File: | build/llvm-toolchain-snapshot-15~++20220420111733+e13d2efed663/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp |
Warning: | line 2835, column 23 Access to field 'IsScheduled' results in a dereference of a null pointer (loaded from variable 'SD') |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | //===- SLPVectorizer.cpp - A bottom up SLP 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 pass implements the Bottom Up SLP vectorizer. It detects consecutive | |||
10 | // stores that can be put together into vector-stores. Next, it attempts to | |||
11 | // construct vectorizable tree using the use-def chains. If a profitable tree | |||
12 | // was found, the SLP vectorizer performs vectorization on the tree. | |||
13 | // | |||
14 | // The pass is inspired by the work described in the paper: | |||
15 | // "Loop-Aware SLP in GCC" by Ira Rosen, Dorit Nuzman, Ayal Zaks. | |||
16 | // | |||
17 | //===----------------------------------------------------------------------===// | |||
18 | ||||
19 | #include "llvm/Transforms/Vectorize/SLPVectorizer.h" | |||
20 | #include "llvm/ADT/DenseMap.h" | |||
21 | #include "llvm/ADT/DenseSet.h" | |||
22 | #include "llvm/ADT/Optional.h" | |||
23 | #include "llvm/ADT/PostOrderIterator.h" | |||
24 | #include "llvm/ADT/PriorityQueue.h" | |||
25 | #include "llvm/ADT/STLExtras.h" | |||
26 | #include "llvm/ADT/SetOperations.h" | |||
27 | #include "llvm/ADT/SetVector.h" | |||
28 | #include "llvm/ADT/SmallBitVector.h" | |||
29 | #include "llvm/ADT/SmallPtrSet.h" | |||
30 | #include "llvm/ADT/SmallSet.h" | |||
31 | #include "llvm/ADT/SmallString.h" | |||
32 | #include "llvm/ADT/Statistic.h" | |||
33 | #include "llvm/ADT/iterator.h" | |||
34 | #include "llvm/ADT/iterator_range.h" | |||
35 | #include "llvm/Analysis/AliasAnalysis.h" | |||
36 | #include "llvm/Analysis/AssumptionCache.h" | |||
37 | #include "llvm/Analysis/CodeMetrics.h" | |||
38 | #include "llvm/Analysis/DemandedBits.h" | |||
39 | #include "llvm/Analysis/GlobalsModRef.h" | |||
40 | #include "llvm/Analysis/IVDescriptors.h" | |||
41 | #include "llvm/Analysis/LoopAccessAnalysis.h" | |||
42 | #include "llvm/Analysis/LoopInfo.h" | |||
43 | #include "llvm/Analysis/MemoryLocation.h" | |||
44 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" | |||
45 | #include "llvm/Analysis/ScalarEvolution.h" | |||
46 | #include "llvm/Analysis/ScalarEvolutionExpressions.h" | |||
47 | #include "llvm/Analysis/TargetLibraryInfo.h" | |||
48 | #include "llvm/Analysis/TargetTransformInfo.h" | |||
49 | #include "llvm/Analysis/ValueTracking.h" | |||
50 | #include "llvm/Analysis/VectorUtils.h" | |||
51 | #include "llvm/IR/Attributes.h" | |||
52 | #include "llvm/IR/BasicBlock.h" | |||
53 | #include "llvm/IR/Constant.h" | |||
54 | #include "llvm/IR/Constants.h" | |||
55 | #include "llvm/IR/DataLayout.h" | |||
56 | #include "llvm/IR/DerivedTypes.h" | |||
57 | #include "llvm/IR/Dominators.h" | |||
58 | #include "llvm/IR/Function.h" | |||
59 | #include "llvm/IR/IRBuilder.h" | |||
60 | #include "llvm/IR/InstrTypes.h" | |||
61 | #include "llvm/IR/Instruction.h" | |||
62 | #include "llvm/IR/Instructions.h" | |||
63 | #include "llvm/IR/IntrinsicInst.h" | |||
64 | #include "llvm/IR/Intrinsics.h" | |||
65 | #include "llvm/IR/Module.h" | |||
66 | #include "llvm/IR/Operator.h" | |||
67 | #include "llvm/IR/PatternMatch.h" | |||
68 | #include "llvm/IR/Type.h" | |||
69 | #include "llvm/IR/Use.h" | |||
70 | #include "llvm/IR/User.h" | |||
71 | #include "llvm/IR/Value.h" | |||
72 | #include "llvm/IR/ValueHandle.h" | |||
73 | #ifdef EXPENSIVE_CHECKS | |||
74 | #include "llvm/IR/Verifier.h" | |||
75 | #endif | |||
76 | #include "llvm/Pass.h" | |||
77 | #include "llvm/Support/Casting.h" | |||
78 | #include "llvm/Support/CommandLine.h" | |||
79 | #include "llvm/Support/Compiler.h" | |||
80 | #include "llvm/Support/DOTGraphTraits.h" | |||
81 | #include "llvm/Support/Debug.h" | |||
82 | #include "llvm/Support/ErrorHandling.h" | |||
83 | #include "llvm/Support/GraphWriter.h" | |||
84 | #include "llvm/Support/InstructionCost.h" | |||
85 | #include "llvm/Support/KnownBits.h" | |||
86 | #include "llvm/Support/MathExtras.h" | |||
87 | #include "llvm/Support/raw_ostream.h" | |||
88 | #include "llvm/Transforms/Utils/InjectTLIMappings.h" | |||
89 | #include "llvm/Transforms/Utils/Local.h" | |||
90 | #include "llvm/Transforms/Utils/LoopUtils.h" | |||
91 | #include "llvm/Transforms/Vectorize.h" | |||
92 | #include <algorithm> | |||
93 | #include <cassert> | |||
94 | #include <cstdint> | |||
95 | #include <iterator> | |||
96 | #include <memory> | |||
97 | #include <set> | |||
98 | #include <string> | |||
99 | #include <tuple> | |||
100 | #include <utility> | |||
101 | #include <vector> | |||
102 | ||||
103 | using namespace llvm; | |||
104 | using namespace llvm::PatternMatch; | |||
105 | using namespace slpvectorizer; | |||
106 | ||||
107 | #define SV_NAME"slp-vectorizer" "slp-vectorizer" | |||
108 | #define DEBUG_TYPE"SLP" "SLP" | |||
109 | ||||
110 | STATISTIC(NumVectorInstructions, "Number of vector instructions generated")static llvm::Statistic NumVectorInstructions = {"SLP", "NumVectorInstructions" , "Number of vector instructions generated"}; | |||
111 | ||||
112 | cl::opt<bool> RunSLPVectorization("vectorize-slp", cl::init(true), cl::Hidden, | |||
113 | cl::desc("Run the SLP vectorization passes")); | |||
114 | ||||
115 | static cl::opt<int> | |||
116 | SLPCostThreshold("slp-threshold", cl::init(0), cl::Hidden, | |||
117 | cl::desc("Only vectorize if you gain more than this " | |||
118 | "number ")); | |||
119 | ||||
120 | static cl::opt<bool> | |||
121 | ShouldVectorizeHor("slp-vectorize-hor", cl::init(true), cl::Hidden, | |||
122 | cl::desc("Attempt to vectorize horizontal reductions")); | |||
123 | ||||
124 | static cl::opt<bool> ShouldStartVectorizeHorAtStore( | |||
125 | "slp-vectorize-hor-store", cl::init(false), cl::Hidden, | |||
126 | cl::desc( | |||
127 | "Attempt to vectorize horizontal reductions feeding into a store")); | |||
128 | ||||
129 | static cl::opt<int> | |||
130 | MaxVectorRegSizeOption("slp-max-reg-size", cl::init(128), cl::Hidden, | |||
131 | cl::desc("Attempt to vectorize for this register size in bits")); | |||
132 | ||||
133 | static cl::opt<unsigned> | |||
134 | MaxVFOption("slp-max-vf", cl::init(0), cl::Hidden, | |||
135 | cl::desc("Maximum SLP vectorization factor (0=unlimited)")); | |||
136 | ||||
137 | static cl::opt<int> | |||
138 | MaxStoreLookup("slp-max-store-lookup", cl::init(32), cl::Hidden, | |||
139 | cl::desc("Maximum depth of the lookup for consecutive stores.")); | |||
140 | ||||
141 | /// Limits the size of scheduling regions in a block. | |||
142 | /// It avoid long compile times for _very_ large blocks where vector | |||
143 | /// instructions are spread over a wide range. | |||
144 | /// This limit is way higher than needed by real-world functions. | |||
145 | static cl::opt<int> | |||
146 | ScheduleRegionSizeBudget("slp-schedule-budget", cl::init(100000), cl::Hidden, | |||
147 | cl::desc("Limit the size of the SLP scheduling region per block")); | |||
148 | ||||
149 | static cl::opt<int> MinVectorRegSizeOption( | |||
150 | "slp-min-reg-size", cl::init(128), cl::Hidden, | |||
151 | cl::desc("Attempt to vectorize for this register size in bits")); | |||
152 | ||||
153 | static cl::opt<unsigned> RecursionMaxDepth( | |||
154 | "slp-recursion-max-depth", cl::init(12), cl::Hidden, | |||
155 | cl::desc("Limit the recursion depth when building a vectorizable tree")); | |||
156 | ||||
157 | static cl::opt<unsigned> MinTreeSize( | |||
158 | "slp-min-tree-size", cl::init(3), cl::Hidden, | |||
159 | cl::desc("Only vectorize small trees if they are fully vectorizable")); | |||
160 | ||||
161 | // The maximum depth that the look-ahead score heuristic will explore. | |||
162 | // The higher this value, the higher the compilation time overhead. | |||
163 | static cl::opt<int> LookAheadMaxDepth( | |||
164 | "slp-max-look-ahead-depth", cl::init(2), cl::Hidden, | |||
165 | cl::desc("The maximum look-ahead depth for operand reordering scores")); | |||
166 | ||||
167 | static cl::opt<bool> | |||
168 | ViewSLPTree("view-slp-tree", cl::Hidden, | |||
169 | cl::desc("Display the SLP trees with Graphviz")); | |||
170 | ||||
171 | // Limit the number of alias checks. The limit is chosen so that | |||
172 | // it has no negative effect on the llvm benchmarks. | |||
173 | static const unsigned AliasedCheckLimit = 10; | |||
174 | ||||
175 | // Another limit for the alias checks: The maximum distance between load/store | |||
176 | // instructions where alias checks are done. | |||
177 | // This limit is useful for very large basic blocks. | |||
178 | static const unsigned MaxMemDepDistance = 160; | |||
179 | ||||
180 | /// If the ScheduleRegionSizeBudget is exhausted, we allow small scheduling | |||
181 | /// regions to be handled. | |||
182 | static const int MinScheduleRegionSize = 16; | |||
183 | ||||
184 | /// Predicate for the element types that the SLP vectorizer supports. | |||
185 | /// | |||
186 | /// The most important thing to filter here are types which are invalid in LLVM | |||
187 | /// vectors. We also filter target specific types which have absolutely no | |||
188 | /// meaningful vectorization path such as x86_fp80 and ppc_f128. This just | |||
189 | /// avoids spending time checking the cost model and realizing that they will | |||
190 | /// be inevitably scalarized. | |||
191 | static bool isValidElementType(Type *Ty) { | |||
192 | return VectorType::isValidElementType(Ty) && !Ty->isX86_FP80Ty() && | |||
193 | !Ty->isPPC_FP128Ty(); | |||
194 | } | |||
195 | ||||
196 | /// \returns True if the value is a constant (but not globals/constant | |||
197 | /// expressions). | |||
198 | static bool isConstant(Value *V) { | |||
199 | return isa<Constant>(V) && !isa<ConstantExpr>(V) && !isa<GlobalValue>(V); | |||
200 | } | |||
201 | ||||
202 | /// Checks if \p V is one of vector-like instructions, i.e. undef, | |||
203 | /// insertelement/extractelement with constant indices for fixed vector type or | |||
204 | /// extractvalue instruction. | |||
205 | static bool isVectorLikeInstWithConstOps(Value *V) { | |||
206 | if (!isa<InsertElementInst, ExtractElementInst>(V) && | |||
207 | !isa<ExtractValueInst, UndefValue>(V)) | |||
208 | return false; | |||
209 | auto *I = dyn_cast<Instruction>(V); | |||
210 | if (!I || isa<ExtractValueInst>(I)) | |||
211 | return true; | |||
212 | if (!isa<FixedVectorType>(I->getOperand(0)->getType())) | |||
213 | return false; | |||
214 | if (isa<ExtractElementInst>(I)) | |||
215 | return isConstant(I->getOperand(1)); | |||
216 | assert(isa<InsertElementInst>(V) && "Expected only insertelement.")(static_cast <bool> (isa<InsertElementInst>(V) && "Expected only insertelement.") ? void (0) : __assert_fail ( "isa<InsertElementInst>(V) && \"Expected only insertelement.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 216, __extension__ __PRETTY_FUNCTION__)); | |||
217 | return isConstant(I->getOperand(2)); | |||
218 | } | |||
219 | ||||
220 | /// \returns true if all of the instructions in \p VL are in the same block or | |||
221 | /// false otherwise. | |||
222 | static bool allSameBlock(ArrayRef<Value *> VL) { | |||
223 | Instruction *I0 = dyn_cast<Instruction>(VL[0]); | |||
224 | if (!I0) | |||
225 | return false; | |||
226 | if (all_of(VL, isVectorLikeInstWithConstOps)) | |||
227 | return true; | |||
228 | ||||
229 | BasicBlock *BB = I0->getParent(); | |||
230 | for (int I = 1, E = VL.size(); I < E; I++) { | |||
231 | auto *II = dyn_cast<Instruction>(VL[I]); | |||
232 | if (!II) | |||
233 | return false; | |||
234 | ||||
235 | if (BB != II->getParent()) | |||
236 | return false; | |||
237 | } | |||
238 | return true; | |||
239 | } | |||
240 | ||||
241 | /// \returns True if all of the values in \p VL are constants (but not | |||
242 | /// globals/constant expressions). | |||
243 | static bool allConstant(ArrayRef<Value *> VL) { | |||
244 | // Constant expressions and globals can't be vectorized like normal integer/FP | |||
