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/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 "VPlanAnalysis.h" |
61 | #include "VPlanCFG.h" |
62 | #include "VPlanHelpers.h" |
63 | #include "VPlanPatternMatch.h" |
64 | #include "VPlanTransforms.h" |
65 | #include "VPlanUtils.h" |
66 | #include "VPlanVerifier.h" |
67 | #include "llvm/ADT/APInt.h" |
68 | #include "llvm/ADT/ArrayRef.h" |
69 | #include "llvm/ADT/DenseMap.h" |
70 | #include "llvm/ADT/DenseMapInfo.h" |
71 | #include "llvm/ADT/Hashing.h" |
72 | #include "llvm/ADT/MapVector.h" |
73 | #include "llvm/ADT/STLExtras.h" |
74 | #include "llvm/ADT/SmallPtrSet.h" |
75 | #include "llvm/ADT/SmallVector.h" |
76 | #include "llvm/ADT/Statistic.h" |
77 | #include "llvm/ADT/StringRef.h" |
78 | #include "llvm/ADT/Twine.h" |
79 | #include "llvm/ADT/TypeSwitch.h" |
80 | #include "llvm/ADT/iterator_range.h" |
81 | #include "llvm/Analysis/AssumptionCache.h" |
82 | #include "llvm/Analysis/BasicAliasAnalysis.h" |
83 | #include "llvm/Analysis/BlockFrequencyInfo.h" |
84 | #include "llvm/Analysis/CFG.h" |
85 | #include "llvm/Analysis/CodeMetrics.h" |
86 | #include "llvm/Analysis/DemandedBits.h" |
87 | #include "llvm/Analysis/GlobalsModRef.h" |
88 | #include "llvm/Analysis/LoopAccessAnalysis.h" |
89 | #include "llvm/Analysis/LoopAnalysisManager.h" |
90 | #include "llvm/Analysis/LoopInfo.h" |
91 | #include "llvm/Analysis/LoopIterator.h" |
92 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" |
93 | #include "llvm/Analysis/ProfileSummaryInfo.h" |
94 | #include "llvm/Analysis/ScalarEvolution.h" |
95 | #include "llvm/Analysis/ScalarEvolutionExpressions.h" |
96 | #include "llvm/Analysis/TargetLibraryInfo.h" |
97 | #include "llvm/Analysis/TargetTransformInfo.h" |
98 | #include "llvm/Analysis/ValueTracking.h" |
99 | #include "llvm/Analysis/VectorUtils.h" |
100 | #include "llvm/IR/Attributes.h" |
101 | #include "llvm/IR/BasicBlock.h" |
102 | #include "llvm/IR/CFG.h" |
103 | #include "llvm/IR/Constant.h" |
104 | #include "llvm/IR/Constants.h" |
105 | #include "llvm/IR/DataLayout.h" |
106 | #include "llvm/IR/DebugInfo.h" |
107 | #include "llvm/IR/DebugLoc.h" |
108 | #include "llvm/IR/DerivedTypes.h" |
109 | #include "llvm/IR/DiagnosticInfo.h" |
110 | #include "llvm/IR/Dominators.h" |
111 | #include "llvm/IR/Function.h" |
112 | #include "llvm/IR/IRBuilder.h" |
113 | #include "llvm/IR/InstrTypes.h" |
114 | #include "llvm/IR/Instruction.h" |
115 | #include "llvm/IR/Instructions.h" |
116 | #include "llvm/IR/IntrinsicInst.h" |
117 | #include "llvm/IR/Intrinsics.h" |
118 | #include "llvm/IR/MDBuilder.h" |
119 | #include "llvm/IR/Metadata.h" |
120 | #include "llvm/IR/Module.h" |
121 | #include "llvm/IR/Operator.h" |
122 | #include "llvm/IR/PatternMatch.h" |
123 | #include "llvm/IR/ProfDataUtils.h" |
124 | #include "llvm/IR/Type.h" |
125 | #include "llvm/IR/Use.h" |
126 | #include "llvm/IR/User.h" |
127 | #include "llvm/IR/Value.h" |
128 | #include "llvm/IR/Verifier.h" |
129 | #include "llvm/Support/Casting.h" |
130 | #include "llvm/Support/CommandLine.h" |
131 | #include "llvm/Support/Debug.h" |
132 | #include "llvm/Support/ErrorHandling.h" |
133 | #include "llvm/Support/InstructionCost.h" |
134 | #include "llvm/Support/MathExtras.h" |
135 | #include "llvm/Support/NativeFormatting.h" |
136 | #include "llvm/Support/raw_ostream.h" |
137 | #include "llvm/Transforms/Utils/BasicBlockUtils.h" |
138 | #include "llvm/Transforms/Utils/InjectTLIMappings.h" |
139 | #include "llvm/Transforms/Utils/Local.h" |
140 | #include "llvm/Transforms/Utils/LoopSimplify.h" |
141 | #include "llvm/Transforms/Utils/LoopUtils.h" |
142 | #include "llvm/Transforms/Utils/LoopVersioning.h" |
143 | #include "llvm/Transforms/Utils/ScalarEvolutionExpander.h" |
144 | #include "llvm/Transforms/Utils/SizeOpts.h" |
145 | #include "llvm/Transforms/Vectorize/LoopVectorizationLegality.h" |
146 | #include <algorithm> |
147 | #include <cassert> |
148 | #include <cstdint> |
149 | #include <functional> |
150 | #include <iterator> |
151 | #include <limits> |
152 | #include <memory> |
153 | #include <string> |
154 | #include <tuple> |
155 | #include <utility> |
156 | |
157 | using namespace llvm; |
158 | |
159 | #define LV_NAME "loop-vectorize" |
160 | #define DEBUG_TYPE LV_NAME |
161 | |
162 | #ifndef NDEBUG |
163 | const char VerboseDebug[] = DEBUG_TYPE "-verbose" ; |
164 | #endif |
165 | |
166 | /// @{ |
167 | /// Metadata attribute names |
168 | const char LLVMLoopVectorizeFollowupAll[] = "llvm.loop.vectorize.followup_all" ; |
169 | const char LLVMLoopVectorizeFollowupVectorized[] = |
170 | "llvm.loop.vectorize.followup_vectorized" ; |
171 | const char LLVMLoopVectorizeFollowupEpilogue[] = |
172 | "llvm.loop.vectorize.followup_epilogue" ; |
173 | /// @} |
174 | |
175 | STATISTIC(LoopsVectorized, "Number of loops vectorized" ); |
176 | STATISTIC(LoopsAnalyzed, "Number of loops analyzed for vectorization" ); |
177 | STATISTIC(LoopsEpilogueVectorized, "Number of epilogues vectorized" ); |
178 | |
179 | static cl::opt<bool> EnableEpilogueVectorization( |
180 | "enable-epilogue-vectorization" , cl::init(Val: true), cl::Hidden, |
181 | cl::desc("Enable vectorization of epilogue loops." )); |
182 | |
183 | static cl::opt<unsigned> EpilogueVectorizationForceVF( |
184 | "epilogue-vectorization-force-VF" , cl::init(Val: 1), cl::Hidden, |
185 | cl::desc("When epilogue vectorization is enabled, and a value greater than " |
186 | "1 is specified, forces the given VF for all applicable epilogue " |
187 | "loops." )); |
188 | |
189 | static cl::opt<unsigned> EpilogueVectorizationMinVF( |
190 | "epilogue-vectorization-minimum-VF" , cl::Hidden, |
191 | cl::desc("Only loops with vectorization factor equal to or larger than " |
192 | "the specified value are considered for epilogue vectorization." )); |
193 | |
194 | /// Loops with a known constant trip count below this number are vectorized only |
195 | /// if no scalar iteration overheads are incurred. |
196 | static cl::opt<unsigned> TinyTripCountVectorThreshold( |
197 | "vectorizer-min-trip-count" , cl::init(Val: 16), cl::Hidden, |
198 | cl::desc("Loops with a constant trip count that is smaller than this " |
199 | "value are vectorized only if no scalar iteration overheads " |
200 | "are incurred." )); |
201 | |
202 | static cl::opt<unsigned> VectorizeMemoryCheckThreshold( |
203 | "vectorize-memory-check-threshold" , cl::init(Val: 128), cl::Hidden, |
204 | cl::desc("The maximum allowed number of runtime memory checks" )); |
205 | |
206 | // Option prefer-predicate-over-epilogue indicates that an epilogue is undesired, |
207 | // that predication is preferred, and this lists all options. I.e., the |
208 | // vectorizer will try to fold the tail-loop (epilogue) into the vector body |
209 | // and predicate the instructions accordingly. If tail-folding fails, there are |
210 | // different fallback strategies depending on these values: |
211 | namespace PreferPredicateTy { |
212 | enum Option { |
213 | ScalarEpilogue = 0, |
214 | PredicateElseScalarEpilogue, |
215 | PredicateOrDontVectorize |
216 | }; |
217 | } // namespace PreferPredicateTy |
218 | |
219 | static cl::opt<PreferPredicateTy::Option> PreferPredicateOverEpilogue( |
220 | "prefer-predicate-over-epilogue" , |
221 | cl::init(Val: PreferPredicateTy::ScalarEpilogue), |
222 | cl::Hidden, |
223 | cl::desc("Tail-folding and predication preferences over creating a scalar " |
224 | "epilogue loop." ), |
225 | cl::values(clEnumValN(PreferPredicateTy::ScalarEpilogue, |
226 | "scalar-epilogue" , |
227 | "Don't tail-predicate loops, create scalar epilogue" ), |
228 | clEnumValN(PreferPredicateTy::PredicateElseScalarEpilogue, |
229 | "predicate-else-scalar-epilogue" , |
230 | "prefer tail-folding, create scalar epilogue if tail " |
231 | "folding fails." ), |
232 | clEnumValN(PreferPredicateTy::PredicateOrDontVectorize, |
233 | "predicate-dont-vectorize" , |
234 | "prefers tail-folding, don't attempt vectorization if " |
235 | "tail-folding fails." ))); |
236 | |
237 | static cl::opt<TailFoldingStyle> ForceTailFoldingStyle( |
238 | "force-tail-folding-style" , cl::desc("Force the tail folding style" ), |
239 | cl::init(Val: TailFoldingStyle::None), |
240 | cl::values( |
241 | clEnumValN(TailFoldingStyle::None, "none" , "Disable tail folding" ), |
242 | clEnumValN( |
243 | TailFoldingStyle::Data, "data" , |
244 | "Create lane mask for data only, using active.lane.mask intrinsic" ), |
245 | clEnumValN(TailFoldingStyle::DataWithoutLaneMask, |
246 | "data-without-lane-mask" , |
247 | "Create lane mask with compare/stepvector" ), |
248 | clEnumValN(TailFoldingStyle::DataAndControlFlow, "data-and-control" , |
249 | "Create lane mask using active.lane.mask intrinsic, and use " |
250 | "it for both data and control flow" ), |
251 | clEnumValN(TailFoldingStyle::DataAndControlFlowWithoutRuntimeCheck, |
252 | "data-and-control-without-rt-check" , |
253 | "Similar to data-and-control, but remove the runtime check" ), |
254 | clEnumValN(TailFoldingStyle::DataWithEVL, "data-with-evl" , |
255 | "Use predicated EVL instructions for tail folding. If EVL " |
256 | "is unsupported, fallback to data-without-lane-mask." ))); |
257 | |
258 | static cl::opt<bool> MaximizeBandwidth( |
259 | "vectorizer-maximize-bandwidth" , cl::init(Val: false), cl::Hidden, |
260 | cl::desc("Maximize bandwidth when selecting vectorization factor which " |
261 | "will be determined by the smallest type in loop." )); |
262 | |
263 | static cl::opt<bool> EnableInterleavedMemAccesses( |
264 | "enable-interleaved-mem-accesses" , cl::init(Val: false), cl::Hidden, |
265 | cl::desc("Enable vectorization on interleaved memory accesses in a loop" )); |
266 | |
267 | /// An interleave-group may need masking if it resides in a block that needs |
268 | /// predication, or in order to mask away gaps. |
269 | static cl::opt<bool> EnableMaskedInterleavedMemAccesses( |
270 | "enable-masked-interleaved-mem-accesses" , cl::init(Val: false), cl::Hidden, |
271 | cl::desc("Enable vectorization on masked interleaved memory accesses in a loop" )); |
272 | |
273 | static cl::opt<unsigned> ForceTargetNumScalarRegs( |
274 | "force-target-num-scalar-regs" , cl::init(Val: 0), cl::Hidden, |
275 | cl::desc("A flag that overrides the target's number of scalar registers." )); |
276 | |
277 | static cl::opt<unsigned> ForceTargetNumVectorRegs( |
278 | "force-target-num-vector-regs" , cl::init(Val: 0), cl::Hidden, |
279 | cl::desc("A flag that overrides the target's number of vector registers." )); |
280 | |
281 | static cl::opt<unsigned> ForceTargetMaxScalarInterleaveFactor( |
282 | "force-target-max-scalar-interleave" , cl::init(Val: 0), cl::Hidden, |
283 | cl::desc("A flag that overrides the target's max interleave factor for " |
284 | "scalar loops." )); |
285 | |
286 | static cl::opt<unsigned> ForceTargetMaxVectorInterleaveFactor( |
287 | "force-target-max-vector-interleave" , cl::init(Val: 0), cl::Hidden, |
288 | cl::desc("A flag that overrides the target's max interleave factor for " |
289 | "vectorized loops." )); |
290 | |
291 | cl::opt<unsigned> llvm::ForceTargetInstructionCost( |
292 | "force-target-instruction-cost" , cl::init(Val: 0), cl::Hidden, |
293 | cl::desc("A flag that overrides the target's expected cost for " |
294 | "an instruction to a single constant value. Mostly " |
295 | "useful for getting consistent testing." )); |
296 | |
297 | static cl::opt<bool> ForceTargetSupportsScalableVectors( |
298 | "force-target-supports-scalable-vectors" , cl::init(Val: false), cl::Hidden, |
299 | cl::desc( |
300 | "Pretend that scalable vectors are supported, even if the target does " |
301 | "not support them. This flag should only be used for testing." )); |
302 | |
303 | static cl::opt<unsigned> SmallLoopCost( |
304 | "small-loop-cost" , cl::init(Val: 20), cl::Hidden, |
305 | cl::desc( |
306 | "The cost of a loop that is considered 'small' by the interleaver." )); |
307 | |
308 | static cl::opt<bool> LoopVectorizeWithBlockFrequency( |
309 | "loop-vectorize-with-block-frequency" , cl::init(Val: true), cl::Hidden, |
310 | cl::desc("Enable the use of the block frequency analysis to access PGO " |
311 | "heuristics minimizing code growth in cold regions and being more " |
312 | "aggressive in hot regions." )); |
313 | |
314 | // Runtime interleave loops for load/store throughput. |
315 | static cl::opt<bool> EnableLoadStoreRuntimeInterleave( |
316 | "enable-loadstore-runtime-interleave" , cl::init(Val: true), cl::Hidden, |
317 | cl::desc( |
318 | "Enable runtime interleaving until load/store ports are saturated" )); |
319 | |
320 | /// The number of stores in a loop that are allowed to need predication. |
321 | static cl::opt<unsigned> NumberOfStoresToPredicate( |
322 | "vectorize-num-stores-pred" , cl::init(Val: 1), cl::Hidden, |
323 | cl::desc("Max number of stores to be predicated behind an if." )); |
324 | |
325 | static cl::opt<bool> EnableIndVarRegisterHeur( |
326 | "enable-ind-var-reg-heur" , cl::init(Val: true), cl::Hidden, |
327 | cl::desc("Count the induction variable only once when interleaving" )); |
328 | |
329 | static cl::opt<bool> EnableCondStoresVectorization( |
330 | "enable-cond-stores-vec" , cl::init(Val: true), cl::Hidden, |
331 | cl::desc("Enable if predication of stores during vectorization." )); |
332 | |
333 | static cl::opt<unsigned> MaxNestedScalarReductionIC( |
334 | "max-nested-scalar-reduction-interleave" , cl::init(Val: 2), cl::Hidden, |
335 | cl::desc("The maximum interleave count to use when interleaving a scalar " |
336 | "reduction in a nested loop." )); |
337 | |
338 | static cl::opt<bool> |
339 | PreferInLoopReductions("prefer-inloop-reductions" , cl::init(Val: false), |
340 | cl::Hidden, |
341 | cl::desc("Prefer in-loop vector reductions, " |
342 | "overriding the targets preference." )); |
343 | |
344 | static cl::opt<bool> ForceOrderedReductions( |
345 | "force-ordered-reductions" , cl::init(Val: false), cl::Hidden, |
346 | cl::desc("Enable the vectorisation of loops with in-order (strict) " |
347 | "FP reductions" )); |
348 | |
349 | static cl::opt<bool> PreferPredicatedReductionSelect( |
350 | "prefer-predicated-reduction-select" , cl::init(Val: false), cl::Hidden, |
351 | cl::desc( |
352 | "Prefer predicating a reduction operation over an after loop select." )); |
353 | |
354 | cl::opt<bool> llvm::EnableVPlanNativePath( |
355 | "enable-vplan-native-path" , cl::Hidden, |
356 | cl::desc("Enable VPlan-native vectorization path with " |
357 | "support for outer loop vectorization." )); |
358 | |
359 | cl::opt<bool> |
360 | llvm::VerifyEachVPlan("vplan-verify-each" , |
361 | #ifdef EXPENSIVE_CHECKS |
362 | cl::init(true), |
363 | #else |
364 | cl::init(Val: false), |
365 | #endif |
366 | cl::Hidden, |
367 | cl::desc("Verfiy VPlans after VPlan transforms." )); |
368 | |
369 | // This flag enables the stress testing of the VPlan H-CFG construction in the |
370 | // VPlan-native vectorization path. It must be used in conjuction with |
371 | // -enable-vplan-native-path. -vplan-verify-hcfg can also be used to enable the |
372 | // verification of the H-CFGs built. |
373 | static cl::opt<bool> VPlanBuildStressTest( |
374 | "vplan-build-stress-test" , cl::init(Val: false), cl::Hidden, |
375 | cl::desc( |
376 | "Build VPlan for every supported loop nest in the function and bail " |
377 | "out right after the build (stress test the VPlan H-CFG construction " |
378 | "in the VPlan-native vectorization path)." )); |
379 | |
380 | cl::opt<bool> llvm::EnableLoopInterleaving( |
381 | "interleave-loops" , cl::init(Val: true), cl::Hidden, |
382 | cl::desc("Enable loop interleaving in Loop vectorization passes" )); |
383 | cl::opt<bool> llvm::EnableLoopVectorization( |
384 | "vectorize-loops" , cl::init(Val: true), cl::Hidden, |
385 | cl::desc("Run the Loop vectorization passes" )); |
386 | |
387 | static cl::opt<cl::boolOrDefault> ForceSafeDivisor( |
388 | "force-widen-divrem-via-safe-divisor" , cl::Hidden, |
389 | cl::desc( |
390 | "Override cost based safe divisor widening for div/rem instructions" )); |
391 | |
392 | static cl::opt<bool> UseWiderVFIfCallVariantsPresent( |
393 | "vectorizer-maximize-bandwidth-for-vector-calls" , cl::init(Val: true), |
394 | cl::Hidden, |
395 | cl::desc("Try wider VFs if they enable the use of vector variants" )); |
396 | |
397 | static cl::opt<bool> EnableEarlyExitVectorization( |
398 | "enable-early-exit-vectorization" , cl::init(Val: true), cl::Hidden, |
399 | cl::desc( |
400 | "Enable vectorization of early exit loops with uncountable exits." )); |
401 | |
402 | // Likelyhood of bypassing the vectorized loop because assumptions about SCEV |
403 | // variables not overflowing do not hold. See `emitSCEVChecks`. |
404 | static constexpr uint32_t SCEVCheckBypassWeights[] = {1, 127}; |
405 | // Likelyhood of bypassing the vectorized loop because pointers overlap. See |
406 | // `emitMemRuntimeChecks`. |
407 | static constexpr uint32_t MemCheckBypassWeights[] = {1, 127}; |
408 | // Likelyhood of bypassing the vectorized loop because there are zero trips left |
409 | // after prolog. See `emitIterationCountCheck`. |
410 | static constexpr uint32_t MinItersBypassWeights[] = {1, 127}; |
411 | |
412 | /// A helper function that returns true if the given type is irregular. The |
413 | /// type is irregular if its allocated size doesn't equal the store size of an |
414 | /// element of the corresponding vector type. |
415 | static bool hasIrregularType(Type *Ty, const DataLayout &DL) { |
416 | // Determine if an array of N elements of type Ty is "bitcast compatible" |
417 | // with a <N x Ty> vector. |
418 | // This is only true if there is no padding between the array elements. |
419 | return DL.getTypeAllocSizeInBits(Ty) != DL.getTypeSizeInBits(Ty); |
420 | } |
421 | |
422 | /// A version of ScalarEvolution::getSmallConstantTripCount that returns an |
423 | /// ElementCount to include loops whose trip count is a function of vscale. |
424 | static ElementCount getSmallConstantTripCount(ScalarEvolution *SE, |
425 | const Loop *L) { |
426 | return ElementCount::getFixed(MinVal: SE->getSmallConstantTripCount(L)); |
427 | } |
428 | |
429 | /// Returns "best known" trip count, which is either a valid positive trip count |
430 | /// or std::nullopt when an estimate cannot be made (including when the trip |
431 | /// count would overflow), for the specified loop \p L as defined by the |
432 | /// following procedure: |
433 | /// 1) Returns exact trip count if it is known. |
434 | /// 2) Returns expected trip count according to profile data if any. |
435 | /// 3) Returns upper bound estimate if known, and if \p CanUseConstantMax. |
436 | /// 4) Returns std::nullopt if all of the above failed. |
437 | static std::optional<ElementCount> |
438 | getSmallBestKnownTC(PredicatedScalarEvolution &PSE, Loop *L, |
439 | bool CanUseConstantMax = true) { |
440 | // Check if exact trip count is known. |
441 | if (auto ExpectedTC = getSmallConstantTripCount(SE: PSE.getSE(), L)) |
442 | return ExpectedTC; |
443 | |
444 | // Check if there is an expected trip count available from profile data. |
445 | if (LoopVectorizeWithBlockFrequency) |
446 | if (auto EstimatedTC = getLoopEstimatedTripCount(L)) |
447 | return ElementCount::getFixed(MinVal: *EstimatedTC); |
448 | |
449 | if (!CanUseConstantMax) |
450 | return std::nullopt; |
451 | |
452 | // Check if upper bound estimate is known. |
453 | if (unsigned ExpectedTC = PSE.getSmallConstantMaxTripCount()) |
454 | return ElementCount::getFixed(MinVal: ExpectedTC); |
455 | |
456 | return std::nullopt; |
457 | } |
458 | |
459 | namespace { |
460 | // Forward declare GeneratedRTChecks. |
461 | class GeneratedRTChecks; |
462 | |
463 | using SCEV2ValueTy = DenseMap<const SCEV *, Value *>; |
464 | } // namespace |
465 | |
466 | namespace llvm { |
467 | |
468 | AnalysisKey ShouldRunExtraVectorPasses::; |
469 | |
470 | /// InnerLoopVectorizer vectorizes loops which contain only one basic |
471 | /// block to a specified vectorization factor (VF). |
472 | /// This class performs the widening of scalars into vectors, or multiple |
473 | /// scalars. This class also implements the following features: |
474 | /// * It inserts an epilogue loop for handling loops that don't have iteration |
475 | /// counts that are known to be a multiple of the vectorization factor. |
476 | /// * It handles the code generation for reduction variables. |
477 | /// * Scalarization (implementation using scalars) of un-vectorizable |
478 | /// instructions. |
479 | /// InnerLoopVectorizer does not perform any vectorization-legality |
480 | /// checks, and relies on the caller to check for the different legality |
481 | /// aspects. The InnerLoopVectorizer relies on the |
482 | /// LoopVectorizationLegality class to provide information about the induction |
483 | /// and reduction variables that were found to a given vectorization factor. |
484 | class InnerLoopVectorizer { |
485 | public: |
486 | InnerLoopVectorizer(Loop *OrigLoop, PredicatedScalarEvolution &PSE, |
487 | LoopInfo *LI, DominatorTree *DT, |
488 | const TargetLibraryInfo *TLI, |
489 | const TargetTransformInfo *TTI, AssumptionCache *AC, |
490 | OptimizationRemarkEmitter *ORE, ElementCount VecWidth, |
491 | ElementCount MinProfitableTripCount, |
492 | unsigned UnrollFactor, LoopVectorizationCostModel *CM, |
493 | BlockFrequencyInfo *BFI, ProfileSummaryInfo *PSI, |
494 | GeneratedRTChecks &RTChecks, VPlan &Plan) |
495 | : OrigLoop(OrigLoop), PSE(PSE), LI(LI), DT(DT), TLI(TLI), TTI(TTI), |
496 | AC(AC), ORE(ORE), VF(VecWidth), |
497 | MinProfitableTripCount(MinProfitableTripCount), UF(UnrollFactor), |
498 | Builder(PSE.getSE()->getContext()), Cost(CM), BFI(BFI), PSI(PSI), |
499 | RTChecks(RTChecks), Plan(Plan), |
500 | VectorPHVPB(Plan.getVectorLoopRegion()->getSinglePredecessor()) {} |
501 | |
502 | virtual ~InnerLoopVectorizer() = default; |
503 | |
504 | /// Create a new empty loop that will contain vectorized instructions later |
505 | /// on, while the old loop will be used as the scalar remainder. Control flow |
506 | /// is generated around the vectorized (and scalar epilogue) loops consisting |
507 | /// of various checks and bypasses. Return the pre-header block of the new |
508 | /// loop. In the case of epilogue vectorization, this function is overriden to |
509 | /// handle the more complex control flow around the loops. |
510 | virtual BasicBlock *createVectorizedLoopSkeleton(); |
511 | |
512 | /// Fix the vectorized code, taking care of header phi's, and more. |
513 | void fixVectorizedLoop(VPTransformState &State); |
514 | |
515 | /// Fix the non-induction PHIs in \p Plan. |
516 | void fixNonInductionPHIs(VPTransformState &State); |
517 | |
518 | /// Returns the original loop trip count. |
519 | Value *getTripCount() const { return TripCount; } |
520 | |
521 | /// Used to set the trip count after ILV's construction and after the |
522 | /// preheader block has been executed. Note that this always holds the trip |
523 | /// count of the original loop for both main loop and epilogue vectorization. |
524 | void setTripCount(Value *TC) { TripCount = TC; } |
525 | |
526 | /// Return the additional bypass block which targets the scalar loop by |
527 | /// skipping the epilogue loop after completing the main loop. |
528 | BasicBlock *getAdditionalBypassBlock() const { |
529 | assert(AdditionalBypassBlock && |
530 | "Trying to access AdditionalBypassBlock but it has not been set" ); |
531 | return AdditionalBypassBlock; |
532 | } |
533 | |
534 | protected: |
535 | friend class LoopVectorizationPlanner; |
536 | |
537 | /// Returns (and creates if needed) the trip count of the widened loop. |
538 | Value *getOrCreateVectorTripCount(BasicBlock *InsertBlock); |
539 | |
540 | // Create a check to see if the vector loop should be executed |
541 | Value *createIterationCountCheck(ElementCount VF, unsigned UF) const; |
542 | |
543 | /// Emit a bypass check to see if the vector trip count is zero, including if |
544 | /// it overflows. |
545 | void emitIterationCountCheck(BasicBlock *Bypass); |
546 | |
547 | /// Emit a bypass check to see if all of the SCEV assumptions we've |
548 | /// had to make are correct. Returns the block containing the checks or |
549 | /// nullptr if no checks have been added. |
550 | BasicBlock *emitSCEVChecks(BasicBlock *Bypass); |
551 | |
552 | /// Emit bypass checks to check any memory assumptions we may have made. |
553 | /// Returns the block containing the checks or nullptr if no checks have been |
554 | /// added. |
555 | BasicBlock *emitMemRuntimeChecks(BasicBlock *Bypass); |
556 | |
557 | /// Emit basic blocks (prefixed with \p Prefix) for the iteration check, |
558 | /// vector loop preheader, middle block and scalar preheader. |
559 | void createVectorLoopSkeleton(StringRef Prefix); |
560 | |
561 | /// Allow subclasses to override and print debug traces before/after vplan |
562 | /// execution, when trace information is requested. |
563 | virtual void printDebugTracesAtStart() {} |
564 | virtual void printDebugTracesAtEnd() {} |
565 | |
566 | /// Introduces a new VPIRBasicBlock for \p CheckIRBB to Plan between the |
567 | /// vector preheader and its predecessor, also connecting the new block to the |
568 | /// scalar preheader. |
569 | void introduceCheckBlockInVPlan(BasicBlock *CheckIRBB); |
570 | |
571 | /// The original loop. |
572 | Loop *OrigLoop; |
573 | |
574 | /// A wrapper around ScalarEvolution used to add runtime SCEV checks. Applies |
575 | /// dynamic knowledge to simplify SCEV expressions and converts them to a |
576 | /// more usable form. |
577 | PredicatedScalarEvolution &PSE; |
578 | |
579 | /// Loop Info. |
580 | LoopInfo *LI; |
581 | |
582 | /// Dominator Tree. |
583 | DominatorTree *DT; |
584 | |
585 | /// Target Library Info. |
586 | const TargetLibraryInfo *TLI; |
587 | |
588 | /// Target Transform Info. |
589 | const TargetTransformInfo *TTI; |
590 | |
591 | /// Assumption Cache. |
592 | AssumptionCache *AC; |
593 | |
594 | /// Interface to emit optimization remarks. |
595 | OptimizationRemarkEmitter *ORE; |
596 | |
597 | /// The vectorization SIMD factor to use. Each vector will have this many |
598 | /// vector elements. |
599 | ElementCount VF; |
600 | |
601 | ElementCount MinProfitableTripCount; |
602 | |
603 | /// The vectorization unroll factor to use. Each scalar is vectorized to this |
604 | /// many different vector instructions. |
605 | unsigned UF; |
606 | |
607 | /// The builder that we use |
608 | IRBuilder<> Builder; |
609 | |
610 | // --- Vectorization state --- |
611 | |
612 | /// The vector-loop preheader. |
613 | BasicBlock * = nullptr; |
614 | |
615 | /// The scalar-loop preheader. |
616 | BasicBlock * = nullptr; |
617 | |
618 | /// Middle Block between the vector and the scalar. |
619 | BasicBlock *LoopMiddleBlock = nullptr; |
620 | |
621 | /// Trip count of the original loop. |
622 | Value *TripCount = nullptr; |
623 | |
624 | /// Trip count of the widened loop (TripCount - TripCount % (VF*UF)) |
625 | Value *VectorTripCount = nullptr; |
626 | |
627 | /// The profitablity analysis. |
628 | LoopVectorizationCostModel *Cost; |
629 | |
630 | /// BFI and PSI are used to check for profile guided size optimizations. |
631 | BlockFrequencyInfo *BFI; |
632 | ProfileSummaryInfo *PSI; |
633 | |
634 | /// Structure to hold information about generated runtime checks, responsible |
635 | /// for cleaning the checks, if vectorization turns out unprofitable. |
636 | GeneratedRTChecks &RTChecks; |
637 | |
638 | /// The additional bypass block which conditionally skips over the epilogue |
639 | /// loop after executing the main loop. Needed to resume inductions and |
640 | /// reductions during epilogue vectorization. |
641 | BasicBlock *AdditionalBypassBlock = nullptr; |
642 | |
643 | VPlan &Plan; |
644 | |
645 | /// The vector preheader block of \p Plan, used as target for check blocks |
646 | /// introduced during skeleton creation. |
647 | VPBlockBase *VectorPHVPB; |
648 | }; |
649 | |
650 | /// Encapsulate information regarding vectorization of a loop and its epilogue. |
651 | /// This information is meant to be updated and used across two stages of |
652 | /// epilogue vectorization. |
653 | struct EpilogueLoopVectorizationInfo { |
654 | ElementCount MainLoopVF = ElementCount::getFixed(MinVal: 0); |
655 | unsigned MainLoopUF = 0; |
656 | ElementCount EpilogueVF = ElementCount::getFixed(MinVal: 0); |
657 | unsigned EpilogueUF = 0; |
658 | BasicBlock *MainLoopIterationCountCheck = nullptr; |
659 | BasicBlock *EpilogueIterationCountCheck = nullptr; |
660 | BasicBlock *SCEVSafetyCheck = nullptr; |
661 | BasicBlock *MemSafetyCheck = nullptr; |
662 | Value *TripCount = nullptr; |
663 | Value *VectorTripCount = nullptr; |
664 | VPlan &EpiloguePlan; |
665 | |
666 | EpilogueLoopVectorizationInfo(ElementCount MVF, unsigned MUF, |
667 | ElementCount EVF, unsigned EUF, |
668 | VPlan &EpiloguePlan) |
669 | : MainLoopVF(MVF), MainLoopUF(MUF), EpilogueVF(EVF), EpilogueUF(EUF), |
670 | EpiloguePlan(EpiloguePlan) { |
671 | assert(EUF == 1 && |
672 | "A high UF for the epilogue loop is likely not beneficial." ); |
673 | } |
674 | }; |
675 | |
676 | /// An extension of the inner loop vectorizer that creates a skeleton for a |
677 | /// vectorized loop that has its epilogue (residual) also vectorized. |
678 | /// The idea is to run the vplan on a given loop twice, firstly to setup the |
679 | /// skeleton and vectorize the main loop, and secondly to complete the skeleton |
680 | /// from the first step and vectorize the epilogue. This is achieved by |
681 | /// deriving two concrete strategy classes from this base class and invoking |
682 | /// them in succession from the loop vectorizer planner. |
683 | class InnerLoopAndEpilogueVectorizer : public InnerLoopVectorizer { |
684 | public: |
685 | InnerLoopAndEpilogueVectorizer( |
686 | Loop *OrigLoop, PredicatedScalarEvolution &PSE, LoopInfo *LI, |
687 | DominatorTree *DT, const TargetLibraryInfo *TLI, |
688 | const TargetTransformInfo *TTI, AssumptionCache *AC, |
689 | OptimizationRemarkEmitter *ORE, EpilogueLoopVectorizationInfo &EPI, |
690 | LoopVectorizationCostModel *CM, BlockFrequencyInfo *BFI, |
691 | ProfileSummaryInfo *PSI, GeneratedRTChecks &Checks, VPlan &Plan) |
692 | : InnerLoopVectorizer(OrigLoop, PSE, LI, DT, TLI, TTI, AC, ORE, |
693 | EPI.MainLoopVF, EPI.MainLoopVF, EPI.MainLoopUF, CM, |
694 | BFI, PSI, Checks, Plan), |
695 | EPI(EPI) {} |
696 | |
697 | // Override this function to handle the more complex control flow around the |
698 | // three loops. |
699 | BasicBlock *createVectorizedLoopSkeleton() final { |
700 | return createEpilogueVectorizedLoopSkeleton(); |
701 | } |
702 | |
703 | /// The interface for creating a vectorized skeleton using one of two |
704 | /// different strategies, each corresponding to one execution of the vplan |
705 | /// as described above. |
706 | virtual BasicBlock *createEpilogueVectorizedLoopSkeleton() = 0; |
707 | |
708 | /// Holds and updates state information required to vectorize the main loop |
709 | /// and its epilogue in two separate passes. This setup helps us avoid |
710 | /// regenerating and recomputing runtime safety checks. It also helps us to |
711 | /// shorten the iteration-count-check path length for the cases where the |
712 | /// iteration count of the loop is so small that the main vector loop is |
713 | /// completely skipped. |
714 | EpilogueLoopVectorizationInfo &EPI; |
715 | }; |
716 | |
717 | /// A specialized derived class of inner loop vectorizer that performs |
718 | /// vectorization of *main* loops in the process of vectorizing loops and their |
719 | /// epilogues. |
720 | class EpilogueVectorizerMainLoop : public InnerLoopAndEpilogueVectorizer { |
721 | public: |
722 | EpilogueVectorizerMainLoop( |
723 | Loop *OrigLoop, PredicatedScalarEvolution &PSE, LoopInfo *LI, |
724 | DominatorTree *DT, const TargetLibraryInfo *TLI, |
725 | const TargetTransformInfo *TTI, AssumptionCache *AC, |
726 | OptimizationRemarkEmitter *ORE, EpilogueLoopVectorizationInfo &EPI, |
727 | LoopVectorizationCostModel *CM, BlockFrequencyInfo *BFI, |
728 | ProfileSummaryInfo *PSI, GeneratedRTChecks &Check, VPlan &Plan) |
729 | : InnerLoopAndEpilogueVectorizer(OrigLoop, PSE, LI, DT, TLI, TTI, AC, ORE, |
730 | EPI, CM, BFI, PSI, Check, Plan) {} |
731 | /// Implements the interface for creating a vectorized skeleton using the |
732 | /// *main loop* strategy (ie the first pass of vplan execution). |
733 | BasicBlock *createEpilogueVectorizedLoopSkeleton() final; |
734 | |
735 | protected: |
736 | /// Emits an iteration count bypass check once for the main loop (when \p |
737 | /// ForEpilogue is false) and once for the epilogue loop (when \p |
738 | /// ForEpilogue is true). |
739 | BasicBlock *emitIterationCountCheck(BasicBlock *Bypass, bool ForEpilogue); |
740 | void printDebugTracesAtStart() override; |
741 | void printDebugTracesAtEnd() override; |
742 | }; |
743 | |
744 | // A specialized derived class of inner loop vectorizer that performs |
745 | // vectorization of *epilogue* loops in the process of vectorizing loops and |
746 | // their epilogues. |
747 | class EpilogueVectorizerEpilogueLoop : public InnerLoopAndEpilogueVectorizer { |
748 | public: |
749 | EpilogueVectorizerEpilogueLoop( |
750 | Loop *OrigLoop, PredicatedScalarEvolution &PSE, LoopInfo *LI, |
751 | DominatorTree *DT, const TargetLibraryInfo *TLI, |
752 | const TargetTransformInfo *TTI, AssumptionCache *AC, |
753 | OptimizationRemarkEmitter *ORE, EpilogueLoopVectorizationInfo &EPI, |
754 | LoopVectorizationCostModel *CM, BlockFrequencyInfo *BFI, |
755 | ProfileSummaryInfo *PSI, GeneratedRTChecks &Checks, VPlan &Plan) |
756 | : InnerLoopAndEpilogueVectorizer(OrigLoop, PSE, LI, DT, TLI, TTI, AC, ORE, |
757 | EPI, CM, BFI, PSI, Checks, Plan) { |
758 | TripCount = EPI.TripCount; |
759 | } |
760 | /// Implements the interface for creating a vectorized skeleton using the |
761 | /// *epilogue loop* strategy (ie the second pass of vplan execution). |
762 | BasicBlock *createEpilogueVectorizedLoopSkeleton() final; |
763 | |
764 | protected: |
765 | /// Emits an iteration count bypass check after the main vector loop has |
766 | /// finished to see if there are any iterations left to execute by either |
767 | /// the vector epilogue or the scalar epilogue. |
768 | BasicBlock *emitMinimumVectorEpilogueIterCountCheck( |
769 | BasicBlock *Bypass, |
770 | BasicBlock *Insert); |
771 | void printDebugTracesAtStart() override; |
772 | void printDebugTracesAtEnd() override; |
773 | }; |
774 | } // end namespace llvm |
775 | |
776 | /// Look for a meaningful debug location on the instruction or its operands. |
777 | static DebugLoc getDebugLocFromInstOrOperands(Instruction *I) { |
778 | if (!I) |
779 | return DebugLoc::getUnknown(); |
780 | |
781 | DebugLoc Empty; |
782 | if (I->getDebugLoc() != Empty) |
783 | return I->getDebugLoc(); |
784 | |
785 | for (Use &Op : I->operands()) { |
786 | if (Instruction *OpInst = dyn_cast<Instruction>(Val&: Op)) |
787 | if (OpInst->getDebugLoc() != Empty) |
788 | return OpInst->getDebugLoc(); |
789 | } |
790 | |
791 | return I->getDebugLoc(); |
792 | } |
793 | |
794 | /// Write a \p DebugMsg about vectorization to the debug output stream. If \p I |
795 | /// is passed, the message relates to that particular instruction. |
796 | #ifndef NDEBUG |
797 | static void debugVectorizationMessage(const StringRef Prefix, |
798 | const StringRef DebugMsg, |
799 | Instruction *I) { |
800 | dbgs() << "LV: " << Prefix << DebugMsg; |
801 | if (I != nullptr) |
802 | dbgs() << " " << *I; |
803 | else |
804 | dbgs() << '.'; |
805 | dbgs() << '\n'; |
806 | } |
807 | #endif |
808 | |
809 | /// Create an analysis remark that explains why vectorization failed |
810 | /// |
811 | /// \p PassName is the name of the pass (e.g. can be AlwaysPrint). \p |
812 | /// RemarkName is the identifier for the remark. If \p I is passed it is an |
813 | /// instruction that prevents vectorization. Otherwise \p TheLoop is used for |
814 | /// the location of the remark. If \p DL is passed, use it as debug location for |
815 | /// the remark. \return the remark object that can be streamed to. |
816 | static OptimizationRemarkAnalysis |
817 | createLVAnalysis(const char *PassName, StringRef , Loop *TheLoop, |
818 | Instruction *I, DebugLoc DL = {}) { |
819 | BasicBlock *CodeRegion = I ? I->getParent() : TheLoop->getHeader(); |
820 | // If debug location is attached to the instruction, use it. Otherwise if DL |
821 | // was not provided, use the loop's. |
822 | if (I && I->getDebugLoc()) |
823 | DL = I->getDebugLoc(); |
824 | else if (!DL) |
825 | DL = TheLoop->getStartLoc(); |
826 | |
827 | return OptimizationRemarkAnalysis(PassName, RemarkName, DL, CodeRegion); |
828 | } |
829 | |
830 | namespace llvm { |
831 | |
832 | /// Return a value for Step multiplied by VF. |
833 | Value *createStepForVF(IRBuilderBase &B, Type *Ty, ElementCount VF, |
834 | int64_t Step) { |
835 | assert(Ty->isIntegerTy() && "Expected an integer step" ); |
836 | return B.CreateElementCount(Ty, EC: VF.multiplyCoefficientBy(RHS: Step)); |
837 | } |
838 | |
839 | /// Return the runtime value for VF. |
840 | Value *getRuntimeVF(IRBuilderBase &B, Type *Ty, ElementCount VF) { |
841 | return B.CreateElementCount(Ty, EC: VF); |
842 | } |
843 | |
844 | void (const StringRef DebugMsg, |
845 | const StringRef OREMsg, const StringRef ORETag, |
846 | OptimizationRemarkEmitter *ORE, Loop *TheLoop, |
847 | Instruction *I) { |
848 | LLVM_DEBUG(debugVectorizationMessage("Not vectorizing: " , DebugMsg, I)); |
849 | LoopVectorizeHints Hints(TheLoop, true /* doesn't matter */, *ORE); |
850 | ORE->emit( |
851 | OptDiag: createLVAnalysis(PassName: Hints.vectorizeAnalysisPassName(), RemarkName: ORETag, TheLoop, I) |
852 | << "loop not vectorized: " << OREMsg); |
853 | } |
854 | |
855 | /// Reports an informative message: print \p Msg for debugging purposes as well |
856 | /// as an optimization remark. Uses either \p I as location of the remark, or |
857 | /// otherwise \p TheLoop. If \p DL is passed, use it as debug location for the |
858 | /// remark. If \p DL is passed, use it as debug location for the remark. |
859 | static void (const StringRef Msg, const StringRef ORETag, |
860 | OptimizationRemarkEmitter *ORE, |
861 | Loop *TheLoop, Instruction *I = nullptr, |
862 | DebugLoc DL = {}) { |
863 | LLVM_DEBUG(debugVectorizationMessage("" , Msg, I)); |
864 | LoopVectorizeHints Hints(TheLoop, true /* doesn't matter */, *ORE); |
865 | ORE->emit(OptDiag: createLVAnalysis(PassName: Hints.vectorizeAnalysisPassName(), RemarkName: ORETag, TheLoop, |
866 | I, DL) |
867 | << Msg); |
868 | } |
869 | |
870 | /// Report successful vectorization of the loop. In case an outer loop is |
871 | /// vectorized, prepend "outer" to the vectorization remark. |
872 | static void (OptimizationRemarkEmitter *ORE, Loop *TheLoop, |
873 | VectorizationFactor VF, unsigned IC) { |
874 | LLVM_DEBUG(debugVectorizationMessage( |
875 | "Vectorizing: " , TheLoop->isInnermost() ? "innermost loop" : "outer loop" , |
876 | nullptr)); |
877 | StringRef LoopType = TheLoop->isInnermost() ? "" : "outer " ; |
878 | ORE->emit(RemarkBuilder: [&]() { |
879 | return OptimizationRemark(LV_NAME, "Vectorized" , TheLoop->getStartLoc(), |
880 | TheLoop->getHeader()) |
881 | << "vectorized " << LoopType << "loop (vectorization width: " |
882 | << ore::NV("VectorizationFactor" , VF.Width) |
883 | << ", interleaved count: " << ore::NV("InterleaveCount" , IC) << ")" ; |
884 | }); |
885 | } |
886 | |
887 | } // end namespace llvm |
888 | |
889 | namespace llvm { |
890 | |
891 | // Loop vectorization cost-model hints how the scalar epilogue loop should be |
892 | // lowered. |
893 | enum ScalarEpilogueLowering { |
894 | |
895 | // The default: allowing scalar epilogues. |
896 | CM_ScalarEpilogueAllowed, |
897 | |
898 | // Vectorization with OptForSize: don't allow epilogues. |
899 | CM_ScalarEpilogueNotAllowedOptSize, |
900 | |
901 | // A special case of vectorisation with OptForSize: loops with a very small |
902 | // trip count are considered for vectorization under OptForSize, thereby |
903 | // making sure the cost of their loop body is dominant, free of runtime |
904 | // guards and scalar iteration overheads. |
905 | CM_ScalarEpilogueNotAllowedLowTripLoop, |
906 | |
907 | // Loop hint predicate indicating an epilogue is undesired. |
908 | CM_ScalarEpilogueNotNeededUsePredicate, |
909 | |
910 | // Directive indicating we must either tail fold or not vectorize |
911 | CM_ScalarEpilogueNotAllowedUsePredicate |
912 | }; |
913 | |
914 | /// LoopVectorizationCostModel - estimates the expected speedups due to |
915 | /// vectorization. |
916 | /// In many cases vectorization is not profitable. This can happen because of |
917 | /// a number of reasons. In this class we mainly attempt to predict the |
918 | /// expected speedup/slowdowns due to the supported instruction set. We use the |
919 | /// TargetTransformInfo to query the different backends for the cost of |
920 | /// different operations. |
921 | class LoopVectorizationCostModel { |
922 | friend class LoopVectorizationPlanner; |
923 | |
924 | public: |
925 | LoopVectorizationCostModel(ScalarEpilogueLowering SEL, Loop *L, |
926 | PredicatedScalarEvolution &PSE, LoopInfo *LI, |
927 | LoopVectorizationLegality *Legal, |
928 | const TargetTransformInfo &TTI, |
929 | const TargetLibraryInfo *TLI, DemandedBits *DB, |
930 | AssumptionCache *AC, |
931 | OptimizationRemarkEmitter *ORE, const Function *F, |
932 | const LoopVectorizeHints *Hints, |
933 | InterleavedAccessInfo &IAI, |
934 | ProfileSummaryInfo *PSI, BlockFrequencyInfo *BFI) |
935 | : ScalarEpilogueStatus(SEL), TheLoop(L), PSE(PSE), LI(LI), Legal(Legal), |
936 | TTI(TTI), TLI(TLI), DB(DB), AC(AC), ORE(ORE), TheFunction(F), |
937 | Hints(Hints), InterleaveInfo(IAI) { |
938 | if (TTI.supportsScalableVectors() || ForceTargetSupportsScalableVectors) |
939 | initializeVScaleForTuning(); |
940 | CostKind = F->hasMinSize() ? TTI::TCK_CodeSize : TTI::TCK_RecipThroughput; |
941 | // Query this against the original loop and save it here because the profile |
942 | // of the original loop header may change as the transformation happens. |
943 | OptForSize = llvm::shouldOptimizeForSize(BB: L->getHeader(), PSI, BFI, |
944 | QueryType: PGSOQueryType::IRPass); |
945 | } |
946 | |
947 | /// \return An upper bound for the vectorization factors (both fixed and |
948 | /// scalable). If the factors are 0, vectorization and interleaving should be |
949 | /// avoided up front. |
950 | FixedScalableVFPair computeMaxVF(ElementCount UserVF, unsigned UserIC); |
951 | |
952 | /// \return True if runtime checks are required for vectorization, and false |
953 | /// otherwise. |
954 | bool runtimeChecksRequired(); |
955 | |
956 | /// Setup cost-based decisions for user vectorization factor. |
957 | /// \return true if the UserVF is a feasible VF to be chosen. |
958 | bool selectUserVectorizationFactor(ElementCount UserVF) { |
959 | collectNonVectorizedAndSetWideningDecisions(VF: UserVF); |
960 | return expectedCost(VF: UserVF).isValid(); |
961 | } |
962 | |
963 | /// \return True if maximizing vector bandwidth is enabled by the target or |
964 | /// user options, for the given register kind. |
965 | bool useMaxBandwidth(TargetTransformInfo::RegisterKind RegKind); |
966 | |
967 | /// \return True if maximizing vector bandwidth is enabled by the target or |
968 | /// user options, for the given vector factor. |
969 | bool useMaxBandwidth(ElementCount VF); |
970 | |
971 | /// \return The size (in bits) of the smallest and widest types in the code |
972 | /// that needs to be vectorized. We ignore values that remain scalar such as |
973 | /// 64 bit loop indices. |
974 | std::pair<unsigned, unsigned> getSmallestAndWidestTypes(); |
975 | |
976 | /// \return The desired interleave count. |
977 | /// If interleave count has been specified by metadata it will be returned. |
978 | /// Otherwise, the interleave count is computed and returned. VF and LoopCost |
979 | /// are the selected vectorization factor and the cost of the selected VF. |
980 | unsigned selectInterleaveCount(VPlan &Plan, ElementCount VF, |
981 | InstructionCost LoopCost); |
982 | |
983 | /// Memory access instruction may be vectorized in more than one way. |
984 | /// Form of instruction after vectorization depends on cost. |
985 | /// This function takes cost-based decisions for Load/Store instructions |
986 | /// and collects them in a map. This decisions map is used for building |
987 | /// the lists of loop-uniform and loop-scalar instructions. |
988 | /// The calculated cost is saved with widening decision in order to |
989 | /// avoid redundant calculations. |
990 | void setCostBasedWideningDecision(ElementCount VF); |
991 | |
992 | /// A call may be vectorized in different ways depending on whether we have |
993 | /// vectorized variants available and whether the target supports masking. |
994 | /// This function analyzes all calls in the function at the supplied VF, |
995 | /// makes a decision based on the costs of available options, and stores that |
996 | /// decision in a map for use in planning and plan execution. |
997 | void setVectorizedCallDecision(ElementCount VF); |
998 | |
999 | /// Collect values we want to ignore in the cost model. |
1000 | void collectValuesToIgnore(); |
1001 | |
1002 | /// Collect all element types in the loop for which widening is needed. |
1003 | void collectElementTypesForWidening(); |
1004 | |
1005 | /// Split reductions into those that happen in the loop, and those that happen |
1006 | /// outside. In loop reductions are collected into InLoopReductions. |
1007 | void collectInLoopReductions(); |
1008 | |
1009 | /// Returns true if we should use strict in-order reductions for the given |
1010 | /// RdxDesc. This is true if the -enable-strict-reductions flag is passed, |
1011 | /// the IsOrdered flag of RdxDesc is set and we do not allow reordering |
1012 | /// of FP operations. |
1013 | bool useOrderedReductions(const RecurrenceDescriptor &RdxDesc) const { |
1014 | return !Hints->allowReordering() && RdxDesc.isOrdered(); |
1015 | } |
1016 | |
1017 | /// \returns The smallest bitwidth each instruction can be represented with. |
1018 | /// The vector equivalents of these instructions should be truncated to this |
1019 | /// type. |
1020 | const MapVector<Instruction *, uint64_t> &getMinimalBitwidths() const { |
1021 | return MinBWs; |
1022 | } |
1023 | |
1024 | /// \returns True if it is more profitable to scalarize instruction \p I for |
1025 | /// vectorization factor \p VF. |
1026 | bool isProfitableToScalarize(Instruction *I, ElementCount VF) const { |
1027 | assert(VF.isVector() && |
1028 | "Profitable to scalarize relevant only for VF > 1." ); |
1029 | assert( |
1030 | TheLoop->isInnermost() && |
1031 | "cost-model should not be used for outer loops (in VPlan-native path)" ); |
1032 | |
1033 | auto Scalars = InstsToScalarize.find(Val: VF); |
1034 | assert(Scalars != InstsToScalarize.end() && |
1035 | "VF not yet analyzed for scalarization profitability" ); |
1036 | return Scalars->second.contains(Val: I); |
1037 | } |
1038 | |
1039 | /// Returns true if \p I is known to be uniform after vectorization. |
1040 | bool isUniformAfterVectorization(Instruction *I, ElementCount VF) const { |
1041 | assert( |
1042 | TheLoop->isInnermost() && |
1043 | "cost-model should not be used for outer loops (in VPlan-native path)" ); |
1044 | // Pseudo probe needs to be duplicated for each unrolled iteration and |
1045 | // vector lane so that profiled loop trip count can be accurately |
1046 | // accumulated instead of being under counted. |
1047 | if (isa<PseudoProbeInst>(Val: I)) |
1048 | return false; |
1049 | |
1050 | if (VF.isScalar()) |
1051 | return true; |
1052 | |
1053 | auto UniformsPerVF = Uniforms.find(Val: VF); |
1054 | assert(UniformsPerVF != Uniforms.end() && |
1055 | "VF not yet analyzed for uniformity" ); |
1056 | return UniformsPerVF->second.count(Ptr: I); |
1057 | } |
1058 | |
1059 | /// Returns true if \p I is known to be scalar after vectorization. |
1060 | bool isScalarAfterVectorization(Instruction *I, ElementCount VF) const { |
1061 | assert( |
1062 | TheLoop->isInnermost() && |
1063 | "cost-model should not be used for outer loops (in VPlan-native path)" ); |
1064 | if (VF.isScalar()) |
1065 | return true; |
1066 | |
1067 | auto ScalarsPerVF = Scalars.find(Val: VF); |
1068 | assert(ScalarsPerVF != Scalars.end() && |
1069 | "Scalar values are not calculated for VF" ); |
1070 | return ScalarsPerVF->second.count(Ptr: I); |
1071 | } |
1072 | |
1073 | /// \returns True if instruction \p I can be truncated to a smaller bitwidth |
1074 | /// for vectorization factor \p VF. |
1075 | bool canTruncateToMinimalBitwidth(Instruction *I, ElementCount VF) const { |
1076 | return VF.isVector() && MinBWs.contains(Key: I) && |
1077 | !isProfitableToScalarize(I, VF) && |
1078 | !isScalarAfterVectorization(I, VF); |
1079 | } |
1080 | |
1081 | /// Decision that was taken during cost calculation for memory instruction. |
1082 | enum InstWidening { |
1083 | CM_Unknown, |
1084 | CM_Widen, // For consecutive accesses with stride +1. |
1085 | CM_Widen_Reverse, // For consecutive accesses with stride -1. |
1086 | CM_Interleave, |
1087 | CM_GatherScatter, |
1088 | CM_Scalarize, |
1089 | CM_VectorCall, |
1090 | CM_IntrinsicCall |
1091 | }; |
1092 | |
1093 | /// Save vectorization decision \p W and \p Cost taken by the cost model for |
1094 | /// instruction \p I and vector width \p VF. |
1095 | void setWideningDecision(Instruction *I, ElementCount VF, InstWidening W, |
1096 | InstructionCost Cost) { |
1097 | assert(VF.isVector() && "Expected VF >=2" ); |
1098 | WideningDecisions[{I, VF}] = {W, Cost}; |
1099 | } |
1100 | |
1101 | /// Save vectorization decision \p W and \p Cost taken by the cost model for |
1102 | /// interleaving group \p Grp and vector width \p VF. |
1103 | void setWideningDecision(const InterleaveGroup<Instruction> *Grp, |
1104 | ElementCount VF, InstWidening W, |
1105 | InstructionCost Cost) { |
1106 | assert(VF.isVector() && "Expected VF >=2" ); |
1107 | /// Broadcast this decicion to all instructions inside the group. |
1108 | /// When interleaving, the cost will only be assigned one instruction, the |
1109 | /// insert position. For other cases, add the appropriate fraction of the |
1110 | /// total cost to each instruction. This ensures accurate costs are used, |
1111 | /// even if the insert position instruction is not used. |
1112 | InstructionCost InsertPosCost = Cost; |
1113 | InstructionCost OtherMemberCost = 0; |
1114 | if (W != CM_Interleave) |
1115 | OtherMemberCost = InsertPosCost = Cost / Grp->getNumMembers(); |
1116 | ; |
1117 | for (unsigned Idx = 0; Idx < Grp->getFactor(); ++Idx) { |
1118 | if (auto *I = Grp->getMember(Index: Idx)) { |
1119 | if (Grp->getInsertPos() == I) |
1120 | WideningDecisions[{I, VF}] = {W, InsertPosCost}; |
1121 | else |
1122 | WideningDecisions[{I, VF}] = {W, OtherMemberCost}; |
1123 | } |
1124 | } |
1125 | } |
1126 | |
1127 | /// Return the cost model decision for the given instruction \p I and vector |
1128 | /// width \p VF. Return CM_Unknown if this instruction did not pass |
1129 | /// through the cost modeling. |
1130 | InstWidening getWideningDecision(Instruction *I, ElementCount VF) const { |
1131 | assert(VF.isVector() && "Expected VF to be a vector VF" ); |
1132 | assert( |
1133 | TheLoop->isInnermost() && |
1134 | "cost-model should not be used for outer loops (in VPlan-native path)" ); |
1135 | |
1136 | std::pair<Instruction *, ElementCount> InstOnVF(I, VF); |
1137 | auto Itr = WideningDecisions.find(Val: InstOnVF); |
1138 | if (Itr == WideningDecisions.end()) |
1139 | return CM_Unknown; |
1140 | return Itr->second.first; |
1141 | } |
1142 | |
1143 | /// Return the vectorization cost for the given instruction \p I and vector |
1144 | /// width \p VF. |
1145 | InstructionCost getWideningCost(Instruction *I, ElementCount VF) { |
1146 | assert(VF.isVector() && "Expected VF >=2" ); |
1147 | std::pair<Instruction *, ElementCount> InstOnVF(I, VF); |
1148 | assert(WideningDecisions.contains(InstOnVF) && |
1149 | "The cost is not calculated" ); |
1150 | return WideningDecisions[InstOnVF].second; |
1151 | } |
1152 | |
1153 | struct CallWideningDecision { |
1154 | InstWidening Kind; |
1155 | Function *Variant; |
1156 | Intrinsic::ID IID; |
1157 | std::optional<unsigned> MaskPos; |
1158 | InstructionCost Cost; |
1159 | }; |
1160 | |
1161 | void setCallWideningDecision(CallInst *CI, ElementCount VF, InstWidening Kind, |
1162 | Function *Variant, Intrinsic::ID IID, |
1163 | std::optional<unsigned> MaskPos, |
1164 | InstructionCost Cost) { |
1165 | assert(!VF.isScalar() && "Expected vector VF" ); |
1166 | CallWideningDecisions[{CI, VF}] = {.Kind: Kind, .Variant: Variant, .IID: IID, .MaskPos: MaskPos, .Cost: Cost}; |
1167 | } |
1168 | |
1169 | CallWideningDecision getCallWideningDecision(CallInst *CI, |
1170 | ElementCount VF) const { |
1171 | assert(!VF.isScalar() && "Expected vector VF" ); |
1172 | return CallWideningDecisions.at(Val: {CI, VF}); |
1173 | } |
1174 | |
1175 | /// Return True if instruction \p I is an optimizable truncate whose operand |
1176 | /// is an induction variable. Such a truncate will be removed by adding a new |
1177 | /// induction variable with the destination type. |
1178 | bool isOptimizableIVTruncate(Instruction *I, ElementCount VF) { |
1179 | // If the instruction is not a truncate, return false. |
1180 | auto *Trunc = dyn_cast<TruncInst>(Val: I); |
1181 | if (!Trunc) |
1182 | return false; |
1183 | |
1184 | // Get the source and destination types of the truncate. |
1185 | Type *SrcTy = toVectorTy(Scalar: Trunc->getSrcTy(), EC: VF); |
1186 | Type *DestTy = toVectorTy(Scalar: Trunc->getDestTy(), EC: VF); |
1187 | |
1188 | // If the truncate is free for the given types, return false. Replacing a |
1189 | // free truncate with an induction variable would add an induction variable |
1190 | // update instruction to each iteration of the loop. We exclude from this |
1191 | // check the primary induction variable since it will need an update |
1192 | // instruction regardless. |
1193 | Value *Op = Trunc->getOperand(i_nocapture: 0); |
1194 | if (Op != Legal->getPrimaryInduction() && TTI.isTruncateFree(Ty1: SrcTy, Ty2: DestTy)) |
1195 | return false; |
1196 | |
1197 | // If the truncated value is not an induction variable, return false. |
1198 | return Legal->isInductionPhi(V: Op); |
1199 | } |
1200 | |
1201 | /// Collects the instructions to scalarize for each predicated instruction in |
1202 | /// the loop. |
1203 | void collectInstsToScalarize(ElementCount VF); |
1204 | |
1205 | /// Collect values that will not be widened, including Uniforms, Scalars, and |
1206 | /// Instructions to Scalarize for the given \p VF. |
1207 | /// The sets depend on CM decision for Load/Store instructions |
1208 | /// that may be vectorized as interleave, gather-scatter or scalarized. |
1209 | /// Also make a decision on what to do about call instructions in the loop |
1210 | /// at that VF -- scalarize, call a known vector routine, or call a |
1211 | /// vector intrinsic. |
1212 | void collectNonVectorizedAndSetWideningDecisions(ElementCount VF) { |
1213 | // Do the analysis once. |
1214 | if (VF.isScalar() || Uniforms.contains(Val: VF)) |
1215 | return; |
1216 | setCostBasedWideningDecision(VF); |
1217 | collectLoopUniforms(VF); |
1218 | setVectorizedCallDecision(VF); |
1219 | collectLoopScalars(VF); |
1220 | collectInstsToScalarize(VF); |
1221 | } |
1222 | |
1223 | /// Returns true if the target machine supports masked store operation |
1224 | /// for the given \p DataType and kind of access to \p Ptr. |
1225 | bool isLegalMaskedStore(Type *DataType, Value *Ptr, Align Alignment, |
1226 | unsigned AddressSpace) const { |
1227 | return Legal->isConsecutivePtr(AccessTy: DataType, Ptr) && |
1228 | TTI.isLegalMaskedStore(DataType, Alignment, AddressSpace); |
1229 | } |
1230 | |
1231 | /// Returns true if the target machine supports masked load operation |
1232 | /// for the given \p DataType and kind of access to \p Ptr. |
1233 | bool isLegalMaskedLoad(Type *DataType, Value *Ptr, Align Alignment, |
1234 | unsigned AddressSpace) const { |
1235 | return Legal->isConsecutivePtr(AccessTy: DataType, Ptr) && |
1236 | TTI.isLegalMaskedLoad(DataType, Alignment, AddressSpace); |
1237 | } |
1238 | |
1239 | /// Returns true if the target machine can represent \p V as a masked gather |
1240 | /// or scatter operation. |
1241 | bool isLegalGatherOrScatter(Value *V, ElementCount VF) { |
1242 | bool LI = isa<LoadInst>(Val: V); |
1243 | bool SI = isa<StoreInst>(Val: V); |
1244 | if (!LI && !SI) |
1245 | return false; |
1246 | auto *Ty = getLoadStoreType(I: V); |
1247 | Align Align = getLoadStoreAlignment(I: V); |
1248 | if (VF.isVector()) |
1249 | Ty = VectorType::get(ElementType: Ty, EC: VF); |
1250 | return (LI && TTI.isLegalMaskedGather(DataType: Ty, Alignment: Align)) || |
1251 | (SI && TTI.isLegalMaskedScatter(DataType: Ty, Alignment: Align)); |
1252 | } |
1253 | |
1254 | /// Returns true if the target machine supports all of the reduction |
1255 | /// variables found for the given VF. |
1256 | bool canVectorizeReductions(ElementCount VF) const { |
1257 | return (all_of(Range: Legal->getReductionVars(), P: [&](auto &Reduction) -> bool { |
1258 | const RecurrenceDescriptor &RdxDesc = Reduction.second; |
1259 | return TTI.isLegalToVectorizeReduction(RdxDesc, VF); |
1260 | })); |
1261 | } |
1262 | |
1263 | /// Given costs for both strategies, return true if the scalar predication |
1264 | /// lowering should be used for div/rem. This incorporates an override |
1265 | /// option so it is not simply a cost comparison. |
1266 | bool isDivRemScalarWithPredication(InstructionCost ScalarCost, |
1267 | InstructionCost SafeDivisorCost) const { |
1268 | switch (ForceSafeDivisor) { |
1269 | case cl::BOU_UNSET: |
1270 | return ScalarCost < SafeDivisorCost; |
1271 | case cl::BOU_TRUE: |
1272 | return false; |
1273 | case cl::BOU_FALSE: |
1274 | return true; |
1275 | } |
1276 | llvm_unreachable("impossible case value" ); |
1277 | } |
1278 | |
1279 | /// Returns true if \p I is an instruction which requires predication and |
1280 | /// for which our chosen predication strategy is scalarization (i.e. we |
1281 | /// don't have an alternate strategy such as masking available). |
1282 | /// \p VF is the vectorization factor that will be used to vectorize \p I. |
1283 | bool isScalarWithPredication(Instruction *I, ElementCount VF) const; |
1284 | |
1285 | /// Returns true if \p I is an instruction that needs to be predicated |
1286 | /// at runtime. The result is independent of the predication mechanism. |
1287 | /// Superset of instructions that return true for isScalarWithPredication. |
1288 | bool isPredicatedInst(Instruction *I) const; |
1289 | |
1290 | /// Return the costs for our two available strategies for lowering a |
1291 | /// div/rem operation which requires speculating at least one lane. |
1292 | /// First result is for scalarization (will be invalid for scalable |
1293 | /// vectors); second is for the safe-divisor strategy. |
1294 | std::pair<InstructionCost, InstructionCost> |
1295 | getDivRemSpeculationCost(Instruction *I, |
1296 | ElementCount VF) const; |
1297 | |
1298 | /// Returns true if \p I is a memory instruction with consecutive memory |
1299 | /// access that can be widened. |
1300 | bool memoryInstructionCanBeWidened(Instruction *I, ElementCount VF); |
1301 | |
1302 | /// Returns true if \p I is a memory instruction in an interleaved-group |
1303 | /// of memory accesses that can be vectorized with wide vector loads/stores |
1304 | /// and shuffles. |
1305 | bool interleavedAccessCanBeWidened(Instruction *I, ElementCount VF) const; |
1306 | |
1307 | /// Check if \p Instr belongs to any interleaved access group. |
1308 | bool isAccessInterleaved(Instruction *Instr) const { |
1309 | return InterleaveInfo.isInterleaved(Instr); |
1310 | } |
1311 | |
1312 | /// Get the interleaved access group that \p Instr belongs to. |
1313 | const InterleaveGroup<Instruction> * |
1314 | getInterleavedAccessGroup(Instruction *Instr) const { |
1315 | return InterleaveInfo.getInterleaveGroup(Instr); |
1316 | } |
1317 | |
1318 | /// Returns true if we're required to use a scalar epilogue for at least |
1319 | /// the final iteration of the original loop. |
1320 | bool requiresScalarEpilogue(bool IsVectorizing) const { |
1321 | if (!isScalarEpilogueAllowed()) { |
1322 | LLVM_DEBUG(dbgs() << "LV: Loop does not require scalar epilogue\n" ); |
1323 | return false; |
1324 | } |
1325 | // If we might exit from anywhere but the latch and early exit vectorization |
1326 | // is disabled, we must run the exiting iteration in scalar form. |
1327 | if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch() && |
1328 | !(EnableEarlyExitVectorization && Legal->hasUncountableEarlyExit())) { |
1329 | LLVM_DEBUG(dbgs() << "LV: Loop requires scalar epilogue: not exiting " |
1330 | "from latch block\n" ); |
1331 | return true; |
1332 | } |
1333 | if (IsVectorizing && InterleaveInfo.requiresScalarEpilogue()) { |
1334 | LLVM_DEBUG(dbgs() << "LV: Loop requires scalar epilogue: " |
1335 | "interleaved group requires scalar epilogue\n" ); |
1336 | return true; |
1337 | } |
1338 | LLVM_DEBUG(dbgs() << "LV: Loop does not require scalar epilogue\n" ); |
1339 | return false; |
1340 | } |
1341 | |
1342 | /// Returns true if a scalar epilogue is not allowed due to optsize or a |
1343 | /// loop hint annotation. |
1344 | bool isScalarEpilogueAllowed() const { |
1345 | return ScalarEpilogueStatus == CM_ScalarEpilogueAllowed; |
1346 | } |
1347 | |
1348 | /// Returns the TailFoldingStyle that is best for the current loop. |
1349 | TailFoldingStyle getTailFoldingStyle(bool IVUpdateMayOverflow = true) const { |
1350 | if (!ChosenTailFoldingStyle) |
1351 | return TailFoldingStyle::None; |
1352 | return IVUpdateMayOverflow ? ChosenTailFoldingStyle->first |
1353 | : ChosenTailFoldingStyle->second; |
1354 | } |
1355 | |
1356 | /// Selects and saves TailFoldingStyle for 2 options - if IV update may |
1357 | /// overflow or not. |
1358 | /// \param IsScalableVF true if scalable vector factors enabled. |
1359 | /// \param UserIC User specific interleave count. |
1360 | void setTailFoldingStyles(bool IsScalableVF, unsigned UserIC) { |
1361 | assert(!ChosenTailFoldingStyle && "Tail folding must not be selected yet." ); |
1362 | if (!Legal->canFoldTailByMasking()) { |
1363 | ChosenTailFoldingStyle = {TailFoldingStyle::None, TailFoldingStyle::None}; |
1364 | return; |
1365 | } |
1366 | |
1367 | // Default to TTI preference, but allow command line override. |
1368 | ChosenTailFoldingStyle = { |
1369 | TTI.getPreferredTailFoldingStyle(/*IVUpdateMayOverflow=*/true), |
1370 | TTI.getPreferredTailFoldingStyle(/*IVUpdateMayOverflow=*/false)}; |
