | AssumptionCache *AC; | ||||||||
1629 | |||||||||
1630 | /// Interface to emit optimization remarks. | ||||||||
1631 | OptimizationRemarkEmitter *ORE; | ||||||||
1632 | |||||||||
1633 | const Function *TheFunction; | ||||||||
1634 | |||||||||
1635 | /// Loop Vectorize Hint. | ||||||||
1636 | const LoopVectorizeHints *Hints; | ||||||||
1637 | |||||||||
1638 | /// The interleave access information contains groups of interleaved accesses | ||||||||
1639 | /// with the same stride and close to each other. | ||||||||
1640 | InterleavedAccessInfo &InterleaveInfo; | ||||||||
1641 | |||||||||
1642 | /// Values to ignore in the cost model. | ||||||||
1643 | SmallPtrSet<const Value *, 16> ValuesToIgnore; | ||||||||
1644 | |||||||||
1645 | /// Values to ignore in the cost model when VF > 1. | ||||||||
1646 | SmallPtrSet<const Value *, 16> VecValuesToIgnore; | ||||||||
1647 | }; | ||||||||
1648 | |||||||||
1649 | } // end namespace llvm | ||||||||
1650 | |||||||||
1651 | // Return true if \p OuterLp is an outer loop annotated with hints for explicit | ||||||||
1652 | // vectorization. The loop needs to be annotated with #pragma omp simd | ||||||||
1653 | // simdlen(#) or #pragma clang vectorize(enable) vectorize_width(#). If the | ||||||||
1654 | // vector length information is not provided, vectorization is not considered | ||||||||
1655 | // explicit. Interleave hints are not allowed either. These limitations will be | ||||||||
1656 | // relaxed in the future. | ||||||||
1657 | // Please, note that we are currently forced to abuse the pragma 'clang | ||||||||
1658 | // vectorize' semantics. This pragma provides *auto-vectorization hints* | ||||||||
1659 | // (i.e., LV must check that vectorization is legal) whereas pragma 'omp simd' | ||||||||
1660 | // provides *explicit vectorization hints* (LV can bypass legal checks and | ||||||||
1661 | // assume that vectorization is legal). However, both hints are implemented | ||||||||
1662 | // using the same metadata (llvm.loop.vectorize, processed by | ||||||||
1663 | // LoopVectorizeHints). This will be fixed in the future when the native IR | ||||||||
1664 | // representation for pragma 'omp simd' is introduced. | ||||||||
1665 | static bool isExplicitVecOuterLoop(Loop *OuterLp, | ||||||||
1666 | OptimizationRemarkEmitter *ORE) { | ||||||||
1667 | assert(!OuterLp->isInnermost() && "This is not an outer loop")((!OuterLp->isInnermost() && "This is not an outer loop" ) ? static_cast<void> (0) : __assert_fail ("!OuterLp->isInnermost() && \"This is not an outer loop\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 1667, __PRETTY_FUNCTION__)); | ||||||||
1668 | LoopVectorizeHints Hints(OuterLp, true /*DisableInterleaving*/, *ORE); | ||||||||
1669 | |||||||||
1670 | // Only outer loops with an explicit vectorization hint are supported. | ||||||||
1671 | // Unannotated outer loops are ignored. | ||||||||
1672 | if (Hints.getForce() == LoopVectorizeHints::FK_Undefined) | ||||||||
1673 | return false; | ||||||||
1674 | |||||||||
1675 | Function *Fn = OuterLp->getHeader()->getParent(); | ||||||||
1676 | if (!Hints.allowVectorization(Fn, OuterLp, | ||||||||
1677 | true /*VectorizeOnlyWhenForced*/)) { | ||||||||
1678 | 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); | ||||||||
1679 | return false; | ||||||||
1680 | } | ||||||||
1681 | |||||||||
1682 | if (Hints.getInterleave() > 1) { | ||||||||
1683 | // TODO: Interleave support is future work. | ||||||||
1684 | 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) | ||||||||
1685 | "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); | ||||||||
1686 | Hints.emitRemarkWithHints(); | ||||||||
1687 | return false; | ||||||||
1688 | } | ||||||||
1689 | |||||||||
1690 | return true; | ||||||||
1691 | } | ||||||||
1692 | |||||||||
1693 | static void collectSupportedLoops(Loop &L, LoopInfo *LI, | ||||||||
1694 | OptimizationRemarkEmitter *ORE, | ||||||||
1695 | SmallVectorImpl<Loop *> &V) { | ||||||||
1696 | // Collect inner loops and outer loops without irreducible control flow. For | ||||||||
1697 | // now, only collect outer loops that have explicit vectorization hints. If we | ||||||||
1698 | // are stress testing the VPlan H-CFG construction, we collect the outermost | ||||||||
1699 | // loop of every loop nest. | ||||||||
1700 | if (L.isInnermost() || VPlanBuildStressTest || | ||||||||
1701 | (EnableVPlanNativePath && isExplicitVecOuterLoop(&L, ORE))) { | ||||||||
1702 | LoopBlocksRPO RPOT(&L); | ||||||||
1703 | RPOT.perform(LI); | ||||||||
1704 | if (!containsIrreducibleCFG<const BasicBlock *>(RPOT, *LI)) { | ||||||||
1705 | V.push_back(&L); | ||||||||
1706 | // TODO: Collect inner loops inside marked outer loops in case | ||||||||
1707 | // vectorization fails for the outer loop. Do not invoke | ||||||||
1708 | // 'containsIrreducibleCFG' again for inner loops when the outer loop is | ||||||||
1709 | // already known to be reducible. We can use an inherited attribute for | ||||||||
1710 | // that. | ||||||||
1711 | return; | ||||||||
1712 | } | ||||||||
1713 | } | ||||||||
1714 | for (Loop *InnerL : L) | ||||||||
1715 | collectSupportedLoops(*InnerL, LI, ORE, V); | ||||||||
1716 | } | ||||||||
1717 | |||||||||
1718 | namespace { | ||||||||
1719 | |||||||||
1720 | /// The LoopVectorize Pass. | ||||||||
1721 | struct LoopVectorize : public FunctionPass { | ||||||||
1722 | /// Pass identification, replacement for typeid | ||||||||
1723 | static char ID; | ||||||||
1724 | |||||||||
1725 | LoopVectorizePass Impl; | ||||||||
1726 | |||||||||
1727 | explicit LoopVectorize(bool InterleaveOnlyWhenForced = false, | ||||||||
1728 | bool VectorizeOnlyWhenForced = false) | ||||||||
1729 | : FunctionPass(ID), | ||||||||
1730 | Impl({InterleaveOnlyWhenForced, VectorizeOnlyWhenForced}) { | ||||||||
1731 | initializeLoopVectorizePass(*PassRegistry::getPassRegistry()); | ||||||||
1732 | } | ||||||||
1733 | |||||||||
1734 | bool runOnFunction(Function &F) override { | ||||||||
1735 | if (skipFunction(F)) | ||||||||
1736 | return false; | ||||||||
1737 | |||||||||
1738 | auto *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE(); | ||||||||
1739 | auto *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); | ||||||||
1740 | auto *TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F); | ||||||||
1741 | auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); | ||||||||
1742 | auto *BFI = &getAnalysis<BlockFrequencyInfoWrapperPass>().getBFI(); | ||||||||
1743 | auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>(); | ||||||||
1744 | auto *TLI = TLIP ? &TLIP->getTLI(F) : nullptr; | ||||||||
1745 | auto *AA = &getAnalysis<AAResultsWrapperPass>().getAAResults(); | ||||||||
1746 | auto *AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F); | ||||||||
1747 | auto *LAA = &getAnalysis<LoopAccessLegacyAnalysis>(); | ||||||||
1748 | auto *DB = &getAnalysis<DemandedBitsWrapperPass>().getDemandedBits(); | ||||||||
1749 | auto *ORE = &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE(); | ||||||||
1750 | auto *PSI = &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI(); | ||||||||
1751 | |||||||||
1752 | std::function<const LoopAccessInfo &(Loop &)> GetLAA = | ||||||||
1753 | [&](Loop &L) -> const LoopAccessInfo & { return LAA->getInfo(&L); }; | ||||||||
1754 | |||||||||
1755 | return Impl.runImpl(F, *SE, *LI, *TTI, *DT, *BFI, TLI, *DB, *AA, *AC, | ||||||||
1756 | GetLAA, *ORE, PSI).MadeAnyChange; | ||||||||
1757 | } | ||||||||
1758 | |||||||||
1759 | void getAnalysisUsage(AnalysisUsage &AU) const override { | ||||||||
1760 | AU.addRequired<AssumptionCacheTracker>(); | ||||||||
1761 | AU.addRequired<BlockFrequencyInfoWrapperPass>(); | ||||||||
1762 | AU.addRequired<DominatorTreeWrapperPass>(); | ||||||||
1763 | AU.addRequired<LoopInfoWrapperPass>(); | ||||||||
1764 | AU.addRequired<ScalarEvolutionWrapperPass>(); | ||||||||
1765 | AU.addRequired<TargetTransformInfoWrapperPass>(); | ||||||||
1766 | AU.addRequired<AAResultsWrapperPass>(); | ||||||||
1767 | AU.addRequired<LoopAccessLegacyAnalysis>(); | ||||||||
1768 | AU.addRequired<DemandedBitsWrapperPass>(); | ||||||||
1769 | AU.addRequired<OptimizationRemarkEmitterWrapperPass>(); | ||||||||
1770 | AU.addRequired<InjectTLIMappingsLegacy>(); | ||||||||
1771 | |||||||||
1772 | // We currently do not preserve loopinfo/dominator analyses with outer loop | ||||||||
1773 | // vectorization. Until this is addressed, mark these analyses as preserved | ||||||||
1774 | // only for non-VPlan-native path. | ||||||||
1775 | // TODO: Preserve Loop and Dominator analyses for VPlan-native path. | ||||||||
1776 | if (!EnableVPlanNativePath) { | ||||||||
1777 | AU.addPreserved<LoopInfoWrapperPass>(); | ||||||||
1778 | AU.addPreserved<DominatorTreeWrapperPass>(); | ||||||||
1779 | } | ||||||||
1780 | |||||||||
1781 | AU.addPreserved<BasicAAWrapperPass>(); | ||||||||
1782 | AU.addPreserved<GlobalsAAWrapperPass>(); | ||||||||
1783 | AU.addRequired<ProfileSummaryInfoWrapperPass>(); | ||||||||
1784 | } | ||||||||
1785 | }; | ||||||||
1786 | |||||||||
1787 | } // end anonymous namespace | ||||||||
1788 | |||||||||
1789 | //===----------------------------------------------------------------------===// | ||||||||
1790 | // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and | ||||||||
1791 | // LoopVectorizationCostModel and LoopVectorizationPlanner. | ||||||||
1792 | //===----------------------------------------------------------------------===// | ||||||||
1793 | |||||||||
1794 | Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) { | ||||||||
1795 | // We need to place the broadcast of invariant variables outside the loop, | ||||||||
1796 | // but only if it's proven safe to do so. Else, broadcast will be inside | ||||||||
1797 | // vector loop body. | ||||||||
1798 | Instruction *Instr = dyn_cast<Instruction>(V); | ||||||||
1799 | bool SafeToHoist = OrigLoop->isLoopInvariant(V) && | ||||||||
1800 | (!Instr || | ||||||||
1801 | DT->dominates(Instr->getParent(), LoopVectorPreHeader)); | ||||||||
1802 | // Place the code for broadcasting invariant variables in the new preheader. | ||||||||
1803 | IRBuilder<>::InsertPointGuard Guard(Builder); | ||||||||
1804 | if (SafeToHoist) | ||||||||
1805 | Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator()); | ||||||||
1806 | |||||||||
1807 | // Broadcast the scalar into all locations in the vector. | ||||||||
1808 | Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast"); | ||||||||
1809 | |||||||||
1810 | return Shuf; | ||||||||
1811 | } | ||||||||
1812 | |||||||||
1813 | void InnerLoopVectorizer::createVectorIntOrFpInductionPHI( | ||||||||
1814 | const InductionDescriptor &II, Value *Step, Instruction *EntryVal) { | ||||||||
1815 | assert((isa<PHINode>(EntryVal) || isa<TruncInst>(EntryVal)) &&(((isa<PHINode>(EntryVal) || isa<TruncInst>(EntryVal )) && "Expected either an induction phi-node or a truncate of it!" ) ? static_cast<void> (0) : __assert_fail ("(isa<PHINode>(EntryVal) || isa<TruncInst>(EntryVal)) && \"Expected either an induction phi-node or a truncate of it!\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 1816, __PRETTY_FUNCTION__)) | ||||||||
1816 | "Expected either an induction phi-node or a truncate of it!")(((isa<PHINode>(EntryVal) || isa<TruncInst>(EntryVal )) && "Expected either an induction phi-node or a truncate of it!" ) ? static_cast<void> (0) : __assert_fail ("(isa<PHINode>(EntryVal) || isa<TruncInst>(EntryVal)) && \"Expected either an induction phi-node or a truncate of it!\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 1816, __PRETTY_FUNCTION__)); | ||||||||
1817 | Value *Start = II.getStartValue(); | ||||||||
1818 | |||||||||
1819 | // Construct the initial value of the vector IV in the vector loop preheader | ||||||||
1820 | auto CurrIP = Builder.saveIP(); | ||||||||
1821 | Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator()); | ||||||||
1822 | if (isa<TruncInst>(EntryVal)) { | ||||||||
1823 | assert(Start->getType()->isIntegerTy() &&((Start->getType()->isIntegerTy() && "Truncation requires an integer type" ) ? static_cast<void> (0) : __assert_fail ("Start->getType()->isIntegerTy() && \"Truncation requires an integer type\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 1824, __PRETTY_FUNCTION__)) | ||||||||
1824 | "Truncation requires an integer type")((Start->getType()->isIntegerTy() && "Truncation requires an integer type" ) ? static_cast<void> (0) : __assert_fail ("Start->getType()->isIntegerTy() && \"Truncation requires an integer type\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 1824, __PRETTY_FUNCTION__)); | ||||||||
1825 | auto *TruncType = cast<IntegerType>(EntryVal->getType()); | ||||||||
1826 | Step = Builder.CreateTrunc(Step, TruncType); | ||||||||
1827 | Start = Builder.CreateCast(Instruction::Trunc, Start, TruncType); | ||||||||
1828 | } | ||||||||
1829 | Value *SplatStart = Builder.CreateVectorSplat(VF, Start); | ||||||||
1830 | Value *SteppedStart = | ||||||||
1831 | getStepVector(SplatStart, 0, Step, II.getInductionOpcode()); | ||||||||
1832 | |||||||||
1833 | // We create vector phi nodes for both integer and floating-point induction | ||||||||
1834 | // variables. Here, we determine the kind of arithmetic we will perform. | ||||||||
1835 | Instruction::BinaryOps AddOp; | ||||||||
1836 | Instruction::BinaryOps MulOp; | ||||||||
1837 | if (Step->getType()->isIntegerTy()) { | ||||||||
1838 | AddOp = Instruction::Add; | ||||||||
1839 | MulOp = Instruction::Mul; | ||||||||
1840 | } else { | ||||||||
1841 | AddOp = II.getInductionOpcode(); | ||||||||
1842 | MulOp = Instruction::FMul; | ||||||||
1843 | } | ||||||||
1844 | |||||||||
1845 | // Multiply the vectorization factor by the step using integer or | ||||||||
1846 | // floating-point arithmetic as appropriate. | ||||||||
1847 | Value *ConstVF = | ||||||||
1848 | getSignedIntOrFpConstant(Step->getType(), VF.getKnownMinValue()); | ||||||||
1849 | Value *Mul = addFastMathFlag(Builder.CreateBinOp(MulOp, Step, ConstVF)); | ||||||||
1850 | |||||||||
1851 | // Create a vector splat to use in the induction update. | ||||||||
1852 | // | ||||||||
1853 | // FIXME: If the step is non-constant, we create the vector splat with | ||||||||
1854 | // IRBuilder. IRBuilder can constant-fold the multiply, but it doesn't | ||||||||
1855 | // handle a constant vector splat. | ||||||||
1856 | assert(!VF.isScalable() && "scalable vectors not yet supported.")((!VF.isScalable() && "scalable vectors not yet supported." ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"scalable vectors not yet supported.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 1856, __PRETTY_FUNCTION__)); | ||||||||
1857 | Value *SplatVF = isa<Constant>(Mul) | ||||||||
1858 | ? ConstantVector::getSplat(VF, cast<Constant>(Mul)) | ||||||||
1859 | : Builder.CreateVectorSplat(VF, Mul); | ||||||||
1860 | Builder.restoreIP(CurrIP); | ||||||||
1861 | |||||||||
1862 | // We may need to add the step a number of times, depending on the unroll | ||||||||
1863 | // factor. The last of those goes into the PHI. | ||||||||
1864 | PHINode *VecInd = PHINode::Create(SteppedStart->getType(), 2, "vec.ind", | ||||||||
1865 | &*LoopVectorBody->getFirstInsertionPt()); | ||||||||
1866 | VecInd->setDebugLoc(EntryVal->getDebugLoc()); | ||||||||
1867 | Instruction *LastInduction = VecInd; | ||||||||
1868 | for (unsigned Part = 0; Part < UF; ++Part) { | ||||||||
1869 | VectorLoopValueMap.setVectorValue(EntryVal, Part, LastInduction); | ||||||||
1870 | |||||||||
1871 | if (isa<TruncInst>(EntryVal)) | ||||||||
1872 | addMetadata(LastInduction, EntryVal); | ||||||||
1873 | recordVectorLoopValueForInductionCast(II, EntryVal, LastInduction, Part); | ||||||||
1874 | |||||||||
1875 | LastInduction = cast<Instruction>(addFastMathFlag( | ||||||||
1876 | Builder.CreateBinOp(AddOp, LastInduction, SplatVF, "step.add"))); | ||||||||
1877 | LastInduction->setDebugLoc(EntryVal->getDebugLoc()); | ||||||||
1878 | } | ||||||||
1879 | |||||||||
1880 | // Move the last step to the end of the latch block. This ensures consistent | ||||||||
1881 | // placement of all induction updates. | ||||||||
1882 | auto *LoopVectorLatch = LI->getLoopFor(LoopVectorBody)->getLoopLatch(); | ||||||||
1883 | auto *Br = cast<BranchInst>(LoopVectorLatch->getTerminator()); | ||||||||
1884 | auto *ICmp = cast<Instruction>(Br->getCondition()); | ||||||||
1885 | LastInduction->moveBefore(ICmp); | ||||||||
1886 | LastInduction->setName("vec.ind.next"); | ||||||||
1887 | |||||||||
1888 | VecInd->addIncoming(SteppedStart, LoopVectorPreHeader); | ||||||||
1889 | VecInd->addIncoming(LastInduction, LoopVectorLatch); | ||||||||
1890 | } | ||||||||
1891 | |||||||||
1892 | bool InnerLoopVectorizer::shouldScalarizeInstruction(Instruction *I) const { | ||||||||
1893 | return Cost->isScalarAfterVectorization(I, VF) || | ||||||||
1894 | Cost->isProfitableToScalarize(I, VF); | ||||||||
1895 | } | ||||||||
1896 | |||||||||
1897 | bool InnerLoopVectorizer::needsScalarInduction(Instruction *IV) const { | ||||||||
1898 | if (shouldScalarizeInstruction(IV)) | ||||||||
1899 | return true; | ||||||||
1900 | auto isScalarInst = [&](User *U) -> bool { | ||||||||
1901 | auto *I = cast<Instruction>(U); | ||||||||
1902 | return (OrigLoop->contains(I) && shouldScalarizeInstruction(I)); | ||||||||
1903 | }; | ||||||||
1904 | return llvm::any_of(IV->users(), isScalarInst); | ||||||||
1905 | } | ||||||||
1906 | |||||||||
1907 | void InnerLoopVectorizer::recordVectorLoopValueForInductionCast( | ||||||||
1908 | const InductionDescriptor &ID, const Instruction *EntryVal, | ||||||||
1909 | Value *VectorLoopVal, unsigned Part, unsigned Lane) { | ||||||||
1910 | assert((isa<PHINode>(EntryVal) || isa<TruncInst>(EntryVal)) &&(((isa<PHINode>(EntryVal) || isa<TruncInst>(EntryVal )) && "Expected either an induction phi-node or a truncate of it!" ) ? static_cast<void> (0) : __assert_fail ("(isa<PHINode>(EntryVal) || isa<TruncInst>(EntryVal)) && \"Expected either an induction phi-node or a truncate of it!\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 1911, __PRETTY_FUNCTION__)) | ||||||||
1911 | "Expected either an induction phi-node or a truncate of it!")(((isa<PHINode>(EntryVal) || isa<TruncInst>(EntryVal )) && "Expected either an induction phi-node or a truncate of it!" ) ? static_cast<void> (0) : __assert_fail ("(isa<PHINode>(EntryVal) || isa<TruncInst>(EntryVal)) && \"Expected either an induction phi-node or a truncate of it!\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 1911, __PRETTY_FUNCTION__)); | ||||||||
1912 | |||||||||
1913 | // This induction variable is not the phi from the original loop but the | ||||||||
1914 | // newly-created IV based on the proof that casted Phi is equal to the | ||||||||
1915 | // uncasted Phi in the vectorized loop (under a runtime guard possibly). It | ||||||||
1916 | // re-uses the same InductionDescriptor that original IV uses but we don't | ||||||||
1917 | // have to do any recording in this case - that is done when original IV is | ||||||||
1918 | // processed. | ||||||||
1919 | if (isa<TruncInst>(EntryVal)) | ||||||||
1920 | return; | ||||||||
1921 | |||||||||
1922 | const SmallVectorImpl<Instruction *> &Casts = ID.getCastInsts(); | ||||||||
1923 | if (Casts.empty()) | ||||||||
1924 | return; | ||||||||
1925 | // Only the first Cast instruction in the Casts vector is of interest. | ||||||||
1926 | // The rest of the Casts (if exist) have no uses outside the | ||||||||
1927 | // induction update chain itself. | ||||||||
1928 | Instruction *CastInst = *Casts.begin(); | ||||||||
1929 | if (Lane < UINT_MAX(2147483647 *2U +1U)) | ||||||||
1930 | VectorLoopValueMap.setScalarValue(CastInst, {Part, Lane}, VectorLoopVal); | ||||||||
1931 | else | ||||||||
1932 | VectorLoopValueMap.setVectorValue(CastInst, Part, VectorLoopVal); | ||||||||
1933 | } | ||||||||
1934 | |||||||||
1935 | void InnerLoopVectorizer::widenIntOrFpInduction(PHINode *IV, TruncInst *Trunc) { | ||||||||
1936 | assert((IV->getType()->isIntegerTy() || IV != OldInduction) &&(((IV->getType()->isIntegerTy() || IV != OldInduction) && "Primary induction variable must have an integer type") ? static_cast <void> (0) : __assert_fail ("(IV->getType()->isIntegerTy() || IV != OldInduction) && \"Primary induction variable must have an integer type\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 1937, __PRETTY_FUNCTION__)) | ||||||||
1937 | "Primary induction variable must have an integer type")(((IV->getType()->isIntegerTy() || IV != OldInduction) && "Primary induction variable must have an integer type") ? static_cast <void> (0) : __assert_fail ("(IV->getType()->isIntegerTy() || IV != OldInduction) && \"Primary induction variable must have an integer type\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 1937, __PRETTY_FUNCTION__)); | ||||||||
1938 | |||||||||
1939 | auto II = Legal->getInductionVars().find(IV); | ||||||||
1940 | assert(II != Legal->getInductionVars().end() && "IV is not an induction")((II != Legal->getInductionVars().end() && "IV is not an induction" ) ? static_cast<void> (0) : __assert_fail ("II != Legal->getInductionVars().end() && \"IV is not an induction\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 1940, __PRETTY_FUNCTION__)); | ||||||||
1941 | |||||||||
1942 | auto ID = II->second; | ||||||||
1943 | assert(IV->getType() == ID.getStartValue()->getType() && "Types must match")((IV->getType() == ID.getStartValue()->getType() && "Types must match") ? static_cast<void> (0) : __assert_fail ("IV->getType() == ID.getStartValue()->getType() && \"Types must match\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 1943, __PRETTY_FUNCTION__)); | ||||||||
1944 | |||||||||
1945 | // The value from the original loop to which we are mapping the new induction | ||||||||
1946 | // variable. | ||||||||
1947 | Instruction *EntryVal = Trunc ? cast<Instruction>(Trunc) : IV; | ||||||||
1948 | |||||||||
1949 | auto &DL = OrigLoop->getHeader()->getModule()->getDataLayout(); | ||||||||
1950 | |||||||||
1951 | // Generate code for the induction step. Note that induction steps are | ||||||||
1952 | // required to be loop-invariant | ||||||||
1953 | auto CreateStepValue = [&](const SCEV *Step) -> Value * { | ||||||||
1954 | assert(PSE.getSE()->isLoopInvariant(Step, OrigLoop) &&((PSE.getSE()->isLoopInvariant(Step, OrigLoop) && "Induction step should be loop invariant" ) ? static_cast<void> (0) : __assert_fail ("PSE.getSE()->isLoopInvariant(Step, OrigLoop) && \"Induction step should be loop invariant\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 1955, __PRETTY_FUNCTION__)) | ||||||||
1955 | "Induction step should be loop invariant")((PSE.getSE()->isLoopInvariant(Step, OrigLoop) && "Induction step should be loop invariant" ) ? static_cast<void> (0) : __assert_fail ("PSE.getSE()->isLoopInvariant(Step, OrigLoop) && \"Induction step should be loop invariant\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 1955, __PRETTY_FUNCTION__)); | ||||||||
1956 | if (PSE.getSE()->isSCEVable(IV->getType())) { | ||||||||
1957 | SCEVExpander Exp(*PSE.getSE(), DL, "induction"); | ||||||||
1958 | return Exp.expandCodeFor(Step, Step->getType(), | ||||||||
1959 | LoopVectorPreHeader->getTerminator()); | ||||||||
1960 | } | ||||||||
1961 | return cast<SCEVUnknown>(Step)->getValue(); | ||||||||
1962 | }; | ||||||||
1963 | |||||||||
1964 | // The scalar value to broadcast. This is derived from the canonical | ||||||||
1965 | // induction variable. If a truncation type is given, truncate the canonical | ||||||||
1966 | // induction variable and step. Otherwise, derive these values from the | ||||||||
1967 | // induction descriptor. | ||||||||
1968 | auto CreateScalarIV = [&](Value *&Step) -> Value * { | ||||||||
1969 | Value *ScalarIV = Induction; | ||||||||
1970 | if (IV != OldInduction) { | ||||||||
1971 | ScalarIV = IV->getType()->isIntegerTy() | ||||||||
1972 | ? Builder.CreateSExtOrTrunc(Induction, IV->getType()) | ||||||||
1973 | : Builder.CreateCast(Instruction::SIToFP, Induction, | ||||||||
1974 | IV->getType()); | ||||||||
1975 | ScalarIV = emitTransformedIndex(Builder, ScalarIV, PSE.getSE(), DL, ID); | ||||||||
1976 | ScalarIV->setName("offset.idx"); | ||||||||
1977 | } | ||||||||
1978 | if (Trunc) { | ||||||||
1979 | auto *TruncType = cast<IntegerType>(Trunc->getType()); | ||||||||
1980 | assert(Step->getType()->isIntegerTy() &&((Step->getType()->isIntegerTy() && "Truncation requires an integer step" ) ? static_cast<void> (0) : __assert_fail ("Step->getType()->isIntegerTy() && \"Truncation requires an integer step\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 1981, __PRETTY_FUNCTION__)) | ||||||||
1981 | "Truncation requires an integer step")((Step->getType()->isIntegerTy() && "Truncation requires an integer step" ) ? static_cast<void> (0) : __assert_fail ("Step->getType()->isIntegerTy() && \"Truncation requires an integer step\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 1981, __PRETTY_FUNCTION__)); | ||||||||
1982 | ScalarIV = Builder.CreateTrunc(ScalarIV, TruncType); | ||||||||
1983 | Step = Builder.CreateTrunc(Step, TruncType); | ||||||||
1984 | } | ||||||||
1985 | return ScalarIV; | ||||||||
1986 | }; | ||||||||
1987 | |||||||||
1988 | // Create the vector values from the scalar IV, in the absence of creating a | ||||||||
1989 | // vector IV. | ||||||||
1990 | auto CreateSplatIV = [&](Value *ScalarIV, Value *Step) { | ||||||||
1991 | Value *Broadcasted = getBroadcastInstrs(ScalarIV); | ||||||||
1992 | for (unsigned Part = 0; Part < UF; ++Part) { | ||||||||
1993 | assert(!VF.isScalable() && "scalable vectors not yet supported.")((!VF.isScalable() && "scalable vectors not yet supported." ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"scalable vectors not yet supported.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 1993, __PRETTY_FUNCTION__)); | ||||||||
1994 | Value *EntryPart = | ||||||||
1995 | getStepVector(Broadcasted, VF.getKnownMinValue() * Part, Step, | ||||||||
1996 | ID.getInductionOpcode()); | ||||||||
1997 | VectorLoopValueMap.setVectorValue(EntryVal, Part, EntryPart); | ||||||||
1998 | if (Trunc) | ||||||||
1999 | addMetadata(EntryPart, Trunc); | ||||||||
2000 | recordVectorLoopValueForInductionCast(ID, EntryVal, EntryPart, Part); | ||||||||
2001 | } | ||||||||
2002 | }; | ||||||||
2003 | |||||||||
2004 | // Now do the actual transformations, and start with creating the step value. | ||||||||
2005 | Value *Step = CreateStepValue(ID.getStep()); | ||||||||
2006 | if (VF.isZero() || VF.isScalar()) { | ||||||||
2007 | Value *ScalarIV = CreateScalarIV(Step); | ||||||||
2008 | CreateSplatIV(ScalarIV, Step); | ||||||||
2009 | return; | ||||||||
2010 | } | ||||||||
2011 | |||||||||
2012 | // Determine if we want a scalar version of the induction variable. This is | ||||||||
2013 | // true if the induction variable itself is not widened, or if it has at | ||||||||
2014 | // least one user in the loop that is not widened. | ||||||||
2015 | auto NeedsScalarIV = needsScalarInduction(EntryVal); | ||||||||
2016 | if (!NeedsScalarIV) { | ||||||||
2017 | createVectorIntOrFpInductionPHI(ID, Step, EntryVal); | ||||||||
2018 | return; | ||||||||
2019 | } | ||||||||
2020 | |||||||||
2021 | // Try to create a new independent vector induction variable. If we can't | ||||||||
2022 | // create the phi node, we will splat the scalar induction variable in each | ||||||||
2023 | // loop iteration. | ||||||||
2024 | if (!shouldScalarizeInstruction(EntryVal)) { | ||||||||
2025 | createVectorIntOrFpInductionPHI(ID, Step, EntryVal); | ||||||||
2026 | Value *ScalarIV = CreateScalarIV(Step); | ||||||||
2027 | // Create scalar steps that can be used by instructions we will later | ||||||||
2028 | // scalarize. Note that the addition of the scalar steps will not increase | ||||||||
2029 | // the number of instructions in the loop in the common case prior to | ||||||||
2030 | // InstCombine. We will be trading one vector extract for each scalar step. | ||||||||
2031 | buildScalarSteps(ScalarIV, Step, EntryVal, ID); | ||||||||
2032 | return; | ||||||||
2033 | } | ||||||||
2034 | |||||||||
2035 | // All IV users are scalar instructions, so only emit a scalar IV, not a | ||||||||
2036 | // vectorised IV. Except when we tail-fold, then the splat IV feeds the | ||||||||
2037 | // predicate used by the masked loads/stores. | ||||||||
2038 | Value *ScalarIV = CreateScalarIV(Step); | ||||||||
2039 | if (!Cost->isScalarEpilogueAllowed()) | ||||||||
2040 | CreateSplatIV(ScalarIV, Step); | ||||||||
2041 | buildScalarSteps(ScalarIV, Step, EntryVal, ID); | ||||||||
2042 | } | ||||||||
2043 | |||||||||
2044 | Value *InnerLoopVectorizer::getStepVector(Value *Val, int StartIdx, Value *Step, | ||||||||
2045 | Instruction::BinaryOps BinOp) { | ||||||||
2046 | // Create and check the types. | ||||||||
2047 | auto *ValVTy = cast<FixedVectorType>(Val->getType()); | ||||||||
2048 | int VLen = ValVTy->getNumElements(); | ||||||||
2049 | |||||||||
2050 | Type *STy = Val->getType()->getScalarType(); | ||||||||
2051 | assert((STy->isIntegerTy() || STy->isFloatingPointTy()) &&(((STy->isIntegerTy() || STy->isFloatingPointTy()) && "Induction Step must be an integer or FP") ? static_cast< void> (0) : __assert_fail ("(STy->isIntegerTy() || STy->isFloatingPointTy()) && \"Induction Step must be an integer or FP\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2052, __PRETTY_FUNCTION__)) | ||||||||
2052 | "Induction Step must be an integer or FP")(((STy->isIntegerTy() || STy->isFloatingPointTy()) && "Induction Step must be an integer or FP") ? static_cast< void> (0) : __assert_fail ("(STy->isIntegerTy() || STy->isFloatingPointTy()) && \"Induction Step must be an integer or FP\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2052, __PRETTY_FUNCTION__)); | ||||||||
2053 | assert(Step->getType() == STy && "Step has wrong type")((Step->getType() == STy && "Step has wrong type") ? static_cast<void> (0) : __assert_fail ("Step->getType() == STy && \"Step has wrong type\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2053, __PRETTY_FUNCTION__)); | ||||||||
2054 | |||||||||
2055 | SmallVector<Constant *, 8> Indices; | ||||||||
2056 | |||||||||
2057 | if (STy->isIntegerTy()) { | ||||||||
2058 | // Create a vector of consecutive numbers from zero to VF. | ||||||||
2059 | for (int i = 0; i < VLen; ++i) | ||||||||
2060 | Indices.push_back(ConstantInt::get(STy, StartIdx + i)); | ||||||||
2061 | |||||||||
2062 | // Add the consecutive indices to the vector value. | ||||||||
2063 | Constant *Cv = ConstantVector::get(Indices); | ||||||||
2064 | assert(Cv->getType() == Val->getType() && "Invalid consecutive vec")((Cv->getType() == Val->getType() && "Invalid consecutive vec" ) ? static_cast<void> (0) : __assert_fail ("Cv->getType() == Val->getType() && \"Invalid consecutive vec\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2064, __PRETTY_FUNCTION__)); | ||||||||
2065 | Step = Builder.CreateVectorSplat(VLen, Step); | ||||||||
2066 | assert(Step->getType() == Val->getType() && "Invalid step vec")((Step->getType() == Val->getType() && "Invalid step vec" ) ? static_cast<void> (0) : __assert_fail ("Step->getType() == Val->getType() && \"Invalid step vec\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2066, __PRETTY_FUNCTION__)); | ||||||||
2067 | // FIXME: The newly created binary instructions should contain nsw/nuw flags, | ||||||||
2068 | // which can be found from the original scalar operations. | ||||||||
2069 | Step = Builder.CreateMul(Cv, Step); | ||||||||
2070 | return Builder.CreateAdd(Val, Step, "induction"); | ||||||||
2071 | } | ||||||||
2072 | |||||||||
2073 | // Floating point induction. | ||||||||
2074 | assert((BinOp == Instruction::FAdd || BinOp == Instruction::FSub) &&(((BinOp == Instruction::FAdd || BinOp == Instruction::FSub) && "Binary Opcode should be specified for FP induction") ? static_cast <void> (0) : __assert_fail ("(BinOp == Instruction::FAdd || BinOp == Instruction::FSub) && \"Binary Opcode should be specified for FP induction\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2075, __PRETTY_FUNCTION__)) | ||||||||
2075 | "Binary Opcode should be specified for FP induction")(((BinOp == Instruction::FAdd || BinOp == Instruction::FSub) && "Binary Opcode should be specified for FP induction") ? static_cast <void> (0) : __assert_fail ("(BinOp == Instruction::FAdd || BinOp == Instruction::FSub) && \"Binary Opcode should be specified for FP induction\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2075, __PRETTY_FUNCTION__)); | ||||||||
2076 | // Create a vector of consecutive numbers from zero to VF. | ||||||||
2077 | for (int i = 0; i < VLen; ++i) | ||||||||
2078 | Indices.push_back(ConstantFP::get(STy, (double)(StartIdx + i))); | ||||||||
2079 | |||||||||
2080 | // Add the consecutive indices to the vector value. | ||||||||
2081 | Constant *Cv = ConstantVector::get(Indices); | ||||||||
2082 | |||||||||
2083 | Step = Builder.CreateVectorSplat(VLen, Step); | ||||||||
2084 | |||||||||
2085 | // Floating point operations had to be 'fast' to enable the induction. | ||||||||
2086 | FastMathFlags Flags; | ||||||||
2087 | Flags.setFast(); | ||||||||
2088 | |||||||||
2089 | Value *MulOp = Builder.CreateFMul(Cv, Step); | ||||||||
2090 | if (isa<Instruction>(MulOp)) | ||||||||
2091 | // Have to check, MulOp may be a constant | ||||||||
2092 | cast<Instruction>(MulOp)->setFastMathFlags(Flags); | ||||||||
2093 | |||||||||
2094 | Value *BOp = Builder.CreateBinOp(BinOp, Val, MulOp, "induction"); | ||||||||
2095 | if (isa<Instruction>(BOp)) | ||||||||
2096 | cast<Instruction>(BOp)->setFastMathFlags(Flags); | ||||||||
2097 | return BOp; | ||||||||
2098 | } | ||||||||
2099 | |||||||||
2100 | void InnerLoopVectorizer::buildScalarSteps(Value *ScalarIV, Value *Step, | ||||||||
2101 | Instruction *EntryVal, | ||||||||
2102 | const InductionDescriptor &ID) { | ||||||||
2103 | // We shouldn't have to build scalar steps if we aren't vectorizing. | ||||||||
2104 | assert(VF.isVector() && "VF should be greater than one")((VF.isVector() && "VF should be greater than one") ? static_cast<void> (0) : __assert_fail ("VF.isVector() && \"VF should be greater than one\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2104, __PRETTY_FUNCTION__)); | ||||||||
2105 | assert(!VF.isScalable() &&((!VF.isScalable() && "the code below assumes a fixed number of elements at compile time" ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"the code below assumes a fixed number of elements at compile time\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2106, __PRETTY_FUNCTION__)) | ||||||||
2106 | "the code below assumes a fixed number of elements at compile time")((!VF.isScalable() && "the code below assumes a fixed number of elements at compile time" ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"the code below assumes a fixed number of elements at compile time\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2106, __PRETTY_FUNCTION__)); | ||||||||
2107 | // Get the value type and ensure it and the step have the same integer type. | ||||||||
2108 | Type *ScalarIVTy = ScalarIV->getType()->getScalarType(); | ||||||||
2109 | assert(ScalarIVTy == Step->getType() &&((ScalarIVTy == Step->getType() && "Val and Step should have the same type" ) ? static_cast<void> (0) : __assert_fail ("ScalarIVTy == Step->getType() && \"Val and Step should have the same type\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2110, __PRETTY_FUNCTION__)) | ||||||||
2110 | "Val and Step should have the same type")((ScalarIVTy == Step->getType() && "Val and Step should have the same type" ) ? static_cast<void> (0) : __assert_fail ("ScalarIVTy == Step->getType() && \"Val and Step should have the same type\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2110, __PRETTY_FUNCTION__)); | ||||||||
2111 | |||||||||
2112 | // We build scalar steps for both integer and floating-point induction | ||||||||
2113 | // variables. Here, we determine the kind of arithmetic we will perform. | ||||||||
2114 | Instruction::BinaryOps AddOp; | ||||||||
2115 | Instruction::BinaryOps MulOp; | ||||||||
2116 | if (ScalarIVTy->isIntegerTy()) { | ||||||||
2117 | AddOp = Instruction::Add; | ||||||||
2118 | MulOp = Instruction::Mul; | ||||||||
2119 | } else { | ||||||||
2120 | AddOp = ID.getInductionOpcode(); | ||||||||
2121 | MulOp = Instruction::FMul; | ||||||||
2122 | } | ||||||||
2123 | |||||||||
2124 | // Determine the number of scalars we need to generate for each unroll | ||||||||
2125 | // iteration. If EntryVal is uniform, we only need to generate the first | ||||||||
2126 | // lane. Otherwise, we generate all VF values. | ||||||||
2127 | unsigned Lanes = | ||||||||
2128 | Cost->isUniformAfterVectorization(cast<Instruction>(EntryVal), VF) | ||||||||
2129 | ? 1 | ||||||||
2130 | : VF.getKnownMinValue(); | ||||||||
2131 | // Compute the scalar steps and save the results in VectorLoopValueMap. | ||||||||
2132 | for (unsigned Part = 0; Part < UF; ++Part) { | ||||||||
2133 | for (unsigned Lane = 0; Lane < Lanes; ++Lane) { | ||||||||
2134 | auto *StartIdx = getSignedIntOrFpConstant( | ||||||||
2135 | ScalarIVTy, VF.getKnownMinValue() * Part + Lane); | ||||||||
2136 | auto *Mul = addFastMathFlag(Builder.CreateBinOp(MulOp, StartIdx, Step)); | ||||||||
2137 | auto *Add = addFastMathFlag(Builder.CreateBinOp(AddOp, ScalarIV, Mul)); | ||||||||
2138 | VectorLoopValueMap.setScalarValue(EntryVal, {Part, Lane}, Add); | ||||||||
2139 | recordVectorLoopValueForInductionCast(ID, EntryVal, Add, Part, Lane); | ||||||||
2140 | } | ||||||||
2141 | } | ||||||||
2142 | } | ||||||||
2143 | |||||||||
2144 | Value *InnerLoopVectorizer::getOrCreateVectorValue(Value *V, unsigned Part) { | ||||||||
2145 | assert(V != Induction && "The new induction variable should not be used.")((V != Induction && "The new induction variable should not be used." ) ? static_cast<void> (0) : __assert_fail ("V != Induction && \"The new induction variable should not be used.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2145, __PRETTY_FUNCTION__)); | ||||||||
2146 | assert(!V->getType()->isVectorTy() && "Can't widen a vector")((!V->getType()->isVectorTy() && "Can't widen a vector" ) ? static_cast<void> (0) : __assert_fail ("!V->getType()->isVectorTy() && \"Can't widen a vector\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2146, __PRETTY_FUNCTION__)); | ||||||||
2147 | assert(!V->getType()->isVoidTy() && "Type does not produce a value")((!V->getType()->isVoidTy() && "Type does not produce a value" ) ? static_cast<void> (0) : __assert_fail ("!V->getType()->isVoidTy() && \"Type does not produce a value\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2147, __PRETTY_FUNCTION__)); | ||||||||
2148 | |||||||||
2149 | // If we have a stride that is replaced by one, do it here. Defer this for | ||||||||
2150 | // the VPlan-native path until we start running Legal checks in that path. | ||||||||
2151 | if (!EnableVPlanNativePath && Legal->hasStride(V)) | ||||||||
2152 | V = ConstantInt::get(V->getType(), 1); | ||||||||
2153 | |||||||||
2154 | // If we have a vector mapped to this value, return it. | ||||||||
2155 | if (VectorLoopValueMap.hasVectorValue(V, Part)) | ||||||||
2156 | return VectorLoopValueMap.getVectorValue(V, Part); | ||||||||
2157 | |||||||||
2158 | // If the value has not been vectorized, check if it has been scalarized | ||||||||
2159 | // instead. If it has been scalarized, and we actually need the value in | ||||||||
2160 | // vector form, we will construct the vector values on demand. | ||||||||
2161 | if (VectorLoopValueMap.hasAnyScalarValue(V)) { | ||||||||
2162 | Value *ScalarValue = VectorLoopValueMap.getScalarValue(V, {Part, 0}); | ||||||||
2163 | |||||||||
2164 | // If we've scalarized a value, that value should be an instruction. | ||||||||
2165 | auto *I = cast<Instruction>(V); | ||||||||
2166 | |||||||||
2167 | // If we aren't vectorizing, we can just copy the scalar map values over to | ||||||||
2168 | // the vector map. | ||||||||
2169 | if (VF == 1) { | ||||||||
2170 | VectorLoopValueMap.setVectorValue(V, Part, ScalarValue); | ||||||||
2171 | return ScalarValue; | ||||||||
2172 | } | ||||||||
2173 | |||||||||
2174 | // Get the last scalar instruction we generated for V and Part. If the value | ||||||||
2175 | // is known to be uniform after vectorization, this corresponds to lane zero | ||||||||
2176 | // of the Part unroll iteration. Otherwise, the last instruction is the one | ||||||||
2177 | // we created for the last vector lane of the Part unroll iteration. | ||||||||
2178 | assert(!VF.isScalable() && "scalable vectors not yet supported.")((!VF.isScalable() && "scalable vectors not yet supported." ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"scalable vectors not yet supported.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2178, __PRETTY_FUNCTION__)); | ||||||||
2179 | unsigned LastLane = Cost->isUniformAfterVectorization(I, VF) | ||||||||
2180 | ? 0 | ||||||||
2181 | : VF.getKnownMinValue() - 1; | ||||||||
2182 | auto *LastInst = cast<Instruction>( | ||||||||
2183 | VectorLoopValueMap.getScalarValue(V, {Part, LastLane})); | ||||||||
2184 | |||||||||
2185 | // Set the insert point after the last scalarized instruction. This ensures | ||||||||
2186 | // the insertelement sequence will directly follow the scalar definitions. | ||||||||
2187 | auto OldIP = Builder.saveIP(); | ||||||||
2188 | auto NewIP = std::next(BasicBlock::iterator(LastInst)); | ||||||||
2189 | Builder.SetInsertPoint(&*NewIP); | ||||||||
2190 | |||||||||
2191 | // However, if we are vectorizing, we need to construct the vector values. | ||||||||
2192 | // If the value is known to be uniform after vectorization, we can just | ||||||||
2193 | // broadcast the scalar value corresponding to lane zero for each unroll | ||||||||
2194 | // iteration. Otherwise, we construct the vector values using insertelement | ||||||||
2195 | // instructions. Since the resulting vectors are stored in | ||||||||
2196 | // VectorLoopValueMap, we will only generate the insertelements once. | ||||||||
2197 | Value *VectorValue = nullptr; | ||||||||
2198 | if (Cost->isUniformAfterVectorization(I, VF)) { | ||||||||
2199 | VectorValue = getBroadcastInstrs(ScalarValue); | ||||||||
2200 | VectorLoopValueMap.setVectorValue(V, Part, VectorValue); | ||||||||
2201 | } else { | ||||||||
2202 | // Initialize packing with insertelements to start from undef. | ||||||||
2203 | assert(!VF.isScalable() && "VF is assumed to be non scalable.")((!VF.isScalable() && "VF is assumed to be non scalable." ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"VF is assumed to be non scalable.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2203, __PRETTY_FUNCTION__)); | ||||||||
2204 | Value *Undef = UndefValue::get(VectorType::get(V->getType(), VF)); | ||||||||
2205 | VectorLoopValueMap.setVectorValue(V, Part, Undef); | ||||||||
2206 | for (unsigned Lane = 0; Lane < VF.getKnownMinValue(); ++Lane) | ||||||||
2207 | packScalarIntoVectorValue(V, {Part, Lane}); | ||||||||
2208 | VectorValue = VectorLoopValueMap.getVectorValue(V, Part); | ||||||||
2209 | } | ||||||||
2210 | Builder.restoreIP(OldIP); | ||||||||
2211 | return VectorValue; | ||||||||
2212 | } | ||||||||
2213 | |||||||||
2214 | // If this scalar is unknown, assume that it is a constant or that it is | ||||||||
2215 | // loop invariant. Broadcast V and save the value for future uses. | ||||||||
2216 | Value *B = getBroadcastInstrs(V); | ||||||||
2217 | VectorLoopValueMap.setVectorValue(V, Part, B); | ||||||||
2218 | return B; | ||||||||
2219 | } | ||||||||
2220 | |||||||||
2221 | Value * | ||||||||
2222 | InnerLoopVectorizer::getOrCreateScalarValue(Value *V, | ||||||||
2223 | const VPIteration &Instance) { | ||||||||
2224 | // If the value is not an instruction contained in the loop, it should | ||||||||
2225 | // already be scalar. | ||||||||
2226 | if (OrigLoop->isLoopInvariant(V)) | ||||||||
2227 | return V; | ||||||||
2228 | |||||||||
2229 | assert(Instance.Lane > 0((Instance.Lane > 0 ? !Cost->isUniformAfterVectorization (cast<Instruction>(V), VF) : true && "Uniform values only have lane zero" ) ? static_cast<void> (0) : __assert_fail ("Instance.Lane > 0 ? !Cost->isUniformAfterVectorization(cast<Instruction>(V), VF) : true && \"Uniform values only have lane zero\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2231, __PRETTY_FUNCTION__)) | ||||||||
2230 | ? !Cost->isUniformAfterVectorization(cast<Instruction>(V), VF)((Instance.Lane > 0 ? !Cost->isUniformAfterVectorization (cast<Instruction>(V), VF) : true && "Uniform values only have lane zero" ) ? static_cast<void> (0) : __assert_fail ("Instance.Lane > 0 ? !Cost->isUniformAfterVectorization(cast<Instruction>(V), VF) : true && \"Uniform values only have lane zero\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2231, __PRETTY_FUNCTION__)) | ||||||||
2231 | : true && "Uniform values only have lane zero")((Instance.Lane > 0 ? !Cost->isUniformAfterVectorization (cast<Instruction>(V), VF) : true && "Uniform values only have lane zero" ) ? static_cast<void> (0) : __assert_fail ("Instance.Lane > 0 ? !Cost->isUniformAfterVectorization(cast<Instruction>(V), VF) : true && \"Uniform values only have lane zero\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2231, __PRETTY_FUNCTION__)); | ||||||||
2232 | |||||||||
2233 | // If the value from the original loop has not been vectorized, it is | ||||||||
2234 | // represented by UF x VF scalar values in the new loop. Return the requested | ||||||||
2235 | // scalar value. | ||||||||
2236 | if (VectorLoopValueMap.hasScalarValue(V, Instance)) | ||||||||
2237 | return VectorLoopValueMap.getScalarValue(V, Instance); | ||||||||
2238 | |||||||||
2239 | // If the value has not been scalarized, get its entry in VectorLoopValueMap | ||||||||
2240 | // for the given unroll part. If this entry is not a vector type (i.e., the | ||||||||
2241 | // vectorization factor is one), there is no need to generate an | ||||||||
2242 | // extractelement instruction. | ||||||||
2243 | auto *U = getOrCreateVectorValue(V, Instance.Part); | ||||||||
2244 | if (!U->getType()->isVectorTy()) { | ||||||||
2245 | assert(VF == 1 && "Value not scalarized has non-vector type")((VF == 1 && "Value not scalarized has non-vector type" ) ? static_cast<void> (0) : __assert_fail ("VF == 1 && \"Value not scalarized has non-vector type\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2245, __PRETTY_FUNCTION__)); | ||||||||
2246 | return U; | ||||||||
2247 | } | ||||||||
2248 | |||||||||
2249 | // Otherwise, the value from the original loop has been vectorized and is | ||||||||
2250 | // represented by UF vector values. Extract and return the requested scalar | ||||||||
2251 | // value from the appropriate vector lane. | ||||||||
2252 | return Builder.CreateExtractElement(U, Builder.getInt32(Instance.Lane)); | ||||||||
2253 | } | ||||||||
2254 | |||||||||
2255 | void InnerLoopVectorizer::packScalarIntoVectorValue( | ||||||||
2256 | Value *V, const VPIteration &Instance) { | ||||||||
2257 | assert(V != Induction && "The new induction variable should not be used.")((V != Induction && "The new induction variable should not be used." ) ? static_cast<void> (0) : __assert_fail ("V != Induction && \"The new induction variable should not be used.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2257, __PRETTY_FUNCTION__)); | ||||||||
2258 | assert(!V->getType()->isVectorTy() && "Can't pack a vector")((!V->getType()->isVectorTy() && "Can't pack a vector" ) ? static_cast<void> (0) : __assert_fail ("!V->getType()->isVectorTy() && \"Can't pack a vector\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2258, __PRETTY_FUNCTION__)); | ||||||||
2259 | assert(!V->getType()->isVoidTy() && "Type does not produce a value")((!V->getType()->isVoidTy() && "Type does not produce a value" ) ? static_cast<void> (0) : __assert_fail ("!V->getType()->isVoidTy() && \"Type does not produce a value\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2259, __PRETTY_FUNCTION__)); | ||||||||
2260 | |||||||||
2261 | Value *ScalarInst = VectorLoopValueMap.getScalarValue(V, Instance); | ||||||||
2262 | Value *VectorValue = VectorLoopValueMap.getVectorValue(V, Instance.Part); | ||||||||
2263 | VectorValue = Builder.CreateInsertElement(VectorValue, ScalarInst, | ||||||||
2264 | Builder.getInt32(Instance.Lane)); | ||||||||
2265 | VectorLoopValueMap.resetVectorValue(V, Instance.Part, VectorValue); | ||||||||
2266 | } | ||||||||
2267 | |||||||||
2268 | Value *InnerLoopVectorizer::reverseVector(Value *Vec) { | ||||||||
2269 | assert(Vec->getType()->isVectorTy() && "Invalid type")((Vec->getType()->isVectorTy() && "Invalid type" ) ? static_cast<void> (0) : __assert_fail ("Vec->getType()->isVectorTy() && \"Invalid type\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2269, __PRETTY_FUNCTION__)); | ||||||||
2270 | assert(!VF.isScalable() && "Cannot reverse scalable vectors")((!VF.isScalable() && "Cannot reverse scalable vectors" ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"Cannot reverse scalable vectors\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2270, __PRETTY_FUNCTION__)); | ||||||||
2271 | SmallVector<int, 8> ShuffleMask; | ||||||||
2272 | for (unsigned i = 0; i < VF.getKnownMinValue(); ++i) | ||||||||
2273 | ShuffleMask.push_back(VF.getKnownMinValue() - i - 1); | ||||||||
2274 | |||||||||
2275 | return Builder.CreateShuffleVector(Vec, ShuffleMask, "reverse"); | ||||||||
2276 | } | ||||||||
2277 | |||||||||
2278 | // Return whether we allow using masked interleave-groups (for dealing with | ||||||||
2279 | // strided loads/stores that reside in predicated blocks, or for dealing | ||||||||
2280 | // with gaps). | ||||||||
2281 | static bool useMaskedInterleavedAccesses(const TargetTransformInfo &TTI) { | ||||||||
2282 | // If an override option has been passed in for interleaved accesses, use it. | ||||||||
2283 | if (EnableMaskedInterleavedMemAccesses.getNumOccurrences() > 0) | ||||||||
2284 | return EnableMaskedInterleavedMemAccesses; | ||||||||
2285 | |||||||||
2286 | return TTI.enableMaskedInterleavedAccessVectorization(); | ||||||||
2287 | } | ||||||||
2288 | |||||||||
2289 | // Try to vectorize the interleave group that \p Instr belongs to. | ||||||||
2290 | // | ||||||||
2291 | // E.g. Translate following interleaved load group (factor = 3): | ||||||||
2292 | // for (i = 0; i < N; i+=3) { | ||||||||
2293 | // R = Pic[i]; // Member of index 0 | ||||||||
2294 | // G = Pic[i+1]; // Member of index 1 | ||||||||
2295 | // B = Pic[i+2]; // Member of index 2 | ||||||||
2296 | // ... // do something to R, G, B | ||||||||
2297 | // } | ||||||||
2298 | // To: | ||||||||
2299 | // %wide.vec = load <12 x i32> ; Read 4 tuples of R,G,B | ||||||||
2300 | // %R.vec = shuffle %wide.vec, undef, <0, 3, 6, 9> ; R elements | ||||||||
2301 | // %G.vec = shuffle %wide.vec, undef, <1, 4, 7, 10> ; G elements | ||||||||
2302 | // %B.vec = shuffle %wide.vec, undef, <2, 5, 8, 11> ; B elements | ||||||||
2303 | // | ||||||||
2304 | // Or translate following interleaved store group (factor = 3): | ||||||||
2305 | // for (i = 0; i < N; i+=3) { | ||||||||
2306 | // ... do something to R, G, B | ||||||||
2307 | // Pic[i] = R; // Member of index 0 | ||||||||
2308 | // Pic[i+1] = G; // Member of index 1 | ||||||||
2309 | // Pic[i+2] = B; // Member of index 2 | ||||||||
2310 | // } | ||||||||
2311 | // To: | ||||||||
2312 | // %R_G.vec = shuffle %R.vec, %G.vec, <0, 1, 2, ..., 7> | ||||||||
2313 | // %B_U.vec = shuffle %B.vec, undef, <0, 1, 2, 3, u, u, u, u> | ||||||||
2314 | // %interleaved.vec = shuffle %R_G.vec, %B_U.vec, | ||||||||
2315 | // <0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11> ; Interleave R,G,B elements | ||||||||
2316 | // store <12 x i32> %interleaved.vec ; Write 4 tuples of R,G,B | ||||||||
2317 | void InnerLoopVectorizer::vectorizeInterleaveGroup( | ||||||||
2318 | const InterleaveGroup<Instruction> *Group, VPTransformState &State, | ||||||||
2319 | VPValue *Addr, VPValue *BlockInMask) { | ||||||||
2320 | Instruction *Instr = Group->getInsertPos(); | ||||||||
2321 | const DataLayout &DL = Instr->getModule()->getDataLayout(); | ||||||||
2322 | |||||||||
2323 | // Prepare for the vector type of the interleaved load/store. | ||||||||
2324 | Type *ScalarTy = getMemInstValueType(Instr); | ||||||||
2325 | unsigned InterleaveFactor = Group->getFactor(); | ||||||||
2326 | assert(!VF.isScalable() && "scalable vectors not yet supported.")((!VF.isScalable() && "scalable vectors not yet supported." ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"scalable vectors not yet supported.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2326, __PRETTY_FUNCTION__)); | ||||||||
2327 | auto *VecTy = VectorType::get(ScalarTy, VF * InterleaveFactor); | ||||||||
2328 | |||||||||
2329 | // Prepare for the new pointers. | ||||||||
2330 | SmallVector<Value *, 2> AddrParts; | ||||||||
2331 | unsigned Index = Group->getIndex(Instr); | ||||||||
2332 | |||||||||
2333 | // TODO: extend the masked interleaved-group support to reversed access. | ||||||||
2334 | assert((!BlockInMask || !Group->isReverse()) &&(((!BlockInMask || !Group->isReverse()) && "Reversed masked interleave-group not supported." ) ? static_cast<void> (0) : __assert_fail ("(!BlockInMask || !Group->isReverse()) && \"Reversed masked interleave-group not supported.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2335, __PRETTY_FUNCTION__)) | ||||||||
2335 | "Reversed masked interleave-group not supported.")(((!BlockInMask || !Group->isReverse()) && "Reversed masked interleave-group not supported." ) ? static_cast<void> (0) : __assert_fail ("(!BlockInMask || !Group->isReverse()) && \"Reversed masked interleave-group not supported.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2335, __PRETTY_FUNCTION__)); | ||||||||
2336 | |||||||||
2337 | // If the group is reverse, adjust the index to refer to the last vector lane | ||||||||
2338 | // instead of the first. We adjust the index from the first vector lane, | ||||||||
2339 | // rather than directly getting the pointer for lane VF - 1, because the | ||||||||
2340 | // pointer operand of the interleaved access is supposed to be uniform. For | ||||||||
2341 | // uniform instructions, we're only required to generate a value for the | ||||||||
2342 | // first vector lane in each unroll iteration. | ||||||||
2343 | assert(!VF.isScalable() &&((!VF.isScalable() && "scalable vector reverse operation is not implemented" ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"scalable vector reverse operation is not implemented\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2344, __PRETTY_FUNCTION__)) | ||||||||
2344 | "scalable vector reverse operation is not implemented")((!VF.isScalable() && "scalable vector reverse operation is not implemented" ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"scalable vector reverse operation is not implemented\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2344, __PRETTY_FUNCTION__)); | ||||||||
2345 | if (Group->isReverse()) | ||||||||
2346 | Index += (VF.getKnownMinValue() - 1) * Group->getFactor(); | ||||||||
2347 | |||||||||
2348 | for (unsigned Part = 0; Part < UF; Part++) { | ||||||||
2349 | Value *AddrPart = State.get(Addr, {Part, 0}); | ||||||||
2350 | setDebugLocFromInst(Builder, AddrPart); | ||||||||
2351 | |||||||||
2352 | // Notice current instruction could be any index. Need to adjust the address | ||||||||
2353 | // to the member of index 0. | ||||||||
2354 | // | ||||||||
2355 | // E.g. a = A[i+1]; // Member of index 1 (Current instruction) | ||||||||
2356 | // b = A[i]; // Member of index 0 | ||||||||
2357 | // Current pointer is pointed to A[i+1], adjust it to A[i]. | ||||||||
2358 | // | ||||||||
2359 | // E.g. A[i+1] = a; // Member of index 1 | ||||||||
2360 | // A[i] = b; // Member of index 0 | ||||||||
2361 | // A[i+2] = c; // Member of index 2 (Current instruction) | ||||||||
2362 | // Current pointer is pointed to A[i+2], adjust it to A[i]. | ||||||||
2363 | |||||||||
2364 | bool InBounds = false; | ||||||||
2365 | if (auto *gep = dyn_cast<GetElementPtrInst>(AddrPart->stripPointerCasts())) | ||||||||
2366 | InBounds = gep->isInBounds(); | ||||||||
2367 | AddrPart = Builder.CreateGEP(ScalarTy, AddrPart, Builder.getInt32(-Index)); | ||||||||
2368 | cast<GetElementPtrInst>(AddrPart)->setIsInBounds(InBounds); | ||||||||
2369 | |||||||||
2370 | // Cast to the vector pointer type. | ||||||||
2371 | unsigned AddressSpace = AddrPart->getType()->getPointerAddressSpace(); | ||||||||
2372 | Type *PtrTy = VecTy->getPointerTo(AddressSpace); | ||||||||
2373 | AddrParts.push_back(Builder.CreateBitCast(AddrPart, PtrTy)); | ||||||||
2374 | } | ||||||||
2375 | |||||||||
2376 | setDebugLocFromInst(Builder, Instr); | ||||||||
2377 | Value *UndefVec = UndefValue::get(VecTy); | ||||||||
2378 | |||||||||
2379 | Value *MaskForGaps = nullptr; | ||||||||
2380 | if (Group->requiresScalarEpilogue() && !Cost->isScalarEpilogueAllowed()) { | ||||||||
2381 | assert(!VF.isScalable() && "scalable vectors not yet supported.")((!VF.isScalable() && "scalable vectors not yet supported." ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"scalable vectors not yet supported.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2381, __PRETTY_FUNCTION__)); | ||||||||
2382 | MaskForGaps = createBitMaskForGaps(Builder, VF.getKnownMinValue(), *Group); | ||||||||
2383 | assert(MaskForGaps && "Mask for Gaps is required but it is null")((MaskForGaps && "Mask for Gaps is required but it is null" ) ? static_cast<void> (0) : __assert_fail ("MaskForGaps && \"Mask for Gaps is required but it is null\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2383, __PRETTY_FUNCTION__)); | ||||||||
2384 | } | ||||||||
2385 | |||||||||
2386 | // Vectorize the interleaved load group. | ||||||||
2387 | if (isa<LoadInst>(Instr)) { | ||||||||
2388 | // For each unroll part, create a wide load for the group. | ||||||||
2389 | SmallVector<Value *, 2> NewLoads; | ||||||||
2390 | for (unsigned Part = 0; Part < UF; Part++) { | ||||||||
2391 | Instruction *NewLoad; | ||||||||
2392 | if (BlockInMask || MaskForGaps) { | ||||||||
2393 | assert(useMaskedInterleavedAccesses(*TTI) &&((useMaskedInterleavedAccesses(*TTI) && "masked interleaved groups are not allowed." ) ? static_cast<void> (0) : __assert_fail ("useMaskedInterleavedAccesses(*TTI) && \"masked interleaved groups are not allowed.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2394, __PRETTY_FUNCTION__)) | ||||||||
2394 | "masked interleaved groups are not allowed.")((useMaskedInterleavedAccesses(*TTI) && "masked interleaved groups are not allowed." ) ? static_cast<void> (0) : __assert_fail ("useMaskedInterleavedAccesses(*TTI) && \"masked interleaved groups are not allowed.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2394, __PRETTY_FUNCTION__)); | ||||||||
2395 | Value *GroupMask = MaskForGaps; | ||||||||
2396 | if (BlockInMask) { | ||||||||
2397 | Value *BlockInMaskPart = State.get(BlockInMask, Part); | ||||||||
2398 | assert(!VF.isScalable() && "scalable vectors not yet supported.")((!VF.isScalable() && "scalable vectors not yet supported." ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"scalable vectors not yet supported.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2398, __PRETTY_FUNCTION__)); | ||||||||
2399 | Value *ShuffledMask = Builder.CreateShuffleVector( | ||||||||
2400 | BlockInMaskPart, | ||||||||
2401 | createReplicatedMask(InterleaveFactor, VF.getKnownMinValue()), | ||||||||
2402 | "interleaved.mask"); | ||||||||
2403 | GroupMask = MaskForGaps | ||||||||
2404 | ? Builder.CreateBinOp(Instruction::And, ShuffledMask, | ||||||||
2405 | MaskForGaps) | ||||||||
2406 | : ShuffledMask; | ||||||||
2407 | } | ||||||||
2408 | NewLoad = | ||||||||
2409 | Builder.CreateMaskedLoad(AddrParts[Part], Group->getAlign(), | ||||||||
2410 | GroupMask, UndefVec, "wide.masked.vec"); | ||||||||
2411 | } | ||||||||
2412 | else | ||||||||
2413 | NewLoad = Builder.CreateAlignedLoad(VecTy, AddrParts[Part], | ||||||||
2414 | Group->getAlign(), "wide.vec"); | ||||||||
2415 | Group->addMetadata(NewLoad); | ||||||||
2416 | NewLoads.push_back(NewLoad); | ||||||||
2417 | } | ||||||||
2418 | |||||||||
2419 | // For each member in the group, shuffle out the appropriate data from the | ||||||||
2420 | // wide loads. | ||||||||
2421 | for (unsigned I = 0; I < InterleaveFactor; ++I) { | ||||||||
2422 | Instruction *Member = Group->getMember(I); | ||||||||
2423 | |||||||||
2424 | // Skip the gaps in the group. | ||||||||
2425 | if (!Member) | ||||||||
2426 | continue; | ||||||||
2427 | |||||||||
2428 | assert(!VF.isScalable() && "scalable vectors not yet supported.")((!VF.isScalable() && "scalable vectors not yet supported." ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"scalable vectors not yet supported.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2428, __PRETTY_FUNCTION__)); | ||||||||
2429 | auto StrideMask = | ||||||||
2430 | createStrideMask(I, InterleaveFactor, VF.getKnownMinValue()); | ||||||||
2431 | for (unsigned Part = 0; Part < UF; Part++) { | ||||||||
2432 | Value *StridedVec = Builder.CreateShuffleVector( | ||||||||
2433 | NewLoads[Part], StrideMask, "strided.vec"); | ||||||||
2434 | |||||||||
2435 | // If this member has different type, cast the result type. | ||||||||
2436 | if (Member->getType() != ScalarTy) { | ||||||||