245 | // constants. | |||
246 | return all_of(VL, isConstant); | |||
247 | } | |||
248 | ||||
249 | /// \returns True if all of the values in \p VL are identical or some of them | |||
250 | /// are UndefValue. | |||
251 | static bool isSplat(ArrayRef<Value *> VL) { | |||
252 | Value *FirstNonUndef = nullptr; | |||
253 | for (Value *V : VL) { | |||
254 | if (isa<UndefValue>(V)) | |||
255 | continue; | |||
256 | if (!FirstNonUndef) { | |||
257 | FirstNonUndef = V; | |||
258 | continue; | |||
259 | } | |||
260 | if (V != FirstNonUndef) | |||
261 | return false; | |||
262 | } | |||
263 | return FirstNonUndef != nullptr; | |||
264 | } | |||
265 | ||||
266 | /// \returns True if \p I is commutative, handles CmpInst and BinaryOperator. | |||
267 | static bool isCommutative(Instruction *I) { | |||
268 | if (auto *Cmp = dyn_cast<CmpInst>(I)) | |||
269 | return Cmp->isCommutative(); | |||
270 | if (auto *BO = dyn_cast<BinaryOperator>(I)) | |||
271 | return BO->isCommutative(); | |||
272 | // TODO: This should check for generic Instruction::isCommutative(), but | |||
273 | // we need to confirm that the caller code correctly handles Intrinsics | |||
274 | // for example (does not have 2 operands). | |||
275 | return false; | |||
276 | } | |||
277 | ||||
278 | /// Checks if the given value is actually an undefined constant vector. | |||
279 | static bool isUndefVector(const Value *V) { | |||
280 | if (isa<UndefValue>(V)) | |||
281 | return true; | |||
282 | auto *C = dyn_cast<Constant>(V); | |||
283 | if (!C) | |||
284 | return false; | |||
285 | if (!C->containsUndefOrPoisonElement()) | |||
286 | return false; | |||
287 | auto *VecTy = dyn_cast<FixedVectorType>(C->getType()); | |||
288 | if (!VecTy) | |||
289 | return false; | |||
290 | for (unsigned I = 0, E = VecTy->getNumElements(); I != E; ++I) { | |||
291 | if (Constant *Elem = C->getAggregateElement(I)) | |||
292 | if (!isa<UndefValue>(Elem)) | |||
293 | return false; | |||
294 | } | |||
295 | return true; | |||
296 | } | |||
297 | ||||
298 | /// Checks if the vector of instructions can be represented as a shuffle, like: | |||
299 | /// %x0 = extractelement <4 x i8> %x, i32 0 | |||
300 | /// %x3 = extractelement <4 x i8> %x, i32 3 | |||
301 | /// %y1 = extractelement <4 x i8> %y, i32 1 | |||
302 | /// %y2 = extractelement <4 x i8> %y, i32 2 | |||
303 | /// %x0x0 = mul i8 %x0, %x0 | |||
304 | /// %x3x3 = mul i8 %x3, %x3 | |||
305 | /// %y1y1 = mul i8 %y1, %y1 | |||
306 | /// %y2y2 = mul i8 %y2, %y2 | |||
307 | /// %ins1 = insertelement <4 x i8> poison, i8 %x0x0, i32 0 | |||
308 | /// %ins2 = insertelement <4 x i8> %ins1, i8 %x3x3, i32 1 | |||
309 | /// %ins3 = insertelement <4 x i8> %ins2, i8 %y1y1, i32 2 | |||
310 | /// %ins4 = insertelement <4 x i8> %ins3, i8 %y2y2, i32 3 | |||
311 | /// ret <4 x i8> %ins4 | |||
312 | /// can be transformed into: | |||
313 | /// %1 = shufflevector <4 x i8> %x, <4 x i8> %y, <4 x i32> <i32 0, i32 3, i32 5, | |||
314 | /// i32 6> | |||
315 | /// %2 = mul <4 x i8> %1, %1 | |||
316 | /// ret <4 x i8> %2 | |||
317 | /// We convert this initially to something like: | |||
318 | /// %x0 = extractelement <4 x i8> %x, i32 0 | |||
319 | /// %x3 = extractelement <4 x i8> %x, i32 3 | |||
320 | /// %y1 = extractelement <4 x i8> %y, i32 1 | |||
321 | /// %y2 = extractelement <4 x i8> %y, i32 2 | |||
322 | /// %1 = insertelement <4 x i8> poison, i8 %x0, i32 0 | |||
323 | /// %2 = insertelement <4 x i8> %1, i8 %x3, i32 1 | |||
324 | /// %3 = insertelement <4 x i8> %2, i8 %y1, i32 2 | |||
325 | /// %4 = insertelement <4 x i8> %3, i8 %y2, i32 3 | |||
326 | /// %5 = mul <4 x i8> %4, %4 | |||
327 | /// %6 = extractelement <4 x i8> %5, i32 0 | |||
328 | /// %ins1 = insertelement <4 x i8> poison, i8 %6, i32 0 | |||
329 | /// %7 = extractelement <4 x i8> %5, i32 1 | |||
330 | /// %ins2 = insertelement <4 x i8> %ins1, i8 %7, i32 1 | |||
331 | /// %8 = extractelement <4 x i8> %5, i32 2 | |||
332 | /// %ins3 = insertelement <4 x i8> %ins2, i8 %8, i32 2 | |||
333 | /// %9 = extractelement <4 x i8> %5, i32 3 | |||
334 | /// %ins4 = insertelement <4 x i8> %ins3, i8 %9, i32 3 | |||
335 | /// ret <4 x i8> %ins4 | |||
336 | /// InstCombiner transforms this into a shuffle and vector mul | |||
337 | /// Mask will return the Shuffle Mask equivalent to the extracted elements. | |||
338 | /// TODO: Can we split off and reuse the shuffle mask detection from | |||
339 | /// TargetTransformInfo::getInstructionThroughput? | |||
340 | static Optional<TargetTransformInfo::ShuffleKind> | |||
341 | isFixedVectorShuffle(ArrayRef<Value *> VL, SmallVectorImpl<int> &Mask) { | |||
342 | const auto *It = | |||
343 | find_if(VL, [](Value *V) { return isa<ExtractElementInst>(V); }); | |||
344 | if (It == VL.end()) | |||
345 | return None; | |||
346 | auto *EI0 = cast<ExtractElementInst>(*It); | |||
347 | if (isa<ScalableVectorType>(EI0->getVectorOperandType())) | |||
348 | return None; | |||
349 | unsigned Size = | |||
350 | cast<FixedVectorType>(EI0->getVectorOperandType())->getNumElements(); | |||
351 | Value *Vec1 = nullptr; | |||
352 | Value *Vec2 = nullptr; | |||
353 | enum ShuffleMode { Unknown, Select, Permute }; | |||
354 | ShuffleMode CommonShuffleMode = Unknown; | |||
355 | Mask.assign(VL.size(), UndefMaskElem); | |||
356 | for (unsigned I = 0, E = VL.size(); I < E; ++I) { | |||
357 | // Undef can be represented as an undef element in a vector. | |||
358 | if (isa<UndefValue>(VL[I])) | |||
359 | continue; | |||
360 | auto *EI = cast<ExtractElementInst>(VL[I]); | |||
361 | if (isa<ScalableVectorType>(EI->getVectorOperandType())) | |||
362 | return None; | |||
363 | auto *Vec = EI->getVectorOperand(); | |||
364 | // We can extractelement from undef or poison vector. | |||
365 | if (isUndefVector(Vec)) | |||
366 | continue; | |||
367 | // All vector operands must have the same number of vector elements. | |||
368 | if (cast<FixedVectorType>(Vec->getType())->getNumElements() != Size) | |||
369 | return None; | |||
370 | if (isa<UndefValue>(EI->getIndexOperand())) | |||
371 | continue; | |||
372 | auto *Idx = dyn_cast<ConstantInt>(EI->getIndexOperand()); | |||
373 | if (!Idx) | |||
374 | return None; | |||
375 | // Undefined behavior if Idx is negative or >= Size. | |||
376 | if (Idx->getValue().uge(Size)) | |||
377 | continue; | |||
378 | unsigned IntIdx = Idx->getValue().getZExtValue(); | |||
379 | Mask[I] = IntIdx; | |||
380 | // For correct shuffling we have to have at most 2 different vector operands | |||
381 | // in all extractelement instructions. | |||
382 | if (!Vec1 || Vec1 == Vec) { | |||
383 | Vec1 = Vec; | |||
384 | } else if (!Vec2 || Vec2 == Vec) { | |||
385 | Vec2 = Vec; | |||
386 | Mask[I] += Size; | |||
387 | } else { | |||
388 | return None; | |||
389 | } | |||
390 | if (CommonShuffleMode == Permute) | |||
391 | continue; | |||
392 | // If the extract index is not the same as the operation number, it is a | |||
393 | // permutation. | |||
394 | if (IntIdx != I) { | |||
395 | CommonShuffleMode = Permute; | |||
396 | continue; | |||
397 | } | |||
398 | CommonShuffleMode = Select; | |||
399 | } | |||
400 | // If we're not crossing lanes in different vectors, consider it as blending. | |||
401 | if (CommonShuffleMode == Select && Vec2) | |||
402 | return TargetTransformInfo::SK_Select; | |||
403 | // If Vec2 was never used, we have a permutation of a single vector, otherwise | |||
404 | // we have permutation of 2 vectors. | |||
405 | return Vec2 ? TargetTransformInfo::SK_PermuteTwoSrc | |||
406 | : TargetTransformInfo::SK_PermuteSingleSrc; | |||
407 | } | |||
408 | ||||
409 | namespace { | |||
410 | ||||
411 | /// Main data required for vectorization of instructions. | |||
412 | struct InstructionsState { | |||
413 | /// The very first instruction in the list with the main opcode. | |||
414 | Value *OpValue = nullptr; | |||
415 | ||||
416 | /// The main/alternate instruction. | |||
417 | Instruction *MainOp = nullptr; | |||
418 | Instruction *AltOp = nullptr; | |||
419 | ||||
420 | /// The main/alternate opcodes for the list of instructions. | |||
421 | unsigned getOpcode() const { | |||
422 | return MainOp ? MainOp->getOpcode() : 0; | |||
423 | } | |||
424 | ||||
425 | unsigned getAltOpcode() const { | |||
426 | return AltOp ? AltOp->getOpcode() : 0; | |||
427 | } | |||
428 | ||||
429 | /// Some of the instructions in the list have alternate opcodes. | |||
430 | bool isAltShuffle() const { return AltOp != MainOp; } | |||
431 | ||||
432 | bool isOpcodeOrAlt(Instruction *I) const { | |||
433 | unsigned CheckedOpcode = I->getOpcode(); | |||
434 | return getOpcode() == CheckedOpcode || getAltOpcode() == CheckedOpcode; | |||
435 | } | |||
436 | ||||
437 | InstructionsState() = delete; | |||
438 | InstructionsState(Value *OpValue, Instruction *MainOp, Instruction *AltOp) | |||
439 | : OpValue(OpValue), MainOp(MainOp), AltOp(AltOp) {} | |||
440 | }; | |||
441 | ||||
442 | } // end anonymous namespace | |||
443 | ||||
444 | /// Chooses the correct key for scheduling data. If \p Op has the same (or | |||
445 | /// alternate) opcode as \p OpValue, the key is \p Op. Otherwise the key is \p | |||
446 | /// OpValue. | |||
447 | static Value *isOneOf(const InstructionsState &S, Value *Op) { | |||
448 | auto *I = dyn_cast<Instruction>(Op); | |||
449 | if (I && S.isOpcodeOrAlt(I)) | |||
450 | return Op; | |||
451 | return S.OpValue; | |||
452 | } | |||
453 | ||||
454 | /// \returns true if \p Opcode is allowed as part of of the main/alternate | |||
455 | /// instruction for SLP vectorization. | |||
456 | /// | |||
457 | /// Example of unsupported opcode is SDIV that can potentially cause UB if the | |||
458 | /// "shuffled out" lane would result in division by zero. | |||
459 | static bool isValidForAlternation(unsigned Opcode) { | |||
460 | if (Instruction::isIntDivRem(Opcode)) | |||
461 | return false; | |||
462 | ||||
463 | return true; | |||
464 | } | |||
465 | ||||
466 | static InstructionsState getSameOpcode(ArrayRef<Value *> VL, | |||
467 | unsigned BaseIndex = 0); | |||
468 | ||||
469 | /// Checks if the provided operands of 2 cmp instructions are compatible, i.e. | |||
470 | /// compatible instructions or constants, or just some other regular values. | |||
471 | static bool areCompatibleCmpOps(Value *BaseOp0, Value *BaseOp1, Value *Op0, | |||
472 | Value *Op1) { | |||
473 | return (isConstant(BaseOp0) && isConstant(Op0)) || | |||
474 | (isConstant(BaseOp1) && isConstant(Op1)) || | |||
475 | (!isa<Instruction>(BaseOp0) && !isa<Instruction>(Op0) && | |||
476 | !isa<Instruction>(BaseOp1) && !isa<Instruction>(Op1)) || | |||
477 | getSameOpcode({BaseOp0, Op0}).getOpcode() || | |||
478 | getSameOpcode({BaseOp1, Op1}).getOpcode(); | |||
479 | } | |||
480 | ||||
481 | /// \returns analysis of the Instructions in \p VL described in | |||
482 | /// InstructionsState, the Opcode that we suppose the whole list | |||
483 | /// could be vectorized even if its structure is diverse. | |||
484 | static InstructionsState getSameOpcode(ArrayRef<Value *> VL, | |||
485 | unsigned BaseIndex) { | |||
486 | // Make sure these are all Instructions. | |||
487 | if (llvm::any_of(VL, [](Value *V) { return !isa<Instruction>(V); })) | |||
488 | return InstructionsState(VL[BaseIndex], nullptr, nullptr); | |||
489 | ||||
490 | bool IsCastOp = isa<CastInst>(VL[BaseIndex]); | |||
491 | bool IsBinOp = isa<BinaryOperator>(VL[BaseIndex]); | |||
492 | bool IsCmpOp = isa<CmpInst>(VL[BaseIndex]); | |||
493 | CmpInst::Predicate BasePred = | |||
494 | IsCmpOp ? cast<CmpInst>(VL[BaseIndex])->getPredicate() | |||
495 | : CmpInst::BAD_ICMP_PREDICATE; | |||
496 | unsigned Opcode = cast<Instruction>(VL[BaseIndex])->getOpcode(); | |||
497 | unsigned AltOpcode = Opcode; | |||
498 | unsigned AltIndex = BaseIndex; | |||
499 | ||||
500 | // Check for one alternate opcode from another BinaryOperator. | |||
501 | // TODO - generalize to support all operators (types, calls etc.). | |||
502 | for (int Cnt = 0, E = VL.size(); Cnt < E; Cnt++) { | |||
503 | unsigned InstOpcode = cast<Instruction>(VL[Cnt])->getOpcode(); | |||
504 | if (IsBinOp && isa<BinaryOperator>(VL[Cnt])) { | |||