1371 | if (ForceTailFoldingStyle.getNumOccurrences()) |
1372 | ChosenTailFoldingStyle = {ForceTailFoldingStyle.getValue(), |
1373 | ForceTailFoldingStyle.getValue()}; |
1374 | |
1375 | if (ForceTailFoldingStyle != TailFoldingStyle::DataWithEVL) |
1376 | return; |
1377 | // Override forced styles if needed. |
1378 | // FIXME: Investigate opportunity for fixed vector factor. |
1379 | bool EVLIsLegal = UserIC <= 1 && IsScalableVF && |
1380 | TTI.hasActiveVectorLength() && !EnableVPlanNativePath; |
1381 | if (EVLIsLegal) |
1382 | return; |
1383 | // If for some reason EVL mode is unsupported, fallback to |
1384 | // DataWithoutLaneMask to try to vectorize the loop with folded tail |
1385 | // in a generic way. |
1386 | ChosenTailFoldingStyle = {TailFoldingStyle::DataWithoutLaneMask, |
1387 | TailFoldingStyle::DataWithoutLaneMask}; |
1388 | LLVM_DEBUG( |
1389 | dbgs() << "LV: Preference for VP intrinsics indicated. Will " |
1390 | "not try to generate VP Intrinsics " |
1391 | << (UserIC > 1 |
1392 | ? "since interleave count specified is greater than 1.\n" |
1393 | : "due to non-interleaving reasons.\n" )); |
1394 | } |
1395 | |
1396 | /// Returns true if all loop blocks should be masked to fold tail loop. |
1397 | bool foldTailByMasking() const { |
1398 | // TODO: check if it is possible to check for None style independent of |
1399 | // IVUpdateMayOverflow flag in getTailFoldingStyle. |
1400 | return getTailFoldingStyle() != TailFoldingStyle::None; |
1401 | } |
1402 | |
1403 | /// Return maximum safe number of elements to be processed per vector |
1404 | /// iteration, which do not prevent store-load forwarding and are safe with |
1405 | /// regard to the memory dependencies. Required for EVL-based VPlans to |
1406 | /// correctly calculate AVL (application vector length) as min(remaining AVL, |
1407 | /// MaxSafeElements). |
1408 | /// TODO: need to consider adjusting cost model to use this value as a |
1409 | /// vectorization factor for EVL-based vectorization. |
1410 | std::optional<unsigned> getMaxSafeElements() const { return MaxSafeElements; } |
1411 | |
1412 | /// Returns true if the instructions in this block requires predication |
1413 | /// for any reason, e.g. because tail folding now requires a predicate |
1414 | /// or because the block in the original loop was predicated. |
1415 | bool blockNeedsPredicationForAnyReason(BasicBlock *BB) const { |
1416 | return foldTailByMasking() || Legal->blockNeedsPredication(BB); |
1417 | } |
1418 | |
1419 | /// Returns true if VP intrinsics with explicit vector length support should |
1420 | /// be generated in the tail folded loop. |
1421 | bool foldTailWithEVL() const { |
1422 | return getTailFoldingStyle() == TailFoldingStyle::DataWithEVL; |
1423 | } |
1424 | |
1425 | /// Returns true if the Phi is part of an inloop reduction. |
1426 | bool isInLoopReduction(PHINode *Phi) const { |
1427 | return InLoopReductions.contains(Ptr: Phi); |
1428 | } |
1429 | |
1430 | /// Returns true if the predicated reduction select should be used to set the |
1431 | /// incoming value for the reduction phi. |
1432 | bool usePredicatedReductionSelect() const { |
1433 | // Force to use predicated reduction select since the EVL of the |
1434 | // second-to-last iteration might not be VF*UF. |
1435 | if (foldTailWithEVL()) |
1436 | return true; |
1437 | return PreferPredicatedReductionSelect || |
1438 | TTI.preferPredicatedReductionSelect(); |
1439 | } |
1440 | |
1441 | /// Estimate cost of an intrinsic call instruction CI if it were vectorized |
1442 | /// with factor VF. Return the cost of the instruction, including |
1443 | /// scalarization overhead if it's needed. |
1444 | InstructionCost getVectorIntrinsicCost(CallInst *CI, ElementCount VF) const; |
1445 | |
1446 | /// Estimate cost of a call instruction CI if it were vectorized with factor |
1447 | /// VF. Return the cost of the instruction, including scalarization overhead |
1448 | /// if it's needed. |
1449 | InstructionCost getVectorCallCost(CallInst *CI, ElementCount VF) const; |
1450 | |
1451 | /// Invalidates decisions already taken by the cost model. |
1452 | void invalidateCostModelingDecisions() { |
1453 | WideningDecisions.clear(); |
1454 | CallWideningDecisions.clear(); |
1455 | Uniforms.clear(); |
1456 | Scalars.clear(); |
1457 | } |
1458 | |
1459 | /// Returns the expected execution cost. The unit of the cost does |
1460 | /// not matter because we use the 'cost' units to compare different |
1461 | /// vector widths. The cost that is returned is *not* normalized by |
1462 | /// the factor width. |
1463 | InstructionCost expectedCost(ElementCount VF); |
1464 | |
1465 | bool hasPredStores() const { return NumPredStores > 0; } |
1466 | |
1467 | /// Returns true if epilogue vectorization is considered profitable, and |
1468 | /// false otherwise. |
1469 | /// \p VF is the vectorization factor chosen for the original loop. |
1470 | /// \p Multiplier is an aditional scaling factor applied to VF before |
1471 | /// comparing to EpilogueVectorizationMinVF. |
1472 | bool isEpilogueVectorizationProfitable(const ElementCount VF, |
1473 | const unsigned IC) const; |
1474 | |
1475 | /// Returns the execution time cost of an instruction for a given vector |
1476 | /// width. Vector width of one means scalar. |
1477 | InstructionCost getInstructionCost(Instruction *I, ElementCount VF); |
1478 | |
1479 | /// Return the cost of instructions in an inloop reduction pattern, if I is |
1480 | /// part of that pattern. |
1481 | std::optional<InstructionCost> getReductionPatternCost(Instruction *I, |
1482 | ElementCount VF, |
1483 | Type *VectorTy) const; |
1484 | |
1485 | /// Returns true if \p Op should be considered invariant and if it is |
1486 | /// trivially hoistable. |
1487 | bool shouldConsiderInvariant(Value *Op); |
1488 | |
1489 | /// Return the value of vscale used for tuning the cost model. |
1490 | std::optional<unsigned> getVScaleForTuning() const { return VScaleForTuning; } |
1491 | |
1492 | private: |
1493 | unsigned NumPredStores = 0; |
1494 | |
1495 | /// Used to store the value of vscale used for tuning the cost model. It is |
1496 | /// initialized during object construction. |
1497 | std::optional<unsigned> VScaleForTuning; |
1498 | |
1499 | /// Initializes the value of vscale used for tuning the cost model. If |
1500 | /// vscale_range.min == vscale_range.max then return vscale_range.max, else |
1501 | /// return the value returned by the corresponding TTI method. |
1502 | void initializeVScaleForTuning() { |
1503 | const Function *Fn = TheLoop->getHeader()->getParent(); |
1504 | if (Fn->hasFnAttribute(Kind: Attribute::VScaleRange)) { |
1505 | auto Attr = Fn->getFnAttribute(Kind: Attribute::VScaleRange); |
1506 | auto Min = Attr.getVScaleRangeMin(); |
1507 | auto Max = Attr.getVScaleRangeMax(); |
1508 | if (Max && Min == Max) { |
1509 | VScaleForTuning = Max; |
1510 | return; |
1511 | } |
1512 | } |
1513 | |
1514 | VScaleForTuning = TTI.getVScaleForTuning(); |
1515 | } |
1516 | |
1517 | /// \return An upper bound for the vectorization factors for both |
1518 | /// fixed and scalable vectorization, where the minimum-known number of |
1519 | /// elements is a power-of-2 larger than zero. If scalable vectorization is |
1520 | /// disabled or unsupported, then the scalable part will be equal to |
1521 | /// ElementCount::getScalable(0). |
1522 | FixedScalableVFPair computeFeasibleMaxVF(unsigned MaxTripCount, |
1523 | ElementCount UserVF, |
1524 | bool FoldTailByMasking); |
1525 | |
1526 | /// \return the maximized element count based on the targets vector |
1527 | /// registers and the loop trip-count, but limited to a maximum safe VF. |
1528 | /// This is a helper function of computeFeasibleMaxVF. |
1529 | ElementCount getMaximizedVFForTarget(unsigned MaxTripCount, |
1530 | unsigned SmallestType, |
1531 | unsigned WidestType, |
1532 | ElementCount MaxSafeVF, |
1533 | bool FoldTailByMasking); |
1534 | |
1535 | /// Checks if scalable vectorization is supported and enabled. Caches the |
1536 | /// result to avoid repeated debug dumps for repeated queries. |
1537 | bool isScalableVectorizationAllowed(); |
1538 | |
1539 | /// \return the maximum legal scalable VF, based on the safe max number |
1540 | /// of elements. |
1541 | ElementCount getMaxLegalScalableVF(unsigned MaxSafeElements); |
1542 | |
1543 | /// Calculate vectorization cost of memory instruction \p I. |
1544 | InstructionCost getMemoryInstructionCost(Instruction *I, ElementCount VF); |
1545 | |
1546 | /// The cost computation for scalarized memory instruction. |
1547 | InstructionCost getMemInstScalarizationCost(Instruction *I, ElementCount VF); |
1548 | |
1549 | /// The cost computation for interleaving group of memory instructions. |
1550 | InstructionCost getInterleaveGroupCost(Instruction *I, ElementCount VF); |
1551 | |
1552 | /// The cost computation for Gather/Scatter instruction. |
1553 | InstructionCost getGatherScatterCost(Instruction *I, ElementCount VF); |
1554 | |
1555 | /// The cost computation for widening instruction \p I with consecutive |
1556 | /// memory access. |
1557 | InstructionCost getConsecutiveMemOpCost(Instruction *I, ElementCount VF); |
1558 | |
1559 | /// The cost calculation for Load/Store instruction \p I with uniform pointer - |
1560 | /// Load: scalar load + broadcast. |
1561 | /// Store: scalar store + (loop invariant value stored? 0 : extract of last |
1562 | /// element) |
1563 | InstructionCost getUniformMemOpCost(Instruction *I, ElementCount VF); |
1564 | |
1565 | /// Estimate the overhead of scalarizing an instruction. This is a |
1566 | /// convenience wrapper for the type-based getScalarizationOverhead API. |
1567 | InstructionCost getScalarizationOverhead(Instruction *I, |
1568 | ElementCount VF) const; |
1569 | |
1570 | /// Returns true if an artificially high cost for emulated masked memrefs |
1571 | /// should be used. |
1572 | bool useEmulatedMaskMemRefHack(Instruction *I, ElementCount VF); |
1573 | |
1574 | /// Map of scalar integer values to the smallest bitwidth they can be legally |
1575 | /// represented as. The vector equivalents of these values should be truncated |
1576 | /// to this type. |
1577 | MapVector<Instruction *, uint64_t> MinBWs; |
1578 | |
1579 | /// A type representing the costs for instructions if they were to be |
1580 | /// scalarized rather than vectorized. The entries are Instruction-Cost |
1581 | /// pairs. |
1582 | using ScalarCostsTy = DenseMap<Instruction *, InstructionCost>; |
1583 | |
1584 | /// A set containing all BasicBlocks that are known to present after |
1585 | /// vectorization as a predicated block. |
1586 | DenseMap<ElementCount, SmallPtrSet<BasicBlock *, 4>> |
1587 | PredicatedBBsAfterVectorization; |
1588 | |
1589 | /// Records whether it is allowed to have the original scalar loop execute at |
1590 | /// least once. This may be needed as a fallback loop in case runtime |
1591 | /// aliasing/dependence checks fail, or to handle the tail/remainder |
1592 | /// iterations when the trip count is unknown or doesn't divide by the VF, |
1593 | /// or as a peel-loop to handle gaps in interleave-groups. |
1594 | /// Under optsize and when the trip count is very small we don't allow any |
1595 | /// iterations to execute in the scalar loop. |
1596 | ScalarEpilogueLowering ScalarEpilogueStatus = CM_ScalarEpilogueAllowed; |
1597 | |
1598 | /// Control finally chosen tail folding style. The first element is used if |
1599 | /// the IV update may overflow, the second element - if it does not. |
1600 | std::optional<std::pair<TailFoldingStyle, TailFoldingStyle>> |
1601 | ChosenTailFoldingStyle; |
1602 | |
1603 | /// true if scalable vectorization is supported and enabled. |
1604 | std::optional<bool> IsScalableVectorizationAllowed; |
1605 | |
1606 | /// Maximum safe number of elements to be processed per vector iteration, |
1607 | /// which do not prevent store-load forwarding and are safe with regard to the |
1608 | /// memory dependencies. Required for EVL-based veectorization, where this |
1609 | /// value is used as the upper bound of the safe AVL. |
1610 | std::optional<unsigned> MaxSafeElements; |
1611 | |
1612 | /// A map holding scalar costs for different vectorization factors. The |
1613 | /// presence of a cost for an instruction in the mapping indicates that the |
1614 | /// instruction will be scalarized when vectorizing with the associated |
1615 | /// vectorization factor. The entries are VF-ScalarCostTy pairs. |
1616 | DenseMap<ElementCount, ScalarCostsTy> InstsToScalarize; |
1617 | |
1618 | /// Holds the instructions known to be uniform after vectorization. |
1619 | /// The data is collected per VF. |
1620 | DenseMap<ElementCount, SmallPtrSet<Instruction *, 4>> Uniforms; |
1621 | |
1622 | /// Holds the instructions known to be scalar after vectorization. |
1623 | /// The data is collected per VF. |
1624 | DenseMap<ElementCount, SmallPtrSet<Instruction *, 4>> Scalars; |
1625 | |
1626 | /// Holds the instructions (address computations) that are forced to be |
1627 | /// scalarized. |
1628 | DenseMap<ElementCount, SmallPtrSet<Instruction *, 4>> ForcedScalars; |
1629 | |
1630 | /// PHINodes of the reductions that should be expanded in-loop. |
1631 | SmallPtrSet<PHINode *, 4> InLoopReductions; |
1632 | |
1633 | /// A Map of inloop reduction operations and their immediate chain operand. |
1634 | /// FIXME: This can be removed once reductions can be costed correctly in |
1635 | /// VPlan. This was added to allow quick lookup of the inloop operations. |
1636 | DenseMap<Instruction *, Instruction *> InLoopReductionImmediateChains; |
1637 | |
1638 | /// Returns the expected difference in cost from scalarizing the expression |
1639 | /// feeding a predicated instruction \p PredInst. The instructions to |
1640 | /// scalarize and their scalar costs are collected in \p ScalarCosts. A |
1641 | /// non-negative return value implies the expression will be scalarized. |
1642 | /// Currently, only single-use chains are considered for scalarization. |
1643 | InstructionCost computePredInstDiscount(Instruction *PredInst, |
1644 | ScalarCostsTy &ScalarCosts, |
1645 | ElementCount VF); |
1646 | |
1647 | /// Collect the instructions that are uniform after vectorization. An |
1648 | /// instruction is uniform if we represent it with a single scalar value in |
1649 | /// the vectorized loop corresponding to each vector iteration. Examples of |
1650 | /// uniform instructions include pointer operands of consecutive or |
1651 | /// interleaved memory accesses. Note that although uniformity implies an |
1652 | /// instruction will be scalar, the reverse is not true. In general, a |
1653 | /// scalarized instruction will be represented by VF scalar values in the |
1654 | /// vectorized loop, each corresponding to an iteration of the original |
1655 | /// scalar loop. |
1656 | void collectLoopUniforms(ElementCount VF); |
1657 | |
1658 | /// Collect the instructions that are scalar after vectorization. An |
1659 | /// instruction is scalar if it is known to be uniform or will be scalarized |
1660 | /// during vectorization. collectLoopScalars should only add non-uniform nodes |
1661 | /// to the list if they are used by a load/store instruction that is marked as |
1662 | /// CM_Scalarize. Non-uniform scalarized instructions will be represented by |
1663 | /// VF values in the vectorized loop, each corresponding to an iteration of |
1664 | /// the original scalar loop. |
1665 | void collectLoopScalars(ElementCount VF); |
1666 | |
1667 | /// Keeps cost model vectorization decision and cost for instructions. |
1668 | /// Right now it is used for memory instructions only. |
1669 | using DecisionList = DenseMap<std::pair<Instruction *, ElementCount>, |
1670 | std::pair<InstWidening, InstructionCost>>; |
1671 | |
1672 | DecisionList WideningDecisions; |
1673 | |
1674 | using CallDecisionList = |
1675 | DenseMap<std::pair<CallInst *, ElementCount>, CallWideningDecision>; |
1676 | |
1677 | CallDecisionList CallWideningDecisions; |
1678 | |
1679 | /// Returns true if \p V is expected to be vectorized and it needs to be |
1680 | /// extracted. |
1681 | bool (Value *V, ElementCount VF) const { |
1682 | Instruction *I = dyn_cast<Instruction>(Val: V); |
1683 | if (VF.isScalar() || !I || !TheLoop->contains(Inst: I) || |
1684 | TheLoop->isLoopInvariant(V: I) || |
1685 | getWideningDecision(I, VF) == CM_Scalarize) |
1686 | return false; |
1687 | |
1688 | // Assume we can vectorize V (and hence we need extraction) if the |
1689 | // scalars are not computed yet. This can happen, because it is called |
1690 | // via getScalarizationOverhead from setCostBasedWideningDecision, before |
1691 | // the scalars are collected. That should be a safe assumption in most |
1692 | // cases, because we check if the operands have vectorizable types |
1693 | // beforehand in LoopVectorizationLegality. |
1694 | return !Scalars.contains(Val: VF) || !isScalarAfterVectorization(I, VF); |
1695 | }; |
1696 | |
1697 | /// Returns a range containing only operands needing to be extracted. |
1698 | SmallVector<Value *, 4> filterExtractingOperands(Instruction::op_range Ops, |
1699 | ElementCount VF) const { |
1700 | return SmallVector<Value *, 4>(make_filter_range( |
1701 | Range&: Ops, Pred: [this, VF](Value *V) { return this->needsExtract(V, VF); })); |
1702 | } |
1703 | |
1704 | public: |
1705 | /// The loop that we evaluate. |
1706 | Loop *TheLoop; |
1707 | |
1708 | /// Predicated scalar evolution analysis. |
1709 | PredicatedScalarEvolution &PSE; |
1710 | |
1711 | /// Loop Info analysis. |
1712 | LoopInfo *LI; |
1713 | |
1714 | /// Vectorization legality. |
1715 | LoopVectorizationLegality *Legal; |
1716 | |
1717 | /// Vector target information. |
1718 | const TargetTransformInfo &TTI; |
1719 | |
1720 | /// Target Library Info. |
1721 | const TargetLibraryInfo *TLI; |
1722 | |
1723 | /// Demanded bits analysis. |
1724 | DemandedBits *DB; |
1725 | |
1726 | /// Assumption cache. |
1727 | AssumptionCache *AC; |
1728 | |
1729 | /// Interface to emit optimization remarks. |
1730 | OptimizationRemarkEmitter *ORE; |
1731 | |
1732 | const Function *TheFunction; |
1733 | |
1734 | /// Loop Vectorize Hint. |
1735 | const LoopVectorizeHints *Hints; |
1736 | |
1737 | /// The interleave access information contains groups of interleaved accesses |
1738 | /// with the same stride and close to each other. |
1739 | InterleavedAccessInfo &InterleaveInfo; |
1740 | |
1741 | /// Values to ignore in the cost model. |
1742 | SmallPtrSet<const Value *, 16> ValuesToIgnore; |
1743 | |
1744 | /// Values to ignore in the cost model when VF > 1. |
1745 | SmallPtrSet<const Value *, 16> VecValuesToIgnore; |
1746 | |
1747 | /// All element types found in the loop. |
1748 | SmallPtrSet<Type *, 16> ElementTypesInLoop; |
1749 | |
1750 | /// The kind of cost that we are calculating |
1751 | TTI::TargetCostKind CostKind; |
1752 | |
1753 | /// Whether this loop should be optimized for size based on function attribute |
1754 | /// or profile information. |
1755 | bool OptForSize; |
1756 | }; |
1757 | } // end namespace llvm |
1758 | |
1759 | namespace { |
1760 | /// Helper struct to manage generating runtime checks for vectorization. |
1761 | /// |
1762 | /// The runtime checks are created up-front in temporary blocks to allow better |
1763 | /// estimating the cost and un-linked from the existing IR. After deciding to |
1764 | /// vectorize, the checks are moved back. If deciding not to vectorize, the |
1765 | /// temporary blocks are completely removed. |
1766 | class GeneratedRTChecks { |
1767 | /// Basic block which contains the generated SCEV checks, if any. |
1768 | BasicBlock *SCEVCheckBlock = nullptr; |
1769 | |
1770 | /// The value representing the result of the generated SCEV checks. If it is |
1771 | /// nullptr no SCEV checks have been generated. |
1772 | Value *SCEVCheckCond = nullptr; |
1773 | |
1774 | /// Basic block which contains the generated memory runtime checks, if any. |
1775 | BasicBlock *MemCheckBlock = nullptr; |
1776 | |
1777 | /// The value representing the result of the generated memory runtime checks. |
1778 | /// If it is nullptr no memory runtime checks have been generated. |
1779 | Value *MemRuntimeCheckCond = nullptr; |
1780 | |
1781 | /// True if any checks have been added. |
1782 | bool AddedAnyChecks = false; |
1783 | |
1784 | DominatorTree *DT; |
1785 | LoopInfo *LI; |
1786 | TargetTransformInfo *TTI; |
1787 | |
1788 | SCEVExpander SCEVExp; |
1789 | SCEVExpander MemCheckExp; |
1790 | |
1791 | bool CostTooHigh = false; |
1792 | const bool AddBranchWeights; |
1793 | |
1794 | Loop *OuterLoop = nullptr; |
1795 | |
1796 | PredicatedScalarEvolution &PSE; |
1797 | |
1798 | /// The kind of cost that we are calculating |
1799 | TTI::TargetCostKind CostKind; |
1800 | |
1801 | public: |
1802 | GeneratedRTChecks(PredicatedScalarEvolution &PSE, DominatorTree *DT, |
1803 | LoopInfo *LI, TargetTransformInfo *TTI, |
1804 | const DataLayout &DL, bool AddBranchWeights, |
1805 | TTI::TargetCostKind CostKind) |
1806 | : DT(DT), LI(LI), TTI(TTI), SCEVExp(*PSE.getSE(), DL, "scev.check" ), |
1807 | MemCheckExp(*PSE.getSE(), DL, "scev.check" ), |
1808 | AddBranchWeights(AddBranchWeights), PSE(PSE), CostKind(CostKind) {} |
1809 | |
1810 | /// Generate runtime checks in SCEVCheckBlock and MemCheckBlock, so we can |
1811 | /// accurately estimate the cost of the runtime checks. The blocks are |
1812 | /// un-linked from the IR and are added back during vector code generation. If |
1813 | /// there is no vector code generation, the check blocks are removed |
1814 | /// completely. |
1815 | void create(Loop *L, const LoopAccessInfo &LAI, |
1816 | const SCEVPredicate &UnionPred, ElementCount VF, unsigned IC) { |
1817 | |
1818 | // Hard cutoff to limit compile-time increase in case a very large number of |
1819 | // runtime checks needs to be generated. |
1820 | // TODO: Skip cutoff if the loop is guaranteed to execute, e.g. due to |
1821 | // profile info. |
1822 | CostTooHigh = |
1823 | LAI.getNumRuntimePointerChecks() > VectorizeMemoryCheckThreshold; |
1824 | if (CostTooHigh) |
1825 | return; |
1826 | |
1827 | BasicBlock * = L->getHeader(); |
1828 | BasicBlock * = L->getLoopPreheader(); |
1829 | |
1830 | // Use SplitBlock to create blocks for SCEV & memory runtime checks to |
1831 | // ensure the blocks are properly added to LoopInfo & DominatorTree. Those |
1832 | // may be used by SCEVExpander. The blocks will be un-linked from their |
1833 | // predecessors and removed from LI & DT at the end of the function. |
1834 | if (!UnionPred.isAlwaysTrue()) { |
1835 | SCEVCheckBlock = SplitBlock(Old: Preheader, SplitPt: Preheader->getTerminator(), DT, LI, |
1836 | MSSAU: nullptr, BBName: "vector.scevcheck" ); |
1837 | |
1838 | SCEVCheckCond = SCEVExp.expandCodeForPredicate( |
1839 | Pred: &UnionPred, Loc: SCEVCheckBlock->getTerminator()); |
1840 | if (isa<Constant>(Val: SCEVCheckCond)) { |
1841 | // Clean up directly after expanding the predicate to a constant, to |
1842 | // avoid further expansions re-using anything left over from SCEVExp. |
1843 | SCEVExpanderCleaner SCEVCleaner(SCEVExp); |
1844 | SCEVCleaner.cleanup(); |
1845 | } |
1846 | } |
1847 | |
1848 | const auto &RtPtrChecking = *LAI.getRuntimePointerChecking(); |
1849 | if (RtPtrChecking.Need) { |
1850 | auto *Pred = SCEVCheckBlock ? SCEVCheckBlock : Preheader; |
1851 | MemCheckBlock = SplitBlock(Old: Pred, SplitPt: Pred->getTerminator(), DT, LI, MSSAU: nullptr, |
1852 | BBName: "vector.memcheck" ); |
1853 | |
1854 | auto DiffChecks = RtPtrChecking.getDiffChecks(); |
1855 | if (DiffChecks) { |
1856 | Value *RuntimeVF = nullptr; |
1857 | MemRuntimeCheckCond = addDiffRuntimeChecks( |
1858 | Loc: MemCheckBlock->getTerminator(), Checks: *DiffChecks, Expander&: MemCheckExp, |
1859 | GetVF: [VF, &RuntimeVF](IRBuilderBase &B, unsigned Bits) { |
1860 | if (!RuntimeVF) |
1861 | RuntimeVF = getRuntimeVF(B, Ty: B.getIntNTy(N: Bits), VF); |
1862 | return RuntimeVF; |
1863 | }, |
1864 | IC); |
1865 | } else { |
1866 | MemRuntimeCheckCond = addRuntimeChecks( |
1867 | Loc: MemCheckBlock->getTerminator(), TheLoop: L, PointerChecks: RtPtrChecking.getChecks(), |
1868 | Expander&: MemCheckExp, HoistRuntimeChecks: VectorizerParams::HoistRuntimeChecks); |
1869 | } |
1870 | assert(MemRuntimeCheckCond && |
1871 | "no RT checks generated although RtPtrChecking " |
1872 | "claimed checks are required" ); |
1873 | } |
1874 | |
1875 | if (!MemCheckBlock && !SCEVCheckBlock) |
1876 | return; |
1877 | |
1878 | // Unhook the temporary block with the checks, update various places |
1879 | // accordingly. |
1880 | if (SCEVCheckBlock) |
1881 | SCEVCheckBlock->replaceAllUsesWith(V: Preheader); |
1882 | if (MemCheckBlock) |
1883 | MemCheckBlock->replaceAllUsesWith(V: Preheader); |
1884 | |
1885 | if (SCEVCheckBlock) { |
1886 | SCEVCheckBlock->getTerminator()->moveBefore( |
1887 | InsertPos: Preheader->getTerminator()->getIterator()); |
1888 | auto *UI = new UnreachableInst(Preheader->getContext(), SCEVCheckBlock); |
1889 | UI->setDebugLoc(DebugLoc::getTemporary()); |
1890 | Preheader->getTerminator()->eraseFromParent(); |
1891 | } |
1892 | if (MemCheckBlock) { |
1893 | MemCheckBlock->getTerminator()->moveBefore( |
1894 | InsertPos: Preheader->getTerminator()->getIterator()); |
1895 | auto *UI = new UnreachableInst(Preheader->getContext(), MemCheckBlock); |
1896 | UI->setDebugLoc(DebugLoc::getTemporary()); |
1897 | Preheader->getTerminator()->eraseFromParent(); |
1898 | } |
1899 | |
1900 | DT->changeImmediateDominator(BB: LoopHeader, NewBB: Preheader); |
1901 | if (MemCheckBlock) { |
1902 | DT->eraseNode(BB: MemCheckBlock); |
1903 | LI->removeBlock(BB: MemCheckBlock); |
1904 | } |
1905 | if (SCEVCheckBlock) { |
1906 | DT->eraseNode(BB: SCEVCheckBlock); |
1907 | LI->removeBlock(BB: SCEVCheckBlock); |
1908 | } |
1909 | |
1910 | // Outer loop is used as part of the later cost calculations. |
1911 | OuterLoop = L->getParentLoop(); |
1912 | } |
1913 | |
1914 | InstructionCost getCost() { |
1915 | if (SCEVCheckBlock || MemCheckBlock) |
1916 | LLVM_DEBUG(dbgs() << "Calculating cost of runtime checks:\n" ); |
1917 | |
1918 | if (CostTooHigh) { |
1919 | InstructionCost Cost; |
1920 | Cost.setInvalid(); |
1921 | LLVM_DEBUG(dbgs() << " number of checks exceeded threshold\n" ); |
1922 | return Cost; |
1923 | } |
1924 | |
1925 | InstructionCost RTCheckCost = 0; |
1926 | if (SCEVCheckBlock) |
1927 | for (Instruction &I : *SCEVCheckBlock) { |
1928 | if (SCEVCheckBlock->getTerminator() == &I) |
1929 | continue; |
1930 | InstructionCost C = TTI->getInstructionCost(U: &I, CostKind); |
1931 | LLVM_DEBUG(dbgs() << " " << C << " for " << I << "\n" ); |
1932 | RTCheckCost += C; |
1933 | } |
1934 | if (MemCheckBlock) { |
1935 | InstructionCost MemCheckCost = 0; |
1936 | for (Instruction &I : *MemCheckBlock) { |
1937 | if (MemCheckBlock->getTerminator() == &I) |
1938 | continue; |
1939 | InstructionCost C = TTI->getInstructionCost(U: &I, CostKind); |
1940 | LLVM_DEBUG(dbgs() << " " << C << " for " << I << "\n" ); |
1941 | MemCheckCost += C; |
1942 | } |
1943 | |
1944 | // If the runtime memory checks are being created inside an outer loop |
1945 | // we should find out if these checks are outer loop invariant. If so, |
1946 | // the checks will likely be hoisted out and so the effective cost will |
1947 | // reduce according to the outer loop trip count. |
1948 | if (OuterLoop) { |
1949 | ScalarEvolution *SE = MemCheckExp.getSE(); |
1950 | // TODO: If profitable, we could refine this further by analysing every |
1951 | // individual memory check, since there could be a mixture of loop |
1952 | // variant and invariant checks that mean the final condition is |
1953 | // variant. |
1954 | const SCEV *Cond = SE->getSCEV(V: MemRuntimeCheckCond); |
1955 | if (SE->isLoopInvariant(S: Cond, L: OuterLoop)) { |
1956 | // It seems reasonable to assume that we can reduce the effective |
1957 | // cost of the checks even when we know nothing about the trip |
1958 | // count. Assume that the outer loop executes at least twice. |
1959 | unsigned BestTripCount = 2; |
1960 | |
1961 | // Get the best known TC estimate. |
1962 | if (auto EstimatedTC = getSmallBestKnownTC( |
1963 | PSE, L: OuterLoop, /* CanUseConstantMax = */ false)) |
1964 | if (EstimatedTC->isFixed()) |
1965 | BestTripCount = EstimatedTC->getFixedValue(); |
1966 | |
1967 | InstructionCost NewMemCheckCost = MemCheckCost / BestTripCount; |
1968 | |
1969 | // Let's ensure the cost is always at least 1. |
1970 | NewMemCheckCost = std::max(a: NewMemCheckCost.getValue(), |
1971 | b: (InstructionCost::CostType)1); |
1972 | |
1973 | if (BestTripCount > 1) |
1974 | LLVM_DEBUG(dbgs() |
1975 | << "We expect runtime memory checks to be hoisted " |
1976 | << "out of the outer loop. Cost reduced from " |
1977 | << MemCheckCost << " to " << NewMemCheckCost << '\n'); |
1978 | |
1979 | MemCheckCost = NewMemCheckCost; |
1980 | } |
1981 | } |
1982 | |
1983 | RTCheckCost += MemCheckCost; |
1984 | } |
1985 | |
1986 | if (SCEVCheckBlock || MemCheckBlock) |
1987 | LLVM_DEBUG(dbgs() << "Total cost of runtime checks: " << RTCheckCost |
1988 | << "\n" ); |
1989 | |
1990 | return RTCheckCost; |
1991 | } |
1992 | |
1993 | /// Remove the created SCEV & memory runtime check blocks & instructions, if |
1994 | /// unused. |
1995 | ~GeneratedRTChecks() { |
1996 | SCEVExpanderCleaner SCEVCleaner(SCEVExp); |
1997 | SCEVExpanderCleaner MemCheckCleaner(MemCheckExp); |
1998 | bool SCEVChecksUsed = !SCEVCheckBlock || !pred_empty(BB: SCEVCheckBlock); |
1999 | bool MemChecksUsed = !MemCheckBlock || !pred_empty(BB: MemCheckBlock); |
2000 | if (SCEVChecksUsed) |
2001 | SCEVCleaner.markResultUsed(); |
2002 | |
2003 | if (MemChecksUsed) { |
2004 | MemCheckCleaner.markResultUsed(); |
2005 | } else { |
2006 | auto &SE = *MemCheckExp.getSE(); |
2007 | // Memory runtime check generation creates compares that use expanded |
2008 | // values. Remove them before running the SCEVExpanderCleaners. |
2009 | for (auto &I : make_early_inc_range(Range: reverse(C&: *MemCheckBlock))) { |
2010 | if (MemCheckExp.isInsertedInstruction(I: &I)) |
2011 | continue; |
2012 | SE.forgetValue(V: &I); |
2013 | I.eraseFromParent(); |
2014 | } |
2015 | } |
2016 | MemCheckCleaner.cleanup(); |
2017 | SCEVCleaner.cleanup(); |
2018 | |
2019 | if (!SCEVChecksUsed) |
2020 | SCEVCheckBlock->eraseFromParent(); |
2021 | if (!MemChecksUsed) |
2022 | MemCheckBlock->eraseFromParent(); |
2023 | } |
2024 | |
2025 | /// Adds the generated SCEVCheckBlock before \p LoopVectorPreHeader and |
2026 | /// adjusts the branches to branch to the vector preheader or \p Bypass, |
2027 | /// depending on the generated condition. |
2028 | BasicBlock *emitSCEVChecks(BasicBlock *Bypass, |
2029 | BasicBlock *) { |
2030 | using namespace llvm::PatternMatch; |
2031 | if (!SCEVCheckCond || match(V: SCEVCheckCond, P: m_ZeroInt())) |
2032 | return nullptr; |
2033 | |
2034 | auto *Pred = LoopVectorPreHeader->getSinglePredecessor(); |
2035 | BranchInst::Create(IfTrue: LoopVectorPreHeader, InsertBefore: SCEVCheckBlock); |
2036 | |
2037 | SCEVCheckBlock->getTerminator()->eraseFromParent(); |
2038 | SCEVCheckBlock->moveBefore(MovePos: LoopVectorPreHeader); |
2039 | Pred->getTerminator()->replaceSuccessorWith(OldBB: LoopVectorPreHeader, |
2040 | NewBB: SCEVCheckBlock); |
2041 | |
2042 | BranchInst &BI = |
2043 | *BranchInst::Create(IfTrue: Bypass, IfFalse: LoopVectorPreHeader, Cond: SCEVCheckCond); |
2044 | if (AddBranchWeights) |
2045 | setBranchWeights(I&: BI, Weights: SCEVCheckBypassWeights, /*IsExpected=*/false); |
2046 | ReplaceInstWithInst(From: SCEVCheckBlock->getTerminator(), To: &BI); |
2047 | AddedAnyChecks = true; |
2048 | return SCEVCheckBlock; |
2049 | } |
2050 | |
2051 | /// Adds the generated MemCheckBlock before \p LoopVectorPreHeader and adjusts |
2052 | /// the branches to branch to the vector preheader or \p Bypass, depending on |
2053 | /// the generated condition. |
2054 | BasicBlock *emitMemRuntimeChecks(BasicBlock *Bypass, |
2055 | BasicBlock *) { |
2056 | // Check if we generated code that checks in runtime if arrays overlap. |
2057 | if (!MemRuntimeCheckCond) |
2058 | return nullptr; |
2059 | |
2060 | auto *Pred = LoopVectorPreHeader->getSinglePredecessor(); |
2061 | Pred->getTerminator()->replaceSuccessorWith(OldBB: LoopVectorPreHeader, |
2062 | NewBB: MemCheckBlock); |
2063 | |
2064 | MemCheckBlock->moveBefore(MovePos: LoopVectorPreHeader); |
2065 | |
2066 | BranchInst &BI = |
2067 | *BranchInst::Create(IfTrue: Bypass, IfFalse: LoopVectorPreHeader, Cond: MemRuntimeCheckCond); |
2068 | if (AddBranchWeights) { |
2069 | setBranchWeights(I&: BI, Weights: MemCheckBypassWeights, /*IsExpected=*/false); |
2070 | } |
2071 | ReplaceInstWithInst(From: MemCheckBlock->getTerminator(), To: &BI); |
2072 | MemCheckBlock->getTerminator()->setDebugLoc( |
2073 | Pred->getTerminator()->getDebugLoc()); |
2074 | |
2075 | AddedAnyChecks = true; |
2076 | return MemCheckBlock; |
2077 | } |
2078 | |
2079 | /// Return true if any runtime checks have been added |
2080 | bool hasChecks() const { return AddedAnyChecks; } |
2081 | }; |
2082 | } // namespace |
2083 | |
2084 | static bool useActiveLaneMask(TailFoldingStyle Style) { |
2085 | return Style == TailFoldingStyle::Data || |
2086 | Style == TailFoldingStyle::DataAndControlFlow || |
2087 | Style == TailFoldingStyle::DataAndControlFlowWithoutRuntimeCheck; |
2088 | } |
2089 | |
2090 | static bool useActiveLaneMaskForControlFlow(TailFoldingStyle Style) { |
2091 | return Style == TailFoldingStyle::DataAndControlFlow || |
2092 | Style == TailFoldingStyle::DataAndControlFlowWithoutRuntimeCheck; |
2093 | } |
2094 | |
2095 | // Return true if \p OuterLp is an outer loop annotated with hints for explicit |
2096 | // vectorization. The loop needs to be annotated with #pragma omp simd |
2097 | // simdlen(#) or #pragma clang vectorize(enable) vectorize_width(#). If the |
2098 | // vector length information is not provided, vectorization is not considered |
2099 | // explicit. Interleave hints are not allowed either. These limitations will be |
2100 | // relaxed in the future. |
2101 | // Please, note that we are currently forced to abuse the pragma 'clang |
2102 | // vectorize' semantics. This pragma provides *auto-vectorization hints* |
2103 | // (i.e., LV must check that vectorization is legal) whereas pragma 'omp simd' |
2104 | // provides *explicit vectorization hints* (LV can bypass legal checks and |
2105 | // assume that vectorization is legal). However, both hints are implemented |
2106 | // using the same metadata (llvm.loop.vectorize, processed by |
2107 | // LoopVectorizeHints). This will be fixed in the future when the native IR |
2108 | // representation for pragma 'omp simd' is introduced. |
2109 | static bool (Loop *OuterLp, |
2110 | OptimizationRemarkEmitter *ORE) { |
2111 | assert(!OuterLp->isInnermost() && "This is not an outer loop" ); |
2112 | LoopVectorizeHints Hints(OuterLp, true /*DisableInterleaving*/, *ORE); |
2113 | |
2114 | // Only outer loops with an explicit vectorization hint are supported. |
2115 | // Unannotated outer loops are ignored. |
2116 | if (Hints.getForce() == LoopVectorizeHints::FK_Undefined) |
2117 | return false; |
2118 | |
2119 | Function *Fn = OuterLp->getHeader()->getParent(); |
2120 | if (!Hints.allowVectorization(F: Fn, L: OuterLp, |
2121 | VectorizeOnlyWhenForced: true /*VectorizeOnlyWhenForced*/)) { |
2122 | LLVM_DEBUG(dbgs() << "LV: Loop hints prevent outer loop vectorization.\n" ); |
2123 | return false; |
2124 | } |
2125 | |
2126 | if (Hints.getInterleave() > 1) { |
2127 | // TODO: Interleave support is future work. |
2128 | LLVM_DEBUG(dbgs() << "LV: Not vectorizing: Interleave is not supported for " |
2129 | "outer loops.\n" ); |
2130 | Hints.emitRemarkWithHints(); |
2131 | return false; |
2132 | } |
2133 | |
2134 | return true; |
2135 | } |
2136 | |
2137 | static void (Loop &L, LoopInfo *LI, |
2138 | OptimizationRemarkEmitter *ORE, |
2139 | SmallVectorImpl<Loop *> &V) { |
2140 | // Collect inner loops and outer loops without irreducible control flow. For |
2141 | // now, only collect outer loops that have explicit vectorization hints. If we |
2142 | // are stress testing the VPlan H-CFG construction, we collect the outermost |
2143 | // loop of every loop nest. |
2144 | if (L.isInnermost() || VPlanBuildStressTest || |
2145 | (EnableVPlanNativePath && isExplicitVecOuterLoop(OuterLp: &L, ORE))) { |
2146 | LoopBlocksRPO RPOT(&L); |
2147 | RPOT.perform(LI); |
2148 | if (!containsIrreducibleCFG<const BasicBlock *>(RPOTraversal&: RPOT, LI: *LI)) { |
2149 | V.push_back(Elt: &L); |
2150 | // TODO: Collect inner loops inside marked outer loops in case |
2151 | // vectorization fails for the outer loop. Do not invoke |
2152 | // 'containsIrreducibleCFG' again for inner loops when the outer loop is |
2153 | // already known to be reducible. We can use an inherited attribute for |
2154 | // that. |
2155 | return; |
2156 | } |
2157 | } |
2158 | for (Loop *InnerL : L) |
2159 | collectSupportedLoops(L&: *InnerL, LI, ORE, V); |
2160 | } |
2161 | |
2162 | //===----------------------------------------------------------------------===// |
2163 | // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and |
2164 | // LoopVectorizationCostModel and LoopVectorizationPlanner. |
2165 | //===----------------------------------------------------------------------===// |
2166 | |
2167 | /// Compute the transformed value of Index at offset StartValue using step |
2168 | /// StepValue. |
2169 | /// For integer induction, returns StartValue + Index * StepValue. |
2170 | /// For pointer induction, returns StartValue[Index * StepValue]. |
2171 | /// FIXME: The newly created binary instructions should contain nsw/nuw |
2172 | /// flags, which can be found from the original scalar operations. |
2173 | static Value * |
2174 | emitTransformedIndex(IRBuilderBase &B, Value *Index, Value *StartValue, |
2175 | Value *Step, |
2176 | InductionDescriptor::InductionKind InductionKind, |
2177 | const BinaryOperator *InductionBinOp) { |
2178 | using namespace llvm::PatternMatch; |
2179 | Type *StepTy = Step->getType(); |
2180 | Value *CastedIndex = StepTy->isIntegerTy() |
2181 | ? B.CreateSExtOrTrunc(V: Index, DestTy: StepTy) |
2182 | : B.CreateCast(Op: Instruction::SIToFP, V: Index, DestTy: StepTy); |
2183 | if (CastedIndex != Index) { |
2184 | CastedIndex->setName(CastedIndex->getName() + ".cast" ); |
2185 | Index = CastedIndex; |
2186 | } |
2187 | |
2188 | // Note: the IR at this point is broken. We cannot use SE to create any new |
2189 | // SCEV and then expand it, hoping that SCEV's simplification will give us |
2190 | // a more optimal code. Unfortunately, attempt of doing so on invalid IR may |
2191 | // lead to various SCEV crashes. So all we can do is to use builder and rely |
2192 | // on InstCombine for future simplifications. Here we handle some trivial |
2193 | // cases only. |
2194 | auto CreateAdd = [&B](Value *X, Value *Y) { |
2195 | assert(X->getType() == Y->getType() && "Types don't match!" ); |
2196 | if (match(V: X, P: m_ZeroInt())) |
2197 | return Y; |
2198 | if (match(V: Y, P: m_ZeroInt())) |
2199 | return X; |
2200 | return B.CreateAdd(LHS: X, RHS: Y); |
2201 | }; |
2202 | |
2203 | // We allow X to be a vector type, in which case Y will potentially be |
2204 | // splatted into a vector with the same element count. |
2205 | auto CreateMul = [&B](Value *X, Value *Y) { |
2206 | assert(X->getType()->getScalarType() == Y->getType() && |
2207 | "Types don't match!" ); |
2208 | if (match(V: X, P: m_One())) |
2209 | return Y; |
2210 | if (match(V: Y, P: m_One())) |
2211 | return X; |
2212 | VectorType *XVTy = dyn_cast<VectorType>(Val: X->getType()); |
2213 | if (XVTy && !isa<VectorType>(Val: Y->getType())) |
2214 | Y = B.CreateVectorSplat(EC: XVTy->getElementCount(), V: Y); |
2215 | return B.CreateMul(LHS: X, RHS: Y); |
2216 | }; |
2217 | |
2218 | switch (InductionKind) { |
2219 | case InductionDescriptor::IK_IntInduction: { |
2220 | assert(!isa<VectorType>(Index->getType()) && |
2221 | "Vector indices not supported for integer inductions yet" ); |
2222 | assert(Index->getType() == StartValue->getType() && |
2223 | "Index type does not match StartValue type" ); |
2224 | if (isa<ConstantInt>(Val: Step) && cast<ConstantInt>(Val: Step)->isMinusOne()) |
2225 | return B.CreateSub(LHS: StartValue, RHS: Index); |
2226 | auto *Offset = CreateMul(Index, Step); |
2227 | return CreateAdd(StartValue, Offset); |
2228 | } |
2229 | case InductionDescriptor::IK_PtrInduction: |
2230 | return B.CreatePtrAdd(Ptr: StartValue, Offset: CreateMul(Index, Step)); |
2231 | case InductionDescriptor::IK_FpInduction: { |
2232 | assert(!isa<VectorType>(Index->getType()) && |
2233 | "Vector indices not supported for FP inductions yet" ); |
2234 | assert(Step->getType()->isFloatingPointTy() && "Expected FP Step value" ); |
2235 | assert(InductionBinOp && |
2236 | (InductionBinOp->getOpcode() == Instruction::FAdd || |
2237 | InductionBinOp->getOpcode() == Instruction::FSub) && |
2238 | "Original bin op should be defined for FP induction" ); |
2239 | |
2240 | Value *MulExp = B.CreateFMul(L: Step, R: Index); |
2241 | return B.CreateBinOp(Opc: InductionBinOp->getOpcode(), LHS: StartValue, RHS: MulExp, |
2242 | Name: "induction" ); |
2243 | } |
2244 | case InductionDescriptor::IK_NoInduction: |
2245 | return nullptr; |
2246 | } |
2247 | llvm_unreachable("invalid enum" ); |
2248 | } |
2249 | |
2250 | static std::optional<unsigned> getMaxVScale(const Function &F, |
2251 | const TargetTransformInfo &TTI) { |
2252 | if (std::optional<unsigned> MaxVScale = TTI.getMaxVScale()) |
2253 | return MaxVScale; |
2254 | |
2255 | if (F.hasFnAttribute(Kind: Attribute::VScaleRange)) |
2256 | return F.getFnAttribute(Kind: Attribute::VScaleRange).getVScaleRangeMax(); |
2257 | |
2258 | return std::nullopt; |
2259 | } |
2260 | |
2261 | /// For the given VF and UF and maximum trip count computed for the loop, return |
2262 | /// whether the induction variable might overflow in the vectorized loop. If not, |
2263 | /// then we know a runtime overflow check always evaluates to false and can be |
2264 | /// removed. |
2265 | static bool isIndvarOverflowCheckKnownFalse( |
2266 | const LoopVectorizationCostModel *Cost, |
2267 | ElementCount VF, std::optional<unsigned> UF = std::nullopt) { |
2268 | // Always be conservative if we don't know the exact unroll factor. |
2269 | unsigned MaxUF = UF ? *UF : Cost->TTI.getMaxInterleaveFactor(VF); |
2270 | |
2271 | IntegerType *IdxTy = Cost->Legal->getWidestInductionType(); |
2272 | APInt MaxUIntTripCount = IdxTy->getMask(); |
2273 | |
2274 | // We know the runtime overflow check is known false iff the (max) trip-count |
2275 | // is known and (max) trip-count + (VF * UF) does not overflow in the type of |
2276 | // the vector loop induction variable. |
2277 | if (unsigned TC = Cost->PSE.getSmallConstantMaxTripCount()) { |
2278 | uint64_t MaxVF = VF.getKnownMinValue(); |
2279 | if (VF.isScalable()) { |
2280 | std::optional<unsigned> MaxVScale = |
2281 | getMaxVScale(F: *Cost->TheFunction, TTI: Cost->TTI); |
2282 | if (!MaxVScale) |
2283 | return false; |
2284 | MaxVF *= *MaxVScale; |
2285 | } |
2286 | |
2287 | return (MaxUIntTripCount - TC).ugt(RHS: MaxVF * MaxUF); |
2288 | } |
2289 | |
2290 | return false; |
2291 | } |
2292 | |
2293 | // Return whether we allow using masked interleave-groups (for dealing with |
2294 | // strided loads/stores that reside in predicated blocks, or for dealing |
2295 | // with gaps). |
2296 | static bool useMaskedInterleavedAccesses(const TargetTransformInfo &TTI) { |
2297 | // If an override option has been passed in for interleaved accesses, use it. |
2298 | if (EnableMaskedInterleavedMemAccesses.getNumOccurrences() > 0) |
2299 | return EnableMaskedInterleavedMemAccesses; |
2300 | |
2301 | return TTI.enableMaskedInterleavedAccessVectorization(); |
2302 | } |
2303 | |
2304 | Value * |
2305 | InnerLoopVectorizer::getOrCreateVectorTripCount(BasicBlock *InsertBlock) { |
2306 | if (VectorTripCount) |
2307 | return VectorTripCount; |
2308 | |
2309 | Value *TC = getTripCount(); |
2310 | IRBuilder<> Builder(InsertBlock->getTerminator()); |
2311 | |
2312 | Type *Ty = TC->getType(); |
2313 | // This is where we can make the step a runtime constant. |
2314 | Value *Step = createStepForVF(B&: Builder, Ty, VF, Step: UF); |
2315 | |
2316 | // If the tail is to be folded by masking, round the number of iterations N |
2317 | // up to a multiple of Step instead of rounding down. This is done by first |
2318 | // adding Step-1 and then rounding down. Note that it's ok if this addition |
2319 | // overflows: the vector induction variable will eventually wrap to zero given |
2320 | // that it starts at zero and its Step is a power of two; the loop will then |
2321 | // exit, with the last early-exit vector comparison also producing all-true. |
2322 | // For scalable vectors the VF is not guaranteed to be a power of 2, but this |
2323 | // is accounted for in emitIterationCountCheck that adds an overflow check. |
2324 | if (Cost->foldTailByMasking()) { |
2325 | assert(isPowerOf2_32(VF.getKnownMinValue() * UF) && |
2326 | "VF*UF must be a power of 2 when folding tail by masking" ); |
2327 | TC = Builder.CreateAdd(LHS: TC, RHS: Builder.CreateSub(LHS: Step, RHS: ConstantInt::get(Ty, V: 1)), |
2328 | Name: "n.rnd.up" ); |
2329 | } |
2330 | |
2331 | // Now we need to generate the expression for the part of the loop that the |
2332 | // vectorized body will execute. This is equal to N - (N % Step) if scalar |
2333 | // iterations are not required for correctness, or N - Step, otherwise. Step |
2334 | // is equal to the vectorization factor (number of SIMD elements) times the |
2335 | // unroll factor (number of SIMD instructions). |
2336 | Value *R = Builder.CreateURem(LHS: TC, RHS: Step, Name: "n.mod.vf" ); |
2337 | |
2338 | // There are cases where we *must* run at least one iteration in the remainder |
2339 | // loop. See the cost model for when this can happen. If the step evenly |
2340 | // divides the trip count, we set the remainder to be equal to the step. If |
2341 | // the step does not evenly divide the trip count, no adjustment is necessary |
2342 | // since there will already be scalar iterations. Note that the minimum |
2343 | // iterations check ensures that N >= Step. |
2344 | if (Cost->requiresScalarEpilogue(IsVectorizing: VF.isVector())) { |
2345 | auto *IsZero = Builder.CreateICmpEQ(LHS: R, RHS: ConstantInt::get(Ty: R->getType(), V: 0)); |
2346 | R = Builder.CreateSelect(C: IsZero, True: Step, False: R); |
2347 | } |
2348 | |
2349 | VectorTripCount = Builder.CreateSub(LHS: TC, RHS: R, Name: "n.vec" ); |
2350 | |
2351 | return VectorTripCount; |
2352 | } |
2353 | |
2354 | void InnerLoopVectorizer::introduceCheckBlockInVPlan(BasicBlock *CheckIRBB) { |
2355 | // Note: The block with the minimum trip-count check is already connected |
2356 | // during earlier VPlan construction. |
2357 | VPBlockBase *ScalarPH = Plan.getScalarPreheader(); |
2358 | VPBlockBase *PreVectorPH = VectorPHVPB->getSinglePredecessor(); |
2359 | assert(PreVectorPH->getNumSuccessors() == 2 && "Expected 2 successors" ); |
2360 | assert(PreVectorPH->getSuccessors()[0] == ScalarPH && "Unexpected successor" ); |
2361 | VPIRBasicBlock *CheckVPIRBB = Plan.createVPIRBasicBlock(IRBB: CheckIRBB); |
2362 | VPBlockUtils::insertOnEdge(From: PreVectorPH, To: VectorPHVPB, BlockPtr: CheckVPIRBB); |
2363 | PreVectorPH = CheckVPIRBB; |
2364 | VPBlockUtils::connectBlocks(From: PreVectorPH, To: ScalarPH); |
2365 | PreVectorPH->swapSuccessors(); |
2366 | |
2367 | // We just connected a new block to the scalar preheader. Update all |
2368 | // VPPhis by adding an incoming value for it, replicating the last value. |
2369 | unsigned NumPredecessors = ScalarPH->getNumPredecessors(); |
2370 | for (VPRecipeBase &R : cast<VPBasicBlock>(Val: ScalarPH)->phis()) { |
2371 | assert(isa<VPPhi>(&R) && "Phi expected to be VPPhi" ); |
2372 | assert(cast<VPPhi>(&R)->getNumIncoming() == NumPredecessors - 1 && |
2373 | "must have incoming values for all operands" ); |
2374 | R.addOperand(Operand: R.getOperand(N: NumPredecessors - 2)); |
2375 | } |
2376 | } |
2377 | |
2378 | Value *InnerLoopVectorizer::createIterationCountCheck(ElementCount VF, |
2379 | unsigned UF) const { |
2380 | // Generate code to check if the loop's trip count is less than VF * UF, or |
2381 | // equal to it in case a scalar epilogue is required; this implies that the |
2382 | // vector trip count is zero. This check also covers the case where adding one |
2383 | // to the backedge-taken count overflowed leading to an incorrect trip count |
2384 | // of zero. In this case we will also jump to the scalar loop. |
2385 | auto P = Cost->requiresScalarEpilogue(IsVectorizing: VF.isVector()) ? ICmpInst::ICMP_ULE |
2386 | : ICmpInst::ICMP_ULT; |
2387 | |
2388 | // Reuse existing vector loop preheader for TC checks. |
2389 | // Note that new preheader block is generated for vector loop. |
2390 | BasicBlock *const TCCheckBlock = LoopVectorPreHeader; |
2391 | IRBuilder<> Builder(TCCheckBlock->getTerminator()); |
2392 | |
2393 | // If tail is to be folded, vector loop takes care of all iterations. |
2394 | Value *Count = getTripCount(); |
2395 | Type *CountTy = Count->getType(); |
2396 | Value *CheckMinIters = Builder.getFalse(); |
2397 | auto CreateStep = [&]() -> Value * { |
2398 | // Create step with max(MinProTripCount, UF * VF). |
2399 | if (UF * VF.getKnownMinValue() >= MinProfitableTripCount.getKnownMinValue()) |
2400 | return createStepForVF(B&: Builder, Ty: CountTy, VF, Step: UF); |
2401 | |
2402 | Value *MinProfTC = |
2403 | createStepForVF(B&: Builder, Ty: CountTy, VF: MinProfitableTripCount, Step: 1); |
2404 | if (!VF.isScalable()) |
2405 | return MinProfTC; |
2406 | return Builder.CreateBinaryIntrinsic( |
2407 | ID: Intrinsic::umax, LHS: MinProfTC, RHS: createStepForVF(B&: Builder, Ty: CountTy, VF, Step: UF)); |
2408 | }; |
2409 | |
2410 | TailFoldingStyle Style = Cost->getTailFoldingStyle(); |
2411 | if (Style == TailFoldingStyle::None) { |
2412 | Value *Step = CreateStep(); |
2413 | ScalarEvolution &SE = *PSE.getSE(); |
2414 | // TODO: Emit unconditional branch to vector preheader instead of |
2415 | // conditional branch with known condition. |
2416 | const SCEV *TripCountSCEV = SE.applyLoopGuards(Expr: SE.getSCEV(V: Count), L: OrigLoop); |
2417 | // Check if the trip count is < the step. |
2418 | if (SE.isKnownPredicate(Pred: P, LHS: TripCountSCEV, RHS: SE.getSCEV(V: Step))) { |
2419 | // TODO: Ensure step is at most the trip count when determining max VF and |
2420 | // UF, w/o tail folding. |
2421 | CheckMinIters = Builder.getTrue(); |
2422 | } else if (!SE.isKnownPredicate(Pred: CmpInst::getInversePredicate(pred: P), |
2423 | LHS: TripCountSCEV, RHS: SE.getSCEV(V: Step))) { |
2424 | // Generate the minimum iteration check only if we cannot prove the |
2425 | // check is known to be true, or known to be false. |
2426 | CheckMinIters = Builder.CreateICmp(P, LHS: Count, RHS: Step, Name: "min.iters.check" ); |
2427 | } // else step known to be < trip count, use CheckMinIters preset to false. |
2428 | } else if (VF.isScalable() && !TTI->isVScaleKnownToBeAPowerOfTwo() && |
2429 | !isIndvarOverflowCheckKnownFalse(Cost, VF, UF) && |
2430 | Style != TailFoldingStyle::DataAndControlFlowWithoutRuntimeCheck) { |
2431 | // vscale is not necessarily a power-of-2, which means we cannot guarantee |
2432 | // an overflow to zero when updating induction variables and so an |
2433 | // additional overflow check is required before entering the vector loop. |
2434 | |
2435 | // Get the maximum unsigned value for the type. |
2436 | Value *MaxUIntTripCount = |
2437 | ConstantInt::get(Ty: CountTy, V: cast<IntegerType>(Val: CountTy)->getMask()); |
2438 | Value *LHS = Builder.CreateSub(LHS: MaxUIntTripCount, RHS: Count); |
2439 | |
2440 | // Don't execute the vector loop if (UMax - n) < (VF * UF). |
2441 | CheckMinIters = Builder.CreateICmp(P: ICmpInst::ICMP_ULT, LHS, RHS: CreateStep()); |
2442 | } |
2443 | return CheckMinIters; |
2444 | } |
2445 | |
2446 | void InnerLoopVectorizer::emitIterationCountCheck(BasicBlock *Bypass) { |
2447 | BasicBlock *const TCCheckBlock = LoopVectorPreHeader; |
2448 | Value *CheckMinIters = createIterationCountCheck(VF, UF); |
2449 | // Create new preheader for vector loop. |
2450 | LoopVectorPreHeader = SplitBlock(Old: TCCheckBlock, SplitPt: TCCheckBlock->getTerminator(), |
2451 | DT: static_cast<DominatorTree *>(nullptr), LI, |
2452 | MSSAU: nullptr, BBName: "vector.ph" ); |
2453 | |
2454 | BranchInst &BI = |
2455 | *BranchInst::Create(IfTrue: Bypass, IfFalse: LoopVectorPreHeader, Cond: CheckMinIters); |
2456 | if (hasBranchWeightMD(I: *OrigLoop->getLoopLatch()->getTerminator())) |
2457 | setBranchWeights(I&: BI, Weights: MinItersBypassWeights, /*IsExpected=*/false); |
2458 | ReplaceInstWithInst(From: TCCheckBlock->getTerminator(), To: &BI); |
2459 | |
2460 | assert(cast<VPIRBasicBlock>(Plan.getEntry())->getIRBasicBlock() == |
2461 | TCCheckBlock && |
2462 | "Plan's entry must be TCCCheckBlock" ); |
2463 | } |
2464 | |
2465 | BasicBlock *InnerLoopVectorizer::emitSCEVChecks(BasicBlock *Bypass) { |
2466 | BasicBlock *const SCEVCheckBlock = |
2467 | RTChecks.emitSCEVChecks(Bypass, LoopVectorPreHeader); |
2468 | if (!SCEVCheckBlock) |
2469 | return nullptr; |
2470 | |
2471 | assert((!Cost->OptForSize || |
2472 | Cost->Hints->getForce() == LoopVectorizeHints::FK_Enabled) && |
2473 | "Cannot SCEV check stride or overflow when optimizing for size" ); |
2474 | |
2475 | introduceCheckBlockInVPlan(CheckIRBB: SCEVCheckBlock); |
2476 | return SCEVCheckBlock; |
2477 | } |
2478 | |
2479 | BasicBlock *InnerLoopVectorizer::emitMemRuntimeChecks(BasicBlock *Bypass) { |
2480 | BasicBlock *const MemCheckBlock = |
2481 | RTChecks.emitMemRuntimeChecks(Bypass, LoopVectorPreHeader); |
2482 | |
2483 | // Check if we generated code that checks in runtime if arrays overlap. We put |
2484 | // the checks into a separate block to make the more common case of few |
2485 | // elements faster. |
2486 | if (!MemCheckBlock) |
2487 | return nullptr; |
2488 | |
2489 | // VPlan-native path does not do any analysis for runtime checks currently. |
2490 | assert((!EnableVPlanNativePath || OrigLoop->begin() == OrigLoop->end()) && |
2491 | "Runtime checks are not supported for outer loops yet" ); |
2492 | |
2493 | if (Cost->OptForSize) { |
2494 | assert(Cost->Hints->getForce() == LoopVectorizeHints::FK_Enabled && |
2495 | "Cannot emit memory checks when optimizing for size, unless forced " |
2496 | "to vectorize." ); |
2497 | ORE->emit(RemarkBuilder: [&]() { |
2498 | return OptimizationRemarkAnalysis(DEBUG_TYPE, "VectorizationCodeSize" , |
2499 | OrigLoop->getStartLoc(), |
2500 | OrigLoop->getHeader()) |
2501 | << "Code-size may be reduced by not forcing " |
2502 | "vectorization, or by source-code modifications " |
2503 | "eliminating the need for runtime checks " |
2504 | "(e.g., adding 'restrict')." ; |
2505 | }); |
2506 | } |
2507 | |
2508 | introduceCheckBlockInVPlan(CheckIRBB: MemCheckBlock); |
2509 | return MemCheckBlock; |
2510 | } |
2511 | |
2512 | /// Replace \p VPBB with a VPIRBasicBlock wrapping \p IRBB. All recipes from \p |
2513 | /// VPBB are moved to the end of the newly created VPIRBasicBlock. VPBB must |
2514 | /// have a single predecessor, which is rewired to the new VPIRBasicBlock. All |
2515 | /// successors of VPBB, if any, are rewired to the new VPIRBasicBlock. |
2516 | static void replaceVPBBWithIRVPBB(VPBasicBlock *VPBB, BasicBlock *IRBB) { |
2517 | VPIRBasicBlock *IRVPBB = VPBB->getPlan()->createVPIRBasicBlock(IRBB); |
2518 | for (auto &R : make_early_inc_range(Range&: *VPBB)) { |
2519 | assert((IRVPBB->empty() || IRVPBB->back().isPhi() || !R.isPhi()) && |
2520 | "Tried to move phi recipe after a non-phi recipe" ); |
2521 | R.moveBefore(BB&: *IRVPBB, I: IRVPBB->end()); |
2522 | } |
2523 | |
2524 | VPBlockUtils::reassociateBlocks(Old: VPBB, New: IRVPBB); |
2525 | // VPBB is now dead and will be cleaned up when the plan gets destroyed. |
2526 | } |
2527 | |
2528 | void InnerLoopVectorizer::createVectorLoopSkeleton(StringRef Prefix) { |
2529 | LoopVectorPreHeader = OrigLoop->getLoopPreheader(); |
2530 | assert(LoopVectorPreHeader && "Invalid loop structure" ); |
2531 | assert((OrigLoop->getUniqueLatchExitBlock() || |
2532 | Cost->requiresScalarEpilogue(VF.isVector())) && |
2533 | "loops not exiting via the latch without required epilogue?" ); |
2534 | |
2535 | LoopScalarPreHeader = |
2536 | SplitBlock(Old: LoopVectorPreHeader, SplitPt: LoopVectorPreHeader->getTerminator(), DT, |
2537 | LI, MSSAU: nullptr, BBName: Twine(Prefix) + "scalar.ph" ); |
2538 | // NOTE: The Plan's scalar preheader VPBB isn't replaced with a VPIRBasicBlock |
2539 | // wrapping LoopScalarPreHeader here at the moment, because the Plan's scalar |
2540 | // preheader may be unreachable at this point. Instead it is replaced in |
2541 | // createVectorizedLoopSkeleton. |
2542 | } |
2543 | |
2544 | /// Return the expanded step for \p ID using \p ExpandedSCEVs to look up SCEV |
2545 | /// expansion results. |
2546 | static Value *getExpandedStep(const InductionDescriptor &ID, |
2547 | const SCEV2ValueTy &ExpandedSCEVs) { |
2548 | const SCEV *Step = ID.getStep(); |
2549 | if (auto *C = dyn_cast<SCEVConstant>(Val: Step)) |
2550 | return C->getValue(); |
2551 | if (auto *U = dyn_cast<SCEVUnknown>(Val: Step)) |
2552 | return U->getValue(); |
2553 | Value *V = ExpandedSCEVs.lookup(Val: Step); |
2554 | assert(V && "SCEV must be expanded at this point" ); |
2555 | return V; |
2556 | } |
2557 | |
2558 | /// Knowing that loop \p L executes a single vector iteration, add instructions |
2559 | /// that will get simplified and thus should not have any cost to \p |
2560 | /// InstsToIgnore. |
2561 | static void addFullyUnrolledInstructionsToIgnore( |
2562 | Loop *L, const LoopVectorizationLegality::InductionList &IL, |
2563 | SmallPtrSetImpl<Instruction *> &InstsToIgnore) { |
2564 | auto *Cmp = L->getLatchCmpInst(); |
2565 | if (Cmp) |
2566 | InstsToIgnore.insert(Ptr: Cmp); |
2567 | for (const auto &KV : IL) { |
2568 | // Extract the key by hand so that it can be used in the lambda below. Note |
2569 | // that captured structured bindings are a C++20 extension. |
2570 | const PHINode *IV = KV.first; |
2571 | |
2572 | // Get next iteration value of the induction variable. |
2573 | Instruction *IVInst = |
2574 | cast<Instruction>(Val: IV->getIncomingValueForBlock(BB: L->getLoopLatch())); |
2575 | if (all_of(Range: IVInst->users(), |
2576 | P: [&](const User *U) { return U == IV || U == Cmp; })) |
2577 | InstsToIgnore.insert(Ptr: IVInst); |
2578 | } |
2579 | } |
2580 | |
2581 | BasicBlock *InnerLoopVectorizer::createVectorizedLoopSkeleton() { |
2582 | /* |
2583 | In this function we generate a new loop. The new loop will contain |
2584 | the vectorized instructions while the old loop will continue to run the |
2585 | scalar remainder. |
2586 | |
2587 | [ ] <-- old preheader - loop iteration number check and SCEVs in Plan's |
2588 | / | preheader are expanded here. Eventually all required SCEV |
2589 | / | expansion should happen here. |
2590 | / v |
2591 | | [ ] <-- vector loop bypass (may consist of multiple blocks). |
2592 | | / | |
2593 | | / v |
2594 | || [ ] <-- vector pre header. |
2595 | |/ | |
2596 | | v |
2597 | | [ ] \ |
2598 | | [ ]_| <-- vector loop (created during VPlan execution). |
2599 | | | |
2600 | | v |
2601 | \ -[ ] <--- middle-block (wrapped in VPIRBasicBlock with the branch to |
2602 | | | successors created during VPlan execution) |
2603 | \/ | |
2604 | /\ v |
2605 | | ->[ ] <--- new preheader (wrapped in VPIRBasicBlock). |
2606 | | | |
2607 | (opt) v <-- edge from middle to exit iff epilogue is not required. |
2608 | | [ ] \ |
2609 | | [ ]_| <-- old scalar loop to handle remainder (scalar epilogue, header |
2610 | | | wrapped in VPIRBasicBlock). |
2611 | \ | |
2612 | \ v |
2613 | >[ ] <-- exit block(s). (wrapped in VPIRBasicBlock) |
2614 | ... |
2615 | */ |
2616 | |
2617 | // Create an empty vector loop, and prepare basic blocks for the runtime |
2618 | // checks. |
2619 | createVectorLoopSkeleton(Prefix: "" ); |
2620 | |
2621 | // Now, compare the new count to zero. If it is zero skip the vector loop and |
2622 | // jump to the scalar loop. This check also covers the case where the |
2623 | // backedge-taken count is uint##_max: adding one to it will overflow leading |
2624 | // to an incorrect trip count of zero. In this (rare) case we will also jump |
2625 | // to the scalar loop. |
2626 | emitIterationCountCheck(Bypass: LoopScalarPreHeader); |
2627 | |
2628 | // Generate the code to check any assumptions that we've made for SCEV |
2629 | // expressions. |
2630 | emitSCEVChecks(Bypass: LoopScalarPreHeader); |
2631 | |
2632 | // Generate the code that checks in runtime if arrays overlap. We put the |
2633 | // checks into a separate block to make the more common case of few elements |
2634 | // faster. |
2635 | emitMemRuntimeChecks(Bypass: LoopScalarPreHeader); |
2636 | |
2637 | replaceVPBBWithIRVPBB(VPBB: Plan.getScalarPreheader(), IRBB: LoopScalarPreHeader); |
2638 | return LoopVectorPreHeader; |
2639 | } |
2640 | |
2641 | namespace { |
2642 | |
2643 | struct CSEDenseMapInfo { |
2644 | static bool canHandle(const Instruction *I) { |
2645 | return isa<InsertElementInst>(Val: I) || isa<ExtractElementInst>(Val: I) || |
2646 | isa<ShuffleVectorInst>(Val: I) || isa<GetElementPtrInst>(Val: I); |
2647 | } |
2648 | |
2649 | static inline Instruction *getEmptyKey() { |
2650 | return DenseMapInfo<Instruction *>::getEmptyKey(); |
2651 | } |
2652 | |
2653 | static inline Instruction *getTombstoneKey() { |
2654 | return DenseMapInfo<Instruction *>::getTombstoneKey(); |
2655 | } |
2656 | |
2657 | static unsigned getHashValue(const Instruction *I) { |
2658 | assert(canHandle(I) && "Unknown instruction!" ); |
2659 | return hash_combine(args: I->getOpcode(), |
2660 | args: hash_combine_range(R: I->operand_values())); |
2661 | } |
2662 | |
2663 | static bool isEqual(const Instruction *LHS, const Instruction *RHS) { |
2664 | if (LHS == getEmptyKey() || RHS == getEmptyKey() || |
2665 | LHS == getTombstoneKey() || RHS == getTombstoneKey()) |
2666 | return LHS == RHS; |
2667 | return LHS->isIdenticalTo(I: RHS); |
2668 | } |
2669 | }; |
2670 | |
2671 | } // end anonymous namespace |
2672 | |
2673 | ///Perform cse of induction variable instructions. |
2674 | static void cse(BasicBlock *BB) { |
2675 | // Perform simple cse. |
2676 | SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap; |
2677 | for (Instruction &In : llvm::make_early_inc_range(Range&: *BB)) { |
2678 | if (!CSEDenseMapInfo::canHandle(I: &In)) |
2679 | continue; |
2680 | |
2681 | // Check if we can replace this instruction with any of the |
2682 | // visited instructions. |
2683 | if (Instruction *V = CSEMap.lookup(Val: &In)) { |