2437 | assert(!VF.isScalable() && "VF is assumed to be non scalable.")((!VF.isScalable() && "VF is assumed to be non scalable." ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"VF is assumed to be non scalable.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2437, __PRETTY_FUNCTION__)); | ||||||||
2438 | VectorType *OtherVTy = VectorType::get(Member->getType(), VF); | ||||||||
2439 | StridedVec = createBitOrPointerCast(StridedVec, OtherVTy, DL); | ||||||||
2440 | } | ||||||||
2441 | |||||||||
2442 | if (Group->isReverse()) | ||||||||
2443 | StridedVec = reverseVector(StridedVec); | ||||||||
2444 | |||||||||
2445 | VectorLoopValueMap.setVectorValue(Member, Part, StridedVec); | ||||||||
2446 | } | ||||||||
2447 | } | ||||||||
2448 | return; | ||||||||
2449 | } | ||||||||
2450 | |||||||||
2451 | // The sub vector type for current instruction. | ||||||||
2452 | assert(!VF.isScalable() && "VF is assumed to be non scalable.")((!VF.isScalable() && "VF is assumed to be non scalable." ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"VF is assumed to be non scalable.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2452, __PRETTY_FUNCTION__)); | ||||||||
2453 | auto *SubVT = VectorType::get(ScalarTy, VF); | ||||||||
2454 | |||||||||
2455 | // Vectorize the interleaved store group. | ||||||||
2456 | for (unsigned Part = 0; Part < UF; Part++) { | ||||||||
2457 | // Collect the stored vector from each member. | ||||||||
2458 | SmallVector<Value *, 4> StoredVecs; | ||||||||
2459 | for (unsigned i = 0; i < InterleaveFactor; i++) { | ||||||||
2460 | // Interleaved store group doesn't allow a gap, so each index has a member | ||||||||
2461 | Instruction *Member = Group->getMember(i); | ||||||||
2462 | assert(Member && "Fail to get a member from an interleaved store group")((Member && "Fail to get a member from an interleaved store group" ) ? static_cast<void> (0) : __assert_fail ("Member && \"Fail to get a member from an interleaved store group\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2462, __PRETTY_FUNCTION__)); | ||||||||
2463 | |||||||||
2464 | Value *StoredVec = getOrCreateVectorValue( | ||||||||
2465 | cast<StoreInst>(Member)->getValueOperand(), Part); | ||||||||
2466 | if (Group->isReverse()) | ||||||||
2467 | StoredVec = reverseVector(StoredVec); | ||||||||
2468 | |||||||||
2469 | // If this member has different type, cast it to a unified type. | ||||||||
2470 | |||||||||
2471 | if (StoredVec->getType() != SubVT) | ||||||||
2472 | StoredVec = createBitOrPointerCast(StoredVec, SubVT, DL); | ||||||||
2473 | |||||||||
2474 | StoredVecs.push_back(StoredVec); | ||||||||
2475 | } | ||||||||
2476 | |||||||||
2477 | // Concatenate all vectors into a wide vector. | ||||||||
2478 | Value *WideVec = concatenateVectors(Builder, StoredVecs); | ||||||||
2479 | |||||||||
2480 | // Interleave the elements in the wide vector. | ||||||||
2481 | assert(!VF.isScalable() && "scalable vectors not yet supported.")((!VF.isScalable() && "scalable vectors not yet supported." ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"scalable vectors not yet supported.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2481, __PRETTY_FUNCTION__)); | ||||||||
2482 | Value *IVec = Builder.CreateShuffleVector( | ||||||||
2483 | WideVec, createInterleaveMask(VF.getKnownMinValue(), InterleaveFactor), | ||||||||
2484 | "interleaved.vec"); | ||||||||
2485 | |||||||||
2486 | Instruction *NewStoreInstr; | ||||||||
2487 | if (BlockInMask) { | ||||||||
2488 | Value *BlockInMaskPart = State.get(BlockInMask, Part); | ||||||||
2489 | Value *ShuffledMask = Builder.CreateShuffleVector( | ||||||||
2490 | BlockInMaskPart, | ||||||||
2491 | createReplicatedMask(InterleaveFactor, VF.getKnownMinValue()), | ||||||||
2492 | "interleaved.mask"); | ||||||||
2493 | NewStoreInstr = Builder.CreateMaskedStore( | ||||||||
2494 | IVec, AddrParts[Part], Group->getAlign(), ShuffledMask); | ||||||||
2495 | } | ||||||||
2496 | else | ||||||||
2497 | NewStoreInstr = | ||||||||
2498 | Builder.CreateAlignedStore(IVec, AddrParts[Part], Group->getAlign()); | ||||||||
2499 | |||||||||
2500 | Group->addMetadata(NewStoreInstr); | ||||||||
2501 | } | ||||||||
2502 | } | ||||||||
2503 | |||||||||
2504 | void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr, | ||||||||
2505 | VPTransformState &State, | ||||||||
2506 | VPValue *Addr, | ||||||||
2507 | VPValue *StoredValue, | ||||||||
2508 | VPValue *BlockInMask) { | ||||||||
2509 | // Attempt to issue a wide load. | ||||||||
2510 | LoadInst *LI = dyn_cast<LoadInst>(Instr); | ||||||||
2511 | StoreInst *SI = dyn_cast<StoreInst>(Instr); | ||||||||
2512 | |||||||||
2513 | assert((LI || SI) && "Invalid Load/Store instruction")(((LI || SI) && "Invalid Load/Store instruction") ? static_cast <void> (0) : __assert_fail ("(LI || SI) && \"Invalid Load/Store instruction\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2513, __PRETTY_FUNCTION__)); | ||||||||
2514 | assert((!SI || StoredValue) && "No stored value provided for widened store")(((!SI || StoredValue) && "No stored value provided for widened store" ) ? static_cast<void> (0) : __assert_fail ("(!SI || StoredValue) && \"No stored value provided for widened store\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2514, __PRETTY_FUNCTION__)); | ||||||||
2515 | assert((!LI || !StoredValue) && "Stored value provided for widened load")(((!LI || !StoredValue) && "Stored value provided for widened load" ) ? static_cast<void> (0) : __assert_fail ("(!LI || !StoredValue) && \"Stored value provided for widened load\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2515, __PRETTY_FUNCTION__)); | ||||||||
2516 | |||||||||
2517 | LoopVectorizationCostModel::InstWidening Decision = | ||||||||
2518 | Cost->getWideningDecision(Instr, VF); | ||||||||
2519 | assert((Decision == LoopVectorizationCostModel::CM_Widen ||(((Decision == LoopVectorizationCostModel::CM_Widen || Decision == LoopVectorizationCostModel::CM_Widen_Reverse || Decision == LoopVectorizationCostModel::CM_GatherScatter) && "CM decision is not to widen the memory instruction" ) ? static_cast<void> (0) : __assert_fail ("(Decision == LoopVectorizationCostModel::CM_Widen || Decision == LoopVectorizationCostModel::CM_Widen_Reverse || Decision == LoopVectorizationCostModel::CM_GatherScatter) && \"CM decision is not to widen the memory instruction\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2522, __PRETTY_FUNCTION__)) | ||||||||
2520 | Decision == LoopVectorizationCostModel::CM_Widen_Reverse ||(((Decision == LoopVectorizationCostModel::CM_Widen || Decision == LoopVectorizationCostModel::CM_Widen_Reverse || Decision == LoopVectorizationCostModel::CM_GatherScatter) && "CM decision is not to widen the memory instruction" ) ? static_cast<void> (0) : __assert_fail ("(Decision == LoopVectorizationCostModel::CM_Widen || Decision == LoopVectorizationCostModel::CM_Widen_Reverse || Decision == LoopVectorizationCostModel::CM_GatherScatter) && \"CM decision is not to widen the memory instruction\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2522, __PRETTY_FUNCTION__)) | ||||||||
2521 | Decision == LoopVectorizationCostModel::CM_GatherScatter) &&(((Decision == LoopVectorizationCostModel::CM_Widen || Decision == LoopVectorizationCostModel::CM_Widen_Reverse || Decision == LoopVectorizationCostModel::CM_GatherScatter) && "CM decision is not to widen the memory instruction" ) ? static_cast<void> (0) : __assert_fail ("(Decision == LoopVectorizationCostModel::CM_Widen || Decision == LoopVectorizationCostModel::CM_Widen_Reverse || Decision == LoopVectorizationCostModel::CM_GatherScatter) && \"CM decision is not to widen the memory instruction\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2522, __PRETTY_FUNCTION__)) | ||||||||
2522 | "CM decision is not to widen the memory instruction")(((Decision == LoopVectorizationCostModel::CM_Widen || Decision == LoopVectorizationCostModel::CM_Widen_Reverse || Decision == LoopVectorizationCostModel::CM_GatherScatter) && "CM decision is not to widen the memory instruction" ) ? static_cast<void> (0) : __assert_fail ("(Decision == LoopVectorizationCostModel::CM_Widen || Decision == LoopVectorizationCostModel::CM_Widen_Reverse || Decision == LoopVectorizationCostModel::CM_GatherScatter) && \"CM decision is not to widen the memory instruction\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2522, __PRETTY_FUNCTION__)); | ||||||||
2523 | |||||||||
2524 | Type *ScalarDataTy = getMemInstValueType(Instr); | ||||||||
2525 | |||||||||
2526 | assert(!VF.isScalable() && "scalable vectors not yet supported.")((!VF.isScalable() && "scalable vectors not yet supported." ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"scalable vectors not yet supported.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2526, __PRETTY_FUNCTION__)); | ||||||||
2527 | auto *DataTy = VectorType::get(ScalarDataTy, VF); | ||||||||
2528 | const Align Alignment = getLoadStoreAlignment(Instr); | ||||||||
2529 | |||||||||
2530 | // Determine if the pointer operand of the access is either consecutive or | ||||||||
2531 | // reverse consecutive. | ||||||||
2532 | bool Reverse = (Decision == LoopVectorizationCostModel::CM_Widen_Reverse); | ||||||||
2533 | bool ConsecutiveStride = | ||||||||
2534 | Reverse || (Decision == LoopVectorizationCostModel::CM_Widen); | ||||||||
2535 | bool CreateGatherScatter = | ||||||||
2536 | (Decision == LoopVectorizationCostModel::CM_GatherScatter); | ||||||||
2537 | |||||||||
2538 | // Either Ptr feeds a vector load/store, or a vector GEP should feed a vector | ||||||||
2539 | // gather/scatter. Otherwise Decision should have been to Scalarize. | ||||||||
2540 | assert((ConsecutiveStride || CreateGatherScatter) &&(((ConsecutiveStride || CreateGatherScatter) && "The instruction should be scalarized" ) ? static_cast<void> (0) : __assert_fail ("(ConsecutiveStride || CreateGatherScatter) && \"The instruction should be scalarized\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2541, __PRETTY_FUNCTION__)) | ||||||||
2541 | "The instruction should be scalarized")(((ConsecutiveStride || CreateGatherScatter) && "The instruction should be scalarized" ) ? static_cast<void> (0) : __assert_fail ("(ConsecutiveStride || CreateGatherScatter) && \"The instruction should be scalarized\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2541, __PRETTY_FUNCTION__)); | ||||||||
2542 | (void)ConsecutiveStride; | ||||||||
2543 | |||||||||
2544 | VectorParts BlockInMaskParts(UF); | ||||||||
2545 | bool isMaskRequired = BlockInMask; | ||||||||
2546 | if (isMaskRequired) | ||||||||
2547 | for (unsigned Part = 0; Part < UF; ++Part) | ||||||||
2548 | BlockInMaskParts[Part] = State.get(BlockInMask, Part); | ||||||||
2549 | |||||||||
2550 | const auto CreateVecPtr = [&](unsigned Part, Value *Ptr) -> Value * { | ||||||||
2551 | // Calculate the pointer for the specific unroll-part. | ||||||||
2552 | GetElementPtrInst *PartPtr = nullptr; | ||||||||
2553 | |||||||||
2554 | bool InBounds = false; | ||||||||
2555 | if (auto *gep = dyn_cast<GetElementPtrInst>(Ptr->stripPointerCasts())) | ||||||||
2556 | InBounds = gep->isInBounds(); | ||||||||
2557 | |||||||||
2558 | if (Reverse) { | ||||||||
2559 | // If the address is consecutive but reversed, then the | ||||||||
2560 | // wide store needs to start at the last vector element. | ||||||||
2561 | PartPtr = cast<GetElementPtrInst>(Builder.CreateGEP( | ||||||||
2562 | ScalarDataTy, Ptr, Builder.getInt32(-Part * VF.getKnownMinValue()))); | ||||||||
2563 | PartPtr->setIsInBounds(InBounds); | ||||||||
2564 | PartPtr = cast<GetElementPtrInst>(Builder.CreateGEP( | ||||||||
2565 | ScalarDataTy, PartPtr, Builder.getInt32(1 - VF.getKnownMinValue()))); | ||||||||
2566 | PartPtr->setIsInBounds(InBounds); | ||||||||
2567 | if (isMaskRequired) // Reverse of a null all-one mask is a null mask. | ||||||||
2568 | BlockInMaskParts[Part] = reverseVector(BlockInMaskParts[Part]); | ||||||||
2569 | } else { | ||||||||
2570 | PartPtr = cast<GetElementPtrInst>(Builder.CreateGEP( | ||||||||
2571 | ScalarDataTy, Ptr, Builder.getInt32(Part * VF.getKnownMinValue()))); | ||||||||
2572 | PartPtr->setIsInBounds(InBounds); | ||||||||
2573 | } | ||||||||
2574 | |||||||||
2575 | unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace(); | ||||||||
2576 | return Builder.CreateBitCast(PartPtr, DataTy->getPointerTo(AddressSpace)); | ||||||||
2577 | }; | ||||||||
2578 | |||||||||
2579 | // Handle Stores: | ||||||||
2580 | if (SI) { | ||||||||
2581 | setDebugLocFromInst(Builder, SI); | ||||||||
2582 | |||||||||
2583 | for (unsigned Part = 0; Part < UF; ++Part) { | ||||||||
2584 | Instruction *NewSI = nullptr; | ||||||||
2585 | Value *StoredVal = State.get(StoredValue, Part); | ||||||||
2586 | if (CreateGatherScatter) { | ||||||||
2587 | Value *MaskPart = isMaskRequired ? BlockInMaskParts[Part] : nullptr; | ||||||||
2588 | Value *VectorGep = State.get(Addr, Part); | ||||||||
2589 | NewSI = Builder.CreateMaskedScatter(StoredVal, VectorGep, Alignment, | ||||||||
2590 | MaskPart); | ||||||||
2591 | } else { | ||||||||
2592 | if (Reverse) { | ||||||||
2593 | // If we store to reverse consecutive memory locations, then we need | ||||||||
2594 | // to reverse the order of elements in the stored value. | ||||||||
2595 | StoredVal = reverseVector(StoredVal); | ||||||||
2596 | // We don't want to update the value in the map as it might be used in | ||||||||
2597 | // another expression. So don't call resetVectorValue(StoredVal). | ||||||||
2598 | } | ||||||||
2599 | auto *VecPtr = CreateVecPtr(Part, State.get(Addr, {0, 0})); | ||||||||
2600 | if (isMaskRequired) | ||||||||
2601 | NewSI = Builder.CreateMaskedStore(StoredVal, VecPtr, Alignment, | ||||||||
2602 | BlockInMaskParts[Part]); | ||||||||
2603 | else | ||||||||
2604 | NewSI = Builder.CreateAlignedStore(StoredVal, VecPtr, Alignment); | ||||||||
2605 | } | ||||||||
2606 | addMetadata(NewSI, SI); | ||||||||
2607 | } | ||||||||
2608 | return; | ||||||||
2609 | } | ||||||||
2610 | |||||||||
2611 | // Handle loads. | ||||||||
2612 | assert(LI && "Must have a load instruction")((LI && "Must have a load instruction") ? static_cast <void> (0) : __assert_fail ("LI && \"Must have a load instruction\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2612, __PRETTY_FUNCTION__)); | ||||||||
2613 | setDebugLocFromInst(Builder, LI); | ||||||||
2614 | for (unsigned Part = 0; Part < UF; ++Part) { | ||||||||
2615 | Value *NewLI; | ||||||||
2616 | if (CreateGatherScatter) { | ||||||||
2617 | Value *MaskPart = isMaskRequired ? BlockInMaskParts[Part] : nullptr; | ||||||||
2618 | Value *VectorGep = State.get(Addr, Part); | ||||||||
2619 | NewLI = Builder.CreateMaskedGather(VectorGep, Alignment, MaskPart, | ||||||||
2620 | nullptr, "wide.masked.gather"); | ||||||||
2621 | addMetadata(NewLI, LI); | ||||||||
2622 | } else { | ||||||||
2623 | auto *VecPtr = CreateVecPtr(Part, State.get(Addr, {0, 0})); | ||||||||
2624 | if (isMaskRequired) | ||||||||
2625 | NewLI = Builder.CreateMaskedLoad( | ||||||||
2626 | VecPtr, Alignment, BlockInMaskParts[Part], UndefValue::get(DataTy), | ||||||||
2627 | "wide.masked.load"); | ||||||||
2628 | else | ||||||||
2629 | NewLI = | ||||||||
2630 | Builder.CreateAlignedLoad(DataTy, VecPtr, Alignment, "wide.load"); | ||||||||
2631 | |||||||||
2632 | // Add metadata to the load, but setVectorValue to the reverse shuffle. | ||||||||
2633 | addMetadata(NewLI, LI); | ||||||||
2634 | if (Reverse) | ||||||||
2635 | NewLI = reverseVector(NewLI); | ||||||||
2636 | } | ||||||||
2637 | VectorLoopValueMap.setVectorValue(Instr, Part, NewLI); | ||||||||
2638 | } | ||||||||
2639 | } | ||||||||
2640 | |||||||||
2641 | void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, VPUser &User, | ||||||||
2642 | const VPIteration &Instance, | ||||||||
2643 | bool IfPredicateInstr, | ||||||||
2644 | VPTransformState &State) { | ||||||||
2645 | assert(!Instr->getType()->isAggregateType() && "Can't handle vectors")((!Instr->getType()->isAggregateType() && "Can't handle vectors" ) ? static_cast<void> (0) : __assert_fail ("!Instr->getType()->isAggregateType() && \"Can't handle vectors\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2645, __PRETTY_FUNCTION__)); | ||||||||
2646 | |||||||||
2647 | setDebugLocFromInst(Builder, Instr); | ||||||||
2648 | |||||||||
2649 | // Does this instruction return a value ? | ||||||||
2650 | bool IsVoidRetTy = Instr->getType()->isVoidTy(); | ||||||||
2651 | |||||||||
2652 | Instruction *Cloned = Instr->clone(); | ||||||||
2653 | if (!IsVoidRetTy) | ||||||||
2654 | Cloned->setName(Instr->getName() + ".cloned"); | ||||||||
2655 | |||||||||
2656 | // Replace the operands of the cloned instructions with their scalar | ||||||||
2657 | // equivalents in the new loop. | ||||||||
2658 | for (unsigned op = 0, e = User.getNumOperands(); op != e; ++op) { | ||||||||
2659 | auto *NewOp = State.get(User.getOperand(op), Instance); | ||||||||
2660 | Cloned->setOperand(op, NewOp); | ||||||||
2661 | } | ||||||||
2662 | addNewMetadata(Cloned, Instr); | ||||||||
2663 | |||||||||
2664 | // Place the cloned scalar in the new loop. | ||||||||
2665 | Builder.Insert(Cloned); | ||||||||
2666 | |||||||||
2667 | // Add the cloned scalar to the scalar map entry. | ||||||||
2668 | VectorLoopValueMap.setScalarValue(Instr, Instance, Cloned); | ||||||||
2669 | |||||||||
2670 | // If we just cloned a new assumption, add it the assumption cache. | ||||||||
2671 | if (auto *II = dyn_cast<IntrinsicInst>(Cloned)) | ||||||||
2672 | if (II->getIntrinsicID() == Intrinsic::assume) | ||||||||
2673 | AC->registerAssumption(II); | ||||||||
2674 | |||||||||
2675 | // End if-block. | ||||||||
2676 | if (IfPredicateInstr) | ||||||||
2677 | PredicatedInstructions.push_back(Cloned); | ||||||||
2678 | } | ||||||||
2679 | |||||||||
2680 | PHINode *InnerLoopVectorizer::createInductionVariable(Loop *L, Value *Start, | ||||||||
2681 | Value *End, Value *Step, | ||||||||
2682 | Instruction *DL) { | ||||||||
2683 | BasicBlock *Header = L->getHeader(); | ||||||||
2684 | BasicBlock *Latch = L->getLoopLatch(); | ||||||||
2685 | // As we're just creating this loop, it's possible no latch exists | ||||||||
2686 | // yet. If so, use the header as this will be a single block loop. | ||||||||
2687 | if (!Latch) | ||||||||
2688 | Latch = Header; | ||||||||
2689 | |||||||||
2690 | IRBuilder<> Builder(&*Header->getFirstInsertionPt()); | ||||||||
2691 | Instruction *OldInst = getDebugLocFromInstOrOperands(OldInduction); | ||||||||
2692 | setDebugLocFromInst(Builder, OldInst); | ||||||||
2693 | auto *Induction = Builder.CreatePHI(Start->getType(), 2, "index"); | ||||||||
2694 | |||||||||
2695 | Builder.SetInsertPoint(Latch->getTerminator()); | ||||||||
2696 | setDebugLocFromInst(Builder, OldInst); | ||||||||
2697 | |||||||||
2698 | // Create i+1 and fill the PHINode. | ||||||||
2699 | Value *Next = Builder.CreateAdd(Induction, Step, "index.next"); | ||||||||
2700 | Induction->addIncoming(Start, L->getLoopPreheader()); | ||||||||
2701 | Induction->addIncoming(Next, Latch); | ||||||||
2702 | // Create the compare. | ||||||||
2703 | Value *ICmp = Builder.CreateICmpEQ(Next, End); | ||||||||
2704 | Builder.CreateCondBr(ICmp, L->getExitBlock(), Header); | ||||||||
2705 | |||||||||
2706 | // Now we have two terminators. Remove the old one from the block. | ||||||||
2707 | Latch->getTerminator()->eraseFromParent(); | ||||||||
2708 | |||||||||
2709 | return Induction; | ||||||||
2710 | } | ||||||||
2711 | |||||||||
2712 | Value *InnerLoopVectorizer::getOrCreateTripCount(Loop *L) { | ||||||||
2713 | if (TripCount) | ||||||||
2714 | return TripCount; | ||||||||
2715 | |||||||||
2716 | assert(L && "Create Trip Count for null loop.")((L && "Create Trip Count for null loop.") ? static_cast <void> (0) : __assert_fail ("L && \"Create Trip Count for null loop.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2716, __PRETTY_FUNCTION__)); | ||||||||
2717 | IRBuilder<> Builder(L->getLoopPreheader()->getTerminator()); | ||||||||
2718 | // Find the loop boundaries. | ||||||||
2719 | ScalarEvolution *SE = PSE.getSE(); | ||||||||
2720 | const SCEV *BackedgeTakenCount = PSE.getBackedgeTakenCount(); | ||||||||
2721 | assert(BackedgeTakenCount != SE->getCouldNotCompute() &&((BackedgeTakenCount != SE->getCouldNotCompute() && "Invalid loop count") ? static_cast<void> (0) : __assert_fail ("BackedgeTakenCount != SE->getCouldNotCompute() && \"Invalid loop count\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2722, __PRETTY_FUNCTION__)) | ||||||||
2722 | "Invalid loop count")((BackedgeTakenCount != SE->getCouldNotCompute() && "Invalid loop count") ? static_cast<void> (0) : __assert_fail ("BackedgeTakenCount != SE->getCouldNotCompute() && \"Invalid loop count\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2722, __PRETTY_FUNCTION__)); | ||||||||
2723 | |||||||||
2724 | Type *IdxTy = Legal->getWidestInductionType(); | ||||||||
2725 | assert(IdxTy && "No type for induction")((IdxTy && "No type for induction") ? static_cast< void> (0) : __assert_fail ("IdxTy && \"No type for induction\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2725, __PRETTY_FUNCTION__)); | ||||||||
2726 | |||||||||
2727 | // The exit count might have the type of i64 while the phi is i32. This can | ||||||||
2728 | // happen if we have an induction variable that is sign extended before the | ||||||||
2729 | // compare. The only way that we get a backedge taken count is that the | ||||||||
2730 | // induction variable was signed and as such will not overflow. In such a case | ||||||||
2731 | // truncation is legal. | ||||||||
2732 | if (SE->getTypeSizeInBits(BackedgeTakenCount->getType()) > | ||||||||
2733 | IdxTy->getPrimitiveSizeInBits()) | ||||||||
2734 | BackedgeTakenCount = SE->getTruncateOrNoop(BackedgeTakenCount, IdxTy); | ||||||||
2735 | BackedgeTakenCount = SE->getNoopOrZeroExtend(BackedgeTakenCount, IdxTy); | ||||||||
2736 | |||||||||
2737 | // Get the total trip count from the count by adding 1. | ||||||||
2738 | const SCEV *ExitCount = SE->getAddExpr( | ||||||||
2739 | BackedgeTakenCount, SE->getOne(BackedgeTakenCount->getType())); | ||||||||
2740 | |||||||||
2741 | const DataLayout &DL = L->getHeader()->getModule()->getDataLayout(); | ||||||||
2742 | |||||||||
2743 | // Expand the trip count and place the new instructions in the preheader. | ||||||||
2744 | // Notice that the pre-header does not change, only the loop body. | ||||||||
2745 | SCEVExpander Exp(*SE, DL, "induction"); | ||||||||
2746 | |||||||||
2747 | // Count holds the overall loop count (N). | ||||||||
2748 | TripCount = Exp.expandCodeFor(ExitCount, ExitCount->getType(), | ||||||||
2749 | L->getLoopPreheader()->getTerminator()); | ||||||||
2750 | |||||||||
2751 | if (TripCount->getType()->isPointerTy()) | ||||||||
2752 | TripCount = | ||||||||
2753 | CastInst::CreatePointerCast(TripCount, IdxTy, "exitcount.ptrcnt.to.int", | ||||||||
2754 | L->getLoopPreheader()->getTerminator()); | ||||||||
2755 | |||||||||
2756 | return TripCount; | ||||||||
2757 | } | ||||||||
2758 | |||||||||
2759 | Value *InnerLoopVectorizer::getOrCreateVectorTripCount(Loop *L) { | ||||||||
2760 | if (VectorTripCount) | ||||||||
2761 | return VectorTripCount; | ||||||||
2762 | |||||||||
2763 | Value *TC = getOrCreateTripCount(L); | ||||||||
2764 | IRBuilder<> Builder(L->getLoopPreheader()->getTerminator()); | ||||||||
2765 | |||||||||
2766 | Type *Ty = TC->getType(); | ||||||||
2767 | // This is where we can make the step a runtime constant. | ||||||||
2768 | assert(!VF.isScalable() && "scalable vectorization is not supported yet")((!VF.isScalable() && "scalable vectorization is not supported yet" ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"scalable vectorization is not supported yet\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2768, __PRETTY_FUNCTION__)); | ||||||||
2769 | Constant *Step = ConstantInt::get(Ty, VF.getKnownMinValue() * UF); | ||||||||
2770 | |||||||||
2771 | // If the tail is to be folded by masking, round the number of iterations N | ||||||||
2772 | // up to a multiple of Step instead of rounding down. This is done by first | ||||||||
2773 | // adding Step-1 and then rounding down. Note that it's ok if this addition | ||||||||
2774 | // overflows: the vector induction variable will eventually wrap to zero given | ||||||||
2775 | // that it starts at zero and its Step is a power of two; the loop will then | ||||||||
2776 | // exit, with the last early-exit vector comparison also producing all-true. | ||||||||
2777 | if (Cost->foldTailByMasking()) { | ||||||||
2778 | assert(isPowerOf2_32(VF.getKnownMinValue() * UF) &&((isPowerOf2_32(VF.getKnownMinValue() * UF) && "VF*UF must be a power of 2 when folding tail by masking" ) ? static_cast<void> (0) : __assert_fail ("isPowerOf2_32(VF.getKnownMinValue() * UF) && \"VF*UF must be a power of 2 when folding tail by masking\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2779, __PRETTY_FUNCTION__)) | ||||||||
2779 | "VF*UF must be a power of 2 when folding tail by masking")((isPowerOf2_32(VF.getKnownMinValue() * UF) && "VF*UF must be a power of 2 when folding tail by masking" ) ? static_cast<void> (0) : __assert_fail ("isPowerOf2_32(VF.getKnownMinValue() * UF) && \"VF*UF must be a power of 2 when folding tail by masking\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2779, __PRETTY_FUNCTION__)); | ||||||||
2780 | TC = Builder.CreateAdd( | ||||||||
2781 | TC, ConstantInt::get(Ty, VF.getKnownMinValue() * UF - 1), "n.rnd.up"); | ||||||||
2782 | } | ||||||||
2783 | |||||||||
2784 | // Now we need to generate the expression for the part of the loop that the | ||||||||
2785 | // vectorized body will execute. This is equal to N - (N % Step) if scalar | ||||||||
2786 | // iterations are not required for correctness, or N - Step, otherwise. Step | ||||||||
2787 | // is equal to the vectorization factor (number of SIMD elements) times the | ||||||||
2788 | // unroll factor (number of SIMD instructions). | ||||||||
2789 | Value *R = Builder.CreateURem(TC, Step, "n.mod.vf"); | ||||||||
2790 | |||||||||
2791 | // If there is a non-reversed interleaved group that may speculatively access | ||||||||
2792 | // memory out-of-bounds, we need to ensure that there will be at least one | ||||||||
2793 | // iteration of the scalar epilogue loop. Thus, if the step evenly divides | ||||||||
2794 | // the trip count, we set the remainder to be equal to the step. If the step | ||||||||
2795 | // does not evenly divide the trip count, no adjustment is necessary since | ||||||||
2796 | // there will already be scalar iterations. Note that the minimum iterations | ||||||||
2797 | // check ensures that N >= Step. | ||||||||
2798 | if (VF.isVector() && Cost->requiresScalarEpilogue()) { | ||||||||
2799 | auto *IsZero = Builder.CreateICmpEQ(R, ConstantInt::get(R->getType(), 0)); | ||||||||
2800 | R = Builder.CreateSelect(IsZero, Step, R); | ||||||||
2801 | } | ||||||||
2802 | |||||||||
2803 | VectorTripCount = Builder.CreateSub(TC, R, "n.vec"); | ||||||||
2804 | |||||||||
2805 | return VectorTripCount; | ||||||||
2806 | } | ||||||||
2807 | |||||||||
2808 | Value *InnerLoopVectorizer::createBitOrPointerCast(Value *V, VectorType *DstVTy, | ||||||||
2809 | const DataLayout &DL) { | ||||||||
2810 | // Verify that V is a vector type with same number of elements as DstVTy. | ||||||||
2811 | auto *DstFVTy = cast<FixedVectorType>(DstVTy); | ||||||||
2812 | unsigned VF = DstFVTy->getNumElements(); | ||||||||
2813 | auto *SrcVecTy = cast<FixedVectorType>(V->getType()); | ||||||||
2814 | assert((VF == SrcVecTy->getNumElements()) && "Vector dimensions do not match")(((VF == SrcVecTy->getNumElements()) && "Vector dimensions do not match" ) ? static_cast<void> (0) : __assert_fail ("(VF == SrcVecTy->getNumElements()) && \"Vector dimensions do not match\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2814, __PRETTY_FUNCTION__)); | ||||||||
2815 | Type *SrcElemTy = SrcVecTy->getElementType(); | ||||||||
2816 | Type *DstElemTy = DstFVTy->getElementType(); | ||||||||
2817 | assert((DL.getTypeSizeInBits(SrcElemTy) == DL.getTypeSizeInBits(DstElemTy)) &&(((DL.getTypeSizeInBits(SrcElemTy) == DL.getTypeSizeInBits(DstElemTy )) && "Vector elements must have same size") ? static_cast <void> (0) : __assert_fail ("(DL.getTypeSizeInBits(SrcElemTy) == DL.getTypeSizeInBits(DstElemTy)) && \"Vector elements must have same size\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2818, __PRETTY_FUNCTION__)) | ||||||||
2818 | "Vector elements must have same size")(((DL.getTypeSizeInBits(SrcElemTy) == DL.getTypeSizeInBits(DstElemTy )) && "Vector elements must have same size") ? static_cast <void> (0) : __assert_fail ("(DL.getTypeSizeInBits(SrcElemTy) == DL.getTypeSizeInBits(DstElemTy)) && \"Vector elements must have same size\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2818, __PRETTY_FUNCTION__)); | ||||||||
2819 | |||||||||
2820 | // Do a direct cast if element types are castable. | ||||||||
2821 | if (CastInst::isBitOrNoopPointerCastable(SrcElemTy, DstElemTy, DL)) { | ||||||||
2822 | return Builder.CreateBitOrPointerCast(V, DstFVTy); | ||||||||
2823 | } | ||||||||
2824 | // V cannot be directly casted to desired vector type. | ||||||||
2825 | // May happen when V is a floating point vector but DstVTy is a vector of | ||||||||
2826 | // pointers or vice-versa. Handle this using a two-step bitcast using an | ||||||||
2827 | // intermediate Integer type for the bitcast i.e. Ptr <-> Int <-> Float. | ||||||||
2828 | assert((DstElemTy->isPointerTy() != SrcElemTy->isPointerTy()) &&(((DstElemTy->isPointerTy() != SrcElemTy->isPointerTy() ) && "Only one type should be a pointer type") ? static_cast <void> (0) : __assert_fail ("(DstElemTy->isPointerTy() != SrcElemTy->isPointerTy()) && \"Only one type should be a pointer type\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2829, __PRETTY_FUNCTION__)) | ||||||||
2829 | "Only one type should be a pointer type")(((DstElemTy->isPointerTy() != SrcElemTy->isPointerTy() ) && "Only one type should be a pointer type") ? static_cast <void> (0) : __assert_fail ("(DstElemTy->isPointerTy() != SrcElemTy->isPointerTy()) && \"Only one type should be a pointer type\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2829, __PRETTY_FUNCTION__)); | ||||||||
2830 | assert((DstElemTy->isFloatingPointTy() != SrcElemTy->isFloatingPointTy()) &&(((DstElemTy->isFloatingPointTy() != SrcElemTy->isFloatingPointTy ()) && "Only one type should be a floating point type" ) ? static_cast<void> (0) : __assert_fail ("(DstElemTy->isFloatingPointTy() != SrcElemTy->isFloatingPointTy()) && \"Only one type should be a floating point type\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2831, __PRETTY_FUNCTION__)) | ||||||||
2831 | "Only one type should be a floating point type")(((DstElemTy->isFloatingPointTy() != SrcElemTy->isFloatingPointTy ()) && "Only one type should be a floating point type" ) ? static_cast<void> (0) : __assert_fail ("(DstElemTy->isFloatingPointTy() != SrcElemTy->isFloatingPointTy()) && \"Only one type should be a floating point type\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2831, __PRETTY_FUNCTION__)); | ||||||||
2832 | Type *IntTy = | ||||||||
2833 | IntegerType::getIntNTy(V->getContext(), DL.getTypeSizeInBits(SrcElemTy)); | ||||||||
2834 | auto *VecIntTy = FixedVectorType::get(IntTy, VF); | ||||||||
2835 | Value *CastVal = Builder.CreateBitOrPointerCast(V, VecIntTy); | ||||||||
2836 | return Builder.CreateBitOrPointerCast(CastVal, DstFVTy); | ||||||||
2837 | } | ||||||||
2838 | |||||||||
2839 | void InnerLoopVectorizer::emitMinimumIterationCountCheck(Loop *L, | ||||||||
2840 | BasicBlock *Bypass) { | ||||||||
2841 | Value *Count = getOrCreateTripCount(L); | ||||||||
2842 | // Reuse existing vector loop preheader for TC checks. | ||||||||
2843 | // Note that new preheader block is generated for vector loop. | ||||||||
2844 | BasicBlock *const TCCheckBlock = LoopVectorPreHeader; | ||||||||
2845 | IRBuilder<> Builder(TCCheckBlock->getTerminator()); | ||||||||
2846 | |||||||||
2847 | // Generate code to check if the loop's trip count is less than VF * UF, or | ||||||||
2848 | // equal to it in case a scalar epilogue is required; this implies that the | ||||||||
2849 | // vector trip count is zero. This check also covers the case where adding one | ||||||||
2850 | // to the backedge-taken count overflowed leading to an incorrect trip count | ||||||||
2851 | // of zero. In this case we will also jump to the scalar loop. | ||||||||
2852 | auto P = Cost->requiresScalarEpilogue() ? ICmpInst::ICMP_ULE | ||||||||
2853 | : ICmpInst::ICMP_ULT; | ||||||||
2854 | |||||||||
2855 | // If tail is to be folded, vector loop takes care of all iterations. | ||||||||
2856 | Value *CheckMinIters = Builder.getFalse(); | ||||||||
2857 | if (!Cost->foldTailByMasking()) { | ||||||||
2858 | assert(!VF.isScalable() && "scalable vectors not yet supported.")((!VF.isScalable() && "scalable vectors not yet supported." ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"scalable vectors not yet supported.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2858, __PRETTY_FUNCTION__)); | ||||||||
2859 | CheckMinIters = Builder.CreateICmp( | ||||||||
2860 | P, Count, | ||||||||
2861 | ConstantInt::get(Count->getType(), VF.getKnownMinValue() * UF), | ||||||||
2862 | "min.iters.check"); | ||||||||
2863 | } | ||||||||
2864 | // Create new preheader for vector loop. | ||||||||
2865 | LoopVectorPreHeader = | ||||||||
2866 | SplitBlock(TCCheckBlock, TCCheckBlock->getTerminator(), DT, LI, nullptr, | ||||||||
2867 | "vector.ph"); | ||||||||
2868 | |||||||||
2869 | assert(DT->properlyDominates(DT->getNode(TCCheckBlock),((DT->properlyDominates(DT->getNode(TCCheckBlock), DT-> getNode(Bypass)->getIDom()) && "TC check is expected to dominate Bypass" ) ? static_cast<void> (0) : __assert_fail ("DT->properlyDominates(DT->getNode(TCCheckBlock), DT->getNode(Bypass)->getIDom()) && \"TC check is expected to dominate Bypass\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2871, __PRETTY_FUNCTION__)) | ||||||||
2870 | DT->getNode(Bypass)->getIDom()) &&((DT->properlyDominates(DT->getNode(TCCheckBlock), DT-> getNode(Bypass)->getIDom()) && "TC check is expected to dominate Bypass" ) ? static_cast<void> (0) : __assert_fail ("DT->properlyDominates(DT->getNode(TCCheckBlock), DT->getNode(Bypass)->getIDom()) && \"TC check is expected to dominate Bypass\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2871, __PRETTY_FUNCTION__)) | ||||||||
2871 | "TC check is expected to dominate Bypass")((DT->properlyDominates(DT->getNode(TCCheckBlock), DT-> getNode(Bypass)->getIDom()) && "TC check is expected to dominate Bypass" ) ? static_cast<void> (0) : __assert_fail ("DT->properlyDominates(DT->getNode(TCCheckBlock), DT->getNode(Bypass)->getIDom()) && \"TC check is expected to dominate Bypass\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2871, __PRETTY_FUNCTION__)); | ||||||||
2872 | |||||||||
2873 | // Update dominator for Bypass & LoopExit. | ||||||||
2874 | DT->changeImmediateDominator(Bypass, TCCheckBlock); | ||||||||
2875 | DT->changeImmediateDominator(LoopExitBlock, TCCheckBlock); | ||||||||
2876 | |||||||||
2877 | ReplaceInstWithInst( | ||||||||
2878 | TCCheckBlock->getTerminator(), | ||||||||
2879 | BranchInst::Create(Bypass, LoopVectorPreHeader, CheckMinIters)); | ||||||||
2880 | LoopBypassBlocks.push_back(TCCheckBlock); | ||||||||
2881 | } | ||||||||
2882 | |||||||||
2883 | void InnerLoopVectorizer::emitSCEVChecks(Loop *L, BasicBlock *Bypass) { | ||||||||
2884 | // Reuse existing vector loop preheader for SCEV checks. | ||||||||
2885 | // Note that new preheader block is generated for vector loop. | ||||||||
2886 | BasicBlock *const SCEVCheckBlock = LoopVectorPreHeader; | ||||||||
2887 | |||||||||
2888 | // Generate the code to check that the SCEV assumptions that we made. | ||||||||
2889 | // We want the new basic block to start at the first instruction in a | ||||||||
2890 | // sequence of instructions that form a check. | ||||||||
2891 | SCEVExpander Exp(*PSE.getSE(), Bypass->getModule()->getDataLayout(), | ||||||||
2892 | "scev.check"); | ||||||||
2893 | Value *SCEVCheck = Exp.expandCodeForPredicate( | ||||||||
2894 | &PSE.getUnionPredicate(), SCEVCheckBlock->getTerminator()); | ||||||||
2895 | |||||||||
2896 | if (auto *C = dyn_cast<ConstantInt>(SCEVCheck)) | ||||||||
2897 | if (C->isZero()) | ||||||||
2898 | return; | ||||||||
2899 | |||||||||
2900 | assert(!(SCEVCheckBlock->getParent()->hasOptSize() ||((!(SCEVCheckBlock->getParent()->hasOptSize() || (OptForSizeBasedOnProfile && Cost->Hints->getForce() != LoopVectorizeHints ::FK_Enabled)) && "Cannot SCEV check stride or overflow when optimizing for size" ) ? static_cast<void> (0) : __assert_fail ("!(SCEVCheckBlock->getParent()->hasOptSize() || (OptForSizeBasedOnProfile && Cost->Hints->getForce() != LoopVectorizeHints::FK_Enabled)) && \"Cannot SCEV check stride or overflow when optimizing for size\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2903, __PRETTY_FUNCTION__)) | ||||||||
2901 | (OptForSizeBasedOnProfile &&((!(SCEVCheckBlock->getParent()->hasOptSize() || (OptForSizeBasedOnProfile && Cost->Hints->getForce() != LoopVectorizeHints ::FK_Enabled)) && "Cannot SCEV check stride or overflow when optimizing for size" ) ? static_cast<void> (0) : __assert_fail ("!(SCEVCheckBlock->getParent()->hasOptSize() || (OptForSizeBasedOnProfile && Cost->Hints->getForce() != LoopVectorizeHints::FK_Enabled)) && \"Cannot SCEV check stride or overflow when optimizing for size\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2903, __PRETTY_FUNCTION__)) | ||||||||
2902 | Cost->Hints->getForce() != LoopVectorizeHints::FK_Enabled)) &&((!(SCEVCheckBlock->getParent()->hasOptSize() || (OptForSizeBasedOnProfile && Cost->Hints->getForce() != LoopVectorizeHints ::FK_Enabled)) && "Cannot SCEV check stride or overflow when optimizing for size" ) ? static_cast<void> (0) : __assert_fail ("!(SCEVCheckBlock->getParent()->hasOptSize() || (OptForSizeBasedOnProfile && Cost->Hints->getForce() != LoopVectorizeHints::FK_Enabled)) && \"Cannot SCEV check stride or overflow when optimizing for size\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2903, __PRETTY_FUNCTION__)) | ||||||||
2903 | "Cannot SCEV check stride or overflow when optimizing for size")((!(SCEVCheckBlock->getParent()->hasOptSize() || (OptForSizeBasedOnProfile && Cost->Hints->getForce() != LoopVectorizeHints ::FK_Enabled)) && "Cannot SCEV check stride or overflow when optimizing for size" ) ? static_cast<void> (0) : __assert_fail ("!(SCEVCheckBlock->getParent()->hasOptSize() || (OptForSizeBasedOnProfile && Cost->Hints->getForce() != LoopVectorizeHints::FK_Enabled)) && \"Cannot SCEV check stride or overflow when optimizing for size\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2903, __PRETTY_FUNCTION__)); | ||||||||
2904 | |||||||||
2905 | SCEVCheckBlock->setName("vector.scevcheck"); | ||||||||
2906 | // Create new preheader for vector loop. | ||||||||
2907 | LoopVectorPreHeader = | ||||||||
2908 | SplitBlock(SCEVCheckBlock, SCEVCheckBlock->getTerminator(), DT, LI, | ||||||||
2909 | nullptr, "vector.ph"); | ||||||||
2910 | |||||||||
2911 | // Update dominator only if this is first RT check. | ||||||||
2912 | if (LoopBypassBlocks.empty()) { | ||||||||
2913 | DT->changeImmediateDominator(Bypass, SCEVCheckBlock); | ||||||||
2914 | DT->changeImmediateDominator(LoopExitBlock, SCEVCheckBlock); | ||||||||
2915 | } | ||||||||
2916 | |||||||||
2917 | ReplaceInstWithInst( | ||||||||
2918 | SCEVCheckBlock->getTerminator(), | ||||||||
2919 | BranchInst::Create(Bypass, LoopVectorPreHeader, SCEVCheck)); | ||||||||
2920 | LoopBypassBlocks.push_back(SCEVCheckBlock); | ||||||||
2921 | AddedSafetyChecks = true; | ||||||||
2922 | } | ||||||||
2923 | |||||||||
2924 | void InnerLoopVectorizer::emitMemRuntimeChecks(Loop *L, BasicBlock *Bypass) { | ||||||||
2925 | // VPlan-native path does not do any analysis for runtime checks currently. | ||||||||
2926 | if (EnableVPlanNativePath) | ||||||||
2927 | return; | ||||||||
2928 | |||||||||
2929 | // Reuse existing vector loop preheader for runtime memory checks. | ||||||||
2930 | // Note that new preheader block is generated for vector loop. | ||||||||
2931 | BasicBlock *const MemCheckBlock = L->getLoopPreheader(); | ||||||||
2932 | |||||||||
2933 | // Generate the code that checks in runtime if arrays overlap. We put the | ||||||||
2934 | // checks into a separate block to make the more common case of few elements | ||||||||
2935 | // faster. | ||||||||
2936 | auto *LAI = Legal->getLAI(); | ||||||||
2937 | const auto &RtPtrChecking = *LAI->getRuntimePointerChecking(); | ||||||||
2938 | if (!RtPtrChecking.Need) | ||||||||
2939 | return; | ||||||||
2940 | |||||||||
2941 | if (MemCheckBlock->getParent()->hasOptSize() || OptForSizeBasedOnProfile) { | ||||||||
2942 | assert(Cost->Hints->getForce() == LoopVectorizeHints::FK_Enabled &&((Cost->Hints->getForce() == LoopVectorizeHints::FK_Enabled && "Cannot emit memory checks when optimizing for size, unless forced " "to vectorize.") ? static_cast<void> (0) : __assert_fail ("Cost->Hints->getForce() == LoopVectorizeHints::FK_Enabled && \"Cannot emit memory checks when optimizing for size, unless forced \" \"to vectorize.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2944, __PRETTY_FUNCTION__)) | ||||||||
2943 | "Cannot emit memory checks when optimizing for size, unless forced "((Cost->Hints->getForce() == LoopVectorizeHints::FK_Enabled && "Cannot emit memory checks when optimizing for size, unless forced " "to vectorize.") ? static_cast<void> (0) : __assert_fail ("Cost->Hints->getForce() == LoopVectorizeHints::FK_Enabled && \"Cannot emit memory checks when optimizing for size, unless forced \" \"to vectorize.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2944, __PRETTY_FUNCTION__)) | ||||||||
2944 | "to vectorize.")((Cost->Hints->getForce() == LoopVectorizeHints::FK_Enabled && "Cannot emit memory checks when optimizing for size, unless forced " "to vectorize.") ? static_cast<void> (0) : __assert_fail ("Cost->Hints->getForce() == LoopVectorizeHints::FK_Enabled && \"Cannot emit memory checks when optimizing for size, unless forced \" \"to vectorize.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2944, __PRETTY_FUNCTION__)); | ||||||||
2945 | ORE->emit([&]() { | ||||||||
2946 | return OptimizationRemarkAnalysis(DEBUG_TYPE"loop-vectorize", "VectorizationCodeSize", | ||||||||
2947 | L->getStartLoc(), L->getHeader()) | ||||||||
2948 | << "Code-size may be reduced by not forcing " | ||||||||
2949 | "vectorization, or by source-code modifications " | ||||||||
2950 | "eliminating the need for runtime checks " | ||||||||
2951 | "(e.g., adding 'restrict')."; | ||||||||
2952 | }); | ||||||||
2953 | } | ||||||||
2954 | |||||||||
2955 | MemCheckBlock->setName("vector.memcheck"); | ||||||||
2956 | // Create new preheader for vector loop. | ||||||||
2957 | LoopVectorPreHeader = | ||||||||
2958 | SplitBlock(MemCheckBlock, MemCheckBlock->getTerminator(), DT, LI, nullptr, | ||||||||
2959 | "vector.ph"); | ||||||||
2960 | |||||||||
2961 | auto *CondBranch = cast<BranchInst>( | ||||||||
2962 | Builder.CreateCondBr(Builder.getTrue(), Bypass, LoopVectorPreHeader)); | ||||||||
2963 | ReplaceInstWithInst(MemCheckBlock->getTerminator(), CondBranch); | ||||||||
2964 | LoopBypassBlocks.push_back(MemCheckBlock); | ||||||||
2965 | AddedSafetyChecks = true; | ||||||||
2966 | |||||||||
2967 | // Update dominator only if this is first RT check. | ||||||||
2968 | if (LoopBypassBlocks.empty()) { | ||||||||
2969 | DT->changeImmediateDominator(Bypass, MemCheckBlock); | ||||||||
2970 | DT->changeImmediateDominator(LoopExitBlock, MemCheckBlock); | ||||||||
2971 | } | ||||||||
2972 | |||||||||
2973 | Instruction *FirstCheckInst; | ||||||||
2974 | Instruction *MemRuntimeCheck; | ||||||||
2975 | std::tie(FirstCheckInst, MemRuntimeCheck) = | ||||||||
2976 | addRuntimeChecks(MemCheckBlock->getTerminator(), OrigLoop, | ||||||||
2977 | RtPtrChecking.getChecks(), RtPtrChecking.getSE()); | ||||||||
2978 | assert(MemRuntimeCheck && "no RT checks generated although RtPtrChecking "((MemRuntimeCheck && "no RT checks generated although RtPtrChecking " "claimed checks are required") ? static_cast<void> (0) : __assert_fail ("MemRuntimeCheck && \"no RT checks generated although RtPtrChecking \" \"claimed checks are required\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2979, __PRETTY_FUNCTION__)) | ||||||||
2979 | "claimed checks are required")((MemRuntimeCheck && "no RT checks generated although RtPtrChecking " "claimed checks are required") ? static_cast<void> (0) : __assert_fail ("MemRuntimeCheck && \"no RT checks generated although RtPtrChecking \" \"claimed checks are required\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2979, __PRETTY_FUNCTION__)); | ||||||||
2980 | CondBranch->setCondition(MemRuntimeCheck); | ||||||||
2981 | |||||||||
2982 | // We currently don't use LoopVersioning for the actual loop cloning but we | ||||||||
2983 | // still use it to add the noalias metadata. | ||||||||
2984 | LVer = std::make_unique<LoopVersioning>(*Legal->getLAI(), OrigLoop, LI, DT, | ||||||||
2985 | PSE.getSE()); | ||||||||
2986 | LVer->prepareNoAliasMetadata(); | ||||||||
2987 | } | ||||||||
2988 | |||||||||
2989 | Value *InnerLoopVectorizer::emitTransformedIndex( | ||||||||
2990 | IRBuilder<> &B, Value *Index, ScalarEvolution *SE, const DataLayout &DL, | ||||||||
2991 | const InductionDescriptor &ID) const { | ||||||||
2992 | |||||||||
2993 | SCEVExpander Exp(*SE, DL, "induction"); | ||||||||
2994 | auto Step = ID.getStep(); | ||||||||
2995 | auto StartValue = ID.getStartValue(); | ||||||||
2996 | assert(Index->getType() == Step->getType() &&((Index->getType() == Step->getType() && "Index type does not match StepValue type" ) ? static_cast<void> (0) : __assert_fail ("Index->getType() == Step->getType() && \"Index type does not match StepValue type\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2997, __PRETTY_FUNCTION__)) | ||||||||
2997 | "Index type does not match StepValue type")((Index->getType() == Step->getType() && "Index type does not match StepValue type" ) ? static_cast<void> (0) : __assert_fail ("Index->getType() == Step->getType() && \"Index type does not match StepValue type\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2997, __PRETTY_FUNCTION__)); | ||||||||
2998 | |||||||||
2999 | // Note: the IR at this point is broken. We cannot use SE to create any new | ||||||||
3000 | // SCEV and then expand it, hoping that SCEV's simplification will give us | ||||||||
3001 | // a more optimal code. Unfortunately, attempt of doing so on invalid IR may | ||||||||
3002 | // lead to various SCEV crashes. So all we can do is to use builder and rely | ||||||||
3003 | // on InstCombine for future simplifications. Here we handle some trivial | ||||||||
3004 | // cases only. | ||||||||
3005 | auto CreateAdd = [&B](Value *X, Value *Y) { | ||||||||
3006 | assert(X->getType() == Y->getType() && "Types don't match!")((X->getType() == Y->getType() && "Types don't match!" ) ? static_cast<void> (0) : __assert_fail ("X->getType() == Y->getType() && \"Types don't match!\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3006, __PRETTY_FUNCTION__)); | ||||||||
3007 | if (auto *CX = dyn_cast<ConstantInt>(X)) | ||||||||
3008 | if (CX->isZero()) | ||||||||
3009 | return Y; | ||||||||
3010 | if (auto *CY = dyn_cast<ConstantInt>(Y)) | ||||||||
3011 | if (CY->isZero()) | ||||||||
3012 | return X; | ||||||||
3013 | return B.CreateAdd(X, Y); | ||||||||
3014 | }; | ||||||||
3015 | |||||||||
3016 | auto CreateMul = [&B](Value *X, Value *Y) { | ||||||||
3017 | assert(X->getType() == Y->getType() && "Types don't match!")((X->getType() == Y->getType() && "Types don't match!" ) ? static_cast<void> (0) : __assert_fail ("X->getType() == Y->getType() && \"Types don't match!\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3017, __PRETTY_FUNCTION__)); | ||||||||
3018 | if (auto *CX = dyn_cast<ConstantInt>(X)) | ||||||||
3019 | if (CX->isOne()) | ||||||||
3020 | return Y; | ||||||||
3021 | if (auto *CY = dyn_cast<ConstantInt>(Y)) | ||||||||
3022 | if (CY->isOne()) | ||||||||
3023 | return X; | ||||||||
3024 | return B.CreateMul(X, Y); | ||||||||
3025 | }; | ||||||||
3026 | |||||||||
3027 | // Get a suitable insert point for SCEV expansion. For blocks in the vector | ||||||||
3028 | // loop, choose the end of the vector loop header (=LoopVectorBody), because | ||||||||
3029 | // the DomTree is not kept up-to-date for additional blocks generated in the | ||||||||
3030 | // vector loop. By using the header as insertion point, we guarantee that the | ||||||||
3031 | // expanded instructions dominate all their uses. | ||||||||
3032 | auto GetInsertPoint = [this, &B]() { | ||||||||
3033 | BasicBlock *InsertBB = B.GetInsertPoint()->getParent(); | ||||||||
3034 | if (InsertBB != LoopVectorBody && | ||||||||
3035 | LI->getLoopFor(LoopVectorBody) == LI->getLoopFor(InsertBB)) | ||||||||
3036 | return LoopVectorBody->getTerminator(); | ||||||||
3037 | return &*B.GetInsertPoint(); | ||||||||
3038 | }; | ||||||||
3039 | switch (ID.getKind()) { | ||||||||
3040 | case InductionDescriptor::IK_IntInduction: { | ||||||||
3041 | assert(Index->getType() == StartValue->getType() &&((Index->getType() == StartValue->getType() && "Index type does not match StartValue type" ) ? static_cast<void> (0) : __assert_fail ("Index->getType() == StartValue->getType() && \"Index type does not match StartValue type\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3042, __PRETTY_FUNCTION__)) | ||||||||
3042 | "Index type does not match StartValue type")((Index->getType() == StartValue->getType() && "Index type does not match StartValue type" ) ? static_cast<void> (0) : __assert_fail ("Index->getType() == StartValue->getType() && \"Index type does not match StartValue type\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3042, __PRETTY_FUNCTION__)); | ||||||||
3043 | if (ID.getConstIntStepValue() && ID.getConstIntStepValue()->isMinusOne()) | ||||||||
3044 | return B.CreateSub(StartValue, Index); | ||||||||
3045 | auto *Offset = CreateMul( | ||||||||
3046 | Index, Exp.expandCodeFor(Step, Index->getType(), GetInsertPoint())); | ||||||||
3047 | return CreateAdd(StartValue, Offset); | ||||||||
3048 | } | ||||||||
3049 | case InductionDescriptor::IK_PtrInduction: { | ||||||||
3050 | assert(isa<SCEVConstant>(Step) &&((isa<SCEVConstant>(Step) && "Expected constant step for pointer induction" ) ? static_cast<void> (0) : __assert_fail ("isa<SCEVConstant>(Step) && \"Expected constant step for pointer induction\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3051, __PRETTY_FUNCTION__)) | ||||||||
3051 | "Expected constant step for pointer induction")((isa<SCEVConstant>(Step) && "Expected constant step for pointer induction" ) ? static_cast<void> (0) : __assert_fail ("isa<SCEVConstant>(Step) && \"Expected constant step for pointer induction\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3051, __PRETTY_FUNCTION__)); | ||||||||
3052 | return B.CreateGEP( | ||||||||
3053 | StartValue->getType()->getPointerElementType(), StartValue, | ||||||||
3054 | CreateMul(Index, | ||||||||
3055 | Exp.expandCodeFor(Step, Index->getType(), GetInsertPoint()))); | ||||||||
3056 | } | ||||||||
3057 | case InductionDescriptor::IK_FpInduction: { | ||||||||
3058 | assert(Step->getType()->isFloatingPointTy() && "Expected FP Step value")((Step->getType()->isFloatingPointTy() && "Expected FP Step value" ) ? static_cast<void> (0) : __assert_fail ("Step->getType()->isFloatingPointTy() && \"Expected FP Step value\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3058, __PRETTY_FUNCTION__)); | ||||||||
3059 | auto InductionBinOp = ID.getInductionBinOp(); | ||||||||
3060 | assert(InductionBinOp &&((InductionBinOp && (InductionBinOp->getOpcode() == Instruction::FAdd || InductionBinOp->getOpcode() == Instruction ::FSub) && "Original bin op should be defined for FP induction" ) ? static_cast<void> (0) : __assert_fail ("InductionBinOp && (InductionBinOp->getOpcode() == Instruction::FAdd || InductionBinOp->getOpcode() == Instruction::FSub) && \"Original bin op should be defined for FP induction\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3063, __PRETTY_FUNCTION__)) | ||||||||
3061 | (InductionBinOp->getOpcode() == Instruction::FAdd ||((InductionBinOp && (InductionBinOp->getOpcode() == Instruction::FAdd || InductionBinOp->getOpcode() == Instruction ::FSub) && "Original bin op should be defined for FP induction" ) ? static_cast<void> (0) : __assert_fail ("InductionBinOp && (InductionBinOp->getOpcode() == Instruction::FAdd || InductionBinOp->getOpcode() == Instruction::FSub) && \"Original bin op should be defined for FP induction\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3063, __PRETTY_FUNCTION__)) | ||||||||
3062 | InductionBinOp->getOpcode() == Instruction::FSub) &&((InductionBinOp && (InductionBinOp->getOpcode() == Instruction::FAdd || InductionBinOp->getOpcode() == Instruction ::FSub) && "Original bin op should be defined for FP induction" ) ? static_cast<void> (0) : __assert_fail ("InductionBinOp && (InductionBinOp->getOpcode() == Instruction::FAdd || InductionBinOp->getOpcode() == Instruction::FSub) && \"Original bin op should be defined for FP induction\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3063, __PRETTY_FUNCTION__)) | ||||||||
3063 | "Original bin op should be defined for FP induction")((InductionBinOp && (InductionBinOp->getOpcode() == Instruction::FAdd || InductionBinOp->getOpcode() == Instruction ::FSub) && "Original bin op should be defined for FP induction" ) ? static_cast<void> (0) : __assert_fail ("InductionBinOp && (InductionBinOp->getOpcode() == Instruction::FAdd || InductionBinOp->getOpcode() == Instruction::FSub) && \"Original bin op should be defined for FP induction\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3063, __PRETTY_FUNCTION__)); | ||||||||
3064 | |||||||||
3065 | Value *StepValue = cast<SCEVUnknown>(Step)->getValue(); | ||||||||
3066 | |||||||||
3067 | // Floating point operations had to be 'fast' to enable the induction. | ||||||||
3068 | FastMathFlags Flags; | ||||||||
3069 | Flags.setFast(); | ||||||||
3070 | |||||||||
3071 | Value *MulExp = B.CreateFMul(StepValue, Index); | ||||||||
3072 | if (isa<Instruction>(MulExp)) | ||||||||
3073 | // We have to check, the MulExp may be a constant. | ||||||||
3074 | cast<Instruction>(MulExp)->setFastMathFlags(Flags); | ||||||||
3075 | |||||||||
3076 | Value *BOp = B.CreateBinOp(InductionBinOp->getOpcode(), StartValue, MulExp, | ||||||||
3077 | "induction"); | ||||||||
3078 | if (isa<Instruction>(BOp)) | ||||||||
3079 | cast<Instruction>(BOp)->setFastMathFlags(Flags); | ||||||||
3080 | |||||||||
3081 | return BOp; | ||||||||
3082 | } | ||||||||
3083 | case InductionDescriptor::IK_NoInduction: | ||||||||
3084 | return nullptr; | ||||||||
3085 | } | ||||||||
3086 | llvm_unreachable("invalid enum")::llvm::llvm_unreachable_internal("invalid enum", "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3086); | ||||||||
3087 | } | ||||||||
3088 | |||||||||
3089 | Loop *InnerLoopVectorizer::createVectorLoopSkeleton(StringRef Prefix) { | ||||||||
3090 | LoopScalarBody = OrigLoop->getHeader(); | ||||||||
3091 | LoopVectorPreHeader = OrigLoop->getLoopPreheader(); | ||||||||
3092 | LoopExitBlock = OrigLoop->getExitBlock(); | ||||||||
3093 | assert(LoopExitBlock && "Must have an exit block")((LoopExitBlock && "Must have an exit block") ? static_cast <void> (0) : __assert_fail ("LoopExitBlock && \"Must have an exit block\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3093, __PRETTY_FUNCTION__)); | ||||||||
3094 | assert(LoopVectorPreHeader && "Invalid loop structure")((LoopVectorPreHeader && "Invalid loop structure") ? static_cast <void> (0) : __assert_fail ("LoopVectorPreHeader && \"Invalid loop structure\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3094, __PRETTY_FUNCTION__)); | ||||||||
3095 | |||||||||
3096 | LoopMiddleBlock = | ||||||||
3097 | SplitBlock(LoopVectorPreHeader, LoopVectorPreHeader->getTerminator(), DT, | ||||||||
3098 | LI, nullptr, Twine(Prefix) + "middle.block"); | ||||||||
3099 | LoopScalarPreHeader = | ||||||||
3100 | SplitBlock(LoopMiddleBlock, LoopMiddleBlock->getTerminator(), DT, LI, | ||||||||
3101 | nullptr, Twine(Prefix) + "scalar.ph"); | ||||||||
3102 | // We intentionally don't let SplitBlock to update LoopInfo since | ||||||||
3103 | // LoopVectorBody should belong to another loop than LoopVectorPreHeader. | ||||||||
3104 | // LoopVectorBody is explicitly added to the correct place few lines later. | ||||||||
3105 | LoopVectorBody = | ||||||||
3106 | SplitBlock(LoopVectorPreHeader, LoopVectorPreHeader->getTerminator(), DT, | ||||||||
3107 | nullptr, nullptr, Twine(Prefix) + "vector.body"); | ||||||||
3108 | |||||||||
3109 | // Update dominator for loop exit. | ||||||||
3110 | DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock); | ||||||||
3111 | |||||||||
3112 | // Create and register the new vector loop. | ||||||||
3113 | Loop *Lp = LI->AllocateLoop(); | ||||||||
3114 | Loop *ParentLoop = OrigLoop->getParentLoop(); | ||||||||
3115 | |||||||||
3116 | // Insert the new loop into the loop nest and register the new basic blocks | ||||||||
3117 | // before calling any utilities such as SCEV that require valid LoopInfo. | ||||||||
3118 | if (ParentLoop) { | ||||||||
3119 | ParentLoop->addChildLoop(Lp); | ||||||||
3120 | } else { | ||||||||
3121 | LI->addTopLevelLoop(Lp); | ||||||||
3122 | } | ||||||||
3123 | Lp->addBasicBlockToLoop(LoopVectorBody, *LI); | ||||||||
3124 | return Lp; | ||||||||
3125 | } | ||||||||
3126 | |||||||||
3127 | void InnerLoopVectorizer::createInductionResumeValues(Loop *L, | ||||||||
3128 | Value *VectorTripCount) { | ||||||||
3129 | assert(VectorTripCount && L && "Expected valid arguments")((VectorTripCount && L && "Expected valid arguments" ) ? static_cast<void> (0) : __assert_fail ("VectorTripCount && L && \"Expected valid arguments\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3129, __PRETTY_FUNCTION__)); | ||||||||
3130 | // We are going to resume the execution of the scalar loop. | ||||||||
3131 | // Go over all of the induction variables that we found and fix the | ||||||||
3132 | // PHIs that are left in the scalar version of the loop. | ||||||||
3133 | // The starting values of PHI nodes depend on the counter of the last | ||||||||
3134 | // iteration in the vectorized loop. | ||||||||
3135 | // If we come from a bypass edge then we need to start from the original | ||||||||
3136 | // start value. | ||||||||
3137 | for (auto &InductionEntry : Legal->getInductionVars()) { | ||||||||
3138 | PHINode *OrigPhi = InductionEntry.first; | ||||||||
3139 | InductionDescriptor II = InductionEntry.second; | ||||||||
3140 | |||||||||
3141 | // Create phi nodes to merge from the backedge-taken check block. | ||||||||
3142 | PHINode *BCResumeVal = | ||||||||
3143 | PHINode::Create(OrigPhi->getType(), 3, "bc.resume.val", | ||||||||
3144 | LoopScalarPreHeader->getTerminator()); | ||||||||
3145 | // Copy original phi DL over to the new one. | ||||||||
3146 | BCResumeVal->setDebugLoc(OrigPhi->getDebugLoc()); | ||||||||
3147 | Value *&EndValue = IVEndValues[OrigPhi]; | ||||||||
3148 | if (OrigPhi == OldInduction) { | ||||||||
3149 | // We know what the end value is. | ||||||||
3150 | EndValue = VectorTripCount; | ||||||||
3151 | } else { | ||||||||
3152 | IRBuilder<> B(L->getLoopPreheader()->getTerminator()); | ||||||||
3153 | Type *StepType = II.getStep()->getType(); | ||||||||
3154 | Instruction::CastOps CastOp = | ||||||||
3155 | CastInst::getCastOpcode(VectorTripCount, true, StepType, true); | ||||||||
3156 | Value *CRD = B.CreateCast(CastOp, VectorTripCount, StepType, "cast.crd"); | ||||||||
3157 | const DataLayout &DL = LoopScalarBody->getModule()->getDataLayout(); | ||||||||
3158 | EndValue = emitTransformedIndex(B, CRD, PSE.getSE(), DL, II); | ||||||||
3159 | EndValue->setName("ind.end"); | ||||||||
3160 | } | ||||||||
3161 | |||||||||
3162 | // The new PHI merges the original incoming value, in case of a bypass, | ||||||||
3163 | // or the value at the end of the vectorized loop. | ||||||||
3164 | BCResumeVal->addIncoming(EndValue, LoopMiddleBlock); | ||||||||
3165 | |||||||||
3166 | // Fix the scalar body counter (PHI node). | ||||||||
3167 | // The old induction's phi node in the scalar body needs the truncated | ||||||||
3168 | // value. | ||||||||
3169 | for (BasicBlock *BB : LoopBypassBlocks) | ||||||||
3170 | BCResumeVal->addIncoming(II.getStartValue(), BB); | ||||||||
3171 | OrigPhi->setIncomingValueForBlock(LoopScalarPreHeader, BCResumeVal); | ||||||||
3172 | } | ||||||||
3173 | } | ||||||||
3174 | |||||||||
3175 | BasicBlock *InnerLoopVectorizer::completeLoopSkeleton(Loop *L, | ||||||||
3176 | MDNode *OrigLoopID) { | ||||||||
3177 | assert(L && "Expected valid loop.")((L && "Expected valid loop.") ? static_cast<void> (0) : __assert_fail ("L && \"Expected valid loop.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3177, __PRETTY_FUNCTION__)); | ||||||||
3178 | |||||||||
3179 | // The trip counts should be cached by now. | ||||||||
3180 | Value *Count = getOrCreateTripCount(L); | ||||||||
3181 | Value *VectorTripCount = getOrCreateVectorTripCount(L); | ||||||||