505 | if (InstOpcode == Opcode || InstOpcode == AltOpcode) | |||
506 | continue; | |||
507 | if (Opcode == AltOpcode && isValidForAlternation(InstOpcode) && | |||
508 | isValidForAlternation(Opcode)) { | |||
509 | AltOpcode = InstOpcode; | |||
510 | AltIndex = Cnt; | |||
511 | continue; | |||
512 | } | |||
513 | } else if (IsCastOp && isa<CastInst>(VL[Cnt])) { | |||
514 | Type *Ty0 = cast<Instruction>(VL[BaseIndex])->getOperand(0)->getType(); | |||
515 | Type *Ty1 = cast<Instruction>(VL[Cnt])->getOperand(0)->getType(); | |||
516 | if (Ty0 == Ty1) { | |||
517 | if (InstOpcode == Opcode || InstOpcode == AltOpcode) | |||
518 | continue; | |||
519 | if (Opcode == AltOpcode) { | |||
520 | assert(isValidForAlternation(Opcode) &&(static_cast <bool> (isValidForAlternation(Opcode) && isValidForAlternation(InstOpcode) && "Cast isn't safe for alternation, logic needs to be updated!" ) ? void (0) : __assert_fail ("isValidForAlternation(Opcode) && isValidForAlternation(InstOpcode) && \"Cast isn't safe for alternation, logic needs to be updated!\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 522, __extension__ __PRETTY_FUNCTION__)) | |||
521 | isValidForAlternation(InstOpcode) &&(static_cast <bool> (isValidForAlternation(Opcode) && isValidForAlternation(InstOpcode) && "Cast isn't safe for alternation, logic needs to be updated!" ) ? void (0) : __assert_fail ("isValidForAlternation(Opcode) && isValidForAlternation(InstOpcode) && \"Cast isn't safe for alternation, logic needs to be updated!\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 522, __extension__ __PRETTY_FUNCTION__)) | |||
522 | "Cast isn't safe for alternation, logic needs to be updated!")(static_cast <bool> (isValidForAlternation(Opcode) && isValidForAlternation(InstOpcode) && "Cast isn't safe for alternation, logic needs to be updated!" ) ? void (0) : __assert_fail ("isValidForAlternation(Opcode) && isValidForAlternation(InstOpcode) && \"Cast isn't safe for alternation, logic needs to be updated!\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 522, __extension__ __PRETTY_FUNCTION__)); | |||
523 | AltOpcode = InstOpcode; | |||
524 | AltIndex = Cnt; | |||
525 | continue; | |||
526 | } | |||
527 | } | |||
528 | } else if (IsCmpOp && isa<CmpInst>(VL[Cnt])) { | |||
529 | auto *BaseInst = cast<Instruction>(VL[BaseIndex]); | |||
530 | auto *Inst = cast<Instruction>(VL[Cnt]); | |||
531 | Type *Ty0 = BaseInst->getOperand(0)->getType(); | |||
532 | Type *Ty1 = Inst->getOperand(0)->getType(); | |||
533 | if (Ty0 == Ty1) { | |||
534 | Value *BaseOp0 = BaseInst->getOperand(0); | |||
535 | Value *BaseOp1 = BaseInst->getOperand(1); | |||
536 | Value *Op0 = Inst->getOperand(0); | |||
537 | Value *Op1 = Inst->getOperand(1); | |||
538 | CmpInst::Predicate CurrentPred = | |||
539 | cast<CmpInst>(VL[Cnt])->getPredicate(); | |||
540 | CmpInst::Predicate SwappedCurrentPred = | |||
541 | CmpInst::getSwappedPredicate(CurrentPred); | |||
542 | // Check for compatible operands. If the corresponding operands are not | |||
543 | // compatible - need to perform alternate vectorization. | |||
544 | if (InstOpcode == Opcode) { | |||
545 | if (BasePred == CurrentPred && | |||
546 | areCompatibleCmpOps(BaseOp0, BaseOp1, Op0, Op1)) | |||
547 | continue; | |||
548 | if (BasePred == SwappedCurrentPred && | |||
549 | areCompatibleCmpOps(BaseOp0, BaseOp1, Op1, Op0)) | |||
550 | continue; | |||
551 | if (E == 2 && | |||
552 | (BasePred == CurrentPred || BasePred == SwappedCurrentPred)) | |||
553 | continue; | |||
554 | auto *AltInst = cast<CmpInst>(VL[AltIndex]); | |||
555 | CmpInst::Predicate AltPred = AltInst->getPredicate(); | |||
556 | Value *AltOp0 = AltInst->getOperand(0); | |||
557 | Value *AltOp1 = AltInst->getOperand(1); | |||
558 | // Check if operands are compatible with alternate operands. | |||
559 | if (AltPred == CurrentPred && | |||
560 | areCompatibleCmpOps(AltOp0, AltOp1, Op0, Op1)) | |||
561 | continue; | |||
562 | if (AltPred == SwappedCurrentPred && | |||
563 | areCompatibleCmpOps(AltOp0, AltOp1, Op1, Op0)) | |||
564 | continue; | |||
565 | } | |||
566 | if (BaseIndex == AltIndex && BasePred != CurrentPred) { | |||
567 | assert(isValidForAlternation(Opcode) &&(static_cast <bool> (isValidForAlternation(Opcode) && isValidForAlternation(InstOpcode) && "Cast isn't safe for alternation, logic needs to be updated!" ) ? void (0) : __assert_fail ("isValidForAlternation(Opcode) && isValidForAlternation(InstOpcode) && \"Cast isn't safe for alternation, logic needs to be updated!\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 569, __extension__ __PRETTY_FUNCTION__)) | |||
568 | isValidForAlternation(InstOpcode) &&(static_cast <bool> (isValidForAlternation(Opcode) && isValidForAlternation(InstOpcode) && "Cast isn't safe for alternation, logic needs to be updated!" ) ? void (0) : __assert_fail ("isValidForAlternation(Opcode) && isValidForAlternation(InstOpcode) && \"Cast isn't safe for alternation, logic needs to be updated!\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 569, __extension__ __PRETTY_FUNCTION__)) | |||
569 | "Cast isn't safe for alternation, logic needs to be updated!")(static_cast <bool> (isValidForAlternation(Opcode) && isValidForAlternation(InstOpcode) && "Cast isn't safe for alternation, logic needs to be updated!" ) ? void (0) : __assert_fail ("isValidForAlternation(Opcode) && isValidForAlternation(InstOpcode) && \"Cast isn't safe for alternation, logic needs to be updated!\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 569, __extension__ __PRETTY_FUNCTION__)); | |||
570 | AltIndex = Cnt; | |||
571 | continue; | |||
572 | } | |||
573 | auto *AltInst = cast<CmpInst>(VL[AltIndex]); | |||
574 | CmpInst::Predicate AltPred = AltInst->getPredicate(); | |||
575 | if (BasePred == CurrentPred || BasePred == SwappedCurrentPred || | |||
576 | AltPred == CurrentPred || AltPred == SwappedCurrentPred) | |||
577 | continue; | |||
578 | } | |||
579 | } else if (InstOpcode == Opcode || InstOpcode == AltOpcode) | |||
580 | continue; | |||
581 | return InstructionsState(VL[BaseIndex], nullptr, nullptr); | |||
582 | } | |||
583 | ||||
584 | return InstructionsState(VL[BaseIndex], cast<Instruction>(VL[BaseIndex]), | |||
585 | cast<Instruction>(VL[AltIndex])); | |||
586 | } | |||
587 | ||||
588 | /// \returns true if all of the values in \p VL have the same type or false | |||
589 | /// otherwise. | |||
590 | static bool allSameType(ArrayRef<Value *> VL) { | |||
591 | Type *Ty = VL[0]->getType(); | |||
592 | for (int i = 1, e = VL.size(); i < e; i++) | |||
593 | if (VL[i]->getType() != Ty) | |||
594 | return false; | |||
595 | ||||
596 | return true; | |||
597 | } | |||
598 | ||||
599 | /// \returns True if Extract{Value,Element} instruction extracts element Idx. | |||
600 | static Optional<unsigned> getExtractIndex(Instruction *E) { | |||
601 | unsigned Opcode = E->getOpcode(); | |||
602 | assert((Opcode == Instruction::ExtractElement ||(static_cast <bool> ((Opcode == Instruction::ExtractElement || Opcode == Instruction::ExtractValue) && "Expected extractelement or extractvalue instruction." ) ? void (0) : __assert_fail ("(Opcode == Instruction::ExtractElement || Opcode == Instruction::ExtractValue) && \"Expected extractelement or extractvalue instruction.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 604, __extension__ __PRETTY_FUNCTION__)) | |||
603 | Opcode == Instruction::ExtractValue) &&(static_cast <bool> ((Opcode == Instruction::ExtractElement || Opcode == Instruction::ExtractValue) && "Expected extractelement or extractvalue instruction." ) ? void (0) : __assert_fail ("(Opcode == Instruction::ExtractElement || Opcode == Instruction::ExtractValue) && \"Expected extractelement or extractvalue instruction.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 604, __extension__ __PRETTY_FUNCTION__)) | |||
604 | "Expected extractelement or extractvalue instruction.")(static_cast <bool> ((Opcode == Instruction::ExtractElement || Opcode == Instruction::ExtractValue) && "Expected extractelement or extractvalue instruction." ) ? void (0) : __assert_fail ("(Opcode == Instruction::ExtractElement || Opcode == Instruction::ExtractValue) && \"Expected extractelement or extractvalue instruction.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 604, __extension__ __PRETTY_FUNCTION__)); | |||
605 | if (Opcode == Instruction::ExtractElement) { | |||
606 | auto *CI = dyn_cast<ConstantInt>(E->getOperand(1)); | |||
607 | if (!CI) | |||
608 | return None; | |||
609 | return CI->getZExtValue(); | |||
610 | } | |||
611 | ExtractValueInst *EI = cast<ExtractValueInst>(E); | |||
612 | if (EI->getNumIndices() != 1) | |||
613 | return None; | |||
614 | return *EI->idx_begin(); | |||
615 | } | |||
616 | ||||
617 | /// \returns True if in-tree use also needs extract. This refers to | |||
618 | /// possible scalar operand in vectorized instruction. | |||
619 | static bool InTreeUserNeedToExtract(Value *Scalar, Instruction *UserInst, | |||
620 | TargetLibraryInfo *TLI) { | |||
621 | unsigned Opcode = UserInst->getOpcode(); | |||
622 | switch (Opcode) { | |||
623 | case Instruction::Load: { | |||
624 | LoadInst *LI = cast<LoadInst>(UserInst); | |||
625 | return (LI->getPointerOperand() == Scalar); | |||
626 | } | |||
627 | case Instruction::Store: { | |||
628 | StoreInst *SI = cast<StoreInst>(UserInst); | |||
629 | return (SI->getPointerOperand() == Scalar); | |||
630 | } | |||
631 | case Instruction::Call: { | |||
632 | CallInst *CI = cast<CallInst>(UserInst); | |||
633 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); | |||
634 | for (unsigned i = 0, e = CI->arg_size(); i != e; ++i) { | |||
635 | if (hasVectorInstrinsicScalarOpd(ID, i)) | |||
636 | return (CI->getArgOperand(i) == Scalar); | |||
637 | } | |||
638 | LLVM_FALLTHROUGH[[gnu::fallthrough]]; | |||
639 | } | |||
640 | default: | |||
641 | return false; | |||
642 | } | |||
643 | } | |||
644 | ||||
645 | /// \returns the AA location that is being access by the instruction. | |||
646 | static MemoryLocation getLocation(Instruction *I) { | |||
647 | if (StoreInst *SI = dyn_cast<StoreInst>(I)) | |||
648 | return MemoryLocation::get(SI); | |||
649 | if (LoadInst *LI = dyn_cast<LoadInst>(I)) | |||
650 | return MemoryLocation::get(LI); | |||
651 | return MemoryLocation(); | |||
652 | } | |||
653 | ||||
654 | /// \returns True if the instruction is not a volatile or atomic load/store. | |||
655 | static bool isSimple(Instruction *I) { | |||
656 | if (LoadInst *LI = dyn_cast<LoadInst>(I)) | |||
657 | return LI->isSimple(); | |||
658 | if (StoreInst *SI = dyn_cast<StoreInst>(I)) | |||
659 | return SI->isSimple(); | |||
660 | if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I)) | |||
661 | return !MI->isVolatile(); | |||
662 | return true; | |||
663 | } | |||
664 | ||||
665 | /// Shuffles \p Mask in accordance with the given \p SubMask. | |||
666 | static void addMask(SmallVectorImpl<int> &Mask, ArrayRef<int> SubMask) { | |||
667 | if (SubMask.empty()) | |||
668 | return; | |||
669 | if (Mask.empty()) { | |||
670 | Mask.append(SubMask.begin(), SubMask.end()); | |||
671 | return; | |||
672 | } | |||
673 | SmallVector<int> NewMask(SubMask.size(), UndefMaskElem); | |||
674 | int TermValue = std::min(Mask.size(), SubMask.size()); | |||
675 | for (int I = 0, E = SubMask.size(); I < E; ++I) { | |||
676 | if (SubMask[I] >= TermValue || SubMask[I] == UndefMaskElem || | |||
677 | Mask[SubMask[I]] >= TermValue) | |||
678 | continue; | |||
679 | NewMask[I] = Mask[SubMask[I]]; | |||
680 | } | |||
681 | Mask.swap(NewMask); | |||
682 | } | |||
683 | ||||
684 | /// Order may have elements assigned special value (size) which is out of | |||
685 | /// bounds. Such indices only appear on places which correspond to undef values | |||
686 | /// (see canReuseExtract for details) and used in order to avoid undef values | |||
687 | /// have effect on operands ordering. | |||
688 | /// The first loop below simply finds all unused indices and then the next loop | |||
689 | /// nest assigns these indices for undef values positions. | |||