2684 | In.replaceAllUsesWith(V); |
2685 | In.eraseFromParent(); |
2686 | continue; |
2687 | } |
2688 | |
2689 | CSEMap[&In] = &In; |
2690 | } |
2691 | } |
2692 | |
2693 | /// This function attempts to return a value that represents the vectorization |
2694 | /// factor at runtime. For fixed-width VFs we know this precisely at compile |
2695 | /// time, but for scalable VFs we calculate it based on an estimate of the |
2696 | /// vscale value. |
2697 | static unsigned getEstimatedRuntimeVF(ElementCount VF, |
2698 | std::optional<unsigned> VScale) { |
2699 | unsigned EstimatedVF = VF.getKnownMinValue(); |
2700 | if (VF.isScalable()) |
2701 | if (VScale) |
2702 | EstimatedVF *= *VScale; |
2703 | assert(EstimatedVF >= 1 && "Estimated VF shouldn't be less than 1" ); |
2704 | return EstimatedVF; |
2705 | } |
2706 | |
2707 | InstructionCost |
2708 | LoopVectorizationCostModel::getVectorCallCost(CallInst *CI, |
2709 | ElementCount VF) const { |
2710 | // We only need to calculate a cost if the VF is scalar; for actual vectors |
2711 | // we should already have a pre-calculated cost at each VF. |
2712 | if (!VF.isScalar()) |
2713 | return getCallWideningDecision(CI, VF).Cost; |
2714 | |
2715 | Type *RetTy = CI->getType(); |
2716 | if (RecurrenceDescriptor::isFMulAddIntrinsic(I: CI)) |
2717 | if (auto RedCost = getReductionPatternCost(I: CI, VF, VectorTy: RetTy)) |
2718 | return *RedCost; |
2719 | |
2720 | SmallVector<Type *, 4> Tys; |
2721 | for (auto &ArgOp : CI->args()) |
2722 | Tys.push_back(Elt: ArgOp->getType()); |
2723 | |
2724 | InstructionCost ScalarCallCost = |
2725 | TTI.getCallInstrCost(F: CI->getCalledFunction(), RetTy, Tys, CostKind); |
2726 | |
2727 | // If this is an intrinsic we may have a lower cost for it. |
2728 | if (getVectorIntrinsicIDForCall(CI, TLI)) { |
2729 | InstructionCost IntrinsicCost = getVectorIntrinsicCost(CI, VF); |
2730 | return std::min(a: ScalarCallCost, b: IntrinsicCost); |
2731 | } |
2732 | return ScalarCallCost; |
2733 | } |
2734 | |
2735 | static Type *maybeVectorizeType(Type *Ty, ElementCount VF) { |
2736 | if (VF.isScalar() || !canVectorizeTy(Ty)) |
2737 | return Ty; |
2738 | return toVectorizedTy(Ty, EC: VF); |
2739 | } |
2740 | |
2741 | InstructionCost |
2742 | LoopVectorizationCostModel::getVectorIntrinsicCost(CallInst *CI, |
2743 | ElementCount VF) const { |
2744 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); |
2745 | assert(ID && "Expected intrinsic call!" ); |
2746 | Type *RetTy = maybeVectorizeType(Ty: CI->getType(), VF); |
2747 | FastMathFlags FMF; |
2748 | if (auto *FPMO = dyn_cast<FPMathOperator>(Val: CI)) |
2749 | FMF = FPMO->getFastMathFlags(); |
2750 | |
2751 | SmallVector<const Value *> Arguments(CI->args()); |
2752 | FunctionType *FTy = CI->getCalledFunction()->getFunctionType(); |
2753 | SmallVector<Type *> ParamTys; |
2754 | std::transform(first: FTy->param_begin(), last: FTy->param_end(), |
2755 | result: std::back_inserter(x&: ParamTys), |
2756 | unary_op: [&](Type *Ty) { return maybeVectorizeType(Ty, VF); }); |
2757 | |
2758 | IntrinsicCostAttributes CostAttrs(ID, RetTy, Arguments, ParamTys, FMF, |
2759 | dyn_cast<IntrinsicInst>(Val: CI), |
2760 | InstructionCost::getInvalid(), TLI); |
2761 | return TTI.getIntrinsicInstrCost(ICA: CostAttrs, CostKind); |
2762 | } |
2763 | |
2764 | void InnerLoopVectorizer::fixVectorizedLoop(VPTransformState &State) { |
2765 | // Fix widened non-induction PHIs by setting up the PHI operands. |
2766 | fixNonInductionPHIs(State); |
2767 | |
2768 | // After vectorization, the exit blocks of the original loop will have |
2769 | // additional predecessors. Invalidate SCEVs for the exit phis in case SE |
2770 | // looked through single-entry phis. |
2771 | SmallVector<BasicBlock *> ExitBlocks; |
2772 | OrigLoop->getExitBlocks(ExitBlocks); |
2773 | for (BasicBlock *Exit : ExitBlocks) |
2774 | for (PHINode &PN : Exit->phis()) |
2775 | PSE.getSE()->forgetLcssaPhiWithNewPredecessor(L: OrigLoop, V: &PN); |
2776 | |
2777 | // Forget the original basic block. |
2778 | PSE.getSE()->forgetLoop(L: OrigLoop); |
2779 | PSE.getSE()->forgetBlockAndLoopDispositions(); |
2780 | |
2781 | // Don't apply optimizations below when no (vector) loop remains, as they all |
2782 | // require one at the moment. |
2783 | VPBasicBlock * = |
2784 | vputils::getFirstLoopHeader(Plan&: *State.Plan, VPDT&: State.VPDT); |
2785 | if (!HeaderVPBB) |
2786 | return; |
2787 | |
2788 | BasicBlock * = State.CFG.VPBB2IRBB[HeaderVPBB]; |
2789 | |
2790 | // Remove redundant induction instructions. |
2791 | cse(BB: HeaderBB); |
2792 | |
2793 | // Set/update profile weights for the vector and remainder loops as original |
2794 | // loop iterations are now distributed among them. Note that original loop |
2795 | // becomes the scalar remainder loop after vectorization. |
2796 | // |
2797 | // For cases like foldTailByMasking() and requiresScalarEpiloque() we may |
2798 | // end up getting slightly roughened result but that should be OK since |
2799 | // profile is not inherently precise anyway. Note also possible bypass of |
2800 | // vector code caused by legality checks is ignored, assigning all the weight |
2801 | // to the vector loop, optimistically. |
2802 | // |
2803 | // For scalable vectorization we can't know at compile time how many |
2804 | // iterations of the loop are handled in one vector iteration, so instead |
2805 | // use the value of vscale used for tuning. |
2806 | Loop *VectorLoop = LI->getLoopFor(BB: HeaderBB); |
2807 | unsigned EstimatedVFxUF = |
2808 | getEstimatedRuntimeVF(VF: VF * UF, VScale: Cost->getVScaleForTuning()); |
2809 | setProfileInfoAfterUnrolling(OrigLoop, UnrolledLoop: VectorLoop, RemainderLoop: OrigLoop, UF: EstimatedVFxUF); |
2810 | } |
2811 | |
2812 | void InnerLoopVectorizer::fixNonInductionPHIs(VPTransformState &State) { |
2813 | auto Iter = vp_depth_first_shallow(G: Plan.getEntry()); |
2814 | for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(Range: Iter)) { |
2815 | for (VPRecipeBase &P : VPBB->phis()) { |
2816 | VPWidenPHIRecipe *VPPhi = dyn_cast<VPWidenPHIRecipe>(Val: &P); |
2817 | if (!VPPhi) |
2818 | continue; |
2819 | PHINode *NewPhi = cast<PHINode>(Val: State.get(Def: VPPhi)); |
2820 | // Make sure the builder has a valid insert point. |
2821 | Builder.SetInsertPoint(NewPhi); |
2822 | for (unsigned Idx = 0; Idx < VPPhi->getNumIncoming(); ++Idx) { |
2823 | VPValue *Inc = VPPhi->getIncomingValue(Idx); |
2824 | const VPBasicBlock *VPBB = VPPhi->getIncomingBlock(Idx); |
2825 | NewPhi->addIncoming(V: State.get(Def: Inc), BB: State.CFG.VPBB2IRBB[VPBB]); |
2826 | } |
2827 | } |
2828 | } |
2829 | } |
2830 | |
2831 | void LoopVectorizationCostModel::collectLoopScalars(ElementCount VF) { |
2832 | // We should not collect Scalars more than once per VF. Right now, this |
2833 | // function is called from collectUniformsAndScalars(), which already does |
2834 | // this check. Collecting Scalars for VF=1 does not make any sense. |
2835 | assert(VF.isVector() && !Scalars.contains(VF) && |
2836 | "This function should not be visited twice for the same VF" ); |
2837 | |
2838 | // This avoids any chances of creating a REPLICATE recipe during planning |
2839 | // since that would result in generation of scalarized code during execution, |
2840 | // which is not supported for scalable vectors. |
2841 | if (VF.isScalable()) { |
2842 | Scalars[VF].insert_range(R&: Uniforms[VF]); |
2843 | return; |
2844 | } |
2845 | |
2846 | SmallSetVector<Instruction *, 8> Worklist; |
2847 | |
2848 | // These sets are used to seed the analysis with pointers used by memory |
2849 | // accesses that will remain scalar. |
2850 | SmallSetVector<Instruction *, 8> ScalarPtrs; |
2851 | SmallPtrSet<Instruction *, 8> PossibleNonScalarPtrs; |
2852 | auto *Latch = TheLoop->getLoopLatch(); |
2853 | |
2854 | // A helper that returns true if the use of Ptr by MemAccess will be scalar. |
2855 | // The pointer operands of loads and stores will be scalar as long as the |
2856 | // memory access is not a gather or scatter operation. The value operand of a |
2857 | // store will remain scalar if the store is scalarized. |
2858 | auto IsScalarUse = [&](Instruction *MemAccess, Value *Ptr) { |
2859 | InstWidening WideningDecision = getWideningDecision(I: MemAccess, VF); |
2860 | assert(WideningDecision != CM_Unknown && |
2861 | "Widening decision should be ready at this moment" ); |
2862 | if (auto *Store = dyn_cast<StoreInst>(Val: MemAccess)) |
2863 | if (Ptr == Store->getValueOperand()) |
2864 | return WideningDecision == CM_Scalarize; |
2865 | assert(Ptr == getLoadStorePointerOperand(MemAccess) && |
2866 | "Ptr is neither a value or pointer operand" ); |
2867 | return WideningDecision != CM_GatherScatter; |
2868 | }; |
2869 | |
2870 | // A helper that returns true if the given value is a getelementptr |
2871 | // instruction contained in the loop. |
2872 | auto IsLoopVaryingGEP = [&](Value *V) { |
2873 | return isa<GetElementPtrInst>(Val: V) && !TheLoop->isLoopInvariant(V); |
2874 | }; |
2875 | |
2876 | // A helper that evaluates a memory access's use of a pointer. If the use will |
2877 | // be a scalar use and the pointer is only used by memory accesses, we place |
2878 | // the pointer in ScalarPtrs. Otherwise, the pointer is placed in |
2879 | // PossibleNonScalarPtrs. |
2880 | auto EvaluatePtrUse = [&](Instruction *MemAccess, Value *Ptr) { |
2881 | // We only care about bitcast and getelementptr instructions contained in |
2882 | // the loop. |
2883 | if (!IsLoopVaryingGEP(Ptr)) |
2884 | return; |
2885 | |
2886 | // If the pointer has already been identified as scalar (e.g., if it was |
2887 | // also identified as uniform), there's nothing to do. |
2888 | auto *I = cast<Instruction>(Val: Ptr); |
2889 | if (Worklist.count(key: I)) |
2890 | return; |
2891 | |
2892 | // If the use of the pointer will be a scalar use, and all users of the |
2893 | // pointer are memory accesses, place the pointer in ScalarPtrs. Otherwise, |
2894 | // place the pointer in PossibleNonScalarPtrs. |
2895 | if (IsScalarUse(MemAccess, Ptr) && |
2896 | all_of(Range: I->users(), P: IsaPred<LoadInst, StoreInst>)) |
2897 | ScalarPtrs.insert(X: I); |
2898 | else |
2899 | PossibleNonScalarPtrs.insert(Ptr: I); |
2900 | }; |
2901 | |
2902 | // We seed the scalars analysis with three classes of instructions: (1) |
2903 | // instructions marked uniform-after-vectorization and (2) bitcast, |
2904 | // getelementptr and (pointer) phi instructions used by memory accesses |
2905 | // requiring a scalar use. |
2906 | // |
2907 | // (1) Add to the worklist all instructions that have been identified as |
2908 | // uniform-after-vectorization. |
2909 | Worklist.insert_range(R&: Uniforms[VF]); |
2910 | |
2911 | // (2) Add to the worklist all bitcast and getelementptr instructions used by |
2912 | // memory accesses requiring a scalar use. The pointer operands of loads and |
2913 | // stores will be scalar unless the operation is a gather or scatter. |
2914 | // The value operand of a store will remain scalar if the store is scalarized. |
2915 | for (auto *BB : TheLoop->blocks()) |
2916 | for (auto &I : *BB) { |
2917 | if (auto *Load = dyn_cast<LoadInst>(Val: &I)) { |
2918 | EvaluatePtrUse(Load, Load->getPointerOperand()); |
2919 | } else if (auto *Store = dyn_cast<StoreInst>(Val: &I)) { |
2920 | EvaluatePtrUse(Store, Store->getPointerOperand()); |
2921 | EvaluatePtrUse(Store, Store->getValueOperand()); |
2922 | } |
2923 | } |
2924 | for (auto *I : ScalarPtrs) |
2925 | if (!PossibleNonScalarPtrs.count(Ptr: I)) { |
2926 | LLVM_DEBUG(dbgs() << "LV: Found scalar instruction: " << *I << "\n" ); |
2927 | Worklist.insert(X: I); |
2928 | } |
2929 | |
2930 | // Insert the forced scalars. |
2931 | // FIXME: Currently VPWidenPHIRecipe() often creates a dead vector |
2932 | // induction variable when the PHI user is scalarized. |
2933 | auto ForcedScalar = ForcedScalars.find(Val: VF); |
2934 | if (ForcedScalar != ForcedScalars.end()) |
2935 | for (auto *I : ForcedScalar->second) { |
2936 | LLVM_DEBUG(dbgs() << "LV: Found (forced) scalar instruction: " << *I << "\n" ); |
2937 | Worklist.insert(X: I); |
2938 | } |
2939 | |
2940 | // Expand the worklist by looking through any bitcasts and getelementptr |
2941 | // instructions we've already identified as scalar. This is similar to the |
2942 | // expansion step in collectLoopUniforms(); however, here we're only |
2943 | // expanding to include additional bitcasts and getelementptr instructions. |
2944 | unsigned Idx = 0; |
2945 | while (Idx != Worklist.size()) { |
2946 | Instruction *Dst = Worklist[Idx++]; |
2947 | if (!IsLoopVaryingGEP(Dst->getOperand(i: 0))) |
2948 | continue; |
2949 | auto *Src = cast<Instruction>(Val: Dst->getOperand(i: 0)); |
2950 | if (llvm::all_of(Range: Src->users(), P: [&](User *U) -> bool { |
2951 | auto *J = cast<Instruction>(Val: U); |
2952 | return !TheLoop->contains(Inst: J) || Worklist.count(key: J) || |
2953 | ((isa<LoadInst>(Val: J) || isa<StoreInst>(Val: J)) && |
2954 | IsScalarUse(J, Src)); |
2955 | })) { |
2956 | Worklist.insert(X: Src); |
2957 | LLVM_DEBUG(dbgs() << "LV: Found scalar instruction: " << *Src << "\n" ); |
2958 | } |
2959 | } |
2960 | |
2961 | // An induction variable will remain scalar if all users of the induction |
2962 | // variable and induction variable update remain scalar. |
2963 | for (const auto &Induction : Legal->getInductionVars()) { |
2964 | auto *Ind = Induction.first; |
2965 | auto *IndUpdate = cast<Instruction>(Val: Ind->getIncomingValueForBlock(BB: Latch)); |
2966 | |
2967 | // If tail-folding is applied, the primary induction variable will be used |
2968 | // to feed a vector compare. |
2969 | if (Ind == Legal->getPrimaryInduction() && foldTailByMasking()) |
2970 | continue; |
2971 | |
2972 | // Returns true if \p Indvar is a pointer induction that is used directly by |
2973 | // load/store instruction \p I. |
2974 | auto IsDirectLoadStoreFromPtrIndvar = [&](Instruction *Indvar, |
2975 | Instruction *I) { |
2976 | return Induction.second.getKind() == |
2977 | InductionDescriptor::IK_PtrInduction && |
2978 | (isa<LoadInst>(Val: I) || isa<StoreInst>(Val: I)) && |
2979 | Indvar == getLoadStorePointerOperand(V: I) && IsScalarUse(I, Indvar); |
2980 | }; |
2981 | |
2982 | // Determine if all users of the induction variable are scalar after |
2983 | // vectorization. |
2984 | bool ScalarInd = all_of(Range: Ind->users(), P: [&](User *U) -> bool { |
2985 | auto *I = cast<Instruction>(Val: U); |
2986 | return I == IndUpdate || !TheLoop->contains(Inst: I) || Worklist.count(key: I) || |
2987 | IsDirectLoadStoreFromPtrIndvar(Ind, I); |
2988 | }); |
2989 | if (!ScalarInd) |
2990 | continue; |
2991 | |
2992 | // If the induction variable update is a fixed-order recurrence, neither the |
2993 | // induction variable or its update should be marked scalar after |
2994 | // vectorization. |
2995 | auto *IndUpdatePhi = dyn_cast<PHINode>(Val: IndUpdate); |
2996 | if (IndUpdatePhi && Legal->isFixedOrderRecurrence(Phi: IndUpdatePhi)) |
2997 | continue; |
2998 | |
2999 | // Determine if all users of the induction variable update instruction are |
3000 | // scalar after vectorization. |
3001 | bool ScalarIndUpdate = all_of(Range: IndUpdate->users(), P: [&](User *U) -> bool { |
3002 | auto *I = cast<Instruction>(Val: U); |
3003 | return I == Ind || !TheLoop->contains(Inst: I) || Worklist.count(key: I) || |
3004 | IsDirectLoadStoreFromPtrIndvar(IndUpdate, I); |
3005 | }); |
3006 | if (!ScalarIndUpdate) |
3007 | continue; |
3008 | |
3009 | // The induction variable and its update instruction will remain scalar. |
3010 | Worklist.insert(X: Ind); |
3011 | Worklist.insert(X: IndUpdate); |
3012 | LLVM_DEBUG(dbgs() << "LV: Found scalar instruction: " << *Ind << "\n" ); |
3013 | LLVM_DEBUG(dbgs() << "LV: Found scalar instruction: " << *IndUpdate |
3014 | << "\n" ); |
3015 | } |
3016 | |
3017 | Scalars[VF].insert_range(R&: Worklist); |
3018 | } |
3019 | |
3020 | bool LoopVectorizationCostModel::isScalarWithPredication( |
3021 | Instruction *I, ElementCount VF) const { |
3022 | if (!isPredicatedInst(I)) |
3023 | return false; |
3024 | |
3025 | // Do we have a non-scalar lowering for this predicated |
3026 | // instruction? No - it is scalar with predication. |
3027 | switch(I->getOpcode()) { |
3028 | default: |
3029 | return true; |
3030 | case Instruction::Call: |
3031 | if (VF.isScalar()) |
3032 | return true; |
3033 | return getCallWideningDecision(CI: cast<CallInst>(Val: I), VF).Kind == CM_Scalarize; |
3034 | case Instruction::Load: |
3035 | case Instruction::Store: { |
3036 | auto *Ptr = getLoadStorePointerOperand(V: I); |
3037 | auto *Ty = getLoadStoreType(I); |
3038 | unsigned AS = getLoadStoreAddressSpace(I); |
3039 | Type *VTy = Ty; |
3040 | if (VF.isVector()) |
3041 | VTy = VectorType::get(ElementType: Ty, EC: VF); |
3042 | const Align Alignment = getLoadStoreAlignment(I); |
3043 | return isa<LoadInst>(Val: I) ? !(isLegalMaskedLoad(DataType: Ty, Ptr, Alignment, AddressSpace: AS) || |
3044 | TTI.isLegalMaskedGather(DataType: VTy, Alignment)) |
3045 | : !(isLegalMaskedStore(DataType: Ty, Ptr, Alignment, AddressSpace: AS) || |
3046 | TTI.isLegalMaskedScatter(DataType: VTy, Alignment)); |
3047 | } |
3048 | case Instruction::UDiv: |
3049 | case Instruction::SDiv: |
3050 | case Instruction::SRem: |
3051 | case Instruction::URem: { |
3052 | // We have the option to use the safe-divisor idiom to avoid predication. |
3053 | // The cost based decision here will always select safe-divisor for |
3054 | // scalable vectors as scalarization isn't legal. |
3055 | const auto [ScalarCost, SafeDivisorCost] = getDivRemSpeculationCost(I, VF); |
3056 | return isDivRemScalarWithPredication(ScalarCost, SafeDivisorCost); |
3057 | } |
3058 | } |
3059 | } |
3060 | |
3061 | // TODO: Fold into LoopVectorizationLegality::isMaskRequired. |
3062 | bool LoopVectorizationCostModel::isPredicatedInst(Instruction *I) const { |
3063 | // TODO: We can use the loop-preheader as context point here and get |
3064 | // context sensitive reasoning for isSafeToSpeculativelyExecute. |
3065 | if (isSafeToSpeculativelyExecute(I) || |
3066 | (isa<LoadInst, StoreInst, CallInst>(Val: I) && !Legal->isMaskRequired(I)) || |
3067 | isa<BranchInst, SwitchInst, PHINode, AllocaInst>(Val: I)) |
3068 | return false; |
3069 | |
3070 | // If the instruction was executed conditionally in the original scalar loop, |
3071 | // predication is needed with a mask whose lanes are all possibly inactive. |
3072 | if (Legal->blockNeedsPredication(BB: I->getParent())) |
3073 | return true; |
3074 | |
3075 | // If we're not folding the tail by masking, predication is unnecessary. |
3076 | if (!foldTailByMasking()) |
3077 | return false; |
3078 | |
3079 | // All that remain are instructions with side-effects originally executed in |
3080 | // the loop unconditionally, but now execute under a tail-fold mask (only) |
3081 | // having at least one active lane (the first). If the side-effects of the |
3082 | // instruction are invariant, executing it w/o (the tail-folding) mask is safe |
3083 | // - it will cause the same side-effects as when masked. |
3084 | switch(I->getOpcode()) { |
3085 | default: |
3086 | llvm_unreachable( |
3087 | "instruction should have been considered by earlier checks" ); |
3088 | case Instruction::Call: |
3089 | // Side-effects of a Call are assumed to be non-invariant, needing a |
3090 | // (fold-tail) mask. |
3091 | assert(Legal->isMaskRequired(I) && |
3092 | "should have returned earlier for calls not needing a mask" ); |
3093 | return true; |
3094 | case Instruction::Load: |
3095 | // If the address is loop invariant no predication is needed. |
3096 | return !Legal->isInvariant(V: getLoadStorePointerOperand(V: I)); |
3097 | case Instruction::Store: { |
3098 | // For stores, we need to prove both speculation safety (which follows from |
3099 | // the same argument as loads), but also must prove the value being stored |
3100 | // is correct. The easiest form of the later is to require that all values |
3101 | // stored are the same. |
3102 | return !(Legal->isInvariant(V: getLoadStorePointerOperand(V: I)) && |
3103 | Legal->isInvariant(V: cast<StoreInst>(Val: I)->getValueOperand())); |
3104 | } |
3105 | case Instruction::UDiv: |
3106 | case Instruction::SDiv: |
3107 | case Instruction::SRem: |
3108 | case Instruction::URem: |
3109 | // If the divisor is loop-invariant no predication is needed. |
3110 | return !Legal->isInvariant(V: I->getOperand(i: 1)); |
3111 | } |
3112 | } |
3113 | |
3114 | std::pair<InstructionCost, InstructionCost> |
3115 | LoopVectorizationCostModel::getDivRemSpeculationCost(Instruction *I, |
3116 | ElementCount VF) const { |
3117 | assert(I->getOpcode() == Instruction::UDiv || |
3118 | I->getOpcode() == Instruction::SDiv || |
3119 | I->getOpcode() == Instruction::SRem || |
3120 | I->getOpcode() == Instruction::URem); |
3121 | assert(!isSafeToSpeculativelyExecute(I)); |
3122 | |
3123 | // Scalarization isn't legal for scalable vector types |
3124 | InstructionCost ScalarizationCost = InstructionCost::getInvalid(); |
3125 | if (!VF.isScalable()) { |
3126 | // Get the scalarization cost and scale this amount by the probability of |
3127 | // executing the predicated block. If the instruction is not predicated, |
3128 | // we fall through to the next case. |
3129 | ScalarizationCost = 0; |
3130 | |
3131 | // These instructions have a non-void type, so account for the phi nodes |
3132 | // that we will create. This cost is likely to be zero. The phi node |
3133 | // cost, if any, should be scaled by the block probability because it |
3134 | // models a copy at the end of each predicated block. |
3135 | ScalarizationCost += |
3136 | VF.getFixedValue() * TTI.getCFInstrCost(Opcode: Instruction::PHI, CostKind); |
3137 | |
3138 | // The cost of the non-predicated instruction. |
3139 | ScalarizationCost += |
3140 | VF.getFixedValue() * |
3141 | TTI.getArithmeticInstrCost(Opcode: I->getOpcode(), Ty: I->getType(), CostKind); |
3142 | |
3143 | // The cost of insertelement and extractelement instructions needed for |
3144 | // scalarization. |
3145 | ScalarizationCost += getScalarizationOverhead(I, VF); |
3146 | |
3147 | // Scale the cost by the probability of executing the predicated blocks. |
3148 | // This assumes the predicated block for each vector lane is equally |
3149 | // likely. |
3150 | ScalarizationCost = ScalarizationCost / getPredBlockCostDivisor(CostKind); |
3151 | } |
3152 | InstructionCost SafeDivisorCost = 0; |
3153 | |
3154 | auto *VecTy = toVectorTy(Scalar: I->getType(), EC: VF); |
3155 | |
3156 | // The cost of the select guard to ensure all lanes are well defined |
3157 | // after we speculate above any internal control flow. |
3158 | SafeDivisorCost += |
3159 | TTI.getCmpSelInstrCost(Opcode: Instruction::Select, ValTy: VecTy, |
3160 | CondTy: toVectorTy(Scalar: Type::getInt1Ty(C&: I->getContext()), EC: VF), |
3161 | VecPred: CmpInst::BAD_ICMP_PREDICATE, CostKind); |
3162 | |
3163 | // Certain instructions can be cheaper to vectorize if they have a constant |
3164 | // second vector operand. One example of this are shifts on x86. |
3165 | Value *Op2 = I->getOperand(i: 1); |
3166 | auto Op2Info = TTI.getOperandInfo(V: Op2); |
3167 | if (Op2Info.Kind == TargetTransformInfo::OK_AnyValue && |
3168 | Legal->isInvariant(V: Op2)) |
3169 | Op2Info.Kind = TargetTransformInfo::OK_UniformValue; |
3170 | |
3171 | SmallVector<const Value *, 4> Operands(I->operand_values()); |
3172 | SafeDivisorCost += TTI.getArithmeticInstrCost( |
3173 | Opcode: I->getOpcode(), Ty: VecTy, CostKind, |
3174 | Opd1Info: {.Kind: TargetTransformInfo::OK_AnyValue, .Properties: TargetTransformInfo::OP_None}, |
3175 | Opd2Info: Op2Info, Args: Operands, CxtI: I); |
3176 | return {ScalarizationCost, SafeDivisorCost}; |
3177 | } |
3178 | |
3179 | bool LoopVectorizationCostModel::interleavedAccessCanBeWidened( |
3180 | Instruction *I, ElementCount VF) const { |
3181 | assert(isAccessInterleaved(I) && "Expecting interleaved access." ); |
3182 | assert(getWideningDecision(I, VF) == CM_Unknown && |
3183 | "Decision should not be set yet." ); |
3184 | auto *Group = getInterleavedAccessGroup(Instr: I); |
3185 | assert(Group && "Must have a group." ); |
3186 | unsigned InterleaveFactor = Group->getFactor(); |
3187 | |
3188 | // If the instruction's allocated size doesn't equal its type size, it |
3189 | // requires padding and will be scalarized. |
3190 | auto &DL = I->getDataLayout(); |
3191 | auto *ScalarTy = getLoadStoreType(I); |
3192 | if (hasIrregularType(Ty: ScalarTy, DL)) |
3193 | return false; |
3194 | |
3195 | // For scalable vectors, the interleave factors must be <= 8 since we require |
3196 | // the (de)interleaveN intrinsics instead of shufflevectors. |
3197 | if (VF.isScalable() && InterleaveFactor > 8) |
3198 | return false; |
3199 | |
3200 | // If the group involves a non-integral pointer, we may not be able to |
3201 | // losslessly cast all values to a common type. |
3202 | bool ScalarNI = DL.isNonIntegralPointerType(Ty: ScalarTy); |
3203 | for (unsigned Idx = 0; Idx < InterleaveFactor; Idx++) { |
3204 | Instruction *Member = Group->getMember(Index: Idx); |
3205 | if (!Member) |
3206 | continue; |
3207 | auto *MemberTy = getLoadStoreType(I: Member); |
3208 | bool MemberNI = DL.isNonIntegralPointerType(Ty: MemberTy); |
3209 | // Don't coerce non-integral pointers to integers or vice versa. |
3210 | if (MemberNI != ScalarNI) |
3211 | // TODO: Consider adding special nullptr value case here |
3212 | return false; |
3213 | if (MemberNI && ScalarNI && |
3214 | ScalarTy->getPointerAddressSpace() != |
3215 | MemberTy->getPointerAddressSpace()) |
3216 | return false; |
3217 | } |
3218 | |
3219 | // Check if masking is required. |
3220 | // A Group may need masking for one of two reasons: it resides in a block that |
3221 | // needs predication, or it was decided to use masking to deal with gaps |
3222 | // (either a gap at the end of a load-access that may result in a speculative |
3223 | // load, or any gaps in a store-access). |
3224 | bool PredicatedAccessRequiresMasking = |
3225 | blockNeedsPredicationForAnyReason(BB: I->getParent()) && |
3226 | Legal->isMaskRequired(I); |
3227 | bool LoadAccessWithGapsRequiresEpilogMasking = |
3228 | isa<LoadInst>(Val: I) && Group->requiresScalarEpilogue() && |
3229 | !isScalarEpilogueAllowed(); |
3230 | bool StoreAccessWithGapsRequiresMasking = |
3231 | isa<StoreInst>(Val: I) && (Group->getNumMembers() < Group->getFactor()); |
3232 | if (!PredicatedAccessRequiresMasking && |
3233 | !LoadAccessWithGapsRequiresEpilogMasking && |
3234 | !StoreAccessWithGapsRequiresMasking) |
3235 | return true; |
3236 | |
3237 | // If masked interleaving is required, we expect that the user/target had |
3238 | // enabled it, because otherwise it either wouldn't have been created or |
3239 | // it should have been invalidated by the CostModel. |
3240 | assert(useMaskedInterleavedAccesses(TTI) && |
3241 | "Masked interleave-groups for predicated accesses are not enabled." ); |
3242 | |
3243 | if (Group->isReverse()) |
3244 | return false; |
3245 | |
3246 | auto *Ty = getLoadStoreType(I); |
3247 | const Align Alignment = getLoadStoreAlignment(I); |
3248 | unsigned AS = getLoadStoreAddressSpace(I); |
3249 | return isa<LoadInst>(Val: I) ? TTI.isLegalMaskedLoad(DataType: Ty, Alignment, AddressSpace: AS) |
3250 | : TTI.isLegalMaskedStore(DataType: Ty, Alignment, AddressSpace: AS); |
3251 | } |
3252 | |
3253 | bool LoopVectorizationCostModel::memoryInstructionCanBeWidened( |
3254 | Instruction *I, ElementCount VF) { |
3255 | // Get and ensure we have a valid memory instruction. |
3256 | assert((isa<LoadInst, StoreInst>(I)) && "Invalid memory instruction" ); |
3257 | |
3258 | auto *Ptr = getLoadStorePointerOperand(V: I); |
3259 | auto *ScalarTy = getLoadStoreType(I); |
3260 | |
3261 | // In order to be widened, the pointer should be consecutive, first of all. |
3262 | if (!Legal->isConsecutivePtr(AccessTy: ScalarTy, Ptr)) |
3263 | return false; |
3264 | |
3265 | // If the instruction is a store located in a predicated block, it will be |
3266 | // scalarized. |
3267 | if (isScalarWithPredication(I, VF)) |
3268 | return false; |
3269 | |
3270 | // If the instruction's allocated size doesn't equal it's type size, it |
3271 | // requires padding and will be scalarized. |
3272 | auto &DL = I->getDataLayout(); |
3273 | if (hasIrregularType(Ty: ScalarTy, DL)) |
3274 | return false; |
3275 | |
3276 | return true; |
3277 | } |
3278 | |
3279 | void LoopVectorizationCostModel::collectLoopUniforms(ElementCount VF) { |
3280 | // We should not collect Uniforms more than once per VF. Right now, |
3281 | // this function is called from collectUniformsAndScalars(), which |
3282 | // already does this check. Collecting Uniforms for VF=1 does not make any |
3283 | // sense. |
3284 | |
3285 | assert(VF.isVector() && !Uniforms.contains(VF) && |
3286 | "This function should not be visited twice for the same VF" ); |
3287 | |
3288 | // Visit the list of Uniforms. If we find no uniform value, we won't |
3289 | // analyze again. Uniforms.count(VF) will return 1. |
3290 | Uniforms[VF].clear(); |
3291 | |
3292 | // Now we know that the loop is vectorizable! |
3293 | // Collect instructions inside the loop that will remain uniform after |
3294 | // vectorization. |
3295 | |
3296 | // Global values, params and instructions outside of current loop are out of |
3297 | // scope. |
3298 | auto IsOutOfScope = [&](Value *V) -> bool { |
3299 | Instruction *I = dyn_cast<Instruction>(Val: V); |
3300 | return (!I || !TheLoop->contains(Inst: I)); |
3301 | }; |
3302 | |
3303 | // Worklist containing uniform instructions demanding lane 0. |
3304 | SetVector<Instruction *> Worklist; |
3305 | |
3306 | // Add uniform instructions demanding lane 0 to the worklist. Instructions |
3307 | // that require predication must not be considered uniform after |
3308 | // vectorization, because that would create an erroneous replicating region |
3309 | // where only a single instance out of VF should be formed. |
3310 | auto AddToWorklistIfAllowed = [&](Instruction *I) -> void { |
3311 | if (IsOutOfScope(I)) { |
3312 | LLVM_DEBUG(dbgs() << "LV: Found not uniform due to scope: " |
3313 | << *I << "\n" ); |
3314 | return; |
3315 | } |
3316 | if (isPredicatedInst(I)) { |
3317 | LLVM_DEBUG( |
3318 | dbgs() << "LV: Found not uniform due to requiring predication: " << *I |
3319 | << "\n" ); |
3320 | return; |
3321 | } |
3322 | LLVM_DEBUG(dbgs() << "LV: Found uniform instruction: " << *I << "\n" ); |
3323 | Worklist.insert(X: I); |
3324 | }; |
3325 | |
3326 | // Start with the conditional branches exiting the loop. If the branch |
3327 | // condition is an instruction contained in the loop that is only used by the |
3328 | // branch, it is uniform. Note conditions from uncountable early exits are not |
3329 | // uniform. |
3330 | SmallVector<BasicBlock *> Exiting; |
3331 | TheLoop->getExitingBlocks(ExitingBlocks&: Exiting); |
3332 | for (BasicBlock *E : Exiting) { |
3333 | if (Legal->hasUncountableEarlyExit() && TheLoop->getLoopLatch() != E) |
3334 | continue; |
3335 | auto *Cmp = dyn_cast<Instruction>(Val: E->getTerminator()->getOperand(i: 0)); |
3336 | if (Cmp && TheLoop->contains(Inst: Cmp) && Cmp->hasOneUse()) |
3337 | AddToWorklistIfAllowed(Cmp); |
3338 | } |
3339 | |
3340 | auto PrevVF = VF.divideCoefficientBy(RHS: 2); |
3341 | // Return true if all lanes perform the same memory operation, and we can |
3342 | // thus choose to execute only one. |
3343 | auto IsUniformMemOpUse = [&](Instruction *I) { |
3344 | // If the value was already known to not be uniform for the previous |
3345 | // (smaller VF), it cannot be uniform for the larger VF. |
3346 | if (PrevVF.isVector()) { |
3347 | auto Iter = Uniforms.find(Val: PrevVF); |
3348 | if (Iter != Uniforms.end() && !Iter->second.contains(Ptr: I)) |
3349 | return false; |
3350 | } |
3351 | if (!Legal->isUniformMemOp(I&: *I, VF)) |
3352 | return false; |
3353 | if (isa<LoadInst>(Val: I)) |
3354 | // Loading the same address always produces the same result - at least |
3355 | // assuming aliasing and ordering which have already been checked. |
3356 | return true; |
3357 | // Storing the same value on every iteration. |
3358 | return TheLoop->isLoopInvariant(V: cast<StoreInst>(Val: I)->getValueOperand()); |
3359 | }; |
3360 | |
3361 | auto IsUniformDecision = [&](Instruction *I, ElementCount VF) { |
3362 | InstWidening WideningDecision = getWideningDecision(I, VF); |
3363 | assert(WideningDecision != CM_Unknown && |
3364 | "Widening decision should be ready at this moment" ); |
3365 | |
3366 | if (IsUniformMemOpUse(I)) |
3367 | return true; |
3368 | |
3369 | return (WideningDecision == CM_Widen || |
3370 | WideningDecision == CM_Widen_Reverse || |
3371 | WideningDecision == CM_Interleave); |
3372 | }; |
3373 | |
3374 | // Returns true if Ptr is the pointer operand of a memory access instruction |
3375 | // I, I is known to not require scalarization, and the pointer is not also |
3376 | // stored. |
3377 | auto IsVectorizedMemAccessUse = [&](Instruction *I, Value *Ptr) -> bool { |
3378 | if (isa<StoreInst>(Val: I) && I->getOperand(i: 0) == Ptr) |
3379 | return false; |
3380 | return getLoadStorePointerOperand(V: I) == Ptr && |
3381 | (IsUniformDecision(I, VF) || Legal->isInvariant(V: Ptr)); |
3382 | }; |
3383 | |
3384 | // Holds a list of values which are known to have at least one uniform use. |
3385 | // Note that there may be other uses which aren't uniform. A "uniform use" |
3386 | // here is something which only demands lane 0 of the unrolled iterations; |
3387 | // it does not imply that all lanes produce the same value (e.g. this is not |
3388 | // the usual meaning of uniform) |
3389 | SetVector<Value *> HasUniformUse; |
3390 | |
3391 | // Scan the loop for instructions which are either a) known to have only |
3392 | // lane 0 demanded or b) are uses which demand only lane 0 of their operand. |
3393 | for (auto *BB : TheLoop->blocks()) |
3394 | for (auto &I : *BB) { |
3395 | if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Val: &I)) { |
3396 | switch (II->getIntrinsicID()) { |
3397 | case Intrinsic::sideeffect: |
3398 | case Intrinsic::experimental_noalias_scope_decl: |
3399 | case Intrinsic::assume: |
3400 | case Intrinsic::lifetime_start: |
3401 | case Intrinsic::lifetime_end: |
3402 | if (TheLoop->hasLoopInvariantOperands(I: &I)) |
3403 | AddToWorklistIfAllowed(&I); |
3404 | break; |
3405 | default: |
3406 | break; |
3407 | } |
3408 | } |
3409 | |
3410 | if (auto *EVI = dyn_cast<ExtractValueInst>(Val: &I)) { |
3411 | if (IsOutOfScope(EVI->getAggregateOperand())) { |
3412 | AddToWorklistIfAllowed(EVI); |
3413 | continue; |
3414 | } |
3415 | // Only ExtractValue instructions where the aggregate value comes from a |
3416 | // call are allowed to be non-uniform. |
3417 | assert(isa<CallInst>(EVI->getAggregateOperand()) && |
3418 | "Expected aggregate value to be call return value" ); |
3419 | } |
3420 | |
3421 | // If there's no pointer operand, there's nothing to do. |
3422 | auto *Ptr = getLoadStorePointerOperand(V: &I); |
3423 | if (!Ptr) |
3424 | continue; |
3425 | |
3426 | if (IsUniformMemOpUse(&I)) |
3427 | AddToWorklistIfAllowed(&I); |
3428 | |
3429 | if (IsVectorizedMemAccessUse(&I, Ptr)) |
3430 | HasUniformUse.insert(X: Ptr); |
3431 | } |
3432 | |
3433 | // Add to the worklist any operands which have *only* uniform (e.g. lane 0 |
3434 | // demanding) users. Since loops are assumed to be in LCSSA form, this |
3435 | // disallows uses outside the loop as well. |
3436 | for (auto *V : HasUniformUse) { |
3437 | if (IsOutOfScope(V)) |
3438 | continue; |
3439 | auto *I = cast<Instruction>(Val: V); |
3440 | bool UsersAreMemAccesses = all_of(Range: I->users(), P: [&](User *U) -> bool { |
3441 | auto *UI = cast<Instruction>(Val: U); |
3442 | return TheLoop->contains(Inst: UI) && IsVectorizedMemAccessUse(UI, V); |
3443 | }); |
3444 | if (UsersAreMemAccesses) |
3445 | AddToWorklistIfAllowed(I); |
3446 | } |
3447 | |
3448 | // Expand Worklist in topological order: whenever a new instruction |
3449 | // is added , its users should be already inside Worklist. It ensures |
3450 | // a uniform instruction will only be used by uniform instructions. |
3451 | unsigned Idx = 0; |
3452 | while (Idx != Worklist.size()) { |
3453 | Instruction *I = Worklist[Idx++]; |
3454 | |
3455 | for (auto *OV : I->operand_values()) { |
3456 | // isOutOfScope operands cannot be uniform instructions. |
3457 | if (IsOutOfScope(OV)) |
3458 | continue; |
3459 | // First order recurrence Phi's should typically be considered |
3460 | // non-uniform. |
3461 | auto *OP = dyn_cast<PHINode>(Val: OV); |
3462 | if (OP && Legal->isFixedOrderRecurrence(Phi: OP)) |
3463 | continue; |
3464 | // If all the users of the operand are uniform, then add the |
3465 | // operand into the uniform worklist. |
3466 | auto *OI = cast<Instruction>(Val: OV); |
3467 | if (llvm::all_of(Range: OI->users(), P: [&](User *U) -> bool { |
3468 | auto *J = cast<Instruction>(Val: U); |
3469 | return Worklist.count(key: J) || IsVectorizedMemAccessUse(J, OI); |
3470 | })) |
3471 | AddToWorklistIfAllowed(OI); |
3472 | } |
3473 | } |
3474 | |
3475 | // For an instruction to be added into Worklist above, all its users inside |
3476 | // the loop should also be in Worklist. However, this condition cannot be |
3477 | // true for phi nodes that form a cyclic dependence. We must process phi |
3478 | // nodes separately. An induction variable will remain uniform if all users |
3479 | // of the induction variable and induction variable update remain uniform. |
3480 | // The code below handles both pointer and non-pointer induction variables. |
3481 | BasicBlock *Latch = TheLoop->getLoopLatch(); |
3482 | for (const auto &Induction : Legal->getInductionVars()) { |
3483 | auto *Ind = Induction.first; |
3484 | auto *IndUpdate = cast<Instruction>(Val: Ind->getIncomingValueForBlock(BB: Latch)); |
3485 | |
3486 | // Determine if all users of the induction variable are uniform after |
3487 | // vectorization. |
3488 | bool UniformInd = all_of(Range: Ind->users(), P: [&](User *U) -> bool { |
3489 | auto *I = cast<Instruction>(Val: U); |
3490 | return I == IndUpdate || !TheLoop->contains(Inst: I) || Worklist.count(key: I) || |
3491 | IsVectorizedMemAccessUse(I, Ind); |
3492 | }); |
3493 | if (!UniformInd) |
3494 | continue; |
3495 | |
3496 | // Determine if all users of the induction variable update instruction are |
3497 | // uniform after vectorization. |
3498 | bool UniformIndUpdate = all_of(Range: IndUpdate->users(), P: [&](User *U) -> bool { |
3499 | auto *I = cast<Instruction>(Val: U); |
3500 | return I == Ind || Worklist.count(key: I) || |
3501 | IsVectorizedMemAccessUse(I, IndUpdate); |
3502 | }); |
3503 | if (!UniformIndUpdate) |
3504 | continue; |
3505 | |
3506 | // The induction variable and its update instruction will remain uniform. |
3507 | AddToWorklistIfAllowed(Ind); |
3508 | AddToWorklistIfAllowed(IndUpdate); |
3509 | } |
3510 | |
3511 | Uniforms[VF].insert_range(R&: Worklist); |
3512 | } |
3513 | |
3514 | bool LoopVectorizationCostModel::runtimeChecksRequired() { |
3515 | LLVM_DEBUG(dbgs() << "LV: Performing code size checks.\n" ); |
3516 | |
3517 | if (Legal->getRuntimePointerChecking()->Need) { |
3518 | reportVectorizationFailure(DebugMsg: "Runtime ptr check is required with -Os/-Oz" , |
3519 | OREMsg: "runtime pointer checks needed. Enable vectorization of this " |
3520 | "loop with '#pragma clang loop vectorize(enable)' when " |
3521 | "compiling with -Os/-Oz" , |
3522 | ORETag: "CantVersionLoopWithOptForSize" , ORE, TheLoop); |
3523 | return true; |
3524 | } |
3525 | |
3526 | if (!PSE.getPredicate().isAlwaysTrue()) { |
3527 | reportVectorizationFailure(DebugMsg: "Runtime SCEV check is required with -Os/-Oz" , |
3528 | OREMsg: "runtime SCEV checks needed. Enable vectorization of this " |
3529 | "loop with '#pragma clang loop vectorize(enable)' when " |
3530 | "compiling with -Os/-Oz" , |
3531 | ORETag: "CantVersionLoopWithOptForSize" , ORE, TheLoop); |
3532 | return true; |
3533 | } |
3534 | |
3535 | // FIXME: Avoid specializing for stride==1 instead of bailing out. |
3536 | if (!Legal->getLAI()->getSymbolicStrides().empty()) { |
3537 | reportVectorizationFailure(DebugMsg: "Runtime stride check for small trip count" , |
3538 | OREMsg: "runtime stride == 1 checks needed. Enable vectorization of " |
3539 | "this loop without such check by compiling with -Os/-Oz" , |
3540 | ORETag: "CantVersionLoopWithOptForSize" , ORE, TheLoop); |
3541 | return true; |
3542 | } |
3543 | |
3544 | return false; |
3545 | } |
3546 | |
3547 | bool LoopVectorizationCostModel::isScalableVectorizationAllowed() { |
3548 | if (IsScalableVectorizationAllowed) |
3549 | return *IsScalableVectorizationAllowed; |
3550 | |
3551 | IsScalableVectorizationAllowed = false; |
3552 | if (!TTI.supportsScalableVectors() && !ForceTargetSupportsScalableVectors) |
3553 | return false; |
3554 | |
3555 | if (Hints->isScalableVectorizationDisabled()) { |
3556 | reportVectorizationInfo(Msg: "Scalable vectorization is explicitly disabled" , |
3557 | ORETag: "ScalableVectorizationDisabled" , ORE, TheLoop); |
3558 | return false; |
3559 | } |
3560 | |
3561 | LLVM_DEBUG(dbgs() << "LV: Scalable vectorization is available\n" ); |
3562 | |
3563 | auto MaxScalableVF = ElementCount::getScalable( |
3564 | MinVal: std::numeric_limits<ElementCount::ScalarTy>::max()); |
3565 | |
3566 | // Test that the loop-vectorizer can legalize all operations for this MaxVF. |
3567 | // FIXME: While for scalable vectors this is currently sufficient, this should |
3568 | // be replaced by a more detailed mechanism that filters out specific VFs, |
3569 | // instead of invalidating vectorization for a whole set of VFs based on the |
3570 | // MaxVF. |
3571 | |
3572 | // Disable scalable vectorization if the loop contains unsupported reductions. |
3573 | if (!canVectorizeReductions(VF: MaxScalableVF)) { |
3574 | reportVectorizationInfo( |
3575 | Msg: "Scalable vectorization not supported for the reduction " |
3576 | "operations found in this loop." , |
3577 | ORETag: "ScalableVFUnfeasible" , ORE, TheLoop); |
3578 | return false; |
3579 | } |
3580 | |
3581 | // Disable scalable vectorization if the loop contains any instructions |
3582 | // with element types not supported for scalable vectors. |
3583 | if (any_of(Range&: ElementTypesInLoop, P: [&](Type *Ty) { |
3584 | return !Ty->isVoidTy() && |
3585 | !this->TTI.isElementTypeLegalForScalableVector(Ty); |
3586 | })) { |
3587 | reportVectorizationInfo(Msg: "Scalable vectorization is not supported " |
3588 | "for all element types found in this loop." , |
3589 | ORETag: "ScalableVFUnfeasible" , ORE, TheLoop); |
3590 | return false; |
3591 | } |
3592 | |
3593 | if (!Legal->isSafeForAnyVectorWidth() && !getMaxVScale(F: *TheFunction, TTI)) { |
3594 | reportVectorizationInfo(Msg: "The target does not provide maximum vscale value " |
3595 | "for safe distance analysis." , |
3596 | ORETag: "ScalableVFUnfeasible" , ORE, TheLoop); |
3597 | return false; |
3598 | } |
3599 | |
3600 | IsScalableVectorizationAllowed = true; |
3601 | return true; |
3602 | } |
3603 | |
3604 | ElementCount |
3605 | LoopVectorizationCostModel::getMaxLegalScalableVF(unsigned MaxSafeElements) { |
3606 | if (!isScalableVectorizationAllowed()) |
3607 | return ElementCount::getScalable(MinVal: 0); |
3608 | |
3609 | auto MaxScalableVF = ElementCount::getScalable( |
3610 | MinVal: std::numeric_limits<ElementCount::ScalarTy>::max()); |
3611 | if (Legal->isSafeForAnyVectorWidth()) |
3612 | return MaxScalableVF; |
3613 | |
3614 | std::optional<unsigned> MaxVScale = getMaxVScale(F: *TheFunction, TTI); |
3615 | // Limit MaxScalableVF by the maximum safe dependence distance. |
3616 | MaxScalableVF = ElementCount::getScalable(MinVal: MaxSafeElements / *MaxVScale); |
3617 | |
3618 | if (!MaxScalableVF) |
3619 | reportVectorizationInfo( |
3620 | Msg: "Max legal vector width too small, scalable vectorization " |
3621 | "unfeasible." , |
3622 | ORETag: "ScalableVFUnfeasible" , ORE, TheLoop); |
3623 | |
3624 | return MaxScalableVF; |
3625 | } |
3626 | |
3627 | FixedScalableVFPair LoopVectorizationCostModel::computeFeasibleMaxVF( |
3628 | unsigned MaxTripCount, ElementCount UserVF, bool FoldTailByMasking) { |
3629 | MinBWs = computeMinimumValueSizes(Blocks: TheLoop->getBlocks(), DB&: *DB, TTI: &TTI); |
3630 | unsigned SmallestType, WidestType; |
3631 | std::tie(args&: SmallestType, args&: WidestType) = getSmallestAndWidestTypes(); |
3632 | |
3633 | // Get the maximum safe dependence distance in bits computed by LAA. |
3634 | // It is computed by MaxVF * sizeOf(type) * 8, where type is taken from |
3635 | // the memory accesses that is most restrictive (involved in the smallest |
3636 | // dependence distance). |
3637 | unsigned MaxSafeElementsPowerOf2 = |
3638 | bit_floor(Value: Legal->getMaxSafeVectorWidthInBits() / WidestType); |
3639 | if (!Legal->isSafeForAnyStoreLoadForwardDistances()) { |
3640 | unsigned SLDist = Legal->getMaxStoreLoadForwardSafeDistanceInBits(); |
3641 | MaxSafeElementsPowerOf2 = |
3642 | std::min(a: MaxSafeElementsPowerOf2, b: SLDist / WidestType); |
3643 | } |
3644 | auto MaxSafeFixedVF = ElementCount::getFixed(MinVal: MaxSafeElementsPowerOf2); |
3645 | auto MaxSafeScalableVF = getMaxLegalScalableVF(MaxSafeElements: MaxSafeElementsPowerOf2); |
3646 | |
3647 | if (!Legal->isSafeForAnyVectorWidth()) |
3648 | this->MaxSafeElements = MaxSafeElementsPowerOf2; |
3649 | |
3650 | LLVM_DEBUG(dbgs() << "LV: The max safe fixed VF is: " << MaxSafeFixedVF |
3651 | << ".\n" ); |
3652 | LLVM_DEBUG(dbgs() << "LV: The max safe scalable VF is: " << MaxSafeScalableVF |
3653 | << ".\n" ); |
3654 | |
3655 | // First analyze the UserVF, fall back if the UserVF should be ignored. |
3656 | if (UserVF) { |
3657 | auto MaxSafeUserVF = |
3658 | UserVF.isScalable() ? MaxSafeScalableVF : MaxSafeFixedVF; |
3659 | |
3660 | if (ElementCount::isKnownLE(LHS: UserVF, RHS: MaxSafeUserVF)) { |
3661 | // If `VF=vscale x N` is safe, then so is `VF=N` |
3662 | if (UserVF.isScalable()) |
3663 | return FixedScalableVFPair( |
3664 | ElementCount::getFixed(MinVal: UserVF.getKnownMinValue()), UserVF); |
3665 | |
3666 | return UserVF; |
3667 | } |
3668 | |
3669 | assert(ElementCount::isKnownGT(UserVF, MaxSafeUserVF)); |
3670 | |
3671 | // Only clamp if the UserVF is not scalable. If the UserVF is scalable, it |
3672 | // is better to ignore the hint and let the compiler choose a suitable VF. |
3673 | if (!UserVF.isScalable()) { |
3674 | LLVM_DEBUG(dbgs() << "LV: User VF=" << UserVF |
3675 | << " is unsafe, clamping to max safe VF=" |
3676 | << MaxSafeFixedVF << ".\n" ); |
3677 | ORE->emit(RemarkBuilder: [&]() { |
3678 | return OptimizationRemarkAnalysis(DEBUG_TYPE, "VectorizationFactor" , |
3679 | TheLoop->getStartLoc(), |
3680 | TheLoop->getHeader()) |
3681 | << "User-specified vectorization factor " |
3682 | << ore::NV("UserVectorizationFactor" , UserVF) |
3683 | << " is unsafe, clamping to maximum safe vectorization factor " |
3684 | << ore::NV("VectorizationFactor" , MaxSafeFixedVF); |
3685 | }); |
3686 | return MaxSafeFixedVF; |
3687 | } |
3688 | |
3689 | if (!TTI.supportsScalableVectors() && !ForceTargetSupportsScalableVectors) { |
3690 | LLVM_DEBUG(dbgs() << "LV: User VF=" << UserVF |
3691 | << " is ignored because scalable vectors are not " |
3692 | "available.\n" ); |
3693 | ORE->emit(RemarkBuilder: [&]() { |
3694 | return OptimizationRemarkAnalysis(DEBUG_TYPE, "VectorizationFactor" , |
3695 | TheLoop->getStartLoc(), |
3696 | TheLoop->getHeader()) |
3697 | << "User-specified vectorization factor " |
3698 | << ore::NV("UserVectorizationFactor" , UserVF) |
3699 | << " is ignored because the target does not support scalable " |
3700 | "vectors. The compiler will pick a more suitable value." ; |
3701 | }); |
3702 | } else { |
3703 | LLVM_DEBUG(dbgs() << "LV: User VF=" << UserVF |
3704 | << " is unsafe. Ignoring scalable UserVF.\n" ); |
3705 | ORE->emit(RemarkBuilder: [&]() { |
3706 | return OptimizationRemarkAnalysis(DEBUG_TYPE, "VectorizationFactor" , |
3707 | TheLoop->getStartLoc(), |
3708 | TheLoop->getHeader()) |
3709 | << "User-specified vectorization factor " |
3710 | << ore::NV("UserVectorizationFactor" , UserVF) |
3711 | << " is unsafe. Ignoring the hint to let the compiler pick a " |
3712 | "more suitable value." ; |
3713 | }); |
3714 | } |
3715 | } |
3716 | |
3717 | LLVM_DEBUG(dbgs() << "LV: The Smallest and Widest types: " << SmallestType |
3718 | << " / " << WidestType << " bits.\n" ); |
3719 | |
3720 | FixedScalableVFPair Result(ElementCount::getFixed(MinVal: 1), |
3721 | ElementCount::getScalable(MinVal: 0)); |
3722 | if (auto MaxVF = |
3723 | getMaximizedVFForTarget(MaxTripCount, SmallestType, WidestType, |
3724 | MaxSafeVF: MaxSafeFixedVF, FoldTailByMasking)) |
3725 | Result.FixedVF = MaxVF; |
3726 | |
3727 | if (auto MaxVF = |
3728 | getMaximizedVFForTarget(MaxTripCount, SmallestType, WidestType, |
3729 | MaxSafeVF: MaxSafeScalableVF, FoldTailByMasking)) |
3730 | if (MaxVF.isScalable()) { |
3731 | Result.ScalableVF = MaxVF; |
3732 | LLVM_DEBUG(dbgs() << "LV: Found feasible scalable VF = " << MaxVF |
3733 | << "\n" ); |
3734 | } |
3735 | |
3736 | return Result; |
3737 | } |
3738 | |
3739 | FixedScalableVFPair |
3740 | LoopVectorizationCostModel::computeMaxVF(ElementCount UserVF, unsigned UserIC) { |
3741 | if (Legal->getRuntimePointerChecking()->Need && TTI.hasBranchDivergence()) { |
3742 | // TODO: It may be useful to do since it's still likely to be dynamically |
3743 | // uniform if the target can skip. |
3744 | reportVectorizationFailure( |
3745 | DebugMsg: "Not inserting runtime ptr check for divergent target" , |
3746 | OREMsg: "runtime pointer checks needed. Not enabled for divergent target" , |
3747 | ORETag: "CantVersionLoopWithDivergentTarget" , ORE, TheLoop); |
3748 | return FixedScalableVFPair::getNone(); |
3749 | } |
3750 | |
3751 | ScalarEvolution *SE = PSE.getSE(); |
3752 | ElementCount TC = getSmallConstantTripCount(SE, L: TheLoop); |
3753 | unsigned MaxTC = PSE.getSmallConstantMaxTripCount(); |
3754 | LLVM_DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n'); |
3755 | if (TC != ElementCount::getFixed(MinVal: MaxTC)) |
3756 | LLVM_DEBUG(dbgs() << "LV: Found maximum trip count: " << MaxTC << '\n'); |
3757 | if (TC.isScalar()) { |
3758 | reportVectorizationFailure(DebugMsg: "Single iteration (non) loop" , |
3759 | OREMsg: "loop trip count is one, irrelevant for vectorization" , |
3760 | ORETag: "SingleIterationLoop" , ORE, TheLoop); |
3761 | return FixedScalableVFPair::getNone(); |
3762 | } |
3763 | |
3764 | // If BTC matches the widest induction type and is -1 then the trip count |
3765 | // computation will wrap to 0 and the vector trip count will be 0. Do not try |
3766 | // to vectorize. |
3767 | const SCEV *BTC = SE->getBackedgeTakenCount(L: TheLoop); |
3768 | if (!isa<SCEVCouldNotCompute>(Val: BTC) && |
3769 | BTC->getType()->getScalarSizeInBits() >= |
3770 | Legal->getWidestInductionType()->getScalarSizeInBits() && |
3771 | SE->isKnownPredicate(Pred: CmpInst::ICMP_EQ, LHS: BTC, |
3772 | RHS: SE->getMinusOne(Ty: BTC->getType()))) { |
3773 | reportVectorizationFailure( |
3774 | DebugMsg: "Trip count computation wrapped" , |
3775 | OREMsg: "backedge-taken count is -1, loop trip count wrapped to 0" , |
3776 | ORETag: "TripCountWrapped" , ORE, TheLoop); |
3777 | return FixedScalableVFPair::getNone(); |
3778 | } |
3779 | |
3780 | switch (ScalarEpilogueStatus) { |
3781 | case CM_ScalarEpilogueAllowed: |
3782 | return computeFeasibleMaxVF(MaxTripCount: MaxTC, UserVF, FoldTailByMasking: false); |
3783 | case CM_ScalarEpilogueNotAllowedUsePredicate: |
3784 | [[fallthrough]]; |
3785 | case CM_ScalarEpilogueNotNeededUsePredicate: |
3786 | LLVM_DEBUG( |
3787 | dbgs() << "LV: vector predicate hint/switch found.\n" |
3788 | << "LV: Not allowing scalar epilogue, creating predicated " |
3789 | << "vector loop.\n" ); |
3790 | break; |
3791 | case CM_ScalarEpilogueNotAllowedLowTripLoop: |
3792 | // fallthrough as a special case of OptForSize |
3793 | case CM_ScalarEpilogueNotAllowedOptSize: |
3794 | if (ScalarEpilogueStatus == CM_ScalarEpilogueNotAllowedOptSize) |
3795 | LLVM_DEBUG( |
3796 | dbgs() << "LV: Not allowing scalar epilogue due to -Os/-Oz.\n" ); |
3797 | else |
3798 | LLVM_DEBUG(dbgs() << "LV: Not allowing scalar epilogue due to low trip " |
3799 | << "count.\n" ); |
3800 | |
3801 | // Bail if runtime checks are required, which are not good when optimising |
3802 | // for size. |
3803 | if (runtimeChecksRequired()) |
3804 | return FixedScalableVFPair::getNone(); |
3805 | |
3806 | break; |
3807 | } |
3808 | |
3809 | // Now try the tail folding |
3810 | |
3811 | // Invalidate interleave groups that require an epilogue if we can't mask |
3812 | // the interleave-group. |
3813 | if (!useMaskedInterleavedAccesses(TTI)) { |
3814 | assert(WideningDecisions.empty() && Uniforms.empty() && Scalars.empty() && |
3815 | "No decisions should have been taken at this point" ); |
3816 | // Note: There is no need to invalidate any cost modeling decisions here, as |
3817 | // none were taken so far. |
3818 | InterleaveInfo.invalidateGroupsRequiringScalarEpilogue(); |
3819 | } |
3820 | |
3821 | FixedScalableVFPair MaxFactors = computeFeasibleMaxVF(MaxTripCount: MaxTC, UserVF, FoldTailByMasking: true); |
3822 | |
3823 | // Avoid tail folding if the trip count is known to be a multiple of any VF |
3824 | // we choose. |
3825 | std::optional<unsigned> MaxPowerOf2RuntimeVF = |
3826 | MaxFactors.FixedVF.getFixedValue(); |
3827 | if (MaxFactors.ScalableVF) { |
3828 | std::optional<unsigned> MaxVScale = getMaxVScale(F: *TheFunction, TTI); |
3829 | if (MaxVScale && TTI.isVScaleKnownToBeAPowerOfTwo()) { |
3830 | MaxPowerOf2RuntimeVF = std::max<unsigned>( |
3831 | a: *MaxPowerOf2RuntimeVF, |
3832 | b: *MaxVScale * MaxFactors.ScalableVF.getKnownMinValue()); |
3833 | } else |
3834 | MaxPowerOf2RuntimeVF = std::nullopt; // Stick with tail-folding for now. |
3835 | } |
3836 | |
3837 | auto NoScalarEpilogueNeeded = [this, &UserIC](unsigned MaxVF) { |
3838 | // Return false if the loop is neither a single-latch-exit loop nor an |
3839 | // early-exit loop as tail-folding is not supported in that case. |
3840 | if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch() && |
3841 | !Legal->hasUncountableEarlyExit()) |
3842 | return false; |
3843 | unsigned MaxVFtimesIC = UserIC ? MaxVF * UserIC : MaxVF; |
3844 | ScalarEvolution *SE = PSE.getSE(); |
3845 | // Calling getSymbolicMaxBackedgeTakenCount enables support for loops |
3846 | // with uncountable exits. For countable loops, the symbolic maximum must |
3847 | // remain identical to the known back-edge taken count. |
3848 | const SCEV *BackedgeTakenCount = PSE.getSymbolicMaxBackedgeTakenCount(); |
3849 | assert((Legal->hasUncountableEarlyExit() || |
3850 | BackedgeTakenCount == PSE.getBackedgeTakenCount()) && |
3851 | "Invalid loop count" ); |
3852 | const SCEV *ExitCount = SE->getAddExpr( |
3853 | LHS: BackedgeTakenCount, RHS: SE->getOne(Ty: BackedgeTakenCount->getType())); |
3854 | const SCEV *Rem = SE->getURemExpr( |
3855 | LHS: SE->applyLoopGuards(Expr: ExitCount, L: TheLoop), |
3856 | RHS: SE->getConstant(Ty: BackedgeTakenCount->getType(), V: MaxVFtimesIC)); |
3857 | return Rem->isZero(); |
3858 | }; |
3859 | |
3860 | if (MaxPowerOf2RuntimeVF > 0u) { |
3861 | assert((UserVF.isNonZero() || isPowerOf2_32(*MaxPowerOf2RuntimeVF)) && |
3862 | "MaxFixedVF must be a power of 2" ); |
3863 | if (NoScalarEpilogueNeeded(*MaxPowerOf2RuntimeVF)) { |
3864 | // Accept MaxFixedVF if we do not have a tail. |
3865 | LLVM_DEBUG(dbgs() << "LV: No tail will remain for any chosen VF.\n" ); |
3866 | return MaxFactors; |
3867 | } |
3868 | } |
3869 | |
3870 | auto ExpectedTC = getSmallBestKnownTC(PSE, L: TheLoop); |
3871 | if (ExpectedTC && ExpectedTC->isFixed() && |
3872 | ExpectedTC->getFixedValue() <= |
3873 | TTI.getMinTripCountTailFoldingThreshold()) { |
3874 | if (MaxPowerOf2RuntimeVF > 0u) { |
3875 | // If we have a low-trip-count, and the fixed-width VF is known to divide |
3876 | // the trip count but the scalable factor does not, use the fixed-width |
3877 | // factor in preference to allow the generation of a non-predicated loop. |
3878 | if (ScalarEpilogueStatus == CM_ScalarEpilogueNotAllowedLowTripLoop && |
3879 | NoScalarEpilogueNeeded(MaxFactors.FixedVF.getFixedValue())) { |
3880 | LLVM_DEBUG(dbgs() << "LV: Picking a fixed-width so that no tail will " |
3881 | "remain for any chosen VF.\n" ); |
3882 | MaxFactors.ScalableVF = ElementCount::getScalable(MinVal: 0); |
3883 | return MaxFactors; |
3884 | } |
3885 | } |
3886 | |
3887 | reportVectorizationFailure( |
3888 | DebugMsg: "The trip count is below the minial threshold value." , |
3889 | OREMsg: "loop trip count is too low, avoiding vectorization" , ORETag: "LowTripCount" , |
3890 | ORE, TheLoop); |
3891 | return FixedScalableVFPair::getNone(); |
3892 | } |
3893 | |
3894 | // If we don't know the precise trip count, or if the trip count that we |
3895 | // found modulo the vectorization factor is not zero, try to fold the tail |
3896 | // by masking. |
3897 | // FIXME: look for a smaller MaxVF that does divide TC rather than masking. |
3898 | bool ContainsScalableVF = MaxFactors.ScalableVF.isNonZero(); |
3899 | setTailFoldingStyles(IsScalableVF: ContainsScalableVF, UserIC); |
3900 | if (foldTailByMasking()) { |
3901 | if (getTailFoldingStyle() == TailFoldingStyle::DataWithEVL) { |
3902 | LLVM_DEBUG( |
3903 | dbgs() |
3904 | << "LV: tail is folded with EVL, forcing unroll factor to be 1. Will " |
3905 | "try to generate VP Intrinsics with scalable vector " |
3906 | "factors only.\n" ); |
3907 | // Tail folded loop using VP intrinsics restricts the VF to be scalable |
3908 | // for now. |
3909 | // TODO: extend it for fixed vectors, if required. |
3910 | assert(ContainsScalableVF && "Expected scalable vector factor." ); |
3911 | |
3912 | MaxFactors.FixedVF = ElementCount::getFixed(MinVal: 1); |
3913 | } |
3914 | return MaxFactors; |
3915 | } |
3916 | |
3917 | // If there was a tail-folding hint/switch, but we can't fold the tail by |
3918 | // masking, fallback to a vectorization with a scalar epilogue. |
3919 | if (ScalarEpilogueStatus == CM_ScalarEpilogueNotNeededUsePredicate) { |
3920 | LLVM_DEBUG(dbgs() << "LV: Cannot fold tail by masking: vectorize with a " |
3921 | "scalar epilogue instead.\n" ); |
3922 | ScalarEpilogueStatus = CM_ScalarEpilogueAllowed; |
3923 | return MaxFactors; |
3924 | } |
3925 | |
3926 | if (ScalarEpilogueStatus == CM_ScalarEpilogueNotAllowedUsePredicate) { |
3927 | LLVM_DEBUG(dbgs() << "LV: Can't fold tail by masking: don't vectorize\n" ); |
3928 | return FixedScalableVFPair::getNone(); |
3929 | } |
3930 | |
3931 | if (TC.isZero()) { |
3932 | reportVectorizationFailure( |
3933 | DebugMsg: "unable to calculate the loop count due to complex control flow" , |
3934 | ORETag: "UnknownLoopCountComplexCFG" , ORE, TheLoop); |
3935 | return FixedScalableVFPair::getNone(); |
3936 | } |
3937 | |
3938 | reportVectorizationFailure( |
3939 | DebugMsg: "Cannot optimize for size and vectorize at the same time." , |
3940 | OREMsg: "cannot optimize for size and vectorize at the same time. " |
3941 | "Enable vectorization of this loop with '#pragma clang loop " |
3942 | "vectorize(enable)' when compiling with -Os/-Oz" , |
3943 | ORETag: "NoTailLoopWithOptForSize" , ORE, TheLoop); |
3944 | return FixedScalableVFPair::getNone(); |
3945 | } |
3946 | |
3947 | bool LoopVectorizationCostModel::useMaxBandwidth(ElementCount VF) { |
3948 | return useMaxBandwidth(RegKind: VF.isScalable() |
3949 | ? TargetTransformInfo::RGK_ScalableVector |
3950 | : TargetTransformInfo::RGK_FixedWidthVector); |
3951 | } |
3952 | |
3953 | bool LoopVectorizationCostModel::useMaxBandwidth( |
3954 | TargetTransformInfo::RegisterKind RegKind) { |
3955 | return MaximizeBandwidth || (MaximizeBandwidth.getNumOccurrences() == 0 && |
3956 | (TTI.shouldMaximizeVectorBandwidth(K: RegKind) || |
3957 | (UseWiderVFIfCallVariantsPresent && |
3958 | Legal->hasVectorCallVariants()))); |
3959 | } |
3960 | |
3961 | ElementCount LoopVectorizationCostModel::getMaximizedVFForTarget( |
3962 | unsigned MaxTripCount, unsigned SmallestType, unsigned WidestType, |
3963 | ElementCount MaxSafeVF, bool FoldTailByMasking) { |
3964 | bool ComputeScalableMaxVF = MaxSafeVF.isScalable(); |
3965 | const TypeSize WidestRegister = TTI.getRegisterBitWidth( |
3966 | K: ComputeScalableMaxVF ? TargetTransformInfo::RGK_ScalableVector |
3967 | : TargetTransformInfo::RGK_FixedWidthVector); |
3968 | |
3969 | // Convenience function to return the minimum of two ElementCounts. |
3970 | auto MinVF = [](const ElementCount &LHS, const ElementCount &RHS) { |
3971 | assert((LHS.isScalable() == RHS.isScalable()) && |
3972 | "Scalable flags must match" ); |
3973 | return ElementCount::isKnownLT(LHS, RHS) ? LHS : RHS; |
3974 | }; |
3975 | |
3976 | // Ensure MaxVF is a power of 2; the dependence distance bound may not be. |
3977 | // Note that both WidestRegister and WidestType may not be a powers of 2. |
3978 | auto MaxVectorElementCount = ElementCount::get( |
3979 | MinVal: llvm::bit_floor(Value: WidestRegister.getKnownMinValue() / WidestType), |
3980 | Scalable: ComputeScalableMaxVF); |
3981 | MaxVectorElementCount = MinVF(MaxVectorElementCount, MaxSafeVF); |
3982 | LLVM_DEBUG(dbgs() << "LV: The Widest register safe to use is: " |
3983 | << (MaxVectorElementCount * WidestType) << " bits.\n" ); |
3984 | |
3985 | if (!MaxVectorElementCount) { |
3986 | LLVM_DEBUG(dbgs() << "LV: The target has no " |
3987 | << (ComputeScalableMaxVF ? "scalable" : "fixed" ) |
3988 | << " vector registers.\n" ); |
3989 | return ElementCount::getFixed(MinVal: 1); |
3990 | } |
3991 | |
3992 | unsigned WidestRegisterMinEC = MaxVectorElementCount.getKnownMinValue(); |
3993 | if (MaxVectorElementCount.isScalable() && |
3994 | TheFunction->hasFnAttribute(Kind: Attribute::VScaleRange)) { |
3995 | auto Attr = TheFunction->getFnAttribute(Kind: Attribute::VScaleRange); |
3996 | auto Min = Attr.getVScaleRangeMin(); |
3997 | WidestRegisterMinEC *= Min; |
3998 | } |
3999 | |
4000 | // When a scalar epilogue is required, at least one iteration of the scalar |
4001 | // loop has to execute. Adjust MaxTripCount accordingly to avoid picking a |
4002 | // max VF that results in a dead vector loop. |
4003 | if (MaxTripCount > 0 && requiresScalarEpilogue(IsVectorizing: true)) |
4004 | MaxTripCount -= 1; |
4005 | |
4006 | if (MaxTripCount && MaxTripCount <= WidestRegisterMinEC && |
4007 | (!FoldTailByMasking || isPowerOf2_32(Value: MaxTripCount))) { |
4008 | // If upper bound loop trip count (TC) is known at compile time there is no |
4009 | // point in choosing VF greater than TC (as done in the loop below). Select |
4010 | // maximum power of two which doesn't exceed TC. If MaxVectorElementCount is |
4011 | // scalable, we only fall back on a fixed VF when the TC is less than or |