3182 | |||||||||
3183 | // We need the OrigLoop (scalar loop part) latch terminator to help | ||||||||
3184 | // produce correct debug info for the middle block BB instructions. | ||||||||
3185 | // The legality check stage guarantees that the loop will have a single | ||||||||
3186 | // latch. | ||||||||
3187 | assert(isa<BranchInst>(OrigLoop->getLoopLatch()->getTerminator()) &&((isa<BranchInst>(OrigLoop->getLoopLatch()->getTerminator ()) && "Scalar loop latch terminator isn't a branch") ? static_cast<void> (0) : __assert_fail ("isa<BranchInst>(OrigLoop->getLoopLatch()->getTerminator()) && \"Scalar loop latch terminator isn't a branch\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3188, __PRETTY_FUNCTION__)) | ||||||||
3188 | "Scalar loop latch terminator isn't a branch")((isa<BranchInst>(OrigLoop->getLoopLatch()->getTerminator ()) && "Scalar loop latch terminator isn't a branch") ? static_cast<void> (0) : __assert_fail ("isa<BranchInst>(OrigLoop->getLoopLatch()->getTerminator()) && \"Scalar loop latch terminator isn't a branch\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3188, __PRETTY_FUNCTION__)); | ||||||||
3189 | BranchInst *ScalarLatchBr = | ||||||||
3190 | cast<BranchInst>(OrigLoop->getLoopLatch()->getTerminator()); | ||||||||
3191 | |||||||||
3192 | // Add a check in the middle block to see if we have completed | ||||||||
3193 | // all of the iterations in the first vector loop. | ||||||||
3194 | // If (N - N%VF) == N, then we *don't* need to run the remainder. | ||||||||
3195 | // If tail is to be folded, we know we don't need to run the remainder. | ||||||||
3196 | Value *CmpN = Builder.getTrue(); | ||||||||
3197 | if (!Cost->foldTailByMasking()) { | ||||||||
3198 | CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, Count, | ||||||||
3199 | VectorTripCount, "cmp.n", | ||||||||
3200 | LoopMiddleBlock->getTerminator()); | ||||||||
3201 | |||||||||
3202 | // Here we use the same DebugLoc as the scalar loop latch branch instead | ||||||||
3203 | // of the corresponding compare because they may have ended up with | ||||||||
3204 | // different line numbers and we want to avoid awkward line stepping while | ||||||||
3205 | // debugging. Eg. if the compare has got a line number inside the loop. | ||||||||
3206 | cast<Instruction>(CmpN)->setDebugLoc(ScalarLatchBr->getDebugLoc()); | ||||||||
3207 | } | ||||||||
3208 | |||||||||
3209 | BranchInst *BrInst = | ||||||||
3210 | BranchInst::Create(LoopExitBlock, LoopScalarPreHeader, CmpN); | ||||||||
3211 | BrInst->setDebugLoc(ScalarLatchBr->getDebugLoc()); | ||||||||
3212 | ReplaceInstWithInst(LoopMiddleBlock->getTerminator(), BrInst); | ||||||||
3213 | |||||||||
3214 | // Get ready to start creating new instructions into the vectorized body. | ||||||||
3215 | assert(LoopVectorPreHeader == L->getLoopPreheader() &&((LoopVectorPreHeader == L->getLoopPreheader() && "Inconsistent vector loop preheader" ) ? static_cast<void> (0) : __assert_fail ("LoopVectorPreHeader == L->getLoopPreheader() && \"Inconsistent vector loop preheader\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3216, __PRETTY_FUNCTION__)) | ||||||||
3216 | "Inconsistent vector loop preheader")((LoopVectorPreHeader == L->getLoopPreheader() && "Inconsistent vector loop preheader" ) ? static_cast<void> (0) : __assert_fail ("LoopVectorPreHeader == L->getLoopPreheader() && \"Inconsistent vector loop preheader\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3216, __PRETTY_FUNCTION__)); | ||||||||
3217 | Builder.SetInsertPoint(&*LoopVectorBody->getFirstInsertionPt()); | ||||||||
3218 | |||||||||
3219 | Optional<MDNode *> VectorizedLoopID = | ||||||||
3220 | makeFollowupLoopID(OrigLoopID, {LLVMLoopVectorizeFollowupAll, | ||||||||
3221 | LLVMLoopVectorizeFollowupVectorized}); | ||||||||
3222 | if (VectorizedLoopID.hasValue()) { | ||||||||
3223 | L->setLoopID(VectorizedLoopID.getValue()); | ||||||||
3224 | |||||||||
3225 | // Do not setAlreadyVectorized if loop attributes have been defined | ||||||||
3226 | // explicitly. | ||||||||
3227 | return LoopVectorPreHeader; | ||||||||
3228 | } | ||||||||
3229 | |||||||||
3230 | // Keep all loop hints from the original loop on the vector loop (we'll | ||||||||
3231 | // replace the vectorizer-specific hints below). | ||||||||
3232 | if (MDNode *LID = OrigLoop->getLoopID()) | ||||||||
3233 | L->setLoopID(LID); | ||||||||
3234 | |||||||||
3235 | LoopVectorizeHints Hints(L, true, *ORE); | ||||||||
3236 | Hints.setAlreadyVectorized(); | ||||||||
3237 | |||||||||
3238 | #ifdef EXPENSIVE_CHECKS | ||||||||
3239 | assert(DT->verify(DominatorTree::VerificationLevel::Fast))((DT->verify(DominatorTree::VerificationLevel::Fast)) ? static_cast <void> (0) : __assert_fail ("DT->verify(DominatorTree::VerificationLevel::Fast)" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3239, __PRETTY_FUNCTION__)); | ||||||||
3240 | LI->verify(*DT); | ||||||||
3241 | #endif | ||||||||
3242 | |||||||||
3243 | return LoopVectorPreHeader; | ||||||||
3244 | } | ||||||||
3245 | |||||||||
3246 | BasicBlock *InnerLoopVectorizer::createVectorizedLoopSkeleton() { | ||||||||
3247 | /* | ||||||||
3248 | In this function we generate a new loop. The new loop will contain | ||||||||
3249 | the vectorized instructions while the old loop will continue to run the | ||||||||
3250 | scalar remainder. | ||||||||
3251 | |||||||||
3252 | [ ] <-- loop iteration number check. | ||||||||
3253 | / | | ||||||||
3254 | / v | ||||||||
3255 | | [ ] <-- vector loop bypass (may consist of multiple blocks). | ||||||||
3256 | | / | | ||||||||
3257 | | / v | ||||||||
3258 | || [ ] <-- vector pre header. | ||||||||
3259 | |/ | | ||||||||
3260 | | v | ||||||||
3261 | | [ ] \ | ||||||||
3262 | | [ ]_| <-- vector loop. | ||||||||
3263 | | | | ||||||||
3264 | | v | ||||||||
3265 | | -[ ] <--- middle-block. | ||||||||
3266 | | / | | ||||||||
3267 | | / v | ||||||||
3268 | -|- >[ ] <--- new preheader. | ||||||||
3269 | | | | ||||||||
3270 | | v | ||||||||
3271 | | [ ] \ | ||||||||
3272 | | [ ]_| <-- old scalar loop to handle remainder. | ||||||||
3273 | \ | | ||||||||
3274 | \ v | ||||||||
3275 | >[ ] <-- exit block. | ||||||||
3276 | ... | ||||||||
3277 | */ | ||||||||
3278 | |||||||||
3279 | // Get the metadata of the original loop before it gets modified. | ||||||||
3280 | MDNode *OrigLoopID = OrigLoop->getLoopID(); | ||||||||
3281 | |||||||||
3282 | // Create an empty vector loop, and prepare basic blocks for the runtime | ||||||||
3283 | // checks. | ||||||||
3284 | Loop *Lp = createVectorLoopSkeleton(""); | ||||||||
3285 | |||||||||
3286 | // Now, compare the new count to zero. If it is zero skip the vector loop and | ||||||||
3287 | // jump to the scalar loop. This check also covers the case where the | ||||||||
3288 | // backedge-taken count is uint##_max: adding one to it will overflow leading | ||||||||
3289 | // to an incorrect trip count of zero. In this (rare) case we will also jump | ||||||||
3290 | // to the scalar loop. | ||||||||
3291 | emitMinimumIterationCountCheck(Lp, LoopScalarPreHeader); | ||||||||
3292 | |||||||||
3293 | // Generate the code to check any assumptions that we've made for SCEV | ||||||||
3294 | // expressions. | ||||||||
3295 | emitSCEVChecks(Lp, LoopScalarPreHeader); | ||||||||
3296 | |||||||||
3297 | // Generate the code that checks in runtime if arrays overlap. We put the | ||||||||
3298 | // checks into a separate block to make the more common case of few elements | ||||||||
3299 | // faster. | ||||||||
3300 | emitMemRuntimeChecks(Lp, LoopScalarPreHeader); | ||||||||
3301 | |||||||||
3302 | // Some loops have a single integer induction variable, while other loops | ||||||||
3303 | // don't. One example is c++ iterators that often have multiple pointer | ||||||||
3304 | // induction variables. In the code below we also support a case where we | ||||||||
3305 | // don't have a single induction variable. | ||||||||
3306 | // | ||||||||
3307 | // We try to obtain an induction variable from the original loop as hard | ||||||||
3308 | // as possible. However if we don't find one that: | ||||||||
3309 | // - is an integer | ||||||||
3310 | // - counts from zero, stepping by one | ||||||||
3311 | // - is the size of the widest induction variable type | ||||||||
3312 | // then we create a new one. | ||||||||
3313 | OldInduction = Legal->getPrimaryInduction(); | ||||||||
3314 | Type *IdxTy = Legal->getWidestInductionType(); | ||||||||
3315 | Value *StartIdx = ConstantInt::get(IdxTy, 0); | ||||||||
3316 | // The loop step is equal to the vectorization factor (num of SIMD elements) | ||||||||
3317 | // times the unroll factor (num of SIMD instructions). | ||||||||
3318 | assert(!VF.isScalable() && "scalable vectors not yet supported.")((!VF.isScalable() && "scalable vectors not yet supported." ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"scalable vectors not yet supported.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3318, __PRETTY_FUNCTION__)); | ||||||||
3319 | Constant *Step = ConstantInt::get(IdxTy, VF.getKnownMinValue() * UF); | ||||||||
3320 | Value *CountRoundDown = getOrCreateVectorTripCount(Lp); | ||||||||
3321 | Induction = | ||||||||
3322 | createInductionVariable(Lp, StartIdx, CountRoundDown, Step, | ||||||||
3323 | getDebugLocFromInstOrOperands(OldInduction)); | ||||||||
3324 | |||||||||
3325 | // Emit phis for the new starting index of the scalar loop. | ||||||||
3326 | createInductionResumeValues(Lp, CountRoundDown); | ||||||||
3327 | |||||||||
3328 | return completeLoopSkeleton(Lp, OrigLoopID); | ||||||||
3329 | } | ||||||||
3330 | |||||||||
3331 | // Fix up external users of the induction variable. At this point, we are | ||||||||
3332 | // in LCSSA form, with all external PHIs that use the IV having one input value, | ||||||||
3333 | // coming from the remainder loop. We need those PHIs to also have a correct | ||||||||
3334 | // value for the IV when arriving directly from the middle block. | ||||||||
3335 | void InnerLoopVectorizer::fixupIVUsers(PHINode *OrigPhi, | ||||||||
3336 | const InductionDescriptor &II, | ||||||||
3337 | Value *CountRoundDown, Value *EndValue, | ||||||||
3338 | BasicBlock *MiddleBlock) { | ||||||||
3339 | // There are two kinds of external IV usages - those that use the value | ||||||||
3340 | // computed in the last iteration (the PHI) and those that use the penultimate | ||||||||
3341 | // value (the value that feeds into the phi from the loop latch). | ||||||||
3342 | // We allow both, but they, obviously, have different values. | ||||||||
3343 | |||||||||
3344 | assert(OrigLoop->getExitBlock() && "Expected a single exit block")((OrigLoop->getExitBlock() && "Expected a single exit block" ) ? static_cast<void> (0) : __assert_fail ("OrigLoop->getExitBlock() && \"Expected a single exit block\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3344, __PRETTY_FUNCTION__)); | ||||||||
3345 | |||||||||
3346 | DenseMap<Value *, Value *> MissingVals; | ||||||||
3347 | |||||||||
3348 | // An external user of the last iteration's value should see the value that | ||||||||
3349 | // the remainder loop uses to initialize its own IV. | ||||||||
3350 | Value *PostInc = OrigPhi->getIncomingValueForBlock(OrigLoop->getLoopLatch()); | ||||||||
3351 | for (User *U : PostInc->users()) { | ||||||||
3352 | Instruction *UI = cast<Instruction>(U); | ||||||||
3353 | if (!OrigLoop->contains(UI)) { | ||||||||
3354 | assert(isa<PHINode>(UI) && "Expected LCSSA form")((isa<PHINode>(UI) && "Expected LCSSA form") ? static_cast <void> (0) : __assert_fail ("isa<PHINode>(UI) && \"Expected LCSSA form\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3354, __PRETTY_FUNCTION__)); | ||||||||
3355 | MissingVals[UI] = EndValue; | ||||||||
3356 | } | ||||||||
3357 | } | ||||||||
3358 | |||||||||
3359 | // An external user of the penultimate value need to see EndValue - Step. | ||||||||
3360 | // The simplest way to get this is to recompute it from the constituent SCEVs, | ||||||||
3361 | // that is Start + (Step * (CRD - 1)). | ||||||||
3362 | for (User *U : OrigPhi->users()) { | ||||||||
3363 | auto *UI = cast<Instruction>(U); | ||||||||
3364 | if (!OrigLoop->contains(UI)) { | ||||||||
3365 | const DataLayout &DL = | ||||||||
3366 | OrigLoop->getHeader()->getModule()->getDataLayout(); | ||||||||
3367 | assert(isa<PHINode>(UI) && "Expected LCSSA form")((isa<PHINode>(UI) && "Expected LCSSA form") ? static_cast <void> (0) : __assert_fail ("isa<PHINode>(UI) && \"Expected LCSSA form\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3367, __PRETTY_FUNCTION__)); | ||||||||
3368 | |||||||||
3369 | IRBuilder<> B(MiddleBlock->getTerminator()); | ||||||||
3370 | Value *CountMinusOne = B.CreateSub( | ||||||||
3371 | CountRoundDown, ConstantInt::get(CountRoundDown->getType(), 1)); | ||||||||
3372 | Value *CMO = | ||||||||
3373 | !II.getStep()->getType()->isIntegerTy() | ||||||||
3374 | ? B.CreateCast(Instruction::SIToFP, CountMinusOne, | ||||||||
3375 | II.getStep()->getType()) | ||||||||
3376 | : B.CreateSExtOrTrunc(CountMinusOne, II.getStep()->getType()); | ||||||||
3377 | CMO->setName("cast.cmo"); | ||||||||
3378 | Value *Escape = emitTransformedIndex(B, CMO, PSE.getSE(), DL, II); | ||||||||
3379 | Escape->setName("ind.escape"); | ||||||||
3380 | MissingVals[UI] = Escape; | ||||||||
3381 | } | ||||||||
3382 | } | ||||||||
3383 | |||||||||
3384 | for (auto &I : MissingVals) { | ||||||||
3385 | PHINode *PHI = cast<PHINode>(I.first); | ||||||||
3386 | // One corner case we have to handle is two IVs "chasing" each-other, | ||||||||
3387 | // that is %IV2 = phi [...], [ %IV1, %latch ] | ||||||||
3388 | // In this case, if IV1 has an external use, we need to avoid adding both | ||||||||
3389 | // "last value of IV1" and "penultimate value of IV2". So, verify that we | ||||||||
3390 | // don't already have an incoming value for the middle block. | ||||||||
3391 | if (PHI->getBasicBlockIndex(MiddleBlock) == -1) | ||||||||
3392 | PHI->addIncoming(I.second, MiddleBlock); | ||||||||
3393 | } | ||||||||
3394 | } | ||||||||
3395 | |||||||||
3396 | namespace { | ||||||||
3397 | |||||||||
3398 | struct CSEDenseMapInfo { | ||||||||
3399 | static bool canHandle(const Instruction *I) { | ||||||||
3400 | return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) || | ||||||||
3401 | isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I); | ||||||||
3402 | } | ||||||||
3403 | |||||||||
3404 | static inline Instruction *getEmptyKey() { | ||||||||
3405 | return DenseMapInfo<Instruction *>::getEmptyKey(); | ||||||||
3406 | } | ||||||||
3407 | |||||||||
3408 | static inline Instruction *getTombstoneKey() { | ||||||||
3409 | return DenseMapInfo<Instruction *>::getTombstoneKey(); | ||||||||
3410 | } | ||||||||
3411 | |||||||||
3412 | static unsigned getHashValue(const Instruction *I) { | ||||||||
3413 | assert(canHandle(I) && "Unknown instruction!")((canHandle(I) && "Unknown instruction!") ? static_cast <void> (0) : __assert_fail ("canHandle(I) && \"Unknown instruction!\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3413, __PRETTY_FUNCTION__)); | ||||||||
3414 | return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(), | ||||||||
3415 | I->value_op_end())); | ||||||||
3416 | } | ||||||||
3417 | |||||||||
3418 | static bool isEqual(const Instruction *LHS, const Instruction *RHS) { | ||||||||
3419 | if (LHS == getEmptyKey() || RHS == getEmptyKey() || | ||||||||
3420 | LHS == getTombstoneKey() || RHS == getTombstoneKey()) | ||||||||
3421 | return LHS == RHS; | ||||||||
3422 | return LHS->isIdenticalTo(RHS); | ||||||||
3423 | } | ||||||||
3424 | }; | ||||||||
3425 | |||||||||
3426 | } // end anonymous namespace | ||||||||
3427 | |||||||||
3428 | ///Perform cse of induction variable instructions. | ||||||||
3429 | static void cse(BasicBlock *BB) { | ||||||||
3430 | // Perform simple cse. | ||||||||
3431 | SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap; | ||||||||
3432 | for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) { | ||||||||
3433 | Instruction *In = &*I++; | ||||||||
3434 | |||||||||
3435 | if (!CSEDenseMapInfo::canHandle(In)) | ||||||||
3436 | continue; | ||||||||
3437 | |||||||||
3438 | // Check if we can replace this instruction with any of the | ||||||||
3439 | // visited instructions. | ||||||||
3440 | if (Instruction *V = CSEMap.lookup(In)) { | ||||||||
3441 | In->replaceAllUsesWith(V); | ||||||||
3442 | In->eraseFromParent(); | ||||||||
3443 | continue; | ||||||||
3444 | } | ||||||||
3445 | |||||||||
3446 | CSEMap[In] = In; | ||||||||
3447 | } | ||||||||
3448 | } | ||||||||
3449 | |||||||||
3450 | unsigned LoopVectorizationCostModel::getVectorCallCost(CallInst *CI, | ||||||||
3451 | ElementCount VF, | ||||||||
3452 | bool &NeedToScalarize) { | ||||||||
3453 | assert(!VF.isScalable() && "scalable vectors not yet supported.")((!VF.isScalable() && "scalable vectors not yet supported." ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"scalable vectors not yet supported.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3453, __PRETTY_FUNCTION__)); | ||||||||
3454 | Function *F = CI->getCalledFunction(); | ||||||||
3455 | Type *ScalarRetTy = CI->getType(); | ||||||||
3456 | SmallVector<Type *, 4> Tys, ScalarTys; | ||||||||
3457 | for (auto &ArgOp : CI->arg_operands()) | ||||||||
3458 | ScalarTys.push_back(ArgOp->getType()); | ||||||||
3459 | |||||||||
3460 | // Estimate cost of scalarized vector call. The source operands are assumed | ||||||||
3461 | // to be vectors, so we need to extract individual elements from there, | ||||||||
3462 | // execute VF scalar calls, and then gather the result into the vector return | ||||||||
3463 | // value. | ||||||||
3464 | unsigned ScalarCallCost = TTI.getCallInstrCost(F, ScalarRetTy, ScalarTys, | ||||||||
3465 | TTI::TCK_RecipThroughput); | ||||||||
3466 | if (VF.isScalar()) | ||||||||
3467 | return ScalarCallCost; | ||||||||
3468 | |||||||||
3469 | // Compute corresponding vector type for return value and arguments. | ||||||||
3470 | Type *RetTy = ToVectorTy(ScalarRetTy, VF); | ||||||||
3471 | for (Type *ScalarTy : ScalarTys) | ||||||||
3472 | Tys.push_back(ToVectorTy(ScalarTy, VF)); | ||||||||
3473 | |||||||||
3474 | // Compute costs of unpacking argument values for the scalar calls and | ||||||||
3475 | // packing the return values to a vector. | ||||||||
3476 | unsigned ScalarizationCost = getScalarizationOverhead(CI, VF); | ||||||||
3477 | |||||||||
3478 | unsigned Cost = ScalarCallCost * VF.getKnownMinValue() + ScalarizationCost; | ||||||||
3479 | |||||||||
3480 | // If we can't emit a vector call for this function, then the currently found | ||||||||
3481 | // cost is the cost we need to return. | ||||||||
3482 | NeedToScalarize = true; | ||||||||
3483 | VFShape Shape = VFShape::get(*CI, VF, false /*HasGlobalPred*/); | ||||||||
3484 | Function *VecFunc = VFDatabase(*CI).getVectorizedFunction(Shape); | ||||||||
3485 | |||||||||
3486 | if (!TLI || CI->isNoBuiltin() || !VecFunc) | ||||||||
3487 | return Cost; | ||||||||
3488 | |||||||||
3489 | // If the corresponding vector cost is cheaper, return its cost. | ||||||||
3490 | unsigned VectorCallCost = TTI.getCallInstrCost(nullptr, RetTy, Tys, | ||||||||
3491 | TTI::TCK_RecipThroughput); | ||||||||
3492 | if (VectorCallCost < Cost) { | ||||||||
3493 | NeedToScalarize = false; | ||||||||
3494 | return VectorCallCost; | ||||||||
3495 | } | ||||||||
3496 | return Cost; | ||||||||
3497 | } | ||||||||
3498 | |||||||||
3499 | unsigned LoopVectorizationCostModel::getVectorIntrinsicCost(CallInst *CI, | ||||||||
3500 | ElementCount VF) { | ||||||||
3501 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); | ||||||||
3502 | assert(ID && "Expected intrinsic call!")((ID && "Expected intrinsic call!") ? static_cast< void> (0) : __assert_fail ("ID && \"Expected intrinsic call!\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3502, __PRETTY_FUNCTION__)); | ||||||||
3503 | |||||||||
3504 | IntrinsicCostAttributes CostAttrs(ID, *CI, VF); | ||||||||
3505 | return TTI.getIntrinsicInstrCost(CostAttrs, | ||||||||
3506 | TargetTransformInfo::TCK_RecipThroughput); | ||||||||
3507 | } | ||||||||
3508 | |||||||||
3509 | static Type *smallestIntegerVectorType(Type *T1, Type *T2) { | ||||||||
3510 | auto *I1 = cast<IntegerType>(cast<VectorType>(T1)->getElementType()); | ||||||||
3511 | auto *I2 = cast<IntegerType>(cast<VectorType>(T2)->getElementType()); | ||||||||
3512 | return I1->getBitWidth() < I2->getBitWidth() ? T1 : T2; | ||||||||
3513 | } | ||||||||
3514 | |||||||||
3515 | static Type *largestIntegerVectorType(Type *T1, Type *T2) { | ||||||||
3516 | auto *I1 = cast<IntegerType>(cast<VectorType>(T1)->getElementType()); | ||||||||
3517 | auto *I2 = cast<IntegerType>(cast<VectorType>(T2)->getElementType()); | ||||||||
3518 | return I1->getBitWidth() > I2->getBitWidth() ? T1 : T2; | ||||||||
3519 | } | ||||||||
3520 | |||||||||
3521 | void InnerLoopVectorizer::truncateToMinimalBitwidths() { | ||||||||
3522 | // For every instruction `I` in MinBWs, truncate the operands, create a | ||||||||
3523 | // truncated version of `I` and reextend its result. InstCombine runs | ||||||||
3524 | // later and will remove any ext/trunc pairs. | ||||||||
3525 | SmallPtrSet<Value *, 4> Erased; | ||||||||
3526 | for (const auto &KV : Cost->getMinimalBitwidths()) { | ||||||||
3527 | // If the value wasn't vectorized, we must maintain the original scalar | ||||||||
3528 | // type. The absence of the value from VectorLoopValueMap indicates that it | ||||||||
3529 | // wasn't vectorized. | ||||||||
3530 | if (!VectorLoopValueMap.hasAnyVectorValue(KV.first)) | ||||||||
3531 | continue; | ||||||||
3532 | for (unsigned Part = 0; Part < UF; ++Part) { | ||||||||
3533 | Value *I = getOrCreateVectorValue(KV.first, Part); | ||||||||
3534 | if (Erased.count(I) || I->use_empty() || !isa<Instruction>(I)) | ||||||||
3535 | continue; | ||||||||
3536 | Type *OriginalTy = I->getType(); | ||||||||
3537 | Type *ScalarTruncatedTy = | ||||||||
3538 | IntegerType::get(OriginalTy->getContext(), KV.second); | ||||||||
3539 | auto *TruncatedTy = FixedVectorType::get( | ||||||||
3540 | ScalarTruncatedTy, | ||||||||
3541 | cast<FixedVectorType>(OriginalTy)->getNumElements()); | ||||||||
3542 | if (TruncatedTy == OriginalTy) | ||||||||
3543 | continue; | ||||||||
3544 | |||||||||
3545 | IRBuilder<> B(cast<Instruction>(I)); | ||||||||
3546 | auto ShrinkOperand = [&](Value *V) -> Value * { | ||||||||
3547 | if (auto *ZI = dyn_cast<ZExtInst>(V)) | ||||||||
3548 | if (ZI->getSrcTy() == TruncatedTy) | ||||||||
3549 | return ZI->getOperand(0); | ||||||||
3550 | return B.CreateZExtOrTrunc(V, TruncatedTy); | ||||||||
3551 | }; | ||||||||
3552 | |||||||||
3553 | // The actual instruction modification depends on the instruction type, | ||||||||
3554 | // unfortunately. | ||||||||
3555 | Value *NewI = nullptr; | ||||||||
3556 | if (auto *BO = dyn_cast<BinaryOperator>(I)) { | ||||||||
3557 | NewI = B.CreateBinOp(BO->getOpcode(), ShrinkOperand(BO->getOperand(0)), | ||||||||
3558 | ShrinkOperand(BO->getOperand(1))); | ||||||||
3559 | |||||||||
3560 | // Any wrapping introduced by shrinking this operation shouldn't be | ||||||||
3561 | // considered undefined behavior. So, we can't unconditionally copy | ||||||||
3562 | // arithmetic wrapping flags to NewI. | ||||||||
3563 | cast<BinaryOperator>(NewI)->copyIRFlags(I, /*IncludeWrapFlags=*/false); | ||||||||
3564 | } else if (auto *CI = dyn_cast<ICmpInst>(I)) { | ||||||||
3565 | NewI = | ||||||||
3566 | B.CreateICmp(CI->getPredicate(), ShrinkOperand(CI->getOperand(0)), | ||||||||
3567 | ShrinkOperand(CI->getOperand(1))); | ||||||||
3568 | } else if (auto *SI = dyn_cast<SelectInst>(I)) { | ||||||||
3569 | NewI = B.CreateSelect(SI->getCondition(), | ||||||||
3570 | ShrinkOperand(SI->getTrueValue()), | ||||||||
3571 | ShrinkOperand(SI->getFalseValue())); | ||||||||
3572 | } else if (auto *CI = dyn_cast<CastInst>(I)) { | ||||||||
3573 | switch (CI->getOpcode()) { | ||||||||
3574 | default: | ||||||||
3575 | llvm_unreachable("Unhandled cast!")::llvm::llvm_unreachable_internal("Unhandled cast!", "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3575); | ||||||||
3576 | case Instruction::Trunc: | ||||||||
3577 | NewI = ShrinkOperand(CI->getOperand(0)); | ||||||||
3578 | break; | ||||||||
3579 | case Instruction::SExt: | ||||||||
3580 | NewI = B.CreateSExtOrTrunc( | ||||||||
3581 | CI->getOperand(0), | ||||||||
3582 | smallestIntegerVectorType(OriginalTy, TruncatedTy)); | ||||||||
3583 | break; | ||||||||
3584 | case Instruction::ZExt: | ||||||||
3585 | NewI = B.CreateZExtOrTrunc( | ||||||||
3586 | CI->getOperand(0), | ||||||||
3587 | smallestIntegerVectorType(OriginalTy, TruncatedTy)); | ||||||||
3588 | break; | ||||||||
3589 | } | ||||||||
3590 | } else if (auto *SI = dyn_cast<ShuffleVectorInst>(I)) { | ||||||||
3591 | auto Elements0 = cast<FixedVectorType>(SI->getOperand(0)->getType()) | ||||||||
3592 | ->getNumElements(); | ||||||||
3593 | auto *O0 = B.CreateZExtOrTrunc( | ||||||||
3594 | SI->getOperand(0), | ||||||||
3595 | FixedVectorType::get(ScalarTruncatedTy, Elements0)); | ||||||||
3596 | auto Elements1 = cast<FixedVectorType>(SI->getOperand(1)->getType()) | ||||||||
3597 | ->getNumElements(); | ||||||||
3598 | auto *O1 = B.CreateZExtOrTrunc( | ||||||||
3599 | SI->getOperand(1), | ||||||||
3600 | FixedVectorType::get(ScalarTruncatedTy, Elements1)); | ||||||||
3601 | |||||||||
3602 | NewI = B.CreateShuffleVector(O0, O1, SI->getShuffleMask()); | ||||||||
3603 | } else if (isa<LoadInst>(I) || isa<PHINode>(I)) { | ||||||||
3604 | // Don't do anything with the operands, just extend the result. | ||||||||
3605 | continue; | ||||||||
3606 | } else if (auto *IE = dyn_cast<InsertElementInst>(I)) { | ||||||||
3607 | auto Elements = cast<FixedVectorType>(IE->getOperand(0)->getType()) | ||||||||
3608 | ->getNumElements(); | ||||||||
3609 | auto *O0 = B.CreateZExtOrTrunc( | ||||||||
3610 | IE->getOperand(0), | ||||||||
3611 | FixedVectorType::get(ScalarTruncatedTy, Elements)); | ||||||||
3612 | auto *O1 = B.CreateZExtOrTrunc(IE->getOperand(1), ScalarTruncatedTy); | ||||||||
3613 | NewI = B.CreateInsertElement(O0, O1, IE->getOperand(2)); | ||||||||
3614 | } else if (auto *EE = dyn_cast<ExtractElementInst>(I)) { | ||||||||
3615 | auto Elements = cast<FixedVectorType>(EE->getOperand(0)->getType()) | ||||||||
3616 | ->getNumElements(); | ||||||||
3617 | auto *O0 = B.CreateZExtOrTrunc( | ||||||||
3618 | EE->getOperand(0), | ||||||||
3619 | FixedVectorType::get(ScalarTruncatedTy, Elements)); | ||||||||
3620 | NewI = B.CreateExtractElement(O0, EE->getOperand(2)); | ||||||||
3621 | } else { | ||||||||
3622 | // If we don't know what to do, be conservative and don't do anything. | ||||||||
3623 | continue; | ||||||||
3624 | } | ||||||||
3625 | |||||||||
3626 | // Lastly, extend the result. | ||||||||
3627 | NewI->takeName(cast<Instruction>(I)); | ||||||||
3628 | Value *Res = B.CreateZExtOrTrunc(NewI, OriginalTy); | ||||||||
3629 | I->replaceAllUsesWith(Res); | ||||||||
3630 | cast<Instruction>(I)->eraseFromParent(); | ||||||||
3631 | Erased.insert(I); | ||||||||
3632 | VectorLoopValueMap.resetVectorValue(KV.first, Part, Res); | ||||||||
3633 | } | ||||||||
3634 | } | ||||||||
3635 | |||||||||
3636 | // We'll have created a bunch of ZExts that are now parentless. Clean up. | ||||||||
3637 | for (const auto &KV : Cost->getMinimalBitwidths()) { | ||||||||
3638 | // If the value wasn't vectorized, we must maintain the original scalar | ||||||||
3639 | // type. The absence of the value from VectorLoopValueMap indicates that it | ||||||||
3640 | // wasn't vectorized. | ||||||||
3641 | if (!VectorLoopValueMap.hasAnyVectorValue(KV.first)) | ||||||||
3642 | continue; | ||||||||
3643 | for (unsigned Part = 0; Part < UF; ++Part) { | ||||||||
3644 | Value *I = getOrCreateVectorValue(KV.first, Part); | ||||||||
3645 | ZExtInst *Inst = dyn_cast<ZExtInst>(I); | ||||||||
3646 | if (Inst && Inst->use_empty()) { | ||||||||
3647 | Value *NewI = Inst->getOperand(0); | ||||||||
3648 | Inst->eraseFromParent(); | ||||||||
3649 | VectorLoopValueMap.resetVectorValue(KV.first, Part, NewI); | ||||||||
3650 | } | ||||||||
3651 | } | ||||||||
3652 | } | ||||||||
3653 | } | ||||||||
3654 | |||||||||
3655 | void InnerLoopVectorizer::fixVectorizedLoop() { | ||||||||
3656 | // Insert truncates and extends for any truncated instructions as hints to | ||||||||
3657 | // InstCombine. | ||||||||
3658 | if (VF.isVector()) | ||||||||
3659 | truncateToMinimalBitwidths(); | ||||||||
3660 | |||||||||
3661 | // Fix widened non-induction PHIs by setting up the PHI operands. | ||||||||
3662 | if (OrigPHIsToFix.size()) { | ||||||||
3663 | assert(EnableVPlanNativePath &&((EnableVPlanNativePath && "Unexpected non-induction PHIs for fixup in non VPlan-native path" ) ? static_cast<void> (0) : __assert_fail ("EnableVPlanNativePath && \"Unexpected non-induction PHIs for fixup in non VPlan-native path\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3664, __PRETTY_FUNCTION__)) | ||||||||
3664 | "Unexpected non-induction PHIs for fixup in non VPlan-native path")((EnableVPlanNativePath && "Unexpected non-induction PHIs for fixup in non VPlan-native path" ) ? static_cast<void> (0) : __assert_fail ("EnableVPlanNativePath && \"Unexpected non-induction PHIs for fixup in non VPlan-native path\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3664, __PRETTY_FUNCTION__)); | ||||||||
3665 | fixNonInductionPHIs(); | ||||||||
3666 | } | ||||||||
3667 | |||||||||
3668 | // At this point every instruction in the original loop is widened to a | ||||||||
3669 | // vector form. Now we need to fix the recurrences in the loop. These PHI | ||||||||
3670 | // nodes are currently empty because we did not want to introduce cycles. | ||||||||
3671 | // This is the second stage of vectorizing recurrences. | ||||||||
3672 | fixCrossIterationPHIs(); | ||||||||
3673 | |||||||||
3674 | // Forget the original basic block. | ||||||||
3675 | PSE.getSE()->forgetLoop(OrigLoop); | ||||||||
3676 | |||||||||
3677 | // Fix-up external users of the induction variables. | ||||||||
3678 | for (auto &Entry : Legal->getInductionVars()) | ||||||||
3679 | fixupIVUsers(Entry.first, Entry.second, | ||||||||
3680 | getOrCreateVectorTripCount(LI->getLoopFor(LoopVectorBody)), | ||||||||
3681 | IVEndValues[Entry.first], LoopMiddleBlock); | ||||||||
3682 | |||||||||
3683 | fixLCSSAPHIs(); | ||||||||
3684 | for (Instruction *PI : PredicatedInstructions) | ||||||||
3685 | sinkScalarOperands(&*PI); | ||||||||
3686 | |||||||||
3687 | // Remove redundant induction instructions. | ||||||||
3688 | cse(LoopVectorBody); | ||||||||
3689 | |||||||||
3690 | // Set/update profile weights for the vector and remainder loops as original | ||||||||
3691 | // loop iterations are now distributed among them. Note that original loop | ||||||||
3692 | // represented by LoopScalarBody becomes remainder loop after vectorization. | ||||||||
3693 | // | ||||||||
3694 | // For cases like foldTailByMasking() and requiresScalarEpiloque() we may | ||||||||
3695 | // end up getting slightly roughened result but that should be OK since | ||||||||
3696 | // profile is not inherently precise anyway. Note also possible bypass of | ||||||||
3697 | // vector code caused by legality checks is ignored, assigning all the weight | ||||||||
3698 | // to the vector loop, optimistically. | ||||||||
3699 | assert(!VF.isScalable() &&((!VF.isScalable() && "cannot use scalable ElementCount to determine unroll factor" ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"cannot use scalable ElementCount to determine unroll factor\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3700, __PRETTY_FUNCTION__)) | ||||||||
3700 | "cannot use scalable ElementCount to determine unroll factor")((!VF.isScalable() && "cannot use scalable ElementCount to determine unroll factor" ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"cannot use scalable ElementCount to determine unroll factor\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3700, __PRETTY_FUNCTION__)); | ||||||||
3701 | setProfileInfoAfterUnrolling( | ||||||||
3702 | LI->getLoopFor(LoopScalarBody), LI->getLoopFor(LoopVectorBody), | ||||||||
3703 | LI->getLoopFor(LoopScalarBody), VF.getKnownMinValue() * UF); | ||||||||
3704 | } | ||||||||
3705 | |||||||||
3706 | void InnerLoopVectorizer::fixCrossIterationPHIs() { | ||||||||
3707 | // In order to support recurrences we need to be able to vectorize Phi nodes. | ||||||||
3708 | // Phi nodes have cycles, so we need to vectorize them in two stages. This is | ||||||||
3709 | // stage #2: We now need to fix the recurrences by adding incoming edges to | ||||||||
3710 | // the currently empty PHI nodes. At this point every instruction in the | ||||||||
3711 | // original loop is widened to a vector form so we can use them to construct | ||||||||
3712 | // the incoming edges. | ||||||||
3713 | for (PHINode &Phi : OrigLoop->getHeader()->phis()) { | ||||||||
3714 | // Handle first-order recurrences and reductions that need to be fixed. | ||||||||
3715 | if (Legal->isFirstOrderRecurrence(&Phi)) | ||||||||
3716 | fixFirstOrderRecurrence(&Phi); | ||||||||
3717 | else if (Legal->isReductionVariable(&Phi)) | ||||||||
3718 | fixReduction(&Phi); | ||||||||
3719 | } | ||||||||
3720 | } | ||||||||
3721 | |||||||||
3722 | void InnerLoopVectorizer::fixFirstOrderRecurrence(PHINode *Phi) { | ||||||||
3723 | // This is the second phase of vectorizing first-order recurrences. An | ||||||||
3724 | // overview of the transformation is described below. Suppose we have the | ||||||||
3725 | // following loop. | ||||||||
3726 | // | ||||||||
3727 | // for (int i = 0; i < n; ++i) | ||||||||
3728 | // b[i] = a[i] - a[i - 1]; | ||||||||
3729 | // | ||||||||
3730 | // There is a first-order recurrence on "a". For this loop, the shorthand | ||||||||
3731 | // scalar IR looks like: | ||||||||
3732 | // | ||||||||
3733 | // scalar.ph: | ||||||||
3734 | // s_init = a[-1] | ||||||||
3735 | // br scalar.body | ||||||||
3736 | // | ||||||||
3737 | // scalar.body: | ||||||||
3738 | // i = phi [0, scalar.ph], [i+1, scalar.body] | ||||||||
3739 | // s1 = phi [s_init, scalar.ph], [s2, scalar.body] | ||||||||
3740 | // s2 = a[i] | ||||||||
3741 | // b[i] = s2 - s1 | ||||||||
3742 | // br cond, scalar.body, ... | ||||||||
3743 | // | ||||||||
3744 | // In this example, s1 is a recurrence because it's value depends on the | ||||||||
3745 | // previous iteration. In the first phase of vectorization, we created a | ||||||||
3746 | // temporary value for s1. We now complete the vectorization and produce the | ||||||||
3747 | // shorthand vector IR shown below (for VF = 4, UF = 1). | ||||||||
3748 | // | ||||||||
3749 | // vector.ph: | ||||||||
3750 | // v_init = vector(..., ..., ..., a[-1]) | ||||||||
3751 | // br vector.body | ||||||||
3752 | // | ||||||||
3753 | // vector.body | ||||||||
3754 | // i = phi [0, vector.ph], [i+4, vector.body] | ||||||||
3755 | // v1 = phi [v_init, vector.ph], [v2, vector.body] | ||||||||
3756 | // v2 = a[i, i+1, i+2, i+3]; | ||||||||
3757 | // v3 = vector(v1(3), v2(0, 1, 2)) | ||||||||
3758 | // b[i, i+1, i+2, i+3] = v2 - v3 | ||||||||
3759 | // br cond, vector.body, middle.block | ||||||||
3760 | // | ||||||||
3761 | // middle.block: | ||||||||
3762 | // x = v2(3) | ||||||||
3763 | // br scalar.ph | ||||||||
3764 | // | ||||||||
3765 | // scalar.ph: | ||||||||
3766 | // s_init = phi [x, middle.block], [a[-1], otherwise] | ||||||||
3767 | // br scalar.body | ||||||||
3768 | // | ||||||||
3769 | // After execution completes the vector loop, we extract the next value of | ||||||||
3770 | // the recurrence (x) to use as the initial value in the scalar loop. | ||||||||
3771 | |||||||||
3772 | // Get the original loop preheader and single loop latch. | ||||||||
3773 | auto *Preheader = OrigLoop->getLoopPreheader(); | ||||||||
3774 | auto *Latch = OrigLoop->getLoopLatch(); | ||||||||
3775 | |||||||||
3776 | // Get the initial and previous values of the scalar recurrence. | ||||||||
3777 | auto *ScalarInit = Phi->getIncomingValueForBlock(Preheader); | ||||||||
3778 | auto *Previous = Phi->getIncomingValueForBlock(Latch); | ||||||||
3779 | |||||||||
3780 | // Create a vector from the initial value. | ||||||||
3781 | auto *VectorInit = ScalarInit; | ||||||||
3782 | if (VF.isVector()) { | ||||||||
3783 | Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator()); | ||||||||
3784 | assert(!VF.isScalable() && "VF is assumed to be non scalable.")((!VF.isScalable() && "VF is assumed to be non scalable." ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"VF is assumed to be non scalable.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3784, __PRETTY_FUNCTION__)); | ||||||||
3785 | VectorInit = Builder.CreateInsertElement( | ||||||||
3786 | UndefValue::get(VectorType::get(VectorInit->getType(), VF)), VectorInit, | ||||||||
3787 | Builder.getInt32(VF.getKnownMinValue() - 1), "vector.recur.init"); | ||||||||
3788 | } | ||||||||
3789 | |||||||||
3790 | // We constructed a temporary phi node in the first phase of vectorization. | ||||||||
3791 | // This phi node will eventually be deleted. | ||||||||
3792 | Builder.SetInsertPoint( | ||||||||
3793 | cast<Instruction>(VectorLoopValueMap.getVectorValue(Phi, 0))); | ||||||||
3794 | |||||||||
3795 | // Create a phi node for the new recurrence. The current value will either be | ||||||||
3796 | // the initial value inserted into a vector or loop-varying vector value. | ||||||||
3797 | auto *VecPhi = Builder.CreatePHI(VectorInit->getType(), 2, "vector.recur"); | ||||||||
3798 | VecPhi->addIncoming(VectorInit, LoopVectorPreHeader); | ||||||||
3799 | |||||||||
3800 | // Get the vectorized previous value of the last part UF - 1. It appears last | ||||||||
3801 | // among all unrolled iterations, due to the order of their construction. | ||||||||
3802 | Value *PreviousLastPart = getOrCreateVectorValue(Previous, UF - 1); | ||||||||
3803 | |||||||||
3804 | // Find and set the insertion point after the previous value if it is an | ||||||||
3805 | // instruction. | ||||||||
3806 | BasicBlock::iterator InsertPt; | ||||||||
3807 | // Note that the previous value may have been constant-folded so it is not | ||||||||
3808 | // guaranteed to be an instruction in the vector loop. | ||||||||
3809 | // FIXME: Loop invariant values do not form recurrences. We should deal with | ||||||||
3810 | // them earlier. | ||||||||
3811 | if (LI->getLoopFor(LoopVectorBody)->isLoopInvariant(PreviousLastPart)) | ||||||||
3812 | InsertPt = LoopVectorBody->getFirstInsertionPt(); | ||||||||
3813 | else { | ||||||||
3814 | Instruction *PreviousInst = cast<Instruction>(PreviousLastPart); | ||||||||
3815 | if (isa<PHINode>(PreviousLastPart)) | ||||||||
3816 | // If the previous value is a phi node, we should insert after all the phi | ||||||||
3817 | // nodes in the block containing the PHI to avoid breaking basic block | ||||||||
3818 | // verification. Note that the basic block may be different to | ||||||||
3819 | // LoopVectorBody, in case we predicate the loop. | ||||||||
3820 | InsertPt = PreviousInst->getParent()->getFirstInsertionPt(); | ||||||||
3821 | else | ||||||||
3822 | InsertPt = ++PreviousInst->getIterator(); | ||||||||
3823 | } | ||||||||
3824 | Builder.SetInsertPoint(&*InsertPt); | ||||||||
3825 | |||||||||
3826 | // We will construct a vector for the recurrence by combining the values for | ||||||||
3827 | // the current and previous iterations. This is the required shuffle mask. | ||||||||
3828 | assert(!VF.isScalable())((!VF.isScalable()) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable()", "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3828, __PRETTY_FUNCTION__)); | ||||||||
3829 | SmallVector<int, 8> ShuffleMask(VF.getKnownMinValue()); | ||||||||
3830 | ShuffleMask[0] = VF.getKnownMinValue() - 1; | ||||||||
3831 | for (unsigned I = 1; I < VF.getKnownMinValue(); ++I) | ||||||||
3832 | ShuffleMask[I] = I + VF.getKnownMinValue() - 1; | ||||||||
3833 | |||||||||
3834 | // The vector from which to take the initial value for the current iteration | ||||||||
3835 | // (actual or unrolled). Initially, this is the vector phi node. | ||||||||
3836 | Value *Incoming = VecPhi; | ||||||||
3837 | |||||||||
3838 | // Shuffle the current and previous vector and update the vector parts. | ||||||||
3839 | for (unsigned Part = 0; Part < UF; ++Part) { | ||||||||
3840 | Value *PreviousPart = getOrCreateVectorValue(Previous, Part); | ||||||||
3841 | Value *PhiPart = VectorLoopValueMap.getVectorValue(Phi, Part); | ||||||||
3842 | auto *Shuffle = | ||||||||
3843 | VF.isVector() | ||||||||
3844 | ? Builder.CreateShuffleVector(Incoming, PreviousPart, ShuffleMask) | ||||||||
3845 | : Incoming; | ||||||||
3846 | PhiPart->replaceAllUsesWith(Shuffle); | ||||||||
3847 | cast<Instruction>(PhiPart)->eraseFromParent(); | ||||||||
3848 | VectorLoopValueMap.resetVectorValue(Phi, Part, Shuffle); | ||||||||
3849 | Incoming = PreviousPart; | ||||||||
3850 | } | ||||||||
3851 | |||||||||
3852 | // Fix the latch value of the new recurrence in the vector loop. | ||||||||
3853 | VecPhi->addIncoming(Incoming, LI->getLoopFor(LoopVectorBody)->getLoopLatch()); | ||||||||
3854 | |||||||||
3855 | // Extract the last vector element in the middle block. This will be the | ||||||||
3856 | // initial value for the recurrence when jumping to the scalar loop. | ||||||||
3857 | auto *ExtractForScalar = Incoming; | ||||||||
3858 | if (VF.isVector()) { | ||||||||
3859 | Builder.SetInsertPoint(LoopMiddleBlock->getTerminator()); | ||||||||
3860 | ExtractForScalar = Builder.CreateExtractElement( | ||||||||
3861 | ExtractForScalar, Builder.getInt32(VF.getKnownMinValue() - 1), | ||||||||
3862 | "vector.recur.extract"); | ||||||||
3863 | } | ||||||||
3864 | // Extract the second last element in the middle block if the | ||||||||
3865 | // Phi is used outside the loop. We need to extract the phi itself | ||||||||
3866 | // and not the last element (the phi update in the current iteration). This | ||||||||
3867 | // will be the value when jumping to the exit block from the LoopMiddleBlock, | ||||||||
3868 | // when the scalar loop is not run at all. | ||||||||
3869 | Value *ExtractForPhiUsedOutsideLoop = nullptr; | ||||||||
3870 | if (VF.isVector()) | ||||||||
3871 | ExtractForPhiUsedOutsideLoop = Builder.CreateExtractElement( | ||||||||
3872 | Incoming, Builder.getInt32(VF.getKnownMinValue() - 2), | ||||||||
3873 | "vector.recur.extract.for.phi"); | ||||||||
3874 | // When loop is unrolled without vectorizing, initialize | ||||||||
3875 | // ExtractForPhiUsedOutsideLoop with the value just prior to unrolled value of | ||||||||
3876 | // `Incoming`. This is analogous to the vectorized case above: extracting the | ||||||||
3877 | // second last element when VF > 1. | ||||||||
3878 | else if (UF > 1) | ||||||||
3879 | ExtractForPhiUsedOutsideLoop = getOrCreateVectorValue(Previous, UF - 2); | ||||||||
3880 | |||||||||
3881 | // Fix the initial value of the original recurrence in the scalar loop. | ||||||||
3882 | Builder.SetInsertPoint(&*LoopScalarPreHeader->begin()); | ||||||||
3883 | auto *Start = Builder.CreatePHI(Phi->getType(), 2, "scalar.recur.init"); | ||||||||
3884 | for (auto *BB : predecessors(LoopScalarPreHeader)) { | ||||||||
3885 | auto *Incoming = BB == LoopMiddleBlock ? ExtractForScalar : ScalarInit; | ||||||||
3886 | Start->addIncoming(Incoming, BB); | ||||||||
3887 | } | ||||||||
3888 | |||||||||
3889 | Phi->setIncomingValueForBlock(LoopScalarPreHeader, Start); | ||||||||
3890 | Phi->setName("scalar.recur"); | ||||||||
3891 | |||||||||
3892 | // Finally, fix users of the recurrence outside the loop. The users will need | ||||||||
3893 | // either the last value of the scalar recurrence or the last value of the | ||||||||
3894 | // vector recurrence we extracted in the middle block. Since the loop is in | ||||||||
3895 | // LCSSA form, we just need to find all the phi nodes for the original scalar | ||||||||
3896 | // recurrence in the exit block, and then add an edge for the middle block. | ||||||||
3897 | for (PHINode &LCSSAPhi : LoopExitBlock->phis()) { | ||||||||
3898 | if (LCSSAPhi.getIncomingValue(0) == Phi) { | ||||||||
3899 | LCSSAPhi.addIncoming(ExtractForPhiUsedOutsideLoop, LoopMiddleBlock); | ||||||||
3900 | } | ||||||||
3901 | } | ||||||||
3902 | } | ||||||||
3903 | |||||||||
3904 | void InnerLoopVectorizer::fixReduction(PHINode *Phi) { | ||||||||
3905 | Constant *Zero = Builder.getInt32(0); | ||||||||
3906 | |||||||||
3907 | // Get it's reduction variable descriptor. | ||||||||
3908 | assert(Legal->isReductionVariable(Phi) &&((Legal->isReductionVariable(Phi) && "Unable to find the reduction variable" ) ? static_cast<void> (0) : __assert_fail ("Legal->isReductionVariable(Phi) && \"Unable to find the reduction variable\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3909, __PRETTY_FUNCTION__)) | ||||||||
3909 | "Unable to find the reduction variable")((Legal->isReductionVariable(Phi) && "Unable to find the reduction variable" ) ? static_cast<void> (0) : __assert_fail ("Legal->isReductionVariable(Phi) && \"Unable to find the reduction variable\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3909, __PRETTY_FUNCTION__)); | ||||||||
3910 | RecurrenceDescriptor RdxDesc = Legal->getReductionVars()[Phi]; | ||||||||
3911 | |||||||||
3912 | RecurrenceDescriptor::RecurrenceKind RK = RdxDesc.getRecurrenceKind(); | ||||||||
3913 | TrackingVH<Value> ReductionStartValue = RdxDesc.getRecurrenceStartValue(); | ||||||||
3914 | Instruction *LoopExitInst = RdxDesc.getLoopExitInstr(); | ||||||||
3915 | RecurrenceDescriptor::MinMaxRecurrenceKind MinMaxKind = | ||||||||
3916 | RdxDesc.getMinMaxRecurrenceKind(); | ||||||||
3917 | setDebugLocFromInst(Builder, ReductionStartValue); | ||||||||
3918 | bool IsInLoopReductionPhi = Cost->isInLoopReduction(Phi); | ||||||||
3919 | |||||||||
3920 | // We need to generate a reduction vector from the incoming scalar. | ||||||||
3921 | // To do so, we need to generate the 'identity' vector and override | ||||||||
3922 | // one of the elements with the incoming scalar reduction. We need | ||||||||
3923 | // to do it in the vector-loop preheader. | ||||||||
3924 | Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator()); | ||||||||
3925 | |||||||||
3926 | // This is the vector-clone of the value that leaves the loop. | ||||||||
3927 | Type *VecTy = getOrCreateVectorValue(LoopExitInst, 0)->getType(); | ||||||||
3928 | |||||||||
3929 | // Find the reduction identity variable. Zero for addition, or, xor, | ||||||||
3930 | // one for multiplication, -1 for And. | ||||||||
3931 | Value *Identity; | ||||||||
3932 | Value *VectorStart; | ||||||||
3933 | if (RK == RecurrenceDescriptor::RK_IntegerMinMax || | ||||||||
3934 | RK == RecurrenceDescriptor::RK_FloatMinMax) { | ||||||||
3935 | // MinMax reduction have the start value as their identify. | ||||||||
3936 | if (VF == 1 || IsInLoopReductionPhi) { | ||||||||
3937 | VectorStart = Identity = ReductionStartValue; | ||||||||
3938 | } else { | ||||||||
3939 | VectorStart = Identity = | ||||||||
3940 | Builder.CreateVectorSplat(VF, ReductionStartValue, "minmax.ident"); | ||||||||
3941 | } | ||||||||
3942 | } else { | ||||||||
3943 | // Handle other reduction kinds: | ||||||||
3944 | Constant *Iden = RecurrenceDescriptor::getRecurrenceIdentity( | ||||||||
3945 | RK, VecTy->getScalarType()); | ||||||||
3946 | if (VF == 1 || IsInLoopReductionPhi) { | ||||||||
3947 | Identity = Iden; | ||||||||
3948 | // This vector is the Identity vector where the first element is the | ||||||||
3949 | // incoming scalar reduction. | ||||||||
3950 | VectorStart = ReductionStartValue; | ||||||||
3951 | } else { | ||||||||
3952 | Identity = ConstantVector::getSplat(VF, Iden); | ||||||||
3953 | |||||||||
3954 | // This vector is the Identity vector where the first element is the | ||||||||
3955 | // incoming scalar reduction. | ||||||||
3956 | VectorStart = | ||||||||
3957 | Builder.CreateInsertElement(Identity, ReductionStartValue, Zero); | ||||||||
3958 | } | ||||||||
3959 | } | ||||||||
3960 | |||||||||
3961 | // Wrap flags are in general invalid after vectorization, clear them. | ||||||||
3962 | clearReductionWrapFlags(RdxDesc); | ||||||||
3963 | |||||||||
3964 | // Fix the vector-loop phi. | ||||||||
3965 | |||||||||
3966 | // Reductions do not have to start at zero. They can start with | ||||||||
3967 | // any loop invariant values. | ||||||||
3968 | BasicBlock *Latch = OrigLoop->getLoopLatch(); | ||||||||
3969 | Value *LoopVal = Phi->getIncomingValueForBlock(Latch); | ||||||||
3970 | |||||||||
3971 | for (unsigned Part = 0; Part < UF; ++Part) { | ||||||||
3972 | Value *VecRdxPhi = getOrCreateVectorValue(Phi, Part); | ||||||||
3973 | Value *Val = getOrCreateVectorValue(LoopVal, Part); | ||||||||
3974 | // Make sure to add the reduction start value only to the | ||||||||
3975 | // first unroll part. | ||||||||
3976 | Value *StartVal = (Part == 0) ? VectorStart : Identity; | ||||||||
3977 | cast<PHINode>(VecRdxPhi)->addIncoming(StartVal, LoopVectorPreHeader); | ||||||||
3978 | cast<PHINode>(VecRdxPhi) | ||||||||
3979 | ->addIncoming(Val, LI->getLoopFor(LoopVectorBody)->getLoopLatch()); | ||||||||
3980 | } | ||||||||
3981 | |||||||||
3982 | // Before each round, move the insertion point right between | ||||||||
3983 | // the PHIs and the values we are going to write. | ||||||||
3984 | // This allows us to write both PHINodes and the extractelement | ||||||||
3985 | // instructions. | ||||||||
3986 | Builder.SetInsertPoint(&*LoopMiddleBlock->getFirstInsertionPt()); | ||||||||
3987 | |||||||||
3988 | setDebugLocFromInst(Builder, LoopExitInst); | ||||||||
3989 | |||||||||
3990 | // If tail is folded by masking, the vector value to leave the loop should be | ||||||||
3991 | // a Select choosing between the vectorized LoopExitInst and vectorized Phi, | ||||||||
3992 | // instead of the former. | ||||||||
3993 | if (Cost->foldTailByMasking()) { | ||||||||
3994 | for (unsigned Part = 0; Part < UF; ++Part) { | ||||||||
3995 | Value *VecLoopExitInst = | ||||||||
3996 | VectorLoopValueMap.getVectorValue(LoopExitInst, Part); | ||||||||
3997 | Value *Sel = nullptr; | ||||||||
3998 | for (User *U : VecLoopExitInst->users()) { | ||||||||
3999 | if (isa<SelectInst>(U)) { | ||||||||
4000 | assert(!Sel && "Reduction exit feeding two selects")((!Sel && "Reduction exit feeding two selects") ? static_cast <void> (0) : __assert_fail ("!Sel && \"Reduction exit feeding two selects\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4000, __PRETTY_FUNCTION__)); | ||||||||
4001 | Sel = U; | ||||||||
4002 | } else | ||||||||
4003 | assert(isa<PHINode>(U) && "Reduction exit must feed Phi's or select")((isa<PHINode>(U) && "Reduction exit must feed Phi's or select" ) ? static_cast<void> (0) : __assert_fail ("isa<PHINode>(U) && \"Reduction exit must feed Phi's or select\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4003, __PRETTY_FUNCTION__)); | ||||||||
4004 | } | ||||||||
4005 | assert(Sel && "Reduction exit feeds no select")((Sel && "Reduction exit feeds no select") ? static_cast <void> (0) : __assert_fail ("Sel && \"Reduction exit feeds no select\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4005, __PRETTY_FUNCTION__)); | ||||||||
4006 | VectorLoopValueMap.resetVectorValue(LoopExitInst, Part, Sel); | ||||||||
4007 | |||||||||
4008 | // If the target can create a predicated operator for the reduction at no | ||||||||
4009 | // extra cost in the loop (for example a predicated vadd), it can be | ||||||||
4010 | // cheaper for the select to remain in the loop than be sunk out of it, | ||||||||
4011 | // and so use the select value for the phi instead of the old | ||||||||
4012 | // LoopExitValue. | ||||||||
4013 | RecurrenceDescriptor RdxDesc = Legal->getReductionVars()[Phi]; | ||||||||
4014 | if (PreferPredicatedReductionSelect || | ||||||||
4015 | TTI->preferPredicatedReductionSelect( | ||||||||
4016 | RdxDesc.getRecurrenceBinOp(RdxDesc.getRecurrenceKind()), | ||||||||
4017 | Phi->getType(), TargetTransformInfo::ReductionFlags())) { | ||||||||
4018 | auto *VecRdxPhi = cast<PHINode>(getOrCreateVectorValue(Phi, Part)); | ||||||||
4019 | VecRdxPhi->setIncomingValueForBlock( | ||||||||
4020 | LI->getLoopFor(LoopVectorBody)->getLoopLatch(), Sel); | ||||||||
4021 | } | ||||||||
4022 | } | ||||||||
4023 | } | ||||||||
4024 | |||||||||
4025 | // If the vector reduction can be performed in a smaller type, we truncate | ||||||||
4026 | // then extend the loop exit value to enable InstCombine to evaluate the | ||||||||
4027 | // entire expression in the smaller type. | ||||||||
4028 | if (VF.isVector() && Phi->getType() != RdxDesc.getRecurrenceType()) { | ||||||||
4029 | assert(!IsInLoopReductionPhi && "Unexpected truncated inloop reduction!")((!IsInLoopReductionPhi && "Unexpected truncated inloop reduction!" ) ? static_cast<void> (0) : __assert_fail ("!IsInLoopReductionPhi && \"Unexpected truncated inloop reduction!\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4029, __PRETTY_FUNCTION__)); | ||||||||
4030 | assert(!VF.isScalable() && "scalable vectors not yet supported.")((!VF.isScalable() && "scalable vectors not yet supported." ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"scalable vectors not yet supported.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4030, __PRETTY_FUNCTION__)); | ||||||||
4031 | Type *RdxVecTy = VectorType::get(RdxDesc.getRecurrenceType(), VF); | ||||||||
4032 | Builder.SetInsertPoint( | ||||||||
4033 | LI->getLoopFor(LoopVectorBody)->getLoopLatch()->getTerminator()); | ||||||||
4034 | VectorParts RdxParts(UF); | ||||||||
4035 | for (unsigned Part = 0; Part < UF; ++Part) { | ||||||||
4036 | RdxParts[Part] = VectorLoopValueMap.getVectorValue(LoopExitInst, Part); | ||||||||
4037 | Value *Trunc = Builder.CreateTrunc(RdxParts[Part], RdxVecTy); | ||||||||
4038 | Value *Extnd = RdxDesc.isSigned() ? Builder.CreateSExt(Trunc, VecTy) | ||||||||
4039 | : Builder.CreateZExt(Trunc, VecTy); | ||||||||
4040 | for (Value::user_iterator UI = RdxParts[Part]->user_begin(); | ||||||||
4041 | UI != RdxParts[Part]->user_end();) | ||||||||
4042 | if (*UI != Trunc) { | ||||||||
4043 | (*UI++)->replaceUsesOfWith(RdxParts[Part], Extnd); | ||||||||
4044 | RdxParts[Part] = Extnd; | ||||||||
4045 | } else { | ||||||||
4046 | ++UI; | ||||||||
4047 | } | ||||||||
4048 | } | ||||||||
4049 | Builder.SetInsertPoint(&*LoopMiddleBlock->getFirstInsertionPt()); | ||||||||
4050 | for (unsigned Part = 0; Part < UF; ++Part) { | ||||||||
4051 | RdxParts[Part] = Builder.CreateTrunc(RdxParts[Part], RdxVecTy); | ||||||||
4052 | VectorLoopValueMap.resetVectorValue(LoopExitInst, Part, RdxParts[Part]); | ||||||||
4053 | } | ||||||||
4054 | } | ||||||||
4055 | |||||||||
4056 | // Reduce all of the unrolled parts into a single vector. | ||||||||
4057 | Value *ReducedPartRdx = VectorLoopValueMap.getVectorValue(LoopExitInst, 0); | ||||||||
4058 | unsigned Op = RecurrenceDescriptor::getRecurrenceBinOp(RK); | ||||||||
4059 | |||||||||
4060 | // The middle block terminator has already been assigned a DebugLoc here (the | ||||||||
4061 | // OrigLoop's single latch terminator). We want the whole middle block to | ||||||||
4062 | // appear to execute on this line because: (a) it is all compiler generated, | ||||||||
4063 | // (b) these instructions are always executed after evaluating the latch | ||||||||
4064 | // conditional branch, and (c) other passes may add new predecessors which | ||||||||
4065 | // terminate on this line. This is the easiest way to ensure we don't | ||||||||
4066 | // accidentally cause an extra step back into the loop while debugging. | ||||||||
4067 | setDebugLocFromInst(Builder, LoopMiddleBlock->getTerminator()); | ||||||||
4068 | for (unsigned Part = 1; Part < UF; ++Part) { | ||||||||
4069 | Value *RdxPart = VectorLoopValueMap.getVectorValue(LoopExitInst, Part); | ||||||||
4070 | if (Op != Instruction::ICmp && Op != Instruction::FCmp) | ||||||||
4071 | // Floating point operations had to be 'fast' to enable the reduction. | ||||||||
4072 | ReducedPartRdx = addFastMathFlag( | ||||||||
4073 | Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxPart, | ||||||||
4074 | ReducedPartRdx, "bin.rdx"), | ||||||||
4075 | RdxDesc.getFastMathFlags()); | ||||||||
4076 | else | ||||||||
4077 | ReducedPartRdx = createMinMaxOp(Builder, MinMaxKind, ReducedPartRdx, | ||||||||
4078 | RdxPart); | ||||||||
4079 | } | ||||||||
4080 | |||||||||
4081 | // Create the reduction after the loop. Note that inloop reductions create the | ||||||||
4082 | // target reduction in the loop using a Reduction recipe. | ||||||||
4083 | if (VF.isVector() && !IsInLoopReductionPhi) { | ||||||||
4084 | bool NoNaN = Legal->hasFunNoNaNAttr(); | ||||||||
4085 | ReducedPartRdx = | ||||||||
4086 | createTargetReduction(Builder, TTI, RdxDesc, ReducedPartRdx, NoNaN); | ||||||||
4087 | // If the reduction can be performed in a smaller type, we need to extend | ||||||||
4088 | // the reduction to the wider type before we branch to the original loop. | ||||||||
4089 | if (Phi->getType() != RdxDesc.getRecurrenceType()) | ||||||||
4090 | ReducedPartRdx = | ||||||||
4091 | RdxDesc.isSigned() | ||||||||
4092 | ? Builder.CreateSExt(ReducedPartRdx, Phi->getType()) | ||||||||
4093 | : Builder.CreateZExt(ReducedPartRdx, Phi->getType()); | ||||||||
4094 | } | ||||||||
4095 | |||||||||
4096 | // Create a phi node that merges control-flow from the backedge-taken check | ||||||||
4097 | // block and the middle block. | ||||||||
4098 | PHINode *BCBlockPhi = PHINode::Create(Phi->getType(), 2, "bc.merge.rdx", | ||||||||
4099 | LoopScalarPreHeader->getTerminator()); | ||||||||
4100 | for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I) | ||||||||
4101 | BCBlockPhi->addIncoming(ReductionStartValue, LoopBypassBlocks[I]); | ||||||||
4102 | BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock); | ||||||||
4103 | |||||||||
4104 | // Now, we need to fix the users of the reduction variable | ||||||||
4105 | // inside and outside of the scalar remainder loop. | ||||||||
4106 | // We know that the loop is in LCSSA form. We need to update the | ||||||||
4107 | // PHI nodes in the exit blocks. | ||||||||
4108 | for (PHINode &LCSSAPhi : LoopExitBlock->phis()) { | ||||||||
4109 | // All PHINodes need to have a single entry edge, or two if | ||||||||
4110 | // we already fixed them. | ||||||||
4111 | assert(LCSSAPhi.getNumIncomingValues() < 3 && "Invalid LCSSA PHI")((LCSSAPhi.getNumIncomingValues() < 3 && "Invalid LCSSA PHI" ) ? static_cast<void> (0) : __assert_fail ("LCSSAPhi.getNumIncomingValues() < 3 && \"Invalid LCSSA PHI\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4111, __PRETTY_FUNCTION__)); | ||||||||
4112 | |||||||||
4113 | // We found a reduction value exit-PHI. Update it with the | ||||||||
4114 | // incoming bypass edge. | ||||||||
4115 | if (LCSSAPhi.getIncomingValue(0) == LoopExitInst) | ||||||||
4116 | LCSSAPhi.addIncoming(ReducedPartRdx, LoopMiddleBlock); | ||||||||
4117 | } // end of the LCSSA phi scan. | ||||||||
4118 | |||||||||
4119 | // Fix the scalar loop reduction variable with the incoming reduction sum | ||||||||
4120 | // from the vector body and from the backedge value. | ||||||||
4121 | int IncomingEdgeBlockIdx = | ||||||||
4122 | Phi->getBasicBlockIndex(OrigLoop->getLoopLatch()); | ||||||||
4123 | assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index")((IncomingEdgeBlockIdx >= 0 && "Invalid block index" ) ? static_cast<void> (0) : __assert_fail ("IncomingEdgeBlockIdx >= 0 && \"Invalid block index\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4123, __PRETTY_FUNCTION__)); | ||||||||
4124 | // Pick the other block. | ||||||||
4125 | int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1); | ||||||||
4126 | Phi->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi); | ||||||||
4127 | Phi->setIncomingValue(IncomingEdgeBlockIdx, LoopExitInst); | ||||||||
4128 | } | ||||||||
4129 | |||||||||
4130 | void InnerLoopVectorizer::clearReductionWrapFlags( | ||||||||
4131 | RecurrenceDescriptor &RdxDesc) { | ||||||||
4132 | RecurrenceDescriptor::RecurrenceKind RK = RdxDesc.getRecurrenceKind(); | ||||||||
4133 | if (RK != RecurrenceDescriptor::RK_IntegerAdd && | ||||||||
4134 | RK != RecurrenceDescriptor::RK_IntegerMult) | ||||||||
4135 | return; | ||||||||
4136 | |||||||||
4137 | Instruction *LoopExitInstr = RdxDesc.getLoopExitInstr(); | ||||||||
4138 | assert(LoopExitInstr && "null loop exit instruction")((LoopExitInstr && "null loop exit instruction") ? static_cast <void> (0) : __assert_fail ("LoopExitInstr && \"null loop exit instruction\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4138, __PRETTY_FUNCTION__)); | ||||||||
4139 | SmallVector<Instruction *, 8> Worklist; | ||||||||
4140 | SmallPtrSet<Instruction *, 8> Visited; | ||||||||
4141 | Worklist.push_back(LoopExitInstr); | ||||||||
4142 | Visited.insert(LoopExitInstr); | ||||||||
4143 | |||||||||
4144 | while (!Worklist.empty()) { | ||||||||
4145 | Instruction *Cur = Worklist.pop_back_val(); | ||||||||
4146 | if (isa<OverflowingBinaryOperator>(Cur)) | ||||||||
4147 | for (unsigned Part = 0; Part < UF; ++Part) { | ||||||||
4148 | Value *V = getOrCreateVectorValue(Cur, Part); | ||||||||
4149 | cast<Instruction>(V)->dropPoisonGeneratingFlags(); | ||||||||
4150 | } | ||||||||
4151 | |||||||||