690 | /// As an example below Order has two undef positions and they have assigned | |||
691 | /// values 3 and 7 respectively: | |||
692 | /// before: 6 9 5 4 9 2 1 0 | |||
693 | /// after: 6 3 5 4 7 2 1 0 | |||
694 | static void fixupOrderingIndices(SmallVectorImpl<unsigned> &Order) { | |||
695 | const unsigned Sz = Order.size(); | |||
696 | SmallBitVector UnusedIndices(Sz, /*t=*/true); | |||
697 | SmallBitVector MaskedIndices(Sz); | |||
698 | for (unsigned I = 0; I < Sz; ++I) { | |||
699 | if (Order[I] < Sz) | |||
700 | UnusedIndices.reset(Order[I]); | |||
701 | else | |||
702 | MaskedIndices.set(I); | |||
703 | } | |||
704 | if (MaskedIndices.none()) | |||
705 | return; | |||
706 | assert(UnusedIndices.count() == MaskedIndices.count() &&(static_cast <bool> (UnusedIndices.count() == MaskedIndices .count() && "Non-synced masked/available indices.") ? void (0) : __assert_fail ("UnusedIndices.count() == MaskedIndices.count() && \"Non-synced masked/available indices.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 707, __extension__ __PRETTY_FUNCTION__)) | |||
707 | "Non-synced masked/available indices.")(static_cast <bool> (UnusedIndices.count() == MaskedIndices .count() && "Non-synced masked/available indices.") ? void (0) : __assert_fail ("UnusedIndices.count() == MaskedIndices.count() && \"Non-synced masked/available indices.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 707, __extension__ __PRETTY_FUNCTION__)); | |||
708 | int Idx = UnusedIndices.find_first(); | |||
709 | int MIdx = MaskedIndices.find_first(); | |||
710 | while (MIdx >= 0) { | |||
711 | assert(Idx >= 0 && "Indices must be synced.")(static_cast <bool> (Idx >= 0 && "Indices must be synced." ) ? void (0) : __assert_fail ("Idx >= 0 && \"Indices must be synced.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 711, __extension__ __PRETTY_FUNCTION__)); | |||
712 | Order[MIdx] = Idx; | |||
713 | Idx = UnusedIndices.find_next(Idx); | |||
714 | MIdx = MaskedIndices.find_next(MIdx); | |||
715 | } | |||
716 | } | |||
717 | ||||
718 | namespace llvm { | |||
719 | ||||
720 | static void inversePermutation(ArrayRef<unsigned> Indices, | |||
721 | SmallVectorImpl<int> &Mask) { | |||
722 | Mask.clear(); | |||
723 | const unsigned E = Indices.size(); | |||
724 | Mask.resize(E, UndefMaskElem); | |||
725 | for (unsigned I = 0; I < E; ++I) | |||
726 | Mask[Indices[I]] = I; | |||
727 | } | |||
728 | ||||
729 | /// \returns inserting index of InsertElement or InsertValue instruction, | |||
730 | /// using Offset as base offset for index. | |||
731 | static Optional<unsigned> getInsertIndex(Value *InsertInst, | |||
732 | unsigned Offset = 0) { | |||
733 | int Index = Offset; | |||
734 | if (auto *IE = dyn_cast<InsertElementInst>(InsertInst)) { | |||
735 | if (auto *CI = dyn_cast<ConstantInt>(IE->getOperand(2))) { | |||
736 | auto *VT = cast<FixedVectorType>(IE->getType()); | |||
737 | if (CI->getValue().uge(VT->getNumElements())) | |||
738 | return None; | |||
739 | Index *= VT->getNumElements(); | |||
740 | Index += CI->getZExtValue(); | |||
741 | return Index; | |||
742 | } | |||
743 | return None; | |||
744 | } | |||
745 | ||||
746 | auto *IV = cast<InsertValueInst>(InsertInst); | |||
747 | Type *CurrentType = IV->getType(); | |||
748 | for (unsigned I : IV->indices()) { | |||
749 | if (auto *ST = dyn_cast<StructType>(CurrentType)) { | |||
750 | Index *= ST->getNumElements(); | |||
751 | CurrentType = ST->getElementType(I); | |||
752 | } else if (auto *AT = dyn_cast<ArrayType>(CurrentType)) { | |||
753 | Index *= AT->getNumElements(); | |||
754 | CurrentType = AT->getElementType(); | |||
755 | } else { | |||
756 | return None; | |||
757 | } | |||
758 | Index += I; | |||
759 | } | |||
760 | return Index; | |||
761 | } | |||
762 | ||||
763 | /// Reorders the list of scalars in accordance with the given \p Mask. | |||
764 | static void reorderScalars(SmallVectorImpl<Value *> &Scalars, | |||
765 | ArrayRef<int> Mask) { | |||
766 | assert(!Mask.empty() && "Expected non-empty mask.")(static_cast <bool> (!Mask.empty() && "Expected non-empty mask." ) ? void (0) : __assert_fail ("!Mask.empty() && \"Expected non-empty mask.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 766, __extension__ __PRETTY_FUNCTION__)); | |||
767 | SmallVector<Value *> Prev(Scalars.size(), | |||
768 | UndefValue::get(Scalars.front()->getType())); | |||
769 | Prev.swap(Scalars); | |||
770 | for (unsigned I = 0, E = Prev.size(); I < E; ++I) | |||
771 | if (Mask[I] != UndefMaskElem) | |||
772 | Scalars[Mask[I]] = Prev[I]; | |||
773 | } | |||
774 | ||||
775 | /// Checks if the provided value does not require scheduling. It does not | |||
776 | /// require scheduling if this is not an instruction or it is an instruction | |||
777 | /// that does not read/write memory and all operands are either not instructions | |||
778 | /// or phi nodes or instructions from different blocks. | |||
779 | static bool areAllOperandsNonInsts(Value *V) { | |||
780 | auto *I = dyn_cast<Instruction>(V); | |||
781 | if (!I) | |||
782 | return true; | |||
783 | return !mayHaveNonDefUseDependency(*I) && | |||
784 | all_of(I->operands(), [I](Value *V) { | |||
785 | auto *IO = dyn_cast<Instruction>(V); | |||
786 | if (!IO) | |||
787 | return true; | |||
788 | return isa<PHINode>(IO) || IO->getParent() != I->getParent(); | |||
789 | }); | |||
790 | } | |||
791 | ||||
792 | /// Checks if the provided value does not require scheduling. It does not | |||
793 | /// require scheduling if this is not an instruction or it is an instruction | |||
794 | /// that does not read/write memory and all users are phi nodes or instructions | |||
795 | /// from the different blocks. | |||
796 | static bool isUsedOutsideBlock(Value *V) { | |||
797 | auto *I = dyn_cast<Instruction>(V); | |||
798 | if (!I) | |||
799 | return true; | |||
800 | // Limits the number of uses to save compile time. | |||
801 | constexpr int UsesLimit = 8; | |||
802 | return !I->mayReadOrWriteMemory() && !I->hasNUsesOrMore(UsesLimit) && | |||
803 | all_of(I->users(), [I](User *U) { | |||
804 | auto *IU = dyn_cast<Instruction>(U); | |||
805 | if (!IU) | |||
806 | return true; | |||
807 | return IU->getParent() != I->getParent() || isa<PHINode>(IU); | |||
808 | }); | |||
809 | } | |||
810 | ||||
811 | /// Checks if the specified value does not require scheduling. It does not | |||
812 | /// require scheduling if all operands and all users do not need to be scheduled | |||
813 | /// in the current basic block. | |||
814 | static bool doesNotNeedToBeScheduled(Value *V) { | |||
815 | return areAllOperandsNonInsts(V) && isUsedOutsideBlock(V); | |||
816 | } | |||
817 | ||||
818 | /// Checks if the specified array of instructions does not require scheduling. | |||
819 | /// It is so if all either instructions have operands that do not require | |||
820 | /// scheduling or their users do not require scheduling since they are phis or | |||
821 | /// in other basic blocks. | |||
822 | static bool doesNotNeedToSchedule(ArrayRef<Value *> VL) { | |||
823 | return !VL.empty() && | |||
824 | (all_of(VL, isUsedOutsideBlock) || all_of(VL, areAllOperandsNonInsts)); | |||
825 | } | |||
826 | ||||
827 | namespace slpvectorizer { | |||
828 | ||||
829 | /// Bottom Up SLP Vectorizer. | |||
830 | class BoUpSLP { | |||
831 | struct TreeEntry; | |||
832 | struct ScheduleData; | |||
833 | ||||
834 | public: | |||
835 | using ValueList = SmallVector<Value *, 8>; | |||
836 | using InstrList = SmallVector<Instruction *, 16>; | |||
837 | using ValueSet = SmallPtrSet<Value *, 16>; | |||
838 | using StoreList = SmallVector<StoreInst *, 8>; | |||
839 | using ExtraValueToDebugLocsMap = | |||
840 | MapVector<Value *, SmallVector<Instruction *, 2>>; | |||
841 | using OrdersType = SmallVector<unsigned, 4>; | |||
842 | ||||
843 | BoUpSLP(Function *Func, ScalarEvolution *Se, TargetTransformInfo *Tti, | |||
844 | TargetLibraryInfo *TLi, AAResults *Aa, LoopInfo *Li, | |||
845 | DominatorTree *Dt, AssumptionCache *AC, DemandedBits *DB, | |||
846 | const DataLayout *DL, OptimizationRemarkEmitter *ORE) | |||
847 | : BatchAA(*Aa), F(Func), SE(Se), TTI(Tti), TLI(TLi), LI(Li), | |||
848 | DT(Dt), AC(AC), DB(DB), DL(DL), ORE(ORE), Builder(Se->getContext()) { | |||
849 | CodeMetrics::collectEphemeralValues(F, AC, EphValues); | |||
850 | // Use the vector register size specified by the target unless overridden | |||
851 | // by a command-line option. | |||
852 | // TODO: It would be better to limit the vectorization factor based on | |||
853 | // data type rather than just register size. For example, x86 AVX has | |||
854 | // 256-bit registers, but it does not support integer operations | |||
855 | // at that width (that requires AVX2). | |||
856 | if (MaxVectorRegSizeOption.getNumOccurrences()) | |||
857 | MaxVecRegSize = MaxVectorRegSizeOption; | |||
858 | else | |||
859 | MaxVecRegSize = | |||
860 | TTI->getRegisterBitWidth(TargetTransformInfo::RGK_FixedWidthVector) | |||
861 | .getFixedSize(); | |||
862 | ||||
863 | if (MinVectorRegSizeOption.getNumOccurrences()) | |||
864 | MinVecRegSize = MinVectorRegSizeOption; | |||
865 | else | |||
866 | MinVecRegSize = TTI->getMinVectorRegisterBitWidth(); | |||
867 | } | |||
868 | ||||
869 | /// Vectorize the tree that starts with the elements in \p VL. | |||
870 | /// Returns the vectorized root. | |||
871 | Value *vectorizeTree(); | |||
872 | ||||
873 | /// Vectorize the tree but with the list of externally used values \p | |||
874 | /// ExternallyUsedValues. Values in this MapVector can be replaced but the | |||
875 | /// generated extractvalue instructions. | |||
876 | Value *vectorizeTree(ExtraValueToDebugLocsMap &ExternallyUsedValues); | |||
877 | ||||
878 | /// \returns the cost incurred by unwanted spills and fills, caused by | |||
879 | /// holding live values over call sites. | |||
880 | InstructionCost getSpillCost() const; | |||
881 | ||||
882 | /// \returns the vectorization cost of the subtree that starts at \p VL. | |||
883 | /// A negative number means that this is profitable. | |||
884 | InstructionCost getTreeCost(ArrayRef<Value *> VectorizedVals = None); | |||
885 | ||||
886 | /// Construct a vectorizable tree that starts at \p Roots, ignoring users for | |||
887 | /// the purpose of scheduling and extraction in the \p UserIgnoreLst. | |||
888 | void buildTree(ArrayRef<Value *> Roots, | |||
889 | ArrayRef<Value *> UserIgnoreLst = None); | |||
890 | ||||
891 | /// Builds external uses of the vectorized scalars, i.e. the list of | |||
892 | /// vectorized scalars to be extracted, their lanes and their scalar users. \p | |||
893 | /// ExternallyUsedValues contains additional list of external uses to handle | |||
894 | /// vectorization of reductions. | |||
895 | void | |||
896 | buildExternalUses(const ExtraValueToDebugLocsMap &ExternallyUsedValues = {}); | |||
897 | ||||
898 | /// Clear the internal data structures that are created by 'buildTree'. | |||
899 | void deleteTree() { | |||
900 | VectorizableTree.clear(); | |||
901 | ScalarToTreeEntry.clear(); | |||
902 | MustGather.clear(); | |||
903 | ExternalUses.clear(); | |||
904 | for (auto &Iter : BlocksSchedules) { | |||
905 | BlockScheduling *BS = Iter.second.get(); | |||
906 | BS->clear(); | |||
907 | } | |||
908 | MinBWs.clear(); | |||
909 | InstrElementSize.clear(); | |||
910 | } | |||
911 | ||||
912 | unsigned getTreeSize() const { return VectorizableTree.size(); } | |||
913 | ||||
914 | /// Perform LICM and CSE on the newly generated gather sequences. | |||
915 | void optimizeGatherSequence(); | |||
916 | ||||
917 | /// Checks if the specified gather tree entry \p TE can be represented as a | |||
918 | /// shuffled vector entry + (possibly) permutation with other gathers. It | |||
919 | /// implements the checks only for possibly ordered scalars (Loads, | |||
920 | /// ExtractElement, ExtractValue), which can be part of the graph. | |||
921 | Optional<OrdersType> findReusedOrderedScalars(const TreeEntry &TE); | |||
922 | ||||
923 | /// Gets reordering data for the given tree entry. If the entry is vectorized | |||