4012 | // equal to the known number of lanes. |
4013 | auto ClampedUpperTripCount = llvm::bit_floor(Value: MaxTripCount); |
4014 | LLVM_DEBUG(dbgs() << "LV: Clamping the MaxVF to maximum power of two not " |
4015 | "exceeding the constant trip count: " |
4016 | << ClampedUpperTripCount << "\n" ); |
4017 | return ElementCount::get( |
4018 | MinVal: ClampedUpperTripCount, |
4019 | Scalable: FoldTailByMasking ? MaxVectorElementCount.isScalable() : false); |
4020 | } |
4021 | |
4022 | TargetTransformInfo::RegisterKind RegKind = |
4023 | ComputeScalableMaxVF ? TargetTransformInfo::RGK_ScalableVector |
4024 | : TargetTransformInfo::RGK_FixedWidthVector; |
4025 | ElementCount MaxVF = MaxVectorElementCount; |
4026 | if (useMaxBandwidth(RegKind)) { |
4027 | auto MaxVectorElementCountMaxBW = ElementCount::get( |
4028 | MinVal: llvm::bit_floor(Value: WidestRegister.getKnownMinValue() / SmallestType), |
4029 | Scalable: ComputeScalableMaxVF); |
4030 | MaxVF = MinVF(MaxVectorElementCountMaxBW, MaxSafeVF); |
4031 | |
4032 | if (ElementCount MinVF = |
4033 | TTI.getMinimumVF(ElemWidth: SmallestType, IsScalable: ComputeScalableMaxVF)) { |
4034 | if (ElementCount::isKnownLT(LHS: MaxVF, RHS: MinVF)) { |
4035 | LLVM_DEBUG(dbgs() << "LV: Overriding calculated MaxVF(" << MaxVF |
4036 | << ") with target's minimum: " << MinVF << '\n'); |
4037 | MaxVF = MinVF; |
4038 | } |
4039 | } |
4040 | |
4041 | // Invalidate any widening decisions we might have made, in case the loop |
4042 | // requires prediction (decided later), but we have already made some |
4043 | // load/store widening decisions. |
4044 | invalidateCostModelingDecisions(); |
4045 | } |
4046 | return MaxVF; |
4047 | } |
4048 | |
4049 | bool LoopVectorizationPlanner::isMoreProfitable(const VectorizationFactor &A, |
4050 | const VectorizationFactor &B, |
4051 | const unsigned MaxTripCount, |
4052 | bool HasTail) const { |
4053 | InstructionCost CostA = A.Cost; |
4054 | InstructionCost CostB = B.Cost; |
4055 | |
4056 | // Improve estimate for the vector width if it is scalable. |
4057 | unsigned EstimatedWidthA = A.Width.getKnownMinValue(); |
4058 | unsigned EstimatedWidthB = B.Width.getKnownMinValue(); |
4059 | if (std::optional<unsigned> VScale = CM.getVScaleForTuning()) { |
4060 | if (A.Width.isScalable()) |
4061 | EstimatedWidthA *= *VScale; |
4062 | if (B.Width.isScalable()) |
4063 | EstimatedWidthB *= *VScale; |
4064 | } |
4065 | |
4066 | // When optimizing for size choose whichever is smallest, which will be the |
4067 | // one with the smallest cost for the whole loop. On a tie pick the larger |
4068 | // vector width, on the assumption that throughput will be greater. |
4069 | if (CM.CostKind == TTI::TCK_CodeSize) |
4070 | return CostA < CostB || |
4071 | (CostA == CostB && EstimatedWidthA > EstimatedWidthB); |
4072 | |
4073 | // Assume vscale may be larger than 1 (or the value being tuned for), |
4074 | // so that scalable vectorization is slightly favorable over fixed-width |
4075 | // vectorization. |
4076 | bool PreferScalable = !TTI.preferFixedOverScalableIfEqualCost() && |
4077 | A.Width.isScalable() && !B.Width.isScalable(); |
4078 | |
4079 | auto CmpFn = [PreferScalable](const InstructionCost &LHS, |
4080 | const InstructionCost &RHS) { |
4081 | return PreferScalable ? LHS <= RHS : LHS < RHS; |
4082 | }; |
4083 | |
4084 | // To avoid the need for FP division: |
4085 | // (CostA / EstimatedWidthA) < (CostB / EstimatedWidthB) |
4086 | // <=> (CostA * EstimatedWidthB) < (CostB * EstimatedWidthA) |
4087 | if (!MaxTripCount) |
4088 | return CmpFn(CostA * EstimatedWidthB, CostB * EstimatedWidthA); |
4089 | |
4090 | auto GetCostForTC = [MaxTripCount, HasTail](unsigned VF, |
4091 | InstructionCost VectorCost, |
4092 | InstructionCost ScalarCost) { |
4093 | // If the trip count is a known (possibly small) constant, the trip count |
4094 | // will be rounded up to an integer number of iterations under |
4095 | // FoldTailByMasking. The total cost in that case will be |
4096 | // VecCost*ceil(TripCount/VF). When not folding the tail, the total |
4097 | // cost will be VecCost*floor(TC/VF) + ScalarCost*(TC%VF). There will be |
4098 | // some extra overheads, but for the purpose of comparing the costs of |
4099 | // different VFs we can use this to compare the total loop-body cost |
4100 | // expected after vectorization. |
4101 | if (HasTail) |
4102 | return VectorCost * (MaxTripCount / VF) + |
4103 | ScalarCost * (MaxTripCount % VF); |
4104 | return VectorCost * divideCeil(Numerator: MaxTripCount, Denominator: VF); |
4105 | }; |
4106 | |
4107 | auto RTCostA = GetCostForTC(EstimatedWidthA, CostA, A.ScalarCost); |
4108 | auto RTCostB = GetCostForTC(EstimatedWidthB, CostB, B.ScalarCost); |
4109 | return CmpFn(RTCostA, RTCostB); |
4110 | } |
4111 | |
4112 | bool LoopVectorizationPlanner::isMoreProfitable(const VectorizationFactor &A, |
4113 | const VectorizationFactor &B, |
4114 | bool HasTail) const { |
4115 | const unsigned MaxTripCount = PSE.getSmallConstantMaxTripCount(); |
4116 | return LoopVectorizationPlanner::isMoreProfitable(A, B, MaxTripCount, |
4117 | HasTail); |
4118 | } |
4119 | |
4120 | void LoopVectorizationPlanner::( |
4121 | OptimizationRemarkEmitter *ORE) { |
4122 | using RecipeVFPair = std::pair<VPRecipeBase *, ElementCount>; |
4123 | SmallVector<RecipeVFPair> InvalidCosts; |
4124 | for (const auto &Plan : VPlans) { |
4125 | for (ElementCount VF : Plan->vectorFactors()) { |
4126 | // The VPlan-based cost model is designed for computing vector cost. |
4127 | // Querying VPlan-based cost model with a scarlar VF will cause some |
4128 | // errors because we expect the VF is vector for most of the widen |
4129 | // recipes. |
4130 | if (VF.isScalar()) |
4131 | continue; |
4132 | |
4133 | VPCostContext CostCtx(CM.TTI, *CM.TLI, Legal->getWidestInductionType(), |
4134 | CM, CM.CostKind); |
4135 | precomputeCosts(Plan&: *Plan, VF, CostCtx); |
4136 | auto Iter = vp_depth_first_deep(G: Plan->getVectorLoopRegion()->getEntry()); |
4137 | for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(Range: Iter)) { |
4138 | for (auto &R : *VPBB) { |
4139 | if (!R.cost(VF, Ctx&: CostCtx).isValid()) |
4140 | InvalidCosts.emplace_back(Args: &R, Args&: VF); |
4141 | } |
4142 | } |
4143 | } |
4144 | } |
4145 | if (InvalidCosts.empty()) |
4146 | return; |
4147 | |
4148 | // Emit a report of VFs with invalid costs in the loop. |
4149 | |
4150 | // Group the remarks per recipe, keeping the recipe order from InvalidCosts. |
4151 | DenseMap<VPRecipeBase *, unsigned> Numbering; |
4152 | unsigned I = 0; |
4153 | for (auto &Pair : InvalidCosts) |
4154 | if (Numbering.try_emplace(Key: Pair.first, Args&: I).second) |
4155 | ++I; |
4156 | |
4157 | // Sort the list, first on recipe(number) then on VF. |
4158 | sort(C&: InvalidCosts, Comp: [&Numbering](RecipeVFPair &A, RecipeVFPair &B) { |
4159 | unsigned NA = Numbering[A.first]; |
4160 | unsigned NB = Numbering[B.first]; |
4161 | if (NA != NB) |
4162 | return NA < NB; |
4163 | return ElementCount::isKnownLT(LHS: A.second, RHS: B.second); |
4164 | }); |
4165 | |
4166 | // For a list of ordered recipe-VF pairs: |
4167 | // [(load, VF1), (load, VF2), (store, VF1)] |
4168 | // group the recipes together to emit separate remarks for: |
4169 | // load (VF1, VF2) |
4170 | // store (VF1) |
4171 | auto Tail = ArrayRef<RecipeVFPair>(InvalidCosts); |
4172 | auto Subset = ArrayRef<RecipeVFPair>(); |
4173 | do { |
4174 | if (Subset.empty()) |
4175 | Subset = Tail.take_front(N: 1); |
4176 | |
4177 | VPRecipeBase *R = Subset.front().first; |
4178 | |
4179 | unsigned Opcode = |
4180 | TypeSwitch<const VPRecipeBase *, unsigned>(R) |
4181 | .Case<VPHeaderPHIRecipe>( |
4182 | caseFn: [](const auto *R) { return Instruction::PHI; }) |
4183 | .Case<VPWidenSelectRecipe>( |
4184 | caseFn: [](const auto *R) { return Instruction::Select; }) |
4185 | .Case<VPWidenStoreRecipe>( |
4186 | caseFn: [](const auto *R) { return Instruction::Store; }) |
4187 | .Case<VPWidenLoadRecipe>( |
4188 | caseFn: [](const auto *R) { return Instruction::Load; }) |
4189 | .Case<VPWidenCallRecipe, VPWidenIntrinsicRecipe>( |
4190 | caseFn: [](const auto *R) { return Instruction::Call; }) |
4191 | .Case<VPInstruction, VPWidenRecipe, VPReplicateRecipe, |
4192 | VPWidenCastRecipe>( |
4193 | caseFn: [](const auto *R) { return R->getOpcode(); }) |
4194 | .Case<VPInterleaveRecipe>(caseFn: [](const VPInterleaveRecipe *R) { |
4195 | return R->getStoredValues().empty() ? Instruction::Load |
4196 | : Instruction::Store; |
4197 | }); |
4198 | |
4199 | // If the next recipe is different, or if there are no other pairs, |
4200 | // emit a remark for the collated subset. e.g. |
4201 | // [(load, VF1), (load, VF2))] |
4202 | // to emit: |
4203 | // remark: invalid costs for 'load' at VF=(VF1, VF2) |
4204 | if (Subset == Tail || Tail[Subset.size()].first != R) { |
4205 | std::string OutString; |
4206 | raw_string_ostream OS(OutString); |
4207 | assert(!Subset.empty() && "Unexpected empty range" ); |
4208 | OS << "Recipe with invalid costs prevented vectorization at VF=(" ; |
4209 | for (const auto &Pair : Subset) |
4210 | OS << (Pair.second == Subset.front().second ? "" : ", " ) << Pair.second; |
4211 | OS << "):" ; |
4212 | if (Opcode == Instruction::Call) { |
4213 | StringRef Name = "" ; |
4214 | if (auto *Int = dyn_cast<VPWidenIntrinsicRecipe>(Val: R)) { |
4215 | Name = Int->getIntrinsicName(); |
4216 | } else { |
4217 | auto *WidenCall = dyn_cast<VPWidenCallRecipe>(Val: R); |
4218 | Function *CalledFn = |
4219 | WidenCall ? WidenCall->getCalledScalarFunction() |
4220 | : cast<Function>(Val: R->getOperand(N: R->getNumOperands() - 1) |
4221 | ->getLiveInIRValue()); |
4222 | Name = CalledFn->getName(); |
4223 | } |
4224 | OS << " call to " << Name; |
4225 | } else |
4226 | OS << " " << Instruction::getOpcodeName(Opcode); |
4227 | reportVectorizationInfo(Msg: OutString, ORETag: "InvalidCost" , ORE, TheLoop: OrigLoop, I: nullptr, |
4228 | DL: R->getDebugLoc()); |
4229 | Tail = Tail.drop_front(N: Subset.size()); |
4230 | Subset = {}; |
4231 | } else |
4232 | // Grow the subset by one element |
4233 | Subset = Tail.take_front(N: Subset.size() + 1); |
4234 | } while (!Tail.empty()); |
4235 | } |
4236 | |
4237 | /// Check if any recipe of \p Plan will generate a vector value, which will be |
4238 | /// assigned a vector register. |
4239 | static bool willGenerateVectors(VPlan &Plan, ElementCount VF, |
4240 | const TargetTransformInfo &TTI) { |
4241 | assert(VF.isVector() && "Checking a scalar VF?" ); |
4242 | VPTypeAnalysis TypeInfo(Plan.getCanonicalIV()->getScalarType()); |
4243 | DenseSet<VPRecipeBase *> EphemeralRecipes; |
4244 | collectEphemeralRecipesForVPlan(Plan, EphRecipes&: EphemeralRecipes); |
4245 | // Set of already visited types. |
4246 | DenseSet<Type *> Visited; |
4247 | for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>( |
4248 | Range: vp_depth_first_shallow(G: Plan.getVectorLoopRegion()->getEntry()))) { |
4249 | for (VPRecipeBase &R : *VPBB) { |
4250 | if (EphemeralRecipes.contains(V: &R)) |
4251 | continue; |
4252 | // Continue early if the recipe is considered to not produce a vector |
4253 | // result. Note that this includes VPInstruction where some opcodes may |
4254 | // produce a vector, to preserve existing behavior as VPInstructions model |
4255 | // aspects not directly mapped to existing IR instructions. |
4256 | switch (R.getVPDefID()) { |
4257 | case VPDef::VPDerivedIVSC: |
4258 | case VPDef::VPScalarIVStepsSC: |
4259 | case VPDef::VPReplicateSC: |
4260 | case VPDef::VPInstructionSC: |
4261 | case VPDef::VPCanonicalIVPHISC: |
4262 | case VPDef::VPVectorPointerSC: |
4263 | case VPDef::VPVectorEndPointerSC: |
4264 | case VPDef::VPExpandSCEVSC: |
4265 | case VPDef::VPEVLBasedIVPHISC: |
4266 | case VPDef::VPPredInstPHISC: |
4267 | case VPDef::VPBranchOnMaskSC: |
4268 | continue; |
4269 | case VPDef::VPReductionSC: |
4270 | case VPDef::VPActiveLaneMaskPHISC: |
4271 | case VPDef::VPWidenCallSC: |
4272 | case VPDef::VPWidenCanonicalIVSC: |
4273 | case VPDef::VPWidenCastSC: |
4274 | case VPDef::VPWidenGEPSC: |
4275 | case VPDef::VPWidenIntrinsicSC: |
4276 | case VPDef::VPWidenSC: |
4277 | case VPDef::VPWidenSelectSC: |
4278 | case VPDef::VPBlendSC: |
4279 | case VPDef::VPFirstOrderRecurrencePHISC: |
4280 | case VPDef::VPHistogramSC: |
4281 | case VPDef::VPWidenPHISC: |
4282 | case VPDef::VPWidenIntOrFpInductionSC: |
4283 | case VPDef::VPWidenPointerInductionSC: |
4284 | case VPDef::VPReductionPHISC: |
4285 | case VPDef::VPInterleaveSC: |
4286 | case VPDef::VPWidenLoadEVLSC: |
4287 | case VPDef::VPWidenLoadSC: |
4288 | case VPDef::VPWidenStoreEVLSC: |
4289 | case VPDef::VPWidenStoreSC: |
4290 | break; |
4291 | default: |
4292 | llvm_unreachable("unhandled recipe" ); |
4293 | } |
4294 | |
4295 | auto WillGenerateTargetVectors = [&TTI, VF](Type *VectorTy) { |
4296 | unsigned NumLegalParts = TTI.getNumberOfParts(Tp: VectorTy); |
4297 | if (!NumLegalParts) |
4298 | return false; |
4299 | if (VF.isScalable()) { |
4300 | // <vscale x 1 x iN> is assumed to be profitable over iN because |
4301 | // scalable registers are a distinct register class from scalar |
4302 | // ones. If we ever find a target which wants to lower scalable |
4303 | // vectors back to scalars, we'll need to update this code to |
4304 | // explicitly ask TTI about the register class uses for each part. |
4305 | return NumLegalParts <= VF.getKnownMinValue(); |
4306 | } |
4307 | // Two or more elements that share a register - are vectorized. |
4308 | return NumLegalParts < VF.getFixedValue(); |
4309 | }; |
4310 | |
4311 | // If no def nor is a store, e.g., branches, continue - no value to check. |
4312 | if (R.getNumDefinedValues() == 0 && |
4313 | !isa<VPWidenStoreRecipe, VPWidenStoreEVLRecipe, VPInterleaveRecipe>( |
4314 | Val: &R)) |
4315 | continue; |
4316 | // For multi-def recipes, currently only interleaved loads, suffice to |
4317 | // check first def only. |
4318 | // For stores check their stored value; for interleaved stores suffice |
4319 | // the check first stored value only. In all cases this is the second |
4320 | // operand. |
4321 | VPValue *ToCheck = |
4322 | R.getNumDefinedValues() >= 1 ? R.getVPValue(I: 0) : R.getOperand(N: 1); |
4323 | Type *ScalarTy = TypeInfo.inferScalarType(V: ToCheck); |
4324 | if (!Visited.insert(V: {ScalarTy}).second) |
4325 | continue; |
4326 | Type *WideTy = toVectorizedTy(Ty: ScalarTy, EC: VF); |
4327 | if (any_of(Range: getContainedTypes(Ty: WideTy), P: WillGenerateTargetVectors)) |
4328 | return true; |
4329 | } |
4330 | } |
4331 | |
4332 | return false; |
4333 | } |
4334 | |
4335 | static bool hasReplicatorRegion(VPlan &Plan) { |
4336 | return any_of(Range: VPBlockUtils::blocksOnly<VPRegionBlock>(Range: vp_depth_first_shallow( |
4337 | G: Plan.getVectorLoopRegion()->getEntry())), |
4338 | P: [](auto *VPRB) { return VPRB->isReplicator(); }); |
4339 | } |
4340 | |
4341 | #ifndef NDEBUG |
4342 | VectorizationFactor LoopVectorizationPlanner::selectVectorizationFactor() { |
4343 | InstructionCost ExpectedCost = CM.expectedCost(ElementCount::getFixed(1)); |
4344 | LLVM_DEBUG(dbgs() << "LV: Scalar loop costs: " << ExpectedCost << ".\n" ); |
4345 | assert(ExpectedCost.isValid() && "Unexpected invalid cost for scalar loop" ); |
4346 | assert( |
4347 | any_of(VPlans, |
4348 | [](std::unique_ptr<VPlan> &P) { return P->hasScalarVFOnly(); }) && |
4349 | "Expected Scalar VF to be a candidate" ); |
4350 | |
4351 | const VectorizationFactor ScalarCost(ElementCount::getFixed(1), ExpectedCost, |
4352 | ExpectedCost); |
4353 | VectorizationFactor ChosenFactor = ScalarCost; |
4354 | |
4355 | bool ForceVectorization = Hints.getForce() == LoopVectorizeHints::FK_Enabled; |
4356 | if (ForceVectorization && |
4357 | (VPlans.size() > 1 || !VPlans[0]->hasScalarVFOnly())) { |
4358 | // Ignore scalar width, because the user explicitly wants vectorization. |
4359 | // Initialize cost to max so that VF = 2 is, at least, chosen during cost |
4360 | // evaluation. |
4361 | ChosenFactor.Cost = InstructionCost::getMax(); |
4362 | } |
4363 | |
4364 | for (auto &P : VPlans) { |
4365 | ArrayRef<ElementCount> VFs(P->vectorFactors().begin(), |
4366 | P->vectorFactors().end()); |
4367 | |
4368 | SmallVector<VPRegisterUsage, 8> RUs; |
4369 | if (CM.useMaxBandwidth(TargetTransformInfo::RGK_ScalableVector) || |
4370 | CM.useMaxBandwidth(TargetTransformInfo::RGK_FixedWidthVector)) |
4371 | RUs = calculateRegisterUsageForPlan(*P, VFs, TTI, CM.ValuesToIgnore); |
4372 | |
4373 | for (unsigned I = 0; I < VFs.size(); I++) { |
4374 | ElementCount VF = VFs[I]; |
4375 | // The cost for scalar VF=1 is already calculated, so ignore it. |
4376 | if (VF.isScalar()) |
4377 | continue; |
4378 | |
4379 | /// Don't consider the VF if it exceeds the number of registers for the |
4380 | /// target. |
4381 | if (CM.useMaxBandwidth(VF) && RUs[I].exceedsMaxNumRegs(TTI)) |
4382 | continue; |
4383 | |
4384 | InstructionCost C = CM.expectedCost(VF); |
4385 | |
4386 | // Add on other costs that are modelled in VPlan, but not in the legacy |
4387 | // cost model. |
4388 | VPCostContext CostCtx(CM.TTI, *CM.TLI, CM.Legal->getWidestInductionType(), |
4389 | CM, CM.CostKind); |
4390 | VPRegionBlock *VectorRegion = P->getVectorLoopRegion(); |
4391 | assert(VectorRegion && "Expected to have a vector region!" ); |
4392 | for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>( |
4393 | vp_depth_first_shallow(VectorRegion->getEntry()))) { |
4394 | for (VPRecipeBase &R : *VPBB) { |
4395 | auto *VPI = dyn_cast<VPInstruction>(&R); |
4396 | if (!VPI) |
4397 | continue; |
4398 | switch (VPI->getOpcode()) { |
4399 | case VPInstruction::ActiveLaneMask: |
4400 | case VPInstruction::ExplicitVectorLength: |
4401 | C += VPI->cost(VF, CostCtx); |
4402 | break; |
4403 | default: |
4404 | break; |
4405 | } |
4406 | } |
4407 | } |
4408 | |
4409 | VectorizationFactor Candidate(VF, C, ScalarCost.ScalarCost); |
4410 | unsigned Width = |
4411 | getEstimatedRuntimeVF(Candidate.Width, CM.getVScaleForTuning()); |
4412 | LLVM_DEBUG(dbgs() << "LV: Vector loop of width " << VF |
4413 | << " costs: " << (Candidate.Cost / Width)); |
4414 | if (VF.isScalable()) |
4415 | LLVM_DEBUG(dbgs() << " (assuming a minimum vscale of " |
4416 | << CM.getVScaleForTuning().value_or(1) << ")" ); |
4417 | LLVM_DEBUG(dbgs() << ".\n" ); |
4418 | |
4419 | if (!ForceVectorization && !willGenerateVectors(*P, VF, TTI)) { |
4420 | LLVM_DEBUG( |
4421 | dbgs() |
4422 | << "LV: Not considering vector loop of width " << VF |
4423 | << " because it will not generate any vector instructions.\n" ); |
4424 | continue; |
4425 | } |
4426 | |
4427 | if (CM.OptForSize && !ForceVectorization && hasReplicatorRegion(*P)) { |
4428 | LLVM_DEBUG( |
4429 | dbgs() |
4430 | << "LV: Not considering vector loop of width " << VF |
4431 | << " because it would cause replicated blocks to be generated," |
4432 | << " which isn't allowed when optimizing for size.\n" ); |
4433 | continue; |
4434 | } |
4435 | |
4436 | if (isMoreProfitable(Candidate, ChosenFactor, P->hasScalarTail())) |
4437 | ChosenFactor = Candidate; |
4438 | } |
4439 | } |
4440 | |
4441 | if (!EnableCondStoresVectorization && CM.hasPredStores()) { |
4442 | reportVectorizationFailure( |
4443 | "There are conditional stores." , |
4444 | "store that is conditionally executed prevents vectorization" , |
4445 | "ConditionalStore" , ORE, OrigLoop); |
4446 | ChosenFactor = ScalarCost; |
4447 | } |
4448 | |
4449 | LLVM_DEBUG(if (ForceVectorization && !ChosenFactor.Width.isScalar() && |
4450 | !isMoreProfitable(ChosenFactor, ScalarCost, |
4451 | !CM.foldTailByMasking())) dbgs() |
4452 | << "LV: Vectorization seems to be not beneficial, " |
4453 | << "but was forced by a user.\n" ); |
4454 | return ChosenFactor; |
4455 | } |
4456 | #endif |
4457 | |
4458 | bool LoopVectorizationPlanner::isCandidateForEpilogueVectorization( |
4459 | ElementCount VF) const { |
4460 | // Cross iteration phis such as reductions need special handling and are |
4461 | // currently unsupported. |
4462 | if (any_of(Range: OrigLoop->getHeader()->phis(), |
4463 | P: [&](PHINode &Phi) { return Legal->isFixedOrderRecurrence(Phi: &Phi); })) |
4464 | return false; |
4465 | |
4466 | // Phis with uses outside of the loop require special handling and are |
4467 | // currently unsupported. |
4468 | for (const auto &Entry : Legal->getInductionVars()) { |
4469 | // Look for uses of the value of the induction at the last iteration. |
4470 | Value *PostInc = |
4471 | Entry.first->getIncomingValueForBlock(BB: OrigLoop->getLoopLatch()); |
4472 | for (User *U : PostInc->users()) |
4473 | if (!OrigLoop->contains(Inst: cast<Instruction>(Val: U))) |
4474 | return false; |
4475 | // Look for uses of penultimate value of the induction. |
4476 | for (User *U : Entry.first->users()) |
4477 | if (!OrigLoop->contains(Inst: cast<Instruction>(Val: U))) |
4478 | return false; |
4479 | } |
4480 | |
4481 | // Epilogue vectorization code has not been auditted to ensure it handles |
4482 | // non-latch exits properly. It may be fine, but it needs auditted and |
4483 | // tested. |
4484 | // TODO: Add support for loops with an early exit. |
4485 | if (OrigLoop->getExitingBlock() != OrigLoop->getLoopLatch()) |
4486 | return false; |
4487 | |
4488 | return true; |
4489 | } |
4490 | |
4491 | bool LoopVectorizationCostModel::isEpilogueVectorizationProfitable( |
4492 | const ElementCount VF, const unsigned IC) const { |
4493 | // FIXME: We need a much better cost-model to take different parameters such |
4494 | // as register pressure, code size increase and cost of extra branches into |
4495 | // account. For now we apply a very crude heuristic and only consider loops |
4496 | // with vectorization factors larger than a certain value. |
4497 | |
4498 | // Allow the target to opt out entirely. |
4499 | if (!TTI.preferEpilogueVectorization()) |
4500 | return false; |
4501 | |
4502 | // We also consider epilogue vectorization unprofitable for targets that don't |
4503 | // consider interleaving beneficial (eg. MVE). |
4504 | if (TTI.getMaxInterleaveFactor(VF) <= 1) |
4505 | return false; |
4506 | |
4507 | // TODO: PR #108190 introduced a discrepancy between fixed-width and scalable |
4508 | // VFs when deciding profitability. |
4509 | // See related "TODO: extend to support scalable VFs." in |
4510 | // selectEpilogueVectorizationFactor. |
4511 | unsigned Multiplier = VF.isFixed() ? IC : 1; |
4512 | unsigned MinVFThreshold = EpilogueVectorizationMinVF.getNumOccurrences() > 0 |
4513 | ? EpilogueVectorizationMinVF |
4514 | : TTI.getEpilogueVectorizationMinVF(); |
4515 | return getEstimatedRuntimeVF(VF: VF * Multiplier, VScale: VScaleForTuning) >= |
4516 | MinVFThreshold; |
4517 | } |
4518 | |
4519 | VectorizationFactor LoopVectorizationPlanner::selectEpilogueVectorizationFactor( |
4520 | const ElementCount MainLoopVF, unsigned IC) { |
4521 | VectorizationFactor Result = VectorizationFactor::Disabled(); |
4522 | if (!EnableEpilogueVectorization) { |
4523 | LLVM_DEBUG(dbgs() << "LEV: Epilogue vectorization is disabled.\n" ); |
4524 | return Result; |
4525 | } |
4526 | |
4527 | if (!CM.isScalarEpilogueAllowed()) { |
4528 | LLVM_DEBUG(dbgs() << "LEV: Unable to vectorize epilogue because no " |
4529 | "epilogue is allowed.\n" ); |
4530 | return Result; |
4531 | } |
4532 | |
4533 | // Not really a cost consideration, but check for unsupported cases here to |
4534 | // simplify the logic. |
4535 | if (!isCandidateForEpilogueVectorization(VF: MainLoopVF)) { |
4536 | LLVM_DEBUG(dbgs() << "LEV: Unable to vectorize epilogue because the loop " |
4537 | "is not a supported candidate.\n" ); |
4538 | return Result; |
4539 | } |
4540 | |
4541 | if (EpilogueVectorizationForceVF > 1) { |
4542 | LLVM_DEBUG(dbgs() << "LEV: Epilogue vectorization factor is forced.\n" ); |
4543 | ElementCount ForcedEC = ElementCount::getFixed(MinVal: EpilogueVectorizationForceVF); |
4544 | if (hasPlanWithVF(VF: ForcedEC)) |
4545 | return {ForcedEC, 0, 0}; |
4546 | |
4547 | LLVM_DEBUG(dbgs() << "LEV: Epilogue vectorization forced factor is not " |
4548 | "viable.\n" ); |
4549 | return Result; |
4550 | } |
4551 | |
4552 | if (OrigLoop->getHeader()->getParent()->hasOptSize()) { |
4553 | LLVM_DEBUG( |
4554 | dbgs() << "LEV: Epilogue vectorization skipped due to opt for size.\n" ); |
4555 | return Result; |
4556 | } |
4557 | |
4558 | if (!CM.isEpilogueVectorizationProfitable(VF: MainLoopVF, IC)) { |
4559 | LLVM_DEBUG(dbgs() << "LEV: Epilogue vectorization is not profitable for " |
4560 | "this loop\n" ); |
4561 | return Result; |
4562 | } |
4563 | |
4564 | // If MainLoopVF = vscale x 2, and vscale is expected to be 4, then we know |
4565 | // the main loop handles 8 lanes per iteration. We could still benefit from |
4566 | // vectorizing the epilogue loop with VF=4. |
4567 | ElementCount EstimatedRuntimeVF = ElementCount::getFixed( |
4568 | MinVal: getEstimatedRuntimeVF(VF: MainLoopVF, VScale: CM.getVScaleForTuning())); |
4569 | |
4570 | ScalarEvolution &SE = *PSE.getSE(); |
4571 | Type *TCType = Legal->getWidestInductionType(); |
4572 | const SCEV *RemainingIterations = nullptr; |
4573 | unsigned MaxTripCount = 0; |
4574 | for (auto &NextVF : ProfitableVFs) { |
4575 | // Skip candidate VFs without a corresponding VPlan. |
4576 | if (!hasPlanWithVF(VF: NextVF.Width)) |
4577 | continue; |
4578 | |
4579 | // Skip candidate VFs with widths >= the (estimated) runtime VF (scalable |
4580 | // vectors) or > the VF of the main loop (fixed vectors). |
4581 | if ((!NextVF.Width.isScalable() && MainLoopVF.isScalable() && |
4582 | ElementCount::isKnownGE(LHS: NextVF.Width, RHS: EstimatedRuntimeVF)) || |
4583 | (NextVF.Width.isScalable() && |
4584 | ElementCount::isKnownGE(LHS: NextVF.Width, RHS: MainLoopVF)) || |
4585 | (!NextVF.Width.isScalable() && !MainLoopVF.isScalable() && |
4586 | ElementCount::isKnownGT(LHS: NextVF.Width, RHS: MainLoopVF))) |
4587 | continue; |
4588 | |
4589 | // If NextVF is greater than the number of remaining iterations, the |
4590 | // epilogue loop would be dead. Skip such factors. |
4591 | if (!MainLoopVF.isScalable() && !NextVF.Width.isScalable()) { |
4592 | // TODO: extend to support scalable VFs. |
4593 | if (!RemainingIterations) { |
4594 | const SCEV *TC = vputils::getSCEVExprForVPValue( |
4595 | V: getPlanFor(VF: NextVF.Width).getTripCount(), SE); |
4596 | assert(!isa<SCEVCouldNotCompute>(TC) && |
4597 | "Trip count SCEV must be computable" ); |
4598 | RemainingIterations = SE.getURemExpr( |
4599 | LHS: TC, RHS: SE.getConstant(Ty: TCType, V: MainLoopVF.getFixedValue() * IC)); |
4600 | MaxTripCount = MainLoopVF.getFixedValue() * IC - 1; |
4601 | if (SE.isKnownPredicate(Pred: CmpInst::ICMP_ULT, LHS: RemainingIterations, |
4602 | RHS: SE.getConstant(Ty: TCType, V: MaxTripCount))) { |
4603 | MaxTripCount = |
4604 | SE.getUnsignedRangeMax(S: RemainingIterations).getZExtValue(); |
4605 | } |
4606 | LLVM_DEBUG(dbgs() << "LEV: Maximum Trip Count for Epilogue: " |
4607 | << MaxTripCount << "\n" ); |
4608 | } |
4609 | if (SE.isKnownPredicate( |
4610 | Pred: CmpInst::ICMP_UGT, |
4611 | LHS: SE.getConstant(Ty: TCType, V: NextVF.Width.getFixedValue()), |
4612 | RHS: RemainingIterations)) |
4613 | continue; |
4614 | } |
4615 | |
4616 | if (Result.Width.isScalar() || |
4617 | isMoreProfitable(A: NextVF, B: Result, MaxTripCount, HasTail: !CM.foldTailByMasking())) |
4618 | Result = NextVF; |
4619 | } |
4620 | |
4621 | if (Result != VectorizationFactor::Disabled()) |
4622 | LLVM_DEBUG(dbgs() << "LEV: Vectorizing epilogue loop with VF = " |
4623 | << Result.Width << "\n" ); |
4624 | return Result; |
4625 | } |
4626 | |
4627 | std::pair<unsigned, unsigned> |
4628 | LoopVectorizationCostModel::getSmallestAndWidestTypes() { |
4629 | unsigned MinWidth = -1U; |
4630 | unsigned MaxWidth = 8; |
4631 | const DataLayout &DL = TheFunction->getDataLayout(); |
4632 | // For in-loop reductions, no element types are added to ElementTypesInLoop |
4633 | // if there are no loads/stores in the loop. In this case, check through the |
4634 | // reduction variables to determine the maximum width. |
4635 | if (ElementTypesInLoop.empty() && !Legal->getReductionVars().empty()) { |
4636 | for (const auto &PhiDescriptorPair : Legal->getReductionVars()) { |
4637 | const RecurrenceDescriptor &RdxDesc = PhiDescriptorPair.second; |
4638 | // When finding the min width used by the recurrence we need to account |
4639 | // for casts on the input operands of the recurrence. |
4640 | MinWidth = std::min( |
4641 | a: MinWidth, |
4642 | b: std::min(a: RdxDesc.getMinWidthCastToRecurrenceTypeInBits(), |
4643 | b: RdxDesc.getRecurrenceType()->getScalarSizeInBits())); |
4644 | MaxWidth = std::max(a: MaxWidth, |
4645 | b: RdxDesc.getRecurrenceType()->getScalarSizeInBits()); |
4646 | } |
4647 | } else { |
4648 | for (Type *T : ElementTypesInLoop) { |
4649 | MinWidth = std::min<unsigned>( |
4650 | a: MinWidth, b: DL.getTypeSizeInBits(Ty: T->getScalarType()).getFixedValue()); |
4651 | MaxWidth = std::max<unsigned>( |
4652 | a: MaxWidth, b: DL.getTypeSizeInBits(Ty: T->getScalarType()).getFixedValue()); |
4653 | } |
4654 | } |
4655 | return {MinWidth, MaxWidth}; |
4656 | } |
4657 | |
4658 | void LoopVectorizationCostModel::collectElementTypesForWidening() { |
4659 | ElementTypesInLoop.clear(); |
4660 | // For each block. |
4661 | for (BasicBlock *BB : TheLoop->blocks()) { |
4662 | // For each instruction in the loop. |
4663 | for (Instruction &I : BB->instructionsWithoutDebug()) { |
4664 | Type *T = I.getType(); |
4665 | |
4666 | // Skip ignored values. |
4667 | if (ValuesToIgnore.count(Ptr: &I)) |
4668 | continue; |
4669 | |
4670 | // Only examine Loads, Stores and PHINodes. |
4671 | if (!isa<LoadInst>(Val: I) && !isa<StoreInst>(Val: I) && !isa<PHINode>(Val: I)) |
4672 | continue; |
4673 | |
4674 | // Examine PHI nodes that are reduction variables. Update the type to |
4675 | // account for the recurrence type. |
4676 | if (auto *PN = dyn_cast<PHINode>(Val: &I)) { |
4677 | if (!Legal->isReductionVariable(PN)) |
4678 | continue; |
4679 | const RecurrenceDescriptor &RdxDesc = |
4680 | Legal->getReductionVars().find(Key: PN)->second; |
4681 | if (PreferInLoopReductions || useOrderedReductions(RdxDesc) || |
4682 | TTI.preferInLoopReduction(Kind: RdxDesc.getRecurrenceKind(), |
4683 | Ty: RdxDesc.getRecurrenceType())) |
4684 | continue; |
4685 | T = RdxDesc.getRecurrenceType(); |
4686 | } |
4687 | |
4688 | // Examine the stored values. |
4689 | if (auto *ST = dyn_cast<StoreInst>(Val: &I)) |
4690 | T = ST->getValueOperand()->getType(); |
4691 | |
4692 | assert(T->isSized() && |
4693 | "Expected the load/store/recurrence type to be sized" ); |
4694 | |
4695 | ElementTypesInLoop.insert(Ptr: T); |
4696 | } |
4697 | } |
4698 | } |
4699 | |
4700 | unsigned |
4701 | LoopVectorizationCostModel::selectInterleaveCount(VPlan &Plan, ElementCount VF, |
4702 | InstructionCost LoopCost) { |
4703 | // -- The interleave heuristics -- |
4704 | // We interleave the loop in order to expose ILP and reduce the loop overhead. |
4705 | // There are many micro-architectural considerations that we can't predict |
4706 | // at this level. For example, frontend pressure (on decode or fetch) due to |
4707 | // code size, or the number and capabilities of the execution ports. |
4708 | // |
4709 | // We use the following heuristics to select the interleave count: |
4710 | // 1. If the code has reductions, then we interleave to break the cross |
4711 | // iteration dependency. |
4712 | // 2. If the loop is really small, then we interleave to reduce the loop |
4713 | // overhead. |
4714 | // 3. We don't interleave if we think that we will spill registers to memory |
4715 | // due to the increased register pressure. |
4716 | |
4717 | if (!isScalarEpilogueAllowed()) |
4718 | return 1; |
4719 | |
4720 | // Do not interleave if EVL is preferred and no User IC is specified. |
4721 | if (foldTailWithEVL()) { |
4722 | LLVM_DEBUG(dbgs() << "LV: Preference for VP intrinsics indicated. " |
4723 | "Unroll factor forced to be 1.\n" ); |
4724 | return 1; |
4725 | } |
4726 | |
4727 | // We used the distance for the interleave count. |
4728 | if (!Legal->isSafeForAnyVectorWidth()) |
4729 | return 1; |
4730 | |
4731 | // We don't attempt to perform interleaving for loops with uncountable early |
4732 | // exits because the VPInstruction::AnyOf code cannot currently handle |
4733 | // multiple parts. |
4734 | if (Legal->hasUncountableEarlyExit()) |
4735 | return 1; |
4736 | |
4737 | const bool HasReductions = !Legal->getReductionVars().empty(); |
4738 | |
4739 | // If we did not calculate the cost for VF (because the user selected the VF) |
4740 | // then we calculate the cost of VF here. |
4741 | if (LoopCost == 0) { |
4742 | LoopCost = expectedCost(VF); |
4743 | assert(LoopCost.isValid() && "Expected to have chosen a VF with valid cost" ); |
4744 | |
4745 | // Loop body is free and there is no need for interleaving. |
4746 | if (LoopCost == 0) |
4747 | return 1; |
4748 | } |
4749 | |
4750 | VPRegisterUsage R = |
4751 | calculateRegisterUsageForPlan(Plan, VFs: {VF}, TTI, ValuesToIgnore)[0]; |
4752 | // We divide by these constants so assume that we have at least one |
4753 | // instruction that uses at least one register. |
4754 | for (auto &Pair : R.MaxLocalUsers) { |
4755 | Pair.second = std::max(a: Pair.second, b: 1U); |
4756 | } |
4757 | |
4758 | // We calculate the interleave count using the following formula. |
4759 | // Subtract the number of loop invariants from the number of available |
4760 | // registers. These registers are used by all of the interleaved instances. |
4761 | // Next, divide the remaining registers by the number of registers that is |
4762 | // required by the loop, in order to estimate how many parallel instances |
4763 | // fit without causing spills. All of this is rounded down if necessary to be |
4764 | // a power of two. We want power of two interleave count to simplify any |
4765 | // addressing operations or alignment considerations. |
4766 | // We also want power of two interleave counts to ensure that the induction |
4767 | // variable of the vector loop wraps to zero, when tail is folded by masking; |
4768 | // this currently happens when OptForSize, in which case IC is set to 1 above. |
4769 | unsigned IC = UINT_MAX; |
4770 | |
4771 | for (const auto &Pair : R.MaxLocalUsers) { |
4772 | unsigned TargetNumRegisters = TTI.getNumberOfRegisters(ClassID: Pair.first); |
4773 | LLVM_DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters |
4774 | << " registers of " |
4775 | << TTI.getRegisterClassName(Pair.first) |
4776 | << " register class\n" ); |
4777 | if (VF.isScalar()) { |
4778 | if (ForceTargetNumScalarRegs.getNumOccurrences() > 0) |
4779 | TargetNumRegisters = ForceTargetNumScalarRegs; |
4780 | } else { |
4781 | if (ForceTargetNumVectorRegs.getNumOccurrences() > 0) |
4782 | TargetNumRegisters = ForceTargetNumVectorRegs; |
4783 | } |
4784 | unsigned MaxLocalUsers = Pair.second; |
4785 | unsigned LoopInvariantRegs = 0; |
4786 | if (R.LoopInvariantRegs.contains(Key: Pair.first)) |
4787 | LoopInvariantRegs = R.LoopInvariantRegs[Pair.first]; |
4788 | |
4789 | unsigned TmpIC = llvm::bit_floor(Value: (TargetNumRegisters - LoopInvariantRegs) / |
4790 | MaxLocalUsers); |
4791 | // Don't count the induction variable as interleaved. |
4792 | if (EnableIndVarRegisterHeur) { |
4793 | TmpIC = llvm::bit_floor(Value: (TargetNumRegisters - LoopInvariantRegs - 1) / |
4794 | std::max(a: 1U, b: (MaxLocalUsers - 1))); |
4795 | } |
4796 | |
4797 | IC = std::min(a: IC, b: TmpIC); |
4798 | } |
4799 | |
4800 | // Clamp the interleave ranges to reasonable counts. |
4801 | unsigned MaxInterleaveCount = TTI.getMaxInterleaveFactor(VF); |
4802 | |
4803 | // Check if the user has overridden the max. |
4804 | if (VF.isScalar()) { |
4805 | if (ForceTargetMaxScalarInterleaveFactor.getNumOccurrences() > 0) |
4806 | MaxInterleaveCount = ForceTargetMaxScalarInterleaveFactor; |
4807 | } else { |
4808 | if (ForceTargetMaxVectorInterleaveFactor.getNumOccurrences() > 0) |
4809 | MaxInterleaveCount = ForceTargetMaxVectorInterleaveFactor; |
4810 | } |
4811 | |
4812 | unsigned EstimatedVF = getEstimatedRuntimeVF(VF, VScale: VScaleForTuning); |
4813 | |
4814 | // Try to get the exact trip count, or an estimate based on profiling data or |
4815 | // ConstantMax from PSE, failing that. |
4816 | if (auto BestKnownTC = getSmallBestKnownTC(PSE, L: TheLoop)) { |
4817 | // At least one iteration must be scalar when this constraint holds. So the |
4818 | // maximum available iterations for interleaving is one less. |
4819 | unsigned AvailableTC = requiresScalarEpilogue(IsVectorizing: VF.isVector()) |
4820 | ? BestKnownTC->getFixedValue() - 1 |
4821 | : BestKnownTC->getFixedValue(); |
4822 | |
4823 | unsigned InterleaveCountLB = bit_floor(Value: std::max( |
4824 | a: 1u, b: std::min(a: AvailableTC / (EstimatedVF * 2), b: MaxInterleaveCount))); |
4825 | |
4826 | if (getSmallConstantTripCount(SE: PSE.getSE(), L: TheLoop).isNonZero()) { |
4827 | // If the best known trip count is exact, we select between two |
4828 | // prospective ICs, where |
4829 | // |
4830 | // 1) the aggressive IC is capped by the trip count divided by VF |
4831 | // 2) the conservative IC is capped by the trip count divided by (VF * 2) |
4832 | // |
4833 | // The final IC is selected in a way that the epilogue loop trip count is |
4834 | // minimized while maximizing the IC itself, so that we either run the |
4835 | // vector loop at least once if it generates a small epilogue loop, or |
4836 | // else we run the vector loop at least twice. |
4837 | |
4838 | unsigned InterleaveCountUB = bit_floor(Value: std::max( |
4839 | a: 1u, b: std::min(a: AvailableTC / EstimatedVF, b: MaxInterleaveCount))); |
4840 | MaxInterleaveCount = InterleaveCountLB; |
4841 | |
4842 | if (InterleaveCountUB != InterleaveCountLB) { |
4843 | unsigned TailTripCountUB = |
4844 | (AvailableTC % (EstimatedVF * InterleaveCountUB)); |
4845 | unsigned TailTripCountLB = |
4846 | (AvailableTC % (EstimatedVF * InterleaveCountLB)); |
4847 | // If both produce same scalar tail, maximize the IC to do the same work |
4848 | // in fewer vector loop iterations |
4849 | if (TailTripCountUB == TailTripCountLB) |
4850 | MaxInterleaveCount = InterleaveCountUB; |
4851 | } |
4852 | } else { |
4853 | // If trip count is an estimated compile time constant, limit the |
4854 | // IC to be capped by the trip count divided by VF * 2, such that the |
4855 | // vector loop runs at least twice to make interleaving seem profitable |
4856 | // when there is an epilogue loop present. Since exact Trip count is not |
4857 | // known we choose to be conservative in our IC estimate. |
4858 | MaxInterleaveCount = InterleaveCountLB; |
4859 | } |
4860 | } |
4861 | |
4862 | assert(MaxInterleaveCount > 0 && |
4863 | "Maximum interleave count must be greater than 0" ); |
4864 | |
4865 | // Clamp the calculated IC to be between the 1 and the max interleave count |
4866 | // that the target and trip count allows. |
4867 | if (IC > MaxInterleaveCount) |
4868 | IC = MaxInterleaveCount; |
4869 | else |
4870 | // Make sure IC is greater than 0. |
4871 | IC = std::max(a: 1u, b: IC); |
4872 | |
4873 | assert(IC > 0 && "Interleave count must be greater than 0." ); |
4874 | |
4875 | // Interleave if we vectorized this loop and there is a reduction that could |
4876 | // benefit from interleaving. |
4877 | if (VF.isVector() && HasReductions) { |
4878 | LLVM_DEBUG(dbgs() << "LV: Interleaving because of reductions.\n" ); |
4879 | return IC; |
4880 | } |
4881 | |
4882 | // For any scalar loop that either requires runtime checks or predication we |
4883 | // are better off leaving this to the unroller. Note that if we've already |
4884 | // vectorized the loop we will have done the runtime check and so interleaving |
4885 | // won't require further checks. |
4886 | bool ScalarInterleavingRequiresPredication = |
4887 | (VF.isScalar() && any_of(Range: TheLoop->blocks(), P: [this](BasicBlock *BB) { |
4888 | return Legal->blockNeedsPredication(BB); |
4889 | })); |
4890 | bool ScalarInterleavingRequiresRuntimePointerCheck = |
4891 | (VF.isScalar() && Legal->getRuntimePointerChecking()->Need); |
4892 | |
4893 | // We want to interleave small loops in order to reduce the loop overhead and |
4894 | // potentially expose ILP opportunities. |
4895 | LLVM_DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n' |
4896 | << "LV: IC is " << IC << '\n' |
4897 | << "LV: VF is " << VF << '\n'); |
4898 | const bool AggressivelyInterleaveReductions = |
4899 | TTI.enableAggressiveInterleaving(LoopHasReductions: HasReductions); |
4900 | if (!ScalarInterleavingRequiresRuntimePointerCheck && |
4901 | !ScalarInterleavingRequiresPredication && LoopCost < SmallLoopCost) { |
4902 | // We assume that the cost overhead is 1 and we use the cost model |
4903 | // to estimate the cost of the loop and interleave until the cost of the |
4904 | // loop overhead is about 5% of the cost of the loop. |
4905 | unsigned SmallIC = std::min(a: IC, b: (unsigned)llvm::bit_floor<uint64_t>( |
4906 | Value: SmallLoopCost / LoopCost.getValue())); |
4907 | |
4908 | // Interleave until store/load ports (estimated by max interleave count) are |
4909 | // saturated. |
4910 | unsigned NumStores = Legal->getNumStores(); |
4911 | unsigned NumLoads = Legal->getNumLoads(); |
4912 | unsigned StoresIC = IC / (NumStores ? NumStores : 1); |
4913 | unsigned LoadsIC = IC / (NumLoads ? NumLoads : 1); |
4914 | |
4915 | // There is little point in interleaving for reductions containing selects |
4916 | // and compares when VF=1 since it may just create more overhead than it's |
4917 | // worth for loops with small trip counts. This is because we still have to |
4918 | // do the final reduction after the loop. |
4919 | bool HasSelectCmpReductions = |
4920 | HasReductions && |
4921 | any_of(Range: Legal->getReductionVars(), P: [&](auto &Reduction) -> bool { |
4922 | const RecurrenceDescriptor &RdxDesc = Reduction.second; |
4923 | RecurKind RK = RdxDesc.getRecurrenceKind(); |
4924 | return RecurrenceDescriptor::isAnyOfRecurrenceKind(Kind: RK) || |
4925 | RecurrenceDescriptor::isFindIVRecurrenceKind(Kind: RK); |
4926 | }); |
4927 | if (HasSelectCmpReductions) { |
4928 | LLVM_DEBUG(dbgs() << "LV: Not interleaving select-cmp reductions.\n" ); |
4929 | return 1; |
4930 | } |
4931 | |
4932 | // If we have a scalar reduction (vector reductions are already dealt with |
4933 | // by this point), we can increase the critical path length if the loop |
4934 | // we're interleaving is inside another loop. For tree-wise reductions |
4935 | // set the limit to 2, and for ordered reductions it's best to disable |
4936 | // interleaving entirely. |
4937 | if (HasReductions && TheLoop->getLoopDepth() > 1) { |
4938 | bool HasOrderedReductions = |
4939 | any_of(Range: Legal->getReductionVars(), P: [&](auto &Reduction) -> bool { |
4940 | const RecurrenceDescriptor &RdxDesc = Reduction.second; |
4941 | return RdxDesc.isOrdered(); |
4942 | }); |
4943 | if (HasOrderedReductions) { |
4944 | LLVM_DEBUG( |
4945 | dbgs() << "LV: Not interleaving scalar ordered reductions.\n" ); |
4946 | return 1; |
4947 | } |
4948 | |
4949 | unsigned F = MaxNestedScalarReductionIC; |
4950 | SmallIC = std::min(a: SmallIC, b: F); |
4951 | StoresIC = std::min(a: StoresIC, b: F); |
4952 | LoadsIC = std::min(a: LoadsIC, b: F); |
4953 | } |
4954 | |
4955 | if (EnableLoadStoreRuntimeInterleave && |
4956 | std::max(a: StoresIC, b: LoadsIC) > SmallIC) { |
4957 | LLVM_DEBUG( |
4958 | dbgs() << "LV: Interleaving to saturate store or load ports.\n" ); |
4959 | return std::max(a: StoresIC, b: LoadsIC); |
4960 | } |
4961 | |
4962 | // If there are scalar reductions and TTI has enabled aggressive |
4963 | // interleaving for reductions, we will interleave to expose ILP. |
4964 | if (VF.isScalar() && AggressivelyInterleaveReductions) { |
4965 | LLVM_DEBUG(dbgs() << "LV: Interleaving to expose ILP.\n" ); |
4966 | // Interleave no less than SmallIC but not as aggressive as the normal IC |
4967 | // to satisfy the rare situation when resources are too limited. |
4968 | return std::max(a: IC / 2, b: SmallIC); |
4969 | } |
4970 | |
4971 | LLVM_DEBUG(dbgs() << "LV: Interleaving to reduce branch cost.\n" ); |
4972 | return SmallIC; |
4973 | } |
4974 | |
4975 | // Interleave if this is a large loop (small loops are already dealt with by |
4976 | // this point) that could benefit from interleaving. |
4977 | if (AggressivelyInterleaveReductions) { |
4978 | LLVM_DEBUG(dbgs() << "LV: Interleaving to expose ILP.\n" ); |
4979 | return IC; |
4980 | } |
4981 | |
4982 | LLVM_DEBUG(dbgs() << "LV: Not Interleaving.\n" ); |
4983 | return 1; |
4984 | } |
4985 | |
4986 | bool LoopVectorizationCostModel::useEmulatedMaskMemRefHack(Instruction *I, |
4987 | ElementCount VF) { |
4988 | // TODO: Cost model for emulated masked load/store is completely |
4989 | // broken. This hack guides the cost model to use an artificially |
4990 | // high enough value to practically disable vectorization with such |
4991 | // operations, except where previously deployed legality hack allowed |
4992 | // using very low cost values. This is to avoid regressions coming simply |
4993 | // from moving "masked load/store" check from legality to cost model. |
4994 | // Masked Load/Gather emulation was previously never allowed. |
4995 | // Limited number of Masked Store/Scatter emulation was allowed. |
4996 | assert((isPredicatedInst(I)) && |
4997 | "Expecting a scalar emulated instruction" ); |
4998 | return isa<LoadInst>(Val: I) || |
4999 | (isa<StoreInst>(Val: I) && |
5000 | NumPredStores > NumberOfStoresToPredicate); |
5001 | } |
5002 | |
5003 | void LoopVectorizationCostModel::collectInstsToScalarize(ElementCount VF) { |
5004 | assert(VF.isVector() && "Expected VF >= 2" ); |
5005 | |
5006 | // If we've already collected the instructions to scalarize or the predicated |
5007 | // BBs after vectorization, there's nothing to do. Collection may already have |
5008 | // occurred if we have a user-selected VF and are now computing the expected |
5009 | // cost for interleaving. |
5010 | if (InstsToScalarize.contains(Val: VF) || |
5011 | PredicatedBBsAfterVectorization.contains(Val: VF)) |
5012 | return; |
5013 | |
5014 | // Initialize a mapping for VF in InstsToScalalarize. If we find that it's |
5015 | // not profitable to scalarize any instructions, the presence of VF in the |
5016 | // map will indicate that we've analyzed it already. |
5017 | ScalarCostsTy &ScalarCostsVF = InstsToScalarize[VF]; |
5018 | |
5019 | // Find all the instructions that are scalar with predication in the loop and |
5020 | // determine if it would be better to not if-convert the blocks they are in. |
5021 | // If so, we also record the instructions to scalarize. |
5022 | for (BasicBlock *BB : TheLoop->blocks()) { |
5023 | if (!blockNeedsPredicationForAnyReason(BB)) |
5024 | continue; |
5025 | for (Instruction &I : *BB) |
5026 | if (isScalarWithPredication(I: &I, VF)) { |
5027 | ScalarCostsTy ScalarCosts; |
5028 | // Do not apply discount logic for: |
5029 | // 1. Scalars after vectorization, as there will only be a single copy |
5030 | // of the instruction. |
5031 | // 2. Scalable VF, as that would lead to invalid scalarization costs. |
5032 | // 3. Emulated masked memrefs, if a hacked cost is needed. |
5033 | if (!isScalarAfterVectorization(I: &I, VF) && !VF.isScalable() && |
5034 | !useEmulatedMaskMemRefHack(I: &I, VF) && |
5035 | computePredInstDiscount(PredInst: &I, ScalarCosts, VF) >= 0) { |
5036 | ScalarCostsVF.insert_range(R&: ScalarCosts); |
5037 | // Check if we decided to scalarize a call. If so, update the widening |
5038 | // decision of the call to CM_Scalarize with the computed scalar cost. |
5039 | for (const auto &[I, Cost] : ScalarCosts) { |
5040 | auto *CI = dyn_cast<CallInst>(Val: I); |
5041 | if (!CI || !CallWideningDecisions.contains(Val: {CI, VF})) |
5042 | continue; |
5043 | CallWideningDecisions[{CI, VF}].Kind = CM_Scalarize; |
5044 | CallWideningDecisions[{CI, VF}].Cost = Cost; |
5045 | } |
5046 | } |
5047 | // Remember that BB will remain after vectorization. |
5048 | PredicatedBBsAfterVectorization[VF].insert(Ptr: BB); |
5049 | for (auto *Pred : predecessors(BB)) { |
5050 | if (Pred->getSingleSuccessor() == BB) |
5051 | PredicatedBBsAfterVectorization[VF].insert(Ptr: Pred); |
5052 | } |
5053 | } |
5054 | } |
5055 | } |
5056 | |
5057 | InstructionCost LoopVectorizationCostModel::computePredInstDiscount( |
5058 | Instruction *PredInst, ScalarCostsTy &ScalarCosts, ElementCount VF) { |
5059 | assert(!isUniformAfterVectorization(PredInst, VF) && |
5060 | "Instruction marked uniform-after-vectorization will be predicated" ); |
5061 | |
5062 | // Initialize the discount to zero, meaning that the scalar version and the |
5063 | // vector version cost the same. |
5064 | InstructionCost Discount = 0; |
5065 | |
5066 | // Holds instructions to analyze. The instructions we visit are mapped in |
5067 | // ScalarCosts. Those instructions are the ones that would be scalarized if |
5068 | // we find that the scalar version costs less. |
5069 | SmallVector<Instruction *, 8> Worklist; |
5070 | |
5071 | // Returns true if the given instruction can be scalarized. |
5072 | auto CanBeScalarized = [&](Instruction *I) -> bool { |
5073 | // We only attempt to scalarize instructions forming a single-use chain |
5074 | // from the original predicated block that would otherwise be vectorized. |
5075 | // Although not strictly necessary, we give up on instructions we know will |
5076 | // already be scalar to avoid traversing chains that are unlikely to be |
5077 | // beneficial. |
5078 | if (!I->hasOneUse() || PredInst->getParent() != I->getParent() || |
5079 | isScalarAfterVectorization(I, VF)) |
5080 | return false; |
5081 | |
5082 | // If the instruction is scalar with predication, it will be analyzed |
5083 | // separately. We ignore it within the context of PredInst. |
5084 | if (isScalarWithPredication(I, VF)) |
5085 | return false; |
5086 | |
5087 | // If any of the instruction's operands are uniform after vectorization, |
5088 | // the instruction cannot be scalarized. This prevents, for example, a |
5089 | // masked load from being scalarized. |
5090 | // |
5091 | // We assume we will only emit a value for lane zero of an instruction |
5092 | // marked uniform after vectorization, rather than VF identical values. |
5093 | // Thus, if we scalarize an instruction that uses a uniform, we would |
5094 | // create uses of values corresponding to the lanes we aren't emitting code |
5095 | // for. This behavior can be changed by allowing getScalarValue to clone |
5096 | // the lane zero values for uniforms rather than asserting. |
5097 | for (Use &U : I->operands()) |
5098 | if (auto *J = dyn_cast<Instruction>(Val: U.get())) |
5099 | if (isUniformAfterVectorization(I: J, VF)) |
5100 | return false; |
5101 | |
5102 | // Otherwise, we can scalarize the instruction. |
5103 | return true; |
5104 | }; |
5105 | |
5106 | // Compute the expected cost discount from scalarizing the entire expression |
5107 | // feeding the predicated instruction. We currently only consider expressions |
5108 | // that are single-use instruction chains. |
5109 | Worklist.push_back(Elt: PredInst); |
5110 | while (!Worklist.empty()) { |
5111 | Instruction *I = Worklist.pop_back_val(); |
5112 | |
5113 | // If we've already analyzed the instruction, there's nothing to do. |
5114 | if (ScalarCosts.contains(Val: I)) |
5115 | continue; |
5116 | |
5117 | // Cannot scalarize fixed-order recurrence phis at the moment. |
5118 | if (isa<PHINode>(Val: I) && Legal->isFixedOrderRecurrence(Phi: cast<PHINode>(Val: I))) |
5119 | continue; |
5120 | |
5121 | // Compute the cost of the vector instruction. Note that this cost already |
5122 | // includes the scalarization overhead of the predicated instruction. |
5123 | InstructionCost VectorCost = getInstructionCost(I, VF); |
5124 | |
5125 | // Compute the cost of the scalarized instruction. This cost is the cost of |
5126 | // the instruction as if it wasn't if-converted and instead remained in the |
5127 | // predicated block. We will scale this cost by block probability after |
5128 | // computing the scalarization overhead. |
5129 | InstructionCost ScalarCost = |
5130 | VF.getFixedValue() * getInstructionCost(I, VF: ElementCount::getFixed(MinVal: 1)); |
5131 | |
5132 | // Compute the scalarization overhead of needed insertelement instructions |
5133 | // and phi nodes. |
5134 | if (isScalarWithPredication(I, VF) && !I->getType()->isVoidTy()) { |
5135 | Type *WideTy = toVectorizedTy(Ty: I->getType(), EC: VF); |
5136 | for (Type *VectorTy : getContainedTypes(Ty: WideTy)) { |
5137 | ScalarCost += TTI.getScalarizationOverhead( |
5138 | Ty: cast<VectorType>(Val: VectorTy), DemandedElts: APInt::getAllOnes(numBits: VF.getFixedValue()), |
5139 | /*Insert=*/true, |
5140 | /*Extract=*/false, CostKind); |
5141 | } |
5142 | ScalarCost += |
5143 | VF.getFixedValue() * TTI.getCFInstrCost(Opcode: Instruction::PHI, CostKind); |
5144 | } |
5145 | |
5146 | // Compute the scalarization overhead of needed extractelement |
5147 | // instructions. For each of the instruction's operands, if the operand can |
5148 | // be scalarized, add it to the worklist; otherwise, account for the |
5149 | // overhead. |
5150 | for (Use &U : I->operands()) |
5151 | if (auto *J = dyn_cast<Instruction>(Val: U.get())) { |
5152 | assert(canVectorizeTy(J->getType()) && |
5153 | "Instruction has non-scalar type" ); |
5154 | if (CanBeScalarized(J)) |
5155 | Worklist.push_back(Elt: J); |
5156 | else if (needsExtract(V: J, VF)) { |
5157 | Type *WideTy = toVectorizedTy(Ty: J->getType(), EC: VF); |
5158 | for (Type *VectorTy : getContainedTypes(Ty: WideTy)) { |
5159 | ScalarCost += TTI.getScalarizationOverhead( |
5160 | Ty: cast<VectorType>(Val: VectorTy), |
5161 | DemandedElts: APInt::getAllOnes(numBits: VF.getFixedValue()), /*Insert*/ false, |
5162 | /*Extract*/ true, CostKind); |
5163 | } |
5164 | } |
5165 | } |
5166 | |
5167 | // Scale the total scalar cost by block probability. |
5168 | ScalarCost /= getPredBlockCostDivisor(CostKind); |
5169 | |
5170 | // Compute the discount. A non-negative discount means the vector version |
5171 | // of the instruction costs more, and scalarizing would be beneficial. |
5172 | Discount += VectorCost - ScalarCost; |
5173 | ScalarCosts[I] = ScalarCost; |
5174 | } |
5175 | |
5176 | return Discount; |
5177 | } |
5178 | |
5179 | InstructionCost LoopVectorizationCostModel::expectedCost(ElementCount VF) { |
5180 | InstructionCost Cost; |
5181 | |
5182 | // If the vector loop gets executed exactly once with the given VF, ignore the |
5183 | // costs of comparison and induction instructions, as they'll get simplified |
5184 | // away. |
5185 | SmallPtrSet<Instruction *, 2> ValuesToIgnoreForVF; |
5186 | auto TC = getSmallConstantTripCount(SE: PSE.getSE(), L: TheLoop); |
5187 | if (TC == VF && !foldTailByMasking()) |
5188 | addFullyUnrolledInstructionsToIgnore(L: TheLoop, IL: Legal->getInductionVars(), |
5189 | InstsToIgnore&: ValuesToIgnoreForVF); |
5190 | |
5191 | // For each block. |
5192 | for (BasicBlock *BB : TheLoop->blocks()) { |
5193 | InstructionCost BlockCost; |
5194 | |
5195 | // For each instruction in the old loop. |
5196 | for (Instruction &I : BB->instructionsWithoutDebug()) { |
5197 | // Skip ignored values. |
5198 | if (ValuesToIgnore.count(Ptr: &I) || ValuesToIgnoreForVF.count(Ptr: &I) || |
5199 | (VF.isVector() && VecValuesToIgnore.count(Ptr: &I))) |
5200 | continue; |
5201 | |
5202 | InstructionCost C = getInstructionCost(I: &I, VF); |
5203 | |
5204 | // Check if we should override the cost. |
5205 | if (C.isValid() && ForceTargetInstructionCost.getNumOccurrences() > 0) |
5206 | C = InstructionCost(ForceTargetInstructionCost); |
5207 | |
5208 | BlockCost += C; |
5209 | LLVM_DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " |
5210 | << VF << " For instruction: " << I << '\n'); |
5211 | } |
5212 | |
5213 | // If we are vectorizing a predicated block, it will have been |
5214 | // if-converted. This means that the block's instructions (aside from |
5215 | // stores and instructions that may divide by zero) will now be |
5216 | // unconditionally executed. For the scalar case, we may not always execute |
5217 | // the predicated block, if it is an if-else block. Thus, scale the block's |
5218 | // cost by the probability of executing it. blockNeedsPredication from |
5219 | // Legal is used so as to not include all blocks in tail folded loops. |
5220 | if (VF.isScalar() && Legal->blockNeedsPredication(BB)) |
5221 | BlockCost /= getPredBlockCostDivisor(CostKind); |
5222 | |
5223 | Cost += BlockCost; |
5224 | } |
5225 | |
5226 | return Cost; |
5227 | } |
5228 | |
5229 | /// Gets Address Access SCEV after verifying that the access pattern |
5230 | /// is loop invariant except the induction variable dependence. |
5231 | /// |
5232 | /// This SCEV can be sent to the Target in order to estimate the address |
5233 | /// calculation cost. |
5234 | static const SCEV *getAddressAccessSCEV( |
5235 | Value *Ptr, |
5236 | LoopVectorizationLegality *Legal, |
5237 | PredicatedScalarEvolution &PSE, |
5238 | const Loop *TheLoop) { |
5239 | |
5240 | auto *Gep = dyn_cast<GetElementPtrInst>(Val: Ptr); |
5241 | if (!Gep) |
5242 | return nullptr; |
5243 | |
5244 | // We are looking for a gep with all loop invariant indices except for one |
5245 | // which should be an induction variable. |
5246 | auto *SE = PSE.getSE(); |
5247 | unsigned NumOperands = Gep->getNumOperands(); |
5248 | for (unsigned Idx = 1; Idx < NumOperands; ++Idx) { |
5249 | Value *Opd = Gep->getOperand(i_nocapture: Idx); |
5250 | if (!SE->isLoopInvariant(S: SE->getSCEV(V: Opd), L: TheLoop) && |
5251 | !Legal->isInductionVariable(V: Opd)) |
5252 | return nullptr; |
5253 | } |
5254 | |
5255 | // Now we know we have a GEP ptr, %inv, %ind, %inv. return the Ptr SCEV. |
5256 | return PSE.getSCEV(V: Ptr); |
5257 | } |
5258 | |
5259 | InstructionCost |
5260 | LoopVectorizationCostModel::getMemInstScalarizationCost(Instruction *I, |
5261 | ElementCount VF) { |
5262 | assert(VF.isVector() && |
5263 | "Scalarization cost of instruction implies vectorization." ); |
5264 | if (VF.isScalable()) |
5265 | return InstructionCost::getInvalid(); |
5266 | |
5267 | Type *ValTy = getLoadStoreType(I); |
5268 | auto *SE = PSE.getSE(); |
5269 | |
5270 | unsigned AS = getLoadStoreAddressSpace(I); |
5271 | Value *Ptr = getLoadStorePointerOperand(V: I); |
5272 | Type *PtrTy = toVectorTy(Scalar: Ptr->getType(), EC: VF); |
5273 | // NOTE: PtrTy is a vector to signal `TTI::getAddressComputationCost` |
5274 | // that it is being called from this specific place. |
5275 | |
5276 | // Figure out whether the access is strided and get the stride value |
5277 | // if it's known in compile time |
5278 | const SCEV *PtrSCEV = getAddressAccessSCEV(Ptr, Legal, PSE, TheLoop); |
5279 | |
5280 | // Get the cost of the scalar memory instruction and address computation. |
5281 | InstructionCost Cost = |
5282 | VF.getFixedValue() * TTI.getAddressComputationCost(Ty: PtrTy, SE, Ptr: PtrSCEV); |
5283 | |
5284 | // Don't pass *I here, since it is scalar but will actually be part of a |
5285 | // vectorized loop where the user of it is a vectorized instruction. |
5286 | const Align Alignment = getLoadStoreAlignment(I); |
5287 | Cost += VF.getFixedValue() * TTI.getMemoryOpCost(Opcode: I->getOpcode(), |
5288 | Src: ValTy->getScalarType(), |
5289 | Alignment, AddressSpace: AS, CostKind); |
5290 | |
5291 | // Get the overhead of the extractelement and insertelement instructions |
5292 | // we might create due to scalarization. |
5293 | Cost += getScalarizationOverhead(I, VF); |
5294 | |
5295 | // If we have a predicated load/store, it will need extra i1 extracts and |
5296 | // conditional branches, but may not be executed for each vector lane. Scale |
5297 | // the cost by the probability of executing the predicated block. |
5298 | if (isPredicatedInst(I)) { |
5299 | Cost /= getPredBlockCostDivisor(CostKind); |
5300 | |
5301 | // Add the cost of an i1 extract and a branch |
5302 | auto *VecI1Ty = |
5303 | VectorType::get(ElementType: IntegerType::getInt1Ty(C&: ValTy->getContext()), EC: VF); |
5304 | Cost += TTI.getScalarizationOverhead( |
5305 | Ty: VecI1Ty, DemandedElts: APInt::getAllOnes(numBits: VF.getFixedValue()), |
5306 | /*Insert=*/false, /*Extract=*/true, CostKind); |
5307 | Cost += TTI.getCFInstrCost(Opcode: Instruction::Br, CostKind); |
5308 | |
5309 | if (useEmulatedMaskMemRefHack(I, VF)) |
5310 | // Artificially setting to a high enough value to practically disable |
5311 | // vectorization with such operations. |
5312 | Cost = 3000000; |
5313 | } |
5314 | |
5315 | return Cost; |
5316 | } |
5317 | |
5318 | InstructionCost |
5319 | LoopVectorizationCostModel::getConsecutiveMemOpCost(Instruction *I, |
5320 | ElementCount VF) { |
5321 | Type *ValTy = getLoadStoreType(I); |
5322 | auto *VectorTy = cast<VectorType>(Val: toVectorTy(Scalar: ValTy, EC: VF)); |
5323 | Value *Ptr = getLoadStorePointerOperand(V: I); |
5324 | unsigned AS = getLoadStoreAddressSpace(I); |
5325 | int ConsecutiveStride = Legal->isConsecutivePtr(AccessTy: ValTy, Ptr); |
5326 | |
5327 | assert((ConsecutiveStride == 1 || ConsecutiveStride == -1) && |
5328 | "Stride should be 1 or -1 for consecutive memory access" ); |
5329 | const Align Alignment = getLoadStoreAlignment(I); |
5330 | InstructionCost Cost = 0; |
5331 | if (Legal->isMaskRequired(I)) { |
5332 | Cost += TTI.getMaskedMemoryOpCost(Opcode: I->getOpcode(), Src: VectorTy, Alignment, AddressSpace: AS, |
5333 | CostKind); |
5334 | } else { |
5335 | TTI::OperandValueInfo OpInfo = TTI::getOperandInfo(V: I->getOperand(i: 0)); |
5336 | Cost += TTI.getMemoryOpCost(Opcode: I->getOpcode(), Src: VectorTy, Alignment, AddressSpace: AS, |
5337 | CostKind, OpdInfo: OpInfo, I); |
5338 | } |
5339 | |
5340 | bool Reverse = ConsecutiveStride < 0; |
5341 | if (Reverse) |
5342 | Cost += TTI.getShuffleCost(Kind: TargetTransformInfo::SK_Reverse, DstTy: VectorTy, |
5343 | SrcTy: VectorTy, Mask: {}, CostKind, Index: 0); |
5344 | return Cost; |
5345 | } |
5346 | |
5347 | InstructionCost |
5348 | LoopVectorizationCostModel::getUniformMemOpCost(Instruction *I, |
5349 | ElementCount VF) { |
5350 | assert(Legal->isUniformMemOp(*I, VF)); |
5351 | |
5352 | Type *ValTy = getLoadStoreType(I); |
5353 | auto *VectorTy = cast<VectorType>(Val: toVectorTy(Scalar: ValTy, EC: VF)); |
5354 | const Align Alignment = getLoadStoreAlignment(I); |
5355 | unsigned AS = getLoadStoreAddressSpace(I); |
5356 | if (isa<LoadInst>(Val: I)) { |
5357 | return TTI.getAddressComputationCost(Ty: ValTy) + |
5358 | TTI.getMemoryOpCost(Opcode: Instruction::Load, Src: ValTy, Alignment, AddressSpace: AS, |
5359 | CostKind) + |
5360 | TTI.getShuffleCost(Kind: TargetTransformInfo::SK_Broadcast, DstTy: VectorTy, |
5361 | SrcTy: VectorTy, Mask: {}, CostKind); |
5362 | } |