4152 | for (User *U : Cur->users()) { | ||||||||
4153 | Instruction *UI = cast<Instruction>(U); | ||||||||
4154 | if ((Cur != LoopExitInstr || OrigLoop->contains(UI->getParent())) && | ||||||||
4155 | Visited.insert(UI).second) | ||||||||
4156 | Worklist.push_back(UI); | ||||||||
4157 | } | ||||||||
4158 | } | ||||||||
4159 | } | ||||||||
4160 | |||||||||
4161 | void InnerLoopVectorizer::fixLCSSAPHIs() { | ||||||||
4162 | assert(!VF.isScalable() && "the code below assumes fixed width vectors")((!VF.isScalable() && "the code below assumes fixed width vectors" ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"the code below assumes fixed width vectors\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4162, __PRETTY_FUNCTION__)); | ||||||||
4163 | for (PHINode &LCSSAPhi : LoopExitBlock->phis()) { | ||||||||
4164 | if (LCSSAPhi.getNumIncomingValues() == 1) { | ||||||||
4165 | auto *IncomingValue = LCSSAPhi.getIncomingValue(0); | ||||||||
4166 | // Non-instruction incoming values will have only one value. | ||||||||
4167 | unsigned LastLane = 0; | ||||||||
4168 | if (isa<Instruction>(IncomingValue)) | ||||||||
4169 | LastLane = Cost->isUniformAfterVectorization( | ||||||||
4170 | cast<Instruction>(IncomingValue), VF) | ||||||||
4171 | ? 0 | ||||||||
4172 | : VF.getKnownMinValue() - 1; | ||||||||
4173 | // Can be a loop invariant incoming value or the last scalar value to be | ||||||||
4174 | // extracted from the vectorized loop. | ||||||||
4175 | Builder.SetInsertPoint(LoopMiddleBlock->getTerminator()); | ||||||||
4176 | Value *lastIncomingValue = | ||||||||
4177 | getOrCreateScalarValue(IncomingValue, { UF - 1, LastLane }); | ||||||||
4178 | LCSSAPhi.addIncoming(lastIncomingValue, LoopMiddleBlock); | ||||||||
4179 | } | ||||||||
4180 | } | ||||||||
4181 | } | ||||||||
4182 | |||||||||
4183 | void InnerLoopVectorizer::sinkScalarOperands(Instruction *PredInst) { | ||||||||
4184 | // The basic block and loop containing the predicated instruction. | ||||||||
4185 | auto *PredBB = PredInst->getParent(); | ||||||||
4186 | auto *VectorLoop = LI->getLoopFor(PredBB); | ||||||||
4187 | |||||||||
4188 | // Initialize a worklist with the operands of the predicated instruction. | ||||||||
4189 | SetVector<Value *> Worklist(PredInst->op_begin(), PredInst->op_end()); | ||||||||
4190 | |||||||||
4191 | // Holds instructions that we need to analyze again. An instruction may be | ||||||||
4192 | // reanalyzed if we don't yet know if we can sink it or not. | ||||||||
4193 | SmallVector<Instruction *, 8> InstsToReanalyze; | ||||||||
4194 | |||||||||
4195 | // Returns true if a given use occurs in the predicated block. Phi nodes use | ||||||||
4196 | // their operands in their corresponding predecessor blocks. | ||||||||
4197 | auto isBlockOfUsePredicated = [&](Use &U) -> bool { | ||||||||
4198 | auto *I = cast<Instruction>(U.getUser()); | ||||||||
4199 | BasicBlock *BB = I->getParent(); | ||||||||
4200 | if (auto *Phi = dyn_cast<PHINode>(I)) | ||||||||
4201 | BB = Phi->getIncomingBlock( | ||||||||
4202 | PHINode::getIncomingValueNumForOperand(U.getOperandNo())); | ||||||||
4203 | return BB == PredBB; | ||||||||
4204 | }; | ||||||||
4205 | |||||||||
4206 | // Iteratively sink the scalarized operands of the predicated instruction | ||||||||
4207 | // into the block we created for it. When an instruction is sunk, it's | ||||||||
4208 | // operands are then added to the worklist. The algorithm ends after one pass | ||||||||
4209 | // through the worklist doesn't sink a single instruction. | ||||||||
4210 | bool Changed; | ||||||||
4211 | do { | ||||||||
4212 | // Add the instructions that need to be reanalyzed to the worklist, and | ||||||||
4213 | // reset the changed indicator. | ||||||||
4214 | Worklist.insert(InstsToReanalyze.begin(), InstsToReanalyze.end()); | ||||||||
4215 | InstsToReanalyze.clear(); | ||||||||
4216 | Changed = false; | ||||||||
4217 | |||||||||
4218 | while (!Worklist.empty()) { | ||||||||
4219 | auto *I = dyn_cast<Instruction>(Worklist.pop_back_val()); | ||||||||
4220 | |||||||||
4221 | // We can't sink an instruction if it is a phi node, is already in the | ||||||||
4222 | // predicated block, is not in the loop, or may have side effects. | ||||||||
4223 | if (!I || isa<PHINode>(I) || I->getParent() == PredBB || | ||||||||
4224 | !VectorLoop->contains(I) || I->mayHaveSideEffects()) | ||||||||
4225 | continue; | ||||||||
4226 | |||||||||
4227 | // It's legal to sink the instruction if all its uses occur in the | ||||||||
4228 | // predicated block. Otherwise, there's nothing to do yet, and we may | ||||||||
4229 | // need to reanalyze the instruction. | ||||||||
4230 | if (!llvm::all_of(I->uses(), isBlockOfUsePredicated)) { | ||||||||
4231 | InstsToReanalyze.push_back(I); | ||||||||
4232 | continue; | ||||||||
4233 | } | ||||||||
4234 | |||||||||
4235 | // Move the instruction to the beginning of the predicated block, and add | ||||||||
4236 | // it's operands to the worklist. | ||||||||
4237 | I->moveBefore(&*PredBB->getFirstInsertionPt()); | ||||||||
4238 | Worklist.insert(I->op_begin(), I->op_end()); | ||||||||
4239 | |||||||||
4240 | // The sinking may have enabled other instructions to be sunk, so we will | ||||||||
4241 | // need to iterate. | ||||||||
4242 | Changed = true; | ||||||||
4243 | } | ||||||||
4244 | } while (Changed); | ||||||||
4245 | } | ||||||||
4246 | |||||||||
4247 | void InnerLoopVectorizer::fixNonInductionPHIs() { | ||||||||
4248 | for (PHINode *OrigPhi : OrigPHIsToFix) { | ||||||||
4249 | PHINode *NewPhi = | ||||||||
4250 | cast<PHINode>(VectorLoopValueMap.getVectorValue(OrigPhi, 0)); | ||||||||
4251 | unsigned NumIncomingValues = OrigPhi->getNumIncomingValues(); | ||||||||
4252 | |||||||||
4253 | SmallVector<BasicBlock *, 2> ScalarBBPredecessors( | ||||||||
4254 | predecessors(OrigPhi->getParent())); | ||||||||
4255 | SmallVector<BasicBlock *, 2> VectorBBPredecessors( | ||||||||
4256 | predecessors(NewPhi->getParent())); | ||||||||
4257 | assert(ScalarBBPredecessors.size() == VectorBBPredecessors.size() &&((ScalarBBPredecessors.size() == VectorBBPredecessors.size() && "Scalar and Vector BB should have the same number of predecessors" ) ? static_cast<void> (0) : __assert_fail ("ScalarBBPredecessors.size() == VectorBBPredecessors.size() && \"Scalar and Vector BB should have the same number of predecessors\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4258, __PRETTY_FUNCTION__)) | ||||||||
4258 | "Scalar and Vector BB should have the same number of predecessors")((ScalarBBPredecessors.size() == VectorBBPredecessors.size() && "Scalar and Vector BB should have the same number of predecessors" ) ? static_cast<void> (0) : __assert_fail ("ScalarBBPredecessors.size() == VectorBBPredecessors.size() && \"Scalar and Vector BB should have the same number of predecessors\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4258, __PRETTY_FUNCTION__)); | ||||||||
4259 | |||||||||
4260 | // The insertion point in Builder may be invalidated by the time we get | ||||||||
4261 | // here. Force the Builder insertion point to something valid so that we do | ||||||||
4262 | // not run into issues during insertion point restore in | ||||||||
4263 | // getOrCreateVectorValue calls below. | ||||||||
4264 | Builder.SetInsertPoint(NewPhi); | ||||||||
4265 | |||||||||
4266 | // The predecessor order is preserved and we can rely on mapping between | ||||||||
4267 | // scalar and vector block predecessors. | ||||||||
4268 | for (unsigned i = 0; i < NumIncomingValues; ++i) { | ||||||||
4269 | BasicBlock *NewPredBB = VectorBBPredecessors[i]; | ||||||||
4270 | |||||||||
4271 | // When looking up the new scalar/vector values to fix up, use incoming | ||||||||
4272 | // values from original phi. | ||||||||
4273 | Value *ScIncV = | ||||||||
4274 | OrigPhi->getIncomingValueForBlock(ScalarBBPredecessors[i]); | ||||||||
4275 | |||||||||
4276 | // Scalar incoming value may need a broadcast | ||||||||
4277 | Value *NewIncV = getOrCreateVectorValue(ScIncV, 0); | ||||||||
4278 | NewPhi->addIncoming(NewIncV, NewPredBB); | ||||||||
4279 | } | ||||||||
4280 | } | ||||||||
4281 | } | ||||||||
4282 | |||||||||
4283 | void InnerLoopVectorizer::widenGEP(GetElementPtrInst *GEP, VPUser &Operands, | ||||||||
4284 | unsigned UF, ElementCount VF, | ||||||||
4285 | bool IsPtrLoopInvariant, | ||||||||
4286 | SmallBitVector &IsIndexLoopInvariant, | ||||||||
4287 | VPTransformState &State) { | ||||||||
4288 | // Construct a vector GEP by widening the operands of the scalar GEP as | ||||||||
4289 | // necessary. We mark the vector GEP 'inbounds' if appropriate. A GEP | ||||||||
4290 | // results in a vector of pointers when at least one operand of the GEP | ||||||||
4291 | // is vector-typed. Thus, to keep the representation compact, we only use | ||||||||
4292 | // vector-typed operands for loop-varying values. | ||||||||
4293 | |||||||||
4294 | if (VF.isVector() && IsPtrLoopInvariant && IsIndexLoopInvariant.all()) { | ||||||||
4295 | // If we are vectorizing, but the GEP has only loop-invariant operands, | ||||||||
4296 | // the GEP we build (by only using vector-typed operands for | ||||||||
4297 | // loop-varying values) would be a scalar pointer. Thus, to ensure we | ||||||||
4298 | // produce a vector of pointers, we need to either arbitrarily pick an | ||||||||
4299 | // operand to broadcast, or broadcast a clone of the original GEP. | ||||||||
4300 | // Here, we broadcast a clone of the original. | ||||||||
4301 | // | ||||||||
4302 | // TODO: If at some point we decide to scalarize instructions having | ||||||||
4303 | // loop-invariant operands, this special case will no longer be | ||||||||
4304 | // required. We would add the scalarization decision to | ||||||||
4305 | // collectLoopScalars() and teach getVectorValue() to broadcast | ||||||||
4306 | // the lane-zero scalar value. | ||||||||
4307 | auto *Clone = Builder.Insert(GEP->clone()); | ||||||||
4308 | for (unsigned Part = 0; Part < UF; ++Part) { | ||||||||
4309 | Value *EntryPart = Builder.CreateVectorSplat(VF, Clone); | ||||||||
4310 | VectorLoopValueMap.setVectorValue(GEP, Part, EntryPart); | ||||||||
4311 | addMetadata(EntryPart, GEP); | ||||||||
4312 | } | ||||||||
4313 | } else { | ||||||||
4314 | // If the GEP has at least one loop-varying operand, we are sure to | ||||||||
4315 | // produce a vector of pointers. But if we are only unrolling, we want | ||||||||
4316 | // to produce a scalar GEP for each unroll part. Thus, the GEP we | ||||||||
4317 | // produce with the code below will be scalar (if VF == 1) or vector | ||||||||
4318 | // (otherwise). Note that for the unroll-only case, we still maintain | ||||||||
4319 | // values in the vector mapping with initVector, as we do for other | ||||||||
4320 | // instructions. | ||||||||
4321 | for (unsigned Part = 0; Part < UF; ++Part) { | ||||||||
4322 | // The pointer operand of the new GEP. If it's loop-invariant, we | ||||||||
4323 | // won't broadcast it. | ||||||||
4324 | auto *Ptr = IsPtrLoopInvariant ? State.get(Operands.getOperand(0), {0, 0}) | ||||||||
4325 | : State.get(Operands.getOperand(0), Part); | ||||||||
4326 | |||||||||
4327 | // Collect all the indices for the new GEP. If any index is | ||||||||
4328 | // loop-invariant, we won't broadcast it. | ||||||||
4329 | SmallVector<Value *, 4> Indices; | ||||||||
4330 | for (unsigned I = 1, E = Operands.getNumOperands(); I < E; I++) { | ||||||||
4331 | VPValue *Operand = Operands.getOperand(I); | ||||||||
4332 | if (IsIndexLoopInvariant[I - 1]) | ||||||||
4333 | Indices.push_back(State.get(Operand, {0, 0})); | ||||||||
4334 | else | ||||||||
4335 | Indices.push_back(State.get(Operand, Part)); | ||||||||
4336 | } | ||||||||
4337 | |||||||||
4338 | // Create the new GEP. Note that this GEP may be a scalar if VF == 1, | ||||||||
4339 | // but it should be a vector, otherwise. | ||||||||
4340 | auto *NewGEP = | ||||||||
4341 | GEP->isInBounds() | ||||||||
4342 | ? Builder.CreateInBoundsGEP(GEP->getSourceElementType(), Ptr, | ||||||||
4343 | Indices) | ||||||||
4344 | : Builder.CreateGEP(GEP->getSourceElementType(), Ptr, Indices); | ||||||||
4345 | assert((VF == 1 || NewGEP->getType()->isVectorTy()) &&(((VF == 1 || NewGEP->getType()->isVectorTy()) && "NewGEP is not a pointer vector") ? static_cast<void> ( 0) : __assert_fail ("(VF == 1 || NewGEP->getType()->isVectorTy()) && \"NewGEP is not a pointer vector\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4346, __PRETTY_FUNCTION__)) | ||||||||
4346 | "NewGEP is not a pointer vector")(((VF == 1 || NewGEP->getType()->isVectorTy()) && "NewGEP is not a pointer vector") ? static_cast<void> ( 0) : __assert_fail ("(VF == 1 || NewGEP->getType()->isVectorTy()) && \"NewGEP is not a pointer vector\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4346, __PRETTY_FUNCTION__)); | ||||||||
4347 | VectorLoopValueMap.setVectorValue(GEP, Part, NewGEP); | ||||||||
4348 | addMetadata(NewGEP, GEP); | ||||||||
4349 | } | ||||||||
4350 | } | ||||||||
4351 | } | ||||||||
4352 | |||||||||
4353 | void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN, unsigned UF, | ||||||||
4354 | ElementCount VF) { | ||||||||
4355 | assert(!VF.isScalable() && "scalable vectors not yet supported.")((!VF.isScalable() && "scalable vectors not yet supported." ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"scalable vectors not yet supported.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4355, __PRETTY_FUNCTION__)); | ||||||||
4356 | PHINode *P = cast<PHINode>(PN); | ||||||||
4357 | if (EnableVPlanNativePath) { | ||||||||
4358 | // Currently we enter here in the VPlan-native path for non-induction | ||||||||
4359 | // PHIs where all control flow is uniform. We simply widen these PHIs. | ||||||||
4360 | // Create a vector phi with no operands - the vector phi operands will be | ||||||||
4361 | // set at the end of vector code generation. | ||||||||
4362 | Type *VecTy = | ||||||||
4363 | (VF.isScalar()) ? PN->getType() : VectorType::get(PN->getType(), VF); | ||||||||
4364 | Value *VecPhi = Builder.CreatePHI(VecTy, PN->getNumOperands(), "vec.phi"); | ||||||||
4365 | VectorLoopValueMap.setVectorValue(P, 0, VecPhi); | ||||||||
4366 | OrigPHIsToFix.push_back(P); | ||||||||
4367 | |||||||||
4368 | return; | ||||||||
4369 | } | ||||||||
4370 | |||||||||
4371 | assert(PN->getParent() == OrigLoop->getHeader() &&((PN->getParent() == OrigLoop->getHeader() && "Non-header phis should have been handled elsewhere" ) ? static_cast<void> (0) : __assert_fail ("PN->getParent() == OrigLoop->getHeader() && \"Non-header phis should have been handled elsewhere\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4372, __PRETTY_FUNCTION__)) | ||||||||
4372 | "Non-header phis should have been handled elsewhere")((PN->getParent() == OrigLoop->getHeader() && "Non-header phis should have been handled elsewhere" ) ? static_cast<void> (0) : __assert_fail ("PN->getParent() == OrigLoop->getHeader() && \"Non-header phis should have been handled elsewhere\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4372, __PRETTY_FUNCTION__)); | ||||||||
4373 | |||||||||
4374 | // In order to support recurrences we need to be able to vectorize Phi nodes. | ||||||||
4375 | // Phi nodes have cycles, so we need to vectorize them in two stages. This is | ||||||||
4376 | // stage #1: We create a new vector PHI node with no incoming edges. We'll use | ||||||||
4377 | // this value when we vectorize all of the instructions that use the PHI. | ||||||||
4378 | if (Legal->isReductionVariable(P) || Legal->isFirstOrderRecurrence(P)) { | ||||||||
4379 | for (unsigned Part = 0; Part < UF; ++Part) { | ||||||||
4380 | // This is phase one of vectorizing PHIs. | ||||||||
4381 | bool ScalarPHI = | ||||||||
4382 | (VF.isScalar()) || Cost->isInLoopReduction(cast<PHINode>(PN)); | ||||||||
4383 | Type *VecTy = | ||||||||
4384 | ScalarPHI ? PN->getType() : VectorType::get(PN->getType(), VF); | ||||||||
4385 | Value *EntryPart = PHINode::Create( | ||||||||
4386 | VecTy, 2, "vec.phi", &*LoopVectorBody->getFirstInsertionPt()); | ||||||||
4387 | VectorLoopValueMap.setVectorValue(P, Part, EntryPart); | ||||||||
4388 | } | ||||||||
4389 | return; | ||||||||
4390 | } | ||||||||
4391 | |||||||||
4392 | setDebugLocFromInst(Builder, P); | ||||||||
4393 | |||||||||
4394 | // This PHINode must be an induction variable. | ||||||||
4395 | // Make sure that we know about it. | ||||||||
4396 | assert(Legal->getInductionVars().count(P) && "Not an induction variable")((Legal->getInductionVars().count(P) && "Not an induction variable" ) ? static_cast<void> (0) : __assert_fail ("Legal->getInductionVars().count(P) && \"Not an induction variable\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4396, __PRETTY_FUNCTION__)); | ||||||||
4397 | |||||||||
4398 | InductionDescriptor II = Legal->getInductionVars().lookup(P); | ||||||||
4399 | const DataLayout &DL = OrigLoop->getHeader()->getModule()->getDataLayout(); | ||||||||
4400 | |||||||||
4401 | // FIXME: The newly created binary instructions should contain nsw/nuw flags, | ||||||||
4402 | // which can be found from the original scalar operations. | ||||||||
4403 | switch (II.getKind()) { | ||||||||
4404 | case InductionDescriptor::IK_NoInduction: | ||||||||
4405 | llvm_unreachable("Unknown induction")::llvm::llvm_unreachable_internal("Unknown induction", "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4405); | ||||||||
4406 | case InductionDescriptor::IK_IntInduction: | ||||||||
4407 | case InductionDescriptor::IK_FpInduction: | ||||||||
4408 | llvm_unreachable("Integer/fp induction is handled elsewhere.")::llvm::llvm_unreachable_internal("Integer/fp induction is handled elsewhere." , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4408); | ||||||||
4409 | case InductionDescriptor::IK_PtrInduction: { | ||||||||
4410 | // Handle the pointer induction variable case. | ||||||||
4411 | assert(P->getType()->isPointerTy() && "Unexpected type.")((P->getType()->isPointerTy() && "Unexpected type." ) ? static_cast<void> (0) : __assert_fail ("P->getType()->isPointerTy() && \"Unexpected type.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4411, __PRETTY_FUNCTION__)); | ||||||||
4412 | |||||||||
4413 | if (Cost->isScalarAfterVectorization(P, VF)) { | ||||||||
4414 | // This is the normalized GEP that starts counting at zero. | ||||||||
4415 | Value *PtrInd = | ||||||||
4416 | Builder.CreateSExtOrTrunc(Induction, II.getStep()->getType()); | ||||||||
4417 | // Determine the number of scalars we need to generate for each unroll | ||||||||
4418 | // iteration. If the instruction is uniform, we only need to generate the | ||||||||
4419 | // first lane. Otherwise, we generate all VF values. | ||||||||
4420 | unsigned Lanes = | ||||||||
4421 | Cost->isUniformAfterVectorization(P, VF) ? 1 : VF.getKnownMinValue(); | ||||||||
4422 | for (unsigned Part = 0; Part < UF; ++Part) { | ||||||||
4423 | for (unsigned Lane = 0; Lane < Lanes; ++Lane) { | ||||||||
4424 | Constant *Idx = ConstantInt::get(PtrInd->getType(), | ||||||||
4425 | Lane + Part * VF.getKnownMinValue()); | ||||||||
4426 | Value *GlobalIdx = Builder.CreateAdd(PtrInd, Idx); | ||||||||
4427 | Value *SclrGep = | ||||||||
4428 | emitTransformedIndex(Builder, GlobalIdx, PSE.getSE(), DL, II); | ||||||||
4429 | SclrGep->setName("next.gep"); | ||||||||
4430 | VectorLoopValueMap.setScalarValue(P, {Part, Lane}, SclrGep); | ||||||||
4431 | } | ||||||||
4432 | } | ||||||||
4433 | return; | ||||||||
4434 | } | ||||||||
4435 | assert(isa<SCEVConstant>(II.getStep()) &&((isa<SCEVConstant>(II.getStep()) && "Induction step not a SCEV constant!" ) ? static_cast<void> (0) : __assert_fail ("isa<SCEVConstant>(II.getStep()) && \"Induction step not a SCEV constant!\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4436, __PRETTY_FUNCTION__)) | ||||||||
4436 | "Induction step not a SCEV constant!")((isa<SCEVConstant>(II.getStep()) && "Induction step not a SCEV constant!" ) ? static_cast<void> (0) : __assert_fail ("isa<SCEVConstant>(II.getStep()) && \"Induction step not a SCEV constant!\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4436, __PRETTY_FUNCTION__)); | ||||||||
4437 | Type *PhiType = II.getStep()->getType(); | ||||||||
4438 | |||||||||
4439 | // Build a pointer phi | ||||||||
4440 | Value *ScalarStartValue = II.getStartValue(); | ||||||||
4441 | Type *ScStValueType = ScalarStartValue->getType(); | ||||||||
4442 | PHINode *NewPointerPhi = | ||||||||
4443 | PHINode::Create(ScStValueType, 2, "pointer.phi", Induction); | ||||||||
4444 | NewPointerPhi->addIncoming(ScalarStartValue, LoopVectorPreHeader); | ||||||||
4445 | |||||||||
4446 | // A pointer induction, performed by using a gep | ||||||||
4447 | BasicBlock *LoopLatch = LI->getLoopFor(LoopVectorBody)->getLoopLatch(); | ||||||||
4448 | Instruction *InductionLoc = LoopLatch->getTerminator(); | ||||||||
4449 | const SCEV *ScalarStep = II.getStep(); | ||||||||
4450 | SCEVExpander Exp(*PSE.getSE(), DL, "induction"); | ||||||||
4451 | Value *ScalarStepValue = | ||||||||
4452 | Exp.expandCodeFor(ScalarStep, PhiType, InductionLoc); | ||||||||
4453 | Value *InductionGEP = GetElementPtrInst::Create( | ||||||||
4454 | ScStValueType->getPointerElementType(), NewPointerPhi, | ||||||||
4455 | Builder.CreateMul( | ||||||||
4456 | ScalarStepValue, | ||||||||
4457 | ConstantInt::get(PhiType, VF.getKnownMinValue() * UF)), | ||||||||
4458 | "ptr.ind", InductionLoc); | ||||||||
4459 | NewPointerPhi->addIncoming(InductionGEP, LoopLatch); | ||||||||
4460 | |||||||||
4461 | // Create UF many actual address geps that use the pointer | ||||||||
4462 | // phi as base and a vectorized version of the step value | ||||||||
4463 | // (<step*0, ..., step*N>) as offset. | ||||||||
4464 | for (unsigned Part = 0; Part < UF; ++Part) { | ||||||||
4465 | SmallVector<Constant *, 8> Indices; | ||||||||
4466 | // Create a vector of consecutive numbers from zero to VF. | ||||||||
4467 | for (unsigned i = 0; i < VF.getKnownMinValue(); ++i) | ||||||||
4468 | Indices.push_back( | ||||||||
4469 | ConstantInt::get(PhiType, i + Part * VF.getKnownMinValue())); | ||||||||
4470 | Constant *StartOffset = ConstantVector::get(Indices); | ||||||||
4471 | |||||||||
4472 | Value *GEP = Builder.CreateGEP( | ||||||||
4473 | ScStValueType->getPointerElementType(), NewPointerPhi, | ||||||||
4474 | Builder.CreateMul( | ||||||||
4475 | StartOffset, | ||||||||
4476 | Builder.CreateVectorSplat(VF.getKnownMinValue(), ScalarStepValue), | ||||||||
4477 | "vector.gep")); | ||||||||
4478 | VectorLoopValueMap.setVectorValue(P, Part, GEP); | ||||||||
4479 | } | ||||||||
4480 | } | ||||||||
4481 | } | ||||||||
4482 | } | ||||||||
4483 | |||||||||
4484 | /// A helper function for checking whether an integer division-related | ||||||||
4485 | /// instruction may divide by zero (in which case it must be predicated if | ||||||||
4486 | /// executed conditionally in the scalar code). | ||||||||
4487 | /// TODO: It may be worthwhile to generalize and check isKnownNonZero(). | ||||||||
4488 | /// Non-zero divisors that are non compile-time constants will not be | ||||||||
4489 | /// converted into multiplication, so we will still end up scalarizing | ||||||||
4490 | /// the division, but can do so w/o predication. | ||||||||
4491 | static bool mayDivideByZero(Instruction &I) { | ||||||||
4492 | assert((I.getOpcode() == Instruction::UDiv ||(((I.getOpcode() == Instruction::UDiv || I.getOpcode() == Instruction ::SDiv || I.getOpcode() == Instruction::URem || I.getOpcode() == Instruction::SRem) && "Unexpected instruction") ? static_cast<void> (0) : __assert_fail ("(I.getOpcode() == Instruction::UDiv || I.getOpcode() == Instruction::SDiv || I.getOpcode() == Instruction::URem || I.getOpcode() == Instruction::SRem) && \"Unexpected instruction\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4496, __PRETTY_FUNCTION__)) | ||||||||
4493 | I.getOpcode() == Instruction::SDiv ||(((I.getOpcode() == Instruction::UDiv || I.getOpcode() == Instruction ::SDiv || I.getOpcode() == Instruction::URem || I.getOpcode() == Instruction::SRem) && "Unexpected instruction") ? static_cast<void> (0) : __assert_fail ("(I.getOpcode() == Instruction::UDiv || I.getOpcode() == Instruction::SDiv || I.getOpcode() == Instruction::URem || I.getOpcode() == Instruction::SRem) && \"Unexpected instruction\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4496, __PRETTY_FUNCTION__)) | ||||||||
4494 | I.getOpcode() == Instruction::URem ||(((I.getOpcode() == Instruction::UDiv || I.getOpcode() == Instruction ::SDiv || I.getOpcode() == Instruction::URem || I.getOpcode() == Instruction::SRem) && "Unexpected instruction") ? static_cast<void> (0) : __assert_fail ("(I.getOpcode() == Instruction::UDiv || I.getOpcode() == Instruction::SDiv || I.getOpcode() == Instruction::URem || I.getOpcode() == Instruction::SRem) && \"Unexpected instruction\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4496, __PRETTY_FUNCTION__)) | ||||||||
4495 | I.getOpcode() == Instruction::SRem) &&(((I.getOpcode() == Instruction::UDiv || I.getOpcode() == Instruction ::SDiv || I.getOpcode() == Instruction::URem || I.getOpcode() == Instruction::SRem) && "Unexpected instruction") ? static_cast<void> (0) : __assert_fail ("(I.getOpcode() == Instruction::UDiv || I.getOpcode() == Instruction::SDiv || I.getOpcode() == Instruction::URem || I.getOpcode() == Instruction::SRem) && \"Unexpected instruction\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4496, __PRETTY_FUNCTION__)) | ||||||||
4496 | "Unexpected instruction")(((I.getOpcode() == Instruction::UDiv || I.getOpcode() == Instruction ::SDiv || I.getOpcode() == Instruction::URem || I.getOpcode() == Instruction::SRem) && "Unexpected instruction") ? static_cast<void> (0) : __assert_fail ("(I.getOpcode() == Instruction::UDiv || I.getOpcode() == Instruction::SDiv || I.getOpcode() == Instruction::URem || I.getOpcode() == Instruction::SRem) && \"Unexpected instruction\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4496, __PRETTY_FUNCTION__)); | ||||||||
4497 | Value *Divisor = I.getOperand(1); | ||||||||
4498 | auto *CInt = dyn_cast<ConstantInt>(Divisor); | ||||||||
4499 | return !CInt || CInt->isZero(); | ||||||||
4500 | } | ||||||||
4501 | |||||||||
4502 | void InnerLoopVectorizer::widenInstruction(Instruction &I, VPUser &User, | ||||||||
4503 | VPTransformState &State) { | ||||||||
4504 | assert(!VF.isScalable() && "scalable vectors not yet supported.")((!VF.isScalable() && "scalable vectors not yet supported." ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"scalable vectors not yet supported.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4504, __PRETTY_FUNCTION__)); | ||||||||
4505 | switch (I.getOpcode()) { | ||||||||
4506 | case Instruction::Call: | ||||||||
4507 | case Instruction::Br: | ||||||||
4508 | case Instruction::PHI: | ||||||||
4509 | case Instruction::GetElementPtr: | ||||||||
4510 | case Instruction::Select: | ||||||||
4511 | llvm_unreachable("This instruction is handled by a different recipe.")::llvm::llvm_unreachable_internal("This instruction is handled by a different recipe." , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4511); | ||||||||
4512 | case Instruction::UDiv: | ||||||||
4513 | case Instruction::SDiv: | ||||||||
4514 | case Instruction::SRem: | ||||||||
4515 | case Instruction::URem: | ||||||||
4516 | case Instruction::Add: | ||||||||
4517 | case Instruction::FAdd: | ||||||||
4518 | case Instruction::Sub: | ||||||||
4519 | case Instruction::FSub: | ||||||||
4520 | case Instruction::FNeg: | ||||||||
4521 | case Instruction::Mul: | ||||||||
4522 | case Instruction::FMul: | ||||||||
4523 | case Instruction::FDiv: | ||||||||
4524 | case Instruction::FRem: | ||||||||
4525 | case Instruction::Shl: | ||||||||
4526 | case Instruction::LShr: | ||||||||
4527 | case Instruction::AShr: | ||||||||
4528 | case Instruction::And: | ||||||||
4529 | case Instruction::Or: | ||||||||
4530 | case Instruction::Xor: { | ||||||||
4531 | // Just widen unops and binops. | ||||||||
4532 | setDebugLocFromInst(Builder, &I); | ||||||||
4533 | |||||||||
4534 | for (unsigned Part = 0; Part < UF; ++Part) { | ||||||||
4535 | SmallVector<Value *, 2> Ops; | ||||||||
4536 | for (VPValue *VPOp : User.operands()) | ||||||||
4537 | Ops.push_back(State.get(VPOp, Part)); | ||||||||
4538 | |||||||||
4539 | Value *V = Builder.CreateNAryOp(I.getOpcode(), Ops); | ||||||||
4540 | |||||||||
4541 | if (auto *VecOp = dyn_cast<Instruction>(V)) | ||||||||
4542 | VecOp->copyIRFlags(&I); | ||||||||
4543 | |||||||||
4544 | // Use this vector value for all users of the original instruction. | ||||||||
4545 | VectorLoopValueMap.setVectorValue(&I, Part, V); | ||||||||
4546 | addMetadata(V, &I); | ||||||||
4547 | } | ||||||||
4548 | |||||||||
4549 | break; | ||||||||
4550 | } | ||||||||
4551 | case Instruction::ICmp: | ||||||||
4552 | case Instruction::FCmp: { | ||||||||
4553 | // Widen compares. Generate vector compares. | ||||||||
4554 | bool FCmp = (I.getOpcode() == Instruction::FCmp); | ||||||||
4555 | auto *Cmp = cast<CmpInst>(&I); | ||||||||
4556 | setDebugLocFromInst(Builder, Cmp); | ||||||||
4557 | for (unsigned Part = 0; Part < UF; ++Part) { | ||||||||
4558 | Value *A = State.get(User.getOperand(0), Part); | ||||||||
4559 | Value *B = State.get(User.getOperand(1), Part); | ||||||||
4560 | Value *C = nullptr; | ||||||||
4561 | if (FCmp) { | ||||||||
4562 | // Propagate fast math flags. | ||||||||
4563 | IRBuilder<>::FastMathFlagGuard FMFG(Builder); | ||||||||
4564 | Builder.setFastMathFlags(Cmp->getFastMathFlags()); | ||||||||
4565 | C = Builder.CreateFCmp(Cmp->getPredicate(), A, B); | ||||||||
4566 | } else { | ||||||||
4567 | C = Builder.CreateICmp(Cmp->getPredicate(), A, B); | ||||||||
4568 | } | ||||||||
4569 | VectorLoopValueMap.setVectorValue(&I, Part, C); | ||||||||
4570 | addMetadata(C, &I); | ||||||||
4571 | } | ||||||||
4572 | |||||||||
4573 | break; | ||||||||
4574 | } | ||||||||
4575 | |||||||||
4576 | case Instruction::ZExt: | ||||||||
4577 | case Instruction::SExt: | ||||||||
4578 | case Instruction::FPToUI: | ||||||||
4579 | case Instruction::FPToSI: | ||||||||
4580 | case Instruction::FPExt: | ||||||||
4581 | case Instruction::PtrToInt: | ||||||||
4582 | case Instruction::IntToPtr: | ||||||||
4583 | case Instruction::SIToFP: | ||||||||
4584 | case Instruction::UIToFP: | ||||||||
4585 | case Instruction::Trunc: | ||||||||
4586 | case Instruction::FPTrunc: | ||||||||
4587 | case Instruction::BitCast: { | ||||||||
4588 | auto *CI = cast<CastInst>(&I); | ||||||||
4589 | setDebugLocFromInst(Builder, CI); | ||||||||
4590 | |||||||||
4591 | /// Vectorize casts. | ||||||||
4592 | assert(!VF.isScalable() && "VF is assumed to be non scalable.")((!VF.isScalable() && "VF is assumed to be non scalable." ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"VF is assumed to be non scalable.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4592, __PRETTY_FUNCTION__)); | ||||||||
4593 | Type *DestTy = | ||||||||
4594 | (VF.isScalar()) ? CI->getType() : VectorType::get(CI->getType(), VF); | ||||||||
4595 | |||||||||
4596 | for (unsigned Part = 0; Part < UF; ++Part) { | ||||||||
4597 | Value *A = State.get(User.getOperand(0), Part); | ||||||||
4598 | Value *Cast = Builder.CreateCast(CI->getOpcode(), A, DestTy); | ||||||||
4599 | VectorLoopValueMap.setVectorValue(&I, Part, Cast); | ||||||||
4600 | addMetadata(Cast, &I); | ||||||||
4601 | } | ||||||||
4602 | break; | ||||||||
4603 | } | ||||||||
4604 | default: | ||||||||
4605 | // This instruction is not vectorized by simple widening. | ||||||||
4606 | LLVM_DEBUG(dbgs() << "LV: Found an unhandled instruction: " << I)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found an unhandled instruction: " << I; } } while (false); | ||||||||
4607 | llvm_unreachable("Unhandled instruction!")::llvm::llvm_unreachable_internal("Unhandled instruction!", "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4607); | ||||||||
4608 | } // end of switch. | ||||||||
4609 | } | ||||||||
4610 | |||||||||
4611 | void InnerLoopVectorizer::widenCallInstruction(CallInst &I, VPUser &ArgOperands, | ||||||||
4612 | VPTransformState &State) { | ||||||||
4613 | assert(!isa<DbgInfoIntrinsic>(I) &&((!isa<DbgInfoIntrinsic>(I) && "DbgInfoIntrinsic should have been dropped during VPlan construction" ) ? static_cast<void> (0) : __assert_fail ("!isa<DbgInfoIntrinsic>(I) && \"DbgInfoIntrinsic should have been dropped during VPlan construction\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4614, __PRETTY_FUNCTION__)) | ||||||||
4614 | "DbgInfoIntrinsic should have been dropped during VPlan construction")((!isa<DbgInfoIntrinsic>(I) && "DbgInfoIntrinsic should have been dropped during VPlan construction" ) ? static_cast<void> (0) : __assert_fail ("!isa<DbgInfoIntrinsic>(I) && \"DbgInfoIntrinsic should have been dropped during VPlan construction\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4614, __PRETTY_FUNCTION__)); | ||||||||
4615 | setDebugLocFromInst(Builder, &I); | ||||||||
4616 | |||||||||
4617 | Module *M = I.getParent()->getParent()->getParent(); | ||||||||
4618 | auto *CI = cast<CallInst>(&I); | ||||||||
4619 | |||||||||
4620 | SmallVector<Type *, 4> Tys; | ||||||||
4621 | for (Value *ArgOperand : CI->arg_operands()) | ||||||||
4622 | Tys.push_back(ToVectorTy(ArgOperand->getType(), VF.getKnownMinValue())); | ||||||||
4623 | |||||||||
4624 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); | ||||||||
4625 | |||||||||
4626 | // The flag shows whether we use Intrinsic or a usual Call for vectorized | ||||||||
4627 | // version of the instruction. | ||||||||
4628 | // Is it beneficial to perform intrinsic call compared to lib call? | ||||||||
4629 | bool NeedToScalarize = false; | ||||||||
4630 | unsigned CallCost = Cost->getVectorCallCost(CI, VF, NeedToScalarize); | ||||||||
4631 | bool UseVectorIntrinsic = | ||||||||
4632 | ID && Cost->getVectorIntrinsicCost(CI, VF) <= CallCost; | ||||||||
4633 | assert((UseVectorIntrinsic || !NeedToScalarize) &&(((UseVectorIntrinsic || !NeedToScalarize) && "Instruction should be scalarized elsewhere." ) ? static_cast<void> (0) : __assert_fail ("(UseVectorIntrinsic || !NeedToScalarize) && \"Instruction should be scalarized elsewhere.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4634, __PRETTY_FUNCTION__)) | ||||||||
4634 | "Instruction should be scalarized elsewhere.")(((UseVectorIntrinsic || !NeedToScalarize) && "Instruction should be scalarized elsewhere." ) ? static_cast<void> (0) : __assert_fail ("(UseVectorIntrinsic || !NeedToScalarize) && \"Instruction should be scalarized elsewhere.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4634, __PRETTY_FUNCTION__)); | ||||||||
4635 | |||||||||
4636 | for (unsigned Part = 0; Part < UF; ++Part) { | ||||||||
4637 | SmallVector<Value *, 4> Args; | ||||||||
4638 | for (auto &I : enumerate(ArgOperands.operands())) { | ||||||||
4639 | // Some intrinsics have a scalar argument - don't replace it with a | ||||||||
4640 | // vector. | ||||||||
4641 | Value *Arg; | ||||||||
4642 | if (!UseVectorIntrinsic || !hasVectorInstrinsicScalarOpd(ID, I.index())) | ||||||||
4643 | Arg = State.get(I.value(), Part); | ||||||||
4644 | else | ||||||||
4645 | Arg = State.get(I.value(), {0, 0}); | ||||||||
4646 | Args.push_back(Arg); | ||||||||
4647 | } | ||||||||
4648 | |||||||||
4649 | Function *VectorF; | ||||||||
4650 | if (UseVectorIntrinsic) { | ||||||||
4651 | // Use vector version of the intrinsic. | ||||||||
4652 | Type *TysForDecl[] = {CI->getType()}; | ||||||||
4653 | if (VF.isVector()) { | ||||||||
4654 | assert(!VF.isScalable() && "VF is assumed to be non scalable.")((!VF.isScalable() && "VF is assumed to be non scalable." ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"VF is assumed to be non scalable.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4654, __PRETTY_FUNCTION__)); | ||||||||
4655 | TysForDecl[0] = VectorType::get(CI->getType()->getScalarType(), VF); | ||||||||
4656 | } | ||||||||
4657 | VectorF = Intrinsic::getDeclaration(M, ID, TysForDecl); | ||||||||
4658 | assert(VectorF && "Can't retrieve vector intrinsic.")((VectorF && "Can't retrieve vector intrinsic.") ? static_cast <void> (0) : __assert_fail ("VectorF && \"Can't retrieve vector intrinsic.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4658, __PRETTY_FUNCTION__)); | ||||||||
4659 | } else { | ||||||||
4660 | // Use vector version of the function call. | ||||||||
4661 | const VFShape Shape = VFShape::get(*CI, VF, false /*HasGlobalPred*/); | ||||||||
4662 | #ifndef NDEBUG | ||||||||
4663 | assert(VFDatabase(*CI).getVectorizedFunction(Shape) != nullptr &&((VFDatabase(*CI).getVectorizedFunction(Shape) != nullptr && "Can't create vector function.") ? static_cast<void> ( 0) : __assert_fail ("VFDatabase(*CI).getVectorizedFunction(Shape) != nullptr && \"Can't create vector function.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4664, __PRETTY_FUNCTION__)) | ||||||||
4664 | "Can't create vector function.")((VFDatabase(*CI).getVectorizedFunction(Shape) != nullptr && "Can't create vector function.") ? static_cast<void> ( 0) : __assert_fail ("VFDatabase(*CI).getVectorizedFunction(Shape) != nullptr && \"Can't create vector function.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4664, __PRETTY_FUNCTION__)); | ||||||||
4665 | #endif | ||||||||
4666 | VectorF = VFDatabase(*CI).getVectorizedFunction(Shape); | ||||||||
4667 | } | ||||||||
4668 | SmallVector<OperandBundleDef, 1> OpBundles; | ||||||||
4669 | CI->getOperandBundlesAsDefs(OpBundles); | ||||||||
4670 | CallInst *V = Builder.CreateCall(VectorF, Args, OpBundles); | ||||||||
4671 | |||||||||
4672 | if (isa<FPMathOperator>(V)) | ||||||||
4673 | V->copyFastMathFlags(CI); | ||||||||
4674 | |||||||||
4675 | VectorLoopValueMap.setVectorValue(&I, Part, V); | ||||||||
4676 | addMetadata(V, &I); | ||||||||
4677 | } | ||||||||
4678 | } | ||||||||
4679 | |||||||||
4680 | void InnerLoopVectorizer::widenSelectInstruction(SelectInst &I, | ||||||||
4681 | VPUser &Operands, | ||||||||
4682 | bool InvariantCond, | ||||||||
4683 | VPTransformState &State) { | ||||||||
4684 | setDebugLocFromInst(Builder, &I); | ||||||||
4685 | |||||||||
4686 | // The condition can be loop invariant but still defined inside the | ||||||||
4687 | // loop. This means that we can't just use the original 'cond' value. | ||||||||
4688 | // We have to take the 'vectorized' value and pick the first lane. | ||||||||
4689 | // Instcombine will make this a no-op. | ||||||||
4690 | auto *InvarCond = | ||||||||
4691 | InvariantCond ? State.get(Operands.getOperand(0), {0, 0}) : nullptr; | ||||||||
4692 | |||||||||
4693 | for (unsigned Part = 0; Part < UF; ++Part) { | ||||||||
4694 | Value *Cond = | ||||||||
4695 | InvarCond ? InvarCond : State.get(Operands.getOperand(0), Part); | ||||||||
4696 | Value *Op0 = State.get(Operands.getOperand(1), Part); | ||||||||
4697 | Value *Op1 = State.get(Operands.getOperand(2), Part); | ||||||||
4698 | Value *Sel = Builder.CreateSelect(Cond, Op0, Op1); | ||||||||
4699 | VectorLoopValueMap.setVectorValue(&I, Part, Sel); | ||||||||
4700 | addMetadata(Sel, &I); | ||||||||
4701 | } | ||||||||
4702 | } | ||||||||
4703 | |||||||||
4704 | void LoopVectorizationCostModel::collectLoopScalars(ElementCount VF) { | ||||||||
4705 | // We should not collect Scalars more than once per VF. Right now, this | ||||||||
4706 | // function is called from collectUniformsAndScalars(), which already does | ||||||||
4707 | // this check. Collecting Scalars for VF=1 does not make any sense. | ||||||||
4708 | assert(VF.isVector() && Scalars.find(VF) == Scalars.end() &&((VF.isVector() && Scalars.find(VF) == Scalars.end() && "This function should not be visited twice for the same VF") ? static_cast<void> (0) : __assert_fail ("VF.isVector() && Scalars.find(VF) == Scalars.end() && \"This function should not be visited twice for the same VF\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4709, __PRETTY_FUNCTION__)) | ||||||||
4709 | "This function should not be visited twice for the same VF")((VF.isVector() && Scalars.find(VF) == Scalars.end() && "This function should not be visited twice for the same VF") ? static_cast<void> (0) : __assert_fail ("VF.isVector() && Scalars.find(VF) == Scalars.end() && \"This function should not be visited twice for the same VF\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4709, __PRETTY_FUNCTION__)); | ||||||||
4710 | |||||||||
4711 | SmallSetVector<Instruction *, 8> Worklist; | ||||||||
4712 | |||||||||
4713 | // These sets are used to seed the analysis with pointers used by memory | ||||||||
4714 | // accesses that will remain scalar. | ||||||||
4715 | SmallSetVector<Instruction *, 8> ScalarPtrs; | ||||||||
4716 | SmallPtrSet<Instruction *, 8> PossibleNonScalarPtrs; | ||||||||
4717 | auto *Latch = TheLoop->getLoopLatch(); | ||||||||
4718 | |||||||||
4719 | // A helper that returns true if the use of Ptr by MemAccess will be scalar. | ||||||||
4720 | // The pointer operands of loads and stores will be scalar as long as the | ||||||||
4721 | // memory access is not a gather or scatter operation. The value operand of a | ||||||||
4722 | // store will remain scalar if the store is scalarized. | ||||||||
4723 | auto isScalarUse = [&](Instruction *MemAccess, Value *Ptr) { | ||||||||
4724 | InstWidening WideningDecision = getWideningDecision(MemAccess, VF); | ||||||||
4725 | assert(WideningDecision != CM_Unknown &&((WideningDecision != CM_Unknown && "Widening decision should be ready at this moment" ) ? static_cast<void> (0) : __assert_fail ("WideningDecision != CM_Unknown && \"Widening decision should be ready at this moment\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4726, __PRETTY_FUNCTION__)) | ||||||||
4726 | "Widening decision should be ready at this moment")((WideningDecision != CM_Unknown && "Widening decision should be ready at this moment" ) ? static_cast<void> (0) : __assert_fail ("WideningDecision != CM_Unknown && \"Widening decision should be ready at this moment\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4726, __PRETTY_FUNCTION__)); | ||||||||
4727 | if (auto *Store = dyn_cast<StoreInst>(MemAccess)) | ||||||||
4728 | if (Ptr == Store->getValueOperand()) | ||||||||
4729 | return WideningDecision == CM_Scalarize; | ||||||||
4730 | assert(Ptr == getLoadStorePointerOperand(MemAccess) &&((Ptr == getLoadStorePointerOperand(MemAccess) && "Ptr is neither a value or pointer operand" ) ? static_cast<void> (0) : __assert_fail ("Ptr == getLoadStorePointerOperand(MemAccess) && \"Ptr is neither a value or pointer operand\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4731, __PRETTY_FUNCTION__)) | ||||||||
4731 | "Ptr is neither a value or pointer operand")((Ptr == getLoadStorePointerOperand(MemAccess) && "Ptr is neither a value or pointer operand" ) ? static_cast<void> (0) : __assert_fail ("Ptr == getLoadStorePointerOperand(MemAccess) && \"Ptr is neither a value or pointer operand\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4731, __PRETTY_FUNCTION__)); | ||||||||
4732 | return WideningDecision != CM_GatherScatter; | ||||||||
4733 | }; | ||||||||
4734 | |||||||||
4735 | // A helper that returns true if the given value is a bitcast or | ||||||||
4736 | // getelementptr instruction contained in the loop. | ||||||||
4737 | auto isLoopVaryingBitCastOrGEP = [&](Value *V) { | ||||||||
4738 | return ((isa<BitCastInst>(V) && V->getType()->isPointerTy()) || | ||||||||
4739 | isa<GetElementPtrInst>(V)) && | ||||||||
4740 | !TheLoop->isLoopInvariant(V); | ||||||||
4741 | }; | ||||||||
4742 | |||||||||
4743 | auto isScalarPtrInduction = [&](Instruction *MemAccess, Value *Ptr) { | ||||||||
4744 | if (!isa<PHINode>(Ptr) || | ||||||||
4745 | !Legal->getInductionVars().count(cast<PHINode>(Ptr))) | ||||||||
4746 | return false; | ||||||||
4747 | auto &Induction = Legal->getInductionVars()[cast<PHINode>(Ptr)]; | ||||||||
4748 | if (Induction.getKind() != InductionDescriptor::IK_PtrInduction) | ||||||||
4749 | return false; | ||||||||
4750 | return isScalarUse(MemAccess, Ptr); | ||||||||
4751 | }; | ||||||||
4752 | |||||||||
4753 | // A helper that evaluates a memory access's use of a pointer. If the | ||||||||
4754 | // pointer is actually the pointer induction of a loop, it is being | ||||||||
4755 | // inserted into Worklist. If the use will be a scalar use, and the | ||||||||
4756 | // pointer is only used by memory accesses, we place the pointer in | ||||||||
4757 | // ScalarPtrs. Otherwise, the pointer is placed in PossibleNonScalarPtrs. | ||||||||
4758 | auto evaluatePtrUse = [&](Instruction *MemAccess, Value *Ptr) { | ||||||||
4759 | if (isScalarPtrInduction(MemAccess, Ptr)) { | ||||||||
4760 | Worklist.insert(cast<Instruction>(Ptr)); | ||||||||
4761 | Instruction *Update = cast<Instruction>( | ||||||||
4762 | cast<PHINode>(Ptr)->getIncomingValueForBlock(Latch)); | ||||||||
4763 | Worklist.insert(Update); | ||||||||
4764 | LLVM_DEBUG(dbgs() << "LV: Found new scalar instruction: " << *Ptrdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found new scalar instruction: " << *Ptr << "\n"; } } while (false) | ||||||||
4765 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found new scalar instruction: " << *Ptr << "\n"; } } while (false); | ||||||||
4766 | LLVM_DEBUG(dbgs() << "LV: Found new scalar instruction: " << *Updatedo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found new scalar instruction: " << *Update << "\n"; } } while (false) | ||||||||
4767 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found new scalar instruction: " << *Update << "\n"; } } while (false); | ||||||||
4768 | return; | ||||||||
4769 | } | ||||||||
4770 | // We only care about bitcast and getelementptr instructions contained in | ||||||||
4771 | // the loop. | ||||||||
4772 | if (!isLoopVaryingBitCastOrGEP(Ptr)) | ||||||||
4773 | return; | ||||||||
4774 | |||||||||
4775 | // If the pointer has already been identified as scalar (e.g., if it was | ||||||||
4776 | // also identified as uniform), there's nothing to do. | ||||||||
4777 | auto *I = cast<Instruction>(Ptr); | ||||||||
4778 | if (Worklist.count(I)) | ||||||||
4779 | return; | ||||||||
4780 | |||||||||
4781 | // If the use of the pointer will be a scalar use, and all users of the | ||||||||
4782 | // pointer are memory accesses, place the pointer in ScalarPtrs. Otherwise, | ||||||||
4783 | // place the pointer in PossibleNonScalarPtrs. | ||||||||
4784 | if (isScalarUse(MemAccess, Ptr) && llvm::all_of(I->users(), [&](User *U) { | ||||||||
4785 | return isa<LoadInst>(U) || isa<StoreInst>(U); | ||||||||
4786 | })) | ||||||||
4787 | ScalarPtrs.insert(I); | ||||||||
4788 | else | ||||||||
4789 | PossibleNonScalarPtrs.insert(I); | ||||||||
4790 | }; | ||||||||
4791 | |||||||||
4792 | // We seed the scalars analysis with three classes of instructions: (1) | ||||||||
4793 | // instructions marked uniform-after-vectorization and (2) bitcast, | ||||||||
4794 | // getelementptr and (pointer) phi instructions used by memory accesses | ||||||||
4795 | // requiring a scalar use. | ||||||||
4796 | // | ||||||||
4797 | // (1) Add to the worklist all instructions that have been identified as | ||||||||
4798 | // uniform-after-vectorization. | ||||||||
4799 | Worklist.insert(Uniforms[VF].begin(), Uniforms[VF].end()); | ||||||||
4800 | |||||||||
4801 | // (2) Add to the worklist all bitcast and getelementptr instructions used by | ||||||||
4802 | // memory accesses requiring a scalar use. The pointer operands of loads and | ||||||||
4803 | // stores will be scalar as long as the memory accesses is not a gather or | ||||||||
4804 | // scatter operation. The value operand of a store will remain scalar if the | ||||||||
4805 | // store is scalarized. | ||||||||
4806 | for (auto *BB : TheLoop->blocks()) | ||||||||
4807 | for (auto &I : *BB) { | ||||||||
4808 | if (auto *Load = dyn_cast<LoadInst>(&I)) { | ||||||||
4809 | evaluatePtrUse(Load, Load->getPointerOperand()); | ||||||||
4810 | } else if (auto *Store = dyn_cast<StoreInst>(&I)) { | ||||||||
4811 | evaluatePtrUse(Store, Store->getPointerOperand()); | ||||||||
4812 | evaluatePtrUse(Store, Store->getValueOperand()); | ||||||||
4813 | } | ||||||||
4814 | } | ||||||||
4815 | for (auto *I : ScalarPtrs) | ||||||||
4816 | if (!PossibleNonScalarPtrs.count(I)) { | ||||||||
4817 | 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); | ||||||||
4818 | Worklist.insert(I); | ||||||||
4819 | } | ||||||||
4820 | |||||||||
4821 | // Insert the forced scalars. | ||||||||
4822 | // FIXME: Currently widenPHIInstruction() often creates a dead vector | ||||||||
4823 | // induction variable when the PHI user is scalarized. | ||||||||
4824 | auto ForcedScalar = ForcedScalars.find(VF); | ||||||||
4825 | if (ForcedScalar != ForcedScalars.end()) | ||||||||
4826 | for (auto *I : ForcedScalar->second) | ||||||||
4827 | Worklist.insert(I); | ||||||||
4828 | |||||||||
4829 | // Expand the worklist by looking through any bitcasts and getelementptr | ||||||||
4830 | // instructions we've already identified as scalar. This is similar to the | ||||||||
4831 | // expansion step in collectLoopUniforms(); however, here we're only | ||||||||
4832 | // expanding to include additional bitcasts and getelementptr instructions. | ||||||||
4833 | unsigned Idx = 0; | ||||||||
4834 | while (Idx != Worklist.size()) { | ||||||||
4835 | Instruction *Dst = Worklist[Idx++]; | ||||||||
4836 | if (!isLoopVaryingBitCastOrGEP(Dst->getOperand(0))) | ||||||||
4837 | continue; | ||||||||
4838 | auto *Src = cast<Instruction>(Dst->getOperand(0)); | ||||||||
4839 | if (llvm::all_of(Src->users(), [&](User *U) -> bool { | ||||||||
4840 | auto *J = cast<Instruction>(U); | ||||||||
4841 | return !TheLoop->contains(J) || Worklist.count(J) || | ||||||||
4842 | ((isa<LoadInst>(J) || isa<StoreInst>(J)) && | ||||||||
4843 | isScalarUse(J, Src)); | ||||||||
4844 | })) { | ||||||||
4845 | Worklist.insert(Src); | ||||||||
4846 | 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); | ||||||||
4847 | } | ||||||||
4848 | } | ||||||||
4849 | |||||||||
4850 | // An induction variable will remain scalar if all users of the induction | ||||||||
4851 | // variable and induction variable update remain scalar. | ||||||||
4852 | for (auto &Induction : Legal->getInductionVars()) { | ||||||||
4853 | auto *Ind = Induction.first; | ||||||||
4854 | auto *IndUpdate = cast<Instruction>(Ind->getIncomingValueForBlock(Latch)); | ||||||||
4855 | |||||||||
4856 | // If tail-folding is applied, the primary induction variable will be used | ||||||||
4857 | // to feed a vector compare. | ||||||||
4858 | if (Ind == Legal->getPrimaryInduction() && foldTailByMasking()) | ||||||||
4859 | continue; | ||||||||
4860 | |||||||||
4861 | // Determine if all users of the induction variable are scalar after | ||||||||
4862 | // vectorization. | ||||||||
4863 | auto ScalarInd = llvm::all_of(Ind->users(), [&](User *U) -> bool { | ||||||||
4864 | auto *I = cast<Instruction>(U); | ||||||||
4865 | return I == IndUpdate || !TheLoop->contains(I) || Worklist.count(I); | ||||||||
4866 | }); | ||||||||
4867 | if (!ScalarInd) | ||||||||
4868 | continue; | ||||||||
4869 | |||||||||
4870 | // Determine if all users of the induction variable update instruction are | ||||||||
4871 | // scalar after vectorization. | ||||||||
4872 | auto ScalarIndUpdate = | ||||||||
4873 | llvm::all_of(IndUpdate->users(), [&](User *U) -> bool { | ||||||||
4874 | auto *I = cast<Instruction>(U); | ||||||||
4875 | return I == Ind || !TheLoop->contains(I) || Worklist.count(I); | ||||||||
4876 | }); | ||||||||
4877 | if (!ScalarIndUpdate) | ||||||||
4878 | continue; | ||||||||
4879 | |||||||||
4880 | // The induction variable and its update instruction will remain scalar. | ||||||||
4881 | Worklist.insert(Ind); | ||||||||
4882 | Worklist.insert(IndUpdate); | ||||||||
4883 | 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); | ||||||||
4884 | LLVM_DEBUG(dbgs() << "LV: Found scalar instruction: " << *IndUpdatedo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found scalar instruction: " << *IndUpdate << "\n"; } } while (false) | ||||||||
4885 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found scalar instruction: " << *IndUpdate << "\n"; } } while (false); | ||||||||
4886 | } | ||||||||
4887 | |||||||||
4888 | Scalars[VF].insert(Worklist.begin(), Worklist.end()); | ||||||||
4889 | } | ||||||||
4890 | |||||||||
4891 | bool LoopVectorizationCostModel::isScalarWithPredication(Instruction *I, | ||||||||
4892 | ElementCount VF) { | ||||||||
4893 | assert(!VF.isScalable() && "scalable vectors not yet supported.")((!VF.isScalable() && "scalable vectors not yet supported." ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"scalable vectors not yet supported.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4893, __PRETTY_FUNCTION__)); | ||||||||
4894 | if (!blockNeedsPredication(I->getParent())) | ||||||||
4895 | return false; | ||||||||
4896 | switch(I->getOpcode()) { | ||||||||
4897 | default: | ||||||||
4898 | break; | ||||||||
4899 | case Instruction::Load: | ||||||||
4900 | case Instruction::Store: { | ||||||||
4901 | if (!Legal->isMaskRequired(I)) | ||||||||
4902 | return false; | ||||||||
4903 | auto *Ptr = getLoadStorePointerOperand(I); | ||||||||
4904 | auto *Ty = getMemInstValueType(I); | ||||||||
4905 | // We have already decided how to vectorize this instruction, get that | ||||||||
4906 | // result. | ||||||||
4907 | if (VF.isVector()) { | ||||||||
4908 | InstWidening WideningDecision = getWideningDecision(I, VF); | ||||||||
4909 | assert(WideningDecision != CM_Unknown &&((WideningDecision != CM_Unknown && "Widening decision should be ready at this moment" ) ? static_cast<void> (0) : __assert_fail ("WideningDecision != CM_Unknown && \"Widening decision should be ready at this moment\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4910, __PRETTY_FUNCTION__)) | ||||||||
4910 | "Widening decision should be ready at this moment")((WideningDecision != CM_Unknown && "Widening decision should be ready at this moment" ) ? static_cast<void> (0) : __assert_fail ("WideningDecision != CM_Unknown && \"Widening decision should be ready at this moment\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4910, __PRETTY_FUNCTION__)); | ||||||||
4911 | return WideningDecision == CM_Scalarize; | ||||||||
4912 | } | ||||||||
4913 | const Align Alignment = getLoadStoreAlignment(I); | ||||||||
4914 | return isa<LoadInst>(I) ? !(isLegalMaskedLoad(Ty, Ptr, Alignment) || | ||||||||
4915 | isLegalMaskedGather(Ty, Alignment)) | ||||||||
4916 | : !(isLegalMaskedStore(Ty, Ptr, Alignment) || | ||||||||
4917 | isLegalMaskedScatter(Ty, Alignment)); | ||||||||
4918 | } | ||||||||
4919 | case Instruction::UDiv: | ||||||||
4920 | case Instruction::SDiv: | ||||||||
4921 | case Instruction::SRem: | ||||||||
4922 | case Instruction::URem: | ||||||||
4923 | return mayDivideByZero(*I); | ||||||||
4924 | } | ||||||||
4925 | return false; | ||||||||
4926 | } | ||||||||
4927 | |||||||||
4928 | bool LoopVectorizationCostModel::interleavedAccessCanBeWidened( | ||||||||
4929 | Instruction *I, ElementCount VF) { | ||||||||
4930 | assert(isAccessInterleaved(I) && "Expecting interleaved access.")((isAccessInterleaved(I) && "Expecting interleaved access." ) ? static_cast<void> (0) : __assert_fail ("isAccessInterleaved(I) && \"Expecting interleaved access.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4930, __PRETTY_FUNCTION__)); | ||||||||
4931 | assert(getWideningDecision(I, VF) == CM_Unknown &&((getWideningDecision(I, VF) == CM_Unknown && "Decision should not be set yet." ) ? static_cast<void> (0) : __assert_fail ("getWideningDecision(I, VF) == CM_Unknown && \"Decision should not be set yet.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4932, __PRETTY_FUNCTION__)) | ||||||||
4932 | "Decision should not be set yet.")((getWideningDecision(I, VF) == CM_Unknown && "Decision should not be set yet." ) ? static_cast<void> (0) : __assert_fail ("getWideningDecision(I, VF) == CM_Unknown && \"Decision should not be set yet.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4932, __PRETTY_FUNCTION__)); | ||||||||
4933 | auto *Group = getInterleavedAccessGroup(I); | ||||||||
4934 | assert(Group && "Must have a group.")((Group && "Must have a group.") ? static_cast<void > (0) : __assert_fail ("Group && \"Must have a group.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4934, __PRETTY_FUNCTION__)); | ||||||||
4935 | |||||||||
4936 | // If the instruction's allocated size doesn't equal it's type size, it | ||||||||
4937 | // requires padding and will be scalarized. | ||||||||
4938 | auto &DL = I->getModule()->getDataLayout(); | ||||||||
4939 | auto *ScalarTy = getMemInstValueType(I); | ||||||||
4940 | if (hasIrregularType(ScalarTy, DL, VF)) | ||||||||
4941 | return false; | ||||||||
4942 | |||||||||
4943 | // Check if masking is required. | ||||||||
4944 | // A Group may need masking for one of two reasons: it resides in a block that | ||||||||
4945 | // needs predication, or it was decided to use masking to deal with gaps. | ||||||||
4946 | bool PredicatedAccessRequiresMasking = | ||||||||
4947 | Legal->blockNeedsPredication(I->getParent()) && Legal->isMaskRequired(I); | ||||||||
4948 | bool AccessWithGapsRequiresMasking = | ||||||||
4949 | Group->requiresScalarEpilogue() && !isScalarEpilogueAllowed(); | ||||||||
4950 | if (!PredicatedAccessRequiresMasking && !AccessWithGapsRequiresMasking) | ||||||||
4951 | return true; | ||||||||
4952 | |||||||||
4953 | // If masked interleaving is required, we expect that the user/target had | ||||||||
4954 | // enabled it, because otherwise it either wouldn't have been created or | ||||||||
4955 | // it should have been invalidated by the CostModel. | ||||||||
4956 | assert(useMaskedInterleavedAccesses(TTI) &&((useMaskedInterleavedAccesses(TTI) && "Masked interleave-groups for predicated accesses are not enabled." ) ? static_cast<void> (0) : __assert_fail ("useMaskedInterleavedAccesses(TTI) && \"Masked interleave-groups for predicated accesses are not enabled.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4957, __PRETTY_FUNCTION__)) | ||||||||
4957 | "Masked interleave-groups for predicated accesses are not enabled.")((useMaskedInterleavedAccesses(TTI) && "Masked interleave-groups for predicated accesses are not enabled." ) ? static_cast<void> (0) : __assert_fail ("useMaskedInterleavedAccesses(TTI) && \"Masked interleave-groups for predicated accesses are not enabled.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4957, __PRETTY_FUNCTION__)); | ||||||||
4958 | |||||||||
4959 | auto *Ty = getMemInstValueType(I); | ||||||||
4960 | const Align Alignment = getLoadStoreAlignment(I); | ||||||||
4961 | return isa<LoadInst>(I) ? TTI.isLegalMaskedLoad(Ty, Alignment) | ||||||||
4962 | : TTI.isLegalMaskedStore(Ty, Alignment); | ||||||||
4963 | } | ||||||||
4964 | |||||||||
4965 | bool LoopVectorizationCostModel::memoryInstructionCanBeWidened( | ||||||||
4966 | Instruction *I, ElementCount VF) { | ||||||||
4967 | // Get and ensure we have a valid memory instruction. | ||||||||
4968 | LoadInst *LI = dyn_cast<LoadInst>(I); | ||||||||
4969 | StoreInst *SI = dyn_cast<StoreInst>(I); | ||||||||
4970 | assert((LI || SI) && "Invalid memory instruction")(((LI || SI) && "Invalid memory instruction") ? static_cast <void> (0) : __assert_fail ("(LI || SI) && \"Invalid memory instruction\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4970, __PRETTY_FUNCTION__)); | ||||||||
4971 | |||||||||
4972 | auto *Ptr = getLoadStorePointerOperand(I); | ||||||||
4973 | |||||||||
4974 | // In order to be widened, the pointer should be consecutive, first of all. | ||||||||
4975 | if (!Legal->isConsecutivePtr(Ptr)) | ||||||||
4976 | return false; | ||||||||
4977 | |||||||||
4978 | // If the instruction is a store located in a predicated block, it will be | ||||||||
4979 | // scalarized. | ||||||||
4980 | if (isScalarWithPredication(I)) | ||||||||
4981 | return false; | ||||||||
4982 | |||||||||
4983 | // If the instruction's allocated size doesn't equal it's type size, it | ||||||||
4984 | // requires padding and will be scalarized. | ||||||||
4985 | auto &DL = I->getModule()->getDataLayout(); | ||||||||
4986 | auto *ScalarTy = LI ? LI->getType() : SI->getValueOperand()->getType(); | ||||||||
4987 | if (hasIrregularType(ScalarTy, DL, VF)) | ||||||||
4988 | return false; | ||||||||
4989 | |||||||||
4990 | return true; | ||||||||
4991 | } | ||||||||
4992 | |||||||||
4993 | void LoopVectorizationCostModel::collectLoopUniforms(ElementCount VF) { | ||||||||
4994 | // We should not collect Uniforms more than once per VF. Right now, | ||||||||
4995 | // this function is called from collectUniformsAndScalars(), which | ||||||||
4996 | // already does this check. Collecting Uniforms for VF=1 does not make any | ||||||||
4997 | // sense. | ||||||||
4998 | |||||||||
4999 | assert(VF.isVector() && Uniforms.find(VF) == Uniforms.end() &&((VF.isVector() && Uniforms.find(VF) == Uniforms.end( ) && "This function should not be visited twice for the same VF" ) ? static_cast<void> (0) : __assert_fail ("VF.isVector() && Uniforms.find(VF) == Uniforms.end() && \"This function should not be visited twice for the same VF\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 5000, __PRETTY_FUNCTION__)) | ||||||||