924 | /// - just return ReorderIndices, otherwise check if the scalars can be | |||
925 | /// reordered and return the most optimal order. | |||
926 | /// \param TopToBottom If true, include the order of vectorized stores and | |||
927 | /// insertelement nodes, otherwise skip them. | |||
928 | Optional<OrdersType> getReorderingData(const TreeEntry &TE, bool TopToBottom); | |||
929 | ||||
930 | /// Reorders the current graph to the most profitable order starting from the | |||
931 | /// root node to the leaf nodes. The best order is chosen only from the nodes | |||
932 | /// of the same size (vectorization factor). Smaller nodes are considered | |||
933 | /// parts of subgraph with smaller VF and they are reordered independently. We | |||
934 | /// can make it because we still need to extend smaller nodes to the wider VF | |||
935 | /// and we can merge reordering shuffles with the widening shuffles. | |||
936 | void reorderTopToBottom(); | |||
937 | ||||
938 | /// Reorders the current graph to the most profitable order starting from | |||
939 | /// leaves to the root. It allows to rotate small subgraphs and reduce the | |||
940 | /// number of reshuffles if the leaf nodes use the same order. In this case we | |||
941 | /// can merge the orders and just shuffle user node instead of shuffling its | |||
942 | /// operands. Plus, even the leaf nodes have different orders, it allows to | |||
943 | /// sink reordering in the graph closer to the root node and merge it later | |||
944 | /// during analysis. | |||
945 | void reorderBottomToTop(bool IgnoreReorder = false); | |||
946 | ||||
947 | /// \return The vector element size in bits to use when vectorizing the | |||
948 | /// expression tree ending at \p V. If V is a store, the size is the width of | |||
949 | /// the stored value. Otherwise, the size is the width of the largest loaded | |||
950 | /// value reaching V. This method is used by the vectorizer to calculate | |||
951 | /// vectorization factors. | |||
952 | unsigned getVectorElementSize(Value *V); | |||
953 | ||||
954 | /// Compute the minimum type sizes required to represent the entries in a | |||
955 | /// vectorizable tree. | |||
956 | void computeMinimumValueSizes(); | |||
957 | ||||
958 | // \returns maximum vector register size as set by TTI or overridden by cl::opt. | |||
959 | unsigned getMaxVecRegSize() const { | |||
960 | return MaxVecRegSize; | |||
961 | } | |||
962 | ||||
963 | // \returns minimum vector register size as set by cl::opt. | |||
964 | unsigned getMinVecRegSize() const { | |||
965 | return MinVecRegSize; | |||
966 | } | |||
967 | ||||
968 | unsigned getMinVF(unsigned Sz) const { | |||
969 | return std::max(2U, getMinVecRegSize() / Sz); | |||
970 | } | |||
971 | ||||
972 | unsigned getMaximumVF(unsigned ElemWidth, unsigned Opcode) const { | |||
973 | unsigned MaxVF = MaxVFOption.getNumOccurrences() ? | |||
974 | MaxVFOption : TTI->getMaximumVF(ElemWidth, Opcode); | |||
975 | return MaxVF ? MaxVF : UINT_MAX(2147483647 *2U +1U); | |||
976 | } | |||
977 | ||||
978 | /// Check if homogeneous aggregate is isomorphic to some VectorType. | |||
979 | /// Accepts homogeneous multidimensional aggregate of scalars/vectors like | |||
980 | /// {[4 x i16], [4 x i16]}, { <2 x float>, <2 x float> }, | |||
981 | /// {{{i16, i16}, {i16, i16}}, {{i16, i16}, {i16, i16}}} and so on. | |||
982 | /// | |||
983 | /// \returns number of elements in vector if isomorphism exists, 0 otherwise. | |||
984 | unsigned canMapToVector(Type *T, const DataLayout &DL) const; | |||
985 | ||||
986 | /// \returns True if the VectorizableTree is both tiny and not fully | |||
987 | /// vectorizable. We do not vectorize such trees. | |||
988 | bool isTreeTinyAndNotFullyVectorizable(bool ForReduction = false) const; | |||
989 | ||||
990 | /// Assume that a legal-sized 'or'-reduction of shifted/zexted loaded values | |||
991 | /// can be load combined in the backend. Load combining may not be allowed in | |||
992 | /// the IR optimizer, so we do not want to alter the pattern. For example, | |||
993 | /// partially transforming a scalar bswap() pattern into vector code is | |||
994 | /// effectively impossible for the backend to undo. | |||
995 | /// TODO: If load combining is allowed in the IR optimizer, this analysis | |||
996 | /// may not be necessary. | |||
997 | bool isLoadCombineReductionCandidate(RecurKind RdxKind) const; | |||
998 | ||||
999 | /// Assume that a vector of stores of bitwise-or/shifted/zexted loaded values | |||
1000 | /// can be load combined in the backend. Load combining may not be allowed in | |||
1001 | /// the IR optimizer, so we do not want to alter the pattern. For example, | |||
1002 | /// partially transforming a scalar bswap() pattern into vector code is | |||
1003 | /// effectively impossible for the backend to undo. | |||
1004 | /// TODO: If load combining is allowed in the IR optimizer, this analysis | |||
1005 | /// may not be necessary. | |||
1006 | bool isLoadCombineCandidate() const; | |||
1007 | ||||
1008 | OptimizationRemarkEmitter *getORE() { return ORE; } | |||
1009 | ||||
1010 | /// This structure holds any data we need about the edges being traversed | |||
1011 | /// during buildTree_rec(). We keep track of: | |||
1012 | /// (i) the user TreeEntry index, and | |||
1013 | /// (ii) the index of the edge. | |||
1014 | struct EdgeInfo { | |||
1015 | EdgeInfo() = default; | |||
1016 | EdgeInfo(TreeEntry *UserTE, unsigned EdgeIdx) | |||
1017 | : UserTE(UserTE), EdgeIdx(EdgeIdx) {} | |||
1018 | /// The user TreeEntry. | |||
1019 | TreeEntry *UserTE = nullptr; | |||
1020 | /// The operand index of the use. | |||
1021 | unsigned EdgeIdx = UINT_MAX(2147483647 *2U +1U); | |||
1022 | #ifndef NDEBUG | |||
1023 | friend inline raw_ostream &operator<<(raw_ostream &OS, | |||
1024 | const BoUpSLP::EdgeInfo &EI) { | |||
1025 | EI.dump(OS); | |||
1026 | return OS; | |||
1027 | } | |||
1028 | /// Debug print. | |||
1029 | void dump(raw_ostream &OS) const { | |||
1030 | OS << "{User:" << (UserTE ? std::to_string(UserTE->Idx) : "null") | |||
1031 | << " EdgeIdx:" << EdgeIdx << "}"; | |||
1032 | } | |||
1033 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void dump() const { dump(dbgs()); } | |||
1034 | #endif | |||
1035 | }; | |||
1036 | ||||
1037 | /// A helper data structure to hold the operands of a vector of instructions. | |||
1038 | /// This supports a fixed vector length for all operand vectors. | |||
1039 | class VLOperands { | |||
1040 | /// For each operand we need (i) the value, and (ii) the opcode that it | |||
1041 | /// would be attached to if the expression was in a left-linearized form. | |||
1042 | /// This is required to avoid illegal operand reordering. | |||
1043 | /// For example: | |||
1044 | /// \verbatim | |||
1045 | /// 0 Op1 | |||
1046 | /// |/ | |||
1047 | /// Op1 Op2 Linearized + Op2 | |||
1048 | /// \ / ----------> |/ | |||
1049 | /// - - | |||
1050 | /// | |||
1051 | /// Op1 - Op2 (0 + Op1) - Op2 | |||
1052 | /// \endverbatim | |||
1053 | /// | |||
1054 | /// Value Op1 is attached to a '+' operation, and Op2 to a '-'. | |||
1055 | /// | |||
1056 | /// Another way to think of this is to track all the operations across the | |||
1057 | /// path from the operand all the way to the root of the tree and to | |||
1058 | /// calculate the operation that corresponds to this path. For example, the | |||
1059 | /// path from Op2 to the root crosses the RHS of the '-', therefore the | |||
1060 | /// corresponding operation is a '-' (which matches the one in the | |||
1061 | /// linearized tree, as shown above). | |||
1062 | /// | |||
1063 | /// For lack of a better term, we refer to this operation as Accumulated | |||
1064 | /// Path Operation (APO). | |||
1065 | struct OperandData { | |||
1066 | OperandData() = default; | |||
1067 | OperandData(Value *V, bool APO, bool IsUsed) | |||
1068 | : V(V), APO(APO), IsUsed(IsUsed) {} | |||
1069 | /// The operand value. | |||
1070 | Value *V = nullptr; | |||
1071 | /// TreeEntries only allow a single opcode, or an alternate sequence of | |||
1072 | /// them (e.g, +, -). Therefore, we can safely use a boolean value for the | |||
1073 | /// APO. It is set to 'true' if 'V' is attached to an inverse operation | |||
1074 | /// in the left-linearized form (e.g., Sub/Div), and 'false' otherwise | |||
1075 | /// (e.g., Add/Mul) | |||
1076 | bool APO = false; | |||
1077 | /// Helper data for the reordering function. | |||
1078 | bool IsUsed = false; | |||
1079 | }; | |||
1080 | ||||
1081 | /// During operand reordering, we are trying to select the operand at lane | |||
1082 | /// that matches best with the operand at the neighboring lane. Our | |||
1083 | /// selection is based on the type of value we are looking for. For example, | |||
1084 | /// if the neighboring lane has a load, we need to look for a load that is | |||
1085 | /// accessing a consecutive address. These strategies are summarized in the | |||
1086 | /// 'ReorderingMode' enumerator. | |||
1087 | enum class ReorderingMode { | |||
1088 | Load, ///< Matching loads to consecutive memory addresses | |||
1089 | Opcode, ///< Matching instructions based on opcode (same or alternate) | |||
1090 | Constant, ///< Matching constants | |||
1091 | Splat, ///< Matching the same instruction multiple times (broadcast) | |||
1092 | Failed, ///< We failed to create a vectorizable group | |||
1093 | }; | |||
1094 | ||||
1095 | using OperandDataVec = SmallVector<OperandData, 2>; | |||
1096 | ||||
1097 | /// A vector of operand vectors. | |||
1098 | SmallVector<OperandDataVec, 4> OpsVec; | |||
1099 | ||||
1100 | const DataLayout &DL; | |||
1101 | ScalarEvolution &SE; | |||
1102 | const BoUpSLP &R; | |||
1103 | ||||
1104 | /// \returns the operand data at \p OpIdx and \p Lane. | |||
1105 | OperandData &getData(unsigned OpIdx, unsigned Lane) { | |||
1106 | return OpsVec[OpIdx][Lane]; | |||
1107 | } | |||
1108 | ||||
1109 | /// \returns the operand data at \p OpIdx and \p Lane. Const version. | |||
1110 | const OperandData &getData(unsigned OpIdx, unsigned Lane) const { | |||
1111 | return OpsVec[OpIdx][Lane]; | |||
1112 | } | |||
1113 | ||||
1114 | /// Clears the used flag for all entries. | |||
1115 | void clearUsed() { | |||
1116 | for (unsigned OpIdx = 0, NumOperands = getNumOperands(); | |||
1117 | OpIdx != NumOperands; ++OpIdx) | |||
1118 | for (unsigned Lane = 0, NumLanes = getNumLanes(); Lane != NumLanes; | |||
1119 | ++Lane) | |||
1120 | OpsVec[OpIdx][Lane].IsUsed = false; | |||
1121 | } | |||
1122 | ||||
1123 | /// Swap the operand at \p OpIdx1 with that one at \p OpIdx2. | |||
1124 | void swap(unsigned OpIdx1, unsigned OpIdx2, unsigned Lane) { | |||
1125 | std::swap(OpsVec[OpIdx1][Lane], OpsVec[OpIdx2][Lane]); | |||
1126 | } | |||
1127 | ||||
1128 | // The hard-coded scores listed here are not very important, though it shall | |||
1129 | // be higher for better matches to improve the resulting cost. When | |||
1130 | // computing the scores of matching one sub-tree with another, we are | |||
1131 | // basically counting the number of values that are matching. So even if all | |||
1132 | // scores are set to 1, we would still get a decent matching result. | |||
1133 | // However, sometimes we have to break ties. For example we may have to | |||
1134 | // choose between matching loads vs matching opcodes. This is what these | |||
1135 | // scores are helping us with: they provide the order of preference. Also, | |||
1136 | // this is important if the scalar is externally used or used in another | |||
1137 | // tree entry node in the different lane. | |||
1138 | ||||
1139 | /// Loads from consecutive memory addresses, e.g. load(A[i]), load(A[i+1]). | |||
1140 | static const int ScoreConsecutiveLoads = 4; | |||
1141 | /// The same load multiple times. This should have a better score than | |||
1142 | /// `ScoreSplat` because it in x86 for a 2-lane vector we can represent it | |||
1143 | /// with `movddup (%reg), xmm0` which has a throughput of 0.5 versus 0.5 for | |||
1144 | /// a vector load and 1.0 for a broadcast. | |||