5363 | StoreInst *SI = cast<StoreInst>(Val: I); |
5364 | |
5365 | bool IsLoopInvariantStoreValue = Legal->isInvariant(V: SI->getValueOperand()); |
5366 | // TODO: We have existing tests that request the cost of extracting element |
5367 | // VF.getKnownMinValue() - 1 from a scalable vector. This does not represent |
5368 | // the actual generated code, which involves extracting the last element of |
5369 | // a scalable vector where the lane to extract is unknown at compile time. |
5370 | return TTI.getAddressComputationCost(Ty: ValTy) + |
5371 | TTI.getMemoryOpCost(Opcode: Instruction::Store, Src: ValTy, Alignment, AddressSpace: AS, |
5372 | CostKind) + |
5373 | (IsLoopInvariantStoreValue |
5374 | ? 0 |
5375 | : TTI.getVectorInstrCost(Opcode: Instruction::ExtractElement, Val: VectorTy, |
5376 | CostKind, Index: VF.getKnownMinValue() - 1)); |
5377 | } |
5378 | |
5379 | InstructionCost |
5380 | LoopVectorizationCostModel::getGatherScatterCost(Instruction *I, |
5381 | ElementCount VF) { |
5382 | Type *ValTy = getLoadStoreType(I); |
5383 | auto *VectorTy = cast<VectorType>(Val: toVectorTy(Scalar: ValTy, EC: VF)); |
5384 | const Align Alignment = getLoadStoreAlignment(I); |
5385 | const Value *Ptr = getLoadStorePointerOperand(V: I); |
5386 | |
5387 | return TTI.getAddressComputationCost(Ty: VectorTy) + |
5388 | TTI.getGatherScatterOpCost(Opcode: I->getOpcode(), DataTy: VectorTy, Ptr, |
5389 | VariableMask: Legal->isMaskRequired(I), Alignment, |
5390 | CostKind, I); |
5391 | } |
5392 | |
5393 | InstructionCost |
5394 | LoopVectorizationCostModel::getInterleaveGroupCost(Instruction *I, |
5395 | ElementCount VF) { |
5396 | const auto *Group = getInterleavedAccessGroup(Instr: I); |
5397 | assert(Group && "Fail to get an interleaved access group." ); |
5398 | |
5399 | Instruction *InsertPos = Group->getInsertPos(); |
5400 | Type *ValTy = getLoadStoreType(I: InsertPos); |
5401 | auto *VectorTy = cast<VectorType>(Val: toVectorTy(Scalar: ValTy, EC: VF)); |
5402 | unsigned AS = getLoadStoreAddressSpace(I: InsertPos); |
5403 | |
5404 | unsigned InterleaveFactor = Group->getFactor(); |
5405 | auto *WideVecTy = VectorType::get(ElementType: ValTy, EC: VF * InterleaveFactor); |
5406 | |
5407 | // Holds the indices of existing members in the interleaved group. |
5408 | SmallVector<unsigned, 4> Indices; |
5409 | for (unsigned IF = 0; IF < InterleaveFactor; IF++) |
5410 | if (Group->getMember(Index: IF)) |
5411 | Indices.push_back(Elt: IF); |
5412 | |
5413 | // Calculate the cost of the whole interleaved group. |
5414 | bool UseMaskForGaps = |
5415 | (Group->requiresScalarEpilogue() && !isScalarEpilogueAllowed()) || |
5416 | (isa<StoreInst>(Val: I) && (Group->getNumMembers() < Group->getFactor())); |
5417 | InstructionCost Cost = TTI.getInterleavedMemoryOpCost( |
5418 | Opcode: InsertPos->getOpcode(), VecTy: WideVecTy, Factor: Group->getFactor(), Indices, |
5419 | Alignment: Group->getAlign(), AddressSpace: AS, CostKind, UseMaskForCond: Legal->isMaskRequired(I), |
5420 | UseMaskForGaps); |
5421 | |
5422 | if (Group->isReverse()) { |
5423 | // TODO: Add support for reversed masked interleaved access. |
5424 | assert(!Legal->isMaskRequired(I) && |
5425 | "Reverse masked interleaved access not supported." ); |
5426 | Cost += Group->getNumMembers() * |
5427 | TTI.getShuffleCost(Kind: TargetTransformInfo::SK_Reverse, DstTy: VectorTy, |
5428 | SrcTy: VectorTy, Mask: {}, CostKind, Index: 0); |
5429 | } |
5430 | return Cost; |
5431 | } |
5432 | |
5433 | std::optional<InstructionCost> |
5434 | LoopVectorizationCostModel::getReductionPatternCost(Instruction *I, |
5435 | ElementCount VF, |
5436 | Type *Ty) const { |
5437 | using namespace llvm::PatternMatch; |
5438 | // Early exit for no inloop reductions |
5439 | if (InLoopReductions.empty() || VF.isScalar() || !isa<VectorType>(Val: Ty)) |
5440 | return std::nullopt; |
5441 | auto *VectorTy = cast<VectorType>(Val: Ty); |
5442 | |
5443 | // We are looking for a pattern of, and finding the minimal acceptable cost: |
5444 | // reduce(mul(ext(A), ext(B))) or |
5445 | // reduce(mul(A, B)) or |
5446 | // reduce(ext(A)) or |
5447 | // reduce(A). |
5448 | // The basic idea is that we walk down the tree to do that, finding the root |
5449 | // reduction instruction in InLoopReductionImmediateChains. From there we find |
5450 | // the pattern of mul/ext and test the cost of the entire pattern vs the cost |
5451 | // of the components. If the reduction cost is lower then we return it for the |
5452 | // reduction instruction and 0 for the other instructions in the pattern. If |
5453 | // it is not we return an invalid cost specifying the orignal cost method |
5454 | // should be used. |
5455 | Instruction *RetI = I; |
5456 | if (match(V: RetI, P: m_ZExtOrSExt(Op: m_Value()))) { |
5457 | if (!RetI->hasOneUser()) |
5458 | return std::nullopt; |
5459 | RetI = RetI->user_back(); |
5460 | } |
5461 | |
5462 | if (match(V: RetI, P: m_OneUse(SubPattern: m_Mul(L: m_Value(), R: m_Value()))) && |
5463 | RetI->user_back()->getOpcode() == Instruction::Add) { |
5464 | RetI = RetI->user_back(); |
5465 | } |
5466 | |
5467 | // Test if the found instruction is a reduction, and if not return an invalid |
5468 | // cost specifying the parent to use the original cost modelling. |
5469 | Instruction *LastChain = InLoopReductionImmediateChains.lookup(Val: RetI); |
5470 | if (!LastChain) |
5471 | return std::nullopt; |
5472 | |
5473 | // Find the reduction this chain is a part of and calculate the basic cost of |
5474 | // the reduction on its own. |
5475 | Instruction *ReductionPhi = LastChain; |
5476 | while (!isa<PHINode>(Val: ReductionPhi)) |
5477 | ReductionPhi = InLoopReductionImmediateChains.at(Val: ReductionPhi); |
5478 | |
5479 | const RecurrenceDescriptor &RdxDesc = |
5480 | Legal->getReductionVars().find(Key: cast<PHINode>(Val: ReductionPhi))->second; |
5481 | |
5482 | InstructionCost BaseCost; |
5483 | RecurKind RK = RdxDesc.getRecurrenceKind(); |
5484 | if (RecurrenceDescriptor::isMinMaxRecurrenceKind(Kind: RK)) { |
5485 | Intrinsic::ID MinMaxID = getMinMaxReductionIntrinsicOp(RK); |
5486 | BaseCost = TTI.getMinMaxReductionCost(IID: MinMaxID, Ty: VectorTy, |
5487 | FMF: RdxDesc.getFastMathFlags(), CostKind); |
5488 | } else { |
5489 | BaseCost = TTI.getArithmeticReductionCost( |
5490 | Opcode: RdxDesc.getOpcode(), Ty: VectorTy, FMF: RdxDesc.getFastMathFlags(), CostKind); |
5491 | } |
5492 | |
5493 | // For a call to the llvm.fmuladd intrinsic we need to add the cost of a |
5494 | // normal fmul instruction to the cost of the fadd reduction. |
5495 | if (RK == RecurKind::FMulAdd) |
5496 | BaseCost += |
5497 | TTI.getArithmeticInstrCost(Opcode: Instruction::FMul, Ty: VectorTy, CostKind); |
5498 | |
5499 | // If we're using ordered reductions then we can just return the base cost |
5500 | // here, since getArithmeticReductionCost calculates the full ordered |
5501 | // reduction cost when FP reassociation is not allowed. |
5502 | if (useOrderedReductions(RdxDesc)) |
5503 | return BaseCost; |
5504 | |
5505 | // Get the operand that was not the reduction chain and match it to one of the |
5506 | // patterns, returning the better cost if it is found. |
5507 | Instruction *RedOp = RetI->getOperand(i: 1) == LastChain |
5508 | ? dyn_cast<Instruction>(Val: RetI->getOperand(i: 0)) |
5509 | : dyn_cast<Instruction>(Val: RetI->getOperand(i: 1)); |
5510 | |
5511 | VectorTy = VectorType::get(ElementType: I->getOperand(i: 0)->getType(), Other: VectorTy); |
5512 | |
5513 | Instruction *Op0, *Op1; |
5514 | if (RedOp && RdxDesc.getOpcode() == Instruction::Add && |
5515 | match(V: RedOp, |
5516 | P: m_ZExtOrSExt(Op: m_Mul(L: m_Instruction(I&: Op0), R: m_Instruction(I&: Op1)))) && |
5517 | match(V: Op0, P: m_ZExtOrSExt(Op: m_Value())) && |
5518 | Op0->getOpcode() == Op1->getOpcode() && |
5519 | Op0->getOperand(i: 0)->getType() == Op1->getOperand(i: 0)->getType() && |
5520 | !TheLoop->isLoopInvariant(V: Op0) && !TheLoop->isLoopInvariant(V: Op1) && |
5521 | (Op0->getOpcode() == RedOp->getOpcode() || Op0 == Op1)) { |
5522 | |
5523 | // Matched reduce.add(ext(mul(ext(A), ext(B))) |
5524 | // Note that the extend opcodes need to all match, or if A==B they will have |
5525 | // been converted to zext(mul(sext(A), sext(A))) as it is known positive, |
5526 | // which is equally fine. |
5527 | bool IsUnsigned = isa<ZExtInst>(Val: Op0); |
5528 | auto *ExtType = VectorType::get(ElementType: Op0->getOperand(i: 0)->getType(), Other: VectorTy); |
5529 | auto *MulType = VectorType::get(ElementType: Op0->getType(), Other: VectorTy); |
5530 | |
5531 | InstructionCost ExtCost = |
5532 | TTI.getCastInstrCost(Opcode: Op0->getOpcode(), Dst: MulType, Src: ExtType, |
5533 | CCH: TTI::CastContextHint::None, CostKind, I: Op0); |
5534 | InstructionCost MulCost = |
5535 | TTI.getArithmeticInstrCost(Opcode: Instruction::Mul, Ty: MulType, CostKind); |
5536 | InstructionCost Ext2Cost = |
5537 | TTI.getCastInstrCost(Opcode: RedOp->getOpcode(), Dst: VectorTy, Src: MulType, |
5538 | CCH: TTI::CastContextHint::None, CostKind, I: RedOp); |
5539 | |
5540 | InstructionCost RedCost = TTI.getMulAccReductionCost( |
5541 | IsUnsigned, ResTy: RdxDesc.getRecurrenceType(), Ty: ExtType, CostKind); |
5542 | |
5543 | if (RedCost.isValid() && |
5544 | RedCost < ExtCost * 2 + MulCost + Ext2Cost + BaseCost) |
5545 | return I == RetI ? RedCost : 0; |
5546 | } else if (RedOp && match(V: RedOp, P: m_ZExtOrSExt(Op: m_Value())) && |
5547 | !TheLoop->isLoopInvariant(V: RedOp)) { |
5548 | // Matched reduce(ext(A)) |
5549 | bool IsUnsigned = isa<ZExtInst>(Val: RedOp); |
5550 | auto *ExtType = VectorType::get(ElementType: RedOp->getOperand(i: 0)->getType(), Other: VectorTy); |
5551 | InstructionCost RedCost = TTI.getExtendedReductionCost( |
5552 | Opcode: RdxDesc.getOpcode(), IsUnsigned, ResTy: RdxDesc.getRecurrenceType(), Ty: ExtType, |
5553 | FMF: RdxDesc.getFastMathFlags(), CostKind); |
5554 | |
5555 | InstructionCost ExtCost = |
5556 | TTI.getCastInstrCost(Opcode: RedOp->getOpcode(), Dst: VectorTy, Src: ExtType, |
5557 | CCH: TTI::CastContextHint::None, CostKind, I: RedOp); |
5558 | if (RedCost.isValid() && RedCost < BaseCost + ExtCost) |
5559 | return I == RetI ? RedCost : 0; |
5560 | } else if (RedOp && RdxDesc.getOpcode() == Instruction::Add && |
5561 | match(V: RedOp, P: m_Mul(L: m_Instruction(I&: Op0), R: m_Instruction(I&: Op1)))) { |
5562 | if (match(V: Op0, P: m_ZExtOrSExt(Op: m_Value())) && |
5563 | Op0->getOpcode() == Op1->getOpcode() && |
5564 | !TheLoop->isLoopInvariant(V: Op0) && !TheLoop->isLoopInvariant(V: Op1)) { |
5565 | bool IsUnsigned = isa<ZExtInst>(Val: Op0); |
5566 | Type *Op0Ty = Op0->getOperand(i: 0)->getType(); |
5567 | Type *Op1Ty = Op1->getOperand(i: 0)->getType(); |
5568 | Type *LargestOpTy = |
5569 | Op0Ty->getIntegerBitWidth() < Op1Ty->getIntegerBitWidth() ? Op1Ty |
5570 | : Op0Ty; |
5571 | auto *ExtType = VectorType::get(ElementType: LargestOpTy, Other: VectorTy); |
5572 | |
5573 | // Matched reduce.add(mul(ext(A), ext(B))), where the two ext may be of |
5574 | // different sizes. We take the largest type as the ext to reduce, and add |
5575 | // the remaining cost as, for example reduce(mul(ext(ext(A)), ext(B))). |
5576 | InstructionCost ExtCost0 = TTI.getCastInstrCost( |
5577 | Opcode: Op0->getOpcode(), Dst: VectorTy, Src: VectorType::get(ElementType: Op0Ty, Other: VectorTy), |
5578 | CCH: TTI::CastContextHint::None, CostKind, I: Op0); |
5579 | InstructionCost ExtCost1 = TTI.getCastInstrCost( |
5580 | Opcode: Op1->getOpcode(), Dst: VectorTy, Src: VectorType::get(ElementType: Op1Ty, Other: VectorTy), |
5581 | CCH: TTI::CastContextHint::None, CostKind, I: Op1); |
5582 | InstructionCost MulCost = |
5583 | TTI.getArithmeticInstrCost(Opcode: Instruction::Mul, Ty: VectorTy, CostKind); |
5584 | |
5585 | InstructionCost RedCost = TTI.getMulAccReductionCost( |
5586 | IsUnsigned, ResTy: RdxDesc.getRecurrenceType(), Ty: ExtType, CostKind); |
5587 | InstructionCost = 0; |
5588 | if (Op0Ty != LargestOpTy || Op1Ty != LargestOpTy) { |
5589 | Instruction * = (Op0Ty != LargestOpTy) ? Op0 : Op1; |
5590 | ExtraExtCost = TTI.getCastInstrCost( |
5591 | Opcode: ExtraExtOp->getOpcode(), Dst: ExtType, |
5592 | Src: VectorType::get(ElementType: ExtraExtOp->getOperand(i: 0)->getType(), Other: VectorTy), |
5593 | CCH: TTI::CastContextHint::None, CostKind, I: ExtraExtOp); |
5594 | } |
5595 | |
5596 | if (RedCost.isValid() && |
5597 | (RedCost + ExtraExtCost) < (ExtCost0 + ExtCost1 + MulCost + BaseCost)) |
5598 | return I == RetI ? RedCost : 0; |
5599 | } else if (!match(V: I, P: m_ZExtOrSExt(Op: m_Value()))) { |
5600 | // Matched reduce.add(mul()) |
5601 | InstructionCost MulCost = |
5602 | TTI.getArithmeticInstrCost(Opcode: Instruction::Mul, Ty: VectorTy, CostKind); |
5603 | |
5604 | InstructionCost RedCost = TTI.getMulAccReductionCost( |
5605 | IsUnsigned: true, ResTy: RdxDesc.getRecurrenceType(), Ty: VectorTy, CostKind); |
5606 | |
5607 | if (RedCost.isValid() && RedCost < MulCost + BaseCost) |
5608 | return I == RetI ? RedCost : 0; |
5609 | } |
5610 | } |
5611 | |
5612 | return I == RetI ? std::optional<InstructionCost>(BaseCost) : std::nullopt; |
5613 | } |
5614 | |
5615 | InstructionCost |
5616 | LoopVectorizationCostModel::getMemoryInstructionCost(Instruction *I, |
5617 | ElementCount VF) { |
5618 | // Calculate scalar cost only. Vectorization cost should be ready at this |
5619 | // moment. |
5620 | if (VF.isScalar()) { |
5621 | Type *ValTy = getLoadStoreType(I); |
5622 | const Align Alignment = getLoadStoreAlignment(I); |
5623 | unsigned AS = getLoadStoreAddressSpace(I); |
5624 | |
5625 | TTI::OperandValueInfo OpInfo = TTI::getOperandInfo(V: I->getOperand(i: 0)); |
5626 | return TTI.getAddressComputationCost(Ty: ValTy) + |
5627 | TTI.getMemoryOpCost(Opcode: I->getOpcode(), Src: ValTy, Alignment, AddressSpace: AS, CostKind, |
5628 | OpdInfo: OpInfo, I); |
5629 | } |
5630 | return getWideningCost(I, VF); |
5631 | } |
5632 | |
5633 | InstructionCost |
5634 | LoopVectorizationCostModel::getScalarizationOverhead(Instruction *I, |
5635 | ElementCount VF) const { |
5636 | |
5637 | // There is no mechanism yet to create a scalable scalarization loop, |
5638 | // so this is currently Invalid. |
5639 | if (VF.isScalable()) |
5640 | return InstructionCost::getInvalid(); |
5641 | |
5642 | if (VF.isScalar()) |
5643 | return 0; |
5644 | |
5645 | InstructionCost Cost = 0; |
5646 | Type *RetTy = toVectorizedTy(Ty: I->getType(), EC: VF); |
5647 | if (!RetTy->isVoidTy() && |
5648 | (!isa<LoadInst>(Val: I) || !TTI.supportsEfficientVectorElementLoadStore())) { |
5649 | |
5650 | for (Type *VectorTy : getContainedTypes(Ty: RetTy)) { |
5651 | Cost += TTI.getScalarizationOverhead( |
5652 | Ty: cast<VectorType>(Val: VectorTy), DemandedElts: APInt::getAllOnes(numBits: VF.getFixedValue()), |
5653 | /*Insert=*/true, |
5654 | /*Extract=*/false, CostKind); |
5655 | } |
5656 | } |
5657 | |
5658 | // Some targets keep addresses scalar. |
5659 | if (isa<LoadInst>(Val: I) && !TTI.prefersVectorizedAddressing()) |
5660 | return Cost; |
5661 | |
5662 | // Some targets support efficient element stores. |
5663 | if (isa<StoreInst>(Val: I) && TTI.supportsEfficientVectorElementLoadStore()) |
5664 | return Cost; |
5665 | |
5666 | // Collect operands to consider. |
5667 | CallInst *CI = dyn_cast<CallInst>(Val: I); |
5668 | Instruction::op_range Ops = CI ? CI->args() : I->operands(); |
5669 | |
5670 | // Skip operands that do not require extraction/scalarization and do not incur |
5671 | // any overhead. |
5672 | SmallVector<Type *> Tys; |
5673 | for (auto *V : filterExtractingOperands(Ops, VF)) |
5674 | Tys.push_back(Elt: maybeVectorizeType(Ty: V->getType(), VF)); |
5675 | return Cost + TTI.getOperandsScalarizationOverhead( |
5676 | Args: filterExtractingOperands(Ops, VF), Tys, CostKind); |
5677 | } |
5678 | |
5679 | void LoopVectorizationCostModel::setCostBasedWideningDecision(ElementCount VF) { |
5680 | if (VF.isScalar()) |
5681 | return; |
5682 | NumPredStores = 0; |
5683 | for (BasicBlock *BB : TheLoop->blocks()) { |
5684 | // For each instruction in the old loop. |
5685 | for (Instruction &I : *BB) { |
5686 | Value *Ptr = getLoadStorePointerOperand(V: &I); |
5687 | if (!Ptr) |
5688 | continue; |
5689 | |
5690 | // TODO: We should generate better code and update the cost model for |
5691 | // predicated uniform stores. Today they are treated as any other |
5692 | // predicated store (see added test cases in |
5693 | // invariant-store-vectorization.ll). |
5694 | if (isa<StoreInst>(Val: &I) && isScalarWithPredication(I: &I, VF)) |
5695 | NumPredStores++; |
5696 | |
5697 | if (Legal->isUniformMemOp(I, VF)) { |
5698 | auto IsLegalToScalarize = [&]() { |
5699 | if (!VF.isScalable()) |
5700 | // Scalarization of fixed length vectors "just works". |
5701 | return true; |
5702 | |
5703 | // We have dedicated lowering for unpredicated uniform loads and |
5704 | // stores. Note that even with tail folding we know that at least |
5705 | // one lane is active (i.e. generalized predication is not possible |
5706 | // here), and the logic below depends on this fact. |
5707 | if (!foldTailByMasking()) |
5708 | return true; |
5709 | |
5710 | // For scalable vectors, a uniform memop load is always |
5711 | // uniform-by-parts and we know how to scalarize that. |
5712 | if (isa<LoadInst>(Val: I)) |
5713 | return true; |
5714 | |
5715 | // A uniform store isn't neccessarily uniform-by-part |
5716 | // and we can't assume scalarization. |
5717 | auto &SI = cast<StoreInst>(Val&: I); |
5718 | return TheLoop->isLoopInvariant(V: SI.getValueOperand()); |
5719 | }; |
5720 | |
5721 | const InstructionCost GatherScatterCost = |
5722 | isLegalGatherOrScatter(V: &I, VF) ? |
5723 | getGatherScatterCost(I: &I, VF) : InstructionCost::getInvalid(); |
5724 | |
5725 | // Load: Scalar load + broadcast |
5726 | // Store: Scalar store + isLoopInvariantStoreValue ? 0 : extract |
5727 | // FIXME: This cost is a significant under-estimate for tail folded |
5728 | // memory ops. |
5729 | const InstructionCost ScalarizationCost = |
5730 | IsLegalToScalarize() ? getUniformMemOpCost(I: &I, VF) |
5731 | : InstructionCost::getInvalid(); |
5732 | |
5733 | // Choose better solution for the current VF, Note that Invalid |
5734 | // costs compare as maximumal large. If both are invalid, we get |
5735 | // scalable invalid which signals a failure and a vectorization abort. |
5736 | if (GatherScatterCost < ScalarizationCost) |
5737 | setWideningDecision(I: &I, VF, W: CM_GatherScatter, Cost: GatherScatterCost); |
5738 | else |
5739 | setWideningDecision(I: &I, VF, W: CM_Scalarize, Cost: ScalarizationCost); |
5740 | continue; |
5741 | } |
5742 | |
5743 | // We assume that widening is the best solution when possible. |
5744 | if (memoryInstructionCanBeWidened(I: &I, VF)) { |
5745 | InstructionCost Cost = getConsecutiveMemOpCost(I: &I, VF); |
5746 | int ConsecutiveStride = Legal->isConsecutivePtr( |
5747 | AccessTy: getLoadStoreType(I: &I), Ptr: getLoadStorePointerOperand(V: &I)); |
5748 | assert((ConsecutiveStride == 1 || ConsecutiveStride == -1) && |
5749 | "Expected consecutive stride." ); |
5750 | InstWidening Decision = |
5751 | ConsecutiveStride == 1 ? CM_Widen : CM_Widen_Reverse; |
5752 | setWideningDecision(I: &I, VF, W: Decision, Cost); |
5753 | continue; |
5754 | } |
5755 | |
5756 | // Choose between Interleaving, Gather/Scatter or Scalarization. |
5757 | InstructionCost InterleaveCost = InstructionCost::getInvalid(); |
5758 | unsigned NumAccesses = 1; |
5759 | if (isAccessInterleaved(Instr: &I)) { |
5760 | const auto *Group = getInterleavedAccessGroup(Instr: &I); |
5761 | assert(Group && "Fail to get an interleaved access group." ); |
5762 | |
5763 | // Make one decision for the whole group. |
5764 | if (getWideningDecision(I: &I, VF) != CM_Unknown) |
5765 | continue; |
5766 | |
5767 | NumAccesses = Group->getNumMembers(); |
5768 | if (interleavedAccessCanBeWidened(I: &I, VF)) |
5769 | InterleaveCost = getInterleaveGroupCost(I: &I, VF); |
5770 | } |
5771 | |
5772 | InstructionCost GatherScatterCost = |
5773 | isLegalGatherOrScatter(V: &I, VF) |
5774 | ? getGatherScatterCost(I: &I, VF) * NumAccesses |
5775 | : InstructionCost::getInvalid(); |
5776 | |
5777 | InstructionCost ScalarizationCost = |
5778 | getMemInstScalarizationCost(I: &I, VF) * NumAccesses; |
5779 | |
5780 | // Choose better solution for the current VF, |
5781 | // write down this decision and use it during vectorization. |
5782 | InstructionCost Cost; |
5783 | InstWidening Decision; |
5784 | if (InterleaveCost <= GatherScatterCost && |
5785 | InterleaveCost < ScalarizationCost) { |
5786 | Decision = CM_Interleave; |
5787 | Cost = InterleaveCost; |
5788 | } else if (GatherScatterCost < ScalarizationCost) { |
5789 | Decision = CM_GatherScatter; |
5790 | Cost = GatherScatterCost; |
5791 | } else { |
5792 | Decision = CM_Scalarize; |
5793 | Cost = ScalarizationCost; |
5794 | } |
5795 | // If the instructions belongs to an interleave group, the whole group |
5796 | // receives the same decision. The whole group receives the cost, but |
5797 | // the cost will actually be assigned to one instruction. |
5798 | if (const auto *Group = getInterleavedAccessGroup(Instr: &I)) |
5799 | setWideningDecision(Grp: Group, VF, W: Decision, Cost); |
5800 | else |
5801 | setWideningDecision(I: &I, VF, W: Decision, Cost); |
5802 | } |
5803 | } |
5804 | |
5805 | // Make sure that any load of address and any other address computation |
5806 | // remains scalar unless there is gather/scatter support. This avoids |
5807 | // inevitable extracts into address registers, and also has the benefit of |
5808 | // activating LSR more, since that pass can't optimize vectorized |
5809 | // addresses. |
5810 | if (TTI.prefersVectorizedAddressing()) |
5811 | return; |
5812 | |
5813 | // Start with all scalar pointer uses. |
5814 | SmallPtrSet<Instruction *, 8> AddrDefs; |
5815 | for (BasicBlock *BB : TheLoop->blocks()) |
5816 | for (Instruction &I : *BB) { |
5817 | Instruction *PtrDef = |
5818 | dyn_cast_or_null<Instruction>(Val: getLoadStorePointerOperand(V: &I)); |
5819 | if (PtrDef && TheLoop->contains(Inst: PtrDef) && |
5820 | getWideningDecision(I: &I, VF) != CM_GatherScatter) |
5821 | AddrDefs.insert(Ptr: PtrDef); |
5822 | } |
5823 | |
5824 | // Add all instructions used to generate the addresses. |
5825 | SmallVector<Instruction *, 4> Worklist; |
5826 | append_range(C&: Worklist, R&: AddrDefs); |
5827 | while (!Worklist.empty()) { |
5828 | Instruction *I = Worklist.pop_back_val(); |
5829 | for (auto &Op : I->operands()) |
5830 | if (auto *InstOp = dyn_cast<Instruction>(Val&: Op)) |
5831 | if ((InstOp->getParent() == I->getParent()) && !isa<PHINode>(Val: InstOp) && |
5832 | AddrDefs.insert(Ptr: InstOp).second) |
5833 | Worklist.push_back(Elt: InstOp); |
5834 | } |
5835 | |
5836 | for (auto *I : AddrDefs) { |
5837 | if (isa<LoadInst>(Val: I)) { |
5838 | // Setting the desired widening decision should ideally be handled in |
5839 | // by cost functions, but since this involves the task of finding out |
5840 | // if the loaded register is involved in an address computation, it is |
5841 | // instead changed here when we know this is the case. |
5842 | InstWidening Decision = getWideningDecision(I, VF); |
5843 | if (Decision == CM_Widen || Decision == CM_Widen_Reverse) |
5844 | // Scalarize a widened load of address. |
5845 | setWideningDecision( |
5846 | I, VF, W: CM_Scalarize, |
5847 | Cost: (VF.getKnownMinValue() * |
5848 | getMemoryInstructionCost(I, VF: ElementCount::getFixed(MinVal: 1)))); |
5849 | else if (const auto *Group = getInterleavedAccessGroup(Instr: I)) { |
5850 | // Scalarize an interleave group of address loads. |
5851 | for (unsigned I = 0; I < Group->getFactor(); ++I) { |
5852 | if (Instruction *Member = Group->getMember(Index: I)) |
5853 | setWideningDecision( |
5854 | I: Member, VF, W: CM_Scalarize, |
5855 | Cost: (VF.getKnownMinValue() * |
5856 | getMemoryInstructionCost(I: Member, VF: ElementCount::getFixed(MinVal: 1)))); |
5857 | } |
5858 | } |
5859 | } else { |
5860 | // Cannot scalarize fixed-order recurrence phis at the moment. |
5861 | if (isa<PHINode>(Val: I) && Legal->isFixedOrderRecurrence(Phi: cast<PHINode>(Val: I))) |
5862 | continue; |
5863 | |
5864 | // Make sure I gets scalarized and a cost estimate without |
5865 | // scalarization overhead. |
5866 | ForcedScalars[VF].insert(Ptr: I); |
5867 | } |
5868 | } |
5869 | } |
5870 | |
5871 | void LoopVectorizationCostModel::setVectorizedCallDecision(ElementCount VF) { |
5872 | assert(!VF.isScalar() && |
5873 | "Trying to set a vectorization decision for a scalar VF" ); |
5874 | |
5875 | auto ForcedScalar = ForcedScalars.find(Val: VF); |
5876 | for (BasicBlock *BB : TheLoop->blocks()) { |
5877 | // For each instruction in the old loop. |
5878 | for (Instruction &I : *BB) { |
5879 | CallInst *CI = dyn_cast<CallInst>(Val: &I); |
5880 | |
5881 | if (!CI) |
5882 | continue; |
5883 | |
5884 | InstructionCost ScalarCost = InstructionCost::getInvalid(); |
5885 | InstructionCost VectorCost = InstructionCost::getInvalid(); |
5886 | InstructionCost IntrinsicCost = InstructionCost::getInvalid(); |
5887 | Function *ScalarFunc = CI->getCalledFunction(); |
5888 | Type *ScalarRetTy = CI->getType(); |
5889 | SmallVector<Type *, 4> Tys, ScalarTys; |
5890 | for (auto &ArgOp : CI->args()) |
5891 | ScalarTys.push_back(Elt: ArgOp->getType()); |
5892 | |
5893 | // Estimate cost of scalarized vector call. The source operands are |
5894 | // assumed to be vectors, so we need to extract individual elements from |
5895 | // there, execute VF scalar calls, and then gather the result into the |
5896 | // vector return value. |
5897 | InstructionCost ScalarCallCost = |
5898 | TTI.getCallInstrCost(F: ScalarFunc, RetTy: ScalarRetTy, Tys: ScalarTys, CostKind); |
5899 | |
5900 | // Compute costs of unpacking argument values for the scalar calls and |
5901 | // packing the return values to a vector. |
5902 | InstructionCost ScalarizationCost = getScalarizationOverhead(I: CI, VF); |
5903 | |
5904 | ScalarCost = ScalarCallCost * VF.getKnownMinValue() + ScalarizationCost; |
5905 | // Honor ForcedScalars and UniformAfterVectorization decisions. |
5906 | // TODO: For calls, it might still be more profitable to widen. Use |
5907 | // VPlan-based cost model to compare different options. |
5908 | if (VF.isVector() && ((ForcedScalar != ForcedScalars.end() && |
5909 | ForcedScalar->second.contains(Ptr: CI)) || |
5910 | isUniformAfterVectorization(I: CI, VF))) { |
5911 | setCallWideningDecision(CI, VF, Kind: CM_Scalarize, Variant: nullptr, |
5912 | IID: Intrinsic::not_intrinsic, MaskPos: std::nullopt, |
5913 | Cost: ScalarCost); |
5914 | continue; |
5915 | } |
5916 | |
5917 | bool MaskRequired = Legal->isMaskRequired(I: CI); |
5918 | // Compute corresponding vector type for return value and arguments. |
5919 | Type *RetTy = toVectorizedTy(Ty: ScalarRetTy, EC: VF); |
5920 | for (Type *ScalarTy : ScalarTys) |
5921 | Tys.push_back(Elt: toVectorizedTy(Ty: ScalarTy, EC: VF)); |
5922 | |
5923 | // An in-loop reduction using an fmuladd intrinsic is a special case; |
5924 | // we don't want the normal cost for that intrinsic. |
5925 | if (RecurrenceDescriptor::isFMulAddIntrinsic(I: CI)) |
5926 | if (auto RedCost = getReductionPatternCost(I: CI, VF, Ty: RetTy)) { |
5927 | setCallWideningDecision(CI, VF, Kind: CM_IntrinsicCall, Variant: nullptr, |
5928 | IID: getVectorIntrinsicIDForCall(CI, TLI), |
5929 | MaskPos: std::nullopt, Cost: *RedCost); |
5930 | continue; |
5931 | } |
5932 | |
5933 | // Find the cost of vectorizing the call, if we can find a suitable |
5934 | // vector variant of the function. |
5935 | VFInfo FuncInfo; |
5936 | Function *VecFunc = nullptr; |
5937 | // Search through any available variants for one we can use at this VF. |
5938 | for (VFInfo &Info : VFDatabase::getMappings(CI: *CI)) { |
5939 | // Must match requested VF. |
5940 | if (Info.Shape.VF != VF) |
5941 | continue; |
5942 | |
5943 | // Must take a mask argument if one is required |
5944 | if (MaskRequired && !Info.isMasked()) |
5945 | continue; |
5946 | |
5947 | // Check that all parameter kinds are supported |
5948 | bool ParamsOk = true; |
5949 | for (VFParameter Param : Info.Shape.Parameters) { |
5950 | switch (Param.ParamKind) { |
5951 | case VFParamKind::Vector: |
5952 | break; |
5953 | case VFParamKind::OMP_Uniform: { |
5954 | Value *ScalarParam = CI->getArgOperand(i: Param.ParamPos); |
5955 | // Make sure the scalar parameter in the loop is invariant. |
5956 | if (!PSE.getSE()->isLoopInvariant(S: PSE.getSCEV(V: ScalarParam), |
5957 | L: TheLoop)) |
5958 | ParamsOk = false; |
5959 | break; |
5960 | } |
5961 | case VFParamKind::OMP_Linear: { |
5962 | Value *ScalarParam = CI->getArgOperand(i: Param.ParamPos); |
5963 | // Find the stride for the scalar parameter in this loop and see if |
5964 | // it matches the stride for the variant. |
5965 | // TODO: do we need to figure out the cost of an extract to get the |
5966 | // first lane? Or do we hope that it will be folded away? |
5967 | ScalarEvolution *SE = PSE.getSE(); |
5968 | const auto *SAR = |
5969 | dyn_cast<SCEVAddRecExpr>(Val: SE->getSCEV(V: ScalarParam)); |
5970 | |
5971 | if (!SAR || SAR->getLoop() != TheLoop) { |
5972 | ParamsOk = false; |
5973 | break; |
5974 | } |
5975 | |
5976 | const SCEVConstant *Step = |
5977 | dyn_cast<SCEVConstant>(Val: SAR->getStepRecurrence(SE&: *SE)); |
5978 | |
5979 | if (!Step || |
5980 | Step->getAPInt().getSExtValue() != Param.LinearStepOrPos) |
5981 | ParamsOk = false; |
5982 | |
5983 | break; |
5984 | } |
5985 | case VFParamKind::GlobalPredicate: |
5986 | break; |
5987 | default: |
5988 | ParamsOk = false; |
5989 | break; |
5990 | } |
5991 | } |
5992 | |
5993 | if (!ParamsOk) |
5994 | continue; |
5995 | |
5996 | // Found a suitable candidate, stop here. |
5997 | VecFunc = CI->getModule()->getFunction(Name: Info.VectorName); |
5998 | FuncInfo = Info; |
5999 | break; |
6000 | } |
6001 | |
6002 | if (TLI && VecFunc && !CI->isNoBuiltin()) |
6003 | VectorCost = TTI.getCallInstrCost(F: nullptr, RetTy, Tys, CostKind); |
6004 | |
6005 | // Find the cost of an intrinsic; some targets may have instructions that |
6006 | // perform the operation without needing an actual call. |
6007 | Intrinsic::ID IID = getVectorIntrinsicIDForCall(CI, TLI); |
6008 | if (IID != Intrinsic::not_intrinsic) |
6009 | IntrinsicCost = getVectorIntrinsicCost(CI, VF); |
6010 | |
6011 | InstructionCost Cost = ScalarCost; |
6012 | InstWidening Decision = CM_Scalarize; |
6013 | |
6014 | if (VectorCost <= Cost) { |
6015 | Cost = VectorCost; |
6016 | Decision = CM_VectorCall; |
6017 | } |
6018 | |
6019 | if (IntrinsicCost <= Cost) { |
6020 | Cost = IntrinsicCost; |
6021 | Decision = CM_IntrinsicCall; |
6022 | } |
6023 | |
6024 | setCallWideningDecision(CI, VF, Kind: Decision, Variant: VecFunc, IID, |
6025 | MaskPos: FuncInfo.getParamIndexForOptionalMask(), Cost); |
6026 | } |
6027 | } |
6028 | } |
6029 | |
6030 | bool LoopVectorizationCostModel::shouldConsiderInvariant(Value *Op) { |
6031 | if (!Legal->isInvariant(V: Op)) |
6032 | return false; |
6033 | // Consider Op invariant, if it or its operands aren't predicated |
6034 | // instruction in the loop. In that case, it is not trivially hoistable. |
6035 | auto *OpI = dyn_cast<Instruction>(Val: Op); |
6036 | return !OpI || !TheLoop->contains(Inst: OpI) || |
6037 | (!isPredicatedInst(I: OpI) && |
6038 | (!isa<PHINode>(Val: OpI) || OpI->getParent() != TheLoop->getHeader()) && |
6039 | all_of(Range: OpI->operands(), |
6040 | P: [this](Value *Op) { return shouldConsiderInvariant(Op); })); |
6041 | } |
6042 | |
6043 | InstructionCost |
6044 | LoopVectorizationCostModel::getInstructionCost(Instruction *I, |
6045 | ElementCount VF) { |
6046 | // If we know that this instruction will remain uniform, check the cost of |
6047 | // the scalar version. |
6048 | if (isUniformAfterVectorization(I, VF)) |
6049 | VF = ElementCount::getFixed(MinVal: 1); |
6050 | |
6051 | if (VF.isVector() && isProfitableToScalarize(I, VF)) |
6052 | return InstsToScalarize[VF][I]; |
6053 | |
6054 | // Forced scalars do not have any scalarization overhead. |
6055 | auto ForcedScalar = ForcedScalars.find(Val: VF); |
6056 | if (VF.isVector() && ForcedScalar != ForcedScalars.end()) { |
6057 | auto InstSet = ForcedScalar->second; |
6058 | if (InstSet.count(Ptr: I)) |
6059 | return getInstructionCost(I, VF: ElementCount::getFixed(MinVal: 1)) * |
6060 | VF.getKnownMinValue(); |
6061 | } |
6062 | |
6063 | Type *RetTy = I->getType(); |
6064 | if (canTruncateToMinimalBitwidth(I, VF)) |
6065 | RetTy = IntegerType::get(C&: RetTy->getContext(), NumBits: MinBWs[I]); |
6066 | auto *SE = PSE.getSE(); |
6067 | |
6068 | Type *VectorTy; |
6069 | if (isScalarAfterVectorization(I, VF)) { |
6070 | [[maybe_unused]] auto HasSingleCopyAfterVectorization = |
6071 | [this](Instruction *I, ElementCount VF) -> bool { |
6072 | if (VF.isScalar()) |
6073 | return true; |
6074 | |
6075 | auto Scalarized = InstsToScalarize.find(Val: VF); |
6076 | assert(Scalarized != InstsToScalarize.end() && |
6077 | "VF not yet analyzed for scalarization profitability" ); |
6078 | return !Scalarized->second.count(Val: I) && |
6079 | llvm::all_of(Range: I->users(), P: [&](User *U) { |
6080 | auto *UI = cast<Instruction>(Val: U); |
6081 | return !Scalarized->second.count(Val: UI); |
6082 | }); |
6083 | }; |
6084 | |
6085 | // With the exception of GEPs and PHIs, after scalarization there should |
6086 | // only be one copy of the instruction generated in the loop. This is |
6087 | // because the VF is either 1, or any instructions that need scalarizing |
6088 | // have already been dealt with by the time we get here. As a result, |
6089 | // it means we don't have to multiply the instruction cost by VF. |
6090 | assert(I->getOpcode() == Instruction::GetElementPtr || |
6091 | I->getOpcode() == Instruction::PHI || |
6092 | (I->getOpcode() == Instruction::BitCast && |
6093 | I->getType()->isPointerTy()) || |
6094 | HasSingleCopyAfterVectorization(I, VF)); |
6095 | VectorTy = RetTy; |
6096 | } else |
6097 | VectorTy = toVectorizedTy(Ty: RetTy, EC: VF); |
6098 | |
6099 | if (VF.isVector() && VectorTy->isVectorTy() && |
6100 | !TTI.getNumberOfParts(Tp: VectorTy)) |
6101 | return InstructionCost::getInvalid(); |
6102 | |
6103 | // TODO: We need to estimate the cost of intrinsic calls. |
6104 | switch (I->getOpcode()) { |
6105 | case Instruction::GetElementPtr: |
6106 | // We mark this instruction as zero-cost because the cost of GEPs in |
6107 | // vectorized code depends on whether the corresponding memory instruction |
6108 | // is scalarized or not. Therefore, we handle GEPs with the memory |
6109 | // instruction cost. |
6110 | return 0; |
6111 | case Instruction::Br: { |
6112 | // In cases of scalarized and predicated instructions, there will be VF |
6113 | // predicated blocks in the vectorized loop. Each branch around these |
6114 | // blocks requires also an extract of its vector compare i1 element. |
6115 | // Note that the conditional branch from the loop latch will be replaced by |
6116 | // a single branch controlling the loop, so there is no extra overhead from |
6117 | // scalarization. |
6118 | bool ScalarPredicatedBB = false; |
6119 | BranchInst *BI = cast<BranchInst>(Val: I); |
6120 | if (VF.isVector() && BI->isConditional() && |
6121 | (PredicatedBBsAfterVectorization[VF].count(Ptr: BI->getSuccessor(i: 0)) || |
6122 | PredicatedBBsAfterVectorization[VF].count(Ptr: BI->getSuccessor(i: 1))) && |
6123 | BI->getParent() != TheLoop->getLoopLatch()) |
6124 | ScalarPredicatedBB = true; |
6125 | |
6126 | if (ScalarPredicatedBB) { |
6127 | // Not possible to scalarize scalable vector with predicated instructions. |
6128 | if (VF.isScalable()) |
6129 | return InstructionCost::getInvalid(); |
6130 | // Return cost for branches around scalarized and predicated blocks. |
6131 | auto *VecI1Ty = |
6132 | VectorType::get(ElementType: IntegerType::getInt1Ty(C&: RetTy->getContext()), EC: VF); |
6133 | return ( |
6134 | TTI.getScalarizationOverhead( |
6135 | Ty: VecI1Ty, DemandedElts: APInt::getAllOnes(numBits: VF.getFixedValue()), |
6136 | /*Insert*/ false, /*Extract*/ true, CostKind) + |
6137 | (TTI.getCFInstrCost(Opcode: Instruction::Br, CostKind) * VF.getFixedValue())); |
6138 | } |
6139 | |
6140 | if (I->getParent() == TheLoop->getLoopLatch() || VF.isScalar()) |
6141 | // The back-edge branch will remain, as will all scalar branches. |
6142 | return TTI.getCFInstrCost(Opcode: Instruction::Br, CostKind); |
6143 | |
6144 | // This branch will be eliminated by if-conversion. |
6145 | return 0; |
6146 | // Note: We currently assume zero cost for an unconditional branch inside |
6147 | // a predicated block since it will become a fall-through, although we |
6148 | // may decide in the future to call TTI for all branches. |
6149 | } |
6150 | case Instruction::Switch: { |
6151 | if (VF.isScalar()) |
6152 | return TTI.getCFInstrCost(Opcode: Instruction::Switch, CostKind); |
6153 | auto *Switch = cast<SwitchInst>(Val: I); |
6154 | return Switch->getNumCases() * |
6155 | TTI.getCmpSelInstrCost( |
6156 | Opcode: Instruction::ICmp, |
6157 | ValTy: toVectorTy(Scalar: Switch->getCondition()->getType(), EC: VF), |
6158 | CondTy: toVectorTy(Scalar: Type::getInt1Ty(C&: I->getContext()), EC: VF), |
6159 | VecPred: CmpInst::ICMP_EQ, CostKind); |
6160 | } |
6161 | case Instruction::PHI: { |
6162 | auto *Phi = cast<PHINode>(Val: I); |
6163 | |
6164 | // First-order recurrences are replaced by vector shuffles inside the loop. |
6165 | if (VF.isVector() && Legal->isFixedOrderRecurrence(Phi)) { |
6166 | SmallVector<int> Mask(VF.getKnownMinValue()); |
6167 | std::iota(first: Mask.begin(), last: Mask.end(), value: VF.getKnownMinValue() - 1); |
6168 | return TTI.getShuffleCost(Kind: TargetTransformInfo::SK_Splice, |
6169 | DstTy: cast<VectorType>(Val: VectorTy), |
6170 | SrcTy: cast<VectorType>(Val: VectorTy), Mask, CostKind, |
6171 | Index: VF.getKnownMinValue() - 1); |
6172 | } |
6173 | |
6174 | // Phi nodes in non-header blocks (not inductions, reductions, etc.) are |
6175 | // converted into select instructions. We require N - 1 selects per phi |
6176 | // node, where N is the number of incoming values. |
6177 | if (VF.isVector() && Phi->getParent() != TheLoop->getHeader()) { |
6178 | Type *ResultTy = Phi->getType(); |
6179 | |
6180 | // All instructions in an Any-of reduction chain are narrowed to bool. |
6181 | // Check if that is the case for this phi node. |
6182 | auto * = cast_if_present<PHINode>( |
6183 | Val: find_singleton<User>(Range: Phi->users(), P: [this](User *U, bool) -> User * { |
6184 | auto *Phi = dyn_cast<PHINode>(Val: U); |
6185 | if (Phi && Phi->getParent() == TheLoop->getHeader()) |
6186 | return Phi; |
6187 | return nullptr; |
6188 | })); |
6189 | if (HeaderUser) { |
6190 | auto &ReductionVars = Legal->getReductionVars(); |
6191 | auto Iter = ReductionVars.find(Key: HeaderUser); |
6192 | if (Iter != ReductionVars.end() && |
6193 | RecurrenceDescriptor::isAnyOfRecurrenceKind( |
6194 | Kind: Iter->second.getRecurrenceKind())) |
6195 | ResultTy = Type::getInt1Ty(C&: Phi->getContext()); |
6196 | } |
6197 | return (Phi->getNumIncomingValues() - 1) * |
6198 | TTI.getCmpSelInstrCost( |
6199 | Opcode: Instruction::Select, ValTy: toVectorTy(Scalar: ResultTy, EC: VF), |
6200 | CondTy: toVectorTy(Scalar: Type::getInt1Ty(C&: Phi->getContext()), EC: VF), |
6201 | VecPred: CmpInst::BAD_ICMP_PREDICATE, CostKind); |
6202 | } |
6203 | |
6204 | // When tail folding with EVL, if the phi is part of an out of loop |
6205 | // reduction then it will be transformed into a wide vp_merge. |
6206 | if (VF.isVector() && foldTailWithEVL() && |
6207 | Legal->getReductionVars().contains(Key: Phi) && !isInLoopReduction(Phi)) { |
6208 | IntrinsicCostAttributes ICA( |
6209 | Intrinsic::vp_merge, toVectorTy(Scalar: Phi->getType(), EC: VF), |
6210 | {toVectorTy(Scalar: Type::getInt1Ty(C&: Phi->getContext()), EC: VF)}); |
6211 | return TTI.getIntrinsicInstrCost(ICA, CostKind); |
6212 | } |
6213 | |
6214 | return TTI.getCFInstrCost(Opcode: Instruction::PHI, CostKind); |
6215 | } |
6216 | case Instruction::UDiv: |
6217 | case Instruction::SDiv: |
6218 | case Instruction::URem: |
6219 | case Instruction::SRem: |
6220 | if (VF.isVector() && isPredicatedInst(I)) { |
6221 | const auto [ScalarCost, SafeDivisorCost] = getDivRemSpeculationCost(I, VF); |
6222 | return isDivRemScalarWithPredication(ScalarCost, SafeDivisorCost) ? |
6223 | ScalarCost : SafeDivisorCost; |
6224 | } |
6225 | // We've proven all lanes safe to speculate, fall through. |
6226 | [[fallthrough]]; |
6227 | case Instruction::Add: |
6228 | case Instruction::Sub: { |
6229 | auto Info = Legal->getHistogramInfo(I); |
6230 | if (Info && VF.isVector()) { |
6231 | const HistogramInfo *HGram = Info.value(); |
6232 | // Assume that a non-constant update value (or a constant != 1) requires |
6233 | // a multiply, and add that into the cost. |
6234 | InstructionCost MulCost = TTI::TCC_Free; |
6235 | ConstantInt *RHS = dyn_cast<ConstantInt>(Val: I->getOperand(i: 1)); |
6236 | if (!RHS || RHS->getZExtValue() != 1) |
6237 | MulCost = |
6238 | TTI.getArithmeticInstrCost(Opcode: Instruction::Mul, Ty: VectorTy, CostKind); |
6239 | |
6240 | // Find the cost of the histogram operation itself. |
6241 | Type *PtrTy = VectorType::get(ElementType: HGram->Load->getPointerOperandType(), EC: VF); |
6242 | Type *ScalarTy = I->getType(); |
6243 | Type *MaskTy = VectorType::get(ElementType: Type::getInt1Ty(C&: I->getContext()), EC: VF); |
6244 | IntrinsicCostAttributes ICA(Intrinsic::experimental_vector_histogram_add, |
6245 | Type::getVoidTy(C&: I->getContext()), |
6246 | {PtrTy, ScalarTy, MaskTy}); |
6247 | |
6248 | // Add the costs together with the add/sub operation. |
6249 | return TTI.getIntrinsicInstrCost(ICA, CostKind) + MulCost + |
6250 | TTI.getArithmeticInstrCost(Opcode: I->getOpcode(), Ty: VectorTy, CostKind); |
6251 | } |
6252 | [[fallthrough]]; |
6253 | } |
6254 | case Instruction::FAdd: |
6255 | case Instruction::FSub: |
6256 | case Instruction::Mul: |
6257 | case Instruction::FMul: |
6258 | case Instruction::FDiv: |
6259 | case Instruction::FRem: |
6260 | case Instruction::Shl: |
6261 | case Instruction::LShr: |
6262 | case Instruction::AShr: |
6263 | case Instruction::And: |
6264 | case Instruction::Or: |
6265 | case Instruction::Xor: { |
6266 | // If we're speculating on the stride being 1, the multiplication may |
6267 | // fold away. We can generalize this for all operations using the notion |
6268 | // of neutral elements. (TODO) |
6269 | if (I->getOpcode() == Instruction::Mul && |
6270 | ((TheLoop->isLoopInvariant(V: I->getOperand(i: 0)) && |
6271 | PSE.getSCEV(V: I->getOperand(i: 0))->isOne()) || |
6272 | (TheLoop->isLoopInvariant(V: I->getOperand(i: 1)) && |
6273 | PSE.getSCEV(V: I->getOperand(i: 1))->isOne()))) |
6274 | return 0; |
6275 | |
6276 | // Detect reduction patterns |
6277 | if (auto RedCost = getReductionPatternCost(I, VF, Ty: VectorTy)) |
6278 | return *RedCost; |
6279 | |
6280 | // Certain instructions can be cheaper to vectorize if they have a constant |
6281 | // second vector operand. One example of this are shifts on x86. |
6282 | Value *Op2 = I->getOperand(i: 1); |
6283 | if (!isa<Constant>(Val: Op2) && TheLoop->isLoopInvariant(V: Op2) && |
6284 | PSE.getSE()->isSCEVable(Ty: Op2->getType()) && |
6285 | isa<SCEVConstant>(Val: PSE.getSCEV(V: Op2))) { |
6286 | Op2 = cast<SCEVConstant>(Val: PSE.getSCEV(V: Op2))->getValue(); |
6287 | } |
6288 | auto Op2Info = TTI.getOperandInfo(V: Op2); |
6289 | if (Op2Info.Kind == TargetTransformInfo::OK_AnyValue && |
6290 | shouldConsiderInvariant(Op: Op2)) |
6291 | Op2Info.Kind = TargetTransformInfo::OK_UniformValue; |
6292 | |
6293 | SmallVector<const Value *, 4> Operands(I->operand_values()); |
6294 | return TTI.getArithmeticInstrCost( |
6295 | Opcode: I->getOpcode(), Ty: VectorTy, CostKind, |
6296 | Opd1Info: {.Kind: TargetTransformInfo::OK_AnyValue, .Properties: TargetTransformInfo::OP_None}, |
6297 | Opd2Info: Op2Info, Args: Operands, CxtI: I, TLibInfo: TLI); |
6298 | } |
6299 | case Instruction::FNeg: { |
6300 | return TTI.getArithmeticInstrCost( |
6301 | Opcode: I->getOpcode(), Ty: VectorTy, CostKind, |
6302 | Opd1Info: {.Kind: TargetTransformInfo::OK_AnyValue, .Properties: TargetTransformInfo::OP_None}, |
6303 | Opd2Info: {.Kind: TargetTransformInfo::OK_AnyValue, .Properties: TargetTransformInfo::OP_None}, |
6304 | Args: I->getOperand(i: 0), CxtI: I); |
6305 | } |
6306 | case Instruction::Select: { |
6307 | SelectInst *SI = cast<SelectInst>(Val: I); |
6308 | const SCEV *CondSCEV = SE->getSCEV(V: SI->getCondition()); |
6309 | bool ScalarCond = (SE->isLoopInvariant(S: CondSCEV, L: TheLoop)); |
6310 | |
6311 | const Value *Op0, *Op1; |
6312 | using namespace llvm::PatternMatch; |
6313 | if (!ScalarCond && (match(V: I, P: m_LogicalAnd(L: m_Value(V&: Op0), R: m_Value(V&: Op1))) || |
6314 | match(V: I, P: m_LogicalOr(L: m_Value(V&: Op0), R: m_Value(V&: Op1))))) { |
6315 | // select x, y, false --> x & y |
6316 | // select x, true, y --> x | y |
6317 | const auto [Op1VK, Op1VP] = TTI::getOperandInfo(V: Op0); |
6318 | const auto [Op2VK, Op2VP] = TTI::getOperandInfo(V: Op1); |
6319 | assert(Op0->getType()->getScalarSizeInBits() == 1 && |
6320 | Op1->getType()->getScalarSizeInBits() == 1); |
6321 | |
6322 | SmallVector<const Value *, 2> Operands{Op0, Op1}; |
6323 | return TTI.getArithmeticInstrCost( |
6324 | Opcode: match(V: I, P: m_LogicalOr()) ? Instruction::Or : Instruction::And, Ty: VectorTy, |
6325 | CostKind, Opd1Info: {.Kind: Op1VK, .Properties: Op1VP}, Opd2Info: {.Kind: Op2VK, .Properties: Op2VP}, Args: Operands, CxtI: I); |
6326 | } |
6327 | |
6328 | Type *CondTy = SI->getCondition()->getType(); |
6329 | if (!ScalarCond) |
6330 | CondTy = VectorType::get(ElementType: CondTy, EC: VF); |
6331 | |
6332 | CmpInst::Predicate Pred = CmpInst::BAD_ICMP_PREDICATE; |
6333 | if (auto *Cmp = dyn_cast<CmpInst>(Val: SI->getCondition())) |
6334 | Pred = Cmp->getPredicate(); |
6335 | return TTI.getCmpSelInstrCost(Opcode: I->getOpcode(), ValTy: VectorTy, CondTy, VecPred: Pred, |
6336 | CostKind, Op1Info: {.Kind: TTI::OK_AnyValue, .Properties: TTI::OP_None}, |
6337 | Op2Info: {.Kind: TTI::OK_AnyValue, .Properties: TTI::OP_None}, I); |
6338 | } |
6339 | case Instruction::ICmp: |
6340 | case Instruction::FCmp: { |
6341 | Type *ValTy = I->getOperand(i: 0)->getType(); |
6342 | |
6343 | if (canTruncateToMinimalBitwidth(I, VF)) { |
6344 | [[maybe_unused]] Instruction *Op0AsInstruction = |
6345 | dyn_cast<Instruction>(Val: I->getOperand(i: 0)); |
6346 | assert((!canTruncateToMinimalBitwidth(Op0AsInstruction, VF) || |
6347 | MinBWs[I] == MinBWs[Op0AsInstruction]) && |
6348 | "if both the operand and the compare are marked for " |
6349 | "truncation, they must have the same bitwidth" ); |
6350 | ValTy = IntegerType::get(C&: ValTy->getContext(), NumBits: MinBWs[I]); |
6351 | } |
6352 | |
6353 | VectorTy = toVectorTy(Scalar: ValTy, EC: VF); |
6354 | return TTI.getCmpSelInstrCost( |
6355 | Opcode: I->getOpcode(), ValTy: VectorTy, CondTy: CmpInst::makeCmpResultType(opnd_type: VectorTy), |
6356 | VecPred: cast<CmpInst>(Val: I)->getPredicate(), CostKind, |
6357 | Op1Info: {.Kind: TTI::OK_AnyValue, .Properties: TTI::OP_None}, Op2Info: {.Kind: TTI::OK_AnyValue, .Properties: TTI::OP_None}, I); |
6358 | } |
6359 | case Instruction::Store: |
6360 | case Instruction::Load: { |
6361 | ElementCount Width = VF; |
6362 | if (Width.isVector()) { |
6363 | InstWidening Decision = getWideningDecision(I, VF: Width); |
6364 | assert(Decision != CM_Unknown && |
6365 | "CM decision should be taken at this point" ); |
6366 | if (getWideningCost(I, VF) == InstructionCost::getInvalid()) |
6367 | return InstructionCost::getInvalid(); |
6368 | if (Decision == CM_Scalarize) |
6369 | Width = ElementCount::getFixed(MinVal: 1); |
6370 | } |
6371 | VectorTy = toVectorTy(Scalar: getLoadStoreType(I), EC: Width); |
6372 | return getMemoryInstructionCost(I, VF); |
6373 | } |
6374 | case Instruction::BitCast: |
6375 | if (I->getType()->isPointerTy()) |
6376 | return 0; |
6377 | [[fallthrough]]; |
6378 | case Instruction::ZExt: |
6379 | case Instruction::SExt: |
6380 | case Instruction::FPToUI: |
6381 | case Instruction::FPToSI: |
6382 | case Instruction::FPExt: |
6383 | case Instruction::PtrToInt: |
6384 | case Instruction::IntToPtr: |
6385 | case Instruction::SIToFP: |
6386 | case Instruction::UIToFP: |
6387 | case Instruction::Trunc: |
6388 | case Instruction::FPTrunc: { |
6389 | // Computes the CastContextHint from a Load/Store instruction. |
6390 | auto ComputeCCH = [&](Instruction *I) -> TTI::CastContextHint { |
6391 | assert((isa<LoadInst>(I) || isa<StoreInst>(I)) && |
6392 | "Expected a load or a store!" ); |
6393 | |
6394 | if (VF.isScalar() || !TheLoop->contains(Inst: I)) |
6395 | return TTI::CastContextHint::Normal; |
6396 | |
6397 | switch (getWideningDecision(I, VF)) { |
6398 | case LoopVectorizationCostModel::CM_GatherScatter: |
6399 | return TTI::CastContextHint::GatherScatter; |
6400 | case LoopVectorizationCostModel::CM_Interleave: |
6401 | return TTI::CastContextHint::Interleave; |
6402 | case LoopVectorizationCostModel::CM_Scalarize: |
6403 | case LoopVectorizationCostModel::CM_Widen: |
6404 | return Legal->isMaskRequired(I) ? TTI::CastContextHint::Masked |
6405 | : TTI::CastContextHint::Normal; |
6406 | case LoopVectorizationCostModel::CM_Widen_Reverse: |
6407 | return TTI::CastContextHint::Reversed; |
6408 | case LoopVectorizationCostModel::CM_Unknown: |
6409 | llvm_unreachable("Instr did not go through cost modelling?" ); |
6410 | case LoopVectorizationCostModel::CM_VectorCall: |
6411 | case LoopVectorizationCostModel::CM_IntrinsicCall: |
6412 | llvm_unreachable_internal(msg: "Instr has invalid widening decision" ); |
6413 | } |
6414 | |
6415 | llvm_unreachable("Unhandled case!" ); |
6416 | }; |
6417 | |
6418 | unsigned Opcode = I->getOpcode(); |
6419 | TTI::CastContextHint CCH = TTI::CastContextHint::None; |
6420 | // For Trunc, the context is the only user, which must be a StoreInst. |
6421 | if (Opcode == Instruction::Trunc || Opcode == Instruction::FPTrunc) { |
6422 | if (I->hasOneUse()) |
6423 | if (StoreInst *Store = dyn_cast<StoreInst>(Val: *I->user_begin())) |
6424 | CCH = ComputeCCH(Store); |
6425 | } |
6426 | // For Z/Sext, the context is the operand, which must be a LoadInst. |
6427 | else if (Opcode == Instruction::ZExt || Opcode == Instruction::SExt || |
6428 | Opcode == Instruction::FPExt) { |
6429 | if (LoadInst *Load = dyn_cast<LoadInst>(Val: I->getOperand(i: 0))) |
6430 | CCH = ComputeCCH(Load); |
6431 | } |
6432 | |
6433 | // We optimize the truncation of induction variables having constant |
6434 | // integer steps. The cost of these truncations is the same as the scalar |
6435 | // operation. |
6436 | if (isOptimizableIVTruncate(I, VF)) { |
6437 | auto *Trunc = cast<TruncInst>(Val: I); |
6438 | return TTI.getCastInstrCost(Opcode: Instruction::Trunc, Dst: Trunc->getDestTy(), |
6439 | Src: Trunc->getSrcTy(), CCH, CostKind, I: Trunc); |
6440 | } |
6441 | |
6442 | // Detect reduction patterns |
6443 | if (auto RedCost = getReductionPatternCost(I, VF, Ty: VectorTy)) |
6444 | return *RedCost; |
6445 | |
6446 | Type *SrcScalarTy = I->getOperand(i: 0)->getType(); |
6447 | Instruction *Op0AsInstruction = dyn_cast<Instruction>(Val: I->getOperand(i: 0)); |
6448 | if (canTruncateToMinimalBitwidth(I: Op0AsInstruction, VF)) |
6449 | SrcScalarTy = |
6450 | IntegerType::get(C&: SrcScalarTy->getContext(), NumBits: MinBWs[Op0AsInstruction]); |
6451 | Type *SrcVecTy = |
6452 | VectorTy->isVectorTy() ? toVectorTy(Scalar: SrcScalarTy, EC: VF) : SrcScalarTy; |
6453 | |
6454 | if (canTruncateToMinimalBitwidth(I, VF)) { |
6455 | // If the result type is <= the source type, there will be no extend |
6456 | // after truncating the users to the minimal required bitwidth. |
6457 | if (VectorTy->getScalarSizeInBits() <= SrcVecTy->getScalarSizeInBits() && |
6458 | (I->getOpcode() == Instruction::ZExt || |
6459 | I->getOpcode() == Instruction::SExt)) |
6460 | return 0; |
6461 | } |
6462 | |
6463 | return TTI.getCastInstrCost(Opcode, Dst: VectorTy, Src: SrcVecTy, CCH, CostKind, I); |
6464 | } |
6465 | case Instruction::Call: |
6466 | return getVectorCallCost(CI: cast<CallInst>(Val: I), VF); |
6467 | case Instruction::ExtractValue: |
6468 | return TTI.getInstructionCost(U: I, CostKind); |
6469 | case Instruction::Alloca: |
6470 | // We cannot easily widen alloca to a scalable alloca, as |
6471 | // the result would need to be a vector of pointers. |
6472 | if (VF.isScalable()) |
6473 | return InstructionCost::getInvalid(); |
6474 | [[fallthrough]]; |
6475 | default: |
6476 | // This opcode is unknown. Assume that it is the same as 'mul'. |
6477 | return TTI.getArithmeticInstrCost(Opcode: Instruction::Mul, Ty: VectorTy, CostKind); |
6478 | } // end of switch. |
6479 | } |
6480 | |
6481 | void LoopVectorizationCostModel::collectValuesToIgnore() { |
6482 | // Ignore ephemeral values. |
6483 | CodeMetrics::collectEphemeralValues(L: TheLoop, AC, EphValues&: ValuesToIgnore); |
6484 | |
6485 | SmallVector<Value *, 4> DeadInterleavePointerOps; |
6486 | SmallVector<Value *, 4> DeadOps; |
6487 | |
6488 | // If a scalar epilogue is required, users outside the loop won't use |
6489 | // live-outs from the vector loop but from the scalar epilogue. Ignore them if |
6490 | // that is the case. |
6491 | bool RequiresScalarEpilogue = requiresScalarEpilogue(IsVectorizing: true); |
6492 | auto IsLiveOutDead = [this, RequiresScalarEpilogue](User *U) { |
6493 | return RequiresScalarEpilogue && |
6494 | !TheLoop->contains(BB: cast<Instruction>(Val: U)->getParent()); |
6495 | }; |
6496 | |
6497 | LoopBlocksDFS DFS(TheLoop); |
6498 | DFS.perform(LI); |
6499 | MapVector<Value *, SmallVector<Value *>> DeadInvariantStoreOps; |
6500 | for (BasicBlock *BB : reverse(C: make_range(x: DFS.beginRPO(), y: DFS.endRPO()))) |
6501 | for (Instruction &I : reverse(C&: *BB)) { |
6502 | // Find all stores to invariant variables. Since they are going to sink |
6503 | // outside the loop we do not need calculate cost for them. |
6504 | StoreInst *SI; |
6505 | if ((SI = dyn_cast<StoreInst>(Val: &I)) && |
6506 | Legal->isInvariantAddressOfReduction(V: SI->getPointerOperand())) { |
6507 | ValuesToIgnore.insert(Ptr: &I); |
6508 | DeadInvariantStoreOps[SI->getPointerOperand()].push_back( |
6509 | Elt: SI->getValueOperand()); |
6510 | } |
6511 | |
6512 | if (VecValuesToIgnore.contains(Ptr: &I) || ValuesToIgnore.contains(Ptr: &I)) |
6513 | continue; |
6514 | |
6515 | // Add instructions that would be trivially dead and are only used by |
6516 | // values already ignored to DeadOps to seed worklist. |
6517 | if (wouldInstructionBeTriviallyDead(I: &I, TLI) && |
6518 | all_of(Range: I.users(), P: [this, IsLiveOutDead](User *U) { |
6519 | return VecValuesToIgnore.contains(Ptr: U) || |
6520 | ValuesToIgnore.contains(Ptr: U) || IsLiveOutDead(U); |
6521 | })) |
6522 | DeadOps.push_back(Elt: &I); |
6523 | |
6524 | // For interleave groups, we only create a pointer for the start of the |
6525 | // interleave group. Queue up addresses of group members except the insert |
6526 | // position for further processing. |
6527 | if (isAccessInterleaved(Instr: &I)) { |
6528 | auto *Group = getInterleavedAccessGroup(Instr: &I); |
6529 | if (Group->getInsertPos() == &I) |
6530 | continue; |
6531 | Value *PointerOp = getLoadStorePointerOperand(V: &I); |
6532 | DeadInterleavePointerOps.push_back(Elt: PointerOp); |
6533 | } |
6534 | |
6535 | // Queue branches for analysis. They are dead, if their successors only |
6536 | // contain dead instructions. |
6537 | if (auto *Br = dyn_cast<BranchInst>(Val: &I)) { |
6538 | if (Br->isConditional()) |
6539 | DeadOps.push_back(Elt: &I); |
6540 | } |
6541 | } |
6542 | |
6543 | // Mark ops feeding interleave group members as free, if they are only used |
6544 | // by other dead computations. |
6545 | for (unsigned I = 0; I != DeadInterleavePointerOps.size(); ++I) { |
6546 | auto *Op = dyn_cast<Instruction>(Val: DeadInterleavePointerOps[I]); |
6547 | if (!Op || !TheLoop->contains(Inst: Op) || any_of(Range: Op->users(), P: [this](User *U) { |
6548 | Instruction *UI = cast<Instruction>(Val: U); |
6549 | return !VecValuesToIgnore.contains(Ptr: U) && |
6550 | (!isAccessInterleaved(Instr: UI) || |
6551 | getInterleavedAccessGroup(Instr: UI)->getInsertPos() == UI); |
6552 | })) |
6553 | continue; |
6554 | VecValuesToIgnore.insert(Ptr: Op); |
6555 | DeadInterleavePointerOps.append(in_start: Op->op_begin(), in_end: Op->op_end()); |
6556 | } |
6557 | |
6558 | for (const auto &[_, Ops] : DeadInvariantStoreOps) |
6559 | llvm::append_range(C&: DeadOps, R: ArrayRef(Ops).drop_back()); |
6560 | |
6561 | // Mark ops that would be trivially dead and are only used by ignored |
6562 | // instructions as free. |
6563 | BasicBlock * = TheLoop->getHeader(); |
6564 | |
6565 | // Returns true if the block contains only dead instructions. Such blocks will |
6566 | // be removed by VPlan-to-VPlan transforms and won't be considered by the |
6567 | // VPlan-based cost model, so skip them in the legacy cost-model as well. |
6568 | auto IsEmptyBlock = [this](BasicBlock *BB) { |
6569 | return all_of(Range&: *BB, P: [this](Instruction &I) { |
6570 | return ValuesToIgnore.contains(Ptr: &I) || VecValuesToIgnore.contains(Ptr: &I) || |
6571 | (isa<BranchInst>(Val: &I) && !cast<BranchInst>(Val: &I)->isConditional()); |
6572 | }); |
6573 | }; |
6574 | for (unsigned I = 0; I != DeadOps.size(); ++I) { |
6575 | auto *Op = dyn_cast<Instruction>(Val: DeadOps[I]); |
6576 | |
6577 | // Check if the branch should be considered dead. |
6578 | if (auto *Br = dyn_cast_or_null<BranchInst>(Val: Op)) { |
6579 | BasicBlock *ThenBB = Br->getSuccessor(i: 0); |
6580 | BasicBlock *ElseBB = Br->getSuccessor(i: 1); |
6581 | // Don't considers branches leaving the loop for simplification. |
6582 | if (!TheLoop->contains(BB: ThenBB) || !TheLoop->contains(BB: ElseBB)) |
6583 | continue; |
6584 | bool ThenEmpty = IsEmptyBlock(ThenBB); |
6585 | bool ElseEmpty = IsEmptyBlock(ElseBB); |
6586 | if ((ThenEmpty && ElseEmpty) || |
6587 | (ThenEmpty && ThenBB->getSingleSuccessor() == ElseBB && |
6588 | ElseBB->phis().empty()) || |
6589 | (ElseEmpty && ElseBB->getSingleSuccessor() == ThenBB && |
6590 | ThenBB->phis().empty())) { |
6591 | VecValuesToIgnore.insert(Ptr: Br); |
6592 | DeadOps.push_back(Elt: Br->getCondition()); |
6593 | } |
6594 | continue; |
6595 | } |
6596 | |
6597 | // Skip any op that shouldn't be considered dead. |
6598 | if (!Op || !TheLoop->contains(Inst: Op) || |
6599 | (isa<PHINode>(Val: Op) && Op->getParent() == Header) || |
6600 | !wouldInstructionBeTriviallyDead(I: Op, TLI) || |
6601 | any_of(Range: Op->users(), P: [this, IsLiveOutDead](User *U) { |
6602 | return !VecValuesToIgnore.contains(Ptr: U) && |
6603 | !ValuesToIgnore.contains(Ptr: U) && !IsLiveOutDead(U); |
6604 | })) |
6605 | continue; |
6606 | |
6607 | // If all of Op's users are in ValuesToIgnore, add it to ValuesToIgnore |
6608 | // which applies for both scalar and vector versions. Otherwise it is only |
6609 | // dead in vector versions, so only add it to VecValuesToIgnore. |
6610 | if (all_of(Range: Op->users(), |
6611 | P: [this](User *U) { return ValuesToIgnore.contains(Ptr: U); })) |
6612 | ValuesToIgnore.insert(Ptr: Op); |
6613 | |
6614 | VecValuesToIgnore.insert(Ptr: Op); |
6615 | DeadOps.append(in_start: Op->op_begin(), in_end: Op->op_end()); |
6616 | } |
6617 | |
6618 | // Ignore type-promoting instructions we identified during reduction |
6619 | // detection. |
6620 | for (const auto &Reduction : Legal->getReductionVars()) { |
6621 | const RecurrenceDescriptor &RedDes = Reduction.second; |
6622 | const SmallPtrSetImpl<Instruction *> &Casts = RedDes.getCastInsts(); |
6623 | VecValuesToIgnore.insert_range(R: Casts); |
6624 | } |
6625 | // Ignore type-casting instructions we identified during induction |
6626 | // detection. |
6627 | for (const auto &Induction : Legal->getInductionVars()) { |
6628 | const InductionDescriptor &IndDes = Induction.second; |
6629 | const SmallVectorImpl<Instruction *> &Casts = IndDes.getCastInsts(); |
6630 | VecValuesToIgnore.insert_range(R: Casts); |
6631 | } |
6632 | } |
6633 | |
6634 | void LoopVectorizationCostModel::collectInLoopReductions() { |
6635 | // Avoid duplicating work finding in-loop reductions. |
6636 | if (!InLoopReductions.empty()) |
6637 | return; |
6638 | |
6639 | for (const auto &Reduction : Legal->getReductionVars()) { |
6640 | PHINode *Phi = Reduction.first; |
6641 | const RecurrenceDescriptor &RdxDesc = Reduction.second; |
6642 | |
6643 | // We don't collect reductions that are type promoted (yet). |
6644 | if (RdxDesc.getRecurrenceType() != Phi->getType()) |
6645 | continue; |
6646 | |
6647 | // If the target would prefer this reduction to happen "in-loop", then we |
6648 | // want to record it as such. |
6649 | RecurKind Kind = RdxDesc.getRecurrenceKind(); |
6650 | if (!PreferInLoopReductions && !useOrderedReductions(RdxDesc) && |
6651 | !TTI.preferInLoopReduction(Kind, Ty: Phi->getType())) |
6652 | continue; |
6653 | |
6654 | // Check that we can correctly put the reductions into the loop, by |
6655 | // finding the chain of operations that leads from the phi to the loop |
6656 | // exit value. |
6657 | SmallVector<Instruction *, 4> ReductionOperations = |
6658 | RdxDesc.getReductionOpChain(Phi, L: TheLoop); |
6659 | bool InLoop = !ReductionOperations.empty(); |
6660 | |
6661 | if (InLoop) { |
6662 | InLoopReductions.insert(Ptr: Phi); |
6663 | // Add the elements to InLoopReductionImmediateChains for cost modelling. |
6664 | Instruction *LastChain = Phi; |
6665 | for (auto *I : ReductionOperations) { |
6666 | InLoopReductionImmediateChains[I] = LastChain; |
6667 | LastChain = I; |
6668 | } |
6669 | } |
6670 | LLVM_DEBUG(dbgs() << "LV: Using " << (InLoop ? "inloop" : "out of loop" ) |
6671 | << " reduction for phi: " << *Phi << "\n" ); |
6672 | } |
6673 | } |
6674 | |
6675 | // This function will select a scalable VF if the target supports scalable |
6676 | // vectors and a fixed one otherwise. |
6677 | // TODO: we could return a pair of values that specify the max VF and |
6678 | // min VF, to be used in `buildVPlans(MinVF, MaxVF)` instead of |
6679 | // `buildVPlans(VF, VF)`. We cannot do it because VPLAN at the moment |
6680 | // doesn't have a cost model that can choose which plan to execute if |
6681 | // more than one is generated. |
6682 | static ElementCount determineVPlanVF(const TargetTransformInfo &TTI, |
6683 | LoopVectorizationCostModel &CM) { |
6684 | unsigned WidestType; |
6685 | std::tie(args: std::ignore, args&: WidestType) = CM.getSmallestAndWidestTypes(); |
6686 | |
6687 | TargetTransformInfo::RegisterKind RegKind = |
6688 | TTI.enableScalableVectorization() |
6689 | ? TargetTransformInfo::RGK_ScalableVector |
6690 | : TargetTransformInfo::RGK_FixedWidthVector; |
6691 | |