5000 | "This function should not be visited twice for the same VF")((VF.isVector() && Uniforms.find(VF) == Uniforms.end( ) && "This function should not be visited twice for the same VF" ) ? static_cast<void> (0) : __assert_fail ("VF.isVector() && Uniforms.find(VF) == Uniforms.end() && \"This function should not be visited twice for the same VF\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 5000, __PRETTY_FUNCTION__)); | ||||||||
5001 | |||||||||
5002 | // Visit the list of Uniforms. If we'll not find any uniform value, we'll | ||||||||
5003 | // not analyze again. Uniforms.count(VF) will return 1. | ||||||||
5004 | Uniforms[VF].clear(); | ||||||||
5005 | |||||||||
5006 | // We now know that the loop is vectorizable! | ||||||||
5007 | // Collect instructions inside the loop that will remain uniform after | ||||||||
5008 | // vectorization. | ||||||||
5009 | |||||||||
5010 | // Global values, params and instructions outside of current loop are out of | ||||||||
5011 | // scope. | ||||||||
5012 | auto isOutOfScope = [&](Value *V) -> bool { | ||||||||
5013 | Instruction *I = dyn_cast<Instruction>(V); | ||||||||
5014 | return (!I || !TheLoop->contains(I)); | ||||||||
5015 | }; | ||||||||
5016 | |||||||||
5017 | SetVector<Instruction *> Worklist; | ||||||||
5018 | BasicBlock *Latch = TheLoop->getLoopLatch(); | ||||||||
5019 | |||||||||
5020 | // Instructions that are scalar with predication must not be considered | ||||||||
5021 | // uniform after vectorization, because that would create an erroneous | ||||||||
5022 | // replicating region where only a single instance out of VF should be formed. | ||||||||
5023 | // TODO: optimize such seldom cases if found important, see PR40816. | ||||||||
5024 | auto addToWorklistIfAllowed = [&](Instruction *I) -> void { | ||||||||
5025 | if (isScalarWithPredication(I, VF)) { | ||||||||
5026 | 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) | ||||||||
5027 | << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found not uniform being ScalarWithPredication: " << *I << "\n"; } } while (false); | ||||||||
5028 | return; | ||||||||
5029 | } | ||||||||
5030 | 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); | ||||||||
5031 | Worklist.insert(I); | ||||||||
5032 | }; | ||||||||
5033 | |||||||||
5034 | // Start with the conditional branch. If the branch condition is an | ||||||||
5035 | // instruction contained in the loop that is only used by the branch, it is | ||||||||
5036 | // uniform. | ||||||||
5037 | auto *Cmp = dyn_cast<Instruction>(Latch->getTerminator()->getOperand(0)); | ||||||||
5038 | if (Cmp && TheLoop->contains(Cmp) && Cmp->hasOneUse()) | ||||||||
5039 | addToWorklistIfAllowed(Cmp); | ||||||||
5040 | |||||||||
5041 | // Holds consecutive and consecutive-like pointers. Consecutive-like pointers | ||||||||
5042 | // are pointers that are treated like consecutive pointers during | ||||||||
5043 | // vectorization. The pointer operands of interleaved accesses are an | ||||||||
5044 | // example. | ||||||||
5045 | SmallSetVector<Instruction *, 8> ConsecutiveLikePtrs; | ||||||||
5046 | |||||||||
5047 | // Holds pointer operands of instructions that are possibly non-uniform. | ||||||||
5048 | SmallPtrSet<Instruction *, 8> PossibleNonUniformPtrs; | ||||||||
5049 | |||||||||
5050 | auto isUniformDecision = [&](Instruction *I, ElementCount VF) { | ||||||||
5051 | InstWidening WideningDecision = getWideningDecision(I, VF); | ||||||||
5052 | assert(WideningDecision != CM_Unknown &&((WideningDecision != CM_Unknown && "Widening decision should be ready at this moment" ) ? static_cast<void> (0) : __assert_fail ("WideningDecision != CM_Unknown && \"Widening decision should be ready at this moment\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 5053, __PRETTY_FUNCTION__)) | ||||||||
5053 | "Widening decision should be ready at this moment")((WideningDecision != CM_Unknown && "Widening decision should be ready at this moment" ) ? static_cast<void> (0) : __assert_fail ("WideningDecision != CM_Unknown && \"Widening decision should be ready at this moment\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 5053, __PRETTY_FUNCTION__)); | ||||||||
5054 | |||||||||
5055 | return (WideningDecision == CM_Widen || | ||||||||
5056 | WideningDecision == CM_Widen_Reverse || | ||||||||
5057 | WideningDecision == CM_Interleave); | ||||||||
5058 | }; | ||||||||
5059 | // Iterate over the instructions in the loop, and collect all | ||||||||
5060 | // consecutive-like pointer operands in ConsecutiveLikePtrs. If it's possible | ||||||||
5061 | // that a consecutive-like pointer operand will be scalarized, we collect it | ||||||||
5062 | // in PossibleNonUniformPtrs instead. We use two sets here because a single | ||||||||
5063 | // getelementptr instruction can be used by both vectorized and scalarized | ||||||||
5064 | // memory instructions. For example, if a loop loads and stores from the same | ||||||||
5065 | // location, but the store is conditional, the store will be scalarized, and | ||||||||
5066 | // the getelementptr won't remain uniform. | ||||||||
5067 | for (auto *BB : TheLoop->blocks()) | ||||||||
5068 | for (auto &I : *BB) { | ||||||||
5069 | // If there's no pointer operand, there's nothing to do. | ||||||||
5070 | auto *Ptr = dyn_cast_or_null<Instruction>(getLoadStorePointerOperand(&I)); | ||||||||
5071 | if (!Ptr) | ||||||||
5072 | continue; | ||||||||
5073 | |||||||||
5074 | // True if all users of Ptr are memory accesses that have Ptr as their | ||||||||
5075 | // pointer operand. | ||||||||
5076 | auto UsersAreMemAccesses = | ||||||||
5077 | llvm::all_of(Ptr->users(), [&](User *U) -> bool { | ||||||||
5078 | return getLoadStorePointerOperand(U) == Ptr; | ||||||||
5079 | }); | ||||||||
5080 | |||||||||
5081 | // Ensure the memory instruction will not be scalarized or used by | ||||||||
5082 | // gather/scatter, making its pointer operand non-uniform. If the pointer | ||||||||
5083 | // operand is used by any instruction other than a memory access, we | ||||||||
5084 | // conservatively assume the pointer operand may be non-uniform. | ||||||||
5085 | if (!UsersAreMemAccesses || !isUniformDecision(&I, VF)) | ||||||||
5086 | PossibleNonUniformPtrs.insert(Ptr); | ||||||||
5087 | |||||||||
5088 | // If the memory instruction will be vectorized and its pointer operand | ||||||||
5089 | // is consecutive-like, or interleaving - the pointer operand should | ||||||||
5090 | // remain uniform. | ||||||||
5091 | else | ||||||||
5092 | ConsecutiveLikePtrs.insert(Ptr); | ||||||||
5093 | } | ||||||||
5094 | |||||||||
5095 | // Add to the Worklist all consecutive and consecutive-like pointers that | ||||||||
5096 | // aren't also identified as possibly non-uniform. | ||||||||
5097 | for (auto *V : ConsecutiveLikePtrs) | ||||||||
5098 | if (!PossibleNonUniformPtrs.count(V)) | ||||||||
5099 | addToWorklistIfAllowed(V); | ||||||||
5100 | |||||||||
5101 | // Expand Worklist in topological order: whenever a new instruction | ||||||||
5102 | // is added , its users should be already inside Worklist. It ensures | ||||||||
5103 | // a uniform instruction will only be used by uniform instructions. | ||||||||
5104 | unsigned idx = 0; | ||||||||
5105 | while (idx != Worklist.size()) { | ||||||||
5106 | Instruction *I = Worklist[idx++]; | ||||||||
5107 | |||||||||
5108 | for (auto OV : I->operand_values()) { | ||||||||
5109 | // isOutOfScope operands cannot be uniform instructions. | ||||||||
5110 | if (isOutOfScope(OV)) | ||||||||
5111 | continue; | ||||||||
5112 | // First order recurrence Phi's should typically be considered | ||||||||
5113 | // non-uniform. | ||||||||
5114 | auto *OP = dyn_cast<PHINode>(OV); | ||||||||
5115 | if (OP && Legal->isFirstOrderRecurrence(OP)) | ||||||||
5116 | continue; | ||||||||
5117 | // If all the users of the operand are uniform, then add the | ||||||||
5118 | // operand into the uniform worklist. | ||||||||
5119 | auto *OI = cast<Instruction>(OV); | ||||||||
5120 | if (llvm::all_of(OI->users(), [&](User *U) -> bool { | ||||||||
5121 | auto *J = cast<Instruction>(U); | ||||||||
5122 | return Worklist.count(J) || | ||||||||
5123 | (OI == getLoadStorePointerOperand(J) && | ||||||||
5124 | isUniformDecision(J, VF)); | ||||||||
5125 | })) | ||||||||
5126 | addToWorklistIfAllowed(OI); | ||||||||
5127 | } | ||||||||
5128 | } | ||||||||
5129 | |||||||||
5130 | // Returns true if Ptr is the pointer operand of a memory access instruction | ||||||||
5131 | // I, and I is known to not require scalarization. | ||||||||
5132 | auto isVectorizedMemAccessUse = [&](Instruction *I, Value *Ptr) -> bool { | ||||||||
5133 | return getLoadStorePointerOperand(I) == Ptr && isUniformDecision(I, VF); | ||||||||
5134 | }; | ||||||||
5135 | |||||||||
5136 | // For an instruction to be added into Worklist above, all its users inside | ||||||||
5137 | // the loop should also be in Worklist. However, this condition cannot be | ||||||||
5138 | // true for phi nodes that form a cyclic dependence. We must process phi | ||||||||
5139 | // nodes separately. An induction variable will remain uniform if all users | ||||||||
5140 | // of the induction variable and induction variable update remain uniform. | ||||||||
5141 | // The code below handles both pointer and non-pointer induction variables. | ||||||||
5142 | for (auto &Induction : Legal->getInductionVars()) { | ||||||||
5143 | auto *Ind = Induction.first; | ||||||||
5144 | auto *IndUpdate = cast<Instruction>(Ind->getIncomingValueForBlock(Latch)); | ||||||||
5145 | |||||||||
5146 | // Determine if all users of the induction variable are uniform after | ||||||||
5147 | // vectorization. | ||||||||
5148 | auto UniformInd = llvm::all_of(Ind->users(), [&](User *U) -> bool { | ||||||||
5149 | auto *I = cast<Instruction>(U); | ||||||||
5150 | return I == IndUpdate || !TheLoop->contains(I) || Worklist.count(I) || | ||||||||
5151 | isVectorizedMemAccessUse(I, Ind); | ||||||||
5152 | }); | ||||||||
5153 | if (!UniformInd) | ||||||||
5154 | continue; | ||||||||
5155 | |||||||||
5156 | // Determine if all users of the induction variable update instruction are | ||||||||
5157 | // uniform after vectorization. | ||||||||
5158 | auto UniformIndUpdate = | ||||||||
5159 | llvm::all_of(IndUpdate->users(), [&](User *U) -> bool { | ||||||||
5160 | auto *I = cast<Instruction>(U); | ||||||||
5161 | return I == Ind || !TheLoop->contains(I) || Worklist.count(I) || | ||||||||
5162 | isVectorizedMemAccessUse(I, IndUpdate); | ||||||||
5163 | }); | ||||||||
5164 | if (!UniformIndUpdate) | ||||||||
5165 | continue; | ||||||||
5166 | |||||||||
5167 | // The induction variable and its update instruction will remain uniform. | ||||||||
5168 | addToWorklistIfAllowed(Ind); | ||||||||
5169 | addToWorklistIfAllowed(IndUpdate); | ||||||||
5170 | } | ||||||||
5171 | |||||||||
5172 | Uniforms[VF].insert(Worklist.begin(), Worklist.end()); | ||||||||
5173 | } | ||||||||
5174 | |||||||||
5175 | bool LoopVectorizationCostModel::runtimeChecksRequired() { | ||||||||
5176 | 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); | ||||||||
5177 | |||||||||
5178 | if (Legal->getRuntimePointerChecking()->Need) { | ||||||||
5179 | reportVectorizationFailure("Runtime ptr check is required with -Os/-Oz", | ||||||||
5180 | "runtime pointer checks needed. Enable vectorization of this " | ||||||||
5181 | "loop with '#pragma clang loop vectorize(enable)' when " | ||||||||
5182 | "compiling with -Os/-Oz", | ||||||||
5183 | "CantVersionLoopWithOptForSize", ORE, TheLoop); | ||||||||
5184 | return true; | ||||||||
5185 | } | ||||||||
5186 | |||||||||
5187 | if (!PSE.getUnionPredicate().getPredicates().empty()) { | ||||||||
5188 | reportVectorizationFailure("Runtime SCEV check is required with -Os/-Oz", | ||||||||
5189 | "runtime SCEV checks needed. Enable vectorization of this " | ||||||||
5190 | "loop with '#pragma clang loop vectorize(enable)' when " | ||||||||
5191 | "compiling with -Os/-Oz", | ||||||||
5192 | "CantVersionLoopWithOptForSize", ORE, TheLoop); | ||||||||
5193 | return true; | ||||||||
5194 | } | ||||||||
5195 | |||||||||
5196 | // FIXME: Avoid specializing for stride==1 instead of bailing out. | ||||||||
5197 | if (!Legal->getLAI()->getSymbolicStrides().empty()) { | ||||||||
5198 | reportVectorizationFailure("Runtime stride check for small trip count", | ||||||||
5199 | "runtime stride == 1 checks needed. Enable vectorization of " | ||||||||
5200 | "this loop without such check by compiling with -Os/-Oz", | ||||||||
5201 | "CantVersionLoopWithOptForSize", ORE, TheLoop); | ||||||||
5202 | return true; | ||||||||
5203 | } | ||||||||
5204 | |||||||||
5205 | return false; | ||||||||
5206 | } | ||||||||
5207 | |||||||||
5208 | Optional<unsigned> LoopVectorizationCostModel::computeMaxVF(unsigned UserVF, | ||||||||
5209 | unsigned UserIC) { | ||||||||
5210 | if (Legal->getRuntimePointerChecking()->Need && TTI.hasBranchDivergence()) { | ||||||||
5211 | // TODO: It may by useful to do since it's still likely to be dynamically | ||||||||
5212 | // uniform if the target can skip. | ||||||||
5213 | reportVectorizationFailure( | ||||||||
5214 | "Not inserting runtime ptr check for divergent target", | ||||||||
5215 | "runtime pointer checks needed. Not enabled for divergent target", | ||||||||
5216 | "CantVersionLoopWithDivergentTarget", ORE, TheLoop); | ||||||||
5217 | return None; | ||||||||
5218 | } | ||||||||
5219 | |||||||||
5220 | unsigned TC = PSE.getSE()->getSmallConstantTripCount(TheLoop); | ||||||||
5221 | 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); | ||||||||
5222 | if (TC == 1) { | ||||||||
5223 | reportVectorizationFailure("Single iteration (non) loop", | ||||||||
5224 | "loop trip count is one, irrelevant for vectorization", | ||||||||
5225 | "SingleIterationLoop", ORE, TheLoop); | ||||||||
5226 | return None; | ||||||||
5227 | } | ||||||||
5228 | |||||||||
5229 | switch (ScalarEpilogueStatus) { | ||||||||
5230 | case CM_ScalarEpilogueAllowed: | ||||||||
5231 | return UserVF ? UserVF : computeFeasibleMaxVF(TC); | ||||||||
5232 | case CM_ScalarEpilogueNotNeededUsePredicate: | ||||||||
5233 | 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) | ||||||||
5234 | 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) | ||||||||
5235 | << "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) | ||||||||
5236 | << "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); | ||||||||
5237 | break; | ||||||||
5238 | case CM_ScalarEpilogueNotAllowedLowTripLoop: | ||||||||
5239 | // fallthrough as a special case of OptForSize | ||||||||
5240 | case CM_ScalarEpilogueNotAllowedOptSize: | ||||||||
5241 | if (ScalarEpilogueStatus == CM_ScalarEpilogueNotAllowedOptSize) | ||||||||
5242 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Not allowing scalar epilogue due to -Os/-Oz.\n" ; } } while (false) | ||||||||
5243 | 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); | ||||||||
5244 | else | ||||||||
5245 | 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) | ||||||||
5246 | << "count.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Not allowing scalar epilogue due to low trip " << "count.\n"; } } while (false); | ||||||||
5247 | |||||||||
5248 | // Bail if runtime checks are required, which are not good when optimising | ||||||||
5249 | // for size. | ||||||||
5250 | if (runtimeChecksRequired()) | ||||||||
5251 | return None; | ||||||||
5252 | break; | ||||||||
5253 | } | ||||||||
5254 | |||||||||
5255 | // Now try the tail folding | ||||||||
5256 | |||||||||
5257 | // Invalidate interleave groups that require an epilogue if we can't mask | ||||||||
5258 | // the interleave-group. | ||||||||
5259 | if (!useMaskedInterleavedAccesses(TTI)) { | ||||||||
5260 | assert(WideningDecisions.empty() && Uniforms.empty() && Scalars.empty() &&((WideningDecisions.empty() && Uniforms.empty() && Scalars.empty() && "No decisions should have been taken at this point" ) ? static_cast<void> (0) : __assert_fail ("WideningDecisions.empty() && Uniforms.empty() && Scalars.empty() && \"No decisions should have been taken at this point\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 5261, __PRETTY_FUNCTION__)) | ||||||||
5261 | "No decisions should have been taken at this point")((WideningDecisions.empty() && Uniforms.empty() && Scalars.empty() && "No decisions should have been taken at this point" ) ? static_cast<void> (0) : __assert_fail ("WideningDecisions.empty() && Uniforms.empty() && Scalars.empty() && \"No decisions should have been taken at this point\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 5261, __PRETTY_FUNCTION__)); | ||||||||
5262 | // Note: There is no need to invalidate any cost modeling decisions here, as | ||||||||
5263 | // non where taken so far. | ||||||||
5264 | InterleaveInfo.invalidateGroupsRequiringScalarEpilogue(); | ||||||||
5265 | } | ||||||||
5266 | |||||||||
5267 | unsigned MaxVF = UserVF ? UserVF : computeFeasibleMaxVF(TC); | ||||||||
5268 | assert((UserVF || isPowerOf2_32(MaxVF)) && "MaxVF must be a power of 2")(((UserVF || isPowerOf2_32(MaxVF)) && "MaxVF must be a power of 2" ) ? static_cast<void> (0) : __assert_fail ("(UserVF || isPowerOf2_32(MaxVF)) && \"MaxVF must be a power of 2\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 5268, __PRETTY_FUNCTION__)); | ||||||||
5269 | unsigned MaxVFtimesIC = UserIC ? MaxVF * UserIC : MaxVF; | ||||||||
5270 | if (TC > 0 && TC % MaxVFtimesIC == 0) { | ||||||||
5271 | // Accept MaxVF if we do not have a tail. | ||||||||
5272 | 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); | ||||||||
5273 | return MaxVF; | ||||||||
5274 | } | ||||||||
5275 | |||||||||
5276 | // If we don't know the precise trip count, or if the trip count that we | ||||||||
5277 | // found modulo the vectorization factor is not zero, try to fold the tail | ||||||||
5278 | // by masking. | ||||||||
5279 | // FIXME: look for a smaller MaxVF that does divide TC rather than masking. | ||||||||
5280 | if (Legal->prepareToFoldTailByMasking()) { | ||||||||
5281 | FoldTailByMasking = true; | ||||||||
5282 | return MaxVF; | ||||||||
5283 | } | ||||||||
5284 | |||||||||
5285 | // If there was a tail-folding hint/switch, but we can't fold the tail by | ||||||||
5286 | // masking, fallback to a vectorization with a scalar epilogue. | ||||||||
5287 | if (ScalarEpilogueStatus == CM_ScalarEpilogueNotNeededUsePredicate) { | ||||||||
5288 | if (PreferPredicateOverEpilogue == PreferPredicateTy::PredicateOrDontVectorize) { | ||||||||
5289 | 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); | ||||||||
5290 | return None; | ||||||||
5291 | } | ||||||||
5292 | 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) | ||||||||
5293 | "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); | ||||||||
5294 | ScalarEpilogueStatus = CM_ScalarEpilogueAllowed; | ||||||||
5295 | return MaxVF; | ||||||||
5296 | } | ||||||||
5297 | |||||||||
5298 | if (TC == 0) { | ||||||||
5299 | reportVectorizationFailure( | ||||||||
5300 | "Unable to calculate the loop count due to complex control flow", | ||||||||
5301 | "unable to calculate the loop count due to complex control flow", | ||||||||
5302 | "UnknownLoopCountComplexCFG", ORE, TheLoop); | ||||||||
5303 | return None; | ||||||||
5304 | } | ||||||||
5305 | |||||||||
5306 | reportVectorizationFailure( | ||||||||
5307 | "Cannot optimize for size and vectorize at the same time.", | ||||||||
5308 | "cannot optimize for size and vectorize at the same time. " | ||||||||
5309 | "Enable vectorization of this loop with '#pragma clang loop " | ||||||||
5310 | "vectorize(enable)' when compiling with -Os/-Oz", | ||||||||
5311 | "NoTailLoopWithOptForSize", ORE, TheLoop); | ||||||||
5312 | return None; | ||||||||
5313 | } | ||||||||
5314 | |||||||||
5315 | unsigned | ||||||||
5316 | LoopVectorizationCostModel::computeFeasibleMaxVF(unsigned ConstTripCount) { | ||||||||
5317 | MinBWs = computeMinimumValueSizes(TheLoop->getBlocks(), *DB, &TTI); | ||||||||
5318 | unsigned SmallestType, WidestType; | ||||||||
5319 | std::tie(SmallestType, WidestType) = getSmallestAndWidestTypes(); | ||||||||
5320 | unsigned WidestRegister = TTI.getRegisterBitWidth(true); | ||||||||
5321 | |||||||||
5322 | // Get the maximum safe dependence distance in bits computed by LAA. | ||||||||
5323 | // It is computed by MaxVF * sizeOf(type) * 8, where type is taken from | ||||||||
5324 | // the memory accesses that is most restrictive (involved in the smallest | ||||||||
5325 | // dependence distance). | ||||||||
5326 | unsigned MaxSafeRegisterWidth = Legal->getMaxSafeRegisterWidth(); | ||||||||
5327 | |||||||||
5328 | WidestRegister = std::min(WidestRegister, MaxSafeRegisterWidth); | ||||||||
5329 | |||||||||
5330 | // Ensure MaxVF is a power of 2; the dependence distance bound may not be. | ||||||||
5331 | // Note that both WidestRegister and WidestType may not be a powers of 2. | ||||||||
5332 | unsigned MaxVectorSize = PowerOf2Floor(WidestRegister / WidestType); | ||||||||
5333 | |||||||||
5334 | 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) | ||||||||
5335 | << " / " << WidestType << " bits.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: The Smallest and Widest types: " << SmallestType << " / " << WidestType << " bits.\n"; } } while (false); | ||||||||
5336 | 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: " << WidestRegister << " bits.\n"; } } while (false ) | ||||||||
5337 | << WidestRegister << " bits.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: The Widest register safe to use is: " << WidestRegister << " bits.\n"; } } while (false ); | ||||||||
5338 | |||||||||
5339 | assert(MaxVectorSize <= 256 && "Did not expect to pack so many elements"((MaxVectorSize <= 256 && "Did not expect to pack so many elements" " into one vector!") ? static_cast<void> (0) : __assert_fail ("MaxVectorSize <= 256 && \"Did not expect to pack so many elements\" \" into one vector!\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 5340, __PRETTY_FUNCTION__)) | ||||||||
5340 | " into one vector!")((MaxVectorSize <= 256 && "Did not expect to pack so many elements" " into one vector!") ? static_cast<void> (0) : __assert_fail ("MaxVectorSize <= 256 && \"Did not expect to pack so many elements\" \" into one vector!\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 5340, __PRETTY_FUNCTION__)); | ||||||||
5341 | if (MaxVectorSize == 0) { | ||||||||
5342 | LLVM_DEBUG(dbgs() << "LV: The target has no vector registers.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: The target has no vector registers.\n" ; } } while (false); | ||||||||
5343 | MaxVectorSize = 1; | ||||||||
5344 | return MaxVectorSize; | ||||||||
5345 | } else if (ConstTripCount && ConstTripCount < MaxVectorSize && | ||||||||
5346 | isPowerOf2_32(ConstTripCount)) { | ||||||||
5347 | // We need to clamp the VF to be the ConstTripCount. There is no point in | ||||||||
5348 | // choosing a higher viable VF as done in the loop below. | ||||||||
5349 | LLVM_DEBUG(dbgs() << "LV: Clamping the MaxVF to the constant trip count: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Clamping the MaxVF to the constant trip count: " << ConstTripCount << "\n"; } } while (false) | ||||||||
5350 | << ConstTripCount << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Clamping the MaxVF to the constant trip count: " << ConstTripCount << "\n"; } } while (false); | ||||||||
5351 | MaxVectorSize = ConstTripCount; | ||||||||
5352 | return MaxVectorSize; | ||||||||
5353 | } | ||||||||
5354 | |||||||||
5355 | unsigned MaxVF = MaxVectorSize; | ||||||||
5356 | if (TTI.shouldMaximizeVectorBandwidth(!isScalarEpilogueAllowed()) || | ||||||||
5357 | (MaximizeBandwidth && isScalarEpilogueAllowed())) { | ||||||||
5358 | // Collect all viable vectorization factors larger than the default MaxVF | ||||||||
5359 | // (i.e. MaxVectorSize). | ||||||||
5360 | SmallVector<ElementCount, 8> VFs; | ||||||||
5361 | unsigned NewMaxVectorSize = WidestRegister / SmallestType; | ||||||||
5362 | for (unsigned VS = MaxVectorSize * 2; VS <= NewMaxVectorSize; VS *= 2) | ||||||||
5363 | VFs.push_back(ElementCount::getFixed(VS)); | ||||||||
5364 | |||||||||
5365 | // For each VF calculate its register usage. | ||||||||
5366 | auto RUs = calculateRegisterUsage(VFs); | ||||||||
5367 | |||||||||
5368 | // Select the largest VF which doesn't require more registers than existing | ||||||||
5369 | // ones. | ||||||||
5370 | for (int i = RUs.size() - 1; i >= 0; --i) { | ||||||||
5371 | bool Selected = true; | ||||||||
5372 | for (auto& pair : RUs[i].MaxLocalUsers) { | ||||||||
5373 | unsigned TargetNumRegisters = TTI.getNumberOfRegisters(pair.first); | ||||||||
5374 | if (pair.second > TargetNumRegisters) | ||||||||
5375 | Selected = false; | ||||||||
5376 | } | ||||||||
5377 | if (Selected) { | ||||||||
5378 | MaxVF = VFs[i].getKnownMinValue(); | ||||||||
5379 | break; | ||||||||
5380 | } | ||||||||
5381 | } | ||||||||
5382 | if (unsigned MinVF = TTI.getMinimumVF(SmallestType)) { | ||||||||
5383 | if (MaxVF < MinVF) { | ||||||||
5384 | 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) | ||||||||
5385 | << ") 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); | ||||||||
5386 | MaxVF = MinVF; | ||||||||
5387 | } | ||||||||
5388 | } | ||||||||
5389 | } | ||||||||
5390 | return MaxVF; | ||||||||
5391 | } | ||||||||
5392 | |||||||||
5393 | VectorizationFactor | ||||||||
5394 | LoopVectorizationCostModel::selectVectorizationFactor(unsigned MaxVF) { | ||||||||
5395 | float Cost = expectedCost(ElementCount::getFixed(1)).first; | ||||||||
5396 | const float ScalarCost = Cost; | ||||||||
5397 | unsigned Width = 1; | ||||||||
5398 | LLVM_DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n"; } } while (false); | ||||||||
5399 | |||||||||
5400 | bool ForceVectorization = Hints->getForce() == LoopVectorizeHints::FK_Enabled; | ||||||||
5401 | if (ForceVectorization && MaxVF > 1) { | ||||||||
5402 | // Ignore scalar width, because the user explicitly wants vectorization. | ||||||||
5403 | // Initialize cost to max so that VF = 2 is, at least, chosen during cost | ||||||||
5404 | // evaluation. | ||||||||
5405 | Cost = std::numeric_limits<float>::max(); | ||||||||
5406 | } | ||||||||
5407 | |||||||||
5408 | for (unsigned i = 2; i <= MaxVF; i *= 2) { | ||||||||
5409 | // Notice that the vector loop needs to be executed less times, so | ||||||||
5410 | // we need to divide the cost of the vector loops by the width of | ||||||||
5411 | // the vector elements. | ||||||||
5412 | VectorizationCostTy C = expectedCost(ElementCount::getFixed(i)); | ||||||||
5413 | float VectorCost = C.first / (float)i; | ||||||||
5414 | LLVM_DEBUG(dbgs() << "LV: Vector loop of width " << ido { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Vector loop of width " << i << " costs: " << (int)VectorCost << ".\n"; } } while (false) | ||||||||
5415 | << " costs: " << (int)VectorCost << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Vector loop of width " << i << " costs: " << (int)VectorCost << ".\n"; } } while (false); | ||||||||
5416 | if (!C.second && !ForceVectorization) { | ||||||||
5417 | 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) | ||||||||
5418 | 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) | ||||||||
5419 | << " 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); | ||||||||
5420 | continue; | ||||||||
5421 | } | ||||||||
5422 | if (VectorCost < Cost) { | ||||||||
5423 | Cost = VectorCost; | ||||||||
5424 | Width = i; | ||||||||
5425 | } | ||||||||
5426 | } | ||||||||
5427 | |||||||||
5428 | if (!EnableCondStoresVectorization && NumPredStores) { | ||||||||
5429 | reportVectorizationFailure("There are conditional stores.", | ||||||||
5430 | "store that is conditionally executed prevents vectorization", | ||||||||
5431 | "ConditionalStore", ORE, TheLoop); | ||||||||
5432 | Width = 1; | ||||||||
5433 | Cost = ScalarCost; | ||||||||
5434 | } | ||||||||
5435 | |||||||||
5436 | LLVM_DEBUG(if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs() << "LV: Vectorization seems to be not beneficial, " << "but was forced by a user.\n"; } } while (false) | ||||||||
5437 | << "LV: Vectorization seems to be not beneficial, "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs() << "LV: Vectorization seems to be not beneficial, " << "but was forced by a user.\n"; } } while (false) | ||||||||
5438 | << "but was forced by a user.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs() << "LV: Vectorization seems to be not beneficial, " << "but was forced by a user.\n"; } } while (false); | ||||||||
5439 | LLVM_DEBUG(dbgs() << "LV: Selecting VF: " << Width << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Selecting VF: " << Width << ".\n"; } } while (false); | ||||||||
5440 | VectorizationFactor Factor = {ElementCount::getFixed(Width), | ||||||||
5441 | (unsigned)(Width * Cost)}; | ||||||||
5442 | return Factor; | ||||||||
5443 | } | ||||||||
5444 | |||||||||
5445 | std::pair<unsigned, unsigned> | ||||||||
5446 | LoopVectorizationCostModel::getSmallestAndWidestTypes() { | ||||||||
5447 | unsigned MinWidth = -1U; | ||||||||
5448 | unsigned MaxWidth = 8; | ||||||||
5449 | const DataLayout &DL = TheFunction->getParent()->getDataLayout(); | ||||||||
5450 | |||||||||
5451 | // For each block. | ||||||||
5452 | for (BasicBlock *BB : TheLoop->blocks()) { | ||||||||
5453 | // For each instruction in the loop. | ||||||||
5454 | for (Instruction &I : BB->instructionsWithoutDebug()) { | ||||||||
5455 | Type *T = I.getType(); | ||||||||
5456 | |||||||||
5457 | // Skip ignored values. | ||||||||
5458 | if (ValuesToIgnore.count(&I)) | ||||||||
5459 | continue; | ||||||||
5460 | |||||||||
5461 | // Only examine Loads, Stores and PHINodes. | ||||||||
5462 | if (!isa<LoadInst>(I) && !isa<StoreInst>(I) && !isa<PHINode>(I)) | ||||||||
5463 | continue; | ||||||||
5464 | |||||||||
5465 | // Examine PHI nodes that are reduction variables. Update the type to | ||||||||
5466 | // account for the recurrence type. | ||||||||
5467 | if (auto *PN = dyn_cast<PHINode>(&I)) { | ||||||||
5468 | if (!Legal->isReductionVariable(PN)) | ||||||||
5469 | continue; | ||||||||
5470 | RecurrenceDescriptor RdxDesc = Legal->getReductionVars()[PN]; | ||||||||
5471 | T = RdxDesc.getRecurrenceType(); | ||||||||
5472 | } | ||||||||
5473 | |||||||||
5474 | // Examine the stored values. | ||||||||
5475 | if (auto *ST = dyn_cast<StoreInst>(&I)) | ||||||||
5476 | T = ST->getValueOperand()->getType(); | ||||||||
5477 | |||||||||
5478 | // Ignore loaded pointer types and stored pointer types that are not | ||||||||
5479 | // vectorizable. | ||||||||
5480 | // | ||||||||
5481 | // FIXME: The check here attempts to predict whether a load or store will | ||||||||
5482 | // be vectorized. We only know this for certain after a VF has | ||||||||
5483 | // been selected. Here, we assume that if an access can be | ||||||||
5484 | // vectorized, it will be. We should also look at extending this | ||||||||
5485 | // optimization to non-pointer types. | ||||||||
5486 | // | ||||||||
5487 | if (T->isPointerTy() && !isConsecutiveLoadOrStore(&I) && | ||||||||
5488 | !isAccessInterleaved(&I) && !isLegalGatherOrScatter(&I)) | ||||||||
5489 | continue; | ||||||||
5490 | |||||||||
5491 | MinWidth = std::min(MinWidth, | ||||||||
5492 | (unsigned)DL.getTypeSizeInBits(T->getScalarType())); | ||||||||
5493 | MaxWidth = std::max(MaxWidth, | ||||||||
5494 | (unsigned)DL.getTypeSizeInBits(T->getScalarType())); | ||||||||
5495 | } | ||||||||
5496 | } | ||||||||
5497 | |||||||||
5498 | return {MinWidth, MaxWidth}; | ||||||||
5499 | } | ||||||||
5500 | |||||||||
5501 | unsigned LoopVectorizationCostModel::selectInterleaveCount(ElementCount VF, | ||||||||
5502 | unsigned LoopCost) { | ||||||||
5503 | // -- The interleave heuristics -- | ||||||||
5504 | // We interleave the loop in order to expose ILP and reduce the loop overhead. | ||||||||
5505 | // There are many micro-architectural considerations that we can't predict | ||||||||
5506 | // at this level. For example, frontend pressure (on decode or fetch) due to | ||||||||
5507 | // code size, or the number and capabilities of the execution ports. | ||||||||
5508 | // | ||||||||
5509 | // We use the following heuristics to select the interleave count: | ||||||||
5510 | // 1. If the code has reductions, then we interleave to break the cross | ||||||||
5511 | // iteration dependency. | ||||||||
5512 | // 2. If the loop is really small, then we interleave to reduce the loop | ||||||||
5513 | // overhead. | ||||||||
5514 | // 3. We don't interleave if we think that we will spill registers to memory | ||||||||
5515 | // due to the increased register pressure. | ||||||||
5516 | |||||||||
5517 | if (!isScalarEpilogueAllowed()) | ||||||||
5518 | return 1; | ||||||||
5519 | |||||||||
5520 | // We used the distance for the interleave count. | ||||||||
5521 | if (Legal->getMaxSafeDepDistBytes() != -1U) | ||||||||
5522 | return 1; | ||||||||
5523 | |||||||||
5524 | auto BestKnownTC = getSmallBestKnownTC(*PSE.getSE(), TheLoop); | ||||||||
5525 | const bool HasReductions = !Legal->getReductionVars().empty(); | ||||||||
5526 | // Do not interleave loops with a relatively small known or estimated trip | ||||||||
5527 | // count. But we will interleave when InterleaveSmallLoopScalarReduction is | ||||||||
5528 | // enabled, and the code has scalar reductions(HasReductions && VF = 1), | ||||||||
5529 | // because with the above conditions interleaving can expose ILP and break | ||||||||
5530 | // cross iteration dependences for reductions. | ||||||||
5531 | if (BestKnownTC && (*BestKnownTC < TinyTripCountInterleaveThreshold) && | ||||||||
5532 | !(InterleaveSmallLoopScalarReduction && HasReductions && VF.isScalar())) | ||||||||
5533 | return 1; | ||||||||
5534 | |||||||||
5535 | RegisterUsage R = calculateRegisterUsage({VF})[0]; | ||||||||
5536 | // We divide by these constants so assume that we have at least one | ||||||||
5537 | // instruction that uses at least one register. | ||||||||
5538 | for (auto& pair : R.MaxLocalUsers) { | ||||||||
5539 | pair.second = std::max(pair.second, 1U); | ||||||||
5540 | } | ||||||||
5541 | |||||||||
5542 | // We calculate the interleave count using the following formula. | ||||||||
5543 | // Subtract the number of loop invariants from the number of available | ||||||||
5544 | // registers. These registers are used by all of the interleaved instances. | ||||||||
5545 | // Next, divide the remaining registers by the number of registers that is | ||||||||
5546 | // required by the loop, in order to estimate how many parallel instances | ||||||||
5547 | // fit without causing spills. All of this is rounded down if necessary to be | ||||||||
5548 | // a power of two. We want power of two interleave count to simplify any | ||||||||
5549 | // addressing operations or alignment considerations. | ||||||||
5550 | // We also want power of two interleave counts to ensure that the induction | ||||||||
5551 | // variable of the vector loop wraps to zero, when tail is folded by masking; | ||||||||
5552 | // this currently happens when OptForSize, in which case IC is set to 1 above. | ||||||||
5553 | unsigned IC = UINT_MAX(2147483647 *2U +1U); | ||||||||
5554 | |||||||||
5555 | for (auto& pair : R.MaxLocalUsers) { | ||||||||
5556 | unsigned TargetNumRegisters = TTI.getNumberOfRegisters(pair.first); | ||||||||
5557 | 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) | ||||||||
5558 | << " 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) | ||||||||
5559 | << 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); | ||||||||
5560 | if (VF.isScalar()) { | ||||||||
5561 | if (ForceTargetNumScalarRegs.getNumOccurrences() > 0) | ||||||||
5562 | TargetNumRegisters = ForceTargetNumScalarRegs; | ||||||||
5563 | } else { | ||||||||
5564 | if (ForceTargetNumVectorRegs.getNumOccurrences() > 0) | ||||||||
5565 | TargetNumRegisters = ForceTargetNumVectorRegs; | ||||||||
5566 | } | ||||||||
5567 | unsigned MaxLocalUsers = pair.second; | ||||||||
5568 | unsigned LoopInvariantRegs = 0; | ||||||||
5569 | if (R.LoopInvariantRegs.find(pair.first) != R.LoopInvariantRegs.end()) | ||||||||
5570 | LoopInvariantRegs = R.LoopInvariantRegs[pair.first]; | ||||||||
5571 | |||||||||
5572 | unsigned TmpIC = PowerOf2Floor((TargetNumRegisters - LoopInvariantRegs) / MaxLocalUsers); | ||||||||
5573 | // Don't count the induction variable as interleaved. | ||||||||
5574 | if (EnableIndVarRegisterHeur) { | ||||||||
5575 | TmpIC = | ||||||||
5576 | PowerOf2Floor((TargetNumRegisters - LoopInvariantRegs - 1) / | ||||||||
5577 | std::max(1U, (MaxLocalUsers - 1))); | ||||||||
5578 | } | ||||||||
5579 | |||||||||
5580 | IC = std::min(IC, TmpIC); | ||||||||
5581 | } | ||||||||
5582 | |||||||||
5583 | // Clamp the interleave ranges to reasonable counts. | ||||||||
5584 | assert(!VF.isScalable() && "scalable vectors not yet supported.")((!VF.isScalable() && "scalable vectors not yet supported." ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"scalable vectors not yet supported.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 5584, __PRETTY_FUNCTION__)); | ||||||||
5585 | unsigned MaxInterleaveCount = | ||||||||
5586 | TTI.getMaxInterleaveFactor(VF.getKnownMinValue()); | ||||||||
5587 | |||||||||
5588 | // Check if the user has overridden the max. | ||||||||
5589 | if (VF.isScalar()) { | ||||||||
5590 | if (ForceTargetMaxScalarInterleaveFactor.getNumOccurrences() > 0) | ||||||||
5591 | MaxInterleaveCount = ForceTargetMaxScalarInterleaveFactor; | ||||||||
5592 | } else { | ||||||||
5593 | if (ForceTargetMaxVectorInterleaveFactor.getNumOccurrences() > 0) | ||||||||
5594 | MaxInterleaveCount = ForceTargetMaxVectorInterleaveFactor; | ||||||||
5595 | } | ||||||||
5596 | |||||||||
5597 | // If trip count is known or estimated compile time constant, limit the | ||||||||
5598 | // interleave count to be less than the trip count divided by VF. | ||||||||
5599 | if (BestKnownTC) { | ||||||||
5600 | MaxInterleaveCount = | ||||||||
5601 | std::min(*BestKnownTC / VF.getKnownMinValue(), MaxInterleaveCount); | ||||||||
5602 | } | ||||||||
5603 | |||||||||
5604 | // If we did not calculate the cost for VF (because the user selected the VF) | ||||||||
5605 | // then we calculate the cost of VF here. | ||||||||
5606 | if (LoopCost == 0) | ||||||||
5607 | LoopCost = expectedCost(VF).first; | ||||||||
5608 | |||||||||
5609 | assert(LoopCost && "Non-zero loop cost expected")((LoopCost && "Non-zero loop cost expected") ? static_cast <void> (0) : __assert_fail ("LoopCost && \"Non-zero loop cost expected\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 5609, __PRETTY_FUNCTION__)); | ||||||||
5610 | |||||||||
5611 | // Clamp the calculated IC to be between the 1 and the max interleave count | ||||||||
5612 | // that the target and trip count allows. | ||||||||
5613 | if (IC > MaxInterleaveCount) | ||||||||
5614 | IC = MaxInterleaveCount; | ||||||||
5615 | else if (IC < 1) | ||||||||
5616 | IC = 1; | ||||||||
5617 | |||||||||
5618 | // Interleave if we vectorized this loop and there is a reduction that could | ||||||||
5619 | // benefit from interleaving. | ||||||||
5620 | if (VF.isVector() && HasReductions) { | ||||||||
5621 | 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); | ||||||||
5622 | return IC; | ||||||||
5623 | } | ||||||||
5624 | |||||||||
5625 | // Note that if we've already vectorized the loop we will have done the | ||||||||
5626 | // runtime check and so interleaving won't require further checks. | ||||||||
5627 | bool InterleavingRequiresRuntimePointerCheck = | ||||||||
5628 | (VF.isScalar() && Legal->getRuntimePointerChecking()->Need); | ||||||||
5629 | |||||||||
5630 | // We want to interleave small loops in order to reduce the loop overhead and | ||||||||
5631 | // potentially expose ILP opportunities. | ||||||||
5632 | 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.getKnownMinValue() << '\n'; } } while (false) | ||||||||
5633 | << "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.getKnownMinValue() << '\n'; } } while (false) | ||||||||
5634 | << "LV: VF is " << VF.getKnownMinValue() << '\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.getKnownMinValue() << '\n'; } } while (false); | ||||||||
5635 | const bool AggressivelyInterleaveReductions = | ||||||||
5636 | TTI.enableAggressiveInterleaving(HasReductions); | ||||||||
5637 | if (!InterleavingRequiresRuntimePointerCheck && LoopCost < SmallLoopCost) { | ||||||||
5638 | // We assume that the cost overhead is 1 and we use the cost model | ||||||||
5639 | // to estimate the cost of the loop and interleave until the cost of the | ||||||||
5640 | // loop overhead is about 5% of the cost of the loop. | ||||||||
5641 | unsigned SmallIC = | ||||||||
5642 | std::min(IC, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost)); | ||||||||
5643 | |||||||||
5644 | // Interleave until store/load ports (estimated by max interleave count) are | ||||||||
5645 | // saturated. | ||||||||
5646 | unsigned NumStores = Legal->getNumStores(); | ||||||||
5647 | unsigned NumLoads = Legal->getNumLoads(); | ||||||||
5648 | unsigned StoresIC = IC / (NumStores ? NumStores : 1); | ||||||||
5649 | unsigned LoadsIC = IC / (NumLoads ? NumLoads : 1); | ||||||||
5650 | |||||||||
5651 | // If we have a scalar reduction (vector reductions are already dealt with | ||||||||
5652 | // by this point), we can increase the critical path length if the loop | ||||||||
5653 | // we're interleaving is inside another loop. Limit, by default to 2, so the | ||||||||
5654 | // critical path only gets increased by one reduction operation. | ||||||||
5655 | if (HasReductions && TheLoop->getLoopDepth() > 1) { | ||||||||
5656 | unsigned F = static_cast<unsigned>(MaxNestedScalarReductionIC); | ||||||||
5657 | SmallIC = std::min(SmallIC, F); | ||||||||
5658 | StoresIC = std::min(StoresIC, F); | ||||||||
5659 | LoadsIC = std::min(LoadsIC, F); | ||||||||
5660 | } | ||||||||
5661 | |||||||||
5662 | if (EnableLoadStoreRuntimeInterleave && | ||||||||
5663 | std::max(StoresIC, LoadsIC) > SmallIC) { | ||||||||
5664 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Interleaving to saturate store or load ports.\n" ; } } while (false) | ||||||||
5665 | 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); | ||||||||
5666 | return std::max(StoresIC, LoadsIC); | ||||||||
5667 | } | ||||||||
5668 | |||||||||
5669 | // If there are scalar reductions and TTI has enabled aggressive | ||||||||
5670 | // interleaving for reductions, we will interleave to expose ILP. | ||||||||
5671 | if (InterleaveSmallLoopScalarReduction && VF.isScalar() && | ||||||||
5672 | AggressivelyInterleaveReductions) { | ||||||||
5673 | 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); | ||||||||
5674 | // Interleave no less than SmallIC but not as aggressive as the normal IC | ||||||||
5675 | // to satisfy the rare situation when resources are too limited. | ||||||||
5676 | return std::max(IC / 2, SmallIC); | ||||||||
5677 | } else { | ||||||||
5678 | 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); | ||||||||
5679 | return SmallIC; | ||||||||
5680 | } | ||||||||
5681 | } | ||||||||
5682 | |||||||||
5683 | // Interleave if this is a large loop (small loops are already dealt with by | ||||||||
5684 | // this point) that could benefit from interleaving. | ||||||||
5685 | if (AggressivelyInterleaveReductions) { | ||||||||
5686 | 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); | ||||||||
5687 | return IC; | ||||||||
5688 | } | ||||||||
5689 | |||||||||
5690 | LLVM_DEBUG(dbgs() << "LV: Not Interleaving.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Not Interleaving.\n" ; } } while (false); | ||||||||
5691 | return 1; | ||||||||
5692 | } | ||||||||
5693 | |||||||||
5694 | SmallVector<LoopVectorizationCostModel::RegisterUsage, 8> | ||||||||
5695 | LoopVectorizationCostModel::calculateRegisterUsage(ArrayRef<ElementCount> VFs) { | ||||||||
5696 | // This function calculates the register usage by measuring the highest number | ||||||||
5697 | // of values that are alive at a single location. Obviously, this is a very | ||||||||
5698 | // rough estimation. We scan the loop in a topological order in order and | ||||||||
5699 | // assign a number to each instruction. We use RPO to ensure that defs are | ||||||||
5700 | // met before their users. We assume that each instruction that has in-loop | ||||||||
5701 | // users starts an interval. We record every time that an in-loop value is | ||||||||
5702 | // used, so we have a list of the first and last occurrences of each | ||||||||
5703 | // instruction. Next, we transpose this data structure into a multi map that | ||||||||
5704 | // holds the list of intervals that *end* at a specific location. This multi | ||||||||
5705 | // map allows us to perform a linear search. We scan the instructions linearly | ||||||||
5706 | // and record each time that a new interval starts, by placing it in a set. | ||||||||
5707 | // If we find this value in the multi-map then we remove it from the set. | ||||||||
5708 | // The max register usage is the maximum size of the set. | ||||||||
5709 | // We also search for instructions that are defined outside the loop, but are | ||||||||
5710 | // used inside the loop. We need this number separately from the max-interval | ||||||||
5711 | // usage number because when we unroll, loop-invariant values do not take | ||||||||
5712 | // more register. | ||||||||
5713 | LoopBlocksDFS DFS(TheLoop); | ||||||||
5714 | DFS.perform(LI); | ||||||||
5715 | |||||||||
5716 | RegisterUsage RU; | ||||||||
5717 | |||||||||
5718 | // Each 'key' in the map opens a new interval. The values | ||||||||
5719 | // of the map are the index of the 'last seen' usage of the | ||||||||
5720 | // instruction that is the key. | ||||||||
5721 | using IntervalMap = DenseMap<Instruction *, unsigned>; | ||||||||
5722 | |||||||||
5723 | // Maps instruction to its index. | ||||||||
5724 | SmallVector<Instruction *, 64> IdxToInstr; | ||||||||
5725 | // Marks the end of each interval. | ||||||||
5726 | IntervalMap EndPoint; | ||||||||
5727 | // Saves the list of instruction indices that are used in the loop. | ||||||||
5728 | SmallPtrSet<Instruction *, 8> Ends; | ||||||||
5729 | // Saves the list of values that are used in the loop but are | ||||||||
5730 | // defined outside the loop, such as arguments and constants. | ||||||||
5731 | SmallPtrSet<Value *, 8> LoopInvariants; | ||||||||
5732 | |||||||||
5733 | for (BasicBlock *BB : make_range(DFS.beginRPO(), DFS.endRPO())) { | ||||||||
5734 | for (Instruction &I : BB->instructionsWithoutDebug()) { | ||||||||
5735 | IdxToInstr.push_back(&I); | ||||||||
5736 | |||||||||
5737 | // Save the end location of each USE. | ||||||||
5738 | for (Value *U : I.operands()) { | ||||||||
5739 | auto *Instr = dyn_cast<Instruction>(U); | ||||||||
5740 | |||||||||
5741 | // Ignore non-instruction values such as arguments, constants, etc. | ||||||||
5742 | if (!Instr) | ||||||||
5743 | continue; | ||||||||
5744 | |||||||||
5745 | // If this instruction is outside the loop then record it and continue. | ||||||||
5746 | if (!TheLoop->contains(Instr)) { | ||||||||
5747 | LoopInvariants.insert(Instr); | ||||||||
5748 | continue; | ||||||||
5749 | } | ||||||||
5750 | |||||||||
5751 | // Overwrite previous end points. | ||||||||
5752 | EndPoint[Instr] = IdxToInstr.size(); | ||||||||
5753 | Ends.insert(Instr); | ||||||||
5754 | } | ||||||||
5755 | } | ||||||||
5756 | } | ||||||||
5757 | |||||||||
5758 | // Saves the list of intervals that end with the index in 'key'. | ||||||||
5759 | using InstrList = SmallVector<Instruction *, 2>; | ||||||||
5760 | DenseMap<unsigned, InstrList> TransposeEnds; | ||||||||
5761 | |||||||||
5762 | // Transpose the EndPoints to a list of values that end at each index. | ||||||||
5763 | for (auto &Interval : EndPoint) | ||||||||
5764 | TransposeEnds[Interval.second].push_back(Interval.first); | ||||||||
5765 | |||||||||
5766 | SmallPtrSet<Instruction *, 8> OpenIntervals; | ||||||||
5767 | |||||||||
5768 | // Get the size of the widest register. | ||||||||
5769 | unsigned MaxSafeDepDist = -1U; | ||||||||
5770 | if (Legal->getMaxSafeDepDistBytes() != -1U) | ||||||||
5771 | MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8; | ||||||||
5772 | unsigned WidestRegister = | ||||||||
5773 | std::min(TTI.getRegisterBitWidth(true), MaxSafeDepDist); | ||||||||
5774 | const DataLayout &DL = TheFunction->getParent()->getDataLayout(); | ||||||||
5775 | |||||||||
5776 | SmallVector<RegisterUsage, 8> RUs(VFs.size()); | ||||||||
5777 | SmallVector<SmallMapVector<unsigned, unsigned, 4>, 8> MaxUsages(VFs.size()); | ||||||||
5778 | |||||||||
5779 | 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); | ||||||||
5780 | |||||||||
5781 | // A lambda that gets the register usage for the given type and VF. | ||||||||
5782 | auto GetRegUsage = [&DL, WidestRegister](Type *Ty, ElementCount VF) { | ||||||||
5783 | if (Ty->isTokenTy()) | ||||||||
5784 | return 0U; | ||||||||
5785 | unsigned TypeSize = DL.getTypeSizeInBits(Ty->getScalarType()); | ||||||||
5786 | assert(!VF.isScalable() && "scalable vectors not yet supported.")((!VF.isScalable() && "scalable vectors not yet supported." ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"scalable vectors not yet supported.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 5786, __PRETTY_FUNCTION__)); | ||||||||
5787 | return std::max<unsigned>(1, VF.getKnownMinValue() * TypeSize / | ||||||||
5788 | WidestRegister); | ||||||||
5789 | }; | ||||||||
5790 | |||||||||
5791 | for (unsigned int i = 0, s = IdxToInstr.size(); i < s; ++i) { | ||||||||
5792 | Instruction *I = IdxToInstr[i]; | ||||||||
5793 | |||||||||
5794 | // Remove all of the instructions that end at this location. | ||||||||
5795 | InstrList &List = TransposeEnds[i]; | ||||||||
5796 | for (Instruction *ToRemove : List) | ||||||||
5797 | OpenIntervals.erase(ToRemove); | ||||||||
5798 | |||||||||
5799 | // Ignore instructions that are never used within the loop. | ||||||||
5800 | if (!Ends.count(I)) | ||||||||
5801 | continue; | ||||||||
5802 | |||||||||
5803 | // Skip ignored values. | ||||||||
5804 | if (ValuesToIgnore.count(I)) | ||||||||
5805 | continue; | ||||||||
5806 | |||||||||
5807 | // For each VF find the maximum usage of registers. | ||||||||
5808 | for (unsigned j = 0, e = VFs.size(); j < e; ++j) { | ||||||||
5809 | // Count the number of live intervals. | ||||||||
5810 | SmallMapVector<unsigned, unsigned, 4> RegUsage; | ||||||||
5811 | |||||||||
5812 | if (VFs[j].isScalar()) { | ||||||||
5813 | for (auto Inst : OpenIntervals) { | ||||||||
5814 | unsigned ClassID = TTI.getRegisterClassForType(false, Inst->getType()); | ||||||||
5815 | if (RegUsage.find(ClassID) == RegUsage.end()) | ||||||||
5816 | RegUsage[ClassID] = 1; | ||||||||
5817 | else | ||||||||
5818 | RegUsage[ClassID] += 1; | ||||||||
5819 | } | ||||||||
5820 | } else { | ||||||||
5821 | collectUniformsAndScalars(VFs[j]); | ||||||||
5822 | for (auto Inst : OpenIntervals) { | ||||||||
5823 | // Skip ignored values for VF > 1. | ||||||||
5824 | if (VecValuesToIgnore.count(Inst)) | ||||||||
5825 | continue; | ||||||||
5826 | if (isScalarAfterVectorization(Inst, VFs[j])) { | ||||||||
5827 | unsigned ClassID = TTI.getRegisterClassForType(false, Inst->getType()); | ||||||||
5828 | if (RegUsage.find(ClassID) == RegUsage.end()) | ||||||||
5829 | RegUsage[ClassID] = 1; | ||||||||
5830 | else | ||||||||
5831 | RegUsage[ClassID] += 1; | ||||||||
5832 | } else { | ||||||||
5833 | unsigned ClassID = TTI.getRegisterClassForType(true, Inst->getType()); | ||||||||
5834 | if (RegUsage.find(ClassID) == RegUsage.end()) | ||||||||
5835 | RegUsage[ClassID] = GetRegUsage(Inst->getType(), VFs[j]); | ||||||||
5836 | else | ||||||||
5837 | RegUsage[ClassID] += GetRegUsage(Inst->getType(), VFs[j]); | ||||||||
5838 | } | ||||||||
5839 | } | ||||||||
5840 | } | ||||||||
5841 | |||||||||
5842 | for (auto& pair : RegUsage) { | ||||||||
5843 | if (MaxUsages[j].find(pair.first) != MaxUsages[j].end()) | ||||||||
5844 | MaxUsages[j][pair.first] = std::max(MaxUsages[j][pair.first], pair.second); | ||||||||
5845 | else | ||||||||
5846 | MaxUsages[j][pair.first] = pair.second; | ||||||||
5847 | } | ||||||||
5848 | } | ||||||||
5849 | |||||||||
5850 | 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) | ||||||||
5851 | << OpenIntervals.size() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV(REG): At #" << i << " Interval # " << OpenIntervals.size() << '\n'; } } while (false); | ||||||||
5852 | |||||||||
5853 | // Add the current instruction to the list of open intervals. | ||||||||
5854 | OpenIntervals.insert(I); | ||||||||
5855 | } | ||||||||
5856 | |||||||||
5857 | for (unsigned i = 0, e = VFs.size(); i < e; ++i) { | ||||||||
5858 | SmallMapVector<unsigned, unsigned, 4> Invariant; | ||||||||
5859 | |||||||||
5860 | for (auto Inst : LoopInvariants) { | ||||||||
5861 | unsigned Usage = | ||||||||
5862 | VFs[i].isScalar() ? 1 : GetRegUsage(Inst->getType(), VFs[i]); | ||||||||
5863 | unsigned ClassID = | ||||||||
5864 | TTI.getRegisterClassForType(VFs[i].isVector(), Inst->getType()); | ||||||||
5865 | if (Invariant.find(ClassID) == Invariant.end()) | ||||||||
5866 | Invariant[ClassID] = Usage; | ||||||||
5867 | else | ||||||||
5868 | Invariant[ClassID] += Usage; | ||||||||
5869 | } | ||||||||
5870 | |||||||||
5871 | 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) | ||||||||
5872 | 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) | ||||||||
5873 | 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) | ||||||||
5874 | << " 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) | ||||||||
5875 | 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) | ||||||||
5876 | 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) | ||||||||
5877 | << 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) | ||||||||
5878 | << " 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) | ||||||||
5879 | }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) | ||||||||
5880 | 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) | ||||||||
5881 | << " 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) | ||||||||
5882 | 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) | ||||||||
5883 | 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) | ||||||||
5884 | << 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) | ||||||||
5885 | << " 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) | ||||||||
5886 | }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) | ||||||||
5887 | })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); | ||||||||
5888 | |||||||||
5889 | RU.LoopInvariantRegs = Invariant; | ||||||||
5890 | RU.MaxLocalUsers = MaxUsages[i]; | ||||||||
5891 | RUs[i] = RU; | ||||||||
5892 | } | ||||||||
5893 | |||||||||
5894 | return RUs; | ||||||||
5895 | } | ||||||||
5896 | |||||||||
5897 | bool LoopVectorizationCostModel::useEmulatedMaskMemRefHack(Instruction *I){ | ||||||||
5898 | // TODO: Cost model for emulated masked load/store is completely | ||||||||
5899 | // broken. This hack guides the cost model to use an artificially | ||||||||
5900 | // high enough value to practically disable vectorization with such | ||||||||
5901 | // operations, except where previously deployed legality hack allowed | ||||||||
5902 | // using very low cost values. This is to avoid regressions coming simply | ||||||||
5903 | // from moving "masked load/store" check from legality to cost model. | ||||||||
5904 | // Masked Load/Gather emulation was previously never allowed. | ||||||||
5905 | // Limited number of Masked Store/Scatter emulation was allowed. | ||||||||
5906 | assert(isPredicatedInst(I) && "Expecting a scalar emulated instruction")((isPredicatedInst(I) && "Expecting a scalar emulated instruction" ) ? static_cast<void> (0) : __assert_fail ("isPredicatedInst(I) && \"Expecting a scalar emulated instruction\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 5906, __PRETTY_FUNCTION__)); | ||||||||
5907 | return isa<LoadInst>(I) || | ||||||||
5908 | (isa<StoreInst>(I) && | ||||||||
5909 | NumPredStores > NumberOfStoresToPredicate); | ||||||||
5910 | } | ||||||||
5911 | |||||||||
5912 | void LoopVectorizationCostModel::collectInstsToScalarize(ElementCount VF) { | ||||||||
5913 | // If we aren't vectorizing the loop, or if we've already collected the | ||||||||
5914 | // instructions to scalarize, there's nothing to do. Collection may already | ||||||||
5915 | // have occurred if we have a user-selected VF and are now computing the | ||||||||
5916 | // expected cost for interleaving. | ||||||||
5917 | if (VF.isScalar() || VF.isZero() || | ||||||||
5918 | InstsToScalarize.find(VF) != InstsToScalarize.end()) | ||||||||
5919 | return; | ||||||||
5920 | |||||||||
5921 | // Initialize a mapping for VF in InstsToScalalarize. If we find that it's | ||||||||
5922 | // not profitable to scalarize any instructions, the presence of VF in the | ||||||||
5923 | // map will indicate that we've analyzed it already. | ||||||||
5924 | ScalarCostsTy &ScalarCostsVF = InstsToScalarize[VF]; | ||||||||
5925 | |||||||||
5926 | // Find all the instructions that are scalar with predication in the loop and | ||||||||
5927 | // determine if it would be better to not if-convert the blocks they are in. | ||||||||
5928 | // If so, we also record the instructions to scalarize. | ||||||||
5929 | for (BasicBlock *BB : TheLoop->blocks()) { | ||||||||
5930 | if (!blockNeedsPredication(BB)) | ||||||||
5931 | continue; | ||||||||
5932 | for (Instruction &I : *BB) | ||||||||
5933 | if (isScalarWithPredication(&I)) { | ||||||||
5934 | ScalarCostsTy ScalarCosts; | ||||||||
5935 | // Do not apply discount logic if hacked cost is needed | ||||||||
5936 | // for emulated masked memrefs. | ||||||||
5937 | if (!useEmulatedMaskMemRefHack(&I) && | ||||||||
5938 | computePredInstDiscount(&I, ScalarCosts, VF) >= 0) | ||||||||
5939 | ScalarCostsVF.insert(ScalarCosts.begin(), ScalarCosts.end()); | ||||||||
5940 | // Remember that BB will remain after vectorization. | ||||||||
5941 | PredicatedBBsAfterVectorization.insert(BB); | ||||||||
5942 | } | ||||||||
5943 | } | ||||||||
5944 | } | ||||||||
5945 | |||||||||
5946 | int LoopVectorizationCostModel::computePredInstDiscount( | ||||||||
5947 | Instruction *PredInst, DenseMap<Instruction *, unsigned> &ScalarCosts, | ||||||||
5948 | ElementCount VF) { | ||||||||
5949 | assert(!isUniformAfterVectorization(PredInst, VF) &&((!isUniformAfterVectorization(PredInst, VF) && "Instruction marked uniform-after-vectorization will be predicated" ) ? static_cast<void> (0) : __assert_fail ("!isUniformAfterVectorization(PredInst, VF) && \"Instruction marked uniform-after-vectorization will be predicated\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 5950, __PRETTY_FUNCTION__)) | ||||||||
5950 | "Instruction marked uniform-after-vectorization will be predicated")((!isUniformAfterVectorization(PredInst, VF) && "Instruction marked uniform-after-vectorization will be predicated" ) ? static_cast<void> (0) : __assert_fail ("!isUniformAfterVectorization(PredInst, VF) && \"Instruction marked uniform-after-vectorization will be predicated\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 5950, __PRETTY_FUNCTION__)); | ||||||||
5951 | |||||||||
5952 | // Initialize the discount to zero, meaning that the scalar version and the | ||||||||
5953 | // vector version cost the same. | ||||||||
5954 | int Discount = 0; | ||||||||
5955 | |||||||||
5956 | // Holds instructions to analyze. The instructions we visit are mapped in | ||||||||
5957 | // ScalarCosts. Those instructions are the ones that would be scalarized if | ||||||||
5958 | // we find that the scalar version costs less. | ||||||||
5959 | SmallVector<Instruction *, 8> Worklist; | ||||||||
5960 | |||||||||
5961 | // Returns true if the given instruction can be scalarized. | ||||||||
5962 | auto canBeScalarized = [&](Instruction *I) -> bool { | ||||||||
5963 | // We only attempt to scalarize instructions forming a single-use chain | ||||||||
5964 | // from the original predicated block that would otherwise be vectorized. | ||||||||
5965 | // Although not strictly necessary, we give up on instructions we know will | ||||||||
5966 | // already be scalar to avoid traversing chains that are unlikely to be | ||||||||
5967 | // beneficial. | ||||||||
5968 | if (!I->hasOneUse() || PredInst->getParent() != I->getParent() || | ||||||||
5969 | isScalarAfterVectorization(I, VF)) | ||||||||
5970 | return false; | ||||||||
5971 | |||||||||
5972 | // If the instruction is scalar with predication, it will be analyzed | ||||||||
5973 | // separately. We ignore it within the context of PredInst. | ||||||||
5974 | if (isScalarWithPredication(I)) | ||||||||
5975 | return false; | ||||||||
5976 | |||||||||
5977 | // If any of the instruction's operands are uniform after vectorization, | ||||||||
5978 | // the instruction cannot be scalarized. This prevents, for example, a | ||||||||
5979 | // masked load from being scalarized. | ||||||||
5980 | // | ||||||||
5981 | // We assume we will only emit a value for lane zero of an instruction | ||||||||
5982 | // marked uniform after vectorization, rather than VF identical values. | ||||||||
5983 | // Thus, if we scalarize an instruction that uses a uniform, we would | ||||||||
5984 | // create uses of values corresponding to the lanes we aren't emitting code | ||||||||