1145 | static const int ScoreSplatLoads = 3; | |||
1146 | /// Loads from reversed memory addresses, e.g. load(A[i+1]), load(A[i]). | |||
1147 | static const int ScoreReversedLoads = 3; | |||
1148 | /// ExtractElementInst from same vector and consecutive indexes. | |||
1149 | static const int ScoreConsecutiveExtracts = 4; | |||
1150 | /// ExtractElementInst from same vector and reversed indices. | |||
1151 | static const int ScoreReversedExtracts = 3; | |||
1152 | /// Constants. | |||
1153 | static const int ScoreConstants = 2; | |||
1154 | /// Instructions with the same opcode. | |||
1155 | static const int ScoreSameOpcode = 2; | |||
1156 | /// Instructions with alt opcodes (e.g, add + sub). | |||
1157 | static const int ScoreAltOpcodes = 1; | |||
1158 | /// Identical instructions (a.k.a. splat or broadcast). | |||
1159 | static const int ScoreSplat = 1; | |||
1160 | /// Matching with an undef is preferable to failing. | |||
1161 | static const int ScoreUndef = 1; | |||
1162 | /// Score for failing to find a decent match. | |||
1163 | static const int ScoreFail = 0; | |||
1164 | /// Score if all users are vectorized. | |||
1165 | static const int ScoreAllUserVectorized = 1; | |||
1166 | ||||
1167 | /// \returns the score of placing \p V1 and \p V2 in consecutive lanes. | |||
1168 | /// \p U1 and \p U2 are the users of \p V1 and \p V2. | |||
1169 | /// Also, checks if \p V1 and \p V2 are compatible with instructions in \p | |||
1170 | /// MainAltOps. | |||
1171 | int getShallowScore(Value *V1, Value *V2, Instruction *U1, Instruction *U2, | |||
1172 | const DataLayout &DL, ScalarEvolution &SE, int NumLanes, | |||
1173 | ArrayRef<Value *> MainAltOps) { | |||
1174 | if (V1 == V2) { | |||
1175 | if (isa<LoadInst>(V1)) { | |||
1176 | // Retruns true if the users of V1 and V2 won't need to be extracted. | |||
1177 | auto AllUsersAreInternal = [U1, U2, this](Value *V1, Value *V2) { | |||
1178 | // Bail out if we have too many uses to save compilation time. | |||
1179 | static constexpr unsigned Limit = 8; | |||
1180 | if (V1->hasNUsesOrMore(Limit) || V2->hasNUsesOrMore(Limit)) | |||
1181 | return false; | |||
1182 | ||||
1183 | auto AllUsersVectorized = [U1, U2, this](Value *V) { | |||
1184 | return llvm::all_of(V->users(), [U1, U2, this](Value *U) { | |||
1185 | return U == U1 || U == U2 || R.getTreeEntry(U) != nullptr; | |||
1186 | }); | |||
1187 | }; | |||
1188 | return AllUsersVectorized(V1) && AllUsersVectorized(V2); | |||
1189 | }; | |||
1190 | // A broadcast of a load can be cheaper on some targets. | |||
1191 | if (R.TTI->isLegalBroadcastLoad(V1->getType(), NumLanes) && | |||
1192 | ((int)V1->getNumUses() == NumLanes || | |||
1193 | AllUsersAreInternal(V1, V2))) | |||
1194 | return VLOperands::ScoreSplatLoads; | |||
1195 | } | |||
1196 | return VLOperands::ScoreSplat; | |||
1197 | } | |||
1198 | ||||
1199 | auto *LI1 = dyn_cast<LoadInst>(V1); | |||
1200 | auto *LI2 = dyn_cast<LoadInst>(V2); | |||
1201 | if (LI1 && LI2) { | |||
1202 | if (LI1->getParent() != LI2->getParent()) | |||
1203 | return VLOperands::ScoreFail; | |||
1204 | ||||
1205 | Optional<int> Dist = getPointersDiff( | |||
1206 | LI1->getType(), LI1->getPointerOperand(), LI2->getType(), | |||
1207 | LI2->getPointerOperand(), DL, SE, /*StrictCheck=*/true); | |||
1208 | if (!Dist || *Dist == 0) | |||
1209 | return VLOperands::ScoreFail; | |||
1210 | // The distance is too large - still may be profitable to use masked | |||
1211 | // loads/gathers. | |||
1212 | if (std::abs(*Dist) > NumLanes / 2) | |||
1213 | return VLOperands::ScoreAltOpcodes; | |||
1214 | // This still will detect consecutive loads, but we might have "holes" | |||
1215 | // in some cases. It is ok for non-power-2 vectorization and may produce | |||
1216 | // better results. It should not affect current vectorization. | |||
1217 | return (*Dist > 0) ? VLOperands::ScoreConsecutiveLoads | |||
1218 | : VLOperands::ScoreReversedLoads; | |||
1219 | } | |||
1220 | ||||
1221 | auto *C1 = dyn_cast<Constant>(V1); | |||
1222 | auto *C2 = dyn_cast<Constant>(V2); | |||
1223 | if (C1 && C2) | |||
1224 | return VLOperands::ScoreConstants; | |||
1225 | ||||
1226 | // Extracts from consecutive indexes of the same vector better score as | |||
1227 | // the extracts could be optimized away. | |||
1228 | Value *EV1; | |||
1229 | ConstantInt *Ex1Idx; | |||
1230 | if (match(V1, m_ExtractElt(m_Value(EV1), m_ConstantInt(Ex1Idx)))) { | |||
1231 | // Undefs are always profitable for extractelements. | |||
1232 | if (isa<UndefValue>(V2)) | |||
1233 | return VLOperands::ScoreConsecutiveExtracts; | |||
1234 | Value *EV2 = nullptr; | |||
1235 | ConstantInt *Ex2Idx = nullptr; | |||
1236 | if (match(V2, | |||
1237 | m_ExtractElt(m_Value(EV2), m_CombineOr(m_ConstantInt(Ex2Idx), | |||
1238 | m_Undef())))) { | |||
1239 | // Undefs are always profitable for extractelements. | |||
1240 | if (!Ex2Idx) | |||
1241 | return VLOperands::ScoreConsecutiveExtracts; | |||
1242 | if (isUndefVector(EV2) && EV2->getType() == EV1->getType()) | |||
1243 | return VLOperands::ScoreConsecutiveExtracts; | |||
1244 | if (EV2 == EV1) { | |||
1245 | int Idx1 = Ex1Idx->getZExtValue(); | |||
1246 | int Idx2 = Ex2Idx->getZExtValue(); | |||
1247 | int Dist = Idx2 - Idx1; | |||
1248 | // The distance is too large - still may be profitable to use | |||
1249 | // shuffles. | |||
1250 | if (std::abs(Dist) == 0) | |||
1251 | return VLOperands::ScoreSplat; | |||
1252 | if (std::abs(Dist) > NumLanes / 2) | |||
1253 | return VLOperands::ScoreSameOpcode; | |||
1254 | return (Dist > 0) ? VLOperands::ScoreConsecutiveExtracts | |||
1255 | : VLOperands::ScoreReversedExtracts; | |||
1256 | } | |||
1257 | return VLOperands::ScoreAltOpcodes; | |||
1258 | } | |||
1259 | return VLOperands::ScoreFail; | |||
1260 | } | |||
1261 | ||||
1262 | auto *I1 = dyn_cast<Instruction>(V1); | |||
1263 | auto *I2 = dyn_cast<Instruction>(V2); | |||
1264 | if (I1 && I2) { | |||
1265 | if (I1->getParent() != I2->getParent()) | |||
1266 | return VLOperands::ScoreFail; | |||
1267 | SmallVector<Value *, 4> Ops(MainAltOps.begin(), MainAltOps.end()); | |||
1268 | Ops.push_back(I1); | |||
1269 | Ops.push_back(I2); | |||
1270 | InstructionsState S = getSameOpcode(Ops); | |||
1271 | // Note: Only consider instructions with <= 2 operands to avoid | |||
1272 | // complexity explosion. | |||
1273 | if (S.getOpcode() && | |||
1274 | (S.MainOp->getNumOperands() <= 2 || !MainAltOps.empty() || | |||
1275 | !S.isAltShuffle()) && | |||
1276 | all_of(Ops, [&S](Value *V) { | |||
1277 | return cast<Instruction>(V)->getNumOperands() == | |||
1278 | S.MainOp->getNumOperands(); | |||
1279 | })) | |||
1280 | return S.isAltShuffle() ? VLOperands::ScoreAltOpcodes | |||
1281 | : VLOperands::ScoreSameOpcode; | |||
1282 | } | |||
1283 | ||||
1284 | if (isa<UndefValue>(V2)) | |||
1285 | return VLOperands::ScoreUndef; | |||
1286 | ||||
1287 | return VLOperands::ScoreFail; | |||
1288 | } | |||
1289 | ||||
1290 | /// \param Lane lane of the operands under analysis. | |||
1291 | /// \param OpIdx operand index in \p Lane lane we're looking the best | |||
1292 | /// candidate for. | |||
1293 | /// \param Idx operand index of the current candidate value. | |||
1294 | /// \returns The additional score due to possible broadcasting of the | |||
1295 | /// elements in the lane. It is more profitable to have power-of-2 unique | |||
1296 | /// elements in the lane, it will be vectorized with higher probability | |||
1297 | /// after removing duplicates. Currently the SLP vectorizer supports only | |||
1298 | /// vectorization of the power-of-2 number of unique scalars. | |||
1299 | int getSplatScore(unsigned Lane, unsigned OpIdx, unsigned Idx) const { | |||
1300 | Value *IdxLaneV = getData(Idx, Lane).V; | |||
1301 | if (!isa<Instruction>(IdxLaneV) || IdxLaneV == getData(OpIdx, Lane).V) | |||
1302 | return 0; | |||
1303 | SmallPtrSet<Value *, 4> Uniques; | |||
1304 | for (unsigned Ln = 0, E = getNumLanes(); Ln < E; ++Ln) { | |||
1305 | if (Ln == Lane) | |||
1306 | continue; | |||
1307 | Value *OpIdxLnV = getData(OpIdx, Ln).V; | |||
1308 | if (!isa<Instruction>(OpIdxLnV)) | |||
1309 | return 0; | |||
1310 | Uniques.insert(OpIdxLnV); | |||
1311 | } | |||
1312 | int UniquesCount = Uniques.size(); | |||
1313 | int UniquesCntWithIdxLaneV = | |||
1314 | Uniques.contains(IdxLaneV) ? UniquesCount : UniquesCount + 1; | |||
1315 | Value *OpIdxLaneV = getData(OpIdx, Lane).V; | |||
1316 | int UniquesCntWithOpIdxLaneV = | |||
1317 | Uniques.contains(OpIdxLaneV) ? UniquesCount : UniquesCount + 1; | |||
1318 | if (UniquesCntWithIdxLaneV == UniquesCntWithOpIdxLaneV) | |||
1319 | return 0; | |||
1320 | return (PowerOf2Ceil(UniquesCntWithOpIdxLaneV) - | |||
1321 | UniquesCntWithOpIdxLaneV) - | |||
1322 | (PowerOf2Ceil(UniquesCntWithIdxLaneV) - UniquesCntWithIdxLaneV); | |||
1323 | } | |||
1324 | ||||
1325 | /// \param Lane lane of the operands under analysis. | |||
1326 | /// \param OpIdx operand index in \p Lane lane we're looking the best | |||
1327 | /// candidate for. | |||
1328 | /// \param Idx operand index of the current candidate value. | |||
1329 | /// \returns The additional score for the scalar which users are all | |||
1330 | /// vectorized. | |||
1331 | int getExternalUseScore(unsigned Lane, unsigned OpIdx, unsigned Idx) const { | |||
1332 | Value *IdxLaneV = getData(Idx, Lane).V; | |||
1333 | Value *OpIdxLaneV = getData(OpIdx, Lane).V; | |||
1334 | // Do not care about number of uses for vector-like instructions | |||
1335 | // (extractelement/extractvalue with constant indices), they are extracts | |||
1336 | // themselves and already externally used. Vectorization of such | |||
1337 | // instructions does not add extra extractelement instruction, just may | |||
1338 | // remove it. | |||
1339 | if (isVectorLikeInstWithConstOps(IdxLaneV) && | |||
1340 | isVectorLikeInstWithConstOps(OpIdxLaneV)) | |||
1341 | return VLOperands::ScoreAllUserVectorized; | |||
1342 | auto *IdxLaneI = dyn_cast<Instruction>(IdxLaneV); | |||
1343 | if (!IdxLaneI || !isa<Instruction>(OpIdxLaneV)) | |||
1344 | return 0; | |||
1345 | return R.areAllUsersVectorized(IdxLaneI, None) | |||
1346 | ? VLOperands::ScoreAllUserVectorized | |||
1347 | : 0; | |||
1348 | } | |||
1349 | ||||
1350 | /// Go through the operands of \p LHS and \p RHS recursively until \p | |||
1351 | /// MaxLevel, and return the cummulative score. \p U1 and \p U2 are | |||
1352 | /// the users of \p LHS and \p RHS (that is \p LHS and \p RHS are operands | |||
1353 | /// of \p U1 and \p U2), except at the beginning of the recursion where | |||
1354 | /// these are set to nullptr. | |||
1355 | /// | |||
1356 | /// For example: | |||
1357 | /// \verbatim | |||
1358 | /// A[0] B[0] A[1] B[1] C[0] D[0] B[1] A[1] | |||
1359 | /// \ / \ / \ / \ / | |||
1360 | /// + + + + | |||
1361 | /// G1 G2 G3 G4 | |||
1362 | /// \endverbatim | |||
1363 | /// The getScoreAtLevelRec(G1, G2) function will try to match the nodes at | |||
1364 | /// each level recursively, accumulating the score. It starts from matching | |||
1365 | /// the additions at level 0, then moves on to the loads (level 1). The | |||
1366 | /// score of G1 and G2 is higher than G1 and G3, because {A[0],A[1]} and | |||
1367 | /// {B[0],B[1]} match with VLOperands::ScoreConsecutiveLoads, while | |||
1368 | /// {A[0],C[0]} has a score of VLOperands::ScoreFail. | |||
1369 | /// Please note that the order of the operands does not matter, as we | |||
1370 | /// evaluate the score of all profitable combinations of operands. In | |||
1371 | /// other words the score of G1 and G4 is the same as G1 and G2. This | |||
1372 | /// heuristic is based on ideas described in: | |||
1373 | /// Look-ahead SLP: Auto-vectorization in the presence of commutative | |||
1374 | /// operations, CGO 2018 by Vasileios Porpodas, Rodrigo C. O. Rocha, | |||
1375 | /// Luís F. W. Góes | |||
1376 | int getScoreAtLevelRec(Value *LHS, Value *RHS, Instruction *U1, | |||
1377 | Instruction *U2, int CurrLevel, int MaxLevel, | |||
1378 | ArrayRef<Value *> MainAltOps) { | |||