6692 | TypeSize RegSize = TTI.getRegisterBitWidth(K: RegKind); |
6693 | unsigned N = RegSize.getKnownMinValue() / WidestType; |
6694 | return ElementCount::get(MinVal: N, Scalable: RegSize.isScalable()); |
6695 | } |
6696 | |
6697 | VectorizationFactor |
6698 | LoopVectorizationPlanner::planInVPlanNativePath(ElementCount UserVF) { |
6699 | ElementCount VF = UserVF; |
6700 | // Outer loop handling: They may require CFG and instruction level |
6701 | // transformations before even evaluating whether vectorization is profitable. |
6702 | // Since we cannot modify the incoming IR, we need to build VPlan upfront in |
6703 | // the vectorization pipeline. |
6704 | if (!OrigLoop->isInnermost()) { |
6705 | // If the user doesn't provide a vectorization factor, determine a |
6706 | // reasonable one. |
6707 | if (UserVF.isZero()) { |
6708 | VF = determineVPlanVF(TTI, CM); |
6709 | LLVM_DEBUG(dbgs() << "LV: VPlan computed VF " << VF << ".\n" ); |
6710 | |
6711 | // Make sure we have a VF > 1 for stress testing. |
6712 | if (VPlanBuildStressTest && (VF.isScalar() || VF.isZero())) { |
6713 | LLVM_DEBUG(dbgs() << "LV: VPlan stress testing: " |
6714 | << "overriding computed VF.\n" ); |
6715 | VF = ElementCount::getFixed(MinVal: 4); |
6716 | } |
6717 | } else if (UserVF.isScalable() && !TTI.supportsScalableVectors() && |
6718 | !ForceTargetSupportsScalableVectors) { |
6719 | LLVM_DEBUG(dbgs() << "LV: Not vectorizing. Scalable VF requested, but " |
6720 | << "not supported by the target.\n" ); |
6721 | reportVectorizationFailure( |
6722 | DebugMsg: "Scalable vectorization requested but not supported by the target" , |
6723 | OREMsg: "the scalable user-specified vectorization width for outer-loop " |
6724 | "vectorization cannot be used because the target does not support " |
6725 | "scalable vectors." , |
6726 | ORETag: "ScalableVFUnfeasible" , ORE, TheLoop: OrigLoop); |
6727 | return VectorizationFactor::Disabled(); |
6728 | } |
6729 | assert(EnableVPlanNativePath && "VPlan-native path is not enabled." ); |
6730 | assert(isPowerOf2_32(VF.getKnownMinValue()) && |
6731 | "VF needs to be a power of two" ); |
6732 | LLVM_DEBUG(dbgs() << "LV: Using " << (!UserVF.isZero() ? "user " : "" ) |
6733 | << "VF " << VF << " to build VPlans.\n" ); |
6734 | buildVPlans(MinVF: VF, MaxVF: VF); |
6735 | |
6736 | if (VPlans.empty()) |
6737 | return VectorizationFactor::Disabled(); |
6738 | |
6739 | // For VPlan build stress testing, we bail out after VPlan construction. |
6740 | if (VPlanBuildStressTest) |
6741 | return VectorizationFactor::Disabled(); |
6742 | |
6743 | return {VF, 0 /*Cost*/, 0 /* ScalarCost */}; |
6744 | } |
6745 | |
6746 | LLVM_DEBUG( |
6747 | dbgs() << "LV: Not vectorizing. Inner loops aren't supported in the " |
6748 | "VPlan-native path.\n" ); |
6749 | return VectorizationFactor::Disabled(); |
6750 | } |
6751 | |
6752 | void LoopVectorizationPlanner::plan(ElementCount UserVF, unsigned UserIC) { |
6753 | assert(OrigLoop->isInnermost() && "Inner loop expected." ); |
6754 | CM.collectValuesToIgnore(); |
6755 | CM.collectElementTypesForWidening(); |
6756 | |
6757 | FixedScalableVFPair MaxFactors = CM.computeMaxVF(UserVF, UserIC); |
6758 | if (!MaxFactors) // Cases that should not to be vectorized nor interleaved. |
6759 | return; |
6760 | |
6761 | // Invalidate interleave groups if all blocks of loop will be predicated. |
6762 | if (CM.blockNeedsPredicationForAnyReason(BB: OrigLoop->getHeader()) && |
6763 | !useMaskedInterleavedAccesses(TTI)) { |
6764 | LLVM_DEBUG( |
6765 | dbgs() |
6766 | << "LV: Invalidate all interleaved groups due to fold-tail by masking " |
6767 | "which requires masked-interleaved support.\n" ); |
6768 | if (CM.InterleaveInfo.invalidateGroups()) |
6769 | // Invalidating interleave groups also requires invalidating all decisions |
6770 | // based on them, which includes widening decisions and uniform and scalar |
6771 | // values. |
6772 | CM.invalidateCostModelingDecisions(); |
6773 | } |
6774 | |
6775 | if (CM.foldTailByMasking()) |
6776 | Legal->prepareToFoldTailByMasking(); |
6777 | |
6778 | ElementCount MaxUserVF = |
6779 | UserVF.isScalable() ? MaxFactors.ScalableVF : MaxFactors.FixedVF; |
6780 | if (UserVF) { |
6781 | if (!ElementCount::isKnownLE(LHS: UserVF, RHS: MaxUserVF)) { |
6782 | reportVectorizationInfo( |
6783 | Msg: "UserVF ignored because it may be larger than the maximal safe VF" , |
6784 | ORETag: "InvalidUserVF" , ORE, TheLoop: OrigLoop); |
6785 | } else { |
6786 | assert(isPowerOf2_32(UserVF.getKnownMinValue()) && |
6787 | "VF needs to be a power of two" ); |
6788 | // Collect the instructions (and their associated costs) that will be more |
6789 | // profitable to scalarize. |
6790 | CM.collectInLoopReductions(); |
6791 | if (CM.selectUserVectorizationFactor(UserVF)) { |
6792 | LLVM_DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n" ); |
6793 | buildVPlansWithVPRecipes(MinVF: UserVF, MaxVF: UserVF); |
6794 | LLVM_DEBUG(printPlans(dbgs())); |
6795 | return; |
6796 | } |
6797 | reportVectorizationInfo(Msg: "UserVF ignored because of invalid costs." , |
6798 | ORETag: "InvalidCost" , ORE, TheLoop: OrigLoop); |
6799 | } |
6800 | } |
6801 | |
6802 | // Collect the Vectorization Factor Candidates. |
6803 | SmallVector<ElementCount> VFCandidates; |
6804 | for (auto VF = ElementCount::getFixed(MinVal: 1); |
6805 | ElementCount::isKnownLE(LHS: VF, RHS: MaxFactors.FixedVF); VF *= 2) |
6806 | VFCandidates.push_back(Elt: VF); |
6807 | for (auto VF = ElementCount::getScalable(MinVal: 1); |
6808 | ElementCount::isKnownLE(LHS: VF, RHS: MaxFactors.ScalableVF); VF *= 2) |
6809 | VFCandidates.push_back(Elt: VF); |
6810 | |
6811 | CM.collectInLoopReductions(); |
6812 | for (const auto &VF : VFCandidates) { |
6813 | // Collect Uniform and Scalar instructions after vectorization with VF. |
6814 | CM.collectNonVectorizedAndSetWideningDecisions(VF); |
6815 | } |
6816 | |
6817 | buildVPlansWithVPRecipes(MinVF: ElementCount::getFixed(MinVal: 1), MaxVF: MaxFactors.FixedVF); |
6818 | buildVPlansWithVPRecipes(MinVF: ElementCount::getScalable(MinVal: 1), MaxVF: MaxFactors.ScalableVF); |
6819 | |
6820 | LLVM_DEBUG(printPlans(dbgs())); |
6821 | } |
6822 | |
6823 | InstructionCost VPCostContext::getLegacyCost(Instruction *UI, |
6824 | ElementCount VF) const { |
6825 | if (ForceTargetInstructionCost.getNumOccurrences()) |
6826 | return InstructionCost(ForceTargetInstructionCost.getNumOccurrences()); |
6827 | return CM.getInstructionCost(I: UI, VF); |
6828 | } |
6829 | |
6830 | bool VPCostContext::isLegacyUniformAfterVectorization(Instruction *I, |
6831 | ElementCount VF) const { |
6832 | return CM.isUniformAfterVectorization(I, VF); |
6833 | } |
6834 | |
6835 | bool VPCostContext::skipCostComputation(Instruction *UI, bool IsVector) const { |
6836 | return CM.ValuesToIgnore.contains(Ptr: UI) || |
6837 | (IsVector && CM.VecValuesToIgnore.contains(Ptr: UI)) || |
6838 | SkipCostComputation.contains(Ptr: UI); |
6839 | } |
6840 | |
6841 | InstructionCost |
6842 | LoopVectorizationPlanner::precomputeCosts(VPlan &Plan, ElementCount VF, |
6843 | VPCostContext &CostCtx) const { |
6844 | InstructionCost Cost; |
6845 | // Cost modeling for inductions is inaccurate in the legacy cost model |
6846 | // compared to the recipes that are generated. To match here initially during |
6847 | // VPlan cost model bring up directly use the induction costs from the legacy |
6848 | // cost model. Note that we do this as pre-processing; the VPlan may not have |
6849 | // any recipes associated with the original induction increment instruction |
6850 | // and may replace truncates with VPWidenIntOrFpInductionRecipe. We precompute |
6851 | // the cost of induction phis and increments (both that are represented by |
6852 | // recipes and those that are not), to avoid distinguishing between them here, |
6853 | // and skip all recipes that represent induction phis and increments (the |
6854 | // former case) later on, if they exist, to avoid counting them twice. |
6855 | // Similarly we pre-compute the cost of any optimized truncates. |
6856 | // TODO: Switch to more accurate costing based on VPlan. |
6857 | for (const auto &[IV, IndDesc] : Legal->getInductionVars()) { |
6858 | Instruction *IVInc = cast<Instruction>( |
6859 | Val: IV->getIncomingValueForBlock(BB: OrigLoop->getLoopLatch())); |
6860 | SmallVector<Instruction *> IVInsts = {IVInc}; |
6861 | for (unsigned I = 0; I != IVInsts.size(); I++) { |
6862 | for (Value *Op : IVInsts[I]->operands()) { |
6863 | auto *OpI = dyn_cast<Instruction>(Val: Op); |
6864 | if (Op == IV || !OpI || !OrigLoop->contains(Inst: OpI) || !Op->hasOneUse()) |
6865 | continue; |
6866 | IVInsts.push_back(Elt: OpI); |
6867 | } |
6868 | } |
6869 | IVInsts.push_back(Elt: IV); |
6870 | for (User *U : IV->users()) { |
6871 | auto *CI = cast<Instruction>(Val: U); |
6872 | if (!CostCtx.CM.isOptimizableIVTruncate(I: CI, VF)) |
6873 | continue; |
6874 | IVInsts.push_back(Elt: CI); |
6875 | } |
6876 | |
6877 | // If the vector loop gets executed exactly once with the given VF, ignore |
6878 | // the costs of comparison and induction instructions, as they'll get |
6879 | // simplified away. |
6880 | // TODO: Remove this code after stepping away from the legacy cost model and |
6881 | // adding code to simplify VPlans before calculating their costs. |
6882 | auto TC = getSmallConstantTripCount(SE: PSE.getSE(), L: OrigLoop); |
6883 | if (TC == VF && !CM.foldTailByMasking()) |
6884 | addFullyUnrolledInstructionsToIgnore(L: OrigLoop, IL: Legal->getInductionVars(), |
6885 | InstsToIgnore&: CostCtx.SkipCostComputation); |
6886 | |
6887 | for (Instruction *IVInst : IVInsts) { |
6888 | if (CostCtx.skipCostComputation(UI: IVInst, IsVector: VF.isVector())) |
6889 | continue; |
6890 | InstructionCost InductionCost = CostCtx.getLegacyCost(UI: IVInst, VF); |
6891 | LLVM_DEBUG({ |
6892 | dbgs() << "Cost of " << InductionCost << " for VF " << VF |
6893 | << ": induction instruction " << *IVInst << "\n" ; |
6894 | }); |
6895 | Cost += InductionCost; |
6896 | CostCtx.SkipCostComputation.insert(Ptr: IVInst); |
6897 | } |
6898 | } |
6899 | |
6900 | /// Compute the cost of all exiting conditions of the loop using the legacy |
6901 | /// cost model. This is to match the legacy behavior, which adds the cost of |
6902 | /// all exit conditions. Note that this over-estimates the cost, as there will |
6903 | /// be a single condition to control the vector loop. |
6904 | SmallVector<BasicBlock *> Exiting; |
6905 | CM.TheLoop->getExitingBlocks(ExitingBlocks&: Exiting); |
6906 | SetVector<Instruction *> ExitInstrs; |
6907 | // Collect all exit conditions. |
6908 | for (BasicBlock *EB : Exiting) { |
6909 | auto *Term = dyn_cast<BranchInst>(Val: EB->getTerminator()); |
6910 | if (!Term || CostCtx.skipCostComputation(UI: Term, IsVector: VF.isVector())) |
6911 | continue; |
6912 | if (auto *CondI = dyn_cast<Instruction>(Val: Term->getOperand(i_nocapture: 0))) { |
6913 | ExitInstrs.insert(X: CondI); |
6914 | } |
6915 | } |
6916 | // Compute the cost of all instructions only feeding the exit conditions. |
6917 | for (unsigned I = 0; I != ExitInstrs.size(); ++I) { |
6918 | Instruction *CondI = ExitInstrs[I]; |
6919 | if (!OrigLoop->contains(Inst: CondI) || |
6920 | !CostCtx.SkipCostComputation.insert(Ptr: CondI).second) |
6921 | continue; |
6922 | InstructionCost CondICost = CostCtx.getLegacyCost(UI: CondI, VF); |
6923 | LLVM_DEBUG({ |
6924 | dbgs() << "Cost of " << CondICost << " for VF " << VF |
6925 | << ": exit condition instruction " << *CondI << "\n" ; |
6926 | }); |
6927 | Cost += CondICost; |
6928 | for (Value *Op : CondI->operands()) { |
6929 | auto *OpI = dyn_cast<Instruction>(Val: Op); |
6930 | if (!OpI || CostCtx.skipCostComputation(UI: OpI, IsVector: VF.isVector()) || |
6931 | any_of(Range: OpI->users(), P: [&ExitInstrs, this](User *U) { |
6932 | return OrigLoop->contains(BB: cast<Instruction>(Val: U)->getParent()) && |
6933 | !ExitInstrs.contains(key: cast<Instruction>(Val: U)); |
6934 | })) |
6935 | continue; |
6936 | ExitInstrs.insert(X: OpI); |
6937 | } |
6938 | } |
6939 | |
6940 | // Pre-compute the costs for branches except for the backedge, as the number |
6941 | // of replicate regions in a VPlan may not directly match the number of |
6942 | // branches, which would lead to different decisions. |
6943 | // TODO: Compute cost of branches for each replicate region in the VPlan, |
6944 | // which is more accurate than the legacy cost model. |
6945 | for (BasicBlock *BB : OrigLoop->blocks()) { |
6946 | if (CostCtx.skipCostComputation(UI: BB->getTerminator(), IsVector: VF.isVector())) |
6947 | continue; |
6948 | CostCtx.SkipCostComputation.insert(Ptr: BB->getTerminator()); |
6949 | if (BB == OrigLoop->getLoopLatch()) |
6950 | continue; |
6951 | auto BranchCost = CostCtx.getLegacyCost(UI: BB->getTerminator(), VF); |
6952 | Cost += BranchCost; |
6953 | } |
6954 | |
6955 | // Pre-compute costs for instructions that are forced-scalar or profitable to |
6956 | // scalarize. Their costs will be computed separately in the legacy cost |
6957 | // model. |
6958 | for (Instruction *ForcedScalar : CM.ForcedScalars[VF]) { |
6959 | if (CostCtx.skipCostComputation(UI: ForcedScalar, IsVector: VF.isVector())) |
6960 | continue; |
6961 | CostCtx.SkipCostComputation.insert(Ptr: ForcedScalar); |
6962 | InstructionCost ForcedCost = CostCtx.getLegacyCost(UI: ForcedScalar, VF); |
6963 | LLVM_DEBUG({ |
6964 | dbgs() << "Cost of " << ForcedCost << " for VF " << VF |
6965 | << ": forced scalar " << *ForcedScalar << "\n" ; |
6966 | }); |
6967 | Cost += ForcedCost; |
6968 | } |
6969 | for (const auto &[Scalarized, ScalarCost] : CM.InstsToScalarize[VF]) { |
6970 | if (CostCtx.skipCostComputation(UI: Scalarized, IsVector: VF.isVector())) |
6971 | continue; |
6972 | CostCtx.SkipCostComputation.insert(Ptr: Scalarized); |
6973 | LLVM_DEBUG({ |
6974 | dbgs() << "Cost of " << ScalarCost << " for VF " << VF |
6975 | << ": profitable to scalarize " << *Scalarized << "\n" ; |
6976 | }); |
6977 | Cost += ScalarCost; |
6978 | } |
6979 | |
6980 | return Cost; |
6981 | } |
6982 | |
6983 | InstructionCost LoopVectorizationPlanner::cost(VPlan &Plan, |
6984 | ElementCount VF) const { |
6985 | VPCostContext CostCtx(CM.TTI, *CM.TLI, Legal->getWidestInductionType(), CM, |
6986 | CM.CostKind); |
6987 | InstructionCost Cost = precomputeCosts(Plan, VF, CostCtx); |
6988 | |
6989 | // Now compute and add the VPlan-based cost. |
6990 | Cost += Plan.cost(VF, Ctx&: CostCtx); |
6991 | #ifndef NDEBUG |
6992 | unsigned EstimatedWidth = getEstimatedRuntimeVF(VF, CM.getVScaleForTuning()); |
6993 | LLVM_DEBUG(dbgs() << "Cost for VF " << VF << ": " << Cost |
6994 | << " (Estimated cost per lane: " ); |
6995 | if (Cost.isValid()) { |
6996 | double CostPerLane = double(Cost.getValue()) / EstimatedWidth; |
6997 | LLVM_DEBUG(dbgs() << format("%.1f" , CostPerLane)); |
6998 | } else /* No point dividing an invalid cost - it will still be invalid */ |
6999 | LLVM_DEBUG(dbgs() << "Invalid" ); |
7000 | LLVM_DEBUG(dbgs() << ")\n" ); |
7001 | #endif |
7002 | return Cost; |
7003 | } |
7004 | |
7005 | #ifndef NDEBUG |
7006 | /// Return true if the original loop \ TheLoop contains any instructions that do |
7007 | /// not have corresponding recipes in \p Plan and are not marked to be ignored |
7008 | /// in \p CostCtx. This means the VPlan contains simplification that the legacy |
7009 | /// cost-model did not account for. |
7010 | static bool planContainsAdditionalSimplifications(VPlan &Plan, |
7011 | VPCostContext &CostCtx, |
7012 | Loop *TheLoop, |
7013 | ElementCount VF) { |
7014 | // First collect all instructions for the recipes in Plan. |
7015 | auto GetInstructionForCost = [](const VPRecipeBase *R) -> Instruction * { |
7016 | if (auto *S = dyn_cast<VPSingleDefRecipe>(R)) |
7017 | return dyn_cast_or_null<Instruction>(S->getUnderlyingValue()); |
7018 | if (auto *WidenMem = dyn_cast<VPWidenMemoryRecipe>(R)) |
7019 | return &WidenMem->getIngredient(); |
7020 | return nullptr; |
7021 | }; |
7022 | |
7023 | DenseSet<Instruction *> SeenInstrs; |
7024 | auto Iter = vp_depth_first_deep(Plan.getVectorLoopRegion()->getEntry()); |
7025 | for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(Iter)) { |
7026 | for (VPRecipeBase &R : *VPBB) { |
7027 | if (auto *IR = dyn_cast<VPInterleaveRecipe>(&R)) { |
7028 | auto *IG = IR->getInterleaveGroup(); |
7029 | unsigned NumMembers = IG->getNumMembers(); |
7030 | for (unsigned I = 0; I != NumMembers; ++I) { |
7031 | if (Instruction *M = IG->getMember(I)) |
7032 | SeenInstrs.insert(M); |
7033 | } |
7034 | continue; |
7035 | } |
7036 | // Unused FOR splices are removed by VPlan transforms, so the VPlan-based |
7037 | // cost model won't cost it whilst the legacy will. |
7038 | if (auto *FOR = dyn_cast<VPFirstOrderRecurrencePHIRecipe>(&R)) { |
7039 | if (none_of(FOR->users(), [](VPUser *U) { |
7040 | auto *VPI = dyn_cast<VPInstruction>(U); |
7041 | return VPI && VPI->getOpcode() == |
7042 | VPInstruction::FirstOrderRecurrenceSplice; |
7043 | })) |
7044 | return true; |
7045 | } |
7046 | // The VPlan-based cost model is more accurate for partial reduction and |
7047 | // comparing against the legacy cost isn't desirable. |
7048 | if (isa<VPPartialReductionRecipe>(&R)) |
7049 | return true; |
7050 | |
7051 | /// If a VPlan transform folded a recipe to one producing a single-scalar, |
7052 | /// but the original instruction wasn't uniform-after-vectorization in the |
7053 | /// legacy cost model, the legacy cost overestimates the actual cost. |
7054 | if (auto *RepR = dyn_cast<VPReplicateRecipe>(&R)) { |
7055 | if (RepR->isSingleScalar() && |
7056 | !CostCtx.isLegacyUniformAfterVectorization( |
7057 | RepR->getUnderlyingInstr(), VF)) |
7058 | return true; |
7059 | } |
7060 | if (Instruction *UI = GetInstructionForCost(&R)) { |
7061 | // If we adjusted the predicate of the recipe, the cost in the legacy |
7062 | // cost model may be different. |
7063 | if (auto *WidenCmp = dyn_cast<VPWidenRecipe>(&R)) { |
7064 | if ((WidenCmp->getOpcode() == Instruction::ICmp || |
7065 | WidenCmp->getOpcode() == Instruction::FCmp) && |
7066 | WidenCmp->getPredicate() != cast<CmpInst>(UI)->getPredicate()) |
7067 | return true; |
7068 | } |
7069 | SeenInstrs.insert(UI); |
7070 | } |
7071 | } |
7072 | } |
7073 | |
7074 | // Return true if the loop contains any instructions that are not also part of |
7075 | // the VPlan or are skipped for VPlan-based cost computations. This indicates |
7076 | // that the VPlan contains extra simplifications. |
7077 | return any_of(TheLoop->blocks(), [&SeenInstrs, &CostCtx, |
7078 | TheLoop](BasicBlock *BB) { |
7079 | return any_of(*BB, [&SeenInstrs, &CostCtx, TheLoop, BB](Instruction &I) { |
7080 | // Skip induction phis when checking for simplifications, as they may not |
7081 | // be lowered directly be lowered to a corresponding PHI recipe. |
7082 | if (isa<PHINode>(&I) && BB == TheLoop->getHeader() && |
7083 | CostCtx.CM.Legal->isInductionPhi(cast<PHINode>(&I))) |
7084 | return false; |
7085 | return !SeenInstrs.contains(&I) && !CostCtx.skipCostComputation(&I, true); |
7086 | }); |
7087 | }); |
7088 | } |
7089 | #endif |
7090 | |
7091 | VectorizationFactor LoopVectorizationPlanner::computeBestVF() { |
7092 | if (VPlans.empty()) |
7093 | return VectorizationFactor::Disabled(); |
7094 | // If there is a single VPlan with a single VF, return it directly. |
7095 | VPlan &FirstPlan = *VPlans[0]; |
7096 | if (VPlans.size() == 1 && size(Range: FirstPlan.vectorFactors()) == 1) |
7097 | return {*FirstPlan.vectorFactors().begin(), 0, 0}; |
7098 | |
7099 | LLVM_DEBUG(dbgs() << "LV: Computing best VF using cost kind: " |
7100 | << (CM.CostKind == TTI::TCK_RecipThroughput |
7101 | ? "Reciprocal Throughput\n" |
7102 | : CM.CostKind == TTI::TCK_Latency |
7103 | ? "Instruction Latency\n" |
7104 | : CM.CostKind == TTI::TCK_CodeSize ? "Code Size\n" |
7105 | : CM.CostKind == TTI::TCK_SizeAndLatency |
7106 | ? "Code Size and Latency\n" |
7107 | : "Unknown\n" )); |
7108 | |
7109 | ElementCount ScalarVF = ElementCount::getFixed(MinVal: 1); |
7110 | assert(hasPlanWithVF(ScalarVF) && |
7111 | "More than a single plan/VF w/o any plan having scalar VF" ); |
7112 | |
7113 | // TODO: Compute scalar cost using VPlan-based cost model. |
7114 | InstructionCost ScalarCost = CM.expectedCost(VF: ScalarVF); |
7115 | LLVM_DEBUG(dbgs() << "LV: Scalar loop costs: " << ScalarCost << ".\n" ); |
7116 | VectorizationFactor ScalarFactor(ScalarVF, ScalarCost, ScalarCost); |
7117 | VectorizationFactor BestFactor = ScalarFactor; |
7118 | |
7119 | bool ForceVectorization = Hints.getForce() == LoopVectorizeHints::FK_Enabled; |
7120 | if (ForceVectorization) { |
7121 | // Ignore scalar width, because the user explicitly wants vectorization. |
7122 | // Initialize cost to max so that VF = 2 is, at least, chosen during cost |
7123 | // evaluation. |
7124 | BestFactor.Cost = InstructionCost::getMax(); |
7125 | } |
7126 | |
7127 | for (auto &P : VPlans) { |
7128 | ArrayRef<ElementCount> VFs(P->vectorFactors().begin(), |
7129 | P->vectorFactors().end()); |
7130 | |
7131 | SmallVector<VPRegisterUsage, 8> RUs; |
7132 | if (CM.useMaxBandwidth(RegKind: TargetTransformInfo::RGK_ScalableVector) || |
7133 | CM.useMaxBandwidth(RegKind: TargetTransformInfo::RGK_FixedWidthVector)) |
7134 | RUs = calculateRegisterUsageForPlan(Plan&: *P, VFs, TTI, ValuesToIgnore: CM.ValuesToIgnore); |
7135 | |
7136 | for (unsigned I = 0; I < VFs.size(); I++) { |
7137 | ElementCount VF = VFs[I]; |
7138 | if (VF.isScalar()) |
7139 | continue; |
7140 | if (!ForceVectorization && !willGenerateVectors(Plan&: *P, VF, TTI)) { |
7141 | LLVM_DEBUG( |
7142 | dbgs() |
7143 | << "LV: Not considering vector loop of width " << VF |
7144 | << " because it will not generate any vector instructions.\n" ); |
7145 | continue; |
7146 | } |
7147 | if (CM.OptForSize && !ForceVectorization && hasReplicatorRegion(Plan&: *P)) { |
7148 | LLVM_DEBUG( |
7149 | dbgs() |
7150 | << "LV: Not considering vector loop of width " << VF |
7151 | << " because it would cause replicated blocks to be generated," |
7152 | << " which isn't allowed when optimizing for size.\n" ); |
7153 | continue; |
7154 | } |
7155 | |
7156 | InstructionCost Cost = cost(Plan&: *P, VF); |
7157 | VectorizationFactor CurrentFactor(VF, Cost, ScalarCost); |
7158 | |
7159 | if (CM.useMaxBandwidth(VF) && RUs[I].exceedsMaxNumRegs(TTI)) { |
7160 | LLVM_DEBUG(dbgs() << "LV(REG): Not considering vector loop of width " |
7161 | << VF << " because it uses too many registers\n" ); |
7162 | continue; |
7163 | } |
7164 | |
7165 | if (isMoreProfitable(A: CurrentFactor, B: BestFactor, HasTail: P->hasScalarTail())) |
7166 | BestFactor = CurrentFactor; |
7167 | |
7168 | // If profitable add it to ProfitableVF list. |
7169 | if (isMoreProfitable(A: CurrentFactor, B: ScalarFactor, HasTail: P->hasScalarTail())) |
7170 | ProfitableVFs.push_back(Elt: CurrentFactor); |
7171 | } |
7172 | } |
7173 | |
7174 | #ifndef NDEBUG |
7175 | // Select the optimal vectorization factor according to the legacy cost-model. |
7176 | // This is now only used to verify the decisions by the new VPlan-based |
7177 | // cost-model and will be retired once the VPlan-based cost-model is |
7178 | // stabilized. |
7179 | VectorizationFactor LegacyVF = selectVectorizationFactor(); |
7180 | VPlan &BestPlan = getPlanFor(BestFactor.Width); |
7181 | |
7182 | // Pre-compute the cost and use it to check if BestPlan contains any |
7183 | // simplifications not accounted for in the legacy cost model. If that's the |
7184 | // case, don't trigger the assertion, as the extra simplifications may cause a |
7185 | // different VF to be picked by the VPlan-based cost model. |
7186 | VPCostContext CostCtx(CM.TTI, *CM.TLI, Legal->getWidestInductionType(), CM, |
7187 | CM.CostKind); |
7188 | precomputeCosts(BestPlan, BestFactor.Width, CostCtx); |
7189 | // Verify that the VPlan-based and legacy cost models agree, except for VPlans |
7190 | // with early exits and plans with additional VPlan simplifications. The |
7191 | // legacy cost model doesn't properly model costs for such loops. |
7192 | assert((BestFactor.Width == LegacyVF.Width || BestPlan.hasEarlyExit() || |
7193 | planContainsAdditionalSimplifications(getPlanFor(BestFactor.Width), |
7194 | CostCtx, OrigLoop, |
7195 | BestFactor.Width) || |
7196 | planContainsAdditionalSimplifications( |
7197 | getPlanFor(LegacyVF.Width), CostCtx, OrigLoop, LegacyVF.Width)) && |
7198 | " VPlan cost model and legacy cost model disagreed" ); |
7199 | assert((BestFactor.Width.isScalar() || BestFactor.ScalarCost > 0) && |
7200 | "when vectorizing, the scalar cost must be computed." ); |
7201 | #endif |
7202 | |
7203 | LLVM_DEBUG(dbgs() << "LV: Selecting VF: " << BestFactor.Width << ".\n" ); |
7204 | return BestFactor; |
7205 | } |
7206 | |
7207 | static void addRuntimeUnrollDisableMetaData(Loop *L) { |
7208 | SmallVector<Metadata *, 4> MDs; |
7209 | // Reserve first location for self reference to the LoopID metadata node. |
7210 | MDs.push_back(Elt: nullptr); |
7211 | bool IsUnrollMetadata = false; |
7212 | MDNode *LoopID = L->getLoopID(); |
7213 | if (LoopID) { |
7214 | // First find existing loop unrolling disable metadata. |
7215 | for (unsigned I = 1, IE = LoopID->getNumOperands(); I < IE; ++I) { |
7216 | auto *MD = dyn_cast<MDNode>(Val: LoopID->getOperand(I)); |
7217 | if (MD) { |
7218 | const auto *S = dyn_cast<MDString>(Val: MD->getOperand(I: 0)); |
7219 | IsUnrollMetadata = |
7220 | S && S->getString().starts_with(Prefix: "llvm.loop.unroll.disable" ); |
7221 | } |
7222 | MDs.push_back(Elt: LoopID->getOperand(I)); |
7223 | } |
7224 | } |
7225 | |
7226 | if (!IsUnrollMetadata) { |
7227 | // Add runtime unroll disable metadata. |
7228 | LLVMContext &Context = L->getHeader()->getContext(); |
7229 | SmallVector<Metadata *, 1> DisableOperands; |
7230 | DisableOperands.push_back( |
7231 | Elt: MDString::get(Context, Str: "llvm.loop.unroll.runtime.disable" )); |
7232 | MDNode *DisableNode = MDNode::get(Context, MDs: DisableOperands); |
7233 | MDs.push_back(Elt: DisableNode); |
7234 | MDNode *NewLoopID = MDNode::get(Context, MDs); |
7235 | // Set operand 0 to refer to the loop id itself. |
7236 | NewLoopID->replaceOperandWith(I: 0, New: NewLoopID); |
7237 | L->setLoopID(NewLoopID); |
7238 | } |
7239 | } |
7240 | |
7241 | static Value *getStartValueFromReductionResult(VPInstruction *RdxResult) { |
7242 | using namespace VPlanPatternMatch; |
7243 | assert(RdxResult->getOpcode() == VPInstruction::ComputeFindIVResult && |
7244 | "RdxResult must be ComputeFindIVResult" ); |
7245 | VPValue *StartVPV = RdxResult->getOperand(N: 1); |
7246 | match(V: StartVPV, P: m_Freeze(Op0: m_VPValue(V&: StartVPV))); |
7247 | return StartVPV->getLiveInIRValue(); |
7248 | } |
7249 | |
7250 | // If \p EpiResumePhiR is resume VPPhi for a reduction when vectorizing the |
7251 | // epilog loop, fix the reduction's scalar PHI node by adding the incoming value |
7252 | // from the main vector loop. |
7253 | static void fixReductionScalarResumeWhenVectorizingEpilog( |
7254 | VPPhi *EpiResumePhiR, VPTransformState &State, BasicBlock *BypassBlock) { |
7255 | // Get the VPInstruction computing the reduction result in the middle block. |
7256 | // The first operand may not be from the middle block if it is not connected |
7257 | // to the scalar preheader. In that case, there's nothing to fix. |
7258 | VPValue *Incoming = EpiResumePhiR->getOperand(N: 0); |
7259 | match(V: Incoming, P: VPlanPatternMatch::m_ZExtOrSExt( |
7260 | Op0: VPlanPatternMatch::m_VPValue(V&: Incoming))); |
7261 | auto *EpiRedResult = dyn_cast<VPInstruction>(Val: Incoming); |
7262 | if (!EpiRedResult || |
7263 | (EpiRedResult->getOpcode() != VPInstruction::ComputeAnyOfResult && |
7264 | EpiRedResult->getOpcode() != VPInstruction::ComputeReductionResult && |
7265 | EpiRedResult->getOpcode() != VPInstruction::ComputeFindIVResult)) |
7266 | return; |
7267 | |
7268 | auto * = |
7269 | cast<VPReductionPHIRecipe>(Val: EpiRedResult->getOperand(N: 0)); |
7270 | const RecurrenceDescriptor &RdxDesc = |
7271 | EpiRedHeaderPhi->getRecurrenceDescriptor(); |
7272 | Value *MainResumeValue; |
7273 | if (auto *VPI = dyn_cast<VPInstruction>(Val: EpiRedHeaderPhi->getStartValue())) { |
7274 | assert((VPI->getOpcode() == VPInstruction::Broadcast || |
7275 | VPI->getOpcode() == VPInstruction::ReductionStartVector) && |
7276 | "unexpected start recipe" ); |
7277 | MainResumeValue = VPI->getOperand(N: 0)->getUnderlyingValue(); |
7278 | } else |
7279 | MainResumeValue = EpiRedHeaderPhi->getStartValue()->getUnderlyingValue(); |
7280 | if (RecurrenceDescriptor::isAnyOfRecurrenceKind( |
7281 | Kind: RdxDesc.getRecurrenceKind())) { |
7282 | [[maybe_unused]] Value *StartV = |
7283 | EpiRedResult->getOperand(N: 1)->getLiveInIRValue(); |
7284 | auto *Cmp = cast<ICmpInst>(Val: MainResumeValue); |
7285 | assert(Cmp->getPredicate() == CmpInst::ICMP_NE && |
7286 | "AnyOf expected to start with ICMP_NE" ); |
7287 | assert(Cmp->getOperand(1) == StartV && |
7288 | "AnyOf expected to start by comparing main resume value to original " |
7289 | "start value" ); |
7290 | MainResumeValue = Cmp->getOperand(i_nocapture: 0); |
7291 | } else if (RecurrenceDescriptor::isFindIVRecurrenceKind( |
7292 | Kind: RdxDesc.getRecurrenceKind())) { |
7293 | Value *StartV = getStartValueFromReductionResult(RdxResult: EpiRedResult); |
7294 | Value *SentinelV = EpiRedResult->getOperand(N: 2)->getLiveInIRValue(); |
7295 | using namespace llvm::PatternMatch; |
7296 | Value *Cmp, *OrigResumeV, *CmpOp; |
7297 | [[maybe_unused]] bool IsExpectedPattern = |
7298 | match(V: MainResumeValue, |
7299 | P: m_Select(C: m_OneUse(SubPattern: m_Value(V&: Cmp)), L: m_Specific(V: SentinelV), |
7300 | R: m_Value(V&: OrigResumeV))) && |
7301 | (match(V: Cmp, P: m_SpecificICmp(MatchPred: ICmpInst::ICMP_EQ, L: m_Specific(V: OrigResumeV), |
7302 | R: m_Value(V&: CmpOp))) && |
7303 | ((CmpOp == StartV && isGuaranteedNotToBeUndefOrPoison(V: CmpOp)))); |
7304 | assert(IsExpectedPattern && "Unexpected reduction resume pattern" ); |
7305 | MainResumeValue = OrigResumeV; |
7306 | } |
7307 | PHINode *MainResumePhi = cast<PHINode>(Val: MainResumeValue); |
7308 | |
7309 | // When fixing reductions in the epilogue loop we should already have |
7310 | // created a bc.merge.rdx Phi after the main vector body. Ensure that we carry |
7311 | // over the incoming values correctly. |
7312 | auto *EpiResumePhi = cast<PHINode>(Val: State.get(Def: EpiResumePhiR, IsScalar: true)); |
7313 | EpiResumePhi->setIncomingValueForBlock( |
7314 | BB: BypassBlock, V: MainResumePhi->getIncomingValueForBlock(BB: BypassBlock)); |
7315 | } |
7316 | |
7317 | DenseMap<const SCEV *, Value *> LoopVectorizationPlanner::executePlan( |
7318 | ElementCount BestVF, unsigned BestUF, VPlan &BestVPlan, |
7319 | InnerLoopVectorizer &ILV, DominatorTree *DT, bool VectorizingEpilogue) { |
7320 | assert(BestVPlan.hasVF(BestVF) && |
7321 | "Trying to execute plan with unsupported VF" ); |
7322 | assert(BestVPlan.hasUF(BestUF) && |
7323 | "Trying to execute plan with unsupported UF" ); |
7324 | // TODO: Move to VPlan transform stage once the transition to the VPlan-based |
7325 | // cost model is complete for better cost estimates. |
7326 | VPlanTransforms::runPass(Fn: VPlanTransforms::unrollByUF, Plan&: BestVPlan, Args&: BestUF, |
7327 | Args&: OrigLoop->getHeader()->getContext()); |
7328 | VPlanTransforms::runPass(Fn: VPlanTransforms::replicateByVF, Plan&: BestVPlan, Args&: BestVF); |
7329 | VPlanTransforms::runPass(Fn: VPlanTransforms::materializeBroadcasts, Plan&: BestVPlan); |
7330 | if (hasBranchWeightMD(I: *OrigLoop->getLoopLatch()->getTerminator())) { |
7331 | std::optional<unsigned> VScale = CM.getVScaleForTuning(); |
7332 | VPlanTransforms::runPass(Fn: VPlanTransforms::addBranchWeightToMiddleTerminator, |
7333 | Plan&: BestVPlan, Args&: BestVF, Args&: VScale); |
7334 | } |
7335 | VPlanTransforms::optimizeForVFAndUF(Plan&: BestVPlan, BestVF, BestUF, PSE); |
7336 | VPlanTransforms::simplifyRecipes(Plan&: BestVPlan, CanonicalIVTy&: *Legal->getWidestInductionType()); |
7337 | VPlanTransforms::narrowInterleaveGroups( |
7338 | Plan&: BestVPlan, VF: BestVF, |
7339 | VectorRegWidth: TTI.getRegisterBitWidth(K: TargetTransformInfo::RGK_FixedWidthVector)); |
7340 | VPlanTransforms::removeDeadRecipes(Plan&: BestVPlan); |
7341 | |
7342 | VPlanTransforms::convertToConcreteRecipes(Plan&: BestVPlan, |
7343 | CanonicalIVTy&: *Legal->getWidestInductionType()); |
7344 | // Regions are dissolved after optimizing for VF and UF, which completely |
7345 | // removes unneeded loop regions first. |
7346 | VPlanTransforms::dissolveLoopRegions(Plan&: BestVPlan); |
7347 | // Perform the actual loop transformation. |
7348 | VPTransformState State(&TTI, BestVF, LI, DT, ILV.AC, ILV.Builder, &BestVPlan, |
7349 | OrigLoop->getParentLoop(), |
7350 | Legal->getWidestInductionType()); |
7351 | |
7352 | #ifdef EXPENSIVE_CHECKS |
7353 | assert(DT->verify(DominatorTree::VerificationLevel::Fast)); |
7354 | #endif |
7355 | |
7356 | // 0. Generate SCEV-dependent code in the entry, including TripCount, before |
7357 | // making any changes to the CFG. |
7358 | DenseMap<const SCEV *, Value *> ExpandedSCEVs; |
7359 | auto *Entry = cast<VPIRBasicBlock>(Val: BestVPlan.getEntry()); |
7360 | State.Builder.SetInsertPoint(Entry->getIRBasicBlock()->getTerminator()); |
7361 | for (VPRecipeBase &R : make_early_inc_range(Range&: *Entry)) { |
7362 | auto *ExpSCEV = dyn_cast<VPExpandSCEVRecipe>(Val: &R); |
7363 | if (!ExpSCEV) |
7364 | continue; |
7365 | ExpSCEV->execute(State); |
7366 | ExpandedSCEVs[ExpSCEV->getSCEV()] = State.get(Def: ExpSCEV, Lane: VPLane(0)); |
7367 | VPValue *Exp = BestVPlan.getOrAddLiveIn(V: ExpandedSCEVs[ExpSCEV->getSCEV()]); |
7368 | ExpSCEV->replaceAllUsesWith(New: Exp); |
7369 | if (BestVPlan.getTripCount() == ExpSCEV) |
7370 | BestVPlan.resetTripCount(NewTripCount: Exp); |
7371 | ExpSCEV->eraseFromParent(); |
7372 | } |
7373 | |
7374 | if (!ILV.getTripCount()) |
7375 | ILV.setTripCount(State.get(Def: BestVPlan.getTripCount(), Lane: VPLane(0))); |
7376 | else |
7377 | assert(VectorizingEpilogue && "should only re-use the existing trip " |
7378 | "count during epilogue vectorization" ); |
7379 | |
7380 | // 1. Set up the skeleton for vectorization, including vector pre-header and |
7381 | // middle block. The vector loop is created during VPlan execution. |
7382 | VPBasicBlock *VectorPH = cast<VPBasicBlock>(Val: Entry->getSuccessors()[1]); |
7383 | State.CFG.PrevBB = ILV.createVectorizedLoopSkeleton(); |
7384 | if (VectorizingEpilogue) |
7385 | VPlanTransforms::removeDeadRecipes(Plan&: BestVPlan); |
7386 | |
7387 | assert(verifyVPlanIsValid(BestVPlan, true /*VerifyLate*/) && |
7388 | "final VPlan is invalid" ); |
7389 | |
7390 | ILV.printDebugTracesAtStart(); |
7391 | |
7392 | //===------------------------------------------------===// |
7393 | // |
7394 | // Notice: any optimization or new instruction that go |
7395 | // into the code below should also be implemented in |
7396 | // the cost-model. |
7397 | // |
7398 | //===------------------------------------------------===// |
7399 | |
7400 | // 2. Copy and widen instructions from the old loop into the new loop. |
7401 | BestVPlan.prepareToExecute( |
7402 | TripCount: ILV.getTripCount(), |
7403 | VectorTripCount: ILV.getOrCreateVectorTripCount(InsertBlock: ILV.LoopVectorPreHeader), State); |
7404 | replaceVPBBWithIRVPBB(VPBB: VectorPH, IRBB: State.CFG.PrevBB); |
7405 | |
7406 | BestVPlan.execute(State: &State); |
7407 | |
7408 | // 2.5 When vectorizing the epilogue, fix reduction resume values from the |
7409 | // additional bypass block. |
7410 | if (VectorizingEpilogue) { |
7411 | assert(!BestVPlan.hasEarlyExit() && |
7412 | "Epilogue vectorisation not yet supported with early exits" ); |
7413 | BasicBlock *PH = OrigLoop->getLoopPreheader(); |
7414 | BasicBlock *BypassBlock = ILV.getAdditionalBypassBlock(); |
7415 | for (auto *Pred : predecessors(BB: PH)) { |
7416 | for (PHINode &Phi : PH->phis()) { |
7417 | if (Phi.getBasicBlockIndex(BB: Pred) != -1) |
7418 | continue; |
7419 | Phi.addIncoming(V: Phi.getIncomingValueForBlock(BB: BypassBlock), BB: Pred); |
7420 | } |
7421 | } |
7422 | VPBasicBlock *ScalarPH = BestVPlan.getScalarPreheader(); |
7423 | if (ScalarPH->getNumPredecessors() > 0) { |
7424 | // If ScalarPH has predecessors, we may need to update its reduction |
7425 | // resume values. |
7426 | for (VPRecipeBase &R : ScalarPH->phis()) { |
7427 | fixReductionScalarResumeWhenVectorizingEpilog(EpiResumePhiR: cast<VPPhi>(Val: &R), State, |
7428 | BypassBlock); |
7429 | } |
7430 | } |
7431 | } |
7432 | |
7433 | // 2.6. Maintain Loop Hints |
7434 | // Keep all loop hints from the original loop on the vector loop (we'll |
7435 | // replace the vectorizer-specific hints below). |
7436 | VPBasicBlock * = vputils::getFirstLoopHeader(Plan&: BestVPlan, VPDT&: State.VPDT); |
7437 | if (HeaderVPBB) { |
7438 | MDNode *OrigLoopID = OrigLoop->getLoopID(); |
7439 | |
7440 | std::optional<MDNode *> VectorizedLoopID = |
7441 | makeFollowupLoopID(OrigLoopID, FollowupAttrs: {LLVMLoopVectorizeFollowupAll, |
7442 | LLVMLoopVectorizeFollowupVectorized}); |
7443 | |
7444 | Loop *L = LI->getLoopFor(BB: State.CFG.VPBB2IRBB[HeaderVPBB]); |
7445 | if (VectorizedLoopID) { |
7446 | L->setLoopID(*VectorizedLoopID); |
7447 | } else { |
7448 | // Keep all loop hints from the original loop on the vector loop (we'll |
7449 | // replace the vectorizer-specific hints below). |
7450 | if (MDNode *LID = OrigLoop->getLoopID()) |
7451 | L->setLoopID(LID); |
7452 | |
7453 | LoopVectorizeHints Hints(L, true, *ORE); |
7454 | Hints.setAlreadyVectorized(); |
7455 | |
7456 | // Check if it's EVL-vectorized and mark the corresponding metadata. |
7457 | bool IsEVLVectorized = |
7458 | llvm::any_of(Range&: *HeaderVPBB, P: [](const VPRecipeBase &Recipe) { |
7459 | // Looking for the ExplictVectorLength VPInstruction. |
7460 | if (const auto *VI = dyn_cast<VPInstruction>(Val: &Recipe)) |
7461 | return VI->getOpcode() == VPInstruction::ExplicitVectorLength; |
7462 | return false; |
7463 | }); |
7464 | if (IsEVLVectorized) { |
7465 | LLVMContext &Context = L->getHeader()->getContext(); |
7466 | MDNode *LoopID = L->getLoopID(); |
7467 | auto *IsEVLVectorizedMD = MDNode::get( |
7468 | Context, |
7469 | MDs: {MDString::get(Context, Str: "llvm.loop.isvectorized.tailfoldingstyle" ), |
7470 | MDString::get(Context, Str: "evl" )}); |
7471 | MDNode *NewLoopID = makePostTransformationMetadata(Context, OrigLoopID: LoopID, RemovePrefixes: {}, |
7472 | AddAttrs: {IsEVLVectorizedMD}); |
7473 | L->setLoopID(NewLoopID); |
7474 | } |
7475 | } |
7476 | TargetTransformInfo::UnrollingPreferences UP; |
7477 | TTI.getUnrollingPreferences(L, *PSE.getSE(), UP, ORE); |
7478 | if (!UP.UnrollVectorizedLoop || VectorizingEpilogue) |
7479 | addRuntimeUnrollDisableMetaData(L); |
7480 | } |
7481 | |
7482 | // 3. Fix the vectorized code: take care of header phi's, live-outs, |
7483 | // predication, updating analyses. |
7484 | ILV.fixVectorizedLoop(State); |
7485 | |
7486 | ILV.printDebugTracesAtEnd(); |
7487 | |
7488 | return ExpandedSCEVs; |
7489 | } |
7490 | |
7491 | //===--------------------------------------------------------------------===// |
7492 | // EpilogueVectorizerMainLoop |
7493 | //===--------------------------------------------------------------------===// |
7494 | |
7495 | /// This function is partially responsible for generating the control flow |
7496 | /// depicted in https://llvm.org/docs/Vectorizers.html#epilogue-vectorization. |
7497 | BasicBlock *EpilogueVectorizerMainLoop::createEpilogueVectorizedLoopSkeleton() { |
7498 | createVectorLoopSkeleton(Prefix: "" ); |
7499 | |
7500 | // Generate the code to check the minimum iteration count of the vector |
7501 | // epilogue (see below). |
7502 | EPI.EpilogueIterationCountCheck = |
7503 | emitIterationCountCheck(Bypass: LoopScalarPreHeader, ForEpilogue: true); |
7504 | EPI.EpilogueIterationCountCheck->setName("iter.check" ); |
7505 | |
7506 | // Generate the code to check any assumptions that we've made for SCEV |
7507 | // expressions. |
7508 | EPI.SCEVSafetyCheck = emitSCEVChecks(Bypass: LoopScalarPreHeader); |
7509 | |
7510 | // Generate the code that checks at runtime if arrays overlap. We put the |
7511 | // checks into a separate block to make the more common case of few elements |
7512 | // faster. |
7513 | EPI.MemSafetyCheck = emitMemRuntimeChecks(Bypass: LoopScalarPreHeader); |
7514 | |
7515 | // Generate the iteration count check for the main loop, *after* the check |
7516 | // for the epilogue loop, so that the path-length is shorter for the case |
7517 | // that goes directly through the vector epilogue. The longer-path length for |
7518 | // the main loop is compensated for, by the gain from vectorizing the larger |
7519 | // trip count. Note: the branch will get updated later on when we vectorize |
7520 | // the epilogue. |
7521 | EPI.MainLoopIterationCountCheck = |
7522 | emitIterationCountCheck(Bypass: LoopScalarPreHeader, ForEpilogue: false); |
7523 | |
7524 | // Generate the induction variable. |
7525 | EPI.VectorTripCount = getOrCreateVectorTripCount(InsertBlock: LoopVectorPreHeader); |
7526 | |
7527 | replaceVPBBWithIRVPBB(VPBB: Plan.getScalarPreheader(), IRBB: LoopScalarPreHeader); |
7528 | return LoopVectorPreHeader; |
7529 | } |
7530 | |
7531 | void EpilogueVectorizerMainLoop::printDebugTracesAtStart() { |
7532 | LLVM_DEBUG({ |
7533 | dbgs() << "Create Skeleton for epilogue vectorized loop (first pass)\n" |
7534 | << "Main Loop VF:" << EPI.MainLoopVF |
7535 | << ", Main Loop UF:" << EPI.MainLoopUF |
7536 | << ", Epilogue Loop VF:" << EPI.EpilogueVF |
7537 | << ", Epilogue Loop UF:" << EPI.EpilogueUF << "\n" ; |
7538 | }); |
7539 | } |
7540 | |
7541 | void EpilogueVectorizerMainLoop::printDebugTracesAtEnd() { |
7542 | DEBUG_WITH_TYPE(VerboseDebug, { |
7543 | dbgs() << "intermediate fn:\n" |
7544 | << *OrigLoop->getHeader()->getParent() << "\n" ; |
7545 | }); |
7546 | } |
7547 | |
7548 | BasicBlock * |
7549 | EpilogueVectorizerMainLoop::emitIterationCountCheck(BasicBlock *Bypass, |
7550 | bool ForEpilogue) { |
7551 | assert(Bypass && "Expected valid bypass basic block." ); |
7552 | Value *Count = getTripCount(); |
7553 | MinProfitableTripCount = ElementCount::getFixed(MinVal: 0); |
7554 | Value *CheckMinIters = createIterationCountCheck( |
7555 | VF: ForEpilogue ? EPI.EpilogueVF : VF, UF: ForEpilogue ? EPI.EpilogueUF : UF); |
7556 | |
7557 | BasicBlock *const TCCheckBlock = LoopVectorPreHeader; |
7558 | if (!ForEpilogue) |
7559 | TCCheckBlock->setName("vector.main.loop.iter.check" ); |
7560 | |
7561 | // Create new preheader for vector loop. |
7562 | LoopVectorPreHeader = SplitBlock(Old: TCCheckBlock, SplitPt: TCCheckBlock->getTerminator(), |
7563 | DT: static_cast<DominatorTree *>(nullptr), LI, |
7564 | MSSAU: nullptr, BBName: "vector.ph" ); |
7565 | |
7566 | if (ForEpilogue) { |
7567 | // Save the trip count so we don't have to regenerate it in the |
7568 | // vec.epilog.iter.check. This is safe to do because the trip count |
7569 | // generated here dominates the vector epilog iter check. |
7570 | EPI.TripCount = Count; |
7571 | } |
7572 | |
7573 | BranchInst &BI = |
7574 | *BranchInst::Create(IfTrue: Bypass, IfFalse: LoopVectorPreHeader, Cond: CheckMinIters); |
7575 | if (hasBranchWeightMD(I: *OrigLoop->getLoopLatch()->getTerminator())) |
7576 | setBranchWeights(I&: BI, Weights: MinItersBypassWeights, /*IsExpected=*/false); |
7577 | ReplaceInstWithInst(From: TCCheckBlock->getTerminator(), To: &BI); |
7578 | |
7579 | // When vectorizing the main loop, its trip-count check is placed in a new |
7580 | // block, whereas the overall trip-count check is placed in the VPlan entry |
7581 | // block. When vectorizing the epilogue loop, its trip-count check is placed |
7582 | // in the VPlan entry block. |
7583 | if (!ForEpilogue) |
7584 | introduceCheckBlockInVPlan(CheckIRBB: TCCheckBlock); |
7585 | return TCCheckBlock; |
7586 | } |
7587 | |
7588 | //===--------------------------------------------------------------------===// |
7589 | // EpilogueVectorizerEpilogueLoop |
7590 | //===--------------------------------------------------------------------===// |
7591 | |
7592 | /// This function is partially responsible for generating the control flow |
7593 | /// depicted in https://llvm.org/docs/Vectorizers.html#epilogue-vectorization. |
7594 | BasicBlock * |
7595 | EpilogueVectorizerEpilogueLoop::createEpilogueVectorizedLoopSkeleton() { |
7596 | createVectorLoopSkeleton(Prefix: "vec.epilog." ); |
7597 | |
7598 | // Now, compare the remaining count and if there aren't enough iterations to |
7599 | // execute the vectorized epilogue skip to the scalar part. |
7600 | LoopVectorPreHeader->setName("vec.epilog.ph" ); |
7601 | BasicBlock *VecEpilogueIterationCountCheck = |
7602 | SplitBlock(Old: LoopVectorPreHeader, SplitPt: LoopVectorPreHeader->begin(), DT, LI, |
7603 | MSSAU: nullptr, BBName: "vec.epilog.iter.check" , Before: true); |
7604 | emitMinimumVectorEpilogueIterCountCheck(Bypass: LoopScalarPreHeader, |
7605 | Insert: VecEpilogueIterationCountCheck); |
7606 | AdditionalBypassBlock = VecEpilogueIterationCountCheck; |
7607 | |
7608 | // Adjust the control flow taking the state info from the main loop |
7609 | // vectorization into account. |
7610 | assert(EPI.MainLoopIterationCountCheck && EPI.EpilogueIterationCountCheck && |
7611 | "expected this to be saved from the previous pass." ); |
7612 | EPI.MainLoopIterationCountCheck->getTerminator()->replaceUsesOfWith( |
7613 | From: VecEpilogueIterationCountCheck, To: LoopVectorPreHeader); |
7614 | |
7615 | EPI.EpilogueIterationCountCheck->getTerminator()->replaceUsesOfWith( |
7616 | From: VecEpilogueIterationCountCheck, To: LoopScalarPreHeader); |
7617 | |
7618 | if (EPI.SCEVSafetyCheck) |
7619 | EPI.SCEVSafetyCheck->getTerminator()->replaceUsesOfWith( |
7620 | From: VecEpilogueIterationCountCheck, To: LoopScalarPreHeader); |
7621 | if (EPI.MemSafetyCheck) |
7622 | EPI.MemSafetyCheck->getTerminator()->replaceUsesOfWith( |
7623 | From: VecEpilogueIterationCountCheck, To: LoopScalarPreHeader); |
7624 | |
7625 | DT->changeImmediateDominator(BB: LoopScalarPreHeader, |
7626 | NewBB: EPI.EpilogueIterationCountCheck); |
7627 | |
7628 | // The vec.epilog.iter.check block may contain Phi nodes from inductions or |
7629 | // reductions which merge control-flow from the latch block and the middle |
7630 | // block. Update the incoming values here and move the Phi into the preheader. |
7631 | SmallVector<PHINode *, 4> PhisInBlock( |
7632 | llvm::make_pointer_range(Range: VecEpilogueIterationCountCheck->phis())); |
7633 | |
7634 | for (PHINode *Phi : PhisInBlock) { |
7635 | Phi->moveBefore(InsertPos: LoopVectorPreHeader->getFirstNonPHIIt()); |
7636 | Phi->replaceIncomingBlockWith( |
7637 | Old: VecEpilogueIterationCountCheck->getSinglePredecessor(), |
7638 | New: VecEpilogueIterationCountCheck); |
7639 | |
7640 | // If the phi doesn't have an incoming value from the |
7641 | // EpilogueIterationCountCheck, we are done. Otherwise remove the incoming |
7642 | // value and also those from other check blocks. This is needed for |
7643 | // reduction phis only. |
7644 | if (none_of(Range: Phi->blocks(), P: [&](BasicBlock *IncB) { |
7645 | return EPI.EpilogueIterationCountCheck == IncB; |
7646 | })) |
7647 | continue; |
7648 | Phi->removeIncomingValue(BB: EPI.EpilogueIterationCountCheck); |
7649 | if (EPI.SCEVSafetyCheck) |
7650 | Phi->removeIncomingValue(BB: EPI.SCEVSafetyCheck); |
7651 | if (EPI.MemSafetyCheck) |
7652 | Phi->removeIncomingValue(BB: EPI.MemSafetyCheck); |
7653 | } |
7654 | |
7655 | replaceVPBBWithIRVPBB(VPBB: Plan.getScalarPreheader(), IRBB: LoopScalarPreHeader); |
7656 | return LoopVectorPreHeader; |
7657 | } |
7658 | |
7659 | BasicBlock * |
7660 | EpilogueVectorizerEpilogueLoop::emitMinimumVectorEpilogueIterCountCheck( |
7661 | BasicBlock *Bypass, BasicBlock *Insert) { |
7662 | |
7663 | assert(EPI.TripCount && |
7664 | "Expected trip count to have been saved in the first pass." ); |
7665 | Value *TC = EPI.TripCount; |
7666 | IRBuilder<> Builder(Insert->getTerminator()); |
7667 | Value *Count = Builder.CreateSub(LHS: TC, RHS: EPI.VectorTripCount, Name: "n.vec.remaining" ); |
7668 | |
7669 | // Generate code to check if the loop's trip count is less than VF * UF of the |
7670 | // vector epilogue loop. |
7671 | auto P = Cost->requiresScalarEpilogue(IsVectorizing: EPI.EpilogueVF.isVector()) |
7672 | ? ICmpInst::ICMP_ULE |
7673 | : ICmpInst::ICMP_ULT; |
7674 | |
7675 | Value *CheckMinIters = |
7676 | Builder.CreateICmp(P, LHS: Count, |
7677 | RHS: createStepForVF(B&: Builder, Ty: Count->getType(), |
7678 | VF: EPI.EpilogueVF, Step: EPI.EpilogueUF), |
7679 | Name: "min.epilog.iters.check" ); |
7680 | |
7681 | BranchInst &BI = |
7682 | *BranchInst::Create(IfTrue: Bypass, IfFalse: LoopVectorPreHeader, Cond: CheckMinIters); |
7683 | if (hasBranchWeightMD(I: *OrigLoop->getLoopLatch()->getTerminator())) { |
7684 | // FIXME: See test Transforms/LoopVectorize/branch-weights.ll. I don't |
7685 | // think the MainLoopStep is correct. |
7686 | unsigned MainLoopStep = UF * VF.getKnownMinValue(); |
7687 | unsigned EpilogueLoopStep = |
7688 | EPI.EpilogueUF * EPI.EpilogueVF.getKnownMinValue(); |
7689 | // We assume the remaining `Count` is equally distributed in |
7690 | // [0, MainLoopStep) |
7691 | // So the probability for `Count < EpilogueLoopStep` should be |
7692 | // min(MainLoopStep, EpilogueLoopStep) / MainLoopStep |
7693 | unsigned EstimatedSkipCount = std::min(a: MainLoopStep, b: EpilogueLoopStep); |
7694 | const uint32_t Weights[] = {EstimatedSkipCount, |
7695 | MainLoopStep - EstimatedSkipCount}; |
7696 | setBranchWeights(I&: BI, Weights, /*IsExpected=*/false); |
7697 | } |
7698 | ReplaceInstWithInst(From: Insert->getTerminator(), To: &BI); |
7699 | |
7700 | // A new entry block has been created for the epilogue VPlan. Hook it in, as |
7701 | // otherwise we would try to modify the entry to the main vector loop. |
7702 | VPIRBasicBlock *NewEntry = Plan.createVPIRBasicBlock(IRBB: Insert); |
7703 | VPBasicBlock *OldEntry = Plan.getEntry(); |
7704 | VPBlockUtils::reassociateBlocks(Old: OldEntry, New: NewEntry); |
7705 | Plan.setEntry(NewEntry); |
7706 | // OldEntry is now dead and will be cleaned up when the plan gets destroyed. |
7707 | |
7708 | return Insert; |
7709 | } |
7710 | |
7711 | void EpilogueVectorizerEpilogueLoop::printDebugTracesAtStart() { |
7712 | LLVM_DEBUG({ |
7713 | dbgs() << "Create Skeleton for epilogue vectorized loop (second pass)\n" |
7714 | << "Epilogue Loop VF:" << EPI.EpilogueVF |
7715 | << ", Epilogue Loop UF:" << EPI.EpilogueUF << "\n" ; |
7716 | }); |
7717 | } |
7718 | |
7719 | void EpilogueVectorizerEpilogueLoop::printDebugTracesAtEnd() { |
7720 | DEBUG_WITH_TYPE(VerboseDebug, { |
7721 | dbgs() << "final fn:\n" << *OrigLoop->getHeader()->getParent() << "\n" ; |
7722 | }); |
7723 | } |
7724 | |
7725 | VPWidenMemoryRecipe * |
7726 | VPRecipeBuilder::tryToWidenMemory(Instruction *I, ArrayRef<VPValue *> Operands, |
7727 | VFRange &Range) { |
7728 | assert((isa<LoadInst>(I) || isa<StoreInst>(I)) && |
7729 | "Must be called with either a load or store" ); |
7730 | |
7731 | auto WillWiden = [&](ElementCount VF) -> bool { |
7732 | LoopVectorizationCostModel::InstWidening Decision = |
7733 | CM.getWideningDecision(I, VF); |
7734 | assert(Decision != LoopVectorizationCostModel::CM_Unknown && |
7735 | "CM decision should be taken at this point." ); |
7736 | if (Decision == LoopVectorizationCostModel::CM_Interleave) |
7737 | return true; |
7738 | if (CM.isScalarAfterVectorization(I, VF) || |
7739 | CM.isProfitableToScalarize(I, VF)) |
7740 | return false; |
7741 | return Decision != LoopVectorizationCostModel::CM_Scalarize; |
7742 | }; |
7743 | |
7744 | if (!LoopVectorizationPlanner::getDecisionAndClampRange(Predicate: WillWiden, Range)) |
7745 | return nullptr; |
7746 | |
7747 | VPValue *Mask = nullptr; |
7748 | if (Legal->isMaskRequired(I)) |
7749 | Mask = getBlockInMask(VPBB: Builder.getInsertBlock()); |
7750 | |
7751 | // Determine if the pointer operand of the access is either consecutive or |
7752 | // reverse consecutive. |
7753 | LoopVectorizationCostModel::InstWidening Decision = |
7754 | CM.getWideningDecision(I, VF: Range.Start); |
7755 | bool Reverse = Decision == LoopVectorizationCostModel::CM_Widen_Reverse; |
7756 | bool Consecutive = |
7757 | Reverse || Decision == LoopVectorizationCostModel::CM_Widen; |
7758 | |
7759 | VPValue *Ptr = isa<LoadInst>(Val: I) ? Operands[0] : Operands[1]; |
7760 | if (Consecutive) { |
7761 | auto *GEP = dyn_cast<GetElementPtrInst>( |
7762 | Val: Ptr->getUnderlyingValue()->stripPointerCasts()); |
7763 | VPSingleDefRecipe *VectorPtr; |
7764 | if (Reverse) { |
7765 | // When folding the tail, we may compute an address that we don't in the |
7766 | // original scalar loop and it may not be inbounds. Drop Inbounds in that |
7767 | // case. |
7768 | GEPNoWrapFlags Flags = |
7769 | (CM.foldTailByMasking() || !GEP || !GEP->isInBounds()) |
7770 | ? GEPNoWrapFlags::none() |
7771 | : GEPNoWrapFlags::inBounds(); |
7772 | VectorPtr = |
7773 | new VPVectorEndPointerRecipe(Ptr, &Plan.getVF(), getLoadStoreType(I), |
7774 | /*Stride*/ -1, Flags, I->getDebugLoc()); |
7775 | } else { |
7776 | VectorPtr = new VPVectorPointerRecipe(Ptr, getLoadStoreType(I), |
7777 | GEP ? GEP->getNoWrapFlags() |
7778 | : GEPNoWrapFlags::none(), |
7779 | I->getDebugLoc()); |
7780 | } |
7781 | Builder.insert(R: VectorPtr); |
7782 | Ptr = VectorPtr; |
7783 | } |
7784 | if (LoadInst *Load = dyn_cast<LoadInst>(Val: I)) |
7785 | return new VPWidenLoadRecipe(*Load, Ptr, Mask, Consecutive, Reverse, |
7786 | VPIRMetadata(*Load, LVer), I->getDebugLoc()); |
7787 | |
7788 | StoreInst *Store = cast<StoreInst>(Val: I); |
7789 | return new VPWidenStoreRecipe(*Store, Ptr, Operands[0], Mask, Consecutive, |
7790 | Reverse, VPIRMetadata(*Store, LVer), |
7791 | I->getDebugLoc()); |
7792 | } |
7793 | |
7794 | /// Creates a VPWidenIntOrFpInductionRecpipe for \p Phi. If needed, it will also |
7795 | /// insert a recipe to expand the step for the induction recipe. |
7796 | static VPWidenIntOrFpInductionRecipe * |
7797 | createWidenInductionRecipes(PHINode *Phi, Instruction *PhiOrTrunc, |
7798 | VPValue *Start, const InductionDescriptor &IndDesc, |
7799 | VPlan &Plan, ScalarEvolution &SE, Loop &OrigLoop) { |
7800 | assert(IndDesc.getStartValue() == |
7801 | Phi->getIncomingValueForBlock(OrigLoop.getLoopPreheader())); |
7802 | assert(SE.isLoopInvariant(IndDesc.getStep(), &OrigLoop) && |
7803 | "step must be loop invariant" ); |
7804 | |
7805 | VPValue *Step = |
7806 | vputils::getOrCreateVPValueForSCEVExpr(Plan, Expr: IndDesc.getStep(), SE); |
7807 | if (auto *TruncI = dyn_cast<TruncInst>(Val: PhiOrTrunc)) { |
7808 | return new VPWidenIntOrFpInductionRecipe(Phi, Start, Step, &Plan.getVF(), |
7809 | IndDesc, TruncI, |
7810 | TruncI->getDebugLoc()); |
7811 | } |
7812 | assert(isa<PHINode>(PhiOrTrunc) && "must be a phi node here" ); |
7813 | return new VPWidenIntOrFpInductionRecipe(Phi, Start, Step, &Plan.getVF(), |
7814 | IndDesc, Phi->getDebugLoc()); |
7815 | } |
7816 | |
7817 | VPHeaderPHIRecipe *VPRecipeBuilder::tryToOptimizeInductionPHI( |
7818 | PHINode *Phi, ArrayRef<VPValue *> Operands, VFRange &Range) { |
7819 | |
7820 | // Check if this is an integer or fp induction. If so, build the recipe that |
7821 | // produces its scalar and vector values. |
7822 | if (auto *II = Legal->getIntOrFpInductionDescriptor(Phi)) |
7823 | return createWidenInductionRecipes(Phi, PhiOrTrunc: Phi, Start: Operands[0], IndDesc: *II, Plan, |
7824 | SE&: *PSE.getSE(), OrigLoop&: *OrigLoop); |
7825 | |
7826 | // Check if this is pointer induction. If so, build the recipe for it. |
7827 | if (auto *II = Legal->getPointerInductionDescriptor(Phi)) { |
7828 | VPValue *Step = vputils::getOrCreateVPValueForSCEVExpr(Plan, Expr: II->getStep(), |
7829 | SE&: *PSE.getSE()); |
7830 | return new VPWidenPointerInductionRecipe( |
7831 | Phi, Operands[0], Step, &Plan.getVFxUF(), *II, |
7832 | LoopVectorizationPlanner::getDecisionAndClampRange( |
7833 | Predicate: [&](ElementCount VF) { |
7834 | return CM.isScalarAfterVectorization(I: Phi, VF); |
7835 | }, |
7836 | Range), |
7837 | Phi->getDebugLoc()); |
7838 | } |
7839 | return nullptr; |
7840 | } |
7841 | |
7842 | VPWidenIntOrFpInductionRecipe *VPRecipeBuilder::tryToOptimizeInductionTruncate( |
7843 | TruncInst *I, ArrayRef<VPValue *> Operands, VFRange &Range) { |
7844 | // Optimize the special case where the source is a constant integer |
7845 | // induction variable. Notice that we can only optimize the 'trunc' case |
7846 | // because (a) FP conversions lose precision, (b) sext/zext may wrap, and |
7847 | // (c) other casts depend on pointer size. |
7848 | |
7849 | // Determine whether \p K is a truncation based on an induction variable that |
7850 | // can be optimized. |
7851 | auto IsOptimizableIVTruncate = |
7852 | [&](Instruction *K) -> std::function<bool(ElementCount)> { |
7853 | return [=](ElementCount VF) -> bool { |
7854 | return CM.isOptimizableIVTruncate(I: K, VF); |
7855 | }; |
7856 | }; |
7857 | |
7858 | if (LoopVectorizationPlanner::getDecisionAndClampRange( |
7859 | Predicate: IsOptimizableIVTruncate(I), Range)) { |
7860 | |
7861 | auto *Phi = cast<PHINode>(Val: I->getOperand(i_nocapture: 0)); |
7862 | const InductionDescriptor &II = *Legal->getIntOrFpInductionDescriptor(Phi); |
7863 | VPValue *Start = Plan.getOrAddLiveIn(V: II.getStartValue()); |
7864 | return createWidenInductionRecipes(Phi, PhiOrTrunc: I, Start, IndDesc: II, Plan, SE&: *PSE.getSE(), |
7865 | OrigLoop&: *OrigLoop); |
7866 | } |
7867 | return nullptr; |
7868 | } |
7869 | |
7870 | VPSingleDefRecipe *VPRecipeBuilder::tryToWidenCall(CallInst *CI, |
7871 | ArrayRef<VPValue *> Operands, |
7872 | VFRange &Range) { |
7873 | bool IsPredicated = LoopVectorizationPlanner::getDecisionAndClampRange( |
7874 | Predicate: [this, CI](ElementCount VF) { |
7875 | return CM.isScalarWithPredication(I: CI, VF); |
7876 | }, |
7877 | Range); |
7878 | |
7879 | if (IsPredicated) |
7880 | return nullptr; |
7881 | |
7882 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); |
7883 | if (ID && (ID == Intrinsic::assume || ID == Intrinsic::lifetime_end || |
7884 | ID == Intrinsic::lifetime_start || ID == Intrinsic::sideeffect || |
7885 | ID == Intrinsic::pseudoprobe || |
7886 | ID == Intrinsic::experimental_noalias_scope_decl)) |
7887 | return nullptr; |
7888 | |
7889 | SmallVector<VPValue *, 4> Ops(Operands.take_front(N: CI->arg_size())); |
7890 | |
7891 | // Is it beneficial to perform intrinsic call compared to lib call? |
7892 | bool ShouldUseVectorIntrinsic = |
7893 | ID && LoopVectorizationPlanner::getDecisionAndClampRange( |
7894 | Predicate: [&](ElementCount VF) -> bool { |
7895 | return CM.getCallWideningDecision(CI, VF).Kind == |
7896 | LoopVectorizationCostModel::CM_IntrinsicCall; |
7897 | }, |
7898 | Range); |
7899 | if (ShouldUseVectorIntrinsic) |
7900 | return new VPWidenIntrinsicRecipe(*CI, ID, Ops, CI->getType(), |
7901 | CI->getDebugLoc()); |
7902 | |
7903 | Function *Variant = nullptr; |
7904 | std::optional<unsigned> MaskPos; |
7905 | // Is better to call a vectorized version of the function than to to scalarize |
7906 | // the call? |
7907 | auto ShouldUseVectorCall = LoopVectorizationPlanner::getDecisionAndClampRange( |
7908 | Predicate: [&](ElementCount VF) -> bool { |
7909 | // The following case may be scalarized depending on the VF. |
7910 | // The flag shows whether we can use a usual Call for vectorized |
7911 | // version of the instruction. |
7912 | |
7913 | // If we've found a variant at a previous VF, then stop looking. A |
7914 | // vectorized variant of a function expects input in a certain shape |
7915 | // -- basically the number of input registers, the number of lanes |
7916 | // per register, and whether there's a mask required. |
7917 | // We store a pointer to the variant in the VPWidenCallRecipe, so |
7918 | // once we have an appropriate variant it's only valid for that VF. |
7919 | // This will force a different vplan to be generated for each VF that |
7920 | // finds a valid variant. |
7921 | if (Variant) |
7922 | return false; |
7923 | LoopVectorizationCostModel::CallWideningDecision Decision = |
7924 | CM.getCallWideningDecision(CI, VF); |
7925 | if (Decision.Kind == LoopVectorizationCostModel::CM_VectorCall) { |
7926 | Variant = Decision.Variant; |
7927 | MaskPos = Decision.MaskPos; |
7928 | return true; |
7929 | } |
7930 | |
7931 | return false; |
7932 | }, |
7933 | Range); |
7934 | if (ShouldUseVectorCall) { |
7935 | if (MaskPos.has_value()) { |
7936 | // We have 2 cases that would require a mask: |
7937 | // 1) The block needs to be predicated, either due to a conditional |
7938 | // in the scalar loop or use of an active lane mask with |
7939 | // tail-folding, and we use the appropriate mask for the block. |
7940 | // 2) No mask is required for the block, but the only available |
7941 | // vector variant at this VF requires a mask, so we synthesize an |
7942 | // all-true mask. |
7943 | VPValue *Mask = nullptr; |
7944 | if (Legal->isMaskRequired(I: CI)) |
7945 | Mask = getBlockInMask(VPBB: Builder.getInsertBlock()); |
7946 | else |
7947 | Mask = Plan.getOrAddLiveIn( |
7948 | V: ConstantInt::getTrue(Ty: IntegerType::getInt1Ty(C&: CI->getContext()))); |
7949 | |
7950 | Ops.insert(I: Ops.begin() + *MaskPos, Elt: Mask); |
7951 | } |
7952 | |
7953 | Ops.push_back(Elt: Operands.back()); |
7954 | return new VPWidenCallRecipe(CI, Variant, Ops, CI->getDebugLoc()); |
7955 | } |
7956 | |
7957 | return nullptr; |
7958 | } |
7959 | |
7960 | bool VPRecipeBuilder::shouldWiden(Instruction *I, VFRange &Range) const { |
7961 | assert(!isa<BranchInst>(I) && !isa<PHINode>(I) && !isa<LoadInst>(I) && |
7962 | !isa<StoreInst>(I) && "Instruction should have been handled earlier" ); |
7963 | // Instruction should be widened, unless it is scalar after vectorization, |
7964 | // scalarization is profitable or it is predicated. |
7965 | auto WillScalarize = [this, I](ElementCount VF) -> bool { |
7966 | return CM.isScalarAfterVectorization(I, VF) || |
7967 | CM.isProfitableToScalarize(I, VF) || |
7968 | CM.isScalarWithPredication(I, VF); |
7969 | }; |
7970 | return !LoopVectorizationPlanner::getDecisionAndClampRange(Predicate: WillScalarize, |
7971 | Range); |
7972 | } |
7973 | |
7974 | VPWidenRecipe *VPRecipeBuilder::tryToWiden(Instruction *I, |
7975 | ArrayRef<VPValue *> Operands) { |
7976 | switch (I->getOpcode()) { |
7977 | default: |
7978 | return nullptr; |
7979 | case Instruction::SDiv: |
7980 | case Instruction::UDiv: |
7981 | case Instruction::SRem: |
7982 | case Instruction::URem: { |
7983 | // If not provably safe, use a select to form a safe divisor before widening the |
7984 | // div/rem operation itself. Otherwise fall through to general handling below. |
7985 | if (CM.isPredicatedInst(I)) { |
7986 | SmallVector<VPValue *> Ops(Operands); |