5985 | // for. This behavior can be changed by allowing getScalarValue to clone | ||||||||
5986 | // the lane zero values for uniforms rather than asserting. | ||||||||
5987 | for (Use &U : I->operands()) | ||||||||
5988 | if (auto *J = dyn_cast<Instruction>(U.get())) | ||||||||
5989 | if (isUniformAfterVectorization(J, VF)) | ||||||||
5990 | return false; | ||||||||
5991 | |||||||||
5992 | // Otherwise, we can scalarize the instruction. | ||||||||
5993 | return true; | ||||||||
5994 | }; | ||||||||
5995 | |||||||||
5996 | // Compute the expected cost discount from scalarizing the entire expression | ||||||||
5997 | // feeding the predicated instruction. We currently only consider expressions | ||||||||
5998 | // that are single-use instruction chains. | ||||||||
5999 | Worklist.push_back(PredInst); | ||||||||
6000 | while (!Worklist.empty()) { | ||||||||
6001 | Instruction *I = Worklist.pop_back_val(); | ||||||||
6002 | |||||||||
6003 | // If we've already analyzed the instruction, there's nothing to do. | ||||||||
6004 | if (ScalarCosts.find(I) != ScalarCosts.end()) | ||||||||
6005 | continue; | ||||||||
6006 | |||||||||
6007 | // Compute the cost of the vector instruction. Note that this cost already | ||||||||
6008 | // includes the scalarization overhead of the predicated instruction. | ||||||||
6009 | unsigned VectorCost = getInstructionCost(I, VF).first; | ||||||||
6010 | |||||||||
6011 | // Compute the cost of the scalarized instruction. This cost is the cost of | ||||||||
6012 | // the instruction as if it wasn't if-converted and instead remained in the | ||||||||
6013 | // predicated block. We will scale this cost by block probability after | ||||||||
6014 | // computing the scalarization overhead. | ||||||||
6015 | assert(!VF.isScalable() && "scalable vectors not yet supported.")((!VF.isScalable() && "scalable vectors not yet supported." ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"scalable vectors not yet supported.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 6015, __PRETTY_FUNCTION__)); | ||||||||
6016 | unsigned ScalarCost = | ||||||||
6017 | VF.getKnownMinValue() * | ||||||||
6018 | getInstructionCost(I, ElementCount::getFixed(1)).first; | ||||||||
6019 | |||||||||
6020 | // Compute the scalarization overhead of needed insertelement instructions | ||||||||
6021 | // and phi nodes. | ||||||||
6022 | if (isScalarWithPredication(I) && !I->getType()->isVoidTy()) { | ||||||||
6023 | ScalarCost += TTI.getScalarizationOverhead( | ||||||||
6024 | cast<VectorType>(ToVectorTy(I->getType(), VF)), | ||||||||
6025 | APInt::getAllOnesValue(VF.getKnownMinValue()), true, false); | ||||||||
6026 | assert(!VF.isScalable() && "scalable vectors not yet supported.")((!VF.isScalable() && "scalable vectors not yet supported." ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"scalable vectors not yet supported.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 6026, __PRETTY_FUNCTION__)); | ||||||||
6027 | ScalarCost += | ||||||||
6028 | VF.getKnownMinValue() * | ||||||||
6029 | TTI.getCFInstrCost(Instruction::PHI, TTI::TCK_RecipThroughput); | ||||||||
6030 | } | ||||||||
6031 | |||||||||
6032 | // Compute the scalarization overhead of needed extractelement | ||||||||
6033 | // instructions. For each of the instruction's operands, if the operand can | ||||||||
6034 | // be scalarized, add it to the worklist; otherwise, account for the | ||||||||
6035 | // overhead. | ||||||||
6036 | for (Use &U : I->operands()) | ||||||||
6037 | if (auto *J = dyn_cast<Instruction>(U.get())) { | ||||||||
6038 | assert(VectorType::isValidElementType(J->getType()) &&((VectorType::isValidElementType(J->getType()) && "Instruction has non-scalar type" ) ? static_cast<void> (0) : __assert_fail ("VectorType::isValidElementType(J->getType()) && \"Instruction has non-scalar type\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 6039, __PRETTY_FUNCTION__)) | ||||||||
6039 | "Instruction has non-scalar type")((VectorType::isValidElementType(J->getType()) && "Instruction has non-scalar type" ) ? static_cast<void> (0) : __assert_fail ("VectorType::isValidElementType(J->getType()) && \"Instruction has non-scalar type\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 6039, __PRETTY_FUNCTION__)); | ||||||||
6040 | if (canBeScalarized(J)) | ||||||||
6041 | Worklist.push_back(J); | ||||||||
6042 | else if (needsExtract(J, VF)) { | ||||||||
6043 | assert(!VF.isScalable() && "scalable vectors not yet supported.")((!VF.isScalable() && "scalable vectors not yet supported." ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"scalable vectors not yet supported.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 6043, __PRETTY_FUNCTION__)); | ||||||||
6044 | ScalarCost += TTI.getScalarizationOverhead( | ||||||||
6045 | cast<VectorType>(ToVectorTy(J->getType(), VF)), | ||||||||
6046 | APInt::getAllOnesValue(VF.getKnownMinValue()), false, true); | ||||||||
6047 | } | ||||||||
6048 | } | ||||||||
6049 | |||||||||
6050 | // Scale the total scalar cost by block probability. | ||||||||
6051 | ScalarCost /= getReciprocalPredBlockProb(); | ||||||||
6052 | |||||||||
6053 | // Compute the discount. A non-negative discount means the vector version | ||||||||
6054 | // of the instruction costs more, and scalarizing would be beneficial. | ||||||||
6055 | Discount += VectorCost - ScalarCost; | ||||||||
6056 | ScalarCosts[I] = ScalarCost; | ||||||||
6057 | } | ||||||||
6058 | |||||||||
6059 | return Discount; | ||||||||
6060 | } | ||||||||
6061 | |||||||||
6062 | LoopVectorizationCostModel::VectorizationCostTy | ||||||||
6063 | LoopVectorizationCostModel::expectedCost(ElementCount VF) { | ||||||||
6064 | assert(!VF.isScalable() && "scalable vectors not yet supported.")((!VF.isScalable() && "scalable vectors not yet supported." ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"scalable vectors not yet supported.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 6064, __PRETTY_FUNCTION__)); | ||||||||
6065 | VectorizationCostTy Cost; | ||||||||
6066 | |||||||||
6067 | // For each block. | ||||||||
6068 | for (BasicBlock *BB : TheLoop->blocks()) { | ||||||||
6069 | VectorizationCostTy BlockCost; | ||||||||
6070 | |||||||||
6071 | // For each instruction in the old loop. | ||||||||
6072 | for (Instruction &I : BB->instructionsWithoutDebug()) { | ||||||||
6073 | // Skip ignored values. | ||||||||
6074 | if (ValuesToIgnore.count(&I) || | ||||||||
6075 | (VF.isVector() && VecValuesToIgnore.count(&I))) | ||||||||
6076 | continue; | ||||||||
6077 | |||||||||
6078 | VectorizationCostTy C = getInstructionCost(&I, VF); | ||||||||
6079 | |||||||||
6080 | // Check if we should override the cost. | ||||||||
6081 | if (ForceTargetInstructionCost.getNumOccurrences() > 0) | ||||||||
6082 | C.first = ForceTargetInstructionCost; | ||||||||
6083 | |||||||||
6084 | BlockCost.first += C.first; | ||||||||
6085 | BlockCost.second |= C.second; | ||||||||
6086 | 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) | ||||||||
6087 | << " 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) | ||||||||
6088 | << '\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); | ||||||||
6089 | } | ||||||||
6090 | |||||||||
6091 | // If we are vectorizing a predicated block, it will have been | ||||||||
6092 | // if-converted. This means that the block's instructions (aside from | ||||||||
6093 | // stores and instructions that may divide by zero) will now be | ||||||||
6094 | // unconditionally executed. For the scalar case, we may not always execute | ||||||||
6095 | // the predicated block. Thus, scale the block's cost by the probability of | ||||||||
6096 | // executing it. | ||||||||
6097 | if (VF.isScalar() && blockNeedsPredication(BB)) | ||||||||
6098 | BlockCost.first /= getReciprocalPredBlockProb(); | ||||||||
6099 | |||||||||
6100 | Cost.first += BlockCost.first; | ||||||||
6101 | Cost.second |= BlockCost.second; | ||||||||
6102 | } | ||||||||
6103 | |||||||||
6104 | return Cost; | ||||||||
6105 | } | ||||||||
6106 | |||||||||
6107 | /// Gets Address Access SCEV after verifying that the access pattern | ||||||||
6108 | /// is loop invariant except the induction variable dependence. | ||||||||
6109 | /// | ||||||||
6110 | /// This SCEV can be sent to the Target in order to estimate the address | ||||||||
6111 | /// calculation cost. | ||||||||
6112 | static const SCEV *getAddressAccessSCEV( | ||||||||
6113 | Value *Ptr, | ||||||||
6114 | LoopVectorizationLegality *Legal, | ||||||||
6115 | PredicatedScalarEvolution &PSE, | ||||||||
6116 | const Loop *TheLoop) { | ||||||||
6117 | |||||||||
6118 | auto *Gep = dyn_cast<GetElementPtrInst>(Ptr); | ||||||||
6119 | if (!Gep) | ||||||||
6120 | return nullptr; | ||||||||
6121 | |||||||||
6122 | // We are looking for a gep with all loop invariant indices except for one | ||||||||
6123 | // which should be an induction variable. | ||||||||
6124 | auto SE = PSE.getSE(); | ||||||||
6125 | unsigned NumOperands = Gep->getNumOperands(); | ||||||||
6126 | for (unsigned i = 1; i < NumOperands; ++i) { | ||||||||
6127 | Value *Opd = Gep->getOperand(i); | ||||||||
6128 | if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) && | ||||||||
6129 | !Legal->isInductionVariable(Opd)) | ||||||||
6130 | return nullptr; | ||||||||
6131 | } | ||||||||
6132 | |||||||||
6133 | // Now we know we have a GEP ptr, %inv, %ind, %inv. return the Ptr SCEV. | ||||||||
6134 | return PSE.getSCEV(Ptr); | ||||||||
6135 | } | ||||||||
6136 | |||||||||
6137 | static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) { | ||||||||
6138 | return Legal->hasStride(I->getOperand(0)) || | ||||||||
6139 | Legal->hasStride(I->getOperand(1)); | ||||||||
6140 | } | ||||||||
6141 | |||||||||
6142 | unsigned | ||||||||
6143 | LoopVectorizationCostModel::getMemInstScalarizationCost(Instruction *I, | ||||||||
6144 | ElementCount VF) { | ||||||||
6145 | assert(VF.isVector() &&((VF.isVector() && "Scalarization cost of instruction implies vectorization." ) ? static_cast<void> (0) : __assert_fail ("VF.isVector() && \"Scalarization cost of instruction implies vectorization.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 6146, __PRETTY_FUNCTION__)) | ||||||||
6146 | "Scalarization cost of instruction implies vectorization.")((VF.isVector() && "Scalarization cost of instruction implies vectorization." ) ? static_cast<void> (0) : __assert_fail ("VF.isVector() && \"Scalarization cost of instruction implies vectorization.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 6146, __PRETTY_FUNCTION__)); | ||||||||
6147 | assert(!VF.isScalable() && "scalable vectors not yet supported.")((!VF.isScalable() && "scalable vectors not yet supported." ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"scalable vectors not yet supported.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 6147, __PRETTY_FUNCTION__)); | ||||||||
6148 | Type *ValTy = getMemInstValueType(I); | ||||||||
6149 | auto SE = PSE.getSE(); | ||||||||
6150 | |||||||||
6151 | unsigned AS = getLoadStoreAddressSpace(I); | ||||||||
6152 | Value *Ptr = getLoadStorePointerOperand(I); | ||||||||
6153 | Type *PtrTy = ToVectorTy(Ptr->getType(), VF); | ||||||||
6154 | |||||||||
6155 | // Figure out whether the access is strided and get the stride value | ||||||||
6156 | // if it's known in compile time | ||||||||
6157 | const SCEV *PtrSCEV = getAddressAccessSCEV(Ptr, Legal, PSE, TheLoop); | ||||||||
6158 | |||||||||
6159 | // Get the cost of the scalar memory instruction and address computation. | ||||||||
6160 | unsigned Cost = | ||||||||
6161 | VF.getKnownMinValue() * TTI.getAddressComputationCost(PtrTy, SE, PtrSCEV); | ||||||||
6162 | |||||||||
6163 | // Don't pass *I here, since it is scalar but will actually be part of a | ||||||||
6164 | // vectorized loop where the user of it is a vectorized instruction. | ||||||||
6165 | const Align Alignment = getLoadStoreAlignment(I); | ||||||||
6166 | Cost += VF.getKnownMinValue() * | ||||||||
6167 | TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(), Alignment, | ||||||||
6168 | AS, TTI::TCK_RecipThroughput); | ||||||||
6169 | |||||||||
6170 | // Get the overhead of the extractelement and insertelement instructions | ||||||||
6171 | // we might create due to scalarization. | ||||||||
6172 | Cost += getScalarizationOverhead(I, VF); | ||||||||
6173 | |||||||||
6174 | // If we have a predicated store, it may not be executed for each vector | ||||||||
6175 | // lane. Scale the cost by the probability of executing the predicated | ||||||||
6176 | // block. | ||||||||
6177 | if (isPredicatedInst(I)) { | ||||||||
6178 | Cost /= getReciprocalPredBlockProb(); | ||||||||
6179 | |||||||||
6180 | if (useEmulatedMaskMemRefHack(I)) | ||||||||
6181 | // Artificially setting to a high enough value to practically disable | ||||||||
6182 | // vectorization with such operations. | ||||||||
6183 | Cost = 3000000; | ||||||||
6184 | } | ||||||||
6185 | |||||||||
6186 | return Cost; | ||||||||
6187 | } | ||||||||
6188 | |||||||||
6189 | unsigned LoopVectorizationCostModel::getConsecutiveMemOpCost(Instruction *I, | ||||||||
6190 | ElementCount VF) { | ||||||||
6191 | Type *ValTy = getMemInstValueType(I); | ||||||||
6192 | auto *VectorTy = cast<VectorType>(ToVectorTy(ValTy, VF)); | ||||||||
6193 | Value *Ptr = getLoadStorePointerOperand(I); | ||||||||
6194 | unsigned AS = getLoadStoreAddressSpace(I); | ||||||||
6195 | int ConsecutiveStride = Legal->isConsecutivePtr(Ptr); | ||||||||
6196 | enum TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput; | ||||||||
6197 | |||||||||
6198 | assert((ConsecutiveStride == 1 || ConsecutiveStride == -1) &&(((ConsecutiveStride == 1 || ConsecutiveStride == -1) && "Stride should be 1 or -1 for consecutive memory access") ? static_cast <void> (0) : __assert_fail ("(ConsecutiveStride == 1 || ConsecutiveStride == -1) && \"Stride should be 1 or -1 for consecutive memory access\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 6199, __PRETTY_FUNCTION__)) | ||||||||
6199 | "Stride should be 1 or -1 for consecutive memory access")(((ConsecutiveStride == 1 || ConsecutiveStride == -1) && "Stride should be 1 or -1 for consecutive memory access") ? static_cast <void> (0) : __assert_fail ("(ConsecutiveStride == 1 || ConsecutiveStride == -1) && \"Stride should be 1 or -1 for consecutive memory access\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 6199, __PRETTY_FUNCTION__)); | ||||||||
6200 | const Align Alignment = getLoadStoreAlignment(I); | ||||||||
6201 | unsigned Cost = 0; | ||||||||
6202 | if (Legal->isMaskRequired(I)) | ||||||||
6203 | Cost += TTI.getMaskedMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS, | ||||||||
6204 | CostKind); | ||||||||
6205 | else | ||||||||
6206 | Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS, | ||||||||
6207 | CostKind, I); | ||||||||
6208 | |||||||||
6209 | bool Reverse = ConsecutiveStride < 0; | ||||||||
6210 | if (Reverse) | ||||||||
6211 | Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, VectorTy, 0); | ||||||||
6212 | return Cost; | ||||||||
6213 | } | ||||||||
6214 | |||||||||
6215 | unsigned LoopVectorizationCostModel::getUniformMemOpCost(Instruction *I, | ||||||||
6216 | ElementCount VF) { | ||||||||
6217 | Type *ValTy = getMemInstValueType(I); | ||||||||
6218 | auto *VectorTy = cast<VectorType>(ToVectorTy(ValTy, VF)); | ||||||||
6219 | const Align Alignment = getLoadStoreAlignment(I); | ||||||||
6220 | unsigned AS = getLoadStoreAddressSpace(I); | ||||||||
6221 | enum TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput; | ||||||||
6222 | if (isa<LoadInst>(I)) { | ||||||||
6223 | return TTI.getAddressComputationCost(ValTy) + | ||||||||
6224 | TTI.getMemoryOpCost(Instruction::Load, ValTy, Alignment, AS, | ||||||||
6225 | CostKind) + | ||||||||
6226 | TTI.getShuffleCost(TargetTransformInfo::SK_Broadcast, VectorTy); | ||||||||
6227 | } | ||||||||
6228 | StoreInst *SI = cast<StoreInst>(I); | ||||||||
6229 | |||||||||
6230 | bool isLoopInvariantStoreValue = Legal->isUniform(SI->getValueOperand()); | ||||||||
6231 | return TTI.getAddressComputationCost(ValTy) + | ||||||||
6232 | TTI.getMemoryOpCost(Instruction::Store, ValTy, Alignment, AS, | ||||||||
6233 | CostKind) + | ||||||||
6234 | (isLoopInvariantStoreValue | ||||||||
6235 | ? 0 | ||||||||
6236 | : TTI.getVectorInstrCost(Instruction::ExtractElement, VectorTy, | ||||||||
6237 | VF.getKnownMinValue() - 1)); | ||||||||
6238 | } | ||||||||
6239 | |||||||||
6240 | unsigned LoopVectorizationCostModel::getGatherScatterCost(Instruction *I, | ||||||||
6241 | ElementCount VF) { | ||||||||
6242 | Type *ValTy = getMemInstValueType(I); | ||||||||
6243 | auto *VectorTy = cast<VectorType>(ToVectorTy(ValTy, VF)); | ||||||||
6244 | const Align Alignment = getLoadStoreAlignment(I); | ||||||||
6245 | const Value *Ptr = getLoadStorePointerOperand(I); | ||||||||
6246 | |||||||||
6247 | return TTI.getAddressComputationCost(VectorTy) + | ||||||||
6248 | TTI.getGatherScatterOpCost( | ||||||||
6249 | I->getOpcode(), VectorTy, Ptr, Legal->isMaskRequired(I), Alignment, | ||||||||
6250 | TargetTransformInfo::TCK_RecipThroughput, I); | ||||||||
6251 | } | ||||||||
6252 | |||||||||
6253 | unsigned LoopVectorizationCostModel::getInterleaveGroupCost(Instruction *I, | ||||||||
6254 | ElementCount VF) { | ||||||||
6255 | Type *ValTy = getMemInstValueType(I); | ||||||||
6256 | auto *VectorTy = cast<VectorType>(ToVectorTy(ValTy, VF)); | ||||||||
6257 | unsigned AS = getLoadStoreAddressSpace(I); | ||||||||
6258 | |||||||||
6259 | auto Group = getInterleavedAccessGroup(I); | ||||||||
6260 | assert(Group && "Fail to get an interleaved access group.")((Group && "Fail to get an interleaved access group." ) ? static_cast<void> (0) : __assert_fail ("Group && \"Fail to get an interleaved access group.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 6260, __PRETTY_FUNCTION__)); | ||||||||
6261 | |||||||||
6262 | unsigned InterleaveFactor = Group->getFactor(); | ||||||||
6263 | assert(!VF.isScalable() && "scalable vectors not yet supported.")((!VF.isScalable() && "scalable vectors not yet supported." ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"scalable vectors not yet supported.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 6263, __PRETTY_FUNCTION__)); | ||||||||
6264 | auto *WideVecTy = VectorType::get(ValTy, VF * InterleaveFactor); | ||||||||
6265 | |||||||||
6266 | // Holds the indices of existing members in an interleaved load group. | ||||||||
6267 | // An interleaved store group doesn't need this as it doesn't allow gaps. | ||||||||
6268 | SmallVector<unsigned, 4> Indices; | ||||||||
6269 | if (isa<LoadInst>(I)) { | ||||||||
6270 | for (unsigned i = 0; i < InterleaveFactor; i++) | ||||||||
6271 | if (Group->getMember(i)) | ||||||||
6272 | Indices.push_back(i); | ||||||||
6273 | } | ||||||||
6274 | |||||||||
6275 | // Calculate the cost of the whole interleaved group. | ||||||||
6276 | bool UseMaskForGaps = | ||||||||
6277 | Group->requiresScalarEpilogue() && !isScalarEpilogueAllowed(); | ||||||||
6278 | unsigned Cost = TTI.getInterleavedMemoryOpCost( | ||||||||
6279 | I->getOpcode(), WideVecTy, Group->getFactor(), Indices, Group->getAlign(), | ||||||||
6280 | AS, TTI::TCK_RecipThroughput, Legal->isMaskRequired(I), UseMaskForGaps); | ||||||||
6281 | |||||||||
6282 | if (Group->isReverse()) { | ||||||||
6283 | // TODO: Add support for reversed masked interleaved access. | ||||||||
6284 | assert(!Legal->isMaskRequired(I) &&((!Legal->isMaskRequired(I) && "Reverse masked interleaved access not supported." ) ? static_cast<void> (0) : __assert_fail ("!Legal->isMaskRequired(I) && \"Reverse masked interleaved access not supported.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 6285, __PRETTY_FUNCTION__)) | ||||||||
6285 | "Reverse masked interleaved access not supported.")((!Legal->isMaskRequired(I) && "Reverse masked interleaved access not supported." ) ? static_cast<void> (0) : __assert_fail ("!Legal->isMaskRequired(I) && \"Reverse masked interleaved access not supported.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 6285, __PRETTY_FUNCTION__)); | ||||||||
6286 | Cost += Group->getNumMembers() * | ||||||||
6287 | TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, VectorTy, 0); | ||||||||
6288 | } | ||||||||
6289 | return Cost; | ||||||||
6290 | } | ||||||||
6291 | |||||||||
6292 | unsigned LoopVectorizationCostModel::getMemoryInstructionCost(Instruction *I, | ||||||||
6293 | ElementCount VF) { | ||||||||
6294 | // Calculate scalar cost only. Vectorization cost should be ready at this | ||||||||
6295 | // moment. | ||||||||
6296 | if (VF.isScalar()) { | ||||||||
6297 | Type *ValTy = getMemInstValueType(I); | ||||||||
6298 | const Align Alignment = getLoadStoreAlignment(I); | ||||||||
6299 | unsigned AS = getLoadStoreAddressSpace(I); | ||||||||
6300 | |||||||||
6301 | return TTI.getAddressComputationCost(ValTy) + | ||||||||
6302 | TTI.getMemoryOpCost(I->getOpcode(), ValTy, Alignment, AS, | ||||||||
6303 | TTI::TCK_RecipThroughput, I); | ||||||||
6304 | } | ||||||||
6305 | return getWideningCost(I, VF); | ||||||||
6306 | } | ||||||||
6307 | |||||||||
6308 | LoopVectorizationCostModel::VectorizationCostTy | ||||||||
6309 | LoopVectorizationCostModel::getInstructionCost(Instruction *I, | ||||||||
6310 | ElementCount VF) { | ||||||||
6311 | assert(!VF.isScalable() &&((!VF.isScalable() && "the cost model is not yet implemented for scalable vectorization" ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"the cost model is not yet implemented for scalable vectorization\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 6312, __PRETTY_FUNCTION__)) | ||||||||
6312 | "the cost model is not yet implemented for scalable vectorization")((!VF.isScalable() && "the cost model is not yet implemented for scalable vectorization" ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"the cost model is not yet implemented for scalable vectorization\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 6312, __PRETTY_FUNCTION__)); | ||||||||
6313 | // If we know that this instruction will remain uniform, check the cost of | ||||||||
6314 | // the scalar version. | ||||||||
6315 | if (isUniformAfterVectorization(I, VF)) | ||||||||
6316 | VF = ElementCount::getFixed(1); | ||||||||
6317 | |||||||||
6318 | if (VF.isVector() && isProfitableToScalarize(I, VF)) | ||||||||
6319 | return VectorizationCostTy(InstsToScalarize[VF][I], false); | ||||||||
6320 | |||||||||
6321 | // Forced scalars do not have any scalarization overhead. | ||||||||
6322 | auto ForcedScalar = ForcedScalars.find(VF); | ||||||||
6323 | if (VF.isVector() && ForcedScalar != ForcedScalars.end()) { | ||||||||
6324 | auto InstSet = ForcedScalar->second; | ||||||||
6325 | if (InstSet.count(I)) | ||||||||
6326 | return VectorizationCostTy( | ||||||||
6327 | (getInstructionCost(I, ElementCount::getFixed(1)).first * | ||||||||
6328 | VF.getKnownMinValue()), | ||||||||
6329 | false); | ||||||||
6330 | } | ||||||||
6331 | |||||||||
6332 | Type *VectorTy; | ||||||||
6333 | unsigned C = getInstructionCost(I, VF, VectorTy); | ||||||||
6334 | |||||||||
6335 | bool TypeNotScalarized = | ||||||||
6336 | VF.isVector() && VectorTy->isVectorTy() && | ||||||||
6337 | TTI.getNumberOfParts(VectorTy) < VF.getKnownMinValue(); | ||||||||
6338 | return VectorizationCostTy(C, TypeNotScalarized); | ||||||||
6339 | } | ||||||||
6340 | |||||||||
6341 | unsigned LoopVectorizationCostModel::getScalarizationOverhead(Instruction *I, | ||||||||
6342 | ElementCount VF) { | ||||||||
6343 | |||||||||
6344 | assert(!VF.isScalable() &&((!VF.isScalable() && "cannot compute scalarization overhead for scalable vectorization" ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"cannot compute scalarization overhead for scalable vectorization\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 6345, __PRETTY_FUNCTION__)) | ||||||||
6345 | "cannot compute scalarization overhead for scalable vectorization")((!VF.isScalable() && "cannot compute scalarization overhead for scalable vectorization" ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"cannot compute scalarization overhead for scalable vectorization\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 6345, __PRETTY_FUNCTION__)); | ||||||||
6346 | if (VF.isScalar()) | ||||||||
6347 | return 0; | ||||||||
6348 | |||||||||
6349 | unsigned Cost = 0; | ||||||||
6350 | Type *RetTy = ToVectorTy(I->getType(), VF); | ||||||||
6351 | if (!RetTy->isVoidTy() && | ||||||||
6352 | (!isa<LoadInst>(I) || !TTI.supportsEfficientVectorElementLoadStore())) | ||||||||
6353 | Cost += TTI.getScalarizationOverhead( | ||||||||
6354 | cast<VectorType>(RetTy), APInt::getAllOnesValue(VF.getKnownMinValue()), | ||||||||
6355 | true, false); | ||||||||
6356 | |||||||||
6357 | // Some targets keep addresses scalar. | ||||||||
6358 | if (isa<LoadInst>(I) && !TTI.prefersVectorizedAddressing()) | ||||||||
6359 | return Cost; | ||||||||
6360 | |||||||||
6361 | // Some targets support efficient element stores. | ||||||||
6362 | if (isa<StoreInst>(I) && TTI.supportsEfficientVectorElementLoadStore()) | ||||||||
6363 | return Cost; | ||||||||
6364 | |||||||||
6365 | // Collect operands to consider. | ||||||||
6366 | CallInst *CI = dyn_cast<CallInst>(I); | ||||||||
6367 | Instruction::op_range Ops = CI ? CI->arg_operands() : I->operands(); | ||||||||
6368 | |||||||||
6369 | // Skip operands that do not require extraction/scalarization and do not incur | ||||||||
6370 | // any overhead. | ||||||||
6371 | return Cost + TTI.getOperandsScalarizationOverhead( | ||||||||
6372 | filterExtractingOperands(Ops, VF), VF.getKnownMinValue()); | ||||||||
6373 | } | ||||||||
6374 | |||||||||
6375 | void LoopVectorizationCostModel::setCostBasedWideningDecision(ElementCount VF) { | ||||||||
6376 | assert(!VF.isScalable() && "scalable vectors not yet supported.")((!VF.isScalable() && "scalable vectors not yet supported." ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"scalable vectors not yet supported.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 6376, __PRETTY_FUNCTION__)); | ||||||||
6377 | if (VF.isScalar()) | ||||||||
6378 | return; | ||||||||
6379 | NumPredStores = 0; | ||||||||
6380 | for (BasicBlock *BB : TheLoop->blocks()) { | ||||||||
6381 | // For each instruction in the old loop. | ||||||||
6382 | for (Instruction &I : *BB) { | ||||||||
6383 | Value *Ptr = getLoadStorePointerOperand(&I); | ||||||||
6384 | if (!Ptr) | ||||||||
6385 | continue; | ||||||||
6386 | |||||||||
6387 | // TODO: We should generate better code and update the cost model for | ||||||||
6388 | // predicated uniform stores. Today they are treated as any other | ||||||||
6389 | // predicated store (see added test cases in | ||||||||
6390 | // invariant-store-vectorization.ll). | ||||||||
6391 | if (isa<StoreInst>(&I) && isScalarWithPredication(&I)) | ||||||||
6392 | NumPredStores++; | ||||||||
6393 | |||||||||
6394 | if (Legal->isUniform(Ptr) && | ||||||||
6395 | // Conditional loads and stores should be scalarized and predicated. | ||||||||
6396 | // isScalarWithPredication cannot be used here since masked | ||||||||
6397 | // gather/scatters are not considered scalar with predication. | ||||||||
6398 | !Legal->blockNeedsPredication(I.getParent())) { | ||||||||
6399 | // TODO: Avoid replicating loads and stores instead of | ||||||||
6400 | // relying on instcombine to remove them. | ||||||||
6401 | // Load: Scalar load + broadcast | ||||||||
6402 | // Store: Scalar store + isLoopInvariantStoreValue ? 0 : extract | ||||||||
6403 | unsigned Cost = getUniformMemOpCost(&I, VF); | ||||||||
6404 | setWideningDecision(&I, VF, CM_Scalarize, Cost); | ||||||||
6405 | continue; | ||||||||
6406 | } | ||||||||
6407 | |||||||||
6408 | // We assume that widening is the best solution when possible. | ||||||||
6409 | if (memoryInstructionCanBeWidened(&I, VF)) { | ||||||||
6410 | unsigned Cost = getConsecutiveMemOpCost(&I, VF); | ||||||||
6411 | int ConsecutiveStride = | ||||||||
6412 | Legal->isConsecutivePtr(getLoadStorePointerOperand(&I)); | ||||||||
6413 | assert((ConsecutiveStride == 1 || ConsecutiveStride == -1) &&(((ConsecutiveStride == 1 || ConsecutiveStride == -1) && "Expected consecutive stride.") ? static_cast<void> (0 ) : __assert_fail ("(ConsecutiveStride == 1 || ConsecutiveStride == -1) && \"Expected consecutive stride.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 6414, __PRETTY_FUNCTION__)) | ||||||||
6414 | "Expected consecutive stride.")(((ConsecutiveStride == 1 || ConsecutiveStride == -1) && "Expected consecutive stride.") ? static_cast<void> (0 ) : __assert_fail ("(ConsecutiveStride == 1 || ConsecutiveStride == -1) && \"Expected consecutive stride.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 6414, __PRETTY_FUNCTION__)); | ||||||||
6415 | InstWidening Decision = | ||||||||
6416 | ConsecutiveStride == 1 ? CM_Widen : CM_Widen_Reverse; | ||||||||
6417 | setWideningDecision(&I, VF, Decision, Cost); | ||||||||
6418 | continue; | ||||||||
6419 | } | ||||||||
6420 | |||||||||
6421 | // Choose between Interleaving, Gather/Scatter or Scalarization. | ||||||||
6422 | unsigned InterleaveCost = std::numeric_limits<unsigned>::max(); | ||||||||
6423 | unsigned NumAccesses = 1; | ||||||||
6424 | if (isAccessInterleaved(&I)) { | ||||||||
6425 | auto Group = getInterleavedAccessGroup(&I); | ||||||||
6426 | assert(Group && "Fail to get an interleaved access group.")((Group && "Fail to get an interleaved access group." ) ? static_cast<void> (0) : __assert_fail ("Group && \"Fail to get an interleaved access group.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 6426, __PRETTY_FUNCTION__)); | ||||||||
6427 | |||||||||
6428 | // Make one decision for the whole group. | ||||||||
6429 | if (getWideningDecision(&I, VF) != CM_Unknown) | ||||||||
6430 | continue; | ||||||||
6431 | |||||||||
6432 | NumAccesses = Group->getNumMembers(); | ||||||||
6433 | if (interleavedAccessCanBeWidened(&I, VF)) | ||||||||
6434 | InterleaveCost = getInterleaveGroupCost(&I, VF); | ||||||||
6435 | } | ||||||||
6436 | |||||||||
6437 | unsigned GatherScatterCost = | ||||||||
6438 | isLegalGatherOrScatter(&I) | ||||||||
6439 | ? getGatherScatterCost(&I, VF) * NumAccesses | ||||||||
6440 | : std::numeric_limits<unsigned>::max(); | ||||||||
6441 | |||||||||
6442 | unsigned ScalarizationCost = | ||||||||
6443 | getMemInstScalarizationCost(&I, VF) * NumAccesses; | ||||||||
6444 | |||||||||
6445 | // Choose better solution for the current VF, | ||||||||
6446 | // write down this decision and use it during vectorization. | ||||||||
6447 | unsigned Cost; | ||||||||
6448 | InstWidening Decision; | ||||||||
6449 | if (InterleaveCost <= GatherScatterCost && | ||||||||
6450 | InterleaveCost < ScalarizationCost) { | ||||||||
6451 | Decision = CM_Interleave; | ||||||||
6452 | Cost = InterleaveCost; | ||||||||
6453 | } else if (GatherScatterCost < ScalarizationCost) { | ||||||||
6454 | Decision = CM_GatherScatter; | ||||||||
6455 | Cost = GatherScatterCost; | ||||||||
6456 | } else { | ||||||||
6457 | Decision = CM_Scalarize; | ||||||||
6458 | Cost = ScalarizationCost; | ||||||||
6459 | } | ||||||||
6460 | // If the instructions belongs to an interleave group, the whole group | ||||||||
6461 | // receives the same decision. The whole group receives the cost, but | ||||||||
6462 | // the cost will actually be assigned to one instruction. | ||||||||
6463 | if (auto Group = getInterleavedAccessGroup(&I)) | ||||||||
6464 | setWideningDecision(Group, VF, Decision, Cost); | ||||||||
6465 | else | ||||||||
6466 | setWideningDecision(&I, VF, Decision, Cost); | ||||||||
6467 | } | ||||||||
6468 | } | ||||||||
6469 | |||||||||
6470 | // Make sure that any load of address and any other address computation | ||||||||
6471 | // remains scalar unless there is gather/scatter support. This avoids | ||||||||
6472 | // inevitable extracts into address registers, and also has the benefit of | ||||||||
6473 | // activating LSR more, since that pass can't optimize vectorized | ||||||||
6474 | // addresses. | ||||||||
6475 | if (TTI.prefersVectorizedAddressing()) | ||||||||
6476 | return; | ||||||||
6477 | |||||||||
6478 | // Start with all scalar pointer uses. | ||||||||
6479 | SmallPtrSet<Instruction *, 8> AddrDefs; | ||||||||
6480 | for (BasicBlock *BB : TheLoop->blocks()) | ||||||||
6481 | for (Instruction &I : *BB) { | ||||||||
6482 | Instruction *PtrDef = | ||||||||
6483 | dyn_cast_or_null<Instruction>(getLoadStorePointerOperand(&I)); | ||||||||
6484 | if (PtrDef && TheLoop->contains(PtrDef) && | ||||||||
6485 | getWideningDecision(&I, VF) != CM_GatherScatter) | ||||||||
6486 | AddrDefs.insert(PtrDef); | ||||||||
6487 | } | ||||||||
6488 | |||||||||
6489 | // Add all instructions used to generate the addresses. | ||||||||
6490 | SmallVector<Instruction *, 4> Worklist; | ||||||||
6491 | for (auto *I : AddrDefs) | ||||||||
6492 | Worklist.push_back(I); | ||||||||
6493 | while (!Worklist.empty()) { | ||||||||
6494 | Instruction *I = Worklist.pop_back_val(); | ||||||||
6495 | for (auto &Op : I->operands()) | ||||||||
6496 | if (auto *InstOp = dyn_cast<Instruction>(Op)) | ||||||||
6497 | if ((InstOp->getParent() == I->getParent()) && !isa<PHINode>(InstOp) && | ||||||||
6498 | AddrDefs.insert(InstOp).second) | ||||||||
6499 | Worklist.push_back(InstOp); | ||||||||
6500 | } | ||||||||
6501 | |||||||||
6502 | for (auto *I : AddrDefs) { | ||||||||
6503 | if (isa<LoadInst>(I)) { | ||||||||
6504 | // Setting the desired widening decision should ideally be handled in | ||||||||
6505 | // by cost functions, but since this involves the task of finding out | ||||||||
6506 | // if the loaded register is involved in an address computation, it is | ||||||||
6507 | // instead changed here when we know this is the case. | ||||||||
6508 | InstWidening Decision = getWideningDecision(I, VF); | ||||||||
6509 | if (Decision == CM_Widen || Decision == CM_Widen_Reverse) | ||||||||
6510 | // Scalarize a widened load of address. | ||||||||
6511 | setWideningDecision( | ||||||||
6512 | I, VF, CM_Scalarize, | ||||||||
6513 | (VF.getKnownMinValue() * | ||||||||
6514 | getMemoryInstructionCost(I, ElementCount::getFixed(1)))); | ||||||||
6515 | else if (auto Group = getInterleavedAccessGroup(I)) { | ||||||||
6516 | // Scalarize an interleave group of address loads. | ||||||||
6517 | for (unsigned I = 0; I < Group->getFactor(); ++I) { | ||||||||
6518 | if (Instruction *Member = Group->getMember(I)) | ||||||||
6519 | setWideningDecision( | ||||||||
6520 | Member, VF, CM_Scalarize, | ||||||||
6521 | (VF.getKnownMinValue() * | ||||||||
6522 | getMemoryInstructionCost(Member, ElementCount::getFixed(1)))); | ||||||||
6523 | } | ||||||||
6524 | } | ||||||||
6525 | } else | ||||||||
6526 | // Make sure I gets scalarized and a cost estimate without | ||||||||
6527 | // scalarization overhead. | ||||||||
6528 | ForcedScalars[VF].insert(I); | ||||||||
6529 | } | ||||||||
6530 | } | ||||||||
6531 | |||||||||
6532 | unsigned LoopVectorizationCostModel::getInstructionCost(Instruction *I, | ||||||||
6533 | ElementCount VF, | ||||||||
6534 | Type *&VectorTy) { | ||||||||
6535 | Type *RetTy = I->getType(); | ||||||||
6536 | if (canTruncateToMinimalBitwidth(I, VF)) | ||||||||
6537 | RetTy = IntegerType::get(RetTy->getContext(), MinBWs[I]); | ||||||||
6538 | VectorTy = isScalarAfterVectorization(I, VF) ? RetTy : ToVectorTy(RetTy, VF); | ||||||||
6539 | auto SE = PSE.getSE(); | ||||||||
6540 | TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput; | ||||||||
6541 | |||||||||
6542 | // TODO: We need to estimate the cost of intrinsic calls. | ||||||||
6543 | switch (I->getOpcode()) { | ||||||||
6544 | case Instruction::GetElementPtr: | ||||||||
6545 | // We mark this instruction as zero-cost because the cost of GEPs in | ||||||||
6546 | // vectorized code depends on whether the corresponding memory instruction | ||||||||
6547 | // is scalarized or not. Therefore, we handle GEPs with the memory | ||||||||
6548 | // instruction cost. | ||||||||
6549 | return 0; | ||||||||
6550 | case Instruction::Br: { | ||||||||
6551 | // In cases of scalarized and predicated instructions, there will be VF | ||||||||
6552 | // predicated blocks in the vectorized loop. Each branch around these | ||||||||
6553 | // blocks requires also an extract of its vector compare i1 element. | ||||||||
6554 | bool ScalarPredicatedBB = false; | ||||||||
6555 | BranchInst *BI = cast<BranchInst>(I); | ||||||||
6556 | if (VF.isVector() && BI->isConditional() && | ||||||||
6557 | (PredicatedBBsAfterVectorization.count(BI->getSuccessor(0)) || | ||||||||
6558 | PredicatedBBsAfterVectorization.count(BI->getSuccessor(1)))) | ||||||||
6559 | ScalarPredicatedBB = true; | ||||||||
6560 | |||||||||
6561 | if (ScalarPredicatedBB) { | ||||||||
6562 | // Return cost for branches around scalarized and predicated blocks. | ||||||||
6563 | assert(!VF.isScalable() && "scalable vectors not yet supported.")((!VF.isScalable() && "scalable vectors not yet supported." ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"scalable vectors not yet supported.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 6563, __PRETTY_FUNCTION__)); | ||||||||
6564 | auto *Vec_i1Ty = | ||||||||
6565 | VectorType::get(IntegerType::getInt1Ty(RetTy->getContext()), VF); | ||||||||
6566 | return (TTI.getScalarizationOverhead( | ||||||||
6567 | Vec_i1Ty, APInt::getAllOnesValue(VF.getKnownMinValue()), | ||||||||
6568 | false, true) + | ||||||||
6569 | (TTI.getCFInstrCost(Instruction::Br, CostKind) * | ||||||||
6570 | VF.getKnownMinValue())); | ||||||||
6571 | } else if (I->getParent() == TheLoop->getLoopLatch() || VF.isScalar()) | ||||||||
6572 | // The back-edge branch will remain, as will all scalar branches. | ||||||||
6573 | return TTI.getCFInstrCost(Instruction::Br, CostKind); | ||||||||
6574 | else | ||||||||
6575 | // This branch will be eliminated by if-conversion. | ||||||||
6576 | return 0; | ||||||||
6577 | // Note: We currently assume zero cost for an unconditional branch inside | ||||||||
6578 | // a predicated block since it will become a fall-through, although we | ||||||||
6579 | // may decide in the future to call TTI for all branches. | ||||||||
6580 | } | ||||||||
6581 | case Instruction::PHI: { | ||||||||
6582 | auto *Phi = cast<PHINode>(I); | ||||||||
6583 | |||||||||
6584 | // First-order recurrences are replaced by vector shuffles inside the loop. | ||||||||
6585 | // NOTE: Don't use ToVectorTy as SK_ExtractSubvector expects a vector type. | ||||||||
6586 | if (VF.isVector() && Legal->isFirstOrderRecurrence(Phi)) | ||||||||
6587 | return TTI.getShuffleCost( | ||||||||
6588 | TargetTransformInfo::SK_ExtractSubvector, cast<VectorType>(VectorTy), | ||||||||
6589 | VF.getKnownMinValue() - 1, FixedVectorType::get(RetTy, 1)); | ||||||||
6590 | |||||||||
6591 | // Phi nodes in non-header blocks (not inductions, reductions, etc.) are | ||||||||
6592 | // converted into select instructions. We require N - 1 selects per phi | ||||||||
6593 | // node, where N is the number of incoming values. | ||||||||
6594 | if (VF.isVector() && Phi->getParent() != TheLoop->getHeader()) | ||||||||
6595 | return (Phi->getNumIncomingValues() - 1) * | ||||||||
6596 | TTI.getCmpSelInstrCost( | ||||||||
6597 | Instruction::Select, ToVectorTy(Phi->getType(), VF), | ||||||||
6598 | ToVectorTy(Type::getInt1Ty(Phi->getContext()), VF), | ||||||||
6599 | CostKind); | ||||||||
6600 | |||||||||
6601 | return TTI.getCFInstrCost(Instruction::PHI, CostKind); | ||||||||
6602 | } | ||||||||
6603 | case Instruction::UDiv: | ||||||||
6604 | case Instruction::SDiv: | ||||||||
6605 | case Instruction::URem: | ||||||||
6606 | case Instruction::SRem: | ||||||||
6607 | // If we have a predicated instruction, it may not be executed for each | ||||||||
6608 | // vector lane. Get the scalarization cost and scale this amount by the | ||||||||
6609 | // probability of executing the predicated block. If the instruction is not | ||||||||
6610 | // predicated, we fall through to the next case. | ||||||||
6611 | if (VF.isVector() && isScalarWithPredication(I)) { | ||||||||
6612 | unsigned Cost = 0; | ||||||||
6613 | |||||||||
6614 | // These instructions have a non-void type, so account for the phi nodes | ||||||||
6615 | // that we will create. This cost is likely to be zero. The phi node | ||||||||
6616 | // cost, if any, should be scaled by the block probability because it | ||||||||
6617 | // models a copy at the end of each predicated block. | ||||||||
6618 | Cost += VF.getKnownMinValue() * | ||||||||
6619 | TTI.getCFInstrCost(Instruction::PHI, CostKind); | ||||||||
6620 | |||||||||
6621 | // The cost of the non-predicated instruction. | ||||||||
6622 | Cost += VF.getKnownMinValue() * | ||||||||
6623 | TTI.getArithmeticInstrCost(I->getOpcode(), RetTy, CostKind); | ||||||||
6624 | |||||||||
6625 | // The cost of insertelement and extractelement instructions needed for | ||||||||
6626 | // scalarization. | ||||||||
6627 | Cost += getScalarizationOverhead(I, VF); | ||||||||
6628 | |||||||||
6629 | // Scale the cost by the probability of executing the predicated blocks. | ||||||||
6630 | // This assumes the predicated block for each vector lane is equally | ||||||||
6631 | // likely. | ||||||||
6632 | return Cost / getReciprocalPredBlockProb(); | ||||||||
6633 | } | ||||||||
6634 | LLVM_FALLTHROUGH[[gnu::fallthrough]]; | ||||||||
6635 | case Instruction::Add: | ||||||||
6636 | case Instruction::FAdd: | ||||||||
6637 | case Instruction::Sub: | ||||||||
6638 | case Instruction::FSub: | ||||||||
6639 | case Instruction::Mul: | ||||||||
6640 | case Instruction::FMul: | ||||||||
6641 | case Instruction::FDiv: | ||||||||
6642 | case Instruction::FRem: | ||||||||
6643 | case Instruction::Shl: | ||||||||
6644 | case Instruction::LShr: | ||||||||
6645 | case Instruction::AShr: | ||||||||
6646 | case Instruction::And: | ||||||||
6647 | case Instruction::Or: | ||||||||
6648 | case Instruction::Xor: { | ||||||||
6649 | // Since we will replace the stride by 1 the multiplication should go away. | ||||||||
6650 | if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal)) | ||||||||
6651 | return 0; | ||||||||
6652 | // Certain instructions can be cheaper to vectorize if they have a constant | ||||||||
6653 | // second vector operand. One example of this are shifts on x86. | ||||||||
6654 | Value *Op2 = I->getOperand(1); | ||||||||
6655 | TargetTransformInfo::OperandValueProperties Op2VP; | ||||||||
6656 | TargetTransformInfo::OperandValueKind Op2VK = | ||||||||
6657 | TTI.getOperandInfo(Op2, Op2VP); | ||||||||
6658 | if (Op2VK == TargetTransformInfo::OK_AnyValue && Legal->isUniform(Op2)) | ||||||||
6659 | Op2VK = TargetTransformInfo::OK_UniformValue; | ||||||||
6660 | |||||||||
6661 | SmallVector<const Value *, 4> Operands(I->operand_values()); | ||||||||
6662 | unsigned N = isScalarAfterVectorization(I, VF) ? VF.getKnownMinValue() : 1; | ||||||||
6663 | return N * TTI.getArithmeticInstrCost( | ||||||||
6664 | I->getOpcode(), VectorTy, CostKind, | ||||||||
6665 | TargetTransformInfo::OK_AnyValue, | ||||||||
6666 | Op2VK, TargetTransformInfo::OP_None, Op2VP, Operands, I); | ||||||||
6667 | } | ||||||||
6668 | case Instruction::FNeg: { | ||||||||
6669 | assert(!VF.isScalable() && "VF is assumed to be non scalable.")((!VF.isScalable() && "VF is assumed to be non scalable." ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"VF is assumed to be non scalable.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 6669, __PRETTY_FUNCTION__)); | ||||||||
6670 | unsigned N = isScalarAfterVectorization(I, VF) ? VF.getKnownMinValue() : 1; | ||||||||
6671 | return N * TTI.getArithmeticInstrCost( | ||||||||
6672 | I->getOpcode(), VectorTy, CostKind, | ||||||||
6673 | TargetTransformInfo::OK_AnyValue, | ||||||||
6674 | TargetTransformInfo::OK_AnyValue, | ||||||||
6675 | TargetTransformInfo::OP_None, TargetTransformInfo::OP_None, | ||||||||
6676 | I->getOperand(0), I); | ||||||||
6677 | } | ||||||||
6678 | case Instruction::Select: { | ||||||||
6679 | SelectInst *SI = cast<SelectInst>(I); | ||||||||
6680 | const SCEV *CondSCEV = SE->getSCEV(SI->getCondition()); | ||||||||
6681 | bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop)); | ||||||||
6682 | Type *CondTy = SI->getCondition()->getType(); | ||||||||
6683 | if (!ScalarCond) { | ||||||||
6684 | assert(!VF.isScalable() && "VF is assumed to be non scalable.")((!VF.isScalable() && "VF is assumed to be non scalable." ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"VF is assumed to be non scalable.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 6684, __PRETTY_FUNCTION__)); | ||||||||
6685 | CondTy = VectorType::get(CondTy, VF); | ||||||||
6686 | } | ||||||||
6687 | return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy, | ||||||||
6688 | CostKind, I); | ||||||||
6689 | } | ||||||||
6690 | case Instruction::ICmp: | ||||||||
6691 | case Instruction::FCmp: { | ||||||||
6692 | Type *ValTy = I->getOperand(0)->getType(); | ||||||||
6693 | Instruction *Op0AsInstruction = dyn_cast<Instruction>(I->getOperand(0)); | ||||||||
6694 | if (canTruncateToMinimalBitwidth(Op0AsInstruction, VF)) | ||||||||
6695 | ValTy = IntegerType::get(ValTy->getContext(), MinBWs[Op0AsInstruction]); | ||||||||
6696 | VectorTy = ToVectorTy(ValTy, VF); | ||||||||
6697 | return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, nullptr, CostKind, | ||||||||
6698 | I); | ||||||||
6699 | } | ||||||||
6700 | case Instruction::Store: | ||||||||
6701 | case Instruction::Load: { | ||||||||
6702 | ElementCount Width = VF; | ||||||||
6703 | if (Width.isVector()) { | ||||||||
6704 | InstWidening Decision = getWideningDecision(I, Width); | ||||||||
6705 | assert(Decision != CM_Unknown &&((Decision != CM_Unknown && "CM decision should be taken at this point" ) ? static_cast<void> (0) : __assert_fail ("Decision != CM_Unknown && \"CM decision should be taken at this point\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 6706, __PRETTY_FUNCTION__)) | ||||||||
6706 | "CM decision should be taken at this point")((Decision != CM_Unknown && "CM decision should be taken at this point" ) ? static_cast<void> (0) : __assert_fail ("Decision != CM_Unknown && \"CM decision should be taken at this point\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 6706, __PRETTY_FUNCTION__)); | ||||||||
6707 | if (Decision == CM_Scalarize) | ||||||||
6708 | Width = ElementCount::getFixed(1); | ||||||||
6709 | } | ||||||||
6710 | VectorTy = ToVectorTy(getMemInstValueType(I), Width); | ||||||||
6711 | return getMemoryInstructionCost(I, VF); | ||||||||
6712 | } | ||||||||
6713 | case Instruction::ZExt: | ||||||||
6714 | case Instruction::SExt: | ||||||||
6715 | case Instruction::FPToUI: | ||||||||
6716 | case Instruction::FPToSI: | ||||||||
6717 | case Instruction::FPExt: | ||||||||
6718 | case Instruction::PtrToInt: | ||||||||
6719 | case Instruction::IntToPtr: | ||||||||
6720 | case Instruction::SIToFP: | ||||||||
6721 | case Instruction::UIToFP: | ||||||||
6722 | case Instruction::Trunc: | ||||||||
6723 | case Instruction::FPTrunc: | ||||||||
6724 | case Instruction::BitCast: { | ||||||||
6725 | // Computes the CastContextHint from a Load/Store instruction. | ||||||||
6726 | auto ComputeCCH = [&](Instruction *I) -> TTI::CastContextHint { | ||||||||
6727 | assert((isa<LoadInst>(I) || isa<StoreInst>(I)) &&(((isa<LoadInst>(I) || isa<StoreInst>(I)) && "Expected a load or a store!") ? static_cast<void> (0) : __assert_fail ("(isa<LoadInst>(I) || isa<StoreInst>(I)) && \"Expected a load or a store!\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 6728, __PRETTY_FUNCTION__)) | ||||||||
6728 | "Expected a load or a store!")(((isa<LoadInst>(I) || isa<StoreInst>(I)) && "Expected a load or a store!") ? static_cast<void> (0) : __assert_fail ("(isa<LoadInst>(I) || isa<StoreInst>(I)) && \"Expected a load or a store!\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 6728, __PRETTY_FUNCTION__)); | ||||||||
6729 | |||||||||
6730 | if (VF.isScalar() || !TheLoop->contains(I)) | ||||||||
6731 | return TTI::CastContextHint::Normal; | ||||||||
6732 | |||||||||
6733 | switch (getWideningDecision(I, VF)) { | ||||||||
6734 | case LoopVectorizationCostModel::CM_GatherScatter: | ||||||||
6735 | return TTI::CastContextHint::GatherScatter; | ||||||||
6736 | case LoopVectorizationCostModel::CM_Interleave: | ||||||||
6737 | return TTI::CastContextHint::Interleave; | ||||||||
6738 | case LoopVectorizationCostModel::CM_Scalarize: | ||||||||
6739 | case LoopVectorizationCostModel::CM_Widen: | ||||||||
6740 | return Legal->isMaskRequired(I) ? TTI::CastContextHint::Masked | ||||||||
6741 | : TTI::CastContextHint::Normal; | ||||||||
6742 | case LoopVectorizationCostModel::CM_Widen_Reverse: | ||||||||
6743 | return TTI::CastContextHint::Reversed; | ||||||||
6744 | case LoopVectorizationCostModel::CM_Unknown: | ||||||||
6745 | llvm_unreachable("Instr did not go through cost modelling?")::llvm::llvm_unreachable_internal("Instr did not go through cost modelling?" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 6745); | ||||||||
6746 | } | ||||||||
6747 | |||||||||
6748 | llvm_unreachable("Unhandled case!")::llvm::llvm_unreachable_internal("Unhandled case!", "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 6748); | ||||||||
6749 | }; | ||||||||
6750 | |||||||||
6751 | unsigned Opcode = I->getOpcode(); | ||||||||
6752 | TTI::CastContextHint CCH = TTI::CastContextHint::None; | ||||||||
6753 | // For Trunc, the context is the only user, which must be a StoreInst. | ||||||||
6754 | if (Opcode == Instruction::Trunc || Opcode == Instruction::FPTrunc) { | ||||||||
6755 | if (I->hasOneUse()) | ||||||||
6756 | if (StoreInst *Store = dyn_cast<StoreInst>(*I->user_begin())) | ||||||||
6757 | CCH = ComputeCCH(Store); | ||||||||
6758 | } | ||||||||
6759 | // For Z/Sext, the context is the operand, which must be a LoadInst. | ||||||||
6760 | else if (Opcode == Instruction::ZExt || Opcode == Instruction::SExt || | ||||||||
6761 | Opcode == Instruction::FPExt) { | ||||||||
6762 | if (LoadInst *Load = dyn_cast<LoadInst>(I->getOperand(0))) | ||||||||
6763 | CCH = ComputeCCH(Load); | ||||||||
6764 | } | ||||||||
6765 | |||||||||
6766 | // We optimize the truncation of induction variables having constant | ||||||||
6767 | // integer steps. The cost of these truncations is the same as the scalar | ||||||||
6768 | // operation. | ||||||||
6769 | if (isOptimizableIVTruncate(I, VF)) { | ||||||||
6770 | auto *Trunc = cast<TruncInst>(I); | ||||||||
6771 | return TTI.getCastInstrCost(Instruction::Trunc, Trunc->getDestTy(), | ||||||||
6772 | Trunc->getSrcTy(), CCH, CostKind, Trunc); | ||||||||
6773 | } | ||||||||
6774 | |||||||||
6775 | Type *SrcScalarTy = I->getOperand(0)->getType(); | ||||||||
6776 | Type *SrcVecTy = | ||||||||
6777 | VectorTy->isVectorTy() ? ToVectorTy(SrcScalarTy, VF) : SrcScalarTy; | ||||||||
6778 | if (canTruncateToMinimalBitwidth(I, VF)) { | ||||||||
6779 | // This cast is going to be shrunk. This may remove the cast or it might | ||||||||
6780 | // turn it into slightly different cast. For example, if MinBW == 16, | ||||||||
6781 | // "zext i8 %1 to i32" becomes "zext i8 %1 to i16". | ||||||||
6782 | // | ||||||||
6783 | // Calculate the modified src and dest types. | ||||||||
6784 | Type *MinVecTy = VectorTy; | ||||||||
6785 | if (Opcode == Instruction::Trunc) { | ||||||||
6786 | SrcVecTy = smallestIntegerVectorType(SrcVecTy, MinVecTy); | ||||||||
6787 | VectorTy = | ||||||||
6788 | largestIntegerVectorType(ToVectorTy(I->getType(), VF), MinVecTy); | ||||||||
6789 | } else if (Opcode == Instruction::ZExt || Opcode == Instruction::SExt) { | ||||||||
6790 | SrcVecTy = largestIntegerVectorType(SrcVecTy, MinVecTy); | ||||||||
6791 | VectorTy = | ||||||||
6792 | smallestIntegerVectorType(ToVectorTy(I->getType(), VF), MinVecTy); | ||||||||
6793 | } | ||||||||
6794 | } | ||||||||
6795 | |||||||||
6796 | assert(!VF.isScalable() && "VF is assumed to be non scalable")((!VF.isScalable() && "VF is assumed to be non scalable" ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"VF is assumed to be non scalable\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 6796, __PRETTY_FUNCTION__)); | ||||||||
6797 | unsigned N = isScalarAfterVectorization(I, VF) ? VF.getKnownMinValue() : 1; | ||||||||
6798 | return N * | ||||||||
6799 | TTI.getCastInstrCost(Opcode, VectorTy, SrcVecTy, CCH, CostKind, I); | ||||||||
6800 | } | ||||||||
6801 | case Instruction::Call: { | ||||||||
6802 | bool NeedToScalarize; | ||||||||
6803 | CallInst *CI = cast<CallInst>(I); | ||||||||
6804 | unsigned CallCost = getVectorCallCost(CI, VF, NeedToScalarize); | ||||||||
6805 | if (getVectorIntrinsicIDForCall(CI, TLI)) | ||||||||
6806 | return std::min(CallCost, getVectorIntrinsicCost(CI, VF)); | ||||||||
6807 | return CallCost; | ||||||||
6808 | } | ||||||||
6809 | default: | ||||||||
6810 | // The cost of executing VF copies of the scalar instruction. This opcode | ||||||||
6811 | // is unknown. Assume that it is the same as 'mul'. | ||||||||
6812 | return VF.getKnownMinValue() * TTI.getArithmeticInstrCost( | ||||||||
6813 | Instruction::Mul, VectorTy, CostKind) + | ||||||||
6814 | getScalarizationOverhead(I, VF); | ||||||||
6815 | } // end of switch. | ||||||||
6816 | } | ||||||||
6817 | |||||||||
6818 | char LoopVectorize::ID = 0; | ||||||||
6819 | |||||||||
6820 | static const char lv_name[] = "Loop Vectorization"; | ||||||||
6821 | |||||||||
6822 | INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)static void *initializeLoopVectorizePassOnce(PassRegistry & Registry) { | ||||||||
6823 | INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)initializeTargetTransformInfoWrapperPassPass(Registry); | ||||||||
6824 | INITIALIZE_PASS_DEPENDENCY(BasicAAWrapperPass)initializeBasicAAWrapperPassPass(Registry); | ||||||||
6825 | INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)initializeAAResultsWrapperPassPass(Registry); | ||||||||
6826 | INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)initializeGlobalsAAWrapperPassPass(Registry); | ||||||||
6827 | INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)initializeAssumptionCacheTrackerPass(Registry); | ||||||||
6828 | INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass)initializeBlockFrequencyInfoWrapperPassPass(Registry); | ||||||||
6829 | INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)initializeDominatorTreeWrapperPassPass(Registry); | ||||||||
6830 | INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)initializeScalarEvolutionWrapperPassPass(Registry); | ||||||||
6831 | INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)initializeLoopInfoWrapperPassPass(Registry); | ||||||||
6832 | INITIALIZE_PASS_DEPENDENCY(LoopAccessLegacyAnalysis)initializeLoopAccessLegacyAnalysisPass(Registry); | ||||||||
6833 | INITIALIZE_PASS_DEPENDENCY(DemandedBitsWrapperPass)initializeDemandedBitsWrapperPassPass(Registry); | ||||||||
6834 | INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass)initializeOptimizationRemarkEmitterWrapperPassPass(Registry); | ||||||||
6835 | INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass)initializeProfileSummaryInfoWrapperPassPass(Registry); | ||||||||
6836 | INITIALIZE_PASS_DEPENDENCY(InjectTLIMappingsLegacy)initializeInjectTLIMappingsLegacyPass(Registry); | ||||||||
6837 | 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)); } | ||||||||
6838 | |||||||||
6839 | namespace llvm { | ||||||||
6840 | |||||||||
6841 | Pass *createLoopVectorizePass() { return new LoopVectorize(); } | ||||||||
6842 | |||||||||
6843 | Pass *createLoopVectorizePass(bool InterleaveOnlyWhenForced, | ||||||||
6844 | bool VectorizeOnlyWhenForced) { | ||||||||
6845 | return new LoopVectorize(InterleaveOnlyWhenForced, VectorizeOnlyWhenForced); | ||||||||
6846 | } | ||||||||
6847 | |||||||||
6848 | } // end namespace llvm | ||||||||
6849 | |||||||||
6850 | bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) { | ||||||||
6851 | // Check if the pointer operand of a load or store instruction is | ||||||||
6852 | // consecutive. | ||||||||
6853 | if (auto *Ptr = getLoadStorePointerOperand(Inst)) | ||||||||
6854 | return Legal->isConsecutivePtr(Ptr); | ||||||||
6855 | return false; | ||||||||
6856 | } | ||||||||
6857 | |||||||||
6858 | void LoopVectorizationCostModel::collectValuesToIgnore() { | ||||||||
6859 | // Ignore ephemeral values. | ||||||||
6860 | CodeMetrics::collectEphemeralValues(TheLoop, AC, ValuesToIgnore); | ||||||||
6861 | |||||||||
6862 | // Ignore type-promoting instructions we identified during reduction | ||||||||
6863 | // detection. | ||||||||
6864 | for (auto &Reduction : Legal->getReductionVars()) { | ||||||||
6865 | RecurrenceDescriptor &RedDes = Reduction.second; | ||||||||
6866 | const SmallPtrSetImpl<Instruction *> &Casts = RedDes.getCastInsts(); | ||||||||
6867 | VecValuesToIgnore.insert(Casts.begin(), Casts.end()); | ||||||||
6868 | } | ||||||||
6869 | // Ignore type-casting instructions we identified during induction | ||||||||
6870 | // detection. | ||||||||
6871 | for (auto &Induction : Legal->getInductionVars()) { | ||||||||
6872 | InductionDescriptor &IndDes = Induction.second; | ||||||||
6873 | const SmallVectorImpl<Instruction *> &Casts = IndDes.getCastInsts(); | ||||||||
6874 | VecValuesToIgnore.insert(Casts.begin(), Casts.end()); | ||||||||
6875 | } | ||||||||
6876 | } | ||||||||
6877 | |||||||||
6878 | void LoopVectorizationCostModel::collectInLoopReductions() { | ||||||||
6879 | // For the moment, without predicated reduction instructions, we do not | ||||||||
6880 | // support inloop reductions whilst folding the tail, and hence in those cases | ||||||||
6881 | // all reductions are currently out of the loop. | ||||||||
6882 | if (foldTailByMasking()) | ||||||||
6883 | return; | ||||||||
6884 | |||||||||
6885 | for (auto &Reduction : Legal->getReductionVars()) { | ||||||||
6886 | PHINode *Phi = Reduction.first; | ||||||||
6887 | RecurrenceDescriptor &RdxDesc = Reduction.second; | ||||||||
6888 | |||||||||
6889 | // We don't collect reductions that are type promoted (yet). | ||||||||
6890 | if (RdxDesc.getRecurrenceType() != Phi->getType()) | ||||||||
6891 | continue; | ||||||||
6892 | |||||||||
6893 | // If the target would prefer this reduction to happen "in-loop", then we | ||||||||
6894 | // want to record it as such. | ||||||||
6895 | unsigned Opcode = RdxDesc.getRecurrenceBinOp(RdxDesc.getRecurrenceKind()); | ||||||||
6896 | if (!PreferInLoopReductions && | ||||||||
6897 | !TTI.preferInLoopReduction(Opcode, Phi->getType(), | ||||||||
6898 | TargetTransformInfo::ReductionFlags())) | ||||||||
6899 | continue; | ||||||||
6900 | |||||||||
6901 | // Check that we can correctly put the reductions into the loop, by | ||||||||
6902 | // finding the chain of operations that leads from the phi to the loop | ||||||||
6903 | // exit value. | ||||||||
6904 | SmallVector<Instruction *, 4> ReductionOperations = | ||||||||
6905 | RdxDesc.getReductionOpChain(Phi, TheLoop); | ||||||||
6906 | bool InLoop = !ReductionOperations.empty(); | ||||||||
6907 | if (InLoop) | ||||||||
6908 | InLoopReductionChains[Phi] = ReductionOperations; | ||||||||
6909 | 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) | ||||||||
6910 | << " 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); | ||||||||
6911 | } | ||||||||
6912 | } | ||||||||
6913 | |||||||||
6914 | // TODO: we could return a pair of values that specify the max VF and | ||||||||
6915 | // min VF, to be used in `buildVPlans(MinVF, MaxVF)` instead of | ||||||||
6916 | // `buildVPlans(VF, VF)`. We cannot do it because VPLAN at the moment | ||||||||
6917 | // doesn't have a cost model that can choose which plan to execute if | ||||||||
6918 | // more than one is generated. | ||||||||
6919 | static unsigned determineVPlanVF(const unsigned WidestVectorRegBits, | ||||||||
6920 | LoopVectorizationCostModel &CM) { | ||||||||
6921 | unsigned WidestType; | ||||||||
6922 | std::tie(std::ignore, WidestType) = CM.getSmallestAndWidestTypes(); | ||||||||
6923 | return WidestVectorRegBits / WidestType; | ||||||||
6924 | } | ||||||||
6925 | |||||||||
6926 | VectorizationFactor | ||||||||
6927 | LoopVectorizationPlanner::planInVPlanNativePath(ElementCount UserVF) { | ||||||||
6928 | assert(!UserVF.isScalable() && "scalable vectors not yet supported")((!UserVF.isScalable() && "scalable vectors not yet supported" ) ? static_cast<void> (0) : __assert_fail ("!UserVF.isScalable() && \"scalable vectors not yet supported\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 6928, __PRETTY_FUNCTION__)); | ||||||||
6929 | ElementCount VF = UserVF; | ||||||||
6930 | // Outer loop handling: They may require CFG and instruction level | ||||||||
6931 | // transformations before even evaluating whether vectorization is profitable. | ||||||||
6932 | // Since we cannot modify the incoming IR, we need to build VPlan upfront in | ||||||||
6933 | // the vectorization pipeline. | ||||||||
6934 | if (!OrigLoop->isInnermost()) { | ||||||||
6935 | // If the user doesn't provide a vectorization factor, determine a | ||||||||
6936 | // reasonable one. | ||||||||
6937 | if (UserVF.isZero()) { | ||||||||
6938 | VF = ElementCount::getFixed( | ||||||||
6939 | determineVPlanVF(TTI->getRegisterBitWidth(true /* Vector*/), CM)); | ||||||||
6940 | 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); | ||||||||