1379 | ||||
1380 | // Get the shallow score of V1 and V2. | |||
1381 | int ShallowScoreAtThisLevel = | |||
1382 | getShallowScore(LHS, RHS, U1, U2, DL, SE, getNumLanes(), MainAltOps); | |||
1383 | ||||
1384 | // If reached MaxLevel, | |||
1385 | // or if V1 and V2 are not instructions, | |||
1386 | // or if they are SPLAT, | |||
1387 | // or if they are not consecutive, | |||
1388 | // or if profitable to vectorize loads or extractelements, early return | |||
1389 | // the current cost. | |||
1390 | auto *I1 = dyn_cast<Instruction>(LHS); | |||
1391 | auto *I2 = dyn_cast<Instruction>(RHS); | |||
1392 | if (CurrLevel == MaxLevel || !(I1 && I2) || I1 == I2 || | |||
1393 | ShallowScoreAtThisLevel == VLOperands::ScoreFail || | |||
1394 | (((isa<LoadInst>(I1) && isa<LoadInst>(I2)) || | |||
1395 | (I1->getNumOperands() > 2 && I2->getNumOperands() > 2) || | |||
1396 | (isa<ExtractElementInst>(I1) && isa<ExtractElementInst>(I2))) && | |||
1397 | ShallowScoreAtThisLevel)) | |||
1398 | return ShallowScoreAtThisLevel; | |||
1399 | assert(I1 && I2 && "Should have early exited.")(static_cast <bool> (I1 && I2 && "Should have early exited." ) ? void (0) : __assert_fail ("I1 && I2 && \"Should have early exited.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 1399, __extension__ __PRETTY_FUNCTION__)); | |||
1400 | ||||
1401 | // Contains the I2 operand indexes that got matched with I1 operands. | |||
1402 | SmallSet<unsigned, 4> Op2Used; | |||
1403 | ||||
1404 | // Recursion towards the operands of I1 and I2. We are trying all possible | |||
1405 | // operand pairs, and keeping track of the best score. | |||
1406 | for (unsigned OpIdx1 = 0, NumOperands1 = I1->getNumOperands(); | |||
1407 | OpIdx1 != NumOperands1; ++OpIdx1) { | |||
1408 | // Try to pair op1I with the best operand of I2. | |||
1409 | int MaxTmpScore = 0; | |||
1410 | unsigned MaxOpIdx2 = 0; | |||
1411 | bool FoundBest = false; | |||
1412 | // If I2 is commutative try all combinations. | |||
1413 | unsigned FromIdx = isCommutative(I2) ? 0 : OpIdx1; | |||
1414 | unsigned ToIdx = isCommutative(I2) | |||
1415 | ? I2->getNumOperands() | |||
1416 | : std::min(I2->getNumOperands(), OpIdx1 + 1); | |||
1417 | assert(FromIdx <= ToIdx && "Bad index")(static_cast <bool> (FromIdx <= ToIdx && "Bad index" ) ? void (0) : __assert_fail ("FromIdx <= ToIdx && \"Bad index\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 1417, __extension__ __PRETTY_FUNCTION__)); | |||
1418 | for (unsigned OpIdx2 = FromIdx; OpIdx2 != ToIdx; ++OpIdx2) { | |||
1419 | // Skip operands already paired with OpIdx1. | |||
1420 | if (Op2Used.count(OpIdx2)) | |||
1421 | continue; | |||
1422 | // Recursively calculate the cost at each level | |||
1423 | int TmpScore = | |||
1424 | getScoreAtLevelRec(I1->getOperand(OpIdx1), I2->getOperand(OpIdx2), | |||
1425 | I1, I2, CurrLevel + 1, MaxLevel, None); | |||
1426 | // Look for the best score. | |||
1427 | if (TmpScore > VLOperands::ScoreFail && TmpScore > MaxTmpScore) { | |||
1428 | MaxTmpScore = TmpScore; | |||
1429 | MaxOpIdx2 = OpIdx2; | |||
1430 | FoundBest = true; | |||
1431 | } | |||
1432 | } | |||
1433 | if (FoundBest) { | |||
1434 | // Pair {OpIdx1, MaxOpIdx2} was found to be best. Never revisit it. | |||
1435 | Op2Used.insert(MaxOpIdx2); | |||
1436 | ShallowScoreAtThisLevel += MaxTmpScore; | |||
1437 | } | |||
1438 | } | |||
1439 | return ShallowScoreAtThisLevel; | |||
1440 | } | |||
1441 | ||||
1442 | /// Score scaling factor for fully compatible instructions but with | |||
1443 | /// different number of external uses. Allows better selection of the | |||
1444 | /// instructions with less external uses. | |||
1445 | static const int ScoreScaleFactor = 10; | |||
1446 | ||||
1447 | /// \Returns the look-ahead score, which tells us how much the sub-trees | |||
1448 | /// rooted at \p LHS and \p RHS match, the more they match the higher the | |||
1449 | /// score. This helps break ties in an informed way when we cannot decide on | |||
1450 | /// the order of the operands by just considering the immediate | |||
1451 | /// predecessors. | |||
1452 | int getLookAheadScore(Value *LHS, Value *RHS, ArrayRef<Value *> MainAltOps, | |||
1453 | int Lane, unsigned OpIdx, unsigned Idx, | |||
1454 | bool &IsUsed) { | |||
1455 | // Keep track of the instruction stack as we recurse into the operands | |||
1456 | // during the look-ahead score exploration. | |||
1457 | int Score = getScoreAtLevelRec(LHS, RHS, /*U1=*/nullptr, /*U2=*/nullptr, | |||
1458 | 1, LookAheadMaxDepth, MainAltOps); | |||
1459 | if (Score) { | |||
1460 | int SplatScore = getSplatScore(Lane, OpIdx, Idx); | |||
1461 | if (Score <= -SplatScore) { | |||
1462 | // Set the minimum score for splat-like sequence to avoid setting | |||
1463 | // failed state. | |||
1464 | Score = 1; | |||
1465 | } else { | |||
1466 | Score += SplatScore; | |||
1467 | // Scale score to see the difference between different operands | |||
1468 | // and similar operands but all vectorized/not all vectorized | |||
1469 | // uses. It does not affect actual selection of the best | |||
1470 | // compatible operand in general, just allows to select the | |||
1471 | // operand with all vectorized uses. | |||
1472 | Score *= ScoreScaleFactor; | |||
1473 | Score += getExternalUseScore(Lane, OpIdx, Idx); | |||
1474 | IsUsed = true; | |||
1475 | } | |||
1476 | } | |||
1477 | return Score; | |||
1478 | } | |||
1479 | ||||
1480 | /// Best defined scores per lanes between the passes. Used to choose the | |||
1481 | /// best operand (with the highest score) between the passes. | |||
1482 | /// The key - {Operand Index, Lane}. | |||
1483 | /// The value - the best score between the passes for the lane and the | |||
1484 | /// operand. | |||
1485 | SmallDenseMap<std::pair<unsigned, unsigned>, unsigned, 8> | |||
1486 | BestScoresPerLanes; | |||
1487 | ||||
1488 | // Search all operands in Ops[*][Lane] for the one that matches best | |||
1489 | // Ops[OpIdx][LastLane] and return its opreand index. | |||
1490 | // If no good match can be found, return None. | |||
1491 | Optional<unsigned> getBestOperand(unsigned OpIdx, int Lane, int LastLane, | |||
1492 | ArrayRef<ReorderingMode> ReorderingModes, | |||
1493 | ArrayRef<Value *> MainAltOps) { | |||
1494 | unsigned NumOperands = getNumOperands(); | |||
1495 | ||||
1496 | // The operand of the previous lane at OpIdx. | |||
1497 | Value *OpLastLane = getData(OpIdx, LastLane).V; | |||
1498 | ||||
1499 | // Our strategy mode for OpIdx. | |||
1500 | ReorderingMode RMode = ReorderingModes[OpIdx]; | |||
1501 | if (RMode == ReorderingMode::Failed) | |||
1502 | return None; | |||
1503 | ||||
1504 | // The linearized opcode of the operand at OpIdx, Lane. | |||
1505 | bool OpIdxAPO = getData(OpIdx, Lane).APO; | |||
1506 | ||||
1507 | // The best operand index and its score. | |||
1508 | // Sometimes we have more than one option (e.g., Opcode and Undefs), so we | |||
1509 | // are using the score to differentiate between the two. | |||
1510 | struct BestOpData { | |||
1511 | Optional<unsigned> Idx = None; | |||
1512 | unsigned Score = 0; | |||
1513 | } BestOp; | |||
1514 | BestOp.Score = | |||
1515 | BestScoresPerLanes.try_emplace(std::make_pair(OpIdx, Lane), 0) | |||
1516 | .first->second; | |||
1517 | ||||
1518 | // Track if the operand must be marked as used. If the operand is set to | |||
1519 | // Score 1 explicitly (because of non power-of-2 unique scalars, we may | |||
1520 | // want to reestimate the operands again on the following iterations). | |||
1521 | bool IsUsed = | |||
1522 | RMode == ReorderingMode::Splat || RMode == ReorderingMode::Constant; | |||
1523 | // Iterate through all unused operands and look for the best. | |||
1524 | for (unsigned Idx = 0; Idx != NumOperands; ++Idx) { | |||
1525 | // Get the operand at Idx and Lane. | |||
1526 | OperandData &OpData = getData(Idx, Lane); | |||
1527 | Value *Op = OpData.V; | |||
1528 | bool OpAPO = OpData.APO; | |||
1529 | ||||
1530 | // Skip already selected operands. | |||
1531 | if (OpData.IsUsed) | |||
1532 | continue; | |||
1533 | ||||
1534 | // Skip if we are trying to move the operand to a position with a | |||
1535 | // different opcode in the linearized tree form. This would break the | |||
1536 | // semantics. | |||
1537 | if (OpAPO != OpIdxAPO) | |||
1538 | continue; | |||
1539 | ||||
1540 | // Look for an operand that matches the current mode. | |||
1541 | switch (RMode) { | |||
1542 | case ReorderingMode::Load: | |||
1543 | case ReorderingMode::Constant: | |||
1544 | case ReorderingMode::Opcode: { | |||
1545 | bool LeftToRight = Lane > LastLane; | |||
1546 | Value *OpLeft = (LeftToRight) ? OpLastLane : Op; | |||
1547 | Value *OpRight = (LeftToRight) ? Op : OpLastLane; | |||
1548 | int Score = getLookAheadScore(OpLeft, OpRight, MainAltOps, Lane, | |||
1549 | OpIdx, Idx, IsUsed); | |||
1550 | if (Score > static_cast<int>(BestOp.Score)) { | |||
1551 | BestOp.Idx = Idx; | |||
1552 | BestOp.Score = Score; | |||
1553 | BestScoresPerLanes[std::make_pair(OpIdx, Lane)] = Score; | |||
1554 | } | |||
1555 | break; | |||
1556 | } | |||
1557 | case ReorderingMode::Splat: | |||
1558 | if (Op == OpLastLane) | |||
1559 | BestOp.Idx = Idx; | |||
1560 | break; | |||
1561 | case ReorderingMode::Failed: | |||
1562 | llvm_unreachable("Not expected Failed reordering mode.")::llvm::llvm_unreachable_internal("Not expected Failed reordering mode." , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 1562); | |||
1563 | } | |||
1564 | } | |||
1565 | ||||
1566 | if (BestOp.Idx) { | |||
1567 | getData(BestOp.Idx.getValue(), Lane).IsUsed = IsUsed; | |||
1568 | return BestOp.Idx; | |||
1569 | } | |||
1570 | // If we could not find a good match return None. | |||
1571 | return None; | |||
1572 | } | |||
1573 | ||||
1574 | /// Helper for reorderOperandVecs. | |||
1575 | /// \returns the lane that we should start reordering from. This is the one | |||
1576 | /// which has the least number of operands that can freely move about or | |||
1577 | /// less profitable because it already has the most optimal set of operands. | |||
1578 | unsigned getBestLaneToStartReordering() const { | |||
1579 | unsigned Min = UINT_MAX(2147483647 *2U +1U); | |||
1580 | unsigned SameOpNumber = 0; | |||
1581 | // std::pair<unsigned, unsigned> is used to implement a simple voting | |||
1582 | // algorithm and choose the lane with the least number of operands that | |||
1583 | // can freely move about or less profitable because it already has the | |||
1584 | // most optimal set of operands. The first unsigned is a counter for | |||
1585 | // voting, the second unsigned is the counter of lanes with instructions | |||
1586 | // with same/alternate opcodes and same parent basic block. | |||
1587 | MapVector<unsigned, std::pair<unsigned, unsigned>> HashMap; | |||
1588 | // Try to be closer to the original results, if we have multiple lanes | |||
1589 | // with same cost. If 2 lanes have the same cost, use the one with the | |||
1590 | // lowest index. | |||
1591 | for (int I = getNumLanes(); I > 0; --I) { | |||
1592 | unsigned Lane = I - 1; | |||
1593 | OperandsOrderData NumFreeOpsHash = | |||
1594 | getMaxNumOperandsThatCanBeReordered(Lane); | |||
1595 | // Compare the number of operands that can move and choose the one with | |||
1596 | // the least number. | |||
1597 | if (NumFreeOpsHash.NumOfAPOs < Min) { | |||
1598 | Min = NumFreeOpsHash.NumOfAPOs; | |||
1599 | SameOpNumber = NumFreeOpsHash.NumOpsWithSameOpcodeParent; | |||
1600 | HashMap.clear(); | |||
1601 | HashMap[NumFreeOpsHash.Hash] = std::make_pair(1, Lane); | |||
1602 | } else if (NumFreeOpsHash.NumOfAPOs == Min && | |||
1603 | NumFreeOpsHash.NumOpsWithSameOpcodeParent < SameOpNumber) { | |||
1604 | // Select the most optimal lane in terms of number of operands that | |||
1605 | // should be moved around. | |||
1606 | SameOpNumber = NumFreeOpsHash.NumOpsWithSameOpcodeParent; | |||
1607 | HashMap[NumFreeOpsHash.Hash] = std::make_pair(1, Lane); | |||
1608 | } else if (NumFreeOpsHash.NumOfAPOs == Min && | |||
1609 | NumFreeOpsHash.NumOpsWithSameOpcodeParent == SameOpNumber) { | |||