7987 | VPValue *Mask = getBlockInMask(VPBB: Builder.getInsertBlock()); |
7988 | VPValue *One = |
7989 | Plan.getOrAddLiveIn(V: ConstantInt::get(Ty: I->getType(), V: 1u, IsSigned: false)); |
7990 | auto *SafeRHS = Builder.createSelect(Cond: Mask, TrueVal: Ops[1], FalseVal: One, DL: I->getDebugLoc()); |
7991 | Ops[1] = SafeRHS; |
7992 | return new VPWidenRecipe(*I, Ops); |
7993 | } |
7994 | [[fallthrough]]; |
7995 | } |
7996 | case Instruction::Add: |
7997 | case Instruction::And: |
7998 | case Instruction::AShr: |
7999 | case Instruction::FAdd: |
8000 | case Instruction::FCmp: |
8001 | case Instruction::FDiv: |
8002 | case Instruction::FMul: |
8003 | case Instruction::FNeg: |
8004 | case Instruction::FRem: |
8005 | case Instruction::FSub: |
8006 | case Instruction::ICmp: |
8007 | case Instruction::LShr: |
8008 | case Instruction::Mul: |
8009 | case Instruction::Or: |
8010 | case Instruction::Select: |
8011 | case Instruction::Shl: |
8012 | case Instruction::Sub: |
8013 | case Instruction::Xor: |
8014 | case Instruction::Freeze: { |
8015 | SmallVector<VPValue *> NewOps(Operands); |
8016 | if (Instruction::isBinaryOp(Opcode: I->getOpcode())) { |
8017 | // The legacy cost model uses SCEV to check if some of the operands are |
8018 | // constants. To match the legacy cost model's behavior, use SCEV to try |
8019 | // to replace operands with constants. |
8020 | ScalarEvolution &SE = *PSE.getSE(); |
8021 | auto GetConstantViaSCEV = [this, &SE](VPValue *Op) { |
8022 | if (!Op->isLiveIn()) |
8023 | return Op; |
8024 | Value *V = Op->getUnderlyingValue(); |
8025 | if (isa<Constant>(Val: V) || !SE.isSCEVable(Ty: V->getType())) |
8026 | return Op; |
8027 | auto *C = dyn_cast<SCEVConstant>(Val: SE.getSCEV(V)); |
8028 | if (!C) |
8029 | return Op; |
8030 | return Plan.getOrAddLiveIn(V: C->getValue()); |
8031 | }; |
8032 | // For Mul, the legacy cost model checks both operands. |
8033 | if (I->getOpcode() == Instruction::Mul) |
8034 | NewOps[0] = GetConstantViaSCEV(NewOps[0]); |
8035 | // For other binops, the legacy cost model only checks the second operand. |
8036 | NewOps[1] = GetConstantViaSCEV(NewOps[1]); |
8037 | } |
8038 | return new VPWidenRecipe(*I, NewOps); |
8039 | } |
8040 | case Instruction::ExtractValue: { |
8041 | SmallVector<VPValue *> NewOps(Operands); |
8042 | Type *I32Ty = IntegerType::getInt32Ty(C&: I->getContext()); |
8043 | auto *EVI = cast<ExtractValueInst>(Val: I); |
8044 | assert(EVI->getNumIndices() == 1 && "Expected one extractvalue index" ); |
8045 | unsigned Idx = EVI->getIndices()[0]; |
8046 | NewOps.push_back(Elt: Plan.getOrAddLiveIn(V: ConstantInt::get(Ty: I32Ty, V: Idx, IsSigned: false))); |
8047 | return new VPWidenRecipe(*I, NewOps); |
8048 | } |
8049 | }; |
8050 | } |
8051 | |
8052 | VPHistogramRecipe * |
8053 | VPRecipeBuilder::tryToWidenHistogram(const HistogramInfo *HI, |
8054 | ArrayRef<VPValue *> Operands) { |
8055 | // FIXME: Support other operations. |
8056 | unsigned Opcode = HI->Update->getOpcode(); |
8057 | assert((Opcode == Instruction::Add || Opcode == Instruction::Sub) && |
8058 | "Histogram update operation must be an Add or Sub" ); |
8059 | |
8060 | SmallVector<VPValue *, 3> HGramOps; |
8061 | // Bucket address. |
8062 | HGramOps.push_back(Elt: Operands[1]); |
8063 | // Increment value. |
8064 | HGramOps.push_back(Elt: getVPValueOrAddLiveIn(V: HI->Update->getOperand(i: 1))); |
8065 | |
8066 | // In case of predicated execution (due to tail-folding, or conditional |
8067 | // execution, or both), pass the relevant mask. |
8068 | if (Legal->isMaskRequired(I: HI->Store)) |
8069 | HGramOps.push_back(Elt: getBlockInMask(VPBB: Builder.getInsertBlock())); |
8070 | |
8071 | return new VPHistogramRecipe(Opcode, HGramOps, HI->Store->getDebugLoc()); |
8072 | } |
8073 | |
8074 | VPReplicateRecipe * |
8075 | VPRecipeBuilder::handleReplication(Instruction *I, ArrayRef<VPValue *> Operands, |
8076 | VFRange &Range) { |
8077 | bool IsUniform = LoopVectorizationPlanner::getDecisionAndClampRange( |
8078 | Predicate: [&](ElementCount VF) { return CM.isUniformAfterVectorization(I, VF); }, |
8079 | Range); |
8080 | |
8081 | bool IsPredicated = CM.isPredicatedInst(I); |
8082 | |
8083 | // Even if the instruction is not marked as uniform, there are certain |
8084 | // intrinsic calls that can be effectively treated as such, so we check for |
8085 | // them here. Conservatively, we only do this for scalable vectors, since |
8086 | // for fixed-width VFs we can always fall back on full scalarization. |
8087 | if (!IsUniform && Range.Start.isScalable() && isa<IntrinsicInst>(Val: I)) { |
8088 | switch (cast<IntrinsicInst>(Val: I)->getIntrinsicID()) { |
8089 | case Intrinsic::assume: |
8090 | case Intrinsic::lifetime_start: |
8091 | case Intrinsic::lifetime_end: |
8092 | // For scalable vectors if one of the operands is variant then we still |
8093 | // want to mark as uniform, which will generate one instruction for just |
8094 | // the first lane of the vector. We can't scalarize the call in the same |
8095 | // way as for fixed-width vectors because we don't know how many lanes |
8096 | // there are. |
8097 | // |
8098 | // The reasons for doing it this way for scalable vectors are: |
8099 | // 1. For the assume intrinsic generating the instruction for the first |
8100 | // lane is still be better than not generating any at all. For |
8101 | // example, the input may be a splat across all lanes. |
8102 | // 2. For the lifetime start/end intrinsics the pointer operand only |
8103 | // does anything useful when the input comes from a stack object, |
8104 | // which suggests it should always be uniform. For non-stack objects |
8105 | // the effect is to poison the object, which still allows us to |
8106 | // remove the call. |
8107 | IsUniform = true; |
8108 | break; |
8109 | default: |
8110 | break; |
8111 | } |
8112 | } |
8113 | VPValue *BlockInMask = nullptr; |
8114 | if (!IsPredicated) { |
8115 | // Finalize the recipe for Instr, first if it is not predicated. |
8116 | LLVM_DEBUG(dbgs() << "LV: Scalarizing:" << *I << "\n" ); |
8117 | } else { |
8118 | LLVM_DEBUG(dbgs() << "LV: Scalarizing and predicating:" << *I << "\n" ); |
8119 | // Instructions marked for predication are replicated and a mask operand is |
8120 | // added initially. Masked replicate recipes will later be placed under an |
8121 | // if-then construct to prevent side-effects. Generate recipes to compute |
8122 | // the block mask for this region. |
8123 | BlockInMask = getBlockInMask(VPBB: Builder.getInsertBlock()); |
8124 | } |
8125 | |
8126 | // Note that there is some custom logic to mark some intrinsics as uniform |
8127 | // manually above for scalable vectors, which this assert needs to account for |
8128 | // as well. |
8129 | assert((Range.Start.isScalar() || !IsUniform || !IsPredicated || |
8130 | (Range.Start.isScalable() && isa<IntrinsicInst>(I))) && |
8131 | "Should not predicate a uniform recipe" ); |
8132 | auto *Recipe = new VPReplicateRecipe(I, Operands, IsUniform, BlockInMask, |
8133 | VPIRMetadata(*I, LVer)); |
8134 | return Recipe; |
8135 | } |
8136 | |
8137 | /// Find all possible partial reductions in the loop and track all of those that |
8138 | /// are valid so recipes can be formed later. |
8139 | void VPRecipeBuilder::collectScaledReductions(VFRange &Range) { |
8140 | // Find all possible partial reductions. |
8141 | SmallVector<std::pair<PartialReductionChain, unsigned>> |
8142 | PartialReductionChains; |
8143 | for (const auto &[Phi, RdxDesc] : Legal->getReductionVars()) { |
8144 | getScaledReductions(PHI: Phi, RdxExitInstr: RdxDesc.getLoopExitInstr(), Range, |
8145 | Chains&: PartialReductionChains); |
8146 | } |
8147 | |
8148 | // A partial reduction is invalid if any of its extends are used by |
8149 | // something that isn't another partial reduction. This is because the |
8150 | // extends are intended to be lowered along with the reduction itself. |
8151 | |
8152 | // Build up a set of partial reduction ops for efficient use checking. |
8153 | SmallSet<User *, 4> PartialReductionOps; |
8154 | for (const auto &[PartialRdx, _] : PartialReductionChains) |
8155 | PartialReductionOps.insert(Ptr: PartialRdx.ExtendUser); |
8156 | |
8157 | auto ExtendIsOnlyUsedByPartialReductions = |
8158 | [&PartialReductionOps](Instruction *Extend) { |
8159 | return all_of(Range: Extend->users(), P: [&](const User *U) { |
8160 | return PartialReductionOps.contains(Ptr: U); |
8161 | }); |
8162 | }; |
8163 | |
8164 | // Check if each use of a chain's two extends is a partial reduction |
8165 | // and only add those that don't have non-partial reduction users. |
8166 | for (auto Pair : PartialReductionChains) { |
8167 | PartialReductionChain Chain = Pair.first; |
8168 | if (ExtendIsOnlyUsedByPartialReductions(Chain.ExtendA) && |
8169 | (!Chain.ExtendB || ExtendIsOnlyUsedByPartialReductions(Chain.ExtendB))) |
8170 | ScaledReductionMap.try_emplace(Key: Chain.Reduction, Args&: Pair.second); |
8171 | } |
8172 | } |
8173 | |
8174 | bool VPRecipeBuilder::getScaledReductions( |
8175 | Instruction *PHI, Instruction *RdxExitInstr, VFRange &Range, |
8176 | SmallVectorImpl<std::pair<PartialReductionChain, unsigned>> &Chains) { |
8177 | if (!CM.TheLoop->contains(Inst: RdxExitInstr)) |
8178 | return false; |
8179 | |
8180 | auto *Update = dyn_cast<BinaryOperator>(Val: RdxExitInstr); |
8181 | if (!Update) |
8182 | return false; |
8183 | |
8184 | Value *Op = Update->getOperand(i_nocapture: 0); |
8185 | Value *PhiOp = Update->getOperand(i_nocapture: 1); |
8186 | if (Op == PHI) |
8187 | std::swap(a&: Op, b&: PhiOp); |
8188 | |
8189 | // Try and get a scaled reduction from the first non-phi operand. |
8190 | // If one is found, we use the discovered reduction instruction in |
8191 | // place of the accumulator for costing. |
8192 | if (auto *OpInst = dyn_cast<Instruction>(Val: Op)) { |
8193 | if (getScaledReductions(PHI, RdxExitInstr: OpInst, Range, Chains)) { |
8194 | PHI = Chains.rbegin()->first.Reduction; |
8195 | |
8196 | Op = Update->getOperand(i_nocapture: 0); |
8197 | PhiOp = Update->getOperand(i_nocapture: 1); |
8198 | if (Op == PHI) |
8199 | std::swap(a&: Op, b&: PhiOp); |
8200 | } |
8201 | } |
8202 | if (PhiOp != PHI) |
8203 | return false; |
8204 | |
8205 | using namespace llvm::PatternMatch; |
8206 | |
8207 | // If the update is a binary operator, check both of its operands to see if |
8208 | // they are extends. Otherwise, see if the update comes directly from an |
8209 | // extend. |
8210 | Instruction *Exts[2] = {nullptr}; |
8211 | BinaryOperator *ExtendUser = dyn_cast<BinaryOperator>(Val: Op); |
8212 | std::optional<unsigned> BinOpc; |
8213 | Type *ExtOpTypes[2] = {nullptr}; |
8214 | |
8215 | auto CollectExtInfo = [&Exts, |
8216 | &ExtOpTypes](SmallVectorImpl<Value *> &Ops) -> bool { |
8217 | unsigned I = 0; |
8218 | for (Value *OpI : Ops) { |
8219 | Value *ExtOp; |
8220 | if (!match(V: OpI, P: m_ZExtOrSExt(Op: m_Value(V&: ExtOp)))) |
8221 | return false; |
8222 | Exts[I] = cast<Instruction>(Val: OpI); |
8223 | ExtOpTypes[I] = ExtOp->getType(); |
8224 | I++; |
8225 | } |
8226 | return true; |
8227 | }; |
8228 | |
8229 | if (ExtendUser) { |
8230 | if (!ExtendUser->hasOneUse()) |
8231 | return false; |
8232 | |
8233 | // Use the side-effect of match to replace BinOp only if the pattern is |
8234 | // matched, we don't care at this point whether it actually matched. |
8235 | match(V: ExtendUser, P: m_Neg(V: m_BinOp(I&: ExtendUser))); |
8236 | |
8237 | SmallVector<Value *> Ops(ExtendUser->operands()); |
8238 | if (!CollectExtInfo(Ops)) |
8239 | return false; |
8240 | |
8241 | BinOpc = std::make_optional(t: ExtendUser->getOpcode()); |
8242 | } else if (match(V: Update, P: m_Add(L: m_Value(), R: m_Value()))) { |
8243 | // We already know the operands for Update are Op and PhiOp. |
8244 | SmallVector<Value *> Ops({Op}); |
8245 | if (!CollectExtInfo(Ops)) |
8246 | return false; |
8247 | |
8248 | ExtendUser = Update; |
8249 | BinOpc = std::nullopt; |
8250 | } else |
8251 | return false; |
8252 | |
8253 | TTI::PartialReductionExtendKind OpAExtend = |
8254 | TTI::getPartialReductionExtendKind(I: Exts[0]); |
8255 | TTI::PartialReductionExtendKind OpBExtend = |
8256 | Exts[1] ? TTI::getPartialReductionExtendKind(I: Exts[1]) : TTI::PR_None; |
8257 | PartialReductionChain Chain(RdxExitInstr, Exts[0], Exts[1], ExtendUser); |
8258 | |
8259 | TypeSize PHISize = PHI->getType()->getPrimitiveSizeInBits(); |
8260 | TypeSize ASize = ExtOpTypes[0]->getPrimitiveSizeInBits(); |
8261 | if (!PHISize.hasKnownScalarFactor(RHS: ASize)) |
8262 | return false; |
8263 | unsigned TargetScaleFactor = PHISize.getKnownScalarFactor(RHS: ASize); |
8264 | |
8265 | if (LoopVectorizationPlanner::getDecisionAndClampRange( |
8266 | Predicate: [&](ElementCount VF) { |
8267 | InstructionCost Cost = TTI->getPartialReductionCost( |
8268 | Opcode: Update->getOpcode(), InputTypeA: ExtOpTypes[0], InputTypeB: ExtOpTypes[1], |
8269 | AccumType: PHI->getType(), VF, OpAExtend, OpBExtend, BinOp: BinOpc, CostKind: CM.CostKind); |
8270 | return Cost.isValid(); |
8271 | }, |
8272 | Range)) { |
8273 | Chains.emplace_back(Args&: Chain, Args&: TargetScaleFactor); |
8274 | return true; |
8275 | } |
8276 | |
8277 | return false; |
8278 | } |
8279 | |
8280 | VPRecipeBase *VPRecipeBuilder::tryToCreateWidenRecipe(VPSingleDefRecipe *R, |
8281 | VFRange &Range) { |
8282 | // First, check for specific widening recipes that deal with inductions, Phi |
8283 | // nodes, calls and memory operations. |
8284 | VPRecipeBase *Recipe; |
8285 | Instruction *Instr = R->getUnderlyingInstr(); |
8286 | SmallVector<VPValue *, 4> Operands(R->operands()); |
8287 | if (auto *PhiR = dyn_cast<VPWidenPHIRecipe>(Val: R)) { |
8288 | VPBasicBlock *Parent = PhiR->getParent(); |
8289 | [[maybe_unused]] VPRegionBlock *LoopRegionOf = |
8290 | Parent->getEnclosingLoopRegion(); |
8291 | assert(LoopRegionOf && LoopRegionOf->getEntry() == Parent && |
8292 | "Non-header phis should have been handled during predication" ); |
8293 | auto *Phi = cast<PHINode>(Val: R->getUnderlyingInstr()); |
8294 | assert(Operands.size() == 2 && "Must have 2 operands for header phis" ); |
8295 | if ((Recipe = tryToOptimizeInductionPHI(Phi, Operands, Range))) |
8296 | return Recipe; |
8297 | |
8298 | VPHeaderPHIRecipe *PhiRecipe = nullptr; |
8299 | assert((Legal->isReductionVariable(Phi) || |
8300 | Legal->isFixedOrderRecurrence(Phi)) && |
8301 | "can only widen reductions and fixed-order recurrences here" ); |
8302 | VPValue *StartV = Operands[0]; |
8303 | if (Legal->isReductionVariable(PN: Phi)) { |
8304 | const RecurrenceDescriptor &RdxDesc = |
8305 | Legal->getReductionVars().find(Key: Phi)->second; |
8306 | assert(RdxDesc.getRecurrenceStartValue() == |
8307 | Phi->getIncomingValueForBlock(OrigLoop->getLoopPreheader())); |
8308 | |
8309 | // If the PHI is used by a partial reduction, set the scale factor. |
8310 | unsigned ScaleFactor = |
8311 | getScalingForReduction(ExitInst: RdxDesc.getLoopExitInstr()).value_or(u: 1); |
8312 | PhiRecipe = new VPReductionPHIRecipe( |
8313 | Phi, RdxDesc, *StartV, CM.isInLoopReduction(Phi), |
8314 | CM.useOrderedReductions(RdxDesc), ScaleFactor); |
8315 | } else { |
8316 | // TODO: Currently fixed-order recurrences are modeled as chains of |
8317 | // first-order recurrences. If there are no users of the intermediate |
8318 | // recurrences in the chain, the fixed order recurrence should be modeled |
8319 | // directly, enabling more efficient codegen. |
8320 | PhiRecipe = new VPFirstOrderRecurrencePHIRecipe(Phi, *StartV); |
8321 | } |
8322 | // Add backedge value. |
8323 | PhiRecipe->addOperand(Operand: Operands[1]); |
8324 | return PhiRecipe; |
8325 | } |
8326 | |
8327 | if (isa<TruncInst>(Val: Instr) && (Recipe = tryToOptimizeInductionTruncate( |
8328 | I: cast<TruncInst>(Val: Instr), Operands, Range))) |
8329 | return Recipe; |
8330 | |
8331 | // All widen recipes below deal only with VF > 1. |
8332 | if (LoopVectorizationPlanner::getDecisionAndClampRange( |
8333 | Predicate: [&](ElementCount VF) { return VF.isScalar(); }, Range)) |
8334 | return nullptr; |
8335 | |
8336 | if (auto *CI = dyn_cast<CallInst>(Val: Instr)) |
8337 | return tryToWidenCall(CI, Operands, Range); |
8338 | |
8339 | if (StoreInst *SI = dyn_cast<StoreInst>(Val: Instr)) |
8340 | if (auto HistInfo = Legal->getHistogramInfo(I: SI)) |
8341 | return tryToWidenHistogram(HI: *HistInfo, Operands); |
8342 | |
8343 | if (isa<LoadInst>(Val: Instr) || isa<StoreInst>(Val: Instr)) |
8344 | return tryToWidenMemory(I: Instr, Operands, Range); |
8345 | |
8346 | if (std::optional<unsigned> ScaleFactor = getScalingForReduction(ExitInst: Instr)) |
8347 | return tryToCreatePartialReduction(Reduction: Instr, Operands, ScaleFactor: ScaleFactor.value()); |
8348 | |
8349 | if (!shouldWiden(I: Instr, Range)) |
8350 | return nullptr; |
8351 | |
8352 | if (auto *GEP = dyn_cast<GetElementPtrInst>(Val: Instr)) |
8353 | return new VPWidenGEPRecipe(GEP, Operands); |
8354 | |
8355 | if (auto *SI = dyn_cast<SelectInst>(Val: Instr)) { |
8356 | return new VPWidenSelectRecipe(*SI, Operands); |
8357 | } |
8358 | |
8359 | if (auto *CI = dyn_cast<CastInst>(Val: Instr)) { |
8360 | return new VPWidenCastRecipe(CI->getOpcode(), Operands[0], CI->getType(), |
8361 | *CI); |
8362 | } |
8363 | |
8364 | return tryToWiden(I: Instr, Operands); |
8365 | } |
8366 | |
8367 | VPRecipeBase * |
8368 | VPRecipeBuilder::tryToCreatePartialReduction(Instruction *Reduction, |
8369 | ArrayRef<VPValue *> Operands, |
8370 | unsigned ScaleFactor) { |
8371 | assert(Operands.size() == 2 && |
8372 | "Unexpected number of operands for partial reduction" ); |
8373 | |
8374 | VPValue *BinOp = Operands[0]; |
8375 | VPValue *Accumulator = Operands[1]; |
8376 | VPRecipeBase *BinOpRecipe = BinOp->getDefiningRecipe(); |
8377 | if (isa<VPReductionPHIRecipe>(Val: BinOpRecipe) || |
8378 | isa<VPPartialReductionRecipe>(Val: BinOpRecipe)) |
8379 | std::swap(a&: BinOp, b&: Accumulator); |
8380 | |
8381 | unsigned ReductionOpcode = Reduction->getOpcode(); |
8382 | if (ReductionOpcode == Instruction::Sub) { |
8383 | auto *const Zero = ConstantInt::get(Ty: Reduction->getType(), V: 0); |
8384 | SmallVector<VPValue *, 2> Ops; |
8385 | Ops.push_back(Elt: Plan.getOrAddLiveIn(V: Zero)); |
8386 | Ops.push_back(Elt: BinOp); |
8387 | BinOp = new VPWidenRecipe(*Reduction, Ops); |
8388 | Builder.insert(R: BinOp->getDefiningRecipe()); |
8389 | ReductionOpcode = Instruction::Add; |
8390 | } |
8391 | |
8392 | VPValue *Cond = nullptr; |
8393 | if (CM.blockNeedsPredicationForAnyReason(BB: Reduction->getParent())) { |
8394 | assert((ReductionOpcode == Instruction::Add || |
8395 | ReductionOpcode == Instruction::Sub) && |
8396 | "Expected an ADD or SUB operation for predicated partial " |
8397 | "reductions (because the neutral element in the mask is zero)!" ); |
8398 | Cond = getBlockInMask(VPBB: Builder.getInsertBlock()); |
8399 | VPValue *Zero = |
8400 | Plan.getOrAddLiveIn(V: ConstantInt::get(Ty: Reduction->getType(), V: 0)); |
8401 | BinOp = Builder.createSelect(Cond, TrueVal: BinOp, FalseVal: Zero, DL: Reduction->getDebugLoc()); |
8402 | } |
8403 | return new VPPartialReductionRecipe(ReductionOpcode, Accumulator, BinOp, Cond, |
8404 | ScaleFactor, Reduction); |
8405 | } |
8406 | |
8407 | void LoopVectorizationPlanner::buildVPlansWithVPRecipes(ElementCount MinVF, |
8408 | ElementCount MaxVF) { |
8409 | if (ElementCount::isKnownGT(LHS: MinVF, RHS: MaxVF)) |
8410 | return; |
8411 | |
8412 | assert(OrigLoop->isInnermost() && "Inner loop expected." ); |
8413 | |
8414 | const LoopAccessInfo *LAI = Legal->getLAI(); |
8415 | LoopVersioning LVer(*LAI, LAI->getRuntimePointerChecking()->getChecks(), |
8416 | OrigLoop, LI, DT, PSE.getSE()); |
8417 | if (!LAI->getRuntimePointerChecking()->getChecks().empty() && |
8418 | !LAI->getRuntimePointerChecking()->getDiffChecks()) { |
8419 | // Only use noalias metadata when using memory checks guaranteeing no |
8420 | // overlap across all iterations. |
8421 | LVer.prepareNoAliasMetadata(); |
8422 | } |
8423 | |
8424 | auto MaxVFTimes2 = MaxVF * 2; |
8425 | auto VPlan0 = VPlanTransforms::buildPlainCFG(TheLoop: OrigLoop, LI&: *LI); |
8426 | for (ElementCount VF = MinVF; ElementCount::isKnownLT(LHS: VF, RHS: MaxVFTimes2);) { |
8427 | VFRange SubRange = {VF, MaxVFTimes2}; |
8428 | if (auto Plan = tryToBuildVPlanWithVPRecipes( |
8429 | InitialPlan: std::unique_ptr<VPlan>(VPlan0->duplicate()), Range&: SubRange, LVer: &LVer)) { |
8430 | bool HasScalarVF = Plan->hasScalarVFOnly(); |
8431 | // Now optimize the initial VPlan. |
8432 | if (!HasScalarVF) |
8433 | VPlanTransforms::runPass(Fn: VPlanTransforms::truncateToMinimalBitwidths, |
8434 | Plan&: *Plan, Args: CM.getMinimalBitwidths()); |
8435 | VPlanTransforms::runPass(Fn: VPlanTransforms::optimize, Plan&: *Plan); |
8436 | // TODO: try to put it close to addActiveLaneMask(). |
8437 | // Discard the plan if it is not EVL-compatible |
8438 | if (CM.foldTailWithEVL() && !HasScalarVF && |
8439 | !VPlanTransforms::runPass(Transform: VPlanTransforms::tryAddExplicitVectorLength, |
8440 | Plan&: *Plan, Args: CM.getMaxSafeElements())) |
8441 | break; |
8442 | assert(verifyVPlanIsValid(*Plan) && "VPlan is invalid" ); |
8443 | VPlans.push_back(Elt: std::move(Plan)); |
8444 | } |
8445 | VF = SubRange.End; |
8446 | } |
8447 | } |
8448 | |
8449 | /// Create and return a ResumePhi for \p WideIV, unless it is truncated. If the |
8450 | /// induction recipe is not canonical, creates a VPDerivedIVRecipe to compute |
8451 | /// the end value of the induction. |
8452 | static VPInstruction *addResumePhiRecipeForInduction( |
8453 | VPWidenInductionRecipe *WideIV, VPBuilder &VectorPHBuilder, |
8454 | VPBuilder &ScalarPHBuilder, VPTypeAnalysis &TypeInfo, VPValue *VectorTC) { |
8455 | auto *WideIntOrFp = dyn_cast<VPWidenIntOrFpInductionRecipe>(Val: WideIV); |
8456 | // Truncated wide inductions resume from the last lane of their vector value |
8457 | // in the last vector iteration which is handled elsewhere. |
8458 | if (WideIntOrFp && WideIntOrFp->getTruncInst()) |
8459 | return nullptr; |
8460 | |
8461 | VPValue *Start = WideIV->getStartValue(); |
8462 | VPValue *Step = WideIV->getStepValue(); |
8463 | const InductionDescriptor &ID = WideIV->getInductionDescriptor(); |
8464 | VPValue *EndValue = VectorTC; |
8465 | if (!WideIntOrFp || !WideIntOrFp->isCanonical()) { |
8466 | EndValue = VectorPHBuilder.createDerivedIV( |
8467 | Kind: ID.getKind(), FPBinOp: dyn_cast_or_null<FPMathOperator>(Val: ID.getInductionBinOp()), |
8468 | Start, Current: VectorTC, Step); |
8469 | } |
8470 | |
8471 | // EndValue is derived from the vector trip count (which has the same type as |
8472 | // the widest induction) and thus may be wider than the induction here. |
8473 | Type *ScalarTypeOfWideIV = TypeInfo.inferScalarType(V: WideIV); |
8474 | if (ScalarTypeOfWideIV != TypeInfo.inferScalarType(V: EndValue)) { |
8475 | EndValue = VectorPHBuilder.createScalarCast(Opcode: Instruction::Trunc, Op: EndValue, |
8476 | ResultTy: ScalarTypeOfWideIV, |
8477 | DL: WideIV->getDebugLoc()); |
8478 | } |
8479 | |
8480 | auto *ResumePhiRecipe = ScalarPHBuilder.createScalarPhi( |
8481 | IncomingValues: {EndValue, Start}, DL: WideIV->getDebugLoc(), Name: "bc.resume.val" ); |
8482 | return ResumePhiRecipe; |
8483 | } |
8484 | |
8485 | /// Create resume phis in the scalar preheader for first-order recurrences, |
8486 | /// reductions and inductions, and update the VPIRInstructions wrapping the |
8487 | /// original phis in the scalar header. End values for inductions are added to |
8488 | /// \p IVEndValues. |
8489 | static void addScalarResumePhis(VPRecipeBuilder &Builder, VPlan &Plan, |
8490 | DenseMap<VPValue *, VPValue *> &IVEndValues) { |
8491 | VPTypeAnalysis TypeInfo(Plan.getCanonicalIV()->getScalarType()); |
8492 | auto *ScalarPH = Plan.getScalarPreheader(); |
8493 | auto *MiddleVPBB = cast<VPBasicBlock>(Val: ScalarPH->getPredecessors()[0]); |
8494 | VPRegionBlock *VectorRegion = Plan.getVectorLoopRegion(); |
8495 | VPBuilder VectorPHBuilder( |
8496 | cast<VPBasicBlock>(Val: VectorRegion->getSinglePredecessor())); |
8497 | VPBuilder MiddleBuilder(MiddleVPBB, MiddleVPBB->getFirstNonPhi()); |
8498 | VPBuilder ScalarPHBuilder(ScalarPH); |
8499 | for (VPRecipeBase &ScalarPhiR : Plan.getScalarHeader()->phis()) { |
8500 | auto *ScalarPhiIRI = cast<VPIRPhi>(Val: &ScalarPhiR); |
8501 | |
8502 | // TODO: Extract final value from induction recipe initially, optimize to |
8503 | // pre-computed end value together in optimizeInductionExitUsers. |
8504 | auto *VectorPhiR = |
8505 | cast<VPHeaderPHIRecipe>(Val: Builder.getRecipe(I: &ScalarPhiIRI->getIRPhi())); |
8506 | if (auto *WideIVR = dyn_cast<VPWidenInductionRecipe>(Val: VectorPhiR)) { |
8507 | if (VPInstruction *ResumePhi = addResumePhiRecipeForInduction( |
8508 | WideIV: WideIVR, VectorPHBuilder, ScalarPHBuilder, TypeInfo, |
8509 | VectorTC: &Plan.getVectorTripCount())) { |
8510 | assert(isa<VPPhi>(ResumePhi) && "Expected a phi" ); |
8511 | IVEndValues[WideIVR] = ResumePhi->getOperand(N: 0); |
8512 | ScalarPhiIRI->addOperand(Operand: ResumePhi); |
8513 | continue; |
8514 | } |
8515 | // TODO: Also handle truncated inductions here. Computing end-values |
8516 | // separately should be done as VPlan-to-VPlan optimization, after |
8517 | // legalizing all resume values to use the last lane from the loop. |
8518 | assert(cast<VPWidenIntOrFpInductionRecipe>(VectorPhiR)->getTruncInst() && |
8519 | "should only skip truncated wide inductions" ); |
8520 | continue; |
8521 | } |
8522 | |
8523 | // The backedge value provides the value to resume coming out of a loop, |
8524 | // which for FORs is a vector whose last element needs to be extracted. The |
8525 | // start value provides the value if the loop is bypassed. |
8526 | bool IsFOR = isa<VPFirstOrderRecurrencePHIRecipe>(Val: VectorPhiR); |
8527 | auto *ResumeFromVectorLoop = VectorPhiR->getBackedgeValue(); |
8528 | assert(VectorRegion->getSingleSuccessor() == Plan.getMiddleBlock() && |
8529 | "Cannot handle loops with uncountable early exits" ); |
8530 | if (IsFOR) |
8531 | ResumeFromVectorLoop = MiddleBuilder.createNaryOp( |
8532 | Opcode: VPInstruction::ExtractLastElement, Operands: {ResumeFromVectorLoop}, Inst: {}, |
8533 | Name: "vector.recur.extract" ); |
8534 | StringRef Name = IsFOR ? "scalar.recur.init" : "bc.merge.rdx" ; |
8535 | auto *ResumePhiR = ScalarPHBuilder.createScalarPhi( |
8536 | IncomingValues: {ResumeFromVectorLoop, VectorPhiR->getStartValue()}, DL: {}, Name); |
8537 | ScalarPhiIRI->addOperand(Operand: ResumePhiR); |
8538 | } |
8539 | } |
8540 | |
8541 | // Collect VPIRInstructions for phis in the exit block from the latch only. |
8542 | static SetVector<VPIRInstruction *> collectUsersInLatchExitBlock(VPlan &Plan) { |
8543 | SetVector<VPIRInstruction *> ExitUsersToFix; |
8544 | for (VPIRBasicBlock *ExitVPBB : Plan.getExitBlocks()) { |
8545 | |
8546 | if (ExitVPBB->getSinglePredecessor() != Plan.getMiddleBlock()) |
8547 | continue; |
8548 | |
8549 | for (VPRecipeBase &R : ExitVPBB->phis()) { |
8550 | auto *ExitIRI = cast<VPIRPhi>(Val: &R); |
8551 | assert(ExitIRI->getNumOperands() == 1 && "must have a single operand" ); |
8552 | VPValue *V = ExitIRI->getOperand(N: 0); |
8553 | if (V->isLiveIn()) |
8554 | continue; |
8555 | assert(V->getDefiningRecipe()->getParent()->getEnclosingLoopRegion() && |
8556 | "Only recipes defined inside a region should need fixing." ); |
8557 | ExitUsersToFix.insert(X: ExitIRI); |
8558 | } |
8559 | } |
8560 | return ExitUsersToFix; |
8561 | } |
8562 | |
8563 | // Add exit values to \p Plan. Extracts are added for each entry in \p |
8564 | // ExitUsersToFix if needed and their operands are updated. |
8565 | static void |
8566 | addUsersInExitBlocks(VPlan &Plan, |
8567 | const SetVector<VPIRInstruction *> &ExitUsersToFix) { |
8568 | if (ExitUsersToFix.empty()) |
8569 | return; |
8570 | |
8571 | auto *MiddleVPBB = Plan.getMiddleBlock(); |
8572 | VPBuilder B(MiddleVPBB, MiddleVPBB->getFirstNonPhi()); |
8573 | |
8574 | // Introduce extract for exiting values and update the VPIRInstructions |
8575 | // modeling the corresponding LCSSA phis. |
8576 | for (VPIRInstruction *ExitIRI : ExitUsersToFix) { |
8577 | assert(ExitIRI->getNumOperands() == 1 && |
8578 | ExitIRI->getParent()->getSinglePredecessor() == MiddleVPBB && |
8579 | "exit values from early exits must be fixed when branch to " |
8580 | "early-exit is added" ); |
8581 | ExitIRI->extractLastLaneOfFirstOperand(Builder&: B); |
8582 | } |
8583 | } |
8584 | |
8585 | /// Handle users in the exit block for first order reductions in the original |
8586 | /// exit block. The penultimate value of recurrences is fed to their LCSSA phi |
8587 | /// users in the original exit block using the VPIRInstruction wrapping to the |
8588 | /// LCSSA phi. |
8589 | static void addExitUsersForFirstOrderRecurrences( |
8590 | VPlan &Plan, SetVector<VPIRInstruction *> &ExitUsersToFix, VFRange &Range) { |
8591 | VPRegionBlock *VectorRegion = Plan.getVectorLoopRegion(); |
8592 | auto *ScalarPHVPBB = Plan.getScalarPreheader(); |
8593 | auto *MiddleVPBB = Plan.getMiddleBlock(); |
8594 | VPBuilder ScalarPHBuilder(ScalarPHVPBB); |
8595 | VPBuilder MiddleBuilder(MiddleVPBB, MiddleVPBB->getFirstNonPhi()); |
8596 | |
8597 | auto IsScalableOne = [](ElementCount VF) -> bool { |
8598 | return VF == ElementCount::getScalable(MinVal: 1); |
8599 | }; |
8600 | |
8601 | for (auto & : VectorRegion->getEntryBasicBlock()->phis()) { |
8602 | auto *FOR = dyn_cast<VPFirstOrderRecurrencePHIRecipe>(Val: &HeaderPhi); |
8603 | if (!FOR) |
8604 | continue; |
8605 | |
8606 | assert(VectorRegion->getSingleSuccessor() == Plan.getMiddleBlock() && |
8607 | "Cannot handle loops with uncountable early exits" ); |
8608 | |
8609 | // This is the second phase of vectorizing first-order recurrences, creating |
8610 | // extract for users outside the loop. An overview of the transformation is |
8611 | // described below. Suppose we have the following loop with some use after |
8612 | // the loop of the last a[i-1], |
8613 | // |
8614 | // for (int i = 0; i < n; ++i) { |
8615 | // t = a[i - 1]; |
8616 | // b[i] = a[i] - t; |
8617 | // } |
8618 | // use t; |
8619 | // |
8620 | // There is a first-order recurrence on "a". For this loop, the shorthand |
8621 | // scalar IR looks like: |
8622 | // |
8623 | // scalar.ph: |
8624 | // s.init = a[-1] |
8625 | // br scalar.body |
8626 | // |
8627 | // scalar.body: |
8628 | // i = phi [0, scalar.ph], [i+1, scalar.body] |
8629 | // s1 = phi [s.init, scalar.ph], [s2, scalar.body] |
8630 | // s2 = a[i] |
8631 | // b[i] = s2 - s1 |
8632 | // br cond, scalar.body, exit.block |
8633 | // |
8634 | // exit.block: |
8635 | // use = lcssa.phi [s1, scalar.body] |
8636 | // |
8637 | // In this example, s1 is a recurrence because it's value depends on the |
8638 | // previous iteration. In the first phase of vectorization, we created a |
8639 | // VPFirstOrderRecurrencePHIRecipe v1 for s1. Now we create the extracts |
8640 | // for users in the scalar preheader and exit block. |
8641 | // |
8642 | // vector.ph: |
8643 | // v_init = vector(..., ..., ..., a[-1]) |
8644 | // br vector.body |
8645 | // |
8646 | // vector.body |
8647 | // i = phi [0, vector.ph], [i+4, vector.body] |
8648 | // v1 = phi [v_init, vector.ph], [v2, vector.body] |
8649 | // v2 = a[i, i+1, i+2, i+3] |
8650 | // b[i] = v2 - v1 |
8651 | // // Next, third phase will introduce v1' = splice(v1(3), v2(0, 1, 2)) |
8652 | // b[i, i+1, i+2, i+3] = v2 - v1 |
8653 | // br cond, vector.body, middle.block |
8654 | // |
8655 | // middle.block: |
8656 | // vector.recur.extract.for.phi = v2(2) |
8657 | // vector.recur.extract = v2(3) |
8658 | // br cond, scalar.ph, exit.block |
8659 | // |
8660 | // scalar.ph: |
8661 | // scalar.recur.init = phi [vector.recur.extract, middle.block], |
8662 | // [s.init, otherwise] |
8663 | // br scalar.body |
8664 | // |
8665 | // scalar.body: |
8666 | // i = phi [0, scalar.ph], [i+1, scalar.body] |
8667 | // s1 = phi [scalar.recur.init, scalar.ph], [s2, scalar.body] |
8668 | // s2 = a[i] |
8669 | // b[i] = s2 - s1 |
8670 | // br cond, scalar.body, exit.block |
8671 | // |
8672 | // exit.block: |
8673 | // lo = lcssa.phi [s1, scalar.body], |
8674 | // [vector.recur.extract.for.phi, middle.block] |
8675 | // |
8676 | // Now update VPIRInstructions modeling LCSSA phis in the exit block. |
8677 | // Extract the penultimate value of the recurrence and use it as operand for |
8678 | // the VPIRInstruction modeling the phi. |
8679 | for (VPIRInstruction *ExitIRI : ExitUsersToFix) { |
8680 | if (ExitIRI->getOperand(N: 0) != FOR) |
8681 | continue; |
8682 | // For VF vscale x 1, if vscale = 1, we are unable to extract the |
8683 | // penultimate value of the recurrence. Instead, we rely on function |
8684 | // addUsersInExitBlocks to extract the last element from the result of |
8685 | // VPInstruction::FirstOrderRecurrenceSplice by leaving the user of the |
8686 | // recurrence phi in ExitUsersToFix. |
8687 | // TODO: Consider vscale_range info and UF. |
8688 | if (LoopVectorizationPlanner::getDecisionAndClampRange(Predicate: IsScalableOne, |
8689 | Range)) |
8690 | return; |
8691 | VPValue *PenultimateElement = MiddleBuilder.createNaryOp( |
8692 | Opcode: VPInstruction::ExtractPenultimateElement, Operands: {FOR->getBackedgeValue()}, |
8693 | Inst: {}, Name: "vector.recur.extract.for.phi" ); |
8694 | ExitIRI->setOperand(I: 0, New: PenultimateElement); |
8695 | ExitUsersToFix.remove(X: ExitIRI); |
8696 | } |
8697 | } |
8698 | } |
8699 | |
8700 | VPlanPtr LoopVectorizationPlanner::tryToBuildVPlanWithVPRecipes( |
8701 | VPlanPtr Plan, VFRange &Range, LoopVersioning *LVer) { |
8702 | |
8703 | using namespace llvm::VPlanPatternMatch; |
8704 | SmallPtrSet<const InterleaveGroup<Instruction> *, 1> InterleaveGroups; |
8705 | |
8706 | // --------------------------------------------------------------------------- |
8707 | // Build initial VPlan: Scan the body of the loop in a topological order to |
8708 | // visit each basic block after having visited its predecessor basic blocks. |
8709 | // --------------------------------------------------------------------------- |
8710 | |
8711 | // Create initial VPlan skeleton, having a basic block for the pre-header |
8712 | // which contains SCEV expansions that need to happen before the CFG is |
8713 | // modified; a basic block for the vector pre-header, followed by a region for |
8714 | // the vector loop, followed by the middle basic block. The skeleton vector |
8715 | // loop region contains a header and latch basic blocks. |
8716 | |
8717 | bool RequiresScalarEpilogueCheck = |
8718 | LoopVectorizationPlanner::getDecisionAndClampRange( |
8719 | Predicate: [this](ElementCount VF) { |
8720 | return !CM.requiresScalarEpilogue(IsVectorizing: VF.isVector()); |
8721 | }, |
8722 | Range); |
8723 | VPlanTransforms::prepareForVectorization( |
8724 | Plan&: *Plan, InductionTy: Legal->getWidestInductionType(), PSE, RequiresScalarEpilogueCheck, |
8725 | TailFolded: CM.foldTailByMasking(), TheLoop: OrigLoop, |
8726 | IVDL: getDebugLocFromInstOrOperands(I: Legal->getPrimaryInduction()), |
8727 | HasUncountableExit: Legal->hasUncountableEarlyExit(), Range); |
8728 | VPlanTransforms::createLoopRegions(Plan&: *Plan); |
8729 | |
8730 | // Don't use getDecisionAndClampRange here, because we don't know the UF |
8731 | // so this function is better to be conservative, rather than to split |
8732 | // it up into different VPlans. |
8733 | // TODO: Consider using getDecisionAndClampRange here to split up VPlans. |
8734 | bool IVUpdateMayOverflow = false; |
8735 | for (ElementCount VF : Range) |
8736 | IVUpdateMayOverflow |= !isIndvarOverflowCheckKnownFalse(Cost: &CM, VF); |
8737 | |
8738 | TailFoldingStyle Style = CM.getTailFoldingStyle(IVUpdateMayOverflow); |
8739 | // Use NUW for the induction increment if we proved that it won't overflow in |
8740 | // the vector loop or when not folding the tail. In the later case, we know |
8741 | // that the canonical induction increment will not overflow as the vector trip |
8742 | // count is >= increment and a multiple of the increment. |
8743 | bool HasNUW = !IVUpdateMayOverflow || Style == TailFoldingStyle::None; |
8744 | if (!HasNUW) { |
8745 | auto *IVInc = Plan->getVectorLoopRegion() |
8746 | ->getExitingBasicBlock() |
8747 | ->getTerminator() |
8748 | ->getOperand(N: 0); |
8749 | assert(match(IVInc, m_VPInstruction<Instruction::Add>( |
8750 | m_Specific(Plan->getCanonicalIV()), m_VPValue())) && |
8751 | "Did not find the canonical IV increment" ); |
8752 | cast<VPRecipeWithIRFlags>(Val: IVInc)->dropPoisonGeneratingFlags(); |
8753 | } |
8754 | |
8755 | // --------------------------------------------------------------------------- |
8756 | // Pre-construction: record ingredients whose recipes we'll need to further |
8757 | // process after constructing the initial VPlan. |
8758 | // --------------------------------------------------------------------------- |
8759 | |
8760 | // For each interleave group which is relevant for this (possibly trimmed) |
8761 | // Range, add it to the set of groups to be later applied to the VPlan and add |
8762 | // placeholders for its members' Recipes which we'll be replacing with a |
8763 | // single VPInterleaveRecipe. |
8764 | for (InterleaveGroup<Instruction> *IG : IAI.getInterleaveGroups()) { |
8765 | auto ApplyIG = [IG, this](ElementCount VF) -> bool { |
8766 | bool Result = (VF.isVector() && // Query is illegal for VF == 1 |
8767 | CM.getWideningDecision(I: IG->getInsertPos(), VF) == |
8768 | LoopVectorizationCostModel::CM_Interleave); |
8769 | // For scalable vectors, the interleave factors must be <= 8 since we |
8770 | // require the (de)interleaveN intrinsics instead of shufflevectors. |
8771 | assert((!Result || !VF.isScalable() || IG->getFactor() <= 8) && |
8772 | "Unsupported interleave factor for scalable vectors" ); |
8773 | return Result; |
8774 | }; |
8775 | if (!getDecisionAndClampRange(Predicate: ApplyIG, Range)) |
8776 | continue; |
8777 | InterleaveGroups.insert(Ptr: IG); |
8778 | } |
8779 | |
8780 | // --------------------------------------------------------------------------- |
8781 | // Predicate and linearize the top-level loop region. |
8782 | // --------------------------------------------------------------------------- |
8783 | auto BlockMaskCache = VPlanTransforms::introduceMasksAndLinearize( |
8784 | Plan&: *Plan, FoldTail: CM.foldTailByMasking()); |
8785 | |
8786 | // --------------------------------------------------------------------------- |
8787 | // Construct wide recipes and apply predication for original scalar |
8788 | // VPInstructions in the loop. |
8789 | // --------------------------------------------------------------------------- |
8790 | VPRecipeBuilder RecipeBuilder(*Plan, OrigLoop, TLI, &TTI, Legal, CM, PSE, |
8791 | Builder, BlockMaskCache, LVer); |
8792 | RecipeBuilder.collectScaledReductions(Range); |
8793 | |
8794 | // Scan the body of the loop in a topological order to visit each basic block |
8795 | // after having visited its predecessor basic blocks. |
8796 | VPRegionBlock *LoopRegion = Plan->getVectorLoopRegion(); |
8797 | VPBasicBlock * = LoopRegion->getEntryBasicBlock(); |
8798 | ReversePostOrderTraversal<VPBlockShallowTraversalWrapper<VPBlockBase *>> RPOT( |
8799 | HeaderVPBB); |
8800 | |
8801 | auto *MiddleVPBB = Plan->getMiddleBlock(); |
8802 | VPBasicBlock::iterator MBIP = MiddleVPBB->getFirstNonPhi(); |
8803 | // Mapping from VPValues in the initial plan to their widened VPValues. Needed |
8804 | // temporarily to update created block masks. |
8805 | DenseMap<VPValue *, VPValue *> Old2New; |
8806 | for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(Range: RPOT)) { |
8807 | // Convert input VPInstructions to widened recipes. |
8808 | for (VPRecipeBase &R : make_early_inc_range(Range&: *VPBB)) { |
8809 | auto *SingleDef = cast<VPSingleDefRecipe>(Val: &R); |
8810 | auto *UnderlyingValue = SingleDef->getUnderlyingValue(); |
8811 | // Skip recipes that do not need transforming, including canonical IV, |
8812 | // wide canonical IV and VPInstructions without underlying values. The |
8813 | // latter are added above for masking. |
8814 | // FIXME: Migrate code relying on the underlying instruction from VPlan0 |
8815 | // to construct recipes below to not use the underlying instruction. |
8816 | if (isa<VPCanonicalIVPHIRecipe, VPWidenCanonicalIVRecipe, VPBlendRecipe>( |
8817 | Val: &R) || |
8818 | (isa<VPInstruction>(Val: &R) && !UnderlyingValue)) |
8819 | continue; |
8820 | |
8821 | // FIXME: VPlan0, which models a copy of the original scalar loop, should |
8822 | // not use VPWidenPHIRecipe to model the phis. |
8823 | assert((isa<VPWidenPHIRecipe>(&R) || isa<VPInstruction>(&R)) && |
8824 | UnderlyingValue && "unsupported recipe" ); |
8825 | |
8826 | // TODO: Gradually replace uses of underlying instruction by analyses on |
8827 | // VPlan. |
8828 | Instruction *Instr = cast<Instruction>(Val: UnderlyingValue); |
8829 | Builder.setInsertPoint(SingleDef); |
8830 | |
8831 | // The stores with invariant address inside the loop will be deleted, and |
8832 | // in the exit block, a uniform store recipe will be created for the final |
8833 | // invariant store of the reduction. |
8834 | StoreInst *SI; |
8835 | if ((SI = dyn_cast<StoreInst>(Val: Instr)) && |
8836 | Legal->isInvariantAddressOfReduction(V: SI->getPointerOperand())) { |
8837 | // Only create recipe for the final invariant store of the reduction. |
8838 | if (Legal->isInvariantStoreOfReduction(SI)) { |
8839 | auto *Recipe = |
8840 | new VPReplicateRecipe(SI, R.operands(), true /* IsUniform */, |
8841 | nullptr /*Mask*/, VPIRMetadata(*SI, LVer)); |
8842 | Recipe->insertBefore(BB&: *MiddleVPBB, IP: MBIP); |
8843 | } |
8844 | R.eraseFromParent(); |
8845 | continue; |
8846 | } |
8847 | |
8848 | VPRecipeBase *Recipe = |
8849 | RecipeBuilder.tryToCreateWidenRecipe(R: SingleDef, Range); |
8850 | if (!Recipe) { |
8851 | SmallVector<VPValue *, 4> Operands(R.operands()); |
8852 | Recipe = RecipeBuilder.handleReplication(I: Instr, Operands, Range); |
8853 | } |
8854 | |
8855 | RecipeBuilder.setRecipe(I: Instr, R: Recipe); |
8856 | if (isa<VPWidenIntOrFpInductionRecipe>(Val: Recipe) && isa<TruncInst>(Val: Instr)) { |
8857 | // Optimized a truncate to VPWidenIntOrFpInductionRecipe. It needs to be |
8858 | // moved to the phi section in the header. |
8859 | Recipe->insertBefore(BB&: *HeaderVPBB, IP: HeaderVPBB->getFirstNonPhi()); |
8860 | } else { |
8861 | Builder.insert(R: Recipe); |
8862 | } |
8863 | if (Recipe->getNumDefinedValues() == 1) { |
8864 | SingleDef->replaceAllUsesWith(New: Recipe->getVPSingleValue()); |
8865 | Old2New[SingleDef] = Recipe->getVPSingleValue(); |
8866 | } else { |
8867 | assert(Recipe->getNumDefinedValues() == 0 && |
8868 | "Unexpected multidef recipe" ); |
8869 | R.eraseFromParent(); |
8870 | } |
8871 | } |
8872 | } |
8873 | |
8874 | // replaceAllUsesWith above may invalidate the block masks. Update them here. |
8875 | // TODO: Include the masks as operands in the predicated VPlan directly |
8876 | // to remove the need to keep a map of masks beyond the predication |
8877 | // transform. |
8878 | RecipeBuilder.updateBlockMaskCache(Old2New); |
8879 | for (const auto &[Old, _] : Old2New) |
8880 | Old->getDefiningRecipe()->eraseFromParent(); |
8881 | |
8882 | assert(isa<VPRegionBlock>(Plan->getVectorLoopRegion()) && |
8883 | !Plan->getVectorLoopRegion()->getEntryBasicBlock()->empty() && |
8884 | "entry block must be set to a VPRegionBlock having a non-empty entry " |
8885 | "VPBasicBlock" ); |
8886 | |
8887 | // Update wide induction increments to use the same step as the corresponding |
8888 | // wide induction. This enables detecting induction increments directly in |
8889 | // VPlan and removes redundant splats. |
8890 | for (const auto &[Phi, ID] : Legal->getInductionVars()) { |
8891 | auto *IVInc = cast<Instruction>( |
8892 | Val: Phi->getIncomingValueForBlock(BB: OrigLoop->getLoopLatch())); |
8893 | if (IVInc->getOperand(i: 0) != Phi || IVInc->getOpcode() != Instruction::Add) |
8894 | continue; |
8895 | VPWidenInductionRecipe *WideIV = |
8896 | cast<VPWidenInductionRecipe>(Val: RecipeBuilder.getRecipe(I: Phi)); |
8897 | VPRecipeBase *R = RecipeBuilder.getRecipe(I: IVInc); |
8898 | R->setOperand(I: 1, New: WideIV->getStepValue()); |
8899 | } |
8900 | |
8901 | DenseMap<VPValue *, VPValue *> IVEndValues; |
8902 | addScalarResumePhis(Builder&: RecipeBuilder, Plan&: *Plan, IVEndValues); |
8903 | SetVector<VPIRInstruction *> ExitUsersToFix = |
8904 | collectUsersInLatchExitBlock(Plan&: *Plan); |
8905 | addExitUsersForFirstOrderRecurrences(Plan&: *Plan, ExitUsersToFix, Range); |
8906 | addUsersInExitBlocks(Plan&: *Plan, ExitUsersToFix); |
8907 | |
8908 | // --------------------------------------------------------------------------- |
8909 | // Transform initial VPlan: Apply previously taken decisions, in order, to |
8910 | // bring the VPlan to its final state. |
8911 | // --------------------------------------------------------------------------- |
8912 | |
8913 | // Adjust the recipes for any inloop reductions. |
8914 | adjustRecipesForReductions(Plan, RecipeBuilder, MinVF: Range.Start); |
8915 | |
8916 | // Transform recipes to abstract recipes if it is legal and beneficial and |
8917 | // clamp the range for better cost estimation. |
8918 | // TODO: Enable following transform when the EVL-version of extended-reduction |
8919 | // and mulacc-reduction are implemented. |
8920 | if (!CM.foldTailWithEVL()) { |
8921 | VPCostContext CostCtx(CM.TTI, *CM.TLI, Legal->getWidestInductionType(), CM, |
8922 | CM.CostKind); |
8923 | VPlanTransforms::runPass(Fn: VPlanTransforms::convertToAbstractRecipes, Plan&: *Plan, |
8924 | Args&: CostCtx, Args&: Range); |
8925 | } |
8926 | |
8927 | for (ElementCount VF : Range) |
8928 | Plan->addVF(VF); |
8929 | Plan->setName("Initial VPlan" ); |
8930 | |
8931 | // Interleave memory: for each Interleave Group we marked earlier as relevant |
8932 | // for this VPlan, replace the Recipes widening its memory instructions with a |
8933 | // single VPInterleaveRecipe at its insertion point. |
8934 | VPlanTransforms::runPass(Fn: VPlanTransforms::createInterleaveGroups, Plan&: *Plan, |
8935 | Args: InterleaveGroups, Args&: RecipeBuilder, |
8936 | Args: CM.isScalarEpilogueAllowed()); |
8937 | |
8938 | // Replace VPValues for known constant strides guaranteed by predicate scalar |
8939 | // evolution. |
8940 | auto CanUseVersionedStride = [&Plan](VPUser &U, unsigned) { |
8941 | auto *R = cast<VPRecipeBase>(Val: &U); |
8942 | return R->getParent()->getParent() || |
8943 | R->getParent() == |
8944 | Plan->getVectorLoopRegion()->getSinglePredecessor(); |
8945 | }; |
8946 | for (auto [_, Stride] : Legal->getLAI()->getSymbolicStrides()) { |
8947 | auto *StrideV = cast<SCEVUnknown>(Val: Stride)->getValue(); |
8948 | auto *ScevStride = dyn_cast<SCEVConstant>(Val: PSE.getSCEV(V: StrideV)); |
8949 | // Only handle constant strides for now. |
8950 | if (!ScevStride) |
8951 | continue; |
8952 | |
8953 | auto *CI = Plan->getOrAddLiveIn( |
8954 | V: ConstantInt::get(Ty: Stride->getType(), V: ScevStride->getAPInt())); |
8955 | if (VPValue *StrideVPV = Plan->getLiveIn(V: StrideV)) |
8956 | StrideVPV->replaceUsesWithIf(New: CI, ShouldReplace: CanUseVersionedStride); |
8957 | |
8958 | // The versioned value may not be used in the loop directly but through a |
8959 | // sext/zext. Add new live-ins in those cases. |
8960 | for (Value *U : StrideV->users()) { |
8961 | if (!isa<SExtInst, ZExtInst>(Val: U)) |
8962 | continue; |
8963 | VPValue *StrideVPV = Plan->getLiveIn(V: U); |
8964 | if (!StrideVPV) |
8965 | continue; |
8966 | unsigned BW = U->getType()->getScalarSizeInBits(); |
8967 | APInt C = isa<SExtInst>(Val: U) ? ScevStride->getAPInt().sext(width: BW) |
8968 | : ScevStride->getAPInt().zext(width: BW); |
8969 | VPValue *CI = Plan->getOrAddLiveIn(V: ConstantInt::get(Ty: U->getType(), V: C)); |
8970 | StrideVPV->replaceUsesWithIf(New: CI, ShouldReplace: CanUseVersionedStride); |
8971 | } |
8972 | } |
8973 | |
8974 | auto BlockNeedsPredication = [this](BasicBlock *BB) { |
8975 | return Legal->blockNeedsPredication(BB); |
8976 | }; |
8977 | VPlanTransforms::runPass(Fn: VPlanTransforms::dropPoisonGeneratingRecipes, Plan&: *Plan, |
8978 | Args: BlockNeedsPredication); |
8979 | |
8980 | // Sink users of fixed-order recurrence past the recipe defining the previous |
8981 | // value and introduce FirstOrderRecurrenceSplice VPInstructions. |
8982 | if (!VPlanTransforms::runPass(Transform: VPlanTransforms::adjustFixedOrderRecurrences, |
8983 | Plan&: *Plan, Args&: Builder)) |
8984 | return nullptr; |
8985 | |
8986 | if (useActiveLaneMask(Style)) { |
8987 | // TODO: Move checks to VPlanTransforms::addActiveLaneMask once |
8988 | // TailFoldingStyle is visible there. |
8989 | bool ForControlFlow = useActiveLaneMaskForControlFlow(Style); |
8990 | bool WithoutRuntimeCheck = |
8991 | Style == TailFoldingStyle::DataAndControlFlowWithoutRuntimeCheck; |
8992 | VPlanTransforms::addActiveLaneMask(Plan&: *Plan, UseActiveLaneMaskForControlFlow: ForControlFlow, |
8993 | DataAndControlFlowWithoutRuntimeCheck: WithoutRuntimeCheck); |
8994 | } |
8995 | VPlanTransforms::optimizeInductionExitUsers(Plan&: *Plan, EndValues&: IVEndValues); |
8996 | |
8997 | assert(verifyVPlanIsValid(*Plan) && "VPlan is invalid" ); |
8998 | return Plan; |
8999 | } |
9000 | |
9001 | VPlanPtr LoopVectorizationPlanner::tryToBuildVPlan(VFRange &Range) { |
9002 | // Outer loop handling: They may require CFG and instruction level |
9003 | // transformations before even evaluating whether vectorization is profitable. |
9004 | // Since we cannot modify the incoming IR, we need to build VPlan upfront in |
9005 | // the vectorization pipeline. |
9006 | assert(!OrigLoop->isInnermost()); |
9007 | assert(EnableVPlanNativePath && "VPlan-native path is not enabled." ); |
9008 | |
9009 | auto Plan = VPlanTransforms::buildPlainCFG(TheLoop: OrigLoop, LI&: *LI); |
9010 | VPlanTransforms::prepareForVectorization( |
9011 | Plan&: *Plan, InductionTy: Legal->getWidestInductionType(), PSE, RequiresScalarEpilogueCheck: true, TailFolded: false, TheLoop: OrigLoop, |
9012 | IVDL: getDebugLocFromInstOrOperands(I: Legal->getPrimaryInduction()), HasUncountableExit: false, |
9013 | Range); |
9014 | VPlanTransforms::createLoopRegions(Plan&: *Plan); |
9015 | |
9016 | for (ElementCount VF : Range) |
9017 | Plan->addVF(VF); |
9018 | |
9019 | if (!VPlanTransforms::tryToConvertVPInstructionsToVPRecipes( |
9020 | Plan, |
9021 | GetIntOrFpInductionDescriptor: [this](PHINode *P) { |
9022 | return Legal->getIntOrFpInductionDescriptor(Phi: P); |
9023 | }, |
9024 | SE&: *PSE.getSE(), TLI: *TLI)) |
9025 | return nullptr; |
9026 | |
9027 | // Collect mapping of IR header phis to header phi recipes, to be used in |
9028 | // addScalarResumePhis. |
9029 | DenseMap<VPBasicBlock *, VPValue *> BlockMaskCache; |
9030 | VPRecipeBuilder RecipeBuilder(*Plan, OrigLoop, TLI, &TTI, Legal, CM, PSE, |
9031 | Builder, BlockMaskCache, nullptr /*LVer*/); |
9032 | for (auto &R : Plan->getVectorLoopRegion()->getEntryBasicBlock()->phis()) { |
9033 | if (isa<VPCanonicalIVPHIRecipe>(Val: &R)) |
9034 | continue; |
9035 | auto * = cast<VPHeaderPHIRecipe>(Val: &R); |
9036 | RecipeBuilder.setRecipe(I: HeaderR->getUnderlyingInstr(), R: HeaderR); |
9037 | } |
9038 | DenseMap<VPValue *, VPValue *> IVEndValues; |
9039 | // TODO: IVEndValues are not used yet in the native path, to optimize exit |
9040 | // values. |
9041 | addScalarResumePhis(Builder&: RecipeBuilder, Plan&: *Plan, IVEndValues); |
9042 | |
9043 | assert(verifyVPlanIsValid(*Plan) && "VPlan is invalid" ); |
9044 | return Plan; |
9045 | } |
9046 | |
9047 | // Adjust the recipes for reductions. For in-loop reductions the chain of |
9048 | // instructions leading from the loop exit instr to the phi need to be converted |
9049 | // to reductions, with one operand being vector and the other being the scalar |
9050 | // reduction chain. For other reductions, a select is introduced between the phi |
9051 | // and users outside the vector region when folding the tail. |
9052 | // |
9053 | // A ComputeReductionResult recipe is added to the middle block, also for |
9054 | // in-loop reductions which compute their result in-loop, because generating |
9055 | // the subsequent bc.merge.rdx phi is driven by ComputeReductionResult recipes. |
9056 | // |
9057 | // Adjust AnyOf reductions; replace the reduction phi for the selected value |
9058 | // with a boolean reduction phi node to check if the condition is true in any |
9059 | // iteration. The final value is selected by the final ComputeReductionResult. |
9060 | void LoopVectorizationPlanner::adjustRecipesForReductions( |
9061 | VPlanPtr &Plan, VPRecipeBuilder &RecipeBuilder, ElementCount MinVF) { |
9062 | using namespace VPlanPatternMatch; |
9063 | VPRegionBlock *VectorLoopRegion = Plan->getVectorLoopRegion(); |
9064 | VPBasicBlock * = VectorLoopRegion->getEntryBasicBlock(); |
9065 | VPBasicBlock *MiddleVPBB = Plan->getMiddleBlock(); |
9066 | SmallVector<VPRecipeBase *> ToDelete; |
9067 | |
9068 | for (VPRecipeBase &R : Header->phis()) { |
9069 | auto *PhiR = dyn_cast<VPReductionPHIRecipe>(Val: &R); |
9070 | if (!PhiR || !PhiR->isInLoop() || (MinVF.isScalar() && !PhiR->isOrdered())) |
9071 | continue; |
9072 | |
9073 | const RecurrenceDescriptor &RdxDesc = PhiR->getRecurrenceDescriptor(); |
9074 | RecurKind Kind = RdxDesc.getRecurrenceKind(); |
9075 | assert( |
9076 | !RecurrenceDescriptor::isAnyOfRecurrenceKind(Kind) && |
9077 | !RecurrenceDescriptor::isFindIVRecurrenceKind(Kind) && |
9078 | "AnyOf and FindIV reductions are not allowed for in-loop reductions" ); |
9079 | |
9080 | // Collect the chain of "link" recipes for the reduction starting at PhiR. |
9081 | SetVector<VPSingleDefRecipe *> Worklist; |
9082 | Worklist.insert(X: PhiR); |
9083 | for (unsigned I = 0; I != Worklist.size(); ++I) { |
9084 | VPSingleDefRecipe *Cur = Worklist[I]; |
9085 | for (VPUser *U : Cur->users()) { |
9086 | auto *UserRecipe = cast<VPSingleDefRecipe>(Val: U); |
9087 | if (!UserRecipe->getParent()->getEnclosingLoopRegion()) { |
9088 | assert((UserRecipe->getParent() == MiddleVPBB || |
9089 | UserRecipe->getParent() == Plan->getScalarPreheader()) && |
9090 | "U must be either in the loop region, the middle block or the " |
9091 | "scalar preheader." ); |
9092 | continue; |
9093 | } |
9094 | Worklist.insert(X: UserRecipe); |
9095 | } |
9096 | } |
9097 | |
9098 | // Visit operation "Links" along the reduction chain top-down starting from |
9099 | // the phi until LoopExitValue. We keep track of the previous item |
9100 | // (PreviousLink) to tell which of the two operands of a Link will remain |
9101 | // scalar and which will be reduced. For minmax by select(cmp), Link will be |
9102 | // the select instructions. Blend recipes of in-loop reduction phi's will |
9103 | // get folded to their non-phi operand, as the reduction recipe handles the |
9104 | // condition directly. |
9105 | VPSingleDefRecipe *PreviousLink = PhiR; // Aka Worklist[0]. |
9106 | for (VPSingleDefRecipe *CurrentLink : Worklist.getArrayRef().drop_front()) { |
9107 | if (auto *Blend = dyn_cast<VPBlendRecipe>(Val: CurrentLink)) { |
9108 | assert(Blend->getNumIncomingValues() == 2 && |
9109 | "Blend must have 2 incoming values" ); |
9110 | if (Blend->getIncomingValue(Idx: 0) == PhiR) { |
9111 | Blend->replaceAllUsesWith(New: Blend->getIncomingValue(Idx: 1)); |
9112 | } else { |
9113 | assert(Blend->getIncomingValue(1) == PhiR && |
9114 | "PhiR must be an operand of the blend" ); |
9115 | Blend->replaceAllUsesWith(New: Blend->getIncomingValue(Idx: 0)); |
9116 | } |
9117 | continue; |
9118 | } |
9119 | |
9120 | Instruction *CurrentLinkI = CurrentLink->getUnderlyingInstr(); |
9121 | |
9122 | // Index of the first operand which holds a non-mask vector operand. |
9123 | unsigned IndexOfFirstOperand; |
9124 | // Recognize a call to the llvm.fmuladd intrinsic. |
9125 | bool IsFMulAdd = (Kind == RecurKind::FMulAdd); |
9126 | VPValue *VecOp; |
9127 | VPBasicBlock *LinkVPBB = CurrentLink->getParent(); |
9128 | if (IsFMulAdd) { |
9129 | assert( |
9130 | RecurrenceDescriptor::isFMulAddIntrinsic(CurrentLinkI) && |
9131 | "Expected instruction to be a call to the llvm.fmuladd intrinsic" ); |
9132 | assert(((MinVF.isScalar() && isa<VPReplicateRecipe>(CurrentLink)) || |
9133 | isa<VPWidenIntrinsicRecipe>(CurrentLink)) && |
9134 | CurrentLink->getOperand(2) == PreviousLink && |
9135 | "expected a call where the previous link is the added operand" ); |
9136 | |
9137 | // If the instruction is a call to the llvm.fmuladd intrinsic then we |
9138 | // need to create an fmul recipe (multiplying the first two operands of |
9139 | // the fmuladd together) to use as the vector operand for the fadd |
9140 | // reduction. |
9141 | VPInstruction *FMulRecipe = new VPInstruction( |
9142 | Instruction::FMul, |
9143 | {CurrentLink->getOperand(N: 0), CurrentLink->getOperand(N: 1)}, |
9144 | CurrentLinkI->getFastMathFlags()); |
9145 | LinkVPBB->insert(Recipe: FMulRecipe, InsertPt: CurrentLink->getIterator()); |
9146 | VecOp = FMulRecipe; |
9147 | } else { |
9148 | if (RecurrenceDescriptor::isMinMaxRecurrenceKind(Kind)) { |
9149 | if (isa<VPWidenRecipe>(Val: CurrentLink)) { |
9150 | assert(isa<CmpInst>(CurrentLinkI) && |
9151 | "need to have the compare of the select" ); |
9152 | continue; |
9153 | } |
9154 | assert(isa<VPWidenSelectRecipe>(CurrentLink) && |
9155 | "must be a select recipe" ); |
9156 | IndexOfFirstOperand = 1; |
9157 | } else { |
9158 | assert((MinVF.isScalar() || isa<VPWidenRecipe>(CurrentLink)) && |
9159 | "Expected to replace a VPWidenSC" ); |
9160 | IndexOfFirstOperand = 0; |
9161 | } |
9162 | // Note that for non-commutable operands (cmp-selects), the semantics of |
9163 | // the cmp-select are captured in the recurrence kind. |
9164 | unsigned VecOpId = |
9165 | CurrentLink->getOperand(N: IndexOfFirstOperand) == PreviousLink |
9166 | ? IndexOfFirstOperand + 1 |
9167 | : IndexOfFirstOperand; |
9168 | VecOp = CurrentLink->getOperand(N: VecOpId); |
9169 | assert(VecOp != PreviousLink && |
9170 | CurrentLink->getOperand(CurrentLink->getNumOperands() - 1 - |
9171 | (VecOpId - IndexOfFirstOperand)) == |
9172 | PreviousLink && |
9173 | "PreviousLink must be the operand other than VecOp" ); |
9174 | } |
9175 | |
9176 | VPValue *CondOp = nullptr; |
9177 | if (CM.blockNeedsPredicationForAnyReason(BB: CurrentLinkI->getParent())) |
9178 | CondOp = RecipeBuilder.getBlockInMask(VPBB: CurrentLink->getParent()); |
9179 | |
9180 | // Non-FP RdxDescs will have all fast math flags set, so clear them. |
9181 | FastMathFlags FMFs = isa<FPMathOperator>(Val: CurrentLinkI) |
9182 | ? RdxDesc.getFastMathFlags() |
9183 | : FastMathFlags(); |
9184 | auto *RedRecipe = new VPReductionRecipe( |
9185 | Kind, FMFs, CurrentLinkI, PreviousLink, VecOp, CondOp, |
9186 | CM.useOrderedReductions(RdxDesc), CurrentLinkI->getDebugLoc()); |
9187 | // Append the recipe to the end of the VPBasicBlock because we need to |
9188 | // ensure that it comes after all of it's inputs, including CondOp. |
9189 | // Delete CurrentLink as it will be invalid if its operand is replaced |
9190 | // with a reduction defined at the bottom of the block in the next link. |
9191 | if (LinkVPBB->getNumSuccessors() == 0) |
9192 | RedRecipe->insertBefore(InsertPos: &*std::prev(x: std::prev(x: LinkVPBB->end()))); |
9193 | else |
9194 | LinkVPBB->appendRecipe(Recipe: RedRecipe); |
9195 | |
9196 | CurrentLink->replaceAllUsesWith(New: RedRecipe); |
9197 | ToDelete.push_back(Elt: CurrentLink); |
9198 | PreviousLink = RedRecipe; |
9199 | } |
9200 | } |
9201 | VPBasicBlock *LatchVPBB = VectorLoopRegion->getExitingBasicBlock(); |
9202 | Builder.setInsertPoint(&*std::prev(x: std::prev(x: LatchVPBB->end()))); |
9203 | VPBasicBlock::iterator IP = MiddleVPBB->getFirstNonPhi(); |
9204 | for (VPRecipeBase &R : |
9205 | Plan->getVectorLoopRegion()->getEntryBasicBlock()->phis()) { |
9206 | VPReductionPHIRecipe *PhiR = dyn_cast<VPReductionPHIRecipe>(Val: &R); |
9207 | if (!PhiR) |
9208 | continue; |
9209 | |
9210 | const RecurrenceDescriptor &RdxDesc = PhiR->getRecurrenceDescriptor(); |
9211 | Type *PhiTy = PhiR->getUnderlyingValue()->getType(); |
9212 | // If tail is folded by masking, introduce selects between the phi |
9213 | // and the users outside the vector region of each reduction, at the |
9214 | // beginning of the dedicated latch block. |
9215 | auto *OrigExitingVPV = PhiR->getBackedgeValue(); |
9216 | auto *NewExitingVPV = PhiR->getBackedgeValue(); |
9217 | // Don't output selects for partial reductions because they have an output |
9218 | // with fewer lanes than the VF. So the operands of the select would have |
9219 | // different numbers of lanes. Partial reductions mask the input instead. |
9220 | if (!PhiR->isInLoop() && CM.foldTailByMasking() && |
9221 | !isa<VPPartialReductionRecipe>(Val: OrigExitingVPV->getDefiningRecipe())) { |
9222 | VPValue *Cond = RecipeBuilder.getBlockInMask(VPBB: PhiR->getParent()); |
9223 | std::optional<FastMathFlags> FMFs = |
9224 | PhiTy->isFloatingPointTy() |
9225 | ? std::make_optional(t: RdxDesc.getFastMathFlags()) |
9226 | : std::nullopt; |
9227 | NewExitingVPV = |
9228 | Builder.createSelect(Cond, TrueVal: OrigExitingVPV, FalseVal: PhiR, DL: {}, Name: "" , FMFs); |
9229 | OrigExitingVPV->replaceUsesWithIf(New: NewExitingVPV, ShouldReplace: [](VPUser &U, unsigned) { |
9230 | return isa<VPInstruction>(Val: &U) && |
9231 | (cast<VPInstruction>(Val: &U)->getOpcode() == |
9232 | VPInstruction::ComputeAnyOfResult || |
9233 | cast<VPInstruction>(Val: &U)->getOpcode() == |
9234 | VPInstruction::ComputeReductionResult || |
9235 | cast<VPInstruction>(Val: &U)->getOpcode() == |
9236 | VPInstruction::ComputeFindIVResult); |
9237 | }); |
9238 | if (CM.usePredicatedReductionSelect()) |
9239 | PhiR->setOperand(I: 1, New: NewExitingVPV); |
9240 | } |
9241 | |
9242 | // We want code in the middle block to appear to execute on the location of |
9243 | // the scalar loop's latch terminator because: (a) it is all compiler |
9244 | // generated, (b) these instructions are always executed after evaluating |
9245 | // the latch conditional branch, and (c) other passes may add new |
9246 | // predecessors which terminate on this line. This is the easiest way to |
9247 | // ensure we don't accidentally cause an extra step back into the loop while |
9248 | // debugging. |
9249 | DebugLoc ExitDL = OrigLoop->getLoopLatch()->getTerminator()->getDebugLoc(); |
9250 | |
9251 | // TODO: At the moment ComputeReductionResult also drives creation of the |
9252 | // bc.merge.rdx phi nodes, hence it needs to be created unconditionally here |
9253 | // even for in-loop reductions, until the reduction resume value handling is |
9254 | // also modeled in VPlan. |
9255 | VPInstruction *FinalReductionResult; |
9256 | VPBuilder::InsertPointGuard Guard(Builder); |
9257 | Builder.setInsertPoint(TheBB: MiddleVPBB, IP); |