6941 | |||||||||
6942 | // Make sure we have a VF > 1 for stress testing. | ||||||||
6943 | if (VPlanBuildStressTest && (VF.isScalar() || VF.isZero())) { | ||||||||
6944 | 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) | ||||||||
6945 | << "overriding computed VF.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: VPlan stress testing: " << "overriding computed VF.\n"; } } while (false); | ||||||||
6946 | VF = ElementCount::getFixed(4); | ||||||||
6947 | } | ||||||||
6948 | } | ||||||||
6949 | assert(EnableVPlanNativePath && "VPlan-native path is not enabled.")((EnableVPlanNativePath && "VPlan-native path is not enabled." ) ? static_cast<void> (0) : __assert_fail ("EnableVPlanNativePath && \"VPlan-native path is not enabled.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 6949, __PRETTY_FUNCTION__)); | ||||||||
6950 | assert(isPowerOf2_32(VF.getKnownMinValue()) &&((isPowerOf2_32(VF.getKnownMinValue()) && "VF needs to be a power of two" ) ? static_cast<void> (0) : __assert_fail ("isPowerOf2_32(VF.getKnownMinValue()) && \"VF needs to be a power of two\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 6951, __PRETTY_FUNCTION__)) | ||||||||
6951 | "VF needs to be a power of two")((isPowerOf2_32(VF.getKnownMinValue()) && "VF needs to be a power of two" ) ? static_cast<void> (0) : __assert_fail ("isPowerOf2_32(VF.getKnownMinValue()) && \"VF needs to be a power of two\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 6951, __PRETTY_FUNCTION__)); | ||||||||
6952 | 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) | ||||||||
6953 | << "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); | ||||||||
6954 | buildVPlans(VF.getKnownMinValue(), VF.getKnownMinValue()); | ||||||||
6955 | |||||||||
6956 | // For VPlan build stress testing, we bail out after VPlan construction. | ||||||||
6957 | if (VPlanBuildStressTest) | ||||||||
6958 | return VectorizationFactor::Disabled(); | ||||||||
6959 | |||||||||
6960 | return {VF, 0 /*Cost*/}; | ||||||||
6961 | } | ||||||||
6962 | |||||||||
6963 | 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) | ||||||||
6964 | 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) | ||||||||
6965 | "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); | ||||||||
6966 | return VectorizationFactor::Disabled(); | ||||||||
6967 | } | ||||||||
6968 | |||||||||
6969 | Optional<VectorizationFactor> | ||||||||
6970 | LoopVectorizationPlanner::plan(ElementCount UserVF, unsigned UserIC) { | ||||||||
6971 | assert(!UserVF.isScalable() && "scalable vectorization not yet handled")((!UserVF.isScalable() && "scalable vectorization not yet handled" ) ? static_cast<void> (0) : __assert_fail ("!UserVF.isScalable() && \"scalable vectorization not yet handled\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 6971, __PRETTY_FUNCTION__)); | ||||||||
6972 | assert(OrigLoop->isInnermost() && "Inner loop expected.")((OrigLoop->isInnermost() && "Inner loop expected." ) ? static_cast<void> (0) : __assert_fail ("OrigLoop->isInnermost() && \"Inner loop expected.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 6972, __PRETTY_FUNCTION__)); | ||||||||
6973 | Optional<unsigned> MaybeMaxVF = | ||||||||
6974 | CM.computeMaxVF(UserVF.getKnownMinValue(), UserIC); | ||||||||
6975 | if (!MaybeMaxVF) // Cases that should not to be vectorized nor interleaved. | ||||||||
6976 | return None; | ||||||||
6977 | |||||||||
6978 | // Invalidate interleave groups if all blocks of loop will be predicated. | ||||||||
6979 | if (CM.blockNeedsPredication(OrigLoop->getHeader()) && | ||||||||
6980 | !useMaskedInterleavedAccesses(*TTI)) { | ||||||||
6981 | 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 ) | ||||||||
6982 | 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 ) | ||||||||
6983 | << "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 ) | ||||||||
6984 | "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 ); | ||||||||
6985 | if (CM.InterleaveInfo.invalidateGroups()) | ||||||||
6986 | // Invalidating interleave groups also requires invalidating all decisions | ||||||||
6987 | // based on them, which includes widening decisions and uniform and scalar | ||||||||
6988 | // values. | ||||||||
6989 | CM.invalidateCostModelingDecisions(); | ||||||||
6990 | } | ||||||||
6991 | |||||||||
6992 | if (!UserVF.isZero()) { | ||||||||
6993 | 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); | ||||||||
6994 | assert(isPowerOf2_32(UserVF.getKnownMinValue()) &&((isPowerOf2_32(UserVF.getKnownMinValue()) && "VF needs to be a power of two" ) ? static_cast<void> (0) : __assert_fail ("isPowerOf2_32(UserVF.getKnownMinValue()) && \"VF needs to be a power of two\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 6995, __PRETTY_FUNCTION__)) | ||||||||
6995 | "VF needs to be a power of two")((isPowerOf2_32(UserVF.getKnownMinValue()) && "VF needs to be a power of two" ) ? static_cast<void> (0) : __assert_fail ("isPowerOf2_32(UserVF.getKnownMinValue()) && \"VF needs to be a power of two\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 6995, __PRETTY_FUNCTION__)); | ||||||||
6996 | // Collect the instructions (and their associated costs) that will be more | ||||||||
6997 | // profitable to scalarize. | ||||||||
6998 | CM.selectUserVectorizationFactor(UserVF); | ||||||||
6999 | CM.collectInLoopReductions(); | ||||||||
7000 | buildVPlansWithVPRecipes(UserVF.getKnownMinValue(), | ||||||||
7001 | UserVF.getKnownMinValue()); | ||||||||
7002 | LLVM_DEBUG(printPlans(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { printPlans(dbgs()); } } while (false); | ||||||||
7003 | return {{UserVF, 0}}; | ||||||||
7004 | } | ||||||||
7005 | |||||||||
7006 | unsigned MaxVF = MaybeMaxVF.getValue(); | ||||||||
7007 | assert(MaxVF != 0 && "MaxVF is zero.")((MaxVF != 0 && "MaxVF is zero.") ? static_cast<void > (0) : __assert_fail ("MaxVF != 0 && \"MaxVF is zero.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 7007, __PRETTY_FUNCTION__)); | ||||||||
7008 | |||||||||
7009 | for (unsigned VF = 1; VF <= MaxVF; VF *= 2) { | ||||||||
7010 | // Collect Uniform and Scalar instructions after vectorization with VF. | ||||||||
7011 | CM.collectUniformsAndScalars(ElementCount::getFixed(VF)); | ||||||||
7012 | |||||||||
7013 | // Collect the instructions (and their associated costs) that will be more | ||||||||
7014 | // profitable to scalarize. | ||||||||
7015 | if (VF > 1) | ||||||||
7016 | CM.collectInstsToScalarize(ElementCount::getFixed(VF)); | ||||||||
7017 | } | ||||||||
7018 | |||||||||
7019 | CM.collectInLoopReductions(); | ||||||||
7020 | |||||||||
7021 | buildVPlansWithVPRecipes(1, MaxVF); | ||||||||
7022 | LLVM_DEBUG(printPlans(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { printPlans(dbgs()); } } while (false); | ||||||||
7023 | if (MaxVF == 1) | ||||||||
7024 | return VectorizationFactor::Disabled(); | ||||||||
7025 | |||||||||
7026 | // Select the optimal vectorization factor. | ||||||||
7027 | return CM.selectVectorizationFactor(MaxVF); | ||||||||
7028 | } | ||||||||
7029 | |||||||||
7030 | void LoopVectorizationPlanner::setBestPlan(ElementCount VF, unsigned UF) { | ||||||||
7031 | LLVM_DEBUG(dbgs() << "Setting best plan to VF=" << VF << ", UF=" << UFdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "Setting best plan to VF=" << VF << ", UF=" << UF << '\n'; } } while (false) | ||||||||
7032 | << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "Setting best plan to VF=" << VF << ", UF=" << UF << '\n'; } } while (false); | ||||||||
7033 | BestVF = VF; | ||||||||
7034 | BestUF = UF; | ||||||||
7035 | |||||||||
7036 | erase_if(VPlans, [VF](const VPlanPtr &Plan) { | ||||||||
7037 | return !Plan->hasVF(VF); | ||||||||
7038 | }); | ||||||||
7039 | assert(VPlans.size() == 1 && "Best VF has not a single VPlan.")((VPlans.size() == 1 && "Best VF has not a single VPlan." ) ? static_cast<void> (0) : __assert_fail ("VPlans.size() == 1 && \"Best VF has not a single VPlan.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 7039, __PRETTY_FUNCTION__)); | ||||||||
7040 | } | ||||||||
7041 | |||||||||
7042 | void LoopVectorizationPlanner::executePlan(InnerLoopVectorizer &ILV, | ||||||||
7043 | DominatorTree *DT) { | ||||||||
7044 | // Perform the actual loop transformation. | ||||||||
7045 | |||||||||
7046 | // 1. Create a new empty loop. Unlink the old loop and connect the new one. | ||||||||
7047 | VPCallbackILV CallbackILV(ILV); | ||||||||
7048 | |||||||||
7049 | assert(BestVF.hasValue() && "Vectorization Factor is missing")((BestVF.hasValue() && "Vectorization Factor is missing" ) ? static_cast<void> (0) : __assert_fail ("BestVF.hasValue() && \"Vectorization Factor is missing\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 7049, __PRETTY_FUNCTION__)); | ||||||||
7050 | |||||||||
7051 | VPTransformState State{*BestVF, BestUF, LI, | ||||||||
7052 | DT, ILV.Builder, ILV.VectorLoopValueMap, | ||||||||
7053 | &ILV, CallbackILV}; | ||||||||
7054 | State.CFG.PrevBB = ILV.createVectorizedLoopSkeleton(); | ||||||||
7055 | State.TripCount = ILV.getOrCreateTripCount(nullptr); | ||||||||
7056 | State.CanonicalIV = ILV.Induction; | ||||||||
7057 | |||||||||
7058 | //===------------------------------------------------===// | ||||||||
7059 | // | ||||||||
7060 | // Notice: any optimization or new instruction that go | ||||||||
7061 | // into the code below should also be implemented in | ||||||||
7062 | // the cost-model. | ||||||||
7063 | // | ||||||||
7064 | //===------------------------------------------------===// | ||||||||
7065 | |||||||||
7066 | // 2. Copy and widen instructions from the old loop into the new loop. | ||||||||
7067 | assert(VPlans.size() == 1 && "Not a single VPlan to execute.")((VPlans.size() == 1 && "Not a single VPlan to execute." ) ? static_cast<void> (0) : __assert_fail ("VPlans.size() == 1 && \"Not a single VPlan to execute.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 7067, __PRETTY_FUNCTION__)); | ||||||||
7068 | VPlans.front()->execute(&State); | ||||||||
7069 | |||||||||
7070 | // 3. Fix the vectorized code: take care of header phi's, live-outs, | ||||||||
7071 | // predication, updating analyses. | ||||||||
7072 | ILV.fixVectorizedLoop(); | ||||||||
7073 | } | ||||||||
7074 | |||||||||
7075 | void LoopVectorizationPlanner::collectTriviallyDeadInstructions( | ||||||||
7076 | SmallPtrSetImpl<Instruction *> &DeadInstructions) { | ||||||||
7077 | BasicBlock *Latch = OrigLoop->getLoopLatch(); | ||||||||
7078 | |||||||||
7079 | // We create new control-flow for the vectorized loop, so the original | ||||||||
7080 | // condition will be dead after vectorization if it's only used by the | ||||||||
7081 | // branch. | ||||||||
7082 | auto *Cmp = dyn_cast<Instruction>(Latch->getTerminator()->getOperand(0)); | ||||||||
7083 | if (Cmp && Cmp->hasOneUse()) | ||||||||
7084 | DeadInstructions.insert(Cmp); | ||||||||
7085 | |||||||||
7086 | // We create new "steps" for induction variable updates to which the original | ||||||||
7087 | // induction variables map. An original update instruction will be dead if | ||||||||
7088 | // all its users except the induction variable are dead. | ||||||||
7089 | for (auto &Induction : Legal->getInductionVars()) { | ||||||||
7090 | PHINode *Ind = Induction.first; | ||||||||
7091 | auto *IndUpdate = cast<Instruction>(Ind->getIncomingValueForBlock(Latch)); | ||||||||
7092 | if (llvm::all_of(IndUpdate->users(), [&](User *U) -> bool { | ||||||||
7093 | return U == Ind || DeadInstructions.count(cast<Instruction>(U)); | ||||||||
7094 | })) | ||||||||
7095 | DeadInstructions.insert(IndUpdate); | ||||||||
7096 | |||||||||
7097 | // We record as "Dead" also the type-casting instructions we had identified | ||||||||
7098 | // during induction analysis. We don't need any handling for them in the | ||||||||
7099 | // vectorized loop because we have proven that, under a proper runtime | ||||||||
7100 | // test guarding the vectorized loop, the value of the phi, and the casted | ||||||||
7101 | // value of the phi, are the same. The last instruction in this casting chain | ||||||||
7102 | // will get its scalar/vector/widened def from the scalar/vector/widened def | ||||||||
7103 | // of the respective phi node. Any other casts in the induction def-use chain | ||||||||
7104 | // have no other uses outside the phi update chain, and will be ignored. | ||||||||
7105 | InductionDescriptor &IndDes = Induction.second; | ||||||||
7106 | const SmallVectorImpl<Instruction *> &Casts = IndDes.getCastInsts(); | ||||||||
7107 | DeadInstructions.insert(Casts.begin(), Casts.end()); | ||||||||
7108 | } | ||||||||
7109 | } | ||||||||
7110 | |||||||||
7111 | Value *InnerLoopUnroller::reverseVector(Value *Vec) { return Vec; } | ||||||||
7112 | |||||||||
7113 | Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) { return V; } | ||||||||
7114 | |||||||||
7115 | Value *InnerLoopUnroller::getStepVector(Value *Val, int StartIdx, Value *Step, | ||||||||
7116 | Instruction::BinaryOps BinOp) { | ||||||||
7117 | // When unrolling and the VF is 1, we only need to add a simple scalar. | ||||||||
7118 | Type *Ty = Val->getType(); | ||||||||
7119 | assert(!Ty->isVectorTy() && "Val must be a scalar")((!Ty->isVectorTy() && "Val must be a scalar") ? static_cast <void> (0) : __assert_fail ("!Ty->isVectorTy() && \"Val must be a scalar\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 7119, __PRETTY_FUNCTION__)); | ||||||||
7120 | |||||||||
7121 | if (Ty->isFloatingPointTy()) { | ||||||||
7122 | Constant *C = ConstantFP::get(Ty, (double)StartIdx); | ||||||||
7123 | |||||||||
7124 | // Floating point operations had to be 'fast' to enable the unrolling. | ||||||||
7125 | Value *MulOp = addFastMathFlag(Builder.CreateFMul(C, Step)); | ||||||||
7126 | return addFastMathFlag(Builder.CreateBinOp(BinOp, Val, MulOp)); | ||||||||
7127 | } | ||||||||
7128 | Constant *C = ConstantInt::get(Ty, StartIdx); | ||||||||
7129 | return Builder.CreateAdd(Val, Builder.CreateMul(C, Step), "induction"); | ||||||||
7130 | } | ||||||||
7131 | |||||||||
7132 | static void AddRuntimeUnrollDisableMetaData(Loop *L) { | ||||||||
7133 | SmallVector<Metadata *, 4> MDs; | ||||||||
7134 | // Reserve first location for self reference to the LoopID metadata node. | ||||||||
7135 | MDs.push_back(nullptr); | ||||||||
7136 | bool IsUnrollMetadata = false; | ||||||||
7137 | MDNode *LoopID = L->getLoopID(); | ||||||||
7138 | if (LoopID) { | ||||||||
7139 | // First find existing loop unrolling disable metadata. | ||||||||
7140 | for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) { | ||||||||
7141 | auto *MD = dyn_cast<MDNode>(LoopID->getOperand(i)); | ||||||||
7142 | if (MD) { | ||||||||
7143 | const auto *S = dyn_cast<MDString>(MD->getOperand(0)); | ||||||||
7144 | IsUnrollMetadata = | ||||||||
7145 | S && S->getString().startswith("llvm.loop.unroll.disable"); | ||||||||
7146 | } | ||||||||
7147 | MDs.push_back(LoopID->getOperand(i)); | ||||||||
7148 | } | ||||||||
7149 | } | ||||||||
7150 | |||||||||
7151 | if (!IsUnrollMetadata) { | ||||||||
7152 | // Add runtime unroll disable metadata. | ||||||||
7153 | LLVMContext &Context = L->getHeader()->getContext(); | ||||||||
7154 | SmallVector<Metadata *, 1> DisableOperands; | ||||||||
7155 | DisableOperands.push_back( | ||||||||
7156 | MDString::get(Context, "llvm.loop.unroll.runtime.disable")); | ||||||||
7157 | MDNode *DisableNode = MDNode::get(Context, DisableOperands); | ||||||||
7158 | MDs.push_back(DisableNode); | ||||||||
7159 | MDNode *NewLoopID = MDNode::get(Context, MDs); | ||||||||
7160 | // Set operand 0 to refer to the loop id itself. | ||||||||
7161 | NewLoopID->replaceOperandWith(0, NewLoopID); | ||||||||
7162 | L->setLoopID(NewLoopID); | ||||||||
7163 | } | ||||||||
7164 | } | ||||||||
7165 | |||||||||
7166 | bool LoopVectorizationPlanner::getDecisionAndClampRange( | ||||||||
7167 | const std::function<bool(ElementCount)> &Predicate, VFRange &Range) { | ||||||||
7168 | assert(Range.End > Range.Start && "Trying to test an empty VF range.")((Range.End > Range.Start && "Trying to test an empty VF range." ) ? static_cast<void> (0) : __assert_fail ("Range.End > Range.Start && \"Trying to test an empty VF range.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 7168, __PRETTY_FUNCTION__)); | ||||||||
7169 | bool PredicateAtRangeStart = Predicate(ElementCount::getFixed(Range.Start)); | ||||||||
7170 | |||||||||
7171 | for (unsigned TmpVF = Range.Start * 2; TmpVF < Range.End; TmpVF *= 2) | ||||||||
7172 | if (Predicate(ElementCount::getFixed(TmpVF)) != PredicateAtRangeStart) { | ||||||||
7173 | Range.End = TmpVF; | ||||||||
7174 | break; | ||||||||
7175 | } | ||||||||
7176 | |||||||||
7177 | return PredicateAtRangeStart; | ||||||||
7178 | } | ||||||||
7179 | |||||||||
7180 | /// Build VPlans for the full range of feasible VF's = {\p MinVF, 2 * \p MinVF, | ||||||||
7181 | /// 4 * \p MinVF, ..., \p MaxVF} by repeatedly building a VPlan for a sub-range | ||||||||
7182 | /// of VF's starting at a given VF and extending it as much as possible. Each | ||||||||
7183 | /// vectorization decision can potentially shorten this sub-range during | ||||||||
7184 | /// buildVPlan(). | ||||||||
7185 | void LoopVectorizationPlanner::buildVPlans(unsigned MinVF, unsigned MaxVF) { | ||||||||
7186 | for (unsigned VF = MinVF; VF < MaxVF + 1;) { | ||||||||
7187 | VFRange SubRange = {VF, MaxVF + 1}; | ||||||||
7188 | VPlans.push_back(buildVPlan(SubRange)); | ||||||||
7189 | VF = SubRange.End; | ||||||||
7190 | } | ||||||||
7191 | } | ||||||||
7192 | |||||||||
7193 | VPValue *VPRecipeBuilder::createEdgeMask(BasicBlock *Src, BasicBlock *Dst, | ||||||||
7194 | VPlanPtr &Plan) { | ||||||||
7195 | assert(is_contained(predecessors(Dst), Src) && "Invalid edge")((is_contained(predecessors(Dst), Src) && "Invalid edge" ) ? static_cast<void> (0) : __assert_fail ("is_contained(predecessors(Dst), Src) && \"Invalid edge\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 7195, __PRETTY_FUNCTION__)); | ||||||||
7196 | |||||||||
7197 | // Look for cached value. | ||||||||
7198 | std::pair<BasicBlock *, BasicBlock *> Edge(Src, Dst); | ||||||||
7199 | EdgeMaskCacheTy::iterator ECEntryIt = EdgeMaskCache.find(Edge); | ||||||||
7200 | if (ECEntryIt != EdgeMaskCache.end()) | ||||||||
7201 | return ECEntryIt->second; | ||||||||
7202 | |||||||||
7203 | VPValue *SrcMask = createBlockInMask(Src, Plan); | ||||||||
7204 | |||||||||
7205 | // The terminator has to be a branch inst! | ||||||||
7206 | BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator()); | ||||||||
7207 | assert(BI && "Unexpected terminator found")((BI && "Unexpected terminator found") ? static_cast< void> (0) : __assert_fail ("BI && \"Unexpected terminator found\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 7207, __PRETTY_FUNCTION__)); | ||||||||
7208 | |||||||||
7209 | if (!BI->isConditional() || BI->getSuccessor(0) == BI->getSuccessor(1)) | ||||||||
7210 | return EdgeMaskCache[Edge] = SrcMask; | ||||||||
7211 | |||||||||
7212 | VPValue *EdgeMask = Plan->getVPValue(BI->getCondition()); | ||||||||
7213 | assert(EdgeMask && "No Edge Mask found for condition")((EdgeMask && "No Edge Mask found for condition") ? static_cast <void> (0) : __assert_fail ("EdgeMask && \"No Edge Mask found for condition\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 7213, __PRETTY_FUNCTION__)); | ||||||||
7214 | |||||||||
7215 | if (BI->getSuccessor(0) != Dst) | ||||||||
7216 | EdgeMask = Builder.createNot(EdgeMask); | ||||||||
7217 | |||||||||
7218 | if (SrcMask) // Otherwise block in-mask is all-one, no need to AND. | ||||||||
7219 | EdgeMask = Builder.createAnd(EdgeMask, SrcMask); | ||||||||
7220 | |||||||||
7221 | return EdgeMaskCache[Edge] = EdgeMask; | ||||||||
7222 | } | ||||||||
7223 | |||||||||
7224 | VPValue *VPRecipeBuilder::createBlockInMask(BasicBlock *BB, VPlanPtr &Plan) { | ||||||||
7225 | assert(OrigLoop->contains(BB) && "Block is not a part of a loop")((OrigLoop->contains(BB) && "Block is not a part of a loop" ) ? static_cast<void> (0) : __assert_fail ("OrigLoop->contains(BB) && \"Block is not a part of a loop\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 7225, __PRETTY_FUNCTION__)); | ||||||||
7226 | |||||||||
7227 | // Look for cached value. | ||||||||
7228 | BlockMaskCacheTy::iterator BCEntryIt = BlockMaskCache.find(BB); | ||||||||
7229 | if (BCEntryIt != BlockMaskCache.end()) | ||||||||
7230 | return BCEntryIt->second; | ||||||||
7231 | |||||||||
7232 | // All-one mask is modelled as no-mask following the convention for masked | ||||||||
7233 | // load/store/gather/scatter. Initialize BlockMask to no-mask. | ||||||||
7234 | VPValue *BlockMask = nullptr; | ||||||||
7235 | |||||||||
7236 | if (OrigLoop->getHeader() == BB) { | ||||||||
7237 | if (!CM.blockNeedsPredication(BB)) | ||||||||
7238 | return BlockMaskCache[BB] = BlockMask; // Loop incoming mask is all-one. | ||||||||
7239 | |||||||||
7240 | // Introduce the early-exit compare IV <= BTC to form header block mask. | ||||||||
7241 | // This is used instead of IV < TC because TC may wrap, unlike BTC. | ||||||||
7242 | // Start by constructing the desired canonical IV. | ||||||||
7243 | VPValue *IV = nullptr; | ||||||||
7244 | if (Legal->getPrimaryInduction()) | ||||||||
7245 | IV = Plan->getVPValue(Legal->getPrimaryInduction()); | ||||||||
7246 | else { | ||||||||
7247 | auto IVRecipe = new VPWidenCanonicalIVRecipe(); | ||||||||
7248 | Builder.getInsertBlock()->appendRecipe(IVRecipe); | ||||||||
7249 | IV = IVRecipe->getVPValue(); | ||||||||
7250 | } | ||||||||
7251 | VPValue *BTC = Plan->getOrCreateBackedgeTakenCount(); | ||||||||
7252 | bool TailFolded = !CM.isScalarEpilogueAllowed(); | ||||||||
7253 | |||||||||
7254 | if (TailFolded && CM.TTI.emitGetActiveLaneMask()) { | ||||||||
7255 | // While ActiveLaneMask is a binary op that consumes the loop tripcount | ||||||||
7256 | // as a second argument, we only pass the IV here and extract the | ||||||||
7257 | // tripcount from the transform state where codegen of the VP instructions | ||||||||
7258 | // happen. | ||||||||
7259 | BlockMask = Builder.createNaryOp(VPInstruction::ActiveLaneMask, {IV}); | ||||||||
7260 | } else { | ||||||||
7261 | BlockMask = Builder.createNaryOp(VPInstruction::ICmpULE, {IV, BTC}); | ||||||||
7262 | } | ||||||||
7263 | return BlockMaskCache[BB] = BlockMask; | ||||||||
7264 | } | ||||||||
7265 | |||||||||
7266 | // This is the block mask. We OR all incoming edges. | ||||||||
7267 | for (auto *Predecessor : predecessors(BB)) { | ||||||||
7268 | VPValue *EdgeMask = createEdgeMask(Predecessor, BB, Plan); | ||||||||
7269 | if (!EdgeMask
| ||||||||
7270 | return BlockMaskCache[BB] = EdgeMask; | ||||||||
| |||||||||
7271 | |||||||||
7272 | if (!BlockMask
| ||||||||
7273 | BlockMask = EdgeMask; | ||||||||
7274 | continue; | ||||||||
7275 | } | ||||||||
7276 | |||||||||
7277 | BlockMask = Builder.createOr(BlockMask, EdgeMask); | ||||||||
7278 | } | ||||||||
7279 | |||||||||
7280 | return BlockMaskCache[BB] = BlockMask; | ||||||||
7281 | } | ||||||||
7282 | |||||||||
7283 | VPWidenMemoryInstructionRecipe * | ||||||||
7284 | VPRecipeBuilder::tryToWidenMemory(Instruction *I, VFRange &Range, | ||||||||
7285 | VPlanPtr &Plan) { | ||||||||
7286 | assert((isa<LoadInst>(I) || isa<StoreInst>(I)) &&(((isa<LoadInst>(I) || isa<StoreInst>(I)) && "Must be called with either a load or store") ? static_cast< void> (0) : __assert_fail ("(isa<LoadInst>(I) || isa<StoreInst>(I)) && \"Must be called with either a load or store\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 7287, __PRETTY_FUNCTION__)) | ||||||||
7287 | "Must be called with either a load or store")(((isa<LoadInst>(I) || isa<StoreInst>(I)) && "Must be called with either a load or store") ? static_cast< void> (0) : __assert_fail ("(isa<LoadInst>(I) || isa<StoreInst>(I)) && \"Must be called with either a load or store\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 7287, __PRETTY_FUNCTION__)); | ||||||||
7288 | |||||||||
7289 | auto willWiden = [&](ElementCount VF) -> bool { | ||||||||
7290 | assert(!VF.isScalable() && "unexpected scalable ElementCount")((!VF.isScalable() && "unexpected scalable ElementCount" ) ? static_cast<void> (0) : __assert_fail ("!VF.isScalable() && \"unexpected scalable ElementCount\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 7290, __PRETTY_FUNCTION__)); | ||||||||
7291 | if (VF.isScalar()) | ||||||||
7292 | return false; | ||||||||
7293 | LoopVectorizationCostModel::InstWidening Decision = | ||||||||
7294 | CM.getWideningDecision(I, VF); | ||||||||
7295 | assert(Decision != LoopVectorizationCostModel::CM_Unknown &&((Decision != LoopVectorizationCostModel::CM_Unknown && "CM decision should be taken at this point.") ? static_cast< void> (0) : __assert_fail ("Decision != LoopVectorizationCostModel::CM_Unknown && \"CM decision should be taken at this point.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 7296, __PRETTY_FUNCTION__)) | ||||||||
7296 | "CM decision should be taken at this point.")((Decision != LoopVectorizationCostModel::CM_Unknown && "CM decision should be taken at this point.") ? static_cast< void> (0) : __assert_fail ("Decision != LoopVectorizationCostModel::CM_Unknown && \"CM decision should be taken at this point.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 7296, __PRETTY_FUNCTION__)); | ||||||||
7297 | if (Decision == LoopVectorizationCostModel::CM_Interleave) | ||||||||
7298 | return true; | ||||||||
7299 | if (CM.isScalarAfterVectorization(I, VF) || | ||||||||
7300 | CM.isProfitableToScalarize(I, VF)) | ||||||||
7301 | return false; | ||||||||
7302 | return Decision != LoopVectorizationCostModel::CM_Scalarize; | ||||||||
7303 | }; | ||||||||
7304 | |||||||||
7305 | if (!LoopVectorizationPlanner::getDecisionAndClampRange(willWiden, Range)) | ||||||||
7306 | return nullptr; | ||||||||
7307 | |||||||||
7308 | VPValue *Mask = nullptr; | ||||||||
7309 | if (Legal->isMaskRequired(I)) | ||||||||
7310 | Mask = createBlockInMask(I->getParent(), Plan); | ||||||||
7311 | |||||||||
7312 | VPValue *Addr = Plan->getOrAddVPValue(getLoadStorePointerOperand(I)); | ||||||||
7313 | if (LoadInst *Load = dyn_cast<LoadInst>(I)) | ||||||||
7314 | return new VPWidenMemoryInstructionRecipe(*Load, Addr, Mask); | ||||||||
7315 | |||||||||
7316 | StoreInst *Store = cast<StoreInst>(I); | ||||||||
7317 | VPValue *StoredValue = Plan->getOrAddVPValue(Store->getValueOperand()); | ||||||||
7318 | return new VPWidenMemoryInstructionRecipe(*Store, Addr, StoredValue, Mask); | ||||||||
7319 | } | ||||||||
7320 | |||||||||
7321 | VPWidenIntOrFpInductionRecipe * | ||||||||
7322 | VPRecipeBuilder::tryToOptimizeInductionPHI(PHINode *Phi) const { | ||||||||
7323 | // Check if this is an integer or fp induction. If so, build the recipe that | ||||||||
7324 | // produces its scalar and vector values. | ||||||||
7325 | InductionDescriptor II = Legal->getInductionVars().lookup(Phi); | ||||||||
7326 | if (II.getKind() == InductionDescriptor::IK_IntInduction || | ||||||||
7327 | II.getKind() == InductionDescriptor::IK_FpInduction) | ||||||||
7328 | return new VPWidenIntOrFpInductionRecipe(Phi); | ||||||||
7329 | |||||||||
7330 | return nullptr; | ||||||||
7331 | } | ||||||||
7332 | |||||||||
7333 | VPWidenIntOrFpInductionRecipe * | ||||||||
7334 | VPRecipeBuilder::tryToOptimizeInductionTruncate(TruncInst *I, | ||||||||
7335 | VFRange &Range) const { | ||||||||
7336 | // Optimize the special case where the source is a constant integer | ||||||||
7337 | // induction variable. Notice that we can only optimize the 'trunc' case | ||||||||
7338 | // because (a) FP conversions lose precision, (b) sext/zext may wrap, and | ||||||||
7339 | // (c) other casts depend on pointer size. | ||||||||
7340 | |||||||||
7341 | // Determine whether \p K is a truncation based on an induction variable that | ||||||||
7342 | // can be optimized. | ||||||||
7343 | auto isOptimizableIVTruncate = | ||||||||
7344 | [&](Instruction *K) -> std::function<bool(ElementCount)> { | ||||||||
7345 | return [=](ElementCount VF) -> bool { | ||||||||
7346 | return CM.isOptimizableIVTruncate(K, VF); | ||||||||
7347 | }; | ||||||||
7348 | }; | ||||||||
7349 | |||||||||
7350 | if (LoopVectorizationPlanner::getDecisionAndClampRange( | ||||||||
7351 | isOptimizableIVTruncate(I), Range)) | ||||||||
7352 | return new VPWidenIntOrFpInductionRecipe(cast<PHINode>(I->getOperand(0)), | ||||||||
7353 | I); | ||||||||
7354 | return nullptr; | ||||||||
7355 | } | ||||||||
7356 | |||||||||
7357 | VPBlendRecipe *VPRecipeBuilder::tryToBlend(PHINode *Phi, VPlanPtr &Plan) { | ||||||||
7358 | // We know that all PHIs in non-header blocks are converted into selects, so | ||||||||
7359 | // we don't have to worry about the insertion order and we can just use the | ||||||||
7360 | // builder. At this point we generate the predication tree. There may be | ||||||||
7361 | // duplications since this is a simple recursive scan, but future | ||||||||
7362 | // optimizations will clean it up. | ||||||||
7363 | |||||||||
7364 | SmallVector<VPValue *, 2> Operands; | ||||||||
7365 | unsigned NumIncoming = Phi->getNumIncomingValues(); | ||||||||
7366 | for (unsigned In = 0; In < NumIncoming; In++) { | ||||||||
7367 | VPValue *EdgeMask = | ||||||||
7368 | createEdgeMask(Phi->getIncomingBlock(In), Phi->getParent(), Plan); | ||||||||
7369 | assert((EdgeMask || NumIncoming == 1) &&(((EdgeMask || NumIncoming == 1) && "Multiple predecessors with one having a full mask" ) ? static_cast<void> (0) : __assert_fail ("(EdgeMask || NumIncoming == 1) && \"Multiple predecessors with one having a full mask\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 7370, __PRETTY_FUNCTION__)) | ||||||||
7370 | "Multiple predecessors with one having a full mask")(((EdgeMask || NumIncoming == 1) && "Multiple predecessors with one having a full mask" ) ? static_cast<void> (0) : __assert_fail ("(EdgeMask || NumIncoming == 1) && \"Multiple predecessors with one having a full mask\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 7370, __PRETTY_FUNCTION__)); | ||||||||
7371 | Operands.push_back(Plan->getOrAddVPValue(Phi->getIncomingValue(In))); | ||||||||
7372 | if (EdgeMask) | ||||||||
7373 | Operands.push_back(EdgeMask); | ||||||||
7374 | } | ||||||||
7375 | return new VPBlendRecipe(Phi, Operands); | ||||||||
7376 | } | ||||||||
7377 | |||||||||
7378 | VPWidenCallRecipe *VPRecipeBuilder::tryToWidenCall(CallInst *CI, VFRange &Range, | ||||||||
7379 | VPlan &Plan) const { | ||||||||
7380 | |||||||||
7381 | bool IsPredicated = LoopVectorizationPlanner::getDecisionAndClampRange( | ||||||||
7382 | [this, CI](ElementCount VF) { | ||||||||
7383 | return CM.isScalarWithPredication(CI, VF); | ||||||||
7384 | }, | ||||||||
7385 | Range); | ||||||||
7386 | |||||||||
7387 | if (IsPredicated) | ||||||||
7388 | return nullptr; | ||||||||
7389 | |||||||||
7390 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); | ||||||||
7391 | if (ID && (ID == Intrinsic::assume || ID == Intrinsic::lifetime_end || | ||||||||
7392 | ID == Intrinsic::lifetime_start || ID == Intrinsic::sideeffect)) | ||||||||
7393 | return nullptr; | ||||||||
7394 | |||||||||
7395 | auto willWiden = [&](ElementCount VF) -> bool { | ||||||||
7396 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); | ||||||||
7397 | // The following case may be scalarized depending on the VF. | ||||||||
7398 | // The flag shows whether we use Intrinsic or a usual Call for vectorized | ||||||||
7399 | // version of the instruction. | ||||||||
7400 | // Is it beneficial to perform intrinsic call compared to lib call? | ||||||||
7401 | bool NeedToScalarize = false; | ||||||||
7402 | unsigned CallCost = CM.getVectorCallCost(CI, VF, NeedToScalarize); | ||||||||
7403 | bool UseVectorIntrinsic = | ||||||||
7404 | ID && CM.getVectorIntrinsicCost(CI, VF) <= CallCost; | ||||||||
7405 | return UseVectorIntrinsic || !NeedToScalarize; | ||||||||
7406 | }; | ||||||||
7407 | |||||||||
7408 | if (!LoopVectorizationPlanner::getDecisionAndClampRange(willWiden, Range)) | ||||||||
7409 | return nullptr; | ||||||||
7410 | |||||||||
7411 | return new VPWidenCallRecipe(*CI, Plan.mapToVPValues(CI->arg_operands())); | ||||||||
7412 | } | ||||||||
7413 | |||||||||
7414 | bool VPRecipeBuilder::shouldWiden(Instruction *I, VFRange &Range) const { | ||||||||
7415 | assert(!isa<BranchInst>(I) && !isa<PHINode>(I) && !isa<LoadInst>(I) &&((!isa<BranchInst>(I) && !isa<PHINode>(I) && !isa<LoadInst>(I) && !isa<StoreInst >(I) && "Instruction should have been handled earlier" ) ? static_cast<void> (0) : __assert_fail ("!isa<BranchInst>(I) && !isa<PHINode>(I) && !isa<LoadInst>(I) && !isa<StoreInst>(I) && \"Instruction should have been handled earlier\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 7416, __PRETTY_FUNCTION__)) | ||||||||
7416 | !isa<StoreInst>(I) && "Instruction should have been handled earlier")((!isa<BranchInst>(I) && !isa<PHINode>(I) && !isa<LoadInst>(I) && !isa<StoreInst >(I) && "Instruction should have been handled earlier" ) ? static_cast<void> (0) : __assert_fail ("!isa<BranchInst>(I) && !isa<PHINode>(I) && !isa<LoadInst>(I) && !isa<StoreInst>(I) && \"Instruction should have been handled earlier\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 7416, __PRETTY_FUNCTION__)); | ||||||||
7417 | // Instruction should be widened, unless it is scalar after vectorization, | ||||||||
7418 | // scalarization is profitable or it is predicated. | ||||||||
7419 | auto WillScalarize = [this, I](ElementCount VF) -> bool { | ||||||||
7420 | return CM.isScalarAfterVectorization(I, VF) || | ||||||||
7421 | CM.isProfitableToScalarize(I, VF) || | ||||||||
7422 | CM.isScalarWithPredication(I, VF); | ||||||||
7423 | }; | ||||||||
7424 | return !LoopVectorizationPlanner::getDecisionAndClampRange(WillScalarize, | ||||||||
7425 | Range); | ||||||||
7426 | } | ||||||||
7427 | |||||||||
7428 | VPWidenRecipe *VPRecipeBuilder::tryToWiden(Instruction *I, VPlan &Plan) const { | ||||||||
7429 | auto IsVectorizableOpcode = [](unsigned Opcode) { | ||||||||
7430 | switch (Opcode) { | ||||||||
7431 | case Instruction::Add: | ||||||||
7432 | case Instruction::And: | ||||||||
7433 | case Instruction::AShr: | ||||||||
7434 | case Instruction::BitCast: | ||||||||
7435 | case Instruction::FAdd: | ||||||||
7436 | case Instruction::FCmp: | ||||||||
7437 | case Instruction::FDiv: | ||||||||
7438 | case Instruction::FMul: | ||||||||
7439 | case Instruction::FNeg: | ||||||||
7440 | case Instruction::FPExt: | ||||||||
7441 | case Instruction::FPToSI: | ||||||||
7442 | case Instruction::FPToUI: | ||||||||
7443 | case Instruction::FPTrunc: | ||||||||
7444 | case Instruction::FRem: | ||||||||
7445 | case Instruction::FSub: | ||||||||
7446 | case Instruction::ICmp: | ||||||||
7447 | case Instruction::IntToPtr: | ||||||||
7448 | case Instruction::LShr: | ||||||||
7449 | case Instruction::Mul: | ||||||||
7450 | case Instruction::Or: | ||||||||
7451 | case Instruction::PtrToInt: | ||||||||
7452 | case Instruction::SDiv: | ||||||||
7453 | case Instruction::Select: | ||||||||
7454 | case Instruction::SExt: | ||||||||
7455 | case Instruction::Shl: | ||||||||
7456 | case Instruction::SIToFP: | ||||||||
7457 | case Instruction::SRem: | ||||||||
7458 | case Instruction::Sub: | ||||||||
7459 | case Instruction::Trunc: | ||||||||
7460 | case Instruction::UDiv: | ||||||||
7461 | case Instruction::UIToFP: | ||||||||
7462 | case Instruction::URem: | ||||||||
7463 | case Instruction::Xor: | ||||||||
7464 | case Instruction::ZExt: | ||||||||
7465 | return true; | ||||||||
7466 | } | ||||||||
7467 | return false; | ||||||||
7468 | }; | ||||||||
7469 | |||||||||
7470 | if (!IsVectorizableOpcode(I->getOpcode())) | ||||||||
7471 | return nullptr; | ||||||||
7472 | |||||||||
7473 | // Success: widen this instruction. | ||||||||
7474 | return new VPWidenRecipe(*I, Plan.mapToVPValues(I->operands())); | ||||||||
7475 | } | ||||||||
7476 | |||||||||
7477 | VPBasicBlock *VPRecipeBuilder::handleReplication( | ||||||||
7478 | Instruction *I, VFRange &Range, VPBasicBlock *VPBB, | ||||||||
7479 | DenseMap<Instruction *, VPReplicateRecipe *> &PredInst2Recipe, | ||||||||
7480 | VPlanPtr &Plan) { | ||||||||
7481 | bool IsUniform = LoopVectorizationPlanner::getDecisionAndClampRange( | ||||||||
7482 | [&](ElementCount VF) { return CM.isUniformAfterVectorization(I, VF); }, | ||||||||
7483 | Range); | ||||||||
7484 | |||||||||
7485 | bool IsPredicated = LoopVectorizationPlanner::getDecisionAndClampRange( | ||||||||
7486 | [&](ElementCount VF) { return CM.isScalarWithPredication(I, VF); }, | ||||||||
7487 | Range); | ||||||||
7488 | |||||||||
7489 | auto *Recipe = new VPReplicateRecipe(I, Plan->mapToVPValues(I->operands()), | ||||||||
7490 | IsUniform, IsPredicated); | ||||||||
7491 | setRecipe(I, Recipe); | ||||||||
7492 | |||||||||
7493 | // Find if I uses a predicated instruction. If so, it will use its scalar | ||||||||
7494 | // value. Avoid hoisting the insert-element which packs the scalar value into | ||||||||
7495 | // a vector value, as that happens iff all users use the vector value. | ||||||||
7496 | for (auto &Op : I->operands()) | ||||||||
7497 | if (auto *PredInst = dyn_cast<Instruction>(Op)) | ||||||||
7498 | if (PredInst2Recipe.find(PredInst) != PredInst2Recipe.end()) | ||||||||
7499 | PredInst2Recipe[PredInst]->setAlsoPack(false); | ||||||||
7500 | |||||||||
7501 | // Finalize the recipe for Instr, first if it is not predicated. | ||||||||
7502 | if (!IsPredicated) { | ||||||||
7503 | LLVM_DEBUG(dbgs() << "LV: Scalarizing:" << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Scalarizing:" << *I << "\n"; } } while (false); | ||||||||
7504 | VPBB->appendRecipe(Recipe); | ||||||||
7505 | return VPBB; | ||||||||
7506 | } | ||||||||
7507 | 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); | ||||||||
7508 | assert(VPBB->getSuccessors().empty() &&((VPBB->getSuccessors().empty() && "VPBB has successors when handling predicated replication." ) ? static_cast<void> (0) : __assert_fail ("VPBB->getSuccessors().empty() && \"VPBB has successors when handling predicated replication.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 7509, __PRETTY_FUNCTION__)) | ||||||||
7509 | "VPBB has successors when handling predicated replication.")((VPBB->getSuccessors().empty() && "VPBB has successors when handling predicated replication." ) ? static_cast<void> (0) : __assert_fail ("VPBB->getSuccessors().empty() && \"VPBB has successors when handling predicated replication.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 7509, __PRETTY_FUNCTION__)); | ||||||||
7510 | // Record predicated instructions for above packing optimizations. | ||||||||
7511 | PredInst2Recipe[I] = Recipe; | ||||||||
7512 | VPBlockBase *Region = createReplicateRegion(I, Recipe, Plan); | ||||||||
7513 | VPBlockUtils::insertBlockAfter(Region, VPBB); | ||||||||
7514 | auto *RegSucc = new VPBasicBlock(); | ||||||||
7515 | VPBlockUtils::insertBlockAfter(RegSucc, Region); | ||||||||
7516 | return RegSucc; | ||||||||
7517 | } | ||||||||
7518 | |||||||||
7519 | VPRegionBlock *VPRecipeBuilder::createReplicateRegion(Instruction *Instr, | ||||||||
7520 | VPRecipeBase *PredRecipe, | ||||||||
7521 | VPlanPtr &Plan) { | ||||||||
7522 | // Instructions marked for predication are replicated and placed under an | ||||||||
7523 | // if-then construct to prevent side-effects. | ||||||||
7524 | |||||||||
7525 | // Generate recipes to compute the block mask for this region. | ||||||||
7526 | VPValue *BlockInMask = createBlockInMask(Instr->getParent(), Plan); | ||||||||
| |||||||||
7527 | |||||||||
7528 | // Build the triangular if-then region. | ||||||||
7529 | std::string RegionName = (Twine("pred.") + Instr->getOpcodeName()).str(); | ||||||||
7530 | assert(Instr->getParent() && "Predicated instruction not in any basic block")((Instr->getParent() && "Predicated instruction not in any basic block" ) ? static_cast<void> (0) : __assert_fail ("Instr->getParent() && \"Predicated instruction not in any basic block\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 7530, __PRETTY_FUNCTION__)); | ||||||||
7531 | auto *BOMRecipe = new VPBranchOnMaskRecipe(BlockInMask); | ||||||||
7532 | auto *Entry = new VPBasicBlock(Twine(RegionName) + ".entry", BOMRecipe); | ||||||||
7533 | auto *PHIRecipe = | ||||||||
7534 | Instr->getType()->isVoidTy() ? nullptr : new VPPredInstPHIRecipe(Instr); | ||||||||
7535 | auto *Exit = new VPBasicBlock(Twine(RegionName) + ".continue", PHIRecipe); | ||||||||
7536 | auto *Pred = new VPBasicBlock(Twine(RegionName) + ".if", PredRecipe); | ||||||||
7537 | VPRegionBlock *Region = new VPRegionBlock(Entry, Exit, RegionName, true); | ||||||||
7538 | |||||||||
7539 | // Note: first set Entry as region entry and then connect successors starting | ||||||||
7540 | // from it in order, to propagate the "parent" of each VPBasicBlock. | ||||||||
7541 | VPBlockUtils::insertTwoBlocksAfter(Pred, Exit, BlockInMask, Entry); | ||||||||
7542 | VPBlockUtils::connectBlocks(Pred, Exit); | ||||||||
7543 | |||||||||
7544 | return Region; | ||||||||
7545 | } | ||||||||
7546 | |||||||||
7547 | VPRecipeBase *VPRecipeBuilder::tryToCreateWidenRecipe(Instruction *Instr, | ||||||||
7548 | VFRange &Range, | ||||||||
7549 | VPlanPtr &Plan) { | ||||||||
7550 | // First, check for specific widening recipes that deal with calls, memory | ||||||||
7551 | // operations, inductions and Phi nodes. | ||||||||
7552 | if (auto *CI = dyn_cast<CallInst>(Instr)) | ||||||||
7553 | return tryToWidenCall(CI, Range, *Plan); | ||||||||
7554 | |||||||||
7555 | if (isa<LoadInst>(Instr) || isa<StoreInst>(Instr)) | ||||||||
7556 | return tryToWidenMemory(Instr, Range, Plan); | ||||||||
7557 | |||||||||
7558 | VPRecipeBase *Recipe; | ||||||||
7559 | if (auto Phi = dyn_cast<PHINode>(Instr)) { | ||||||||
7560 | if (Phi->getParent() != OrigLoop->getHeader()) | ||||||||
7561 | return tryToBlend(Phi, Plan); | ||||||||
7562 | if ((Recipe = tryToOptimizeInductionPHI(Phi))) | ||||||||
7563 | return Recipe; | ||||||||
7564 | return new VPWidenPHIRecipe(Phi); | ||||||||
7565 | } | ||||||||
7566 | |||||||||
7567 | if (isa<TruncInst>(Instr) && | ||||||||
7568 | (Recipe = tryToOptimizeInductionTruncate(cast<TruncInst>(Instr), Range))) | ||||||||
7569 | return Recipe; | ||||||||
7570 | |||||||||
7571 | if (!shouldWiden(Instr, Range)) | ||||||||
7572 | return nullptr; | ||||||||
7573 | |||||||||
7574 | if (auto GEP = dyn_cast<GetElementPtrInst>(Instr)) | ||||||||
7575 | return new VPWidenGEPRecipe(GEP, Plan->mapToVPValues(GEP->operands()), | ||||||||
7576 | OrigLoop); | ||||||||
7577 | |||||||||
7578 | if (auto *SI = dyn_cast<SelectInst>(Instr)) { | ||||||||
7579 | bool InvariantCond = | ||||||||
7580 | PSE.getSE()->isLoopInvariant(PSE.getSCEV(SI->getOperand(0)), OrigLoop); | ||||||||
7581 | return new VPWidenSelectRecipe(*SI, Plan->mapToVPValues(SI->operands()), | ||||||||
7582 | InvariantCond); | ||||||||
7583 | } | ||||||||
7584 | |||||||||
7585 | return tryToWiden(Instr, *Plan); | ||||||||
7586 | } | ||||||||
7587 | |||||||||
7588 | void LoopVectorizationPlanner::buildVPlansWithVPRecipes(unsigned MinVF, | ||||||||
7589 | unsigned MaxVF) { | ||||||||
7590 | assert(OrigLoop->isInnermost() && "Inner loop expected.")((OrigLoop->isInnermost() && "Inner loop expected." ) ? static_cast<void> (0) : __assert_fail ("OrigLoop->isInnermost() && \"Inner loop expected.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 7590, __PRETTY_FUNCTION__)); | ||||||||
7591 | |||||||||
7592 | // Collect conditions feeding internal conditional branches; they need to be | ||||||||
7593 | // represented in VPlan for it to model masking. | ||||||||
7594 | SmallPtrSet<Value *, 1> NeedDef; | ||||||||
7595 | |||||||||
7596 | auto *Latch = OrigLoop->getLoopLatch(); | ||||||||
7597 | for (BasicBlock *BB : OrigLoop->blocks()) { | ||||||||
7598 | if (BB == Latch) | ||||||||
7599 | continue; | ||||||||
7600 | BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator()); | ||||||||
7601 | if (Branch && Branch->isConditional()) | ||||||||
7602 | NeedDef.insert(Branch->getCondition()); | ||||||||
7603 | } | ||||||||
7604 | |||||||||
7605 | // If the tail is to be folded by masking, the primary induction variable, if | ||||||||
7606 | // exists needs to be represented in VPlan for it to model early-exit masking. | ||||||||
7607 | // Also, both the Phi and the live-out instruction of each reduction are | ||||||||
7608 | // required in order to introduce a select between them in VPlan. | ||||||||
7609 | if (CM.foldTailByMasking()) { | ||||||||
7610 | if (Legal->getPrimaryInduction()) | ||||||||
7611 | NeedDef.insert(Legal->getPrimaryInduction()); | ||||||||
7612 | for (auto &Reduction : Legal->getReductionVars()) { | ||||||||
7613 | NeedDef.insert(Reduction.first); | ||||||||
7614 | NeedDef.insert(Reduction.second.getLoopExitInstr()); | ||||||||
7615 | } | ||||||||
7616 | } | ||||||||
7617 | |||||||||
7618 | // Collect instructions from the original loop that will become trivially dead | ||||||||
7619 | // in the vectorized loop. We don't need to vectorize these instructions. For | ||||||||
7620 | // example, original induction update instructions can become dead because we | ||||||||
7621 | // separately emit induction "steps" when generating code for the new loop. | ||||||||
7622 | // Similarly, we create a new latch condition when setting up the structure | ||||||||
7623 | // of the new loop, so the old one can become dead. | ||||||||
7624 | SmallPtrSet<Instruction *, 4> DeadInstructions; | ||||||||
7625 | collectTriviallyDeadInstructions(DeadInstructions); | ||||||||
7626 | |||||||||
7627 | // Add assume instructions we need to drop to DeadInstructions, to prevent | ||||||||
7628 | // them from being added to the VPlan. | ||||||||
7629 | // TODO: We only need to drop assumes in blocks that get flattend. If the | ||||||||
7630 | // control flow is preserved, we should keep them. | ||||||||
7631 | auto &ConditionalAssumes = Legal->getConditionalAssumes(); | ||||||||
7632 | DeadInstructions.insert(ConditionalAssumes.begin(), ConditionalAssumes.end()); | ||||||||
7633 | |||||||||
7634 | DenseMap<Instruction *, Instruction *> &SinkAfter = Legal->getSinkAfter(); | ||||||||
7635 | // Dead instructions do not need sinking. Remove them from SinkAfter. | ||||||||
7636 | for (Instruction *I : DeadInstructions) | ||||||||
7637 | SinkAfter.erase(I); | ||||||||
7638 | |||||||||
7639 | for (unsigned VF = MinVF; VF < MaxVF + 1;) { | ||||||||
7640 | VFRange SubRange = {VF, MaxVF + 1}; | ||||||||
7641 | VPlans.push_back(buildVPlanWithVPRecipes(SubRange, NeedDef, | ||||||||
7642 | DeadInstructions, SinkAfter)); | ||||||||
7643 | VF = SubRange.End; | ||||||||
7644 | } | ||||||||
7645 | } | ||||||||
7646 | |||||||||
7647 | VPlanPtr LoopVectorizationPlanner::buildVPlanWithVPRecipes( | ||||||||
7648 | VFRange &Range, SmallPtrSetImpl<Value *> &NeedDef, | ||||||||
7649 | SmallPtrSetImpl<Instruction *> &DeadInstructions, | ||||||||
7650 | const DenseMap<Instruction *, Instruction *> &SinkAfter) { | ||||||||
7651 | |||||||||
7652 | // Hold a mapping from predicated instructions to their recipes, in order to | ||||||||
7653 | // fix their AlsoPack behavior if a user is determined to replicate and use a | ||||||||
7654 | // scalar instead of vector value. | ||||||||
7655 | DenseMap<Instruction *, VPReplicateRecipe *> PredInst2Recipe; | ||||||||
7656 | |||||||||
7657 | SmallPtrSet<const InterleaveGroup<Instruction> *, 1> InterleaveGroups; | ||||||||
7658 | |||||||||
7659 | VPRecipeBuilder RecipeBuilder(OrigLoop, TLI, Legal, CM, PSE, Builder); | ||||||||
7660 | |||||||||
7661 | // --------------------------------------------------------------------------- | ||||||||
7662 | // Pre-construction: record ingredients whose recipes we'll need to further | ||||||||
7663 | // process after constructing the initial VPlan. | ||||||||
7664 | // --------------------------------------------------------------------------- | ||||||||
7665 | |||||||||
7666 | // Mark instructions we'll need to sink later and their targets as | ||||||||
7667 | // ingredients whose recipe we'll need to record. | ||||||||
7668 | for (auto &Entry : SinkAfter) { | ||||||||
7669 | RecipeBuilder.recordRecipeOf(Entry.first); | ||||||||
7670 | RecipeBuilder.recordRecipeOf(Entry.second); | ||||||||
7671 | } | ||||||||
7672 | for (auto &Reduction : CM.getInLoopReductionChains()) { | ||||||||
7673 | PHINode *Phi = Reduction.first; | ||||||||
7674 | RecurrenceDescriptor::RecurrenceKind Kind = | ||||||||
7675 | Legal->getReductionVars()[Phi].getRecurrenceKind(); | ||||||||
7676 | const SmallVector<Instruction *, 4> &ReductionOperations = Reduction.second; | ||||||||
7677 | |||||||||
7678 | RecipeBuilder.recordRecipeOf(Phi); | ||||||||
7679 | for (auto &R : ReductionOperations) { | ||||||||
7680 | RecipeBuilder.recordRecipeOf(R); | ||||||||
7681 | // For min/max reducitons, where we have a pair of icmp/select, we also | ||||||||
7682 | // need to record the ICmp recipe, so it can be removed later. | ||||||||
7683 | if (Kind == RecurrenceDescriptor::RK_IntegerMinMax || | ||||||||
7684 | Kind == RecurrenceDescriptor::RK_FloatMinMax) { | ||||||||
7685 | RecipeBuilder.recordRecipeOf(cast<Instruction>(R->getOperand(0))); | ||||||||
7686 | } | ||||||||
7687 | } | ||||||||
7688 | } | ||||||||
7689 | |||||||||
7690 | // For each interleave group which is relevant for this (possibly trimmed) | ||||||||
7691 | // Range, add it to the set of groups to be later applied to the VPlan and add | ||||||||
7692 | // placeholders for its members' Recipes which we'll be replacing with a | ||||||||
7693 | // single VPInterleaveRecipe. | ||||||||
7694 | for (InterleaveGroup<Instruction> *IG : IAI.getInterleaveGroups()) { | ||||||||
7695 | auto applyIG = [IG, this](ElementCount VF) -> bool { | ||||||||
7696 | return (VF.isVector() && // Query is illegal for VF == 1 | ||||||||
7697 | CM.getWideningDecision(IG->getInsertPos(), VF) == | ||||||||
7698 | LoopVectorizationCostModel::CM_Interleave); | ||||||||
7699 | }; | ||||||||
7700 | if (!getDecisionAndClampRange(applyIG, Range)) | ||||||||
7701 | continue; | ||||||||
7702 | InterleaveGroups.insert(IG); | ||||||||
7703 | for (unsigned i = 0; i < IG->getFactor(); i++) | ||||||||
7704 | if (Instruction *Member = IG->getMember(i)) | ||||||||
7705 | RecipeBuilder.recordRecipeOf(Member); | ||||||||
7706 | }; | ||||||||
7707 | |||||||||
7708 | // --------------------------------------------------------------------------- | ||||||||
7709 | // Build initial VPlan: Scan the body of the loop in a topological order to | ||||||||
7710 | // visit each basic block after having visited its predecessor basic blocks. | ||||||||
7711 | // --------------------------------------------------------------------------- | ||||||||
7712 | |||||||||
7713 | // Create a dummy pre-entry VPBasicBlock to start building the VPlan. | ||||||||
7714 | auto Plan = std::make_unique<VPlan>(); | ||||||||
7715 | VPBasicBlock *VPBB = new VPBasicBlock("Pre-Entry"); | ||||||||
7716 | Plan->setEntry(VPBB); | ||||||||
7717 | |||||||||
7718 | // Represent values that will have defs inside VPlan. | ||||||||
7719 | for (Value *V : NeedDef) | ||||||||
7720 | Plan->addVPValue(V); | ||||||||
7721 | |||||||||
7722 | // Scan the body of the loop in a topological order to visit each basic block | ||||||||
7723 | // after having visited its predecessor basic blocks. | ||||||||
7724 | LoopBlocksDFS DFS(OrigLoop); | ||||||||
7725 | DFS.perform(LI); | ||||||||
7726 | |||||||||
7727 | for (BasicBlock *BB : make_range(DFS.beginRPO(), DFS.endRPO())) { | ||||||||
7728 | // Relevant instructions from basic block BB will be grouped into VPRecipe | ||||||||
7729 | // ingredients and fill a new VPBasicBlock. | ||||||||
7730 | unsigned VPBBsForBB = 0; | ||||||||
7731 | auto *FirstVPBBForBB = new VPBasicBlock(BB->getName()); | ||||||||
7732 | VPBlockUtils::insertBlockAfter(FirstVPBBForBB, VPBB); | ||||||||
7733 | VPBB = FirstVPBBForBB; | ||||||||
7734 | Builder.setInsertPoint(VPBB); | ||||||||
7735 | |||||||||
7736 | // Introduce each ingredient into VPlan. | ||||||||
7737 | // TODO: Model and preserve debug instrinsics in VPlan. | ||||||||
7738 | for (Instruction &I : BB->instructionsWithoutDebug()) { | ||||||||
7739 | Instruction *Instr = &I; | ||||||||
7740 | |||||||||
7741 | // First filter out irrelevant instructions, to ensure no recipes are | ||||||||
7742 | // built for them. | ||||||||
7743 | if (isa<BranchInst>(Instr) || DeadInstructions.count(Instr)) | ||||||||
7744 | continue; | ||||||||
7745 | |||||||||
7746 | if (auto Recipe = | ||||||||
7747 | RecipeBuilder.tryToCreateWidenRecipe(Instr, Range, Plan)) { | ||||||||
7748 | RecipeBuilder.setRecipe(Instr, Recipe); | ||||||||
7749 | VPBB->appendRecipe(Recipe); | ||||||||
7750 | continue; | ||||||||
7751 | } | ||||||||
7752 | |||||||||
7753 | // Otherwise, if all widening options failed, Instruction is to be | ||||||||
7754 | // replicated. This may create a successor for VPBB. | ||||||||
7755 | VPBasicBlock *NextVPBB = RecipeBuilder.handleReplication( | ||||||||
7756 | Instr, Range, VPBB, PredInst2Recipe, Plan); | ||||||||
7757 | if (NextVPBB != VPBB) { | ||||||||
7758 | VPBB = NextVPBB; | ||||||||
7759 | VPBB->setName(BB->hasName() ? BB->getName() + "." + Twine(VPBBsForBB++) | ||||||||
7760 | : ""); | ||||||||
7761 | } | ||||||||
7762 | } | ||||||||
7763 | } | ||||||||
7764 | |||||||||
7765 | // Discard empty dummy pre-entry VPBasicBlock. Note that other VPBasicBlocks | ||||||||
7766 | // may also be empty, such as the last one VPBB, reflecting original | ||||||||
7767 | // basic-blocks with no recipes. | ||||||||
7768 | VPBasicBlock *PreEntry = cast<VPBasicBlock>(Plan->getEntry()); | ||||||||
7769 | assert(PreEntry->empty() && "Expecting empty pre-entry block.")((PreEntry->empty() && "Expecting empty pre-entry block." ) ? static_cast<void> (0) : __assert_fail ("PreEntry->empty() && \"Expecting empty pre-entry block.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 7769, __PRETTY_FUNCTION__)); | ||||||||
7770 | VPBlockBase *Entry = Plan->setEntry(PreEntry->getSingleSuccessor()); | ||||||||
7771 | VPBlockUtils::disconnectBlocks(PreEntry, Entry); | ||||||||
7772 | delete PreEntry; | ||||||||
7773 | |||||||||
7774 | // --------------------------------------------------------------------------- | ||||||||
7775 | // Transform initial VPlan: Apply previously taken decisions, in order, to | ||||||||
7776 | // bring the VPlan to its final state. | ||||||||
7777 | // --------------------------------------------------------------------------- | ||||||||
7778 | |||||||||
7779 | // Apply Sink-After legal constraints. | ||||||||
7780 | for (auto &Entry : SinkAfter) { | ||||||||
7781 | VPRecipeBase *Sink = RecipeBuilder.getRecipe(Entry.first); | ||||||||
7782 | VPRecipeBase *Target = RecipeBuilder.getRecipe(Entry.second); | ||||||||
7783 | Sink->moveAfter(Target); | ||||||||
7784 | } | ||||||||
7785 | |||||||||
7786 | // Interleave memory: for each Interleave Group we marked earlier as relevant | ||||||||
7787 | // for this VPlan, replace the Recipes widening its memory instructions with a | ||||||||
7788 | // single VPInterleaveRecipe at its insertion point. | ||||||||
7789 | for (auto IG : InterleaveGroups) { | ||||||||
7790 | auto *Recipe = cast<VPWidenMemoryInstructionRecipe>( | ||||||||
7791 | RecipeBuilder.getRecipe(IG->getInsertPos())); | ||||||||
7792 | (new VPInterleaveRecipe(IG, Recipe->getAddr(), Recipe->getMask())) | ||||||||
7793 | ->insertBefore(Recipe); | ||||||||
7794 | |||||||||
7795 | for (unsigned i = 0; i < IG->getFactor(); ++i) | ||||||||
7796 | if (Instruction *Member = IG->getMember(i)) { | ||||||||
7797 | RecipeBuilder.getRecipe(Member)->eraseFromParent(); | ||||||||
7798 | } | ||||||||
7799 | } | ||||||||
7800 | |||||||||
7801 | // Adjust the recipes for any inloop reductions. | ||||||||
7802 | if (Range.Start > 1) | ||||||||
7803 | adjustRecipesForInLoopReductions(Plan, RecipeBuilder); | ||||||||
7804 | |||||||||
7805 | // Finally, if tail is folded by masking, introduce selects between the phi | ||||||||
7806 | // and the live-out instruction of each reduction, at the end of the latch. | ||||||||
7807 | if (CM.foldTailByMasking() && !Legal->getReductionVars().empty()) { | ||||||||
7808 | Builder.setInsertPoint(VPBB); | ||||||||
7809 | auto *Cond = RecipeBuilder.createBlockInMask(OrigLoop->getHeader(), Plan); | ||||||||
7810 | for (auto &Reduction : Legal->getReductionVars()) { | ||||||||
7811 | assert(!CM.isInLoopReduction(Reduction.first) &&((!CM.isInLoopReduction(Reduction.first) && "Didn't expect inloop tail folded reduction yet!" ) ? static_cast<void> (0) : __assert_fail ("!CM.isInLoopReduction(Reduction.first) && \"Didn't expect inloop tail folded reduction yet!\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 7812, __PRETTY_FUNCTION__)) | ||||||||
7812 | "Didn't expect inloop tail folded reduction yet!")((!CM.isInLoopReduction(Reduction.first) && "Didn't expect inloop tail folded reduction yet!" ) ? static_cast<void> (0) : __assert_fail ("!CM.isInLoopReduction(Reduction.first) && \"Didn't expect inloop tail folded reduction yet!\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 7812, __PRETTY_FUNCTION__)); | ||||||||
7813 | VPValue *Phi = Plan->getVPValue(Reduction.first); | ||||||||
7814 | VPValue *Red = Plan->getVPValue(Reduction.second.getLoopExitInstr()); | ||||||||
7815 | Builder.createNaryOp(Instruction::Select, {Cond, Red, Phi}); | ||||||||
7816 | } | ||||||||
7817 | } | ||||||||
7818 | |||||||||
7819 | std::string PlanName; | ||||||||
7820 | raw_string_ostream RSO(PlanName); | ||||||||
7821 | ElementCount VF = ElementCount::getFixed(Range.Start); | ||||||||
7822 | Plan->addVF(VF); | ||||||||
7823 | RSO << "Initial VPlan for VF={" << VF; | ||||||||
7824 | for (VF *= 2; VF.getKnownMinValue() < Range.End; VF *= 2) { | ||||||||
7825 | Plan->addVF(VF); | ||||||||
7826 | RSO << "," << VF; | ||||||||
7827 | } | ||||||||
7828 | RSO << "},UF>=1"; | ||||||||
7829 | RSO.flush(); | ||||||||
7830 | Plan->setName(PlanName); | ||||||||
7831 | |||||||||
7832 | return Plan; | ||||||||
7833 | } | ||||||||
7834 | |||||||||
7835 | VPlanPtr LoopVectorizationPlanner::buildVPlan(VFRange &Range) { | ||||||||
7836 | // Outer loop handling: They may require CFG and instruction level | ||||||||
7837 | // transformations before even evaluating whether vectorization is profitable. | ||||||||
7838 | // Since we cannot modify the incoming IR, we need to build VPlan upfront in | ||||||||
7839 | // the vectorization pipeline. | ||||||||
7840 | assert(!OrigLoop->isInnermost())((!OrigLoop->isInnermost()) ? static_cast<void> (0) : __assert_fail ("!OrigLoop->isInnermost()", "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 7840, __PRETTY_FUNCTION__)); | ||||||||
7841 | assert(EnableVPlanNativePath && "VPlan-native path is not enabled.")((EnableVPlanNativePath && "VPlan-native path is not enabled." ) ? static_cast<void> (0) : __assert_fail ("EnableVPlanNativePath && \"VPlan-native path is not enabled.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 7841, __PRETTY_FUNCTION__)); | ||||||||
7842 | |||||||||
7843 | // Create new empty VPlan | ||||||||
7844 | auto Plan = std::make_unique<VPlan>(); | ||||||||
7845 | |||||||||
7846 | // Build hierarchical CFG | ||||||||
7847 | VPlanHCFGBuilder HCFGBuilder(OrigLoop, LI, *Plan); | ||||||||
7848 | HCFGBuilder.buildHierarchicalCFG(); | ||||||||
7849 | |||||||||
7850 | for (unsigned VF = Range.Start; VF < Range.End; VF *= 2) | ||||||||
7851 | Plan->addVF(ElementCount::getFixed(VF)); | ||||||||
7852 | |||||||||
7853 | if (EnableVPlanPredication) { | ||||||||
7854 | VPlanPredicator VPP(*Plan); | ||||||||
7855 | VPP.predicate(); | ||||||||
7856 | |||||||||
7857 | // Avoid running transformation to recipes until masked code generation in | ||||||||
7858 | // VPlan-native path is in place. | ||||||||
7859 | return Plan; | ||||||||
7860 | } | ||||||||
7861 | |||||||||
7862 | SmallPtrSet<Instruction *, 1> DeadInstructions; | ||||||||
7863 | VPlanTransforms::VPInstructionsToVPRecipes( | ||||||||
7864 | OrigLoop, Plan, Legal->getInductionVars(), DeadInstructions); | ||||||||
7865 | return Plan; | ||||||||
7866 | } | ||||||||
7867 | |||||||||
7868 | // Adjust the recipes for any inloop reductions. The chain of instructions | ||||||||
7869 | // leading from the loop exit instr to the phi need to be converted to | ||||||||
7870 | // reductions, with one operand being vector and the other being the scalar | ||||||||
7871 | // reduction chain. | ||||||||
7872 | void LoopVectorizationPlanner::adjustRecipesForInLoopReductions( | ||||||||
7873 | VPlanPtr &Plan, VPRecipeBuilder &RecipeBuilder) { | ||||||||
7874 | for (auto &Reduction : CM.getInLoopReductionChains()) { | ||||||||
7875 | PHINode *Phi = Reduction.first; | ||||||||
7876 | RecurrenceDescriptor &RdxDesc = Legal->getReductionVars()[Phi]; | ||||||||
7877 | const SmallVector<Instruction *, 4> &ReductionOperations = Reduction.second; | ||||||||
7878 | |||||||||
7879 | // ReductionOperations are orders top-down from the phi's use to the | ||||||||
7880 | // LoopExitValue. We keep a track of the previous item (the Chain) to tell | ||||||||
7881 | // which of the two operands will remain scalar and which will be reduced. | ||||||||
7882 | // For minmax the chain will be the select instructions. | ||||||||
7883 | Instruction *Chain = Phi; | ||||||||
7884 | for (Instruction *R : ReductionOperations) { | ||||||||