1610 | auto It = HashMap.find(NumFreeOpsHash.Hash); | |||
1611 | if (It == HashMap.end()) | |||
1612 | HashMap[NumFreeOpsHash.Hash] = std::make_pair(1, Lane); | |||
1613 | else | |||
1614 | ++It->second.first; | |||
1615 | } | |||
1616 | } | |||
1617 | // Select the lane with the minimum counter. | |||
1618 | unsigned BestLane = 0; | |||
1619 | unsigned CntMin = UINT_MAX(2147483647 *2U +1U); | |||
1620 | for (const auto &Data : reverse(HashMap)) { | |||
1621 | if (Data.second.first < CntMin) { | |||
1622 | CntMin = Data.second.first; | |||
1623 | BestLane = Data.second.second; | |||
1624 | } | |||
1625 | } | |||
1626 | return BestLane; | |||
1627 | } | |||
1628 | ||||
1629 | /// Data structure that helps to reorder operands. | |||
1630 | struct OperandsOrderData { | |||
1631 | /// The best number of operands with the same APOs, which can be | |||
1632 | /// reordered. | |||
1633 | unsigned NumOfAPOs = UINT_MAX(2147483647 *2U +1U); | |||
1634 | /// Number of operands with the same/alternate instruction opcode and | |||
1635 | /// parent. | |||
1636 | unsigned NumOpsWithSameOpcodeParent = 0; | |||
1637 | /// Hash for the actual operands ordering. | |||
1638 | /// Used to count operands, actually their position id and opcode | |||
1639 | /// value. It is used in the voting mechanism to find the lane with the | |||
1640 | /// least number of operands that can freely move about or less profitable | |||
1641 | /// because it already has the most optimal set of operands. Can be | |||
1642 | /// replaced with SmallVector<unsigned> instead but hash code is faster | |||
1643 | /// and requires less memory. | |||
1644 | unsigned Hash = 0; | |||
1645 | }; | |||
1646 | /// \returns the maximum number of operands that are allowed to be reordered | |||
1647 | /// for \p Lane and the number of compatible instructions(with the same | |||
1648 | /// parent/opcode). This is used as a heuristic for selecting the first lane | |||
1649 | /// to start operand reordering. | |||
1650 | OperandsOrderData getMaxNumOperandsThatCanBeReordered(unsigned Lane) const { | |||
1651 | unsigned CntTrue = 0; | |||
1652 | unsigned NumOperands = getNumOperands(); | |||
1653 | // Operands with the same APO can be reordered. We therefore need to count | |||
1654 | // how many of them we have for each APO, like this: Cnt[APO] = x. | |||
1655 | // Since we only have two APOs, namely true and false, we can avoid using | |||
1656 | // a map. Instead we can simply count the number of operands that | |||
1657 | // correspond to one of them (in this case the 'true' APO), and calculate | |||
1658 | // the other by subtracting it from the total number of operands. | |||
1659 | // Operands with the same instruction opcode and parent are more | |||
1660 | // profitable since we don't need to move them in many cases, with a high | |||
1661 | // probability such lane already can be vectorized effectively. | |||
1662 | bool AllUndefs = true; | |||
1663 | unsigned NumOpsWithSameOpcodeParent = 0; | |||
1664 | Instruction *OpcodeI = nullptr; | |||
1665 | BasicBlock *Parent = nullptr; | |||
1666 | unsigned Hash = 0; | |||
1667 | for (unsigned OpIdx = 0; OpIdx != NumOperands; ++OpIdx) { | |||
1668 | const OperandData &OpData = getData(OpIdx, Lane); | |||
1669 | if (OpData.APO) | |||
1670 | ++CntTrue; | |||
1671 | // Use Boyer-Moore majority voting for finding the majority opcode and | |||
1672 | // the number of times it occurs. | |||
1673 | if (auto *I = dyn_cast<Instruction>(OpData.V)) { | |||
1674 | if (!OpcodeI || !getSameOpcode({OpcodeI, I}).getOpcode() || | |||
1675 | I->getParent() != Parent) { | |||
1676 | if (NumOpsWithSameOpcodeParent == 0) { | |||
1677 | NumOpsWithSameOpcodeParent = 1; | |||
1678 | OpcodeI = I; | |||
1679 | Parent = I->getParent(); | |||
1680 | } else { | |||
1681 | --NumOpsWithSameOpcodeParent; | |||
1682 | } | |||
1683 | } else { | |||
1684 | ++NumOpsWithSameOpcodeParent; | |||
1685 | } | |||
1686 | } | |||
1687 | Hash = hash_combine( | |||
1688 | Hash, hash_value((OpIdx + 1) * (OpData.V->getValueID() + 1))); | |||
1689 | AllUndefs = AllUndefs && isa<UndefValue>(OpData.V); | |||
1690 | } | |||
1691 | if (AllUndefs) | |||
1692 | return {}; | |||
1693 | OperandsOrderData Data; | |||
1694 | Data.NumOfAPOs = std::max(CntTrue, NumOperands - CntTrue); | |||
1695 | Data.NumOpsWithSameOpcodeParent = NumOpsWithSameOpcodeParent; | |||
1696 | Data.Hash = Hash; | |||
1697 | return Data; | |||
1698 | } | |||
1699 | ||||
1700 | /// Go through the instructions in VL and append their operands. | |||
1701 | void appendOperandsOfVL(ArrayRef<Value *> VL) { | |||
1702 | assert(!VL.empty() && "Bad VL")(static_cast <bool> (!VL.empty() && "Bad VL") ? void (0) : __assert_fail ("!VL.empty() && \"Bad VL\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 1702, __extension__ __PRETTY_FUNCTION__)); | |||
1703 | assert((empty() || VL.size() == getNumLanes()) &&(static_cast <bool> ((empty() || VL.size() == getNumLanes ()) && "Expected same number of lanes") ? void (0) : __assert_fail ("(empty() || VL.size() == getNumLanes()) && \"Expected same number of lanes\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 1704, __extension__ __PRETTY_FUNCTION__)) | |||
1704 | "Expected same number of lanes")(static_cast <bool> ((empty() || VL.size() == getNumLanes ()) && "Expected same number of lanes") ? void (0) : __assert_fail ("(empty() || VL.size() == getNumLanes()) && \"Expected same number of lanes\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 1704, __extension__ __PRETTY_FUNCTION__)); | |||
1705 | assert(isa<Instruction>(VL[0]) && "Expected instruction")(static_cast <bool> (isa<Instruction>(VL[0]) && "Expected instruction") ? void (0) : __assert_fail ("isa<Instruction>(VL[0]) && \"Expected instruction\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 1705, __extension__ __PRETTY_FUNCTION__)); | |||
1706 | unsigned NumOperands = cast<Instruction>(VL[0])->getNumOperands(); | |||
1707 | OpsVec.resize(NumOperands); | |||
1708 | unsigned NumLanes = VL.size(); | |||
1709 | for (unsigned OpIdx = 0; OpIdx != NumOperands; ++OpIdx) { | |||
1710 | OpsVec[OpIdx].resize(NumLanes); | |||
1711 | for (unsigned Lane = 0; Lane != NumLanes; ++Lane) { | |||
1712 | assert(isa<Instruction>(VL[Lane]) && "Expected instruction")(static_cast <bool> (isa<Instruction>(VL[Lane]) && "Expected instruction") ? void (0) : __assert_fail ("isa<Instruction>(VL[Lane]) && \"Expected instruction\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 1712, __extension__ __PRETTY_FUNCTION__)); | |||
1713 | // Our tree has just 3 nodes: the root and two operands. | |||
1714 | // It is therefore trivial to get the APO. We only need to check the | |||
1715 | // opcode of VL[Lane] and whether the operand at OpIdx is the LHS or | |||
1716 | // RHS operand. The LHS operand of both add and sub is never attached | |||
1717 | // to an inversese operation in the linearized form, therefore its APO | |||
1718 | // is false. The RHS is true only if VL[Lane] is an inverse operation. | |||
1719 | ||||
1720 | // Since operand reordering is performed on groups of commutative | |||
1721 | // operations or alternating sequences (e.g., +, -), we can safely | |||
1722 | // tell the inverse operations by checking commutativity. | |||
1723 | bool IsInverseOperation = !isCommutative(cast<Instruction>(VL[Lane])); | |||
1724 | bool APO = (OpIdx == 0) ? false : IsInverseOperation; | |||
1725 | OpsVec[OpIdx][Lane] = {cast<Instruction>(VL[Lane])->getOperand(OpIdx), | |||
1726 | APO, false}; | |||
1727 | } | |||
1728 | } | |||
1729 | } | |||
1730 | ||||
1731 | /// \returns the number of operands. | |||
1732 | unsigned getNumOperands() const { return OpsVec.size(); } | |||
1733 | ||||
1734 | /// \returns the number of lanes. | |||
1735 | unsigned getNumLanes() const { return OpsVec[0].size(); } | |||
1736 | ||||
1737 | /// \returns the operand value at \p OpIdx and \p Lane. | |||
1738 | Value *getValue(unsigned OpIdx, unsigned Lane) const { | |||
1739 | return getData(OpIdx, Lane).V; | |||
1740 | } | |||
1741 | ||||
1742 | /// \returns true if the data structure is empty. | |||
1743 | bool empty() const { return OpsVec.empty(); } | |||
1744 | ||||
1745 | /// Clears the data. | |||
1746 | void clear() { OpsVec.clear(); } | |||
1747 | ||||
1748 | /// \Returns true if there are enough operands identical to \p Op to fill | |||
1749 | /// the whole vector. | |||
1750 | /// Note: This modifies the 'IsUsed' flag, so a cleanUsed() must follow. | |||
1751 | bool shouldBroadcast(Value *Op, unsigned OpIdx, unsigned Lane) { | |||
1752 | bool OpAPO = getData(OpIdx, Lane).APO; | |||
1753 | for (unsigned Ln = 0, Lns = getNumLanes(); Ln != Lns; ++Ln) { | |||
1754 | if (Ln == Lane) | |||
1755 | continue; | |||
1756 | // This is set to true if we found a candidate for broadcast at Lane. | |||
1757 | bool FoundCandidate = false; | |||
1758 | for (unsigned OpI = 0, OpE = getNumOperands(); OpI != OpE; ++OpI) { | |||
1759 | OperandData &Data = getData(OpI, Ln); | |||
1760 | if (Data.APO != OpAPO || Data.IsUsed) | |||
1761 | continue; | |||
1762 | if (Data.V == Op) { | |||
1763 | FoundCandidate = true; | |||
1764 | Data.IsUsed = true; | |||
1765 | break; | |||
1766 | } | |||
1767 | } | |||
1768 | if (!FoundCandidate) | |||
1769 | return false; | |||
1770 | } | |||
1771 | return true; | |||
1772 | } | |||
1773 | ||||
1774 | public: | |||
1775 | /// Initialize with all the operands of the instruction vector \p RootVL. | |||
1776 | VLOperands(ArrayRef<Value *> RootVL, const DataLayout &DL, | |||
1777 | ScalarEvolution &SE, const BoUpSLP &R) | |||
1778 | : DL(DL), SE(SE), R(R) { | |||
1779 | // Append all the operands of RootVL. | |||
1780 | appendOperandsOfVL(RootVL); | |||
1781 | } | |||
1782 | ||||
1783 | /// \Returns a value vector with the operands across all lanes for the | |||
1784 | /// opearnd at \p OpIdx. | |||
1785 | ValueList getVL(unsigned OpIdx) const { | |||
1786 | ValueList OpVL(OpsVec[OpIdx].size()); | |||
1787 | assert(OpsVec[OpIdx].size() == getNumLanes() &&(static_cast <bool> (OpsVec[OpIdx].size() == getNumLanes () && "Expected same num of lanes across all operands" ) ? void (0) : __assert_fail ("OpsVec[OpIdx].size() == getNumLanes() && \"Expected same num of lanes across all operands\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 1788, __extension__ __PRETTY_FUNCTION__)) | |||
1788 | "Expected same num of lanes across all operands")(static_cast <bool> (OpsVec[OpIdx].size() == getNumLanes () && "Expected same num of lanes across all operands" ) ? void (0) : __assert_fail ("OpsVec[OpIdx].size() == getNumLanes() && \"Expected same num of lanes across all operands\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 1788, __extension__ __PRETTY_FUNCTION__)); | |||
1789 | for (unsigned Lane = 0, Lanes = getNumLanes(); Lane != Lanes; ++Lane) | |||
1790 | OpVL[Lane] = OpsVec[OpIdx][Lane].V; | |||
1791 | return OpVL; | |||
1792 | } | |||
1793 | ||||
1794 | // Performs operand reordering for 2 or more operands. | |||
1795 | // The original operands are in OrigOps[OpIdx][Lane]. | |||
1796 | // The reordered operands are returned in 'SortedOps[OpIdx][Lane]'. | |||
1797 | void reorder() { | |||
1798 | unsigned NumOperands = getNumOperands(); | |||
1799 | unsigned NumLanes = getNumLanes(); | |||
1800 | // Each operand has its own mode. We are using this mode to help us select | |||
1801 | // the instructions for each lane, so that they match best with the ones | |||
1802 | // we have selected so far. | |||
1803 | SmallVector<ReorderingMode, 2> ReorderingModes(NumOperands); | |||
1804 | ||||
1805 | // This is a greedy single-pass algorithm. We are going over each lane | |||
1806 | // once and deciding on the best order right away with no back-tracking. | |||
1807 | // However, in order to increase its effectiveness, we start with the lane | |||
1808 | // that has operands that can move the least. For example, given the | |||
1809 | // following lanes: | |||
1810 | // Lane 0 : A[0] = B[0] + C[0] // Visited 3rd | |||
1811 | // Lane 1 : A[1] = C