9258 | if (RecurrenceDescriptor::isFindIVRecurrenceKind( |
9259 | Kind: RdxDesc.getRecurrenceKind())) { |
9260 | VPValue *Start = PhiR->getStartValue(); |
9261 | VPValue *Sentinel = Plan->getOrAddLiveIn(V: RdxDesc.getSentinelValue()); |
9262 | FinalReductionResult = |
9263 | Builder.createNaryOp(Opcode: VPInstruction::ComputeFindIVResult, |
9264 | Operands: {PhiR, Start, Sentinel, NewExitingVPV}, DL: ExitDL); |
9265 | } else if (RecurrenceDescriptor::isAnyOfRecurrenceKind( |
9266 | Kind: RdxDesc.getRecurrenceKind())) { |
9267 | VPValue *Start = PhiR->getStartValue(); |
9268 | FinalReductionResult = |
9269 | Builder.createNaryOp(Opcode: VPInstruction::ComputeAnyOfResult, |
9270 | Operands: {PhiR, Start, NewExitingVPV}, DL: ExitDL); |
9271 | } else { |
9272 | VPIRFlags Flags = RecurrenceDescriptor::isFloatingPointRecurrenceKind( |
9273 | Kind: RdxDesc.getRecurrenceKind()) |
9274 | ? VPIRFlags(RdxDesc.getFastMathFlags()) |
9275 | : VPIRFlags(); |
9276 | FinalReductionResult = |
9277 | Builder.createNaryOp(Opcode: VPInstruction::ComputeReductionResult, |
9278 | Operands: {PhiR, NewExitingVPV}, Flags, DL: ExitDL); |
9279 | } |
9280 | // If the vector reduction can be performed in a smaller type, we truncate |
9281 | // then extend the loop exit value to enable InstCombine to evaluate the |
9282 | // entire expression in the smaller type. |
9283 | if (MinVF.isVector() && PhiTy != RdxDesc.getRecurrenceType() && |
9284 | !RecurrenceDescriptor::isAnyOfRecurrenceKind( |
9285 | Kind: RdxDesc.getRecurrenceKind())) { |
9286 | assert(!PhiR->isInLoop() && "Unexpected truncated inloop reduction!" ); |
9287 | assert(!RecurrenceDescriptor::isMinMaxRecurrenceKind( |
9288 | RdxDesc.getRecurrenceKind()) && |
9289 | "Unexpected truncated min-max recurrence!" ); |
9290 | Type *RdxTy = RdxDesc.getRecurrenceType(); |
9291 | auto *Trunc = |
9292 | new VPWidenCastRecipe(Instruction::Trunc, NewExitingVPV, RdxTy); |
9293 | Instruction::CastOps ExtendOpc = |
9294 | RdxDesc.isSigned() ? Instruction::SExt : Instruction::ZExt; |
9295 | auto *Extnd = new VPWidenCastRecipe(ExtendOpc, Trunc, PhiTy); |
9296 | Trunc->insertAfter(InsertPos: NewExitingVPV->getDefiningRecipe()); |
9297 | Extnd->insertAfter(InsertPos: Trunc); |
9298 | if (PhiR->getOperand(N: 1) == NewExitingVPV) |
9299 | PhiR->setOperand(I: 1, New: Extnd->getVPSingleValue()); |
9300 | |
9301 | // Update ComputeReductionResult with the truncated exiting value and |
9302 | // extend its result. |
9303 | FinalReductionResult->setOperand(I: 1, New: Trunc); |
9304 | FinalReductionResult = |
9305 | Builder.createScalarCast(Opcode: ExtendOpc, Op: FinalReductionResult, ResultTy: PhiTy, DL: {}); |
9306 | } |
9307 | |
9308 | // Update all users outside the vector region. Also replace redundant |
9309 | // ExtractLastElement. |
9310 | for (auto *U : to_vector(Range: OrigExitingVPV->users())) { |
9311 | auto *Parent = cast<VPRecipeBase>(Val: U)->getParent(); |
9312 | if (FinalReductionResult == U || Parent->getParent()) |
9313 | continue; |
9314 | U->replaceUsesOfWith(From: OrigExitingVPV, To: FinalReductionResult); |
9315 | if (match(U, P: m_VPInstruction<VPInstruction::ExtractLastElement>( |
9316 | Op0: m_VPValue()))) |
9317 | cast<VPInstruction>(Val: U)->replaceAllUsesWith(New: FinalReductionResult); |
9318 | } |
9319 | |
9320 | // Adjust AnyOf reductions; replace the reduction phi for the selected value |
9321 | // with a boolean reduction phi node to check if the condition is true in |
9322 | // any iteration. The final value is selected by the final |
9323 | // ComputeReductionResult. |
9324 | if (RecurrenceDescriptor::isAnyOfRecurrenceKind( |
9325 | Kind: RdxDesc.getRecurrenceKind())) { |
9326 | auto *Select = cast<VPRecipeBase>(Val: *find_if(Range: PhiR->users(), P: [](VPUser *U) { |
9327 | return isa<VPWidenSelectRecipe>(Val: U) || |
9328 | (isa<VPReplicateRecipe>(Val: U) && |
9329 | cast<VPReplicateRecipe>(Val: U)->getUnderlyingInstr()->getOpcode() == |
9330 | Instruction::Select); |
9331 | })); |
9332 | VPValue *Cmp = Select->getOperand(N: 0); |
9333 | // If the compare is checking the reduction PHI node, adjust it to check |
9334 | // the start value. |
9335 | if (VPRecipeBase *CmpR = Cmp->getDefiningRecipe()) |
9336 | CmpR->replaceUsesOfWith(From: PhiR, To: PhiR->getStartValue()); |
9337 | Builder.setInsertPoint(Select); |
9338 | |
9339 | // If the true value of the select is the reduction phi, the new value is |
9340 | // selected if the negated condition is true in any iteration. |
9341 | if (Select->getOperand(N: 1) == PhiR) |
9342 | Cmp = Builder.createNot(Operand: Cmp); |
9343 | VPValue *Or = Builder.createOr(LHS: PhiR, RHS: Cmp); |
9344 | Select->getVPSingleValue()->replaceAllUsesWith(New: Or); |
9345 | // Delete Select now that it has invalid types. |
9346 | ToDelete.push_back(Elt: Select); |
9347 | |
9348 | // Convert the reduction phi to operate on bools. |
9349 | PhiR->setOperand(I: 0, New: Plan->getOrAddLiveIn(V: ConstantInt::getFalse( |
9350 | Context&: OrigLoop->getHeader()->getContext()))); |
9351 | continue; |
9352 | } |
9353 | |
9354 | if (RecurrenceDescriptor::isFindIVRecurrenceKind( |
9355 | Kind: RdxDesc.getRecurrenceKind())) { |
9356 | // Adjust the start value for FindFirstIV/FindLastIV recurrences to use |
9357 | // the sentinel value after generating the ResumePhi recipe, which uses |
9358 | // the original start value. |
9359 | PhiR->setOperand(I: 0, New: Plan->getOrAddLiveIn(V: RdxDesc.getSentinelValue())); |
9360 | } |
9361 | RecurKind RK = RdxDesc.getRecurrenceKind(); |
9362 | if ((!RecurrenceDescriptor::isAnyOfRecurrenceKind(Kind: RK) && |
9363 | !RecurrenceDescriptor::isFindIVRecurrenceKind(Kind: RK) && |
9364 | !RecurrenceDescriptor::isMinMaxRecurrenceKind(Kind: RK))) { |
9365 | VPBuilder PHBuilder(Plan->getVectorPreheader()); |
9366 | VPValue *Iden = Plan->getOrAddLiveIn( |
9367 | V: getRecurrenceIdentity(K: RK, Tp: PhiTy, FMF: RdxDesc.getFastMathFlags())); |
9368 | // If the PHI is used by a partial reduction, set the scale factor. |
9369 | unsigned ScaleFactor = |
9370 | RecipeBuilder.getScalingForReduction(ExitInst: RdxDesc.getLoopExitInstr()) |
9371 | .value_or(u: 1); |
9372 | Type *I32Ty = IntegerType::getInt32Ty(C&: PhiTy->getContext()); |
9373 | auto *ScaleFactorVPV = |
9374 | Plan->getOrAddLiveIn(V: ConstantInt::get(Ty: I32Ty, V: ScaleFactor)); |
9375 | VPValue *StartV = PHBuilder.createNaryOp( |
9376 | Opcode: VPInstruction::ReductionStartVector, |
9377 | Operands: {PhiR->getStartValue(), Iden, ScaleFactorVPV}, |
9378 | Flags: PhiTy->isFloatingPointTy() ? RdxDesc.getFastMathFlags() |
9379 | : FastMathFlags()); |
9380 | PhiR->setOperand(I: 0, New: StartV); |
9381 | } |
9382 | } |
9383 | for (VPRecipeBase *R : ToDelete) |
9384 | R->eraseFromParent(); |
9385 | |
9386 | VPlanTransforms::runPass(Fn: VPlanTransforms::clearReductionWrapFlags, Plan&: *Plan); |
9387 | } |
9388 | |
9389 | void VPDerivedIVRecipe::execute(VPTransformState &State) { |
9390 | assert(!State.Lane && "VPDerivedIVRecipe being replicated." ); |
9391 | |
9392 | // Fast-math-flags propagate from the original induction instruction. |
9393 | IRBuilder<>::FastMathFlagGuard FMFG(State.Builder); |
9394 | if (FPBinOp) |
9395 | State.Builder.setFastMathFlags(FPBinOp->getFastMathFlags()); |
9396 | |
9397 | Value *Step = State.get(Def: getStepValue(), Lane: VPLane(0)); |
9398 | Value *Index = State.get(Def: getOperand(N: 1), Lane: VPLane(0)); |
9399 | Value *DerivedIV = emitTransformedIndex( |
9400 | B&: State.Builder, Index, StartValue: getStartValue()->getLiveInIRValue(), Step, InductionKind: Kind, |
9401 | InductionBinOp: cast_if_present<BinaryOperator>(Val: FPBinOp)); |
9402 | DerivedIV->setName(Name); |
9403 | // If index is the vector trip count, the concrete value will only be set in |
9404 | // prepareToExecute, leading to missed simplifications, e.g. if it is 0. |
9405 | // TODO: Remove the special case for the vector trip count once it is computed |
9406 | // in VPlan and can be used during VPlan simplification. |
9407 | assert((DerivedIV != Index || |
9408 | getOperand(1) == &getParent()->getPlan()->getVectorTripCount()) && |
9409 | "IV didn't need transforming?" ); |
9410 | State.set(Def: this, V: DerivedIV, Lane: VPLane(0)); |
9411 | } |
9412 | |
9413 | // Determine how to lower the scalar epilogue, which depends on 1) optimising |
9414 | // for minimum code-size, 2) predicate compiler options, 3) loop hints forcing |
9415 | // predication, and 4) a TTI hook that analyses whether the loop is suitable |
9416 | // for predication. |
9417 | static ScalarEpilogueLowering getScalarEpilogueLowering( |
9418 | Function *F, Loop *L, LoopVectorizeHints &Hints, ProfileSummaryInfo *PSI, |
9419 | BlockFrequencyInfo *BFI, TargetTransformInfo *TTI, TargetLibraryInfo *TLI, |
9420 | LoopVectorizationLegality &LVL, InterleavedAccessInfo *IAI) { |
9421 | // 1) OptSize takes precedence over all other options, i.e. if this is set, |
9422 | // don't look at hints or options, and don't request a scalar epilogue. |
9423 | // (For PGSO, as shouldOptimizeForSize isn't currently accessible from |
9424 | // LoopAccessInfo (due to code dependency and not being able to reliably get |
9425 | // PSI/BFI from a loop analysis under NPM), we cannot suppress the collection |
9426 | // of strides in LoopAccessInfo::analyzeLoop() and vectorize without |
9427 | // versioning when the vectorization is forced, unlike hasOptSize. So revert |
9428 | // back to the old way and vectorize with versioning when forced. See D81345.) |
9429 | if (F->hasOptSize() || (llvm::shouldOptimizeForSize(BB: L->getHeader(), PSI, BFI, |
9430 | QueryType: PGSOQueryType::IRPass) && |
9431 | Hints.getForce() != LoopVectorizeHints::FK_Enabled)) |
9432 | return CM_ScalarEpilogueNotAllowedOptSize; |
9433 | |
9434 | // 2) If set, obey the directives |
9435 | if (PreferPredicateOverEpilogue.getNumOccurrences()) { |
9436 | switch (PreferPredicateOverEpilogue) { |
9437 | case PreferPredicateTy::ScalarEpilogue: |
9438 | return CM_ScalarEpilogueAllowed; |
9439 | case PreferPredicateTy::PredicateElseScalarEpilogue: |
9440 | return CM_ScalarEpilogueNotNeededUsePredicate; |
9441 | case PreferPredicateTy::PredicateOrDontVectorize: |
9442 | return CM_ScalarEpilogueNotAllowedUsePredicate; |
9443 | }; |
9444 | } |
9445 | |
9446 | // 3) If set, obey the hints |
9447 | switch (Hints.getPredicate()) { |
9448 | case LoopVectorizeHints::FK_Enabled: |
9449 | return CM_ScalarEpilogueNotNeededUsePredicate; |
9450 | case LoopVectorizeHints::FK_Disabled: |
9451 | return CM_ScalarEpilogueAllowed; |
9452 | }; |
9453 | |
9454 | // 4) if the TTI hook indicates this is profitable, request predication. |
9455 | TailFoldingInfo TFI(TLI, &LVL, IAI); |
9456 | if (TTI->preferPredicateOverEpilogue(TFI: &TFI)) |
9457 | return CM_ScalarEpilogueNotNeededUsePredicate; |
9458 | |
9459 | return CM_ScalarEpilogueAllowed; |
9460 | } |
9461 | |
9462 | // Process the loop in the VPlan-native vectorization path. This path builds |
9463 | // VPlan upfront in the vectorization pipeline, which allows to apply |
9464 | // VPlan-to-VPlan transformations from the very beginning without modifying the |
9465 | // input LLVM IR. |
9466 | static bool processLoopInVPlanNativePath( |
9467 | Loop *L, PredicatedScalarEvolution &PSE, LoopInfo *LI, DominatorTree *DT, |
9468 | LoopVectorizationLegality *LVL, TargetTransformInfo *TTI, |
9469 | TargetLibraryInfo *TLI, DemandedBits *DB, AssumptionCache *AC, |
9470 | OptimizationRemarkEmitter *ORE, BlockFrequencyInfo *BFI, |
9471 | ProfileSummaryInfo *PSI, LoopVectorizeHints &Hints, |
9472 | LoopVectorizationRequirements &Requirements) { |
9473 | |
9474 | if (isa<SCEVCouldNotCompute>(Val: PSE.getBackedgeTakenCount())) { |
9475 | LLVM_DEBUG(dbgs() << "LV: cannot compute the outer-loop trip count\n" ); |
9476 | return false; |
9477 | } |
9478 | assert(EnableVPlanNativePath && "VPlan-native path is disabled." ); |
9479 | Function *F = L->getHeader()->getParent(); |
9480 | InterleavedAccessInfo IAI(PSE, L, DT, LI, LVL->getLAI()); |
9481 | |
9482 | ScalarEpilogueLowering SEL = |
9483 | getScalarEpilogueLowering(F, L, Hints, PSI, BFI, TTI, TLI, LVL&: *LVL, IAI: &IAI); |
9484 | |
9485 | LoopVectorizationCostModel CM(SEL, L, PSE, LI, LVL, *TTI, TLI, DB, AC, ORE, F, |
9486 | &Hints, IAI, PSI, BFI); |
9487 | // Use the planner for outer loop vectorization. |
9488 | // TODO: CM is not used at this point inside the planner. Turn CM into an |
9489 | // optional argument if we don't need it in the future. |
9490 | LoopVectorizationPlanner LVP(L, LI, DT, TLI, *TTI, LVL, CM, IAI, PSE, Hints, |
9491 | ORE); |
9492 | |
9493 | // Get user vectorization factor. |
9494 | ElementCount UserVF = Hints.getWidth(); |
9495 | |
9496 | CM.collectElementTypesForWidening(); |
9497 | |
9498 | // Plan how to best vectorize, return the best VF and its cost. |
9499 | const VectorizationFactor VF = LVP.planInVPlanNativePath(UserVF); |
9500 | |
9501 | // If we are stress testing VPlan builds, do not attempt to generate vector |
9502 | // code. Masked vector code generation support will follow soon. |
9503 | // Also, do not attempt to vectorize if no vector code will be produced. |
9504 | if (VPlanBuildStressTest || VectorizationFactor::Disabled() == VF) |
9505 | return false; |
9506 | |
9507 | VPlan &BestPlan = LVP.getPlanFor(VF: VF.Width); |
9508 | |
9509 | { |
9510 | bool AddBranchWeights = |
9511 | hasBranchWeightMD(I: *L->getLoopLatch()->getTerminator()); |
9512 | GeneratedRTChecks Checks(PSE, DT, LI, TTI, F->getDataLayout(), |
9513 | AddBranchWeights, CM.CostKind); |
9514 | InnerLoopVectorizer LB(L, PSE, LI, DT, TLI, TTI, AC, ORE, VF.Width, |
9515 | VF.Width, 1, &CM, BFI, PSI, Checks, BestPlan); |
9516 | LLVM_DEBUG(dbgs() << "Vectorizing outer loop in \"" |
9517 | << L->getHeader()->getParent()->getName() << "\"\n" ); |
9518 | LVP.executePlan(BestVF: VF.Width, BestUF: 1, BestVPlan&: BestPlan, ILV&: LB, DT, VectorizingEpilogue: false); |
9519 | } |
9520 | |
9521 | reportVectorization(ORE, TheLoop: L, VF, IC: 1); |
9522 | |
9523 | // Mark the loop as already vectorized to avoid vectorizing again. |
9524 | Hints.setAlreadyVectorized(); |
9525 | assert(!verifyFunction(*L->getHeader()->getParent(), &dbgs())); |
9526 | return true; |
9527 | } |
9528 | |
9529 | // Emit a remark if there are stores to floats that required a floating point |
9530 | // extension. If the vectorized loop was generated with floating point there |
9531 | // will be a performance penalty from the conversion overhead and the change in |
9532 | // the vector width. |
9533 | static void (Loop *L, OptimizationRemarkEmitter *ORE) { |
9534 | SmallVector<Instruction *, 4> Worklist; |
9535 | for (BasicBlock *BB : L->getBlocks()) { |
9536 | for (Instruction &Inst : *BB) { |
9537 | if (auto *S = dyn_cast<StoreInst>(Val: &Inst)) { |
9538 | if (S->getValueOperand()->getType()->isFloatTy()) |
9539 | Worklist.push_back(Elt: S); |
9540 | } |
9541 | } |
9542 | } |
9543 | |
9544 | // Traverse the floating point stores upwards searching, for floating point |
9545 | // conversions. |
9546 | SmallPtrSet<const Instruction *, 4> Visited; |
9547 | SmallPtrSet<const Instruction *, 4> ; |
9548 | while (!Worklist.empty()) { |
9549 | auto *I = Worklist.pop_back_val(); |
9550 | if (!L->contains(Inst: I)) |
9551 | continue; |
9552 | if (!Visited.insert(Ptr: I).second) |
9553 | continue; |
9554 | |
9555 | // Emit a remark if the floating point store required a floating |
9556 | // point conversion. |
9557 | // TODO: More work could be done to identify the root cause such as a |
9558 | // constant or a function return type and point the user to it. |
9559 | if (isa<FPExtInst>(Val: I) && EmittedRemark.insert(Ptr: I).second) |
9560 | ORE->emit(RemarkBuilder: [&]() { |
9561 | return OptimizationRemarkAnalysis(LV_NAME, "VectorMixedPrecision" , |
9562 | I->getDebugLoc(), L->getHeader()) |
9563 | << "floating point conversion changes vector width. " |
9564 | << "Mixed floating point precision requires an up/down " |
9565 | << "cast that will negatively impact performance." ; |
9566 | }); |
9567 | |
9568 | for (Use &Op : I->operands()) |
9569 | if (auto *OpI = dyn_cast<Instruction>(Val&: Op)) |
9570 | Worklist.push_back(Elt: OpI); |
9571 | } |
9572 | } |
9573 | |
9574 | /// For loops with uncountable early exits, find the cost of doing work when |
9575 | /// exiting the loop early, such as calculating the final exit values of |
9576 | /// variables used outside the loop. |
9577 | /// TODO: This is currently overly pessimistic because the loop may not take |
9578 | /// the early exit, but better to keep this conservative for now. In future, |
9579 | /// it might be possible to relax this by using branch probabilities. |
9580 | static InstructionCost calculateEarlyExitCost(VPCostContext &CostCtx, |
9581 | VPlan &Plan, ElementCount VF) { |
9582 | InstructionCost Cost = 0; |
9583 | for (auto *ExitVPBB : Plan.getExitBlocks()) { |
9584 | for (auto *PredVPBB : ExitVPBB->getPredecessors()) { |
9585 | // If the predecessor is not the middle.block, then it must be the |
9586 | // vector.early.exit block, which may contain work to calculate the exit |
9587 | // values of variables used outside the loop. |
9588 | if (PredVPBB != Plan.getMiddleBlock()) { |
9589 | LLVM_DEBUG(dbgs() << "Calculating cost of work in exit block " |
9590 | << PredVPBB->getName() << ":\n" ); |
9591 | Cost += PredVPBB->cost(VF, Ctx&: CostCtx); |
9592 | } |
9593 | } |
9594 | } |
9595 | return Cost; |
9596 | } |
9597 | |
9598 | /// This function determines whether or not it's still profitable to vectorize |
9599 | /// the loop given the extra work we have to do outside of the loop: |
9600 | /// 1. Perform the runtime checks before entering the loop to ensure it's safe |
9601 | /// to vectorize. |
9602 | /// 2. In the case of loops with uncountable early exits, we may have to do |
9603 | /// extra work when exiting the loop early, such as calculating the final |
9604 | /// exit values of variables used outside the loop. |
9605 | static bool isOutsideLoopWorkProfitable(GeneratedRTChecks &Checks, |
9606 | VectorizationFactor &VF, Loop *L, |
9607 | PredicatedScalarEvolution &PSE, |
9608 | VPCostContext &CostCtx, VPlan &Plan, |
9609 | ScalarEpilogueLowering SEL, |
9610 | std::optional<unsigned> VScale) { |
9611 | InstructionCost TotalCost = Checks.getCost(); |
9612 | if (!TotalCost.isValid()) |
9613 | return false; |
9614 | |
9615 | // Add on the cost of any work required in the vector early exit block, if |
9616 | // one exists. |
9617 | TotalCost += calculateEarlyExitCost(CostCtx, Plan, VF: VF.Width); |
9618 | |
9619 | // When interleaving only scalar and vector cost will be equal, which in turn |
9620 | // would lead to a divide by 0. Fall back to hard threshold. |
9621 | if (VF.Width.isScalar()) { |
9622 | // TODO: Should we rename VectorizeMemoryCheckThreshold? |
9623 | if (TotalCost > VectorizeMemoryCheckThreshold) { |
9624 | LLVM_DEBUG( |
9625 | dbgs() |
9626 | << "LV: Interleaving only is not profitable due to runtime checks\n" ); |
9627 | return false; |
9628 | } |
9629 | return true; |
9630 | } |
9631 | |
9632 | // The scalar cost should only be 0 when vectorizing with a user specified |
9633 | // VF/IC. In those cases, runtime checks should always be generated. |
9634 | uint64_t ScalarC = VF.ScalarCost.getValue(); |
9635 | if (ScalarC == 0) |
9636 | return true; |
9637 | |
9638 | // First, compute the minimum iteration count required so that the vector |
9639 | // loop outperforms the scalar loop. |
9640 | // The total cost of the scalar loop is |
9641 | // ScalarC * TC |
9642 | // where |
9643 | // * TC is the actual trip count of the loop. |
9644 | // * ScalarC is the cost of a single scalar iteration. |
9645 | // |
9646 | // The total cost of the vector loop is |
9647 | // RtC + VecC * (TC / VF) + EpiC |
9648 | // where |
9649 | // * RtC is the cost of the generated runtime checks plus the cost of |
9650 | // performing any additional work in the vector.early.exit block for loops |
9651 | // with uncountable early exits. |
9652 | // * VecC is the cost of a single vector iteration. |
9653 | // * TC is the actual trip count of the loop |
9654 | // * VF is the vectorization factor |
9655 | // * EpiCost is the cost of the generated epilogue, including the cost |
9656 | // of the remaining scalar operations. |
9657 | // |
9658 | // Vectorization is profitable once the total vector cost is less than the |
9659 | // total scalar cost: |
9660 | // RtC + VecC * (TC / VF) + EpiC < ScalarC * TC |
9661 | // |
9662 | // Now we can compute the minimum required trip count TC as |
9663 | // VF * (RtC + EpiC) / (ScalarC * VF - VecC) < TC |
9664 | // |
9665 | // For now we assume the epilogue cost EpiC = 0 for simplicity. Note that |
9666 | // the computations are performed on doubles, not integers and the result |
9667 | // is rounded up, hence we get an upper estimate of the TC. |
9668 | unsigned IntVF = getEstimatedRuntimeVF(VF: VF.Width, VScale); |
9669 | uint64_t RtC = TotalCost.getValue(); |
9670 | uint64_t Div = ScalarC * IntVF - VF.Cost.getValue(); |
9671 | uint64_t MinTC1 = Div == 0 ? 0 : divideCeil(Numerator: RtC * IntVF, Denominator: Div); |
9672 | |
9673 | // Second, compute a minimum iteration count so that the cost of the |
9674 | // runtime checks is only a fraction of the total scalar loop cost. This |
9675 | // adds a loop-dependent bound on the overhead incurred if the runtime |
9676 | // checks fail. In case the runtime checks fail, the cost is RtC + ScalarC |
9677 | // * TC. To bound the runtime check to be a fraction 1/X of the scalar |
9678 | // cost, compute |
9679 | // RtC < ScalarC * TC * (1 / X) ==> RtC * X / ScalarC < TC |
9680 | uint64_t MinTC2 = divideCeil(Numerator: RtC * 10, Denominator: ScalarC); |
9681 | |
9682 | // Now pick the larger minimum. If it is not a multiple of VF and a scalar |
9683 | // epilogue is allowed, choose the next closest multiple of VF. This should |
9684 | // partly compensate for ignoring the epilogue cost. |
9685 | uint64_t MinTC = std::max(a: MinTC1, b: MinTC2); |
9686 | if (SEL == CM_ScalarEpilogueAllowed) |
9687 | MinTC = alignTo(Value: MinTC, Align: IntVF); |
9688 | VF.MinProfitableTripCount = ElementCount::getFixed(MinVal: MinTC); |
9689 | |
9690 | LLVM_DEBUG( |
9691 | dbgs() << "LV: Minimum required TC for runtime checks to be profitable:" |
9692 | << VF.MinProfitableTripCount << "\n" ); |
9693 | |
9694 | // Skip vectorization if the expected trip count is less than the minimum |
9695 | // required trip count. |
9696 | if (auto ExpectedTC = getSmallBestKnownTC(PSE, L)) { |
9697 | if (ElementCount::isKnownLT(LHS: *ExpectedTC, RHS: VF.MinProfitableTripCount)) { |
9698 | LLVM_DEBUG(dbgs() << "LV: Vectorization is not beneficial: expected " |
9699 | "trip count < minimum profitable VF (" |
9700 | << *ExpectedTC << " < " << VF.MinProfitableTripCount |
9701 | << ")\n" ); |
9702 | |
9703 | return false; |
9704 | } |
9705 | } |
9706 | return true; |
9707 | } |
9708 | |
9709 | LoopVectorizePass::LoopVectorizePass(LoopVectorizeOptions Opts) |
9710 | : InterleaveOnlyWhenForced(Opts.InterleaveOnlyWhenForced || |
9711 | !EnableLoopInterleaving), |
9712 | VectorizeOnlyWhenForced(Opts.VectorizeOnlyWhenForced || |
9713 | !EnableLoopVectorization) {} |
9714 | |
9715 | /// Prepare \p MainPlan for vectorizing the main vector loop during epilogue |
9716 | /// vectorization. Remove ResumePhis from \p MainPlan for inductions that |
9717 | /// don't have a corresponding wide induction in \p EpiPlan. |
9718 | static void preparePlanForMainVectorLoop(VPlan &MainPlan, VPlan &EpiPlan) { |
9719 | // Collect PHI nodes of widened phis in the VPlan for the epilogue. Those |
9720 | // will need their resume-values computed in the main vector loop. Others |
9721 | // can be removed from the main VPlan. |
9722 | SmallPtrSet<PHINode *, 2> EpiWidenedPhis; |
9723 | for (VPRecipeBase &R : |
9724 | EpiPlan.getVectorLoopRegion()->getEntryBasicBlock()->phis()) { |
9725 | if (isa<VPCanonicalIVPHIRecipe>(Val: &R)) |
9726 | continue; |
9727 | EpiWidenedPhis.insert( |
9728 | Ptr: cast<PHINode>(Val: R.getVPSingleValue()->getUnderlyingValue())); |
9729 | } |
9730 | for (VPRecipeBase &R : |
9731 | make_early_inc_range(Range: MainPlan.getScalarHeader()->phis())) { |
9732 | auto *VPIRInst = cast<VPIRPhi>(Val: &R); |
9733 | if (EpiWidenedPhis.contains(Ptr: &VPIRInst->getIRPhi())) |
9734 | continue; |
9735 | // There is no corresponding wide induction in the epilogue plan that would |
9736 | // need a resume value. Remove the VPIRInst wrapping the scalar header phi |
9737 | // together with the corresponding ResumePhi. The resume values for the |
9738 | // scalar loop will be created during execution of EpiPlan. |
9739 | VPRecipeBase *ResumePhi = VPIRInst->getOperand(N: 0)->getDefiningRecipe(); |
9740 | VPIRInst->eraseFromParent(); |
9741 | ResumePhi->eraseFromParent(); |
9742 | } |
9743 | VPlanTransforms::runPass(Fn: VPlanTransforms::removeDeadRecipes, Plan&: MainPlan); |
9744 | |
9745 | using namespace VPlanPatternMatch; |
9746 | // When vectorizing the epilogue, FindFirstIV & FindLastIV reductions can |
9747 | // introduce multiple uses of undef/poison. If the reduction start value may |
9748 | // be undef or poison it needs to be frozen and the frozen start has to be |
9749 | // used when computing the reduction result. We also need to use the frozen |
9750 | // value in the resume phi generated by the main vector loop, as this is also |
9751 | // used to compute the reduction result after the epilogue vector loop. |
9752 | auto AddFreezeForFindLastIVReductions = [](VPlan &Plan, |
9753 | bool UpdateResumePhis) { |
9754 | VPBuilder Builder(Plan.getEntry()); |
9755 | for (VPRecipeBase &R : *Plan.getMiddleBlock()) { |
9756 | auto *VPI = dyn_cast<VPInstruction>(Val: &R); |
9757 | if (!VPI || VPI->getOpcode() != VPInstruction::ComputeFindIVResult) |
9758 | continue; |
9759 | VPValue *OrigStart = VPI->getOperand(N: 1); |
9760 | if (isGuaranteedNotToBeUndefOrPoison(V: OrigStart->getLiveInIRValue())) |
9761 | continue; |
9762 | VPInstruction *Freeze = |
9763 | Builder.createNaryOp(Opcode: Instruction::Freeze, Operands: {OrigStart}, Inst: {}, Name: "fr" ); |
9764 | VPI->setOperand(I: 1, New: Freeze); |
9765 | if (UpdateResumePhis) |
9766 | OrigStart->replaceUsesWithIf(New: Freeze, ShouldReplace: [Freeze](VPUser &U, unsigned) { |
9767 | return Freeze != &U && isa<VPPhi>(Val: &U); |
9768 | }); |
9769 | } |
9770 | }; |
9771 | AddFreezeForFindLastIVReductions(MainPlan, true); |
9772 | AddFreezeForFindLastIVReductions(EpiPlan, false); |
9773 | |
9774 | VPBasicBlock *MainScalarPH = MainPlan.getScalarPreheader(); |
9775 | VPValue *VectorTC = &MainPlan.getVectorTripCount(); |
9776 | // If there is a suitable resume value for the canonical induction in the |
9777 | // scalar (which will become vector) epilogue loop we are done. Otherwise |
9778 | // create it below. |
9779 | if (any_of(Range&: *MainScalarPH, P: [VectorTC](VPRecipeBase &R) { |
9780 | return match(V: &R, P: m_VPInstruction<Instruction::PHI>(Op0: m_Specific(VPV: VectorTC), |
9781 | Op1: m_SpecificInt(V: 0))); |
9782 | })) |
9783 | return; |
9784 | VPBuilder ScalarPHBuilder(MainScalarPH, MainScalarPH->begin()); |
9785 | ScalarPHBuilder.createScalarPhi( |
9786 | IncomingValues: {VectorTC, MainPlan.getCanonicalIV()->getStartValue()}, DL: {}, |
9787 | Name: "vec.epilog.resume.val" ); |
9788 | } |
9789 | |
9790 | /// Prepare \p Plan for vectorizing the epilogue loop. That is, re-use expanded |
9791 | /// SCEVs from \p ExpandedSCEVs and set resume values for header recipes. |
9792 | static void |
9793 | preparePlanForEpilogueVectorLoop(VPlan &Plan, Loop *L, |
9794 | const SCEV2ValueTy &ExpandedSCEVs, |
9795 | const EpilogueLoopVectorizationInfo &EPI) { |
9796 | VPRegionBlock *VectorLoop = Plan.getVectorLoopRegion(); |
9797 | VPBasicBlock * = VectorLoop->getEntryBasicBlock(); |
9798 | Header->setName("vec.epilog.vector.body" ); |
9799 | |
9800 | DenseMap<Value *, Value *> ToFrozen; |
9801 | // Ensure that the start values for all header phi recipes are updated before |
9802 | // vectorizing the epilogue loop. |
9803 | for (VPRecipeBase &R : Header->phis()) { |
9804 | if (auto *IV = dyn_cast<VPCanonicalIVPHIRecipe>(Val: &R)) { |
9805 | // When vectorizing the epilogue loop, the canonical induction start |
9806 | // value needs to be changed from zero to the value after the main |
9807 | // vector loop. Find the resume value created during execution of the main |
9808 | // VPlan. |
9809 | // FIXME: Improve modeling for canonical IV start values in the epilogue |
9810 | // loop. |
9811 | using namespace llvm::PatternMatch; |
9812 | Type *IdxTy = IV->getScalarType(); |
9813 | PHINode *EPResumeVal = find_singleton<PHINode>( |
9814 | Range: L->getLoopPreheader()->phis(), |
9815 | P: [&EPI, IdxTy](PHINode &P, bool) -> PHINode * { |
9816 | if (P.getType() == IdxTy && |
9817 | match( |
9818 | V: P.getIncomingValueForBlock(BB: EPI.MainLoopIterationCountCheck), |
9819 | P: m_SpecificInt(V: 0)) && |
9820 | all_of(Range: P.incoming_values(), P: [&EPI](Value *Inc) { |
9821 | return Inc == EPI.VectorTripCount || |
9822 | match(V: Inc, P: m_SpecificInt(V: 0)); |
9823 | })) |
9824 | return &P; |
9825 | return nullptr; |
9826 | }); |
9827 | assert(EPResumeVal && "must have a resume value for the canonical IV" ); |
9828 | VPValue *VPV = Plan.getOrAddLiveIn(V: EPResumeVal); |
9829 | assert(all_of(IV->users(), |
9830 | [](const VPUser *U) { |
9831 | return isa<VPScalarIVStepsRecipe>(U) || |
9832 | isa<VPDerivedIVRecipe>(U) || |
9833 | cast<VPRecipeBase>(U)->isScalarCast() || |
9834 | cast<VPInstruction>(U)->getOpcode() == |
9835 | Instruction::Add; |
9836 | }) && |
9837 | "the canonical IV should only be used by its increment or " |
9838 | "ScalarIVSteps when resetting the start value" ); |
9839 | IV->setOperand(I: 0, New: VPV); |
9840 | continue; |
9841 | } |
9842 | |
9843 | Value *ResumeV = nullptr; |
9844 | // TODO: Move setting of resume values to prepareToExecute. |
9845 | if (auto *ReductionPhi = dyn_cast<VPReductionPHIRecipe>(Val: &R)) { |
9846 | auto *RdxResult = |
9847 | cast<VPInstruction>(Val: *find_if(Range: ReductionPhi->users(), P: [](VPUser *U) { |
9848 | auto *VPI = dyn_cast<VPInstruction>(Val: U); |
9849 | return VPI && |
9850 | (VPI->getOpcode() == VPInstruction::ComputeAnyOfResult || |
9851 | VPI->getOpcode() == VPInstruction::ComputeReductionResult || |
9852 | VPI->getOpcode() == VPInstruction::ComputeFindIVResult); |
9853 | })); |
9854 | ResumeV = cast<PHINode>(Val: ReductionPhi->getUnderlyingInstr()) |
9855 | ->getIncomingValueForBlock(BB: L->getLoopPreheader()); |
9856 | const RecurrenceDescriptor &RdxDesc = |
9857 | ReductionPhi->getRecurrenceDescriptor(); |
9858 | RecurKind RK = RdxDesc.getRecurrenceKind(); |
9859 | if (RecurrenceDescriptor::isAnyOfRecurrenceKind(Kind: RK)) { |
9860 | Value *StartV = RdxResult->getOperand(N: 1)->getLiveInIRValue(); |
9861 | assert(RdxDesc.getRecurrenceStartValue() == StartV && |
9862 | "start value from ComputeAnyOfResult must match" ); |
9863 | |
9864 | // VPReductionPHIRecipes for AnyOf reductions expect a boolean as |
9865 | // start value; compare the final value from the main vector loop |
9866 | // to the start value. |
9867 | BasicBlock *PBB = cast<Instruction>(Val: ResumeV)->getParent(); |
9868 | IRBuilder<> Builder(PBB, PBB->getFirstNonPHIIt()); |
9869 | ResumeV = Builder.CreateICmpNE(LHS: ResumeV, RHS: StartV); |
9870 | } else if (RecurrenceDescriptor::isFindIVRecurrenceKind(Kind: RK)) { |
9871 | Value *StartV = getStartValueFromReductionResult(RdxResult); |
9872 | assert(RdxDesc.getRecurrenceStartValue() == StartV && |
9873 | "start value from ComputeFinIVResult must match" ); |
9874 | |
9875 | ToFrozen[StartV] = cast<PHINode>(Val: ResumeV)->getIncomingValueForBlock( |
9876 | BB: EPI.MainLoopIterationCountCheck); |
9877 | |
9878 | // VPReductionPHIRecipe for FindFirstIV/FindLastIV reductions requires |
9879 | // an adjustment to the resume value. The resume value is adjusted to |
9880 | // the sentinel value when the final value from the main vector loop |
9881 | // equals the start value. This ensures correctness when the start value |
9882 | // might not be less than the minimum value of a monotonically |
9883 | // increasing induction variable. |
9884 | BasicBlock *ResumeBB = cast<Instruction>(Val: ResumeV)->getParent(); |
9885 | IRBuilder<> Builder(ResumeBB, ResumeBB->getFirstNonPHIIt()); |
9886 | Value *Cmp = Builder.CreateICmpEQ(LHS: ResumeV, RHS: ToFrozen[StartV]); |
9887 | Value *Sentinel = RdxResult->getOperand(N: 2)->getLiveInIRValue(); |
9888 | ResumeV = Builder.CreateSelect(C: Cmp, True: Sentinel, False: ResumeV); |
9889 | } else { |
9890 | VPValue *StartVal = Plan.getOrAddLiveIn(V: ResumeV); |
9891 | auto *PhiR = dyn_cast<VPReductionPHIRecipe>(Val: &R); |
9892 | if (auto *VPI = dyn_cast<VPInstruction>(Val: PhiR->getStartValue())) { |
9893 | assert(VPI->getOpcode() == VPInstruction::ReductionStartVector && |
9894 | "unexpected start value" ); |
9895 | VPI->setOperand(I: 0, New: StartVal); |
9896 | continue; |
9897 | } |
9898 | } |
9899 | } else { |
9900 | // Retrieve the induction resume values for wide inductions from |
9901 | // their original phi nodes in the scalar loop. |
9902 | PHINode *IndPhi = cast<VPWidenInductionRecipe>(Val: &R)->getPHINode(); |
9903 | // Hook up to the PHINode generated by a ResumePhi recipe of main |
9904 | // loop VPlan, which feeds the scalar loop. |
9905 | ResumeV = IndPhi->getIncomingValueForBlock(BB: L->getLoopPreheader()); |
9906 | } |
9907 | assert(ResumeV && "Must have a resume value" ); |
9908 | VPValue *StartVal = Plan.getOrAddLiveIn(V: ResumeV); |
9909 | cast<VPHeaderPHIRecipe>(Val: &R)->setStartValue(StartVal); |
9910 | } |
9911 | |
9912 | // For some VPValues in the epilogue plan we must re-use the generated IR |
9913 | // values from the main plan. Replace them with live-in VPValues. |
9914 | // TODO: This is a workaround needed for epilogue vectorization and it |
9915 | // should be removed once induction resume value creation is done |
9916 | // directly in VPlan. |
9917 | for (auto &R : make_early_inc_range(Range&: *Plan.getEntry())) { |
9918 | // Re-use frozen values from the main plan for Freeze VPInstructions in the |
9919 | // epilogue plan. This ensures all users use the same frozen value. |
9920 | auto *VPI = dyn_cast<VPInstruction>(Val: &R); |
9921 | if (VPI && VPI->getOpcode() == Instruction::Freeze) { |
9922 | VPI->replaceAllUsesWith(New: Plan.getOrAddLiveIn( |
9923 | V: ToFrozen.lookup(Val: VPI->getOperand(N: 0)->getLiveInIRValue()))); |
9924 | continue; |
9925 | } |
9926 | |
9927 | // Re-use the trip count and steps expanded for the main loop, as |
9928 | // skeleton creation needs it as a value that dominates both the scalar |
9929 | // and vector epilogue loops |
9930 | auto *ExpandR = dyn_cast<VPExpandSCEVRecipe>(Val: &R); |
9931 | if (!ExpandR) |
9932 | continue; |
9933 | VPValue *ExpandedVal = |
9934 | Plan.getOrAddLiveIn(V: ExpandedSCEVs.lookup(Val: ExpandR->getSCEV())); |
9935 | ExpandR->replaceAllUsesWith(New: ExpandedVal); |
9936 | if (Plan.getTripCount() == ExpandR) |
9937 | Plan.resetTripCount(NewTripCount: ExpandedVal); |
9938 | ExpandR->eraseFromParent(); |
9939 | } |
9940 | } |
9941 | |
9942 | // Generate bypass values from the additional bypass block. Note that when the |
9943 | // vectorized epilogue is skipped due to iteration count check, then the |
9944 | // resume value for the induction variable comes from the trip count of the |
9945 | // main vector loop, passed as the second argument. |
9946 | static Value *createInductionAdditionalBypassValues( |
9947 | PHINode *OrigPhi, const InductionDescriptor &II, IRBuilder<> &BypassBuilder, |
9948 | const SCEV2ValueTy &ExpandedSCEVs, Value *MainVectorTripCount, |
9949 | Instruction *OldInduction) { |
9950 | Value *Step = getExpandedStep(ID: II, ExpandedSCEVs); |
9951 | // For the primary induction the additional bypass end value is known. |
9952 | // Otherwise it is computed. |
9953 | Value *EndValueFromAdditionalBypass = MainVectorTripCount; |
9954 | if (OrigPhi != OldInduction) { |
9955 | auto *BinOp = II.getInductionBinOp(); |
9956 | // Fast-math-flags propagate from the original induction instruction. |
9957 | if (isa_and_nonnull<FPMathOperator>(Val: BinOp)) |
9958 | BypassBuilder.setFastMathFlags(BinOp->getFastMathFlags()); |
9959 | |
9960 | // Compute the end value for the additional bypass. |
9961 | EndValueFromAdditionalBypass = |
9962 | emitTransformedIndex(B&: BypassBuilder, Index: MainVectorTripCount, |
9963 | StartValue: II.getStartValue(), Step, InductionKind: II.getKind(), InductionBinOp: BinOp); |
9964 | EndValueFromAdditionalBypass->setName("ind.end" ); |
9965 | } |
9966 | return EndValueFromAdditionalBypass; |
9967 | } |
9968 | |
9969 | bool LoopVectorizePass::processLoop(Loop *L) { |
9970 | assert((EnableVPlanNativePath || L->isInnermost()) && |
9971 | "VPlan-native path is not enabled. Only process inner loops." ); |
9972 | |
9973 | LLVM_DEBUG(dbgs() << "\nLV: Checking a loop in '" |
9974 | << L->getHeader()->getParent()->getName() << "' from " |
9975 | << L->getLocStr() << "\n" ); |
9976 | |
9977 | LoopVectorizeHints Hints(L, InterleaveOnlyWhenForced, *ORE, TTI); |
9978 | |
9979 | LLVM_DEBUG( |
9980 | dbgs() << "LV: Loop hints:" |
9981 | << " force=" |
9982 | << (Hints.getForce() == LoopVectorizeHints::FK_Disabled |
9983 | ? "disabled" |
9984 | : (Hints.getForce() == LoopVectorizeHints::FK_Enabled |
9985 | ? "enabled" |
9986 | : "?" )) |
9987 | << " width=" << Hints.getWidth() |
9988 | << " interleave=" << Hints.getInterleave() << "\n" ); |
9989 | |
9990 | // Function containing loop |
9991 | Function *F = L->getHeader()->getParent(); |
9992 | |
9993 | // Looking at the diagnostic output is the only way to determine if a loop |
9994 | // was vectorized (other than looking at the IR or machine code), so it |
9995 | // is important to generate an optimization remark for each loop. Most of |
9996 | // these messages are generated as OptimizationRemarkAnalysis. Remarks |
9997 | // generated as OptimizationRemark and OptimizationRemarkMissed are |
9998 | // less verbose reporting vectorized loops and unvectorized loops that may |
9999 | // benefit from vectorization, respectively. |
10000 | |
10001 | if (!Hints.allowVectorization(F, L, VectorizeOnlyWhenForced)) { |
10002 | LLVM_DEBUG(dbgs() << "LV: Loop hints prevent vectorization.\n" ); |
10003 | return false; |
10004 | } |
10005 | |
10006 | PredicatedScalarEvolution PSE(*SE, *L); |
10007 | |
10008 | // Check if it is legal to vectorize the loop. |
10009 | LoopVectorizationRequirements Requirements; |
10010 | LoopVectorizationLegality LVL(L, PSE, DT, TTI, TLI, F, *LAIs, LI, ORE, |
10011 | &Requirements, &Hints, DB, AC, BFI, PSI); |
10012 | if (!LVL.canVectorize(UseVPlanNativePath: EnableVPlanNativePath)) { |
10013 | LLVM_DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n" ); |
10014 | Hints.emitRemarkWithHints(); |
10015 | return false; |
10016 | } |
10017 | |
10018 | if (LVL.hasUncountableEarlyExit() && !EnableEarlyExitVectorization) { |
10019 | reportVectorizationFailure(DebugMsg: "Auto-vectorization of loops with uncountable " |
10020 | "early exit is not enabled" , |
10021 | ORETag: "UncountableEarlyExitLoopsDisabled" , ORE, TheLoop: L); |
10022 | return false; |
10023 | } |
10024 | |
10025 | // Entrance to the VPlan-native vectorization path. Outer loops are processed |
10026 | // here. They may require CFG and instruction level transformations before |
10027 | // even evaluating whether vectorization is profitable. Since we cannot modify |
10028 | // the incoming IR, we need to build VPlan upfront in the vectorization |
10029 | // pipeline. |
10030 | if (!L->isInnermost()) |
10031 | return processLoopInVPlanNativePath(L, PSE, LI, DT, LVL: &LVL, TTI, TLI, DB, AC, |
10032 | ORE, BFI, PSI, Hints, Requirements); |
10033 | |
10034 | assert(L->isInnermost() && "Inner loop expected." ); |
10035 | |
10036 | InterleavedAccessInfo IAI(PSE, L, DT, LI, LVL.getLAI()); |
10037 | bool UseInterleaved = TTI->enableInterleavedAccessVectorization(); |
10038 | |
10039 | // If an override option has been passed in for interleaved accesses, use it. |
10040 | if (EnableInterleavedMemAccesses.getNumOccurrences() > 0) |
10041 | UseInterleaved = EnableInterleavedMemAccesses; |
10042 | |
10043 | // Analyze interleaved memory accesses. |
10044 | if (UseInterleaved) |
10045 | IAI.analyzeInterleaving(EnableMaskedInterleavedGroup: useMaskedInterleavedAccesses(TTI: *TTI)); |
10046 | |
10047 | if (LVL.hasUncountableEarlyExit()) { |
10048 | BasicBlock *LoopLatch = L->getLoopLatch(); |
10049 | if (IAI.requiresScalarEpilogue() || |
10050 | any_of(Range: LVL.getCountableExitingBlocks(), |
10051 | P: [LoopLatch](BasicBlock *BB) { return BB != LoopLatch; })) { |
10052 | reportVectorizationFailure(DebugMsg: "Auto-vectorization of early exit loops " |
10053 | "requiring a scalar epilogue is unsupported" , |
10054 | ORETag: "UncountableEarlyExitUnsupported" , ORE, TheLoop: L); |
10055 | return false; |
10056 | } |
10057 | } |
10058 | |
10059 | // Check the function attributes and profiles to find out if this function |
10060 | // should be optimized for size. |
10061 | ScalarEpilogueLowering SEL = |
10062 | getScalarEpilogueLowering(F, L, Hints, PSI, BFI, TTI, TLI, LVL, IAI: &IAI); |
10063 | |
10064 | // Check the loop for a trip count threshold: vectorize loops with a tiny trip |
10065 | // count by optimizing for size, to minimize overheads. |
10066 | auto ExpectedTC = getSmallBestKnownTC(PSE, L); |
10067 | if (ExpectedTC && ExpectedTC->isFixed() && |
10068 | ExpectedTC->getFixedValue() < TinyTripCountVectorThreshold) { |
10069 | LLVM_DEBUG(dbgs() << "LV: Found a loop with a very small trip count. " |
10070 | << "This loop is worth vectorizing only if no scalar " |
10071 | << "iteration overheads are incurred." ); |
10072 | if (Hints.getForce() == LoopVectorizeHints::FK_Enabled) |
10073 | LLVM_DEBUG(dbgs() << " But vectorizing was explicitly forced.\n" ); |
10074 | else { |
10075 | LLVM_DEBUG(dbgs() << "\n" ); |
10076 | // Predicate tail-folded loops are efficient even when the loop |
10077 | // iteration count is low. However, setting the epilogue policy to |
10078 | // `CM_ScalarEpilogueNotAllowedLowTripLoop` prevents vectorizing loops |
10079 | // with runtime checks. It's more effective to let |
10080 | // `isOutsideLoopWorkProfitable` determine if vectorization is |
10081 | // beneficial for the loop. |
10082 | if (SEL != CM_ScalarEpilogueNotNeededUsePredicate) |
10083 | SEL = CM_ScalarEpilogueNotAllowedLowTripLoop; |
10084 | } |
10085 | } |
10086 | |
10087 | // Check the function attributes to see if implicit floats or vectors are |
10088 | // allowed. |
10089 | if (F->hasFnAttribute(Kind: Attribute::NoImplicitFloat)) { |
10090 | reportVectorizationFailure( |
10091 | DebugMsg: "Can't vectorize when the NoImplicitFloat attribute is used" , |
10092 | OREMsg: "loop not vectorized due to NoImplicitFloat attribute" , |
10093 | ORETag: "NoImplicitFloat" , ORE, TheLoop: L); |
10094 | Hints.emitRemarkWithHints(); |
10095 | return false; |
10096 | } |
10097 | |
10098 | // Check if the target supports potentially unsafe FP vectorization. |
10099 | // FIXME: Add a check for the type of safety issue (denormal, signaling) |
10100 | // for the target we're vectorizing for, to make sure none of the |
10101 | // additional fp-math flags can help. |
10102 | if (Hints.isPotentiallyUnsafe() && |
10103 | TTI->isFPVectorizationPotentiallyUnsafe()) { |
10104 | reportVectorizationFailure( |
10105 | DebugMsg: "Potentially unsafe FP op prevents vectorization" , |
10106 | OREMsg: "loop not vectorized due to unsafe FP support." , |
10107 | ORETag: "UnsafeFP" , ORE, TheLoop: L); |
10108 | Hints.emitRemarkWithHints(); |
10109 | return false; |
10110 | } |
10111 | |
10112 | bool AllowOrderedReductions; |
10113 | // If the flag is set, use that instead and override the TTI behaviour. |
10114 | if (ForceOrderedReductions.getNumOccurrences() > 0) |
10115 | AllowOrderedReductions = ForceOrderedReductions; |
10116 | else |
10117 | AllowOrderedReductions = TTI->enableOrderedReductions(); |
10118 | if (!LVL.canVectorizeFPMath(EnableStrictReductions: AllowOrderedReductions)) { |
10119 | ORE->emit(RemarkBuilder: [&]() { |
10120 | auto *ExactFPMathInst = Requirements.getExactFPInst(); |
10121 | return OptimizationRemarkAnalysisFPCommute(DEBUG_TYPE, "CantReorderFPOps" , |
10122 | ExactFPMathInst->getDebugLoc(), |
10123 | ExactFPMathInst->getParent()) |
10124 | << "loop not vectorized: cannot prove it is safe to reorder " |
10125 | "floating-point operations" ; |
10126 | }); |
10127 | LLVM_DEBUG(dbgs() << "LV: loop not vectorized: cannot prove it is safe to " |
10128 | "reorder floating-point operations\n" ); |
10129 | Hints.emitRemarkWithHints(); |
10130 | return false; |
10131 | } |
10132 | |
10133 | // Use the cost model. |
10134 | LoopVectorizationCostModel CM(SEL, L, PSE, LI, &LVL, *TTI, TLI, DB, AC, ORE, |
10135 | F, &Hints, IAI, PSI, BFI); |
10136 | // Use the planner for vectorization. |
10137 | LoopVectorizationPlanner LVP(L, LI, DT, TLI, *TTI, &LVL, CM, IAI, PSE, Hints, |
10138 | ORE); |
10139 | |
10140 | // Get user vectorization factor and interleave count. |
10141 | ElementCount UserVF = Hints.getWidth(); |
10142 | unsigned UserIC = Hints.getInterleave(); |
10143 | if (LVL.hasUncountableEarlyExit() && UserIC != 1 && |
10144 | !VectorizerParams::isInterleaveForced()) { |
10145 | UserIC = 1; |
10146 | reportVectorizationInfo(Msg: "Interleaving not supported for loops " |
10147 | "with uncountable early exits" , |
10148 | ORETag: "InterleaveEarlyExitDisabled" , ORE, TheLoop: L); |
10149 | } |
10150 | |
10151 | // Plan how to best vectorize. |
10152 | LVP.plan(UserVF, UserIC); |
10153 | VectorizationFactor VF = LVP.computeBestVF(); |
10154 | unsigned IC = 1; |
10155 | |
10156 | if (ORE->allowExtraAnalysis(LV_NAME)) |
10157 | LVP.emitInvalidCostRemarks(ORE); |
10158 | |
10159 | bool AddBranchWeights = |
10160 | hasBranchWeightMD(I: *L->getLoopLatch()->getTerminator()); |
10161 | GeneratedRTChecks Checks(PSE, DT, LI, TTI, F->getDataLayout(), |
10162 | AddBranchWeights, CM.CostKind); |
10163 | if (LVP.hasPlanWithVF(VF: VF.Width)) { |
10164 | // Select the interleave count. |
10165 | IC = CM.selectInterleaveCount(Plan&: LVP.getPlanFor(VF: VF.Width), VF: VF.Width, LoopCost: VF.Cost); |
10166 | |
10167 | unsigned SelectedIC = std::max(a: IC, b: UserIC); |
10168 | // Optimistically generate runtime checks if they are needed. Drop them if |
10169 | // they turn out to not be profitable. |
10170 | if (VF.Width.isVector() || SelectedIC > 1) |
10171 | Checks.create(L, LAI: *LVL.getLAI(), UnionPred: PSE.getPredicate(), VF: VF.Width, IC: SelectedIC); |
10172 | |
10173 | // Check if it is profitable to vectorize with runtime checks. |
10174 | bool ForceVectorization = |
10175 | Hints.getForce() == LoopVectorizeHints::FK_Enabled; |
10176 | VPCostContext CostCtx(CM.TTI, *CM.TLI, CM.Legal->getWidestInductionType(), |
10177 | CM, CM.CostKind); |
10178 | if (!ForceVectorization && |
10179 | !isOutsideLoopWorkProfitable(Checks, VF, L, PSE, CostCtx, |
10180 | Plan&: LVP.getPlanFor(VF: VF.Width), SEL, |
10181 | VScale: CM.getVScaleForTuning())) { |
10182 | ORE->emit(RemarkBuilder: [&]() { |
10183 | return OptimizationRemarkAnalysisAliasing( |
10184 | DEBUG_TYPE, "CantReorderMemOps" , L->getStartLoc(), |
10185 | L->getHeader()) |
10186 | << "loop not vectorized: cannot prove it is safe to reorder " |
10187 | "memory operations" ; |
10188 | }); |
10189 | LLVM_DEBUG(dbgs() << "LV: Too many memory checks needed.\n" ); |
10190 | Hints.emitRemarkWithHints(); |
10191 | return false; |
10192 | } |
10193 | } |
10194 | |
10195 | // Identify the diagnostic messages that should be produced. |
10196 | std::pair<StringRef, std::string> VecDiagMsg, IntDiagMsg; |
10197 | bool VectorizeLoop = true, InterleaveLoop = true; |
10198 | if (VF.Width.isScalar()) { |
10199 | LLVM_DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n" ); |
10200 | VecDiagMsg = { |
10201 | "VectorizationNotBeneficial" , |
10202 | "the cost-model indicates that vectorization is not beneficial" }; |
10203 | VectorizeLoop = false; |
10204 | } |
10205 | |
10206 | if (!LVP.hasPlanWithVF(VF: VF.Width) && UserIC > 1) { |
10207 | // Tell the user interleaving was avoided up-front, despite being explicitly |
10208 | // requested. |
10209 | LLVM_DEBUG(dbgs() << "LV: Ignoring UserIC, because vectorization and " |
10210 | "interleaving should be avoided up front\n" ); |
10211 | IntDiagMsg = {"InterleavingAvoided" , |
10212 | "Ignoring UserIC, because interleaving was avoided up front" }; |
10213 | InterleaveLoop = false; |
10214 | } else if (IC == 1 && UserIC <= 1) { |
10215 | // Tell the user interleaving is not beneficial. |
10216 | LLVM_DEBUG(dbgs() << "LV: Interleaving is not beneficial.\n" ); |
10217 | IntDiagMsg = { |
10218 | "InterleavingNotBeneficial" , |
10219 | "the cost-model indicates that interleaving is not beneficial" }; |
10220 | InterleaveLoop = false; |
10221 | if (UserIC == 1) { |
10222 | IntDiagMsg.first = "InterleavingNotBeneficialAndDisabled" ; |
10223 | IntDiagMsg.second += |
10224 | " and is explicitly disabled or interleave count is set to 1" ; |
10225 | } |
10226 | } else if (IC > 1 && UserIC == 1) { |
10227 | // Tell the user interleaving is beneficial, but it explicitly disabled. |
10228 | LLVM_DEBUG( |
10229 | dbgs() << "LV: Interleaving is beneficial but is explicitly disabled." ); |
10230 | IntDiagMsg = {"InterleavingBeneficialButDisabled" , |
10231 | "the cost-model indicates that interleaving is beneficial " |
10232 | "but is explicitly disabled or interleave count is set to 1" }; |
10233 | InterleaveLoop = false; |
10234 | } |
10235 | |
10236 | // If there is a histogram in the loop, do not just interleave without |
10237 | // vectorizing. The order of operations will be incorrect without the |
10238 | // histogram intrinsics, which are only used for recipes with VF > 1. |
10239 | if (!VectorizeLoop && InterleaveLoop && LVL.hasHistograms()) { |
10240 | LLVM_DEBUG(dbgs() << "LV: Not interleaving without vectorization due " |
10241 | << "to histogram operations.\n" ); |
10242 | IntDiagMsg = { |
10243 | "HistogramPreventsScalarInterleaving" , |
10244 | "Unable to interleave without vectorization due to constraints on " |
10245 | "the order of histogram operations" }; |
10246 | InterleaveLoop = false; |
10247 | } |
10248 | |
10249 | // Override IC if user provided an interleave count. |
10250 | IC = UserIC > 0 ? UserIC : IC; |
10251 | |
10252 | // Emit diagnostic messages, if any. |
10253 | const char *VAPassName = Hints.vectorizeAnalysisPassName(); |
10254 | if (!VectorizeLoop && !InterleaveLoop) { |
10255 | // Do not vectorize or interleaving the loop. |
10256 | ORE->emit(RemarkBuilder: [&]() { |
10257 | return OptimizationRemarkMissed(VAPassName, VecDiagMsg.first, |
10258 | L->getStartLoc(), L->getHeader()) |
10259 | << VecDiagMsg.second; |
10260 | }); |
10261 | ORE->emit(RemarkBuilder: [&]() { |
10262 | return OptimizationRemarkMissed(LV_NAME, IntDiagMsg.first, |
10263 | L->getStartLoc(), L->getHeader()) |
10264 | << IntDiagMsg.second; |
10265 | }); |
10266 | return false; |
10267 | } |
10268 | |
10269 | if (!VectorizeLoop && InterleaveLoop) { |
10270 | LLVM_DEBUG(dbgs() << "LV: Interleave Count is " << IC << '\n'); |
10271 | ORE->emit(RemarkBuilder: [&]() { |
10272 | return OptimizationRemarkAnalysis(VAPassName, VecDiagMsg.first, |
10273 | L->getStartLoc(), L->getHeader()) |
10274 | << VecDiagMsg.second; |
10275 | }); |
10276 | } else if (VectorizeLoop && !InterleaveLoop) { |
10277 | LLVM_DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width |
10278 | << ") in " << L->getLocStr() << '\n'); |
10279 | ORE->emit(RemarkBuilder: [&]() { |
10280 | return OptimizationRemarkAnalysis(LV_NAME, IntDiagMsg.first, |
10281 | L->getStartLoc(), L->getHeader()) |
10282 | << IntDiagMsg.second; |
10283 | }); |
10284 | } else if (VectorizeLoop && InterleaveLoop) { |
10285 | LLVM_DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width |
10286 | << ") in " << L->getLocStr() << '\n'); |
10287 | LLVM_DEBUG(dbgs() << "LV: Interleave Count is " << IC << '\n'); |
10288 | } |
10289 | |
10290 | bool DisableRuntimeUnroll = false; |
10291 | MDNode *OrigLoopID = L->getLoopID(); |
10292 | { |
10293 | using namespace ore; |
10294 | if (!VectorizeLoop) { |
10295 | assert(IC > 1 && "interleave count should not be 1 or 0" ); |
10296 | // If we decided that it is not legal to vectorize the loop, then |
10297 | // interleave it. |
10298 | VPlan &BestPlan = LVP.getPlanFor(VF: VF.Width); |
10299 | InnerLoopVectorizer Unroller( |
10300 | L, PSE, LI, DT, TLI, TTI, AC, ORE, ElementCount::getFixed(MinVal: 1), |
10301 | ElementCount::getFixed(MinVal: 1), IC, &CM, BFI, PSI, Checks, BestPlan); |
10302 | |
10303 | LVP.executePlan(BestVF: VF.Width, BestUF: IC, BestVPlan&: BestPlan, ILV&: Unroller, DT, VectorizingEpilogue: false); |
10304 | |
10305 | ORE->emit(RemarkBuilder: [&]() { |
10306 | return OptimizationRemark(LV_NAME, "Interleaved" , L->getStartLoc(), |
10307 | L->getHeader()) |
10308 | << "interleaved loop (interleaved count: " |
10309 | << NV("InterleaveCount" , IC) << ")" ; |
10310 | }); |
10311 | } else { |
10312 | // If we decided that it is *legal* to vectorize the loop, then do it. |
10313 | |
10314 | VPlan &BestPlan = LVP.getPlanFor(VF: VF.Width); |
10315 | // Consider vectorizing the epilogue too if it's profitable. |
10316 | VectorizationFactor EpilogueVF = |
10317 | LVP.selectEpilogueVectorizationFactor(MainLoopVF: VF.Width, IC); |
10318 | if (EpilogueVF.Width.isVector()) { |
10319 | std::unique_ptr<VPlan> BestMainPlan(BestPlan.duplicate()); |
10320 | |
10321 | // The first pass vectorizes the main loop and creates a scalar epilogue |
10322 | // to be vectorized by executing the plan (potentially with a different |
10323 | // factor) again shortly afterwards. |
10324 | VPlan &BestEpiPlan = LVP.getPlanFor(VF: EpilogueVF.Width); |
10325 | BestEpiPlan.getMiddleBlock()->setName("vec.epilog.middle.block" ); |
10326 | preparePlanForMainVectorLoop(MainPlan&: *BestMainPlan, EpiPlan&: BestEpiPlan); |
10327 | EpilogueLoopVectorizationInfo EPI(VF.Width, IC, EpilogueVF.Width, 1, |
10328 | BestEpiPlan); |
10329 | EpilogueVectorizerMainLoop MainILV(L, PSE, LI, DT, TLI, TTI, AC, ORE, |
10330 | EPI, &CM, BFI, PSI, Checks, |
10331 | *BestMainPlan); |
10332 | auto ExpandedSCEVs = LVP.executePlan(BestVF: EPI.MainLoopVF, BestUF: EPI.MainLoopUF, |
10333 | BestVPlan&: *BestMainPlan, ILV&: MainILV, DT, VectorizingEpilogue: false); |
10334 | ++LoopsVectorized; |
10335 | |
10336 | // Second pass vectorizes the epilogue and adjusts the control flow |
10337 | // edges from the first pass. |
10338 | EPI.MainLoopVF = EPI.EpilogueVF; |
10339 | EPI.MainLoopUF = EPI.EpilogueUF; |
10340 | EpilogueVectorizerEpilogueLoop EpilogILV(L, PSE, LI, DT, TLI, TTI, AC, |
10341 | ORE, EPI, &CM, BFI, PSI, |
10342 | Checks, BestEpiPlan); |
10343 | EpilogILV.setTripCount(MainILV.getTripCount()); |
10344 | preparePlanForEpilogueVectorLoop(Plan&: BestEpiPlan, L, ExpandedSCEVs, EPI); |
10345 | |
10346 | LVP.executePlan(BestVF: EPI.EpilogueVF, BestUF: EPI.EpilogueUF, BestVPlan&: BestEpiPlan, ILV&: EpilogILV, |
10347 | DT, VectorizingEpilogue: true); |
10348 | |
10349 | // Fix induction resume values from the additional bypass block. |
10350 | BasicBlock *BypassBlock = EpilogILV.getAdditionalBypassBlock(); |
10351 | IRBuilder<> BypassBuilder(BypassBlock, |
10352 | BypassBlock->getFirstInsertionPt()); |
10353 | BasicBlock *PH = L->getLoopPreheader(); |
10354 | for (const auto &[IVPhi, II] : LVL.getInductionVars()) { |
10355 | auto *Inc = cast<PHINode>(Val: IVPhi->getIncomingValueForBlock(BB: PH)); |
10356 | Value *V = createInductionAdditionalBypassValues( |
10357 | OrigPhi: IVPhi, II, BypassBuilder, ExpandedSCEVs, MainVectorTripCount: EPI.VectorTripCount, |
10358 | OldInduction: LVL.getPrimaryInduction()); |
10359 | // TODO: Directly add as extra operand to the VPResumePHI recipe. |
10360 | Inc->setIncomingValueForBlock(BB: BypassBlock, V); |
10361 | } |
10362 | ++LoopsEpilogueVectorized; |
10363 | |
10364 | if (!Checks.hasChecks()) |
10365 | DisableRuntimeUnroll = true; |
10366 | } else { |
10367 | InnerLoopVectorizer LB(L, PSE, LI, DT, TLI, TTI, AC, ORE, VF.Width, |
10368 | VF.MinProfitableTripCount, IC, &CM, BFI, PSI, |
10369 | Checks, BestPlan); |
10370 | LVP.executePlan(BestVF: VF.Width, BestUF: IC, BestVPlan&: BestPlan, ILV&: LB, DT, VectorizingEpilogue: false); |
10371 | ++LoopsVectorized; |
10372 | |
10373 | // Add metadata to disable runtime unrolling a scalar loop when there |
10374 | // are no runtime checks about strides and memory. A scalar loop that is |
10375 | // rarely used is not worth unrolling. |
10376 | if (!Checks.hasChecks()) |
10377 | DisableRuntimeUnroll = true; |
10378 | } |
10379 | // Report the vectorization decision. |
10380 | reportVectorization(ORE, TheLoop: L, VF, IC); |
10381 | } |
10382 | |
10383 | if (ORE->allowExtraAnalysis(LV_NAME)) |
10384 | checkMixedPrecision(L, ORE); |
10385 | } |
10386 | |
10387 | assert(DT->verify(DominatorTree::VerificationLevel::Fast) && |
10388 | "DT not preserved correctly" ); |
10389 | |
10390 | std::optional<MDNode *> RemainderLoopID = |
10391 | makeFollowupLoopID(OrigLoopID, FollowupAttrs: {LLVMLoopVectorizeFollowupAll, |
10392 | LLVMLoopVectorizeFollowupEpilogue}); |
10393 | if (RemainderLoopID) { |
10394 | L->setLoopID(*RemainderLoopID); |
10395 | } else { |
10396 | if (DisableRuntimeUnroll) |
10397 | addRuntimeUnrollDisableMetaData(L); |
10398 | |
10399 | // Mark the loop as already vectorized to avoid vectorizing again. |
10400 | Hints.setAlreadyVectorized(); |
10401 | } |
10402 | |
10403 | assert(!verifyFunction(*L->getHeader()->getParent(), &dbgs())); |
10404 | return true; |
10405 | } |
10406 | |
10407 | LoopVectorizeResult LoopVectorizePass::runImpl(Function &F) { |
10408 | |
10409 | // Don't attempt if |
10410 | // 1. the target claims to have no vector registers, and |
10411 | // 2. interleaving won't help ILP. |
10412 | // |
10413 | // The second condition is necessary because, even if the target has no |
10414 | // vector registers, loop vectorization may still enable scalar |
10415 | // interleaving. |
10416 | if (!TTI->getNumberOfRegisters(ClassID: TTI->getRegisterClassForType(Vector: true)) && |
10417 | TTI->getMaxInterleaveFactor(VF: ElementCount::getFixed(MinVal: 1)) < 2) |
10418 | return LoopVectorizeResult(false, false); |
10419 | |
10420 | bool Changed = false, CFGChanged = false; |
10421 | |
10422 | // The vectorizer requires loops to be in simplified form. |
10423 | // Since simplification may add new inner loops, it has to run before the |
10424 | // legality and profitability checks. This means running the loop vectorizer |
10425 | // will simplify all loops, regardless of whether anything end up being |
10426 | // vectorized. |
10427 | for (const auto &L : *LI) |
10428 | Changed |= CFGChanged |= |
10429 | simplifyLoop(L, DT, LI, SE, AC, MSSAU: nullptr, PreserveLCSSA: false /* PreserveLCSSA */); |
10430 | |
10431 | // Build up a worklist of inner-loops to vectorize. This is necessary as |
10432 | // the act of vectorizing or partially unrolling a loop creates new loops |
10433 | // and can invalidate iterators across the loops. |
10434 | SmallVector<Loop *, 8> Worklist; |
10435 | |
10436 | for (Loop *L : *LI) |
10437 | collectSupportedLoops(L&: *L, LI, ORE, V&: Worklist); |
10438 | |
10439 | LoopsAnalyzed += Worklist.size(); |
10440 | |
10441 | // Now walk the identified inner loops. |
10442 | while (!Worklist.empty()) { |
10443 | Loop *L = Worklist.pop_back_val(); |
10444 | |
10445 | // For the inner loops we actually process, form LCSSA to simplify the |
10446 | // transform. |
10447 | Changed |= formLCSSARecursively(L&: *L, DT: *DT, LI, SE); |
10448 | |
10449 | Changed |= CFGChanged |= processLoop(L); |
10450 | |
10451 | if (Changed) { |
10452 | LAIs->clear(); |
10453 | |
10454 | #ifndef NDEBUG |
10455 | if (VerifySCEV) |
10456 | SE->verify(); |
10457 | #endif |
10458 | } |
10459 | } |
10460 | |
10461 | // Process each loop nest in the function. |
10462 | return LoopVectorizeResult(Changed, CFGChanged); |
10463 | } |
10464 | |
10465 | PreservedAnalyses LoopVectorizePass::run(Function &F, |
10466 | FunctionAnalysisManager &AM) { |
10467 | LI = &AM.getResult<LoopAnalysis>(IR&: F); |
10468 | // There are no loops in the function. Return before computing other |
10469 | // expensive analyses. |
10470 | if (LI->empty()) |
10471 | return PreservedAnalyses::all(); |
10472 | SE = &AM.getResult<ScalarEvolutionAnalysis>(IR&: F); |
10473 | TTI = &AM.getResult<TargetIRAnalysis>(IR&: F); |
10474 | DT = &AM.getResult<DominatorTreeAnalysis>(IR&: F); |
10475 | TLI = &AM.getResult<TargetLibraryAnalysis>(IR&: F); |
10476 | AC = &AM.getResult<AssumptionAnalysis>(IR&: F); |
10477 | DB = &AM.getResult<DemandedBitsAnalysis>(IR&: F); |
10478 | ORE = &AM.getResult<OptimizationRemarkEmitterAnalysis>(IR&: F); |
10479 | LAIs = &AM.getResult<LoopAccessAnalysis>(IR&: F); |
10480 | |
10481 | auto &MAMProxy = AM.getResult<ModuleAnalysisManagerFunctionProxy>(IR&: F); |
10482 | PSI = MAMProxy.getCachedResult<ProfileSummaryAnalysis>(IR&: *F.getParent()); |
10483 | BFI = nullptr; |
10484 | if (PSI && PSI->hasProfileSummary()) |
10485 | BFI = &AM.getResult<BlockFrequencyAnalysis>(IR&: F); |
10486 | LoopVectorizeResult Result = runImpl(F); |
10487 | if (!Result.MadeAnyChange) |
10488 | return PreservedAnalyses::all(); |
10489 | PreservedAnalyses PA; |
10490 | |
10491 | if (isAssignmentTrackingEnabled(M: *F.getParent())) { |
10492 | for (auto &BB : F) |
10493 | RemoveRedundantDbgInstrs(BB: &BB); |
10494 | } |
10495 | |
10496 | PA.preserve<LoopAnalysis>(); |
10497 | PA.preserve<DominatorTreeAnalysis>(); |
10498 | PA.preserve<ScalarEvolutionAnalysis>(); |
10499 | PA.preserve<LoopAccessAnalysis>(); |
10500 | |
10501 | if (Result.MadeCFGChange) { |
10502 | // Making CFG changes likely means a loop got vectorized. Indicate that |
10503 | // extra simplification passes should be run. |
10504 | // TODO: MadeCFGChanges is not a prefect proxy. Extra passes should only |
10505 | // be run if runtime checks have been added. |
10506 | AM.getResult<ShouldRunExtraVectorPasses>(IR&: F); |
10507 | PA.preserve<ShouldRunExtraVectorPasses>(); |
10508 | } else { |
10509 | PA.preserveSet<CFGAnalyses>(); |
10510 | } |
10511 | return PA; |
10512 | } |
10513 | |
10514 | void LoopVectorizePass::printPipeline( |
10515 | raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) { |
10516 | static_cast<PassInfoMixin<LoopVectorizePass> *>(this)->printPipeline( |
10517 | OS, MapClassName2PassName); |
10518 | |
10519 | OS << '<'; |
10520 | OS << (InterleaveOnlyWhenForced ? "" : "no-" ) << "interleave-forced-only;" ; |
10521 | OS << (VectorizeOnlyWhenForced ? "" : "no-" ) << "vectorize-forced-only;" ; |
10522 | OS << '>'; |
10523 | } |
10524 | |