7885 | VPRecipeBase *WidenRecipe = RecipeBuilder.getRecipe(R); | ||||||||
7886 | RecurrenceDescriptor::RecurrenceKind Kind = RdxDesc.getRecurrenceKind(); | ||||||||
7887 | |||||||||
7888 | VPValue *ChainOp = Plan->getVPValue(Chain); | ||||||||
7889 | unsigned FirstOpId; | ||||||||
7890 | if (Kind == RecurrenceDescriptor::RK_IntegerMinMax || | ||||||||
7891 | Kind == RecurrenceDescriptor::RK_FloatMinMax) { | ||||||||
7892 | assert(WidenRecipe->getVPRecipeID() == VPRecipeBase::VPWidenSelectSC &&((WidenRecipe->getVPRecipeID() == VPRecipeBase::VPWidenSelectSC && "Expected to replace a VPWidenSelectSC") ? static_cast <void> (0) : __assert_fail ("WidenRecipe->getVPRecipeID() == VPRecipeBase::VPWidenSelectSC && \"Expected to replace a VPWidenSelectSC\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 7893, __PRETTY_FUNCTION__)) | ||||||||
7893 | "Expected to replace a VPWidenSelectSC")((WidenRecipe->getVPRecipeID() == VPRecipeBase::VPWidenSelectSC && "Expected to replace a VPWidenSelectSC") ? static_cast <void> (0) : __assert_fail ("WidenRecipe->getVPRecipeID() == VPRecipeBase::VPWidenSelectSC && \"Expected to replace a VPWidenSelectSC\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 7893, __PRETTY_FUNCTION__)); | ||||||||
7894 | FirstOpId = 1; | ||||||||
7895 | } else { | ||||||||
7896 | assert(WidenRecipe->getVPRecipeID() == VPRecipeBase::VPWidenSC &&((WidenRecipe->getVPRecipeID() == VPRecipeBase::VPWidenSC && "Expected to replace a VPWidenSC") ? static_cast<void> (0) : __assert_fail ("WidenRecipe->getVPRecipeID() == VPRecipeBase::VPWidenSC && \"Expected to replace a VPWidenSC\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 7897, __PRETTY_FUNCTION__)) | ||||||||
7897 | "Expected to replace a VPWidenSC")((WidenRecipe->getVPRecipeID() == VPRecipeBase::VPWidenSC && "Expected to replace a VPWidenSC") ? static_cast<void> (0) : __assert_fail ("WidenRecipe->getVPRecipeID() == VPRecipeBase::VPWidenSC && \"Expected to replace a VPWidenSC\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 7897, __PRETTY_FUNCTION__)); | ||||||||
7898 | FirstOpId = 0; | ||||||||
7899 | } | ||||||||
7900 | unsigned VecOpId = | ||||||||
7901 | R->getOperand(FirstOpId) == Chain ? FirstOpId + 1 : FirstOpId; | ||||||||
7902 | VPValue *VecOp = Plan->getVPValue(R->getOperand(VecOpId)); | ||||||||
7903 | |||||||||
7904 | VPReductionRecipe *RedRecipe = new VPReductionRecipe( | ||||||||
7905 | &RdxDesc, R, ChainOp, VecOp, Legal->hasFunNoNaNAttr(), TTI); | ||||||||
7906 | WidenRecipe->getParent()->insert(RedRecipe, WidenRecipe->getIterator()); | ||||||||
7907 | WidenRecipe->eraseFromParent(); | ||||||||
7908 | |||||||||
7909 | if (Kind == RecurrenceDescriptor::RK_IntegerMinMax || | ||||||||
7910 | Kind == RecurrenceDescriptor::RK_FloatMinMax) { | ||||||||
7911 | VPRecipeBase *CompareRecipe = | ||||||||
7912 | RecipeBuilder.getRecipe(cast<Instruction>(R->getOperand(0))); | ||||||||
7913 | assert(CompareRecipe->getVPRecipeID() == VPRecipeBase::VPWidenSC &&((CompareRecipe->getVPRecipeID() == VPRecipeBase::VPWidenSC && "Expected to replace a VPWidenSC") ? static_cast< void> (0) : __assert_fail ("CompareRecipe->getVPRecipeID() == VPRecipeBase::VPWidenSC && \"Expected to replace a VPWidenSC\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 7914, __PRETTY_FUNCTION__)) | ||||||||
7914 | "Expected to replace a VPWidenSC")((CompareRecipe->getVPRecipeID() == VPRecipeBase::VPWidenSC && "Expected to replace a VPWidenSC") ? static_cast< void> (0) : __assert_fail ("CompareRecipe->getVPRecipeID() == VPRecipeBase::VPWidenSC && \"Expected to replace a VPWidenSC\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 7914, __PRETTY_FUNCTION__)); | ||||||||
7915 | CompareRecipe->eraseFromParent(); | ||||||||
7916 | } | ||||||||
7917 | Chain = R; | ||||||||
7918 | } | ||||||||
7919 | } | ||||||||
7920 | } | ||||||||
7921 | |||||||||
7922 | Value* LoopVectorizationPlanner::VPCallbackILV:: | ||||||||
7923 | getOrCreateVectorValues(Value *V, unsigned Part) { | ||||||||
7924 | return ILV.getOrCreateVectorValue(V, Part); | ||||||||
7925 | } | ||||||||
7926 | |||||||||
7927 | Value *LoopVectorizationPlanner::VPCallbackILV::getOrCreateScalarValue( | ||||||||
7928 | Value *V, const VPIteration &Instance) { | ||||||||
7929 | return ILV.getOrCreateScalarValue(V, Instance); | ||||||||
7930 | } | ||||||||
7931 | |||||||||
7932 | void VPInterleaveRecipe::print(raw_ostream &O, const Twine &Indent, | ||||||||
7933 | VPSlotTracker &SlotTracker) const { | ||||||||
7934 | O << "\"INTERLEAVE-GROUP with factor " << IG->getFactor() << " at "; | ||||||||
7935 | IG->getInsertPos()->printAsOperand(O, false); | ||||||||
7936 | O << ", "; | ||||||||
7937 | getAddr()->printAsOperand(O, SlotTracker); | ||||||||
7938 | VPValue *Mask = getMask(); | ||||||||
7939 | if (Mask) { | ||||||||
7940 | O << ", "; | ||||||||
7941 | Mask->printAsOperand(O, SlotTracker); | ||||||||
7942 | } | ||||||||
7943 | for (unsigned i = 0; i < IG->getFactor(); ++i) | ||||||||
7944 | if (Instruction *I = IG->getMember(i)) | ||||||||
7945 | O << "\\l\" +\n" << Indent << "\" " << VPlanIngredient(I) << " " << i; | ||||||||
7946 | } | ||||||||
7947 | |||||||||
7948 | void VPWidenCallRecipe::execute(VPTransformState &State) { | ||||||||
7949 | State.ILV->widenCallInstruction(Ingredient, User, State); | ||||||||
7950 | } | ||||||||
7951 | |||||||||
7952 | void VPWidenSelectRecipe::execute(VPTransformState &State) { | ||||||||
7953 | State.ILV->widenSelectInstruction(Ingredient, User, InvariantCond, State); | ||||||||
7954 | } | ||||||||
7955 | |||||||||
7956 | void VPWidenRecipe::execute(VPTransformState &State) { | ||||||||
7957 | State.ILV->widenInstruction(Ingredient, User, State); | ||||||||
7958 | } | ||||||||
7959 | |||||||||
7960 | void VPWidenGEPRecipe::execute(VPTransformState &State) { | ||||||||
7961 | State.ILV->widenGEP(GEP, User, State.UF, State.VF, IsPtrLoopInvariant, | ||||||||
7962 | IsIndexLoopInvariant, State); | ||||||||
7963 | } | ||||||||
7964 | |||||||||
7965 | void VPWidenIntOrFpInductionRecipe::execute(VPTransformState &State) { | ||||||||
7966 | assert(!State.Instance && "Int or FP induction being replicated.")((!State.Instance && "Int or FP induction being replicated." ) ? static_cast<void> (0) : __assert_fail ("!State.Instance && \"Int or FP induction being replicated.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 7966, __PRETTY_FUNCTION__)); | ||||||||
7967 | State.ILV->widenIntOrFpInduction(IV, Trunc); | ||||||||
7968 | } | ||||||||
7969 | |||||||||
7970 | void VPWidenPHIRecipe::execute(VPTransformState &State) { | ||||||||
7971 | State.ILV->widenPHIInstruction(Phi, State.UF, State.VF); | ||||||||
7972 | } | ||||||||
7973 | |||||||||
7974 | void VPBlendRecipe::execute(VPTransformState &State) { | ||||||||
7975 | State.ILV->setDebugLocFromInst(State.Builder, Phi); | ||||||||
7976 | // We know that all PHIs in non-header blocks are converted into | ||||||||
7977 | // selects, so we don't have to worry about the insertion order and we | ||||||||
7978 | // can just use the builder. | ||||||||
7979 | // At this point we generate the predication tree. There may be | ||||||||
7980 | // duplications since this is a simple recursive scan, but future | ||||||||
7981 | // optimizations will clean it up. | ||||||||
7982 | |||||||||
7983 | unsigned NumIncoming = getNumIncomingValues(); | ||||||||
7984 | |||||||||
7985 | // Generate a sequence of selects of the form: | ||||||||
7986 | // SELECT(Mask3, In3, | ||||||||
7987 | // SELECT(Mask2, In2, | ||||||||
7988 | // SELECT(Mask1, In1, | ||||||||
7989 | // In0))) | ||||||||
7990 | // Note that Mask0 is never used: lanes for which no path reaches this phi and | ||||||||
7991 | // are essentially undef are taken from In0. | ||||||||
7992 | InnerLoopVectorizer::VectorParts Entry(State.UF); | ||||||||
7993 | for (unsigned In = 0; In < NumIncoming; ++In) { | ||||||||
7994 | for (unsigned Part = 0; Part < State.UF; ++Part) { | ||||||||
7995 | // We might have single edge PHIs (blocks) - use an identity | ||||||||
7996 | // 'select' for the first PHI operand. | ||||||||
7997 | Value *In0 = State.get(getIncomingValue(In), Part); | ||||||||
7998 | if (In == 0) | ||||||||
7999 | Entry[Part] = In0; // Initialize with the first incoming value. | ||||||||
8000 | else { | ||||||||
8001 | // Select between the current value and the previous incoming edge | ||||||||
8002 | // based on the incoming mask. | ||||||||
8003 | Value *Cond = State.get(getMask(In), Part); | ||||||||
8004 | Entry[Part] = | ||||||||
8005 | State.Builder.CreateSelect(Cond, In0, Entry[Part], "predphi"); | ||||||||
8006 | } | ||||||||
8007 | } | ||||||||
8008 | } | ||||||||
8009 | for (unsigned Part = 0; Part < State.UF; ++Part) | ||||||||
8010 | State.ValueMap.setVectorValue(Phi, Part, Entry[Part]); | ||||||||
8011 | } | ||||||||
8012 | |||||||||
8013 | void VPInterleaveRecipe::execute(VPTransformState &State) { | ||||||||
8014 | assert(!State.Instance && "Interleave group being replicated.")((!State.Instance && "Interleave group being replicated." ) ? static_cast<void> (0) : __assert_fail ("!State.Instance && \"Interleave group being replicated.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 8014, __PRETTY_FUNCTION__)); | ||||||||
8015 | State.ILV->vectorizeInterleaveGroup(IG, State, getAddr(), getMask()); | ||||||||
8016 | } | ||||||||
8017 | |||||||||
8018 | void VPReductionRecipe::execute(VPTransformState &State) { | ||||||||
8019 | assert(!State.Instance && "Reduction being replicated.")((!State.Instance && "Reduction being replicated.") ? static_cast<void> (0) : __assert_fail ("!State.Instance && \"Reduction being replicated.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 8019, __PRETTY_FUNCTION__)); | ||||||||
8020 | for (unsigned Part = 0; Part < State.UF; ++Part) { | ||||||||
8021 | unsigned Kind = RdxDesc->getRecurrenceKind(); | ||||||||
8022 | Value *NewVecOp = State.get(VecOp, Part); | ||||||||
8023 | Value *NewRed = | ||||||||
8024 | createTargetReduction(State.Builder, TTI, *RdxDesc, NewVecOp, NoNaN); | ||||||||
8025 | Value *PrevInChain = State.get(ChainOp, Part); | ||||||||
8026 | Value *NextInChain; | ||||||||
8027 | if (Kind == RecurrenceDescriptor::RK_IntegerMinMax || | ||||||||
8028 | Kind == RecurrenceDescriptor::RK_FloatMinMax) { | ||||||||
8029 | NextInChain = | ||||||||
8030 | createMinMaxOp(State.Builder, RdxDesc->getMinMaxRecurrenceKind(), | ||||||||
8031 | NewRed, PrevInChain); | ||||||||
8032 | } else { | ||||||||
8033 | NextInChain = State.Builder.CreateBinOp( | ||||||||
8034 | (Instruction::BinaryOps)I->getOpcode(), NewRed, PrevInChain); | ||||||||
8035 | } | ||||||||
8036 | State.ValueMap.setVectorValue(I, Part, NextInChain); | ||||||||
8037 | } | ||||||||
8038 | } | ||||||||
8039 | |||||||||
8040 | void VPReplicateRecipe::execute(VPTransformState &State) { | ||||||||
8041 | if (State.Instance) { // Generate a single instance. | ||||||||
8042 | State.ILV->scalarizeInstruction(Ingredient, User, *State.Instance, | ||||||||
8043 | IsPredicated, State); | ||||||||
8044 | // Insert scalar instance packing it into a vector. | ||||||||
8045 | if (AlsoPack && State.VF.isVector()) { | ||||||||
8046 | // If we're constructing lane 0, initialize to start from undef. | ||||||||
8047 | if (State.Instance->Lane == 0) { | ||||||||
8048 | assert(!State.VF.isScalable() && "VF is assumed to be non scalable.")((!State.VF.isScalable() && "VF is assumed to be non scalable." ) ? static_cast<void> (0) : __assert_fail ("!State.VF.isScalable() && \"VF is assumed to be non scalable.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 8048, __PRETTY_FUNCTION__)); | ||||||||
8049 | Value *Undef = | ||||||||
8050 | UndefValue::get(VectorType::get(Ingredient->getType(), State.VF)); | ||||||||
8051 | State.ValueMap.setVectorValue(Ingredient, State.Instance->Part, Undef); | ||||||||
8052 | } | ||||||||
8053 | State.ILV->packScalarIntoVectorValue(Ingredient, *State.Instance); | ||||||||
8054 | } | ||||||||
8055 | return; | ||||||||
8056 | } | ||||||||
8057 | |||||||||
8058 | // Generate scalar instances for all VF lanes of all UF parts, unless the | ||||||||
8059 | // instruction is uniform inwhich case generate only the first lane for each | ||||||||
8060 | // of the UF parts. | ||||||||
8061 | unsigned EndLane = IsUniform ? 1 : State.VF.getKnownMinValue(); | ||||||||
8062 | for (unsigned Part = 0; Part < State.UF; ++Part) | ||||||||
8063 | for (unsigned Lane = 0; Lane < EndLane; ++Lane) | ||||||||
8064 | State.ILV->scalarizeInstruction(Ingredient, User, {Part, Lane}, | ||||||||
8065 | IsPredicated, State); | ||||||||
8066 | } | ||||||||
8067 | |||||||||
8068 | void VPBranchOnMaskRecipe::execute(VPTransformState &State) { | ||||||||
8069 | assert(State.Instance && "Branch on Mask works only on single instance.")((State.Instance && "Branch on Mask works only on single instance." ) ? static_cast<void> (0) : __assert_fail ("State.Instance && \"Branch on Mask works only on single instance.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 8069, __PRETTY_FUNCTION__)); | ||||||||
8070 | |||||||||
8071 | unsigned Part = State.Instance->Part; | ||||||||
8072 | unsigned Lane = State.Instance->Lane; | ||||||||
8073 | |||||||||
8074 | Value *ConditionBit = nullptr; | ||||||||
8075 | VPValue *BlockInMask = getMask(); | ||||||||
8076 | if (BlockInMask) { | ||||||||
8077 | ConditionBit = State.get(BlockInMask, Part); | ||||||||
8078 | if (ConditionBit->getType()->isVectorTy()) | ||||||||
8079 | ConditionBit = State.Builder.CreateExtractElement( | ||||||||
8080 | ConditionBit, State.Builder.getInt32(Lane)); | ||||||||
8081 | } else // Block in mask is all-one. | ||||||||
8082 | ConditionBit = State.Builder.getTrue(); | ||||||||
8083 | |||||||||
8084 | // Replace the temporary unreachable terminator with a new conditional branch, | ||||||||
8085 | // whose two destinations will be set later when they are created. | ||||||||
8086 | auto *CurrentTerminator = State.CFG.PrevBB->getTerminator(); | ||||||||
8087 | assert(isa<UnreachableInst>(CurrentTerminator) &&((isa<UnreachableInst>(CurrentTerminator) && "Expected to replace unreachable terminator with conditional branch." ) ? static_cast<void> (0) : __assert_fail ("isa<UnreachableInst>(CurrentTerminator) && \"Expected to replace unreachable terminator with conditional branch.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 8088, __PRETTY_FUNCTION__)) | ||||||||
8088 | "Expected to replace unreachable terminator with conditional branch.")((isa<UnreachableInst>(CurrentTerminator) && "Expected to replace unreachable terminator with conditional branch." ) ? static_cast<void> (0) : __assert_fail ("isa<UnreachableInst>(CurrentTerminator) && \"Expected to replace unreachable terminator with conditional branch.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 8088, __PRETTY_FUNCTION__)); | ||||||||
8089 | auto *CondBr = BranchInst::Create(State.CFG.PrevBB, nullptr, ConditionBit); | ||||||||
8090 | CondBr->setSuccessor(0, nullptr); | ||||||||
8091 | ReplaceInstWithInst(CurrentTerminator, CondBr); | ||||||||
8092 | } | ||||||||
8093 | |||||||||
8094 | void VPPredInstPHIRecipe::execute(VPTransformState &State) { | ||||||||
8095 | assert(State.Instance && "Predicated instruction PHI works per instance.")((State.Instance && "Predicated instruction PHI works per instance." ) ? static_cast<void> (0) : __assert_fail ("State.Instance && \"Predicated instruction PHI works per instance.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 8095, __PRETTY_FUNCTION__)); | ||||||||
8096 | Instruction *ScalarPredInst = cast<Instruction>( | ||||||||
8097 | State.ValueMap.getScalarValue(PredInst, *State.Instance)); | ||||||||
8098 | BasicBlock *PredicatedBB = ScalarPredInst->getParent(); | ||||||||
8099 | BasicBlock *PredicatingBB = PredicatedBB->getSinglePredecessor(); | ||||||||
8100 | assert(PredicatingBB && "Predicated block has no single predecessor.")((PredicatingBB && "Predicated block has no single predecessor." ) ? static_cast<void> (0) : __assert_fail ("PredicatingBB && \"Predicated block has no single predecessor.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 8100, __PRETTY_FUNCTION__)); | ||||||||
8101 | |||||||||
8102 | // By current pack/unpack logic we need to generate only a single phi node: if | ||||||||
8103 | // a vector value for the predicated instruction exists at this point it means | ||||||||
8104 | // the instruction has vector users only, and a phi for the vector value is | ||||||||
8105 | // needed. In this case the recipe of the predicated instruction is marked to | ||||||||
8106 | // also do that packing, thereby "hoisting" the insert-element sequence. | ||||||||
8107 | // Otherwise, a phi node for the scalar value is needed. | ||||||||
8108 | unsigned Part = State.Instance->Part; | ||||||||
8109 | if (State.ValueMap.hasVectorValue(PredInst, Part)) { | ||||||||
8110 | Value *VectorValue = State.ValueMap.getVectorValue(PredInst, Part); | ||||||||
8111 | InsertElementInst *IEI = cast<InsertElementInst>(VectorValue); | ||||||||
8112 | PHINode *VPhi = State.Builder.CreatePHI(IEI->getType(), 2); | ||||||||
8113 | VPhi->addIncoming(IEI->getOperand(0), PredicatingBB); // Unmodified vector. | ||||||||
8114 | VPhi->addIncoming(IEI, PredicatedBB); // New vector with inserted element. | ||||||||
8115 | State.ValueMap.resetVectorValue(PredInst, Part, VPhi); // Update cache. | ||||||||
8116 | } else { | ||||||||
8117 | Type *PredInstType = PredInst->getType(); | ||||||||
8118 | PHINode *Phi = State.Builder.CreatePHI(PredInstType, 2); | ||||||||
8119 | Phi->addIncoming(UndefValue::get(ScalarPredInst->getType()), PredicatingBB); | ||||||||
8120 | Phi->addIncoming(ScalarPredInst, PredicatedBB); | ||||||||
8121 | State.ValueMap.resetScalarValue(PredInst, *State.Instance, Phi); | ||||||||
8122 | } | ||||||||
8123 | } | ||||||||
8124 | |||||||||
8125 | void VPWidenMemoryInstructionRecipe::execute(VPTransformState &State) { | ||||||||
8126 | VPValue *StoredValue = isa<StoreInst>(Instr) ? getStoredValue() : nullptr; | ||||||||
8127 | State.ILV->vectorizeMemoryInstruction(&Instr, State, getAddr(), StoredValue, | ||||||||
8128 | getMask()); | ||||||||
8129 | } | ||||||||
8130 | |||||||||
8131 | // Determine how to lower the scalar epilogue, which depends on 1) optimising | ||||||||
8132 | // for minimum code-size, 2) predicate compiler options, 3) loop hints forcing | ||||||||
8133 | // predication, and 4) a TTI hook that analyses whether the loop is suitable | ||||||||
8134 | // for predication. | ||||||||
8135 | static ScalarEpilogueLowering getScalarEpilogueLowering( | ||||||||
8136 | Function *F, Loop *L, LoopVectorizeHints &Hints, ProfileSummaryInfo *PSI, | ||||||||
8137 | BlockFrequencyInfo *BFI, TargetTransformInfo *TTI, TargetLibraryInfo *TLI, | ||||||||
8138 | AssumptionCache *AC, LoopInfo *LI, ScalarEvolution *SE, DominatorTree *DT, | ||||||||
8139 | LoopVectorizationLegality &LVL) { | ||||||||
8140 | // 1) OptSize takes precedence over all other options, i.e. if this is set, | ||||||||
8141 | // don't look at hints or options, and don't request a scalar epilogue. | ||||||||
8142 | // (For PGSO, as shouldOptimizeForSize isn't currently accessible from | ||||||||
8143 | // LoopAccessInfo (due to code dependency and not being able to reliably get | ||||||||
8144 | // PSI/BFI from a loop analysis under NPM), we cannot suppress the collection | ||||||||
8145 | // of strides in LoopAccessInfo::analyzeLoop() and vectorize without | ||||||||
8146 | // versioning when the vectorization is forced, unlike hasOptSize. So revert | ||||||||
8147 | // back to the old way and vectorize with versioning when forced. See D81345.) | ||||||||
8148 | if (F->hasOptSize() || (llvm::shouldOptimizeForSize(L->getHeader(), PSI, BFI, | ||||||||
8149 | PGSOQueryType::IRPass) && | ||||||||
8150 | Hints.getForce() != LoopVectorizeHints::FK_Enabled)) | ||||||||
8151 | return CM_ScalarEpilogueNotAllowedOptSize; | ||||||||
8152 | |||||||||
8153 | bool PredicateOptDisabled = PreferPredicateOverEpilogue.getNumOccurrences() && | ||||||||
8154 | !PreferPredicateOverEpilogue; | ||||||||
8155 | |||||||||
8156 | // 2) Next, if disabling predication is requested on the command line, honour | ||||||||
8157 | // this and request a scalar epilogue. | ||||||||
8158 | if (PredicateOptDisabled) | ||||||||
8159 | return CM_ScalarEpilogueAllowed; | ||||||||
8160 | |||||||||
8161 | // 3) and 4) look if enabling predication is requested on the command line, | ||||||||
8162 | // with a loop hint, or if the TTI hook indicates this is profitable, request | ||||||||
8163 | // predication. | ||||||||
8164 | if (PreferPredicateOverEpilogue || | ||||||||
8165 | Hints.getPredicate() == LoopVectorizeHints::FK_Enabled || | ||||||||
8166 | (TTI->preferPredicateOverEpilogue(L, LI, *SE, *AC, TLI, DT, | ||||||||
8167 | LVL.getLAI()) && | ||||||||
8168 | Hints.getPredicate() != LoopVectorizeHints::FK_Disabled)) | ||||||||
8169 | return CM_ScalarEpilogueNotNeededUsePredicate; | ||||||||
8170 | |||||||||
8171 | return CM_ScalarEpilogueAllowed; | ||||||||
8172 | } | ||||||||
8173 | |||||||||
8174 | // Process the loop in the VPlan-native vectorization path. This path builds | ||||||||
8175 | // VPlan upfront in the vectorization pipeline, which allows to apply | ||||||||
8176 | // VPlan-to-VPlan transformations from the very beginning without modifying the | ||||||||
8177 | // input LLVM IR. | ||||||||
8178 | static bool processLoopInVPlanNativePath( | ||||||||
8179 | Loop *L, PredicatedScalarEvolution &PSE, LoopInfo *LI, DominatorTree *DT, | ||||||||
8180 | LoopVectorizationLegality *LVL, TargetTransformInfo *TTI, | ||||||||
8181 | TargetLibraryInfo *TLI, DemandedBits *DB, AssumptionCache *AC, | ||||||||
8182 | OptimizationRemarkEmitter *ORE, BlockFrequencyInfo *BFI, | ||||||||
8183 | ProfileSummaryInfo *PSI, LoopVectorizeHints &Hints) { | ||||||||
8184 | |||||||||
8185 | if (PSE.getBackedgeTakenCount() == PSE.getSE()->getCouldNotCompute()) { | ||||||||
8186 | 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); | ||||||||
8187 | return false; | ||||||||
8188 | } | ||||||||
8189 | assert(EnableVPlanNativePath && "VPlan-native path is disabled.")((EnableVPlanNativePath && "VPlan-native path is disabled." ) ? static_cast<void> (0) : __assert_fail ("EnableVPlanNativePath && \"VPlan-native path is disabled.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 8189, __PRETTY_FUNCTION__)); | ||||||||
8190 | Function *F = L->getHeader()->getParent(); | ||||||||
8191 | InterleavedAccessInfo IAI(PSE, L, DT, LI, LVL->getLAI()); | ||||||||
8192 | |||||||||
8193 | ScalarEpilogueLowering SEL = getScalarEpilogueLowering( | ||||||||
8194 | F, L, Hints, PSI, BFI, TTI, TLI, AC, LI, PSE.getSE(), DT, *LVL); | ||||||||
8195 | |||||||||
8196 | LoopVectorizationCostModel CM(SEL, L, PSE, LI, LVL, *TTI, TLI, DB, AC, ORE, F, | ||||||||
8197 | &Hints, IAI); | ||||||||
8198 | // Use the planner for outer loop vectorization. | ||||||||
8199 | // TODO: CM is not used at this point inside the planner. Turn CM into an | ||||||||
8200 | // optional argument if we don't need it in the future. | ||||||||
8201 | LoopVectorizationPlanner LVP(L, LI, TLI, TTI, LVL, CM, IAI, PSE); | ||||||||
8202 | |||||||||
8203 | // Get user vectorization factor. | ||||||||
8204 | const unsigned UserVF = Hints.getWidth(); | ||||||||
8205 | |||||||||
8206 | // Plan how to best vectorize, return the best VF and its cost. | ||||||||
8207 | const VectorizationFactor VF = | ||||||||
8208 | LVP.planInVPlanNativePath(ElementCount::getFixed(UserVF)); | ||||||||
8209 | |||||||||
8210 | // If we are stress testing VPlan builds, do not attempt to generate vector | ||||||||
8211 | // code. Masked vector code generation support will follow soon. | ||||||||
8212 | // Also, do not attempt to vectorize if no vector code will be produced. | ||||||||
8213 | if (VPlanBuildStressTest || EnableVPlanPredication || | ||||||||
8214 | VectorizationFactor::Disabled() == VF) | ||||||||
8215 | return false; | ||||||||
8216 | |||||||||
8217 | LVP.setBestPlan(VF.Width, 1); | ||||||||
8218 | |||||||||
8219 | InnerLoopVectorizer LB(L, PSE, LI, DT, TLI, TTI, AC, ORE, VF.Width, 1, LVL, | ||||||||
8220 | &CM, BFI, PSI); | ||||||||
8221 | 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) | ||||||||
8222 | << 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); | ||||||||
8223 | LVP.executePlan(LB, DT); | ||||||||
8224 | |||||||||
8225 | // Mark the loop as already vectorized to avoid vectorizing again. | ||||||||
8226 | Hints.setAlreadyVectorized(); | ||||||||
8227 | |||||||||
8228 | assert(!verifyFunction(*L->getHeader()->getParent(), &dbgs()))((!verifyFunction(*L->getHeader()->getParent(), &dbgs ())) ? static_cast<void> (0) : __assert_fail ("!verifyFunction(*L->getHeader()->getParent(), &dbgs())" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 8228, __PRETTY_FUNCTION__)); | ||||||||
8229 | return true; | ||||||||
8230 | } | ||||||||
8231 | |||||||||
8232 | LoopVectorizePass::LoopVectorizePass(LoopVectorizeOptions Opts) | ||||||||
8233 | : InterleaveOnlyWhenForced(Opts.InterleaveOnlyWhenForced || | ||||||||
8234 | !EnableLoopInterleaving), | ||||||||
8235 | VectorizeOnlyWhenForced(Opts.VectorizeOnlyWhenForced || | ||||||||
8236 | !EnableLoopVectorization) {} | ||||||||
8237 | |||||||||
8238 | bool LoopVectorizePass::processLoop(Loop *L) { | ||||||||
8239 | assert((EnableVPlanNativePath || L->isInnermost()) &&(((EnableVPlanNativePath || L->isInnermost()) && "VPlan-native path is not enabled. Only process inner loops." ) ? static_cast<void> (0) : __assert_fail ("(EnableVPlanNativePath || L->isInnermost()) && \"VPlan-native path is not enabled. Only process inner loops.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 8240, __PRETTY_FUNCTION__)) | ||||||||
8240 | "VPlan-native path is not enabled. Only process inner loops.")(((EnableVPlanNativePath || L->isInnermost()) && "VPlan-native path is not enabled. Only process inner loops." ) ? static_cast<void> (0) : __assert_fail ("(EnableVPlanNativePath || L->isInnermost()) && \"VPlan-native path is not enabled. Only process inner loops.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 8240, __PRETTY_FUNCTION__)); | ||||||||
8241 | |||||||||
8242 | #ifndef NDEBUG | ||||||||
8243 | const std::string DebugLocStr = getDebugLocString(L); | ||||||||
8244 | #endif /* NDEBUG */ | ||||||||
8245 | |||||||||
8246 | 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 ) | ||||||||
8247 | << 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 ) | ||||||||
8248 | << 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 ); | ||||||||
8249 | |||||||||
8250 | LoopVectorizeHints Hints(L, InterleaveOnlyWhenForced, *ORE); | ||||||||
8251 | |||||||||
8252 | 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() << " unroll=" << Hints.getInterleave( ) << "\n"; } } while (false) | ||||||||
8253 | 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() << " unroll=" << Hints.getInterleave( ) << "\n"; } } while (false) | ||||||||
8254 | << " 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() << " unroll=" << Hints.getInterleave( ) << "\n"; } } while (false) | ||||||||
8255 | << (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() << " unroll=" << Hints.getInterleave( ) << "\n"; } } while (false) | ||||||||
8256 | ? "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() << " unroll=" << Hints.getInterleave( ) << "\n"; } } while (false) | ||||||||
8257 | : (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() << " unroll=" << Hints.getInterleave( ) << "\n"; } } while (false) | ||||||||
8258 | ? "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() << " unroll=" << Hints.getInterleave( ) << "\n"; } } while (false) | ||||||||
8259 | : "?"))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() << " unroll=" << Hints.getInterleave( ) << "\n"; } } while (false) | ||||||||
8260 | << " 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() << " unroll=" << Hints.getInterleave( ) << "\n"; } } while (false) | ||||||||
8261 | << " unroll=" << 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() << " unroll=" << Hints.getInterleave( ) << "\n"; } } while (false); | ||||||||
8262 | |||||||||
8263 | // Function containing loop | ||||||||
8264 | Function *F = L->getHeader()->getParent(); | ||||||||
8265 | |||||||||
8266 | // Looking at the diagnostic output is the only way to determine if a loop | ||||||||
8267 | // was vectorized (other than looking at the IR or machine code), so it | ||||||||
8268 | // is important to generate an optimization remark for each loop. Most of | ||||||||
8269 | // these messages are generated as OptimizationRemarkAnalysis. Remarks | ||||||||
8270 | // generated as OptimizationRemark and OptimizationRemarkMissed are | ||||||||
8271 | // less verbose reporting vectorized loops and unvectorized loops that may | ||||||||
8272 | // benefit from vectorization, respectively. | ||||||||
8273 | |||||||||
8274 | if (!Hints.allowVectorization(F, L, VectorizeOnlyWhenForced)) { | ||||||||
8275 | 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); | ||||||||
8276 | return false; | ||||||||
8277 | } | ||||||||
8278 | |||||||||
8279 | PredicatedScalarEvolution PSE(*SE, *L); | ||||||||
8280 | |||||||||
8281 | // Check if it is legal to vectorize the loop. | ||||||||
8282 | LoopVectorizationRequirements Requirements(*ORE); | ||||||||
8283 | LoopVectorizationLegality LVL(L, PSE, DT, TTI, TLI, AA, F, GetLAA, LI, ORE, | ||||||||
8284 | &Requirements, &Hints, DB, AC, BFI, PSI); | ||||||||
8285 | if (!LVL.canVectorize(EnableVPlanNativePath)) { | ||||||||
8286 | 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); | ||||||||
8287 | Hints.emitRemarkWithHints(); | ||||||||
8288 | return false; | ||||||||
8289 | } | ||||||||
8290 | |||||||||
8291 | // Check the function attributes and profiles to find out if this function | ||||||||
8292 | // should be optimized for size. | ||||||||
8293 | ScalarEpilogueLowering SEL = getScalarEpilogueLowering( | ||||||||
8294 | F, L, Hints, PSI, BFI, TTI, TLI, AC, LI, PSE.getSE(), DT, LVL); | ||||||||
8295 | |||||||||
8296 | // Entrance to the VPlan-native vectorization path. Outer loops are processed | ||||||||
8297 | // here. They may require CFG and instruction level transformations before | ||||||||
8298 | // even evaluating whether vectorization is profitable. Since we cannot modify | ||||||||
8299 | // the incoming IR, we need to build VPlan upfront in the vectorization | ||||||||
8300 | // pipeline. | ||||||||
8301 | if (!L->isInnermost()) | ||||||||
8302 | return processLoopInVPlanNativePath(L, PSE, LI, DT, &LVL, TTI, TLI, DB, AC, | ||||||||
8303 | ORE, BFI, PSI, Hints); | ||||||||
8304 | |||||||||
8305 | assert(L->isInnermost() && "Inner loop expected.")((L->isInnermost() && "Inner loop expected.") ? static_cast <void> (0) : __assert_fail ("L->isInnermost() && \"Inner loop expected.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 8305, __PRETTY_FUNCTION__)); | ||||||||
8306 | |||||||||
8307 | // Check the loop for a trip count threshold: vectorize loops with a tiny trip | ||||||||
8308 | // count by optimizing for size, to minimize overheads. | ||||||||
8309 | auto ExpectedTC = getSmallBestKnownTC(*SE, L); | ||||||||
8310 | if (ExpectedTC && *ExpectedTC < TinyTripCountVectorThreshold) { | ||||||||
8311 | 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 ) | ||||||||
8312 | << "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 ) | ||||||||
8313 | << "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 ); | ||||||||
8314 | if (Hints.getForce() == LoopVectorizeHints::FK_Enabled) | ||||||||
8315 | 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); | ||||||||
8316 | else { | ||||||||
8317 | LLVM_DEBUG(dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "\n"; } } while (false); | ||||||||
8318 | SEL = CM_ScalarEpilogueNotAllowedLowTripLoop; | ||||||||
8319 | } | ||||||||
8320 | } | ||||||||
8321 | |||||||||
8322 | // Check the function attributes to see if implicit floats are allowed. | ||||||||
8323 | // FIXME: This check doesn't seem possibly correct -- what if the loop is | ||||||||
8324 | // an integer loop and the vector instructions selected are purely integer | ||||||||
8325 | // vector instructions? | ||||||||
8326 | if (F->hasFnAttribute(Attribute::NoImplicitFloat)) { | ||||||||
8327 | reportVectorizationFailure( | ||||||||
8328 | "Can't vectorize when the NoImplicitFloat attribute is used", | ||||||||
8329 | "loop not vectorized due to NoImplicitFloat attribute", | ||||||||
8330 | "NoImplicitFloat", ORE, L); | ||||||||
8331 | Hints.emitRemarkWithHints(); | ||||||||
8332 | return false; | ||||||||
8333 | } | ||||||||
8334 | |||||||||
8335 | // Check if the target supports potentially unsafe FP vectorization. | ||||||||
8336 | // FIXME: Add a check for the type of safety issue (denormal, signaling) | ||||||||
8337 | // for the target we're vectorizing for, to make sure none of the | ||||||||
8338 | // additional fp-math flags can help. | ||||||||
8339 | if (Hints.isPotentiallyUnsafe() && | ||||||||
8340 | TTI->isFPVectorizationPotentiallyUnsafe()) { | ||||||||
8341 | reportVectorizationFailure( | ||||||||
8342 | "Potentially unsafe FP op prevents vectorization", | ||||||||
8343 | "loop not vectorized due to unsafe FP support.", | ||||||||
8344 | "UnsafeFP", ORE, L); | ||||||||
8345 | Hints.emitRemarkWithHints(); | ||||||||
8346 | return false; | ||||||||
8347 | } | ||||||||
8348 | |||||||||
8349 | bool UseInterleaved = TTI->enableInterleavedAccessVectorization(); | ||||||||
8350 | InterleavedAccessInfo IAI(PSE, L, DT, LI, LVL.getLAI()); | ||||||||
8351 | |||||||||
8352 | // If an override option has been passed in for interleaved accesses, use it. | ||||||||
8353 | if (EnableInterleavedMemAccesses.getNumOccurrences() > 0) | ||||||||
8354 | UseInterleaved = EnableInterleavedMemAccesses; | ||||||||
8355 | |||||||||
8356 | // Analyze interleaved memory accesses. | ||||||||
8357 | if (UseInterleaved) { | ||||||||
8358 | IAI.analyzeInterleaving(useMaskedInterleavedAccesses(*TTI)); | ||||||||
8359 | } | ||||||||
8360 | |||||||||
8361 | // Use the cost model. | ||||||||
8362 | LoopVectorizationCostModel CM(SEL, L, PSE, LI, &LVL, *TTI, TLI, DB, AC, ORE, | ||||||||
8363 | F, &Hints, IAI); | ||||||||
8364 | CM.collectValuesToIgnore(); | ||||||||
8365 | |||||||||
8366 | // Use the planner for vectorization. | ||||||||
8367 | LoopVectorizationPlanner LVP(L, LI, TLI, TTI, &LVL, CM, IAI, PSE); | ||||||||
8368 | |||||||||
8369 | // Get user vectorization factor and interleave count. | ||||||||
8370 | unsigned UserVF = Hints.getWidth(); | ||||||||
8371 | unsigned UserIC = Hints.getInterleave(); | ||||||||
8372 | |||||||||
8373 | // Plan how to best vectorize, return the best VF and its cost. | ||||||||
8374 | Optional<VectorizationFactor> MaybeVF = | ||||||||
8375 | LVP.plan(ElementCount::getFixed(UserVF), UserIC); | ||||||||
8376 | |||||||||
8377 | VectorizationFactor VF = VectorizationFactor::Disabled(); | ||||||||
8378 | unsigned IC = 1; | ||||||||
8379 | |||||||||
8380 | if (MaybeVF) { | ||||||||
8381 | VF = *MaybeVF; | ||||||||
8382 | // Select the interleave count. | ||||||||
8383 | IC = CM.selectInterleaveCount(VF.Width, VF.Cost); | ||||||||
8384 | } | ||||||||
8385 | |||||||||
8386 | // Identify the diagnostic messages that should be produced. | ||||||||
8387 | std::pair<StringRef, std::string> VecDiagMsg, IntDiagMsg; | ||||||||
8388 | bool VectorizeLoop = true, InterleaveLoop = true; | ||||||||
8389 | if (Requirements.doesNotMeet(F, L, Hints)) { | ||||||||
8390 | LLVM_DEBUG(dbgs() << "LV: Not vectorizing: loop did not meet vectorization "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Not vectorizing: loop did not meet vectorization " "requirements.\n"; } } while (false) | ||||||||
8391 | "requirements.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Not vectorizing: loop did not meet vectorization " "requirements.\n"; } } while (false); | ||||||||
8392 | Hints.emitRemarkWithHints(); | ||||||||
8393 | return false; | ||||||||
8394 | } | ||||||||
8395 | |||||||||
8396 | if (VF.Width == 1) { | ||||||||
8397 | 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); | ||||||||
8398 | VecDiagMsg = std::make_pair( | ||||||||
8399 | "VectorizationNotBeneficial", | ||||||||
8400 | "the cost-model indicates that vectorization is not beneficial"); | ||||||||
8401 | VectorizeLoop = false; | ||||||||
8402 | } | ||||||||
8403 | |||||||||
8404 | if (!MaybeVF && UserIC > 1) { | ||||||||
8405 | // Tell the user interleaving was avoided up-front, despite being explicitly | ||||||||
8406 | // requested. | ||||||||
8407 | 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 ) | ||||||||
8408 | "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 ); | ||||||||
8409 | IntDiagMsg = std::make_pair( | ||||||||
8410 | "InterleavingAvoided", | ||||||||
8411 | "Ignoring UserIC, because interleaving was avoided up front"); | ||||||||
8412 | InterleaveLoop = false; | ||||||||
8413 | } else if (IC == 1 && UserIC <= 1) { | ||||||||
8414 | // Tell the user interleaving is not beneficial. | ||||||||
8415 | 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); | ||||||||
8416 | IntDiagMsg = std::make_pair( | ||||||||
8417 | "InterleavingNotBeneficial", | ||||||||
8418 | "the cost-model indicates that interleaving is not beneficial"); | ||||||||
8419 | InterleaveLoop = false; | ||||||||
8420 | if (UserIC == 1) { | ||||||||
8421 | IntDiagMsg.first = "InterleavingNotBeneficialAndDisabled"; | ||||||||
8422 | IntDiagMsg.second += | ||||||||
8423 | " and is explicitly disabled or interleave count is set to 1"; | ||||||||
8424 | } | ||||||||
8425 | } else if (IC > 1 && UserIC == 1) { | ||||||||
8426 | // Tell the user interleaving is beneficial, but it explicitly disabled. | ||||||||
8427 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Interleaving is beneficial but is explicitly disabled." ; } } while (false) | ||||||||
8428 | 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); | ||||||||
8429 | IntDiagMsg = std::make_pair( | ||||||||
8430 | "InterleavingBeneficialButDisabled", | ||||||||
8431 | "the cost-model indicates that interleaving is beneficial " | ||||||||
8432 | "but is explicitly disabled or interleave count is set to 1"); | ||||||||
8433 | InterleaveLoop = false; | ||||||||
8434 | } | ||||||||
8435 | |||||||||
8436 | // Override IC if user provided an interleave count. | ||||||||
8437 | IC = UserIC > 0 ? UserIC : IC; | ||||||||
8438 | |||||||||
8439 | // Emit diagnostic messages, if any. | ||||||||
8440 | const char *VAPassName = Hints.vectorizeAnalysisPassName(); | ||||||||
8441 | if (!VectorizeLoop && !InterleaveLoop) { | ||||||||
8442 | // Do not vectorize or interleaving the loop. | ||||||||
8443 | ORE->emit([&]() { | ||||||||
8444 | return OptimizationRemarkMissed(VAPassName, VecDiagMsg.first, | ||||||||
8445 | L->getStartLoc(), L->getHeader()) | ||||||||
8446 | << VecDiagMsg.second; | ||||||||
8447 | }); | ||||||||
8448 | ORE->emit([&]() { | ||||||||
8449 | return OptimizationRemarkMissed(LV_NAME"loop-vectorize", IntDiagMsg.first, | ||||||||
8450 | L->getStartLoc(), L->getHeader()) | ||||||||
8451 | << IntDiagMsg.second; | ||||||||
8452 | }); | ||||||||
8453 | return false; | ||||||||
8454 | } else if (!VectorizeLoop && InterleaveLoop) { | ||||||||
8455 | 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); | ||||||||
8456 | ORE->emit([&]() { | ||||||||
8457 | return OptimizationRemarkAnalysis(VAPassName, VecDiagMsg.first, | ||||||||
8458 | L->getStartLoc(), L->getHeader()) | ||||||||
8459 | << VecDiagMsg.second; | ||||||||
8460 | }); | ||||||||
8461 | } else if (VectorizeLoop && !InterleaveLoop) { | ||||||||
8462 | 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) | ||||||||
8463 | << ") in " << DebugLocStr << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in " << DebugLocStr << '\n'; } } while (false); | ||||||||
8464 | ORE->emit([&]() { | ||||||||
8465 | return OptimizationRemarkAnalysis(LV_NAME"loop-vectorize", IntDiagMsg.first, | ||||||||
8466 | L->getStartLoc(), L->getHeader()) | ||||||||
8467 | << IntDiagMsg.second; | ||||||||
8468 | }); | ||||||||
8469 | } else if (VectorizeLoop && InterleaveLoop) { | ||||||||
8470 | 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) | ||||||||
8471 | << ") in " << DebugLocStr << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in " << DebugLocStr << '\n'; } } while (false); | ||||||||
8472 | 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); | ||||||||
8473 | } | ||||||||
8474 | |||||||||
8475 | LVP.setBestPlan(VF.Width, IC); | ||||||||
8476 | |||||||||
8477 | using namespace ore; | ||||||||
8478 | bool DisableRuntimeUnroll = false; | ||||||||
8479 | MDNode *OrigLoopID = L->getLoopID(); | ||||||||
8480 | |||||||||
8481 | if (!VectorizeLoop) { | ||||||||
8482 | assert(IC > 1 && "interleave count should not be 1 or 0")((IC > 1 && "interleave count should not be 1 or 0" ) ? static_cast<void> (0) : __assert_fail ("IC > 1 && \"interleave count should not be 1 or 0\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 8482, __PRETTY_FUNCTION__)); | ||||||||
8483 | // If we decided that it is not legal to vectorize the loop, then | ||||||||
8484 | // interleave it. | ||||||||
8485 | InnerLoopUnroller Unroller(L, PSE, LI, DT, TLI, TTI, AC, ORE, IC, &LVL, &CM, | ||||||||
8486 | BFI, PSI); | ||||||||
8487 | LVP.executePlan(Unroller, DT); | ||||||||
8488 | |||||||||
8489 | ORE->emit([&]() { | ||||||||
8490 | return OptimizationRemark(LV_NAME"loop-vectorize", "Interleaved", L->getStartLoc(), | ||||||||
8491 | L->getHeader()) | ||||||||
8492 | << "interleaved loop (interleaved count: " | ||||||||
8493 | << NV("InterleaveCount", IC) << ")"; | ||||||||
8494 | }); | ||||||||
8495 | } else { | ||||||||
8496 | // If we decided that it is *legal* to vectorize the loop, then do it. | ||||||||
8497 | InnerLoopVectorizer LB(L, PSE, LI, DT, TLI, TTI, AC, ORE, VF.Width, IC, | ||||||||
8498 | &LVL, &CM, BFI, PSI); | ||||||||
8499 | LVP.executePlan(LB, DT); | ||||||||
8500 | ++LoopsVectorized; | ||||||||
8501 | |||||||||
8502 | // Add metadata to disable runtime unrolling a scalar loop when there are | ||||||||
8503 | // no runtime checks about strides and memory. A scalar loop that is | ||||||||
8504 | // rarely used is not worth unrolling. | ||||||||
8505 | if (!LB.areSafetyChecksAdded()) | ||||||||
8506 | DisableRuntimeUnroll = true; | ||||||||
8507 | |||||||||
8508 | // Report the vectorization decision. | ||||||||
8509 | ORE->emit([&]() { | ||||||||
8510 | return OptimizationRemark(LV_NAME"loop-vectorize", "Vectorized", L->getStartLoc(), | ||||||||
8511 | L->getHeader()) | ||||||||
8512 | << "vectorized loop (vectorization width: " | ||||||||
8513 | << NV("VectorizationFactor", VF.Width) | ||||||||
8514 | << ", interleaved count: " << NV("InterleaveCount", IC) << ")"; | ||||||||
8515 | }); | ||||||||
8516 | } | ||||||||
8517 | |||||||||
8518 | Optional<MDNode *> RemainderLoopID = | ||||||||
8519 | makeFollowupLoopID(OrigLoopID, {LLVMLoopVectorizeFollowupAll, | ||||||||
8520 | LLVMLoopVectorizeFollowupEpilogue}); | ||||||||
8521 | if (RemainderLoopID.hasValue()) { | ||||||||
8522 | L->setLoopID(RemainderLoopID.getValue()); | ||||||||
8523 | } else { | ||||||||
8524 | if (DisableRuntimeUnroll) | ||||||||
8525 | AddRuntimeUnrollDisableMetaData(L); | ||||||||
8526 | |||||||||
8527 | // Mark the loop as already vectorized to avoid vectorizing again. | ||||||||
8528 | Hints.setAlreadyVectorized(); | ||||||||
8529 | } | ||||||||
8530 | |||||||||
8531 | assert(!verifyFunction(*L->getHeader()->getParent(), &dbgs()))((!verifyFunction(*L->getHeader()->getParent(), &dbgs ())) ? static_cast<void> (0) : __assert_fail ("!verifyFunction(*L->getHeader()->getParent(), &dbgs())" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 8531, __PRETTY_FUNCTION__)); | ||||||||
8532 | return true; | ||||||||
8533 | } | ||||||||
8534 | |||||||||
8535 | LoopVectorizeResult LoopVectorizePass::runImpl( | ||||||||
8536 | Function &F, ScalarEvolution &SE_, LoopInfo &LI_, TargetTransformInfo &TTI_, | ||||||||
8537 | DominatorTree &DT_, BlockFrequencyInfo &BFI_, TargetLibraryInfo *TLI_, | ||||||||
8538 | DemandedBits &DB_, AAResults &AA_, AssumptionCache &AC_, | ||||||||
8539 | std::function<const LoopAccessInfo &(Loop &)> &GetLAA_, | ||||||||
8540 | OptimizationRemarkEmitter &ORE_, ProfileSummaryInfo *PSI_) { | ||||||||
8541 | SE = &SE_; | ||||||||
8542 | LI = &LI_; | ||||||||
8543 | TTI = &TTI_; | ||||||||
8544 | DT = &DT_; | ||||||||
8545 | BFI = &BFI_; | ||||||||
8546 | TLI = TLI_; | ||||||||
8547 | AA = &AA_; | ||||||||
8548 | AC = &AC_; | ||||||||
8549 | GetLAA = &GetLAA_; | ||||||||
8550 | DB = &DB_; | ||||||||
8551 | ORE = &ORE_; | ||||||||
8552 | PSI = PSI_; | ||||||||
8553 | |||||||||
8554 | // Don't attempt if | ||||||||
8555 | // 1. the target claims to have no vector registers, and | ||||||||
8556 | // 2. interleaving won't help ILP. | ||||||||
8557 | // | ||||||||
8558 | // The second condition is necessary because, even if the target has no | ||||||||
8559 | // vector registers, loop vectorization may still enable scalar | ||||||||
8560 | // interleaving. | ||||||||
8561 | if (!TTI->getNumberOfRegisters(TTI->getRegisterClassForType(true)) && | ||||||||
8562 | TTI->getMaxInterleaveFactor(1) < 2) | ||||||||
8563 | return LoopVectorizeResult(false, false); | ||||||||
8564 | |||||||||
8565 | bool Changed = false, CFGChanged = false; | ||||||||
8566 | |||||||||
8567 | // The vectorizer requires loops to be in simplified form. | ||||||||
8568 | // Since simplification may add new inner loops, it has to run before the | ||||||||
8569 | // legality and profitability checks. This means running the loop vectorizer | ||||||||
8570 | // will simplify all loops, regardless of whether anything end up being | ||||||||
8571 | // vectorized. | ||||||||
8572 | for (auto &L : *LI) | ||||||||
8573 | Changed |= CFGChanged |= | ||||||||
8574 | simplifyLoop(L, DT, LI, SE, AC, nullptr, false /* PreserveLCSSA */); | ||||||||
8575 | |||||||||
8576 | // Build up a worklist of inner-loops to vectorize. This is necessary as | ||||||||
8577 | // the act of vectorizing or partially unrolling a loop creates new loops | ||||||||
8578 | // and can invalidate iterators across the loops. | ||||||||
8579 | SmallVector<Loop *, 8> Worklist; | ||||||||
8580 | |||||||||
8581 | for (Loop *L : *LI) | ||||||||
8582 | collectSupportedLoops(*L, LI, ORE, Worklist); | ||||||||
8583 | |||||||||
8584 | LoopsAnalyzed += Worklist.size(); | ||||||||
8585 | |||||||||
8586 | // Now walk the identified inner loops. | ||||||||
8587 | while (!Worklist.empty()) { | ||||||||
8588 | Loop *L = Worklist.pop_back_val(); | ||||||||
8589 | |||||||||
8590 | // For the inner loops we actually process, form LCSSA to simplify the | ||||||||
8591 | // transform. | ||||||||
8592 | Changed |= formLCSSARecursively(*L, *DT, LI, SE); | ||||||||
8593 | |||||||||
8594 | Changed |= CFGChanged |= processLoop(L); | ||||||||
8595 | } | ||||||||
8596 | |||||||||
8597 | // Process each loop nest in the function. | ||||||||
8598 | return LoopVectorizeResult(Changed, CFGChanged); | ||||||||
8599 | } | ||||||||
8600 | |||||||||
8601 | PreservedAnalyses LoopVectorizePass::run(Function &F, | ||||||||
8602 | FunctionAnalysisManager &AM) { | ||||||||
8603 | auto &SE = AM.getResult<ScalarEvolutionAnalysis>(F); | ||||||||
8604 | auto &LI = AM.getResult<LoopAnalysis>(F); | ||||||||
8605 | auto &TTI = AM.getResult<TargetIRAnalysis>(F); | ||||||||
8606 | auto &DT = AM.getResult<DominatorTreeAnalysis>(F); | ||||||||
8607 | auto &BFI = AM.getResult<BlockFrequencyAnalysis>(F); | ||||||||
8608 | auto &TLI = AM.getResult<TargetLibraryAnalysis>(F); | ||||||||
8609 | auto &AA = AM.getResult<AAManager>(F); | ||||||||
8610 | auto &AC = AM.getResult<AssumptionAnalysis>(F); | ||||||||
8611 | auto &DB = AM.getResult<DemandedBitsAnalysis>(F); | ||||||||
8612 | auto &ORE = AM.getResult<OptimizationRemarkEmitterAnalysis>(F); | ||||||||
8613 | MemorySSA *MSSA = EnableMSSALoopDependency | ||||||||
8614 | ? &AM.getResult<MemorySSAAnalysis>(F).getMSSA() | ||||||||
8615 | : nullptr; | ||||||||
8616 | |||||||||
8617 | auto &LAM = AM.getResult<LoopAnalysisManagerFunctionProxy>(F).getManager(); | ||||||||
8618 | std::function<const LoopAccessInfo &(Loop &)> GetLAA = | ||||||||
8619 | [&](Loop &L) -> const LoopAccessInfo & { | ||||||||
8620 | LoopStandardAnalysisResults AR = {AA, AC, DT, LI, SE, | ||||||||
8621 | TLI, TTI, nullptr, MSSA}; | ||||||||
8622 | return LAM.getResult<LoopAccessAnalysis>(L, AR); | ||||||||
8623 | }; | ||||||||
8624 | auto &MAMProxy = AM.getResult<ModuleAnalysisManagerFunctionProxy>(F); | ||||||||
8625 | ProfileSummaryInfo *PSI = | ||||||||
8626 | MAMProxy.getCachedResult<ProfileSummaryAnalysis>(*F.getParent()); | ||||||||
8627 | LoopVectorizeResult Result = | ||||||||
8628 | runImpl(F, SE, LI, TTI, DT, BFI, &TLI, DB, AA, AC, GetLAA, ORE, PSI); | ||||||||
8629 | if (!Result.MadeAnyChange) | ||||||||
8630 | return PreservedAnalyses::all(); | ||||||||
8631 | PreservedAnalyses PA; | ||||||||
8632 | |||||||||
8633 | // We currently do not preserve loopinfo/dominator analyses with outer loop | ||||||||
8634 | // vectorization. Until this is addressed, mark these analyses as preserved | ||||||||
8635 | // only for non-VPlan-native path. | ||||||||
8636 | // TODO: Preserve Loop and Dominator analyses for VPlan-native path. | ||||||||
8637 | if (!EnableVPlanNativePath) { | ||||||||
8638 | PA.preserve<LoopAnalysis>(); | ||||||||
8639 | PA.preserve<DominatorTreeAnalysis>(); | ||||||||
8640 | } | ||||||||
8641 | PA.preserve<BasicAA>(); | ||||||||
8642 | PA.preserve<GlobalsAA>(); | ||||||||
8643 | if (!Result.MadeCFGChange) | ||||||||
8644 | PA.preserveSet<CFGAnalyses>(); | ||||||||
8645 | return PA; | ||||||||
8646 | } |
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/Analysis/LoopInfo.h" |
29 | #include "llvm/Analysis/TargetLibraryInfo.h" |
30 | #include "llvm/Analysis/TargetTransformInfo.h" |
31 | |
32 | namespace llvm { |
33 | |
34 | class LoopVectorizationLegality; |
35 | class LoopVectorizationCostModel; |
36 | class PredicatedScalarEvolution; |
37 | class VPRecipeBuilder; |
38 | |
39 | /// VPlan-based builder utility analogous to IRBuilder. |
40 | class VPBuilder { |
41 | VPBasicBlock *BB = nullptr; |
42 | VPBasicBlock::iterator InsertPt = VPBasicBlock::iterator(); |
43 | |
44 | VPInstruction *createInstruction(unsigned Opcode, |
45 | ArrayRef<VPValue *> Operands) { |
46 | VPInstruction *Instr = new VPInstruction(Opcode, Operands); |
47 | if (BB) |
48 | BB->insert(Instr, InsertPt); |
49 | return Instr; |
50 | } |
51 | |
52 | VPInstruction *createInstruction(unsigned Opcode, |
53 | std::initializer_list<VPValue *> Operands) { |
54 | return createInstruction(Opcode, ArrayRef<VPValue *>(Operands)); |
55 | } |
56 | |
57 | public: |
58 | VPBuilder() {} |
59 | |
60 | /// Clear the insertion point: created instructions will not be inserted into |
61 | /// a block. |
62 | void clearInsertionPoint() { |
63 | BB = nullptr; |
64 | InsertPt = VPBasicBlock::iterator(); |
65 | } |
66 | |
67 | VPBasicBlock *getInsertBlock() const { return BB; } |
68 | VPBasicBlock::iterator getInsertPoint() const { return InsertPt; } |
69 | |
70 | /// InsertPoint - A saved insertion point. |
71 | class VPInsertPoint { |
72 | VPBasicBlock *Block = nullptr; |
73 | VPBasicBlock::iterator Point; |
74 | |
75 | public: |
76 | /// Creates a new insertion point which doesn't point to anything. |
77 | VPInsertPoint() = default; |
78 | |
79 | /// Creates a new insertion point at the given location. |
80 | VPInsertPoint(VPBasicBlock *InsertBlock, VPBasicBlock::iterator InsertPoint) |
81 | : Block(InsertBlock), Point(InsertPoint) {} |
82 | |
83 | /// Returns true if this insert point is set. |
84 | bool isSet() const { return Block != nullptr; } |
85 | |
86 | VPBasicBlock *getBlock() const { return Block; } |
87 | VPBasicBlock::iterator getPoint() const { return Point; } |
88 | }; |
89 | |
90 | /// Sets the current insert point to a previously-saved location. |
91 | void restoreIP(VPInsertPoint IP) { |
92 | if (IP.isSet()) |
93 | setInsertPoint(IP.getBlock(), IP.getPoint()); |
94 | else |
95 | clearInsertionPoint(); |
96 | } |
97 | |
98 | /// This specifies that created VPInstructions should be appended to the end |
99 | /// of the specified block. |
100 | void setInsertPoint(VPBasicBlock *TheBB) { |
101 | assert(TheBB && "Attempting to set a null insert point")((TheBB && "Attempting to set a null insert point") ? static_cast<void> (0) : __assert_fail ("TheBB && \"Attempting to set a null insert point\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/LoopVectorizationPlanner.h" , 101, __PRETTY_FUNCTION__)); |
102 | BB = TheBB; |
103 | InsertPt = BB->end(); |
104 | } |
105 | |
106 | /// This specifies that created instructions should be inserted at the |
107 | /// specified point. |
108 | void setInsertPoint(VPBasicBlock *TheBB, VPBasicBlock::iterator IP) { |
109 | BB = TheBB; |
110 | InsertPt = IP; |
111 | } |
112 | |
113 | /// Insert and return the specified instruction. |
114 | VPInstruction *insert(VPInstruction *I) const { |
115 | BB->insert(I, InsertPt); |
116 | return I; |
117 | } |
118 | |
119 | /// Create an N-ary operation with \p Opcode, \p Operands and set \p Inst as |
120 | /// its underlying Instruction. |
121 | VPValue *createNaryOp(unsigned Opcode, ArrayRef<VPValue *> Operands, |
122 | Instruction *Inst = nullptr) { |
123 | VPInstruction *NewVPInst = createInstruction(Opcode, Operands); |
124 | NewVPInst->setUnderlyingValue(Inst); |
125 | return NewVPInst; |
126 | } |
127 | VPValue *createNaryOp(unsigned Opcode, |
128 | std::initializer_list<VPValue *> Operands, |
129 | Instruction *Inst = nullptr) { |
130 | return createNaryOp(Opcode, ArrayRef<VPValue *>(Operands), Inst); |
131 | } |
132 | |
133 | VPValue *createNot(VPValue *Operand) { |
134 | return createInstruction(VPInstruction::Not, {Operand}); |
135 | } |
136 | |
137 | VPValue *createAnd(VPValue *LHS, VPValue *RHS) { |
138 | return createInstruction(Instruction::BinaryOps::And, {LHS, RHS}); |
139 | } |
140 | |
141 | VPValue *createOr(VPValue *LHS, VPValue *RHS) { |
142 | return createInstruction(Instruction::BinaryOps::Or, {LHS, RHS}); |
143 | } |
144 | |
145 | //===--------------------------------------------------------------------===// |
146 | // RAII helpers. |
147 | //===--------------------------------------------------------------------===// |
148 | |
149 | /// RAII object that stores the current insertion point and restores it when |
150 | /// the object is destroyed. |
151 | class InsertPointGuard { |
152 | VPBuilder &Builder; |
153 | VPBasicBlock *Block; |
154 | VPBasicBlock::iterator Point; |
155 | |
156 | public: |
157 | InsertPointGuard(VPBuilder &B) |
158 | : Builder(B), Block(B.getInsertBlock()), Point(B.getInsertPoint()) {} |
159 | |
160 | InsertPointGuard(const InsertPointGuard &) = delete; |
161 | InsertPointGuard &operator=(const InsertPointGuard &) = delete; |
162 | |
163 | ~InsertPointGuard() { Builder.restoreIP(VPInsertPoint(Block, Point)); } |
164 | }; |
165 | }; |
166 | |
167 | /// TODO: The following VectorizationFactor was pulled out of |
168 | /// LoopVectorizationCostModel class. LV also deals with |
169 | /// VectorizerParams::VectorizationFactor and VectorizationCostTy. |
170 | /// We need to streamline them. |
171 | |
172 | /// Information about vectorization costs |
173 | struct VectorizationFactor { |
174 | // Vector width with best cost |
175 | ElementCount Width; |
176 | // Cost of the loop with that width |
177 | unsigned Cost; |
178 | |
179 | // Width 1 means no vectorization, cost 0 means uncomputed cost. |
180 | static VectorizationFactor Disabled() { |
181 | return {ElementCount::getFixed(1), 0}; |
182 | } |
183 | |
184 | bool operator==(const VectorizationFactor &rhs) const { |
185 | return Width == rhs.Width && Cost == rhs.Cost; |
186 | } |
187 | }; |
188 | |
189 | /// Planner drives the vectorization process after having passed |
190 | /// Legality checks. |
191 | class LoopVectorizationPlanner { |
192 | /// The loop that we evaluate. |
193 | Loop *OrigLoop; |
194 | |
195 | /// Loop Info analysis. |
196 | LoopInfo *LI; |
197 | |
198 | /// Target Library Info. |
199 | const TargetLibraryInfo *TLI; |
200 | |
201 | /// Target Transform Info. |
202 | const TargetTransformInfo *TTI; |
203 | |
204 | /// The legality analysis. |
205 | LoopVectorizationLegality *Legal; |
206 | |
207 | /// The profitability analysis. |
208 | LoopVectorizationCostModel &CM; |
209 | |
210 | /// The interleaved access analysis. |
211 | InterleavedAccessInfo &IAI; |
212 | |
213 | PredicatedScalarEvolution &PSE; |
214 | |
215 | SmallVector<VPlanPtr, 4> VPlans; |
216 | |
217 | /// This class is used to enable the VPlan to invoke a method of ILV. This is |
218 | /// needed until the method is refactored out of ILV and becomes reusable. |
219 | struct VPCallbackILV : public VPCallback { |
220 | InnerLoopVectorizer &ILV; |
221 | |
222 | VPCallbackILV(InnerLoopVectorizer &ILV) : ILV(ILV) {} |
223 | |
224 | Value *getOrCreateVectorValues(Value *V, unsigned Part) override; |
225 | Value *getOrCreateScalarValue(Value *V, |
226 | const VPIteration &Instance) override; |
227 | }; |
228 | |
229 | /// A builder used to construct the current plan. |
230 | VPBuilder Builder; |
231 | |
232 | /// The best number of elements of the vector types used in the |
233 | /// transformed loop. BestVF = None means that vectorization is |
234 | /// disabled. |
235 | Optional<ElementCount> BestVF = None; |
236 | unsigned BestUF = 0; |
237 | |
238 | public: |
239 | LoopVectorizationPlanner(Loop *L, LoopInfo *LI, const TargetLibraryInfo *TLI, |
240 | const TargetTransformInfo *TTI, |
241 | LoopVectorizationLegality *Legal, |
242 | LoopVectorizationCostModel &CM, |
243 | InterleavedAccessInfo &IAI, |
244 | PredicatedScalarEvolution &PSE) |
245 | : OrigLoop(L), LI(LI), TLI(TLI), TTI(TTI), Legal(Legal), CM(CM), IAI(IAI), |
246 | PSE(PSE) {} |
247 | |
248 | /// Plan how to best vectorize, return the best VF and its cost, or None if |
249 | /// vectorization and interleaving should be avoided up front. |
250 | Optional<VectorizationFactor> plan(ElementCount UserVF, unsigned UserIC); |
251 | |
252 | /// Use the VPlan-native path to plan how to best vectorize, return the best |
253 | /// VF and its cost. |
254 | VectorizationFactor planInVPlanNativePath(ElementCount UserVF); |
255 | |
256 | /// Finalize the best decision and dispose of all other VPlans. |
257 | void setBestPlan(ElementCount VF, unsigned UF); |
258 | |
259 | /// Generate the IR code for the body of the vectorized loop according to the |
260 | /// best selected VPlan. |
261 | void executePlan(InnerLoopVectorizer &LB, DominatorTree *DT); |
262 | |
263 | void printPlans(raw_ostream &O) { |
264 | for (const auto &Plan : VPlans) |
265 | O << *Plan; |
266 | } |
267 | |
268 | /// Test a \p Predicate on a \p Range of VF's. Return the value of applying |
269 | /// \p Predicate on Range.Start, possibly decreasing Range.End such that the |
270 | /// returned value holds for the entire \p Range. |
271 | static bool |
272 | getDecisionAndClampRange(const std::function<bool(ElementCount)> &Predicate, |
273 | VFRange &Range); |
274 | |
275 | protected: |
276 | /// Collect the instructions from the original loop that would be trivially |
277 | /// dead in the vectorized loop if generated. |
278 | void collectTriviallyDeadInstructions( |
279 | SmallPtrSetImpl<Instruction *> &DeadInstructions); |
280 | |
281 | /// Build VPlans for power-of-2 VF's between \p MinVF and \p MaxVF inclusive, |
282 | /// according to the information gathered by Legal when it checked if it is |
283 | /// legal to vectorize the loop. |
284 | void buildVPlans(unsigned MinVF, unsigned MaxVF); |
285 | |
286 | private: |
287 | /// Build a VPlan according to the information gathered by Legal. \return a |
288 | /// VPlan for vectorization factors \p Range.Start and up to \p Range.End |
289 | /// exclusive, possibly decreasing \p Range.End. |
290 | VPlanPtr buildVPlan(VFRange &Range); |
291 | |
292 | /// Build a VPlan using VPRecipes according to the information gather by |
293 | /// Legal. This method is only used for the legacy inner loop vectorizer. |
294 | VPlanPtr buildVPlanWithVPRecipes( |
295 | VFRange &Range, SmallPtrSetImpl<Value *> &NeedDef, |
296 | SmallPtrSetImpl<Instruction *> &DeadInstructions, |
297 | const DenseMap<Instruction *, Instruction *> &SinkAfter); |
298 | |
299 | /// Build VPlans for power-of-2 VF's between \p MinVF and \p MaxVF inclusive, |
300 | /// according to the information gathered by Legal when it checked if it is |
301 | /// legal to vectorize the loop. This method creates VPlans using VPRecipes. |
302 | void buildVPlansWithVPRecipes(unsigned MinVF, unsigned MaxVF); |
303 | |
304 | /// Adjust the recipes for any inloop reductions. The chain of instructions |
305 | /// leading from the loop exit instr to the phi need to be converted to |
306 | /// reductions, with one operand being vector and the other being the scalar |
307 | /// reduction chain. |
308 | void adjustRecipesForInLoopReductions(VPlanPtr &Plan, |
309 | VPRecipeBuilder &RecipeBuilder); |
310 | }; |
311 | |
312 | } // namespace llvm |
313 | |
314 | #endif // LLVM_TRANSFORMS_VECTORIZE_LOOPVECTORIZATIONPLANNER_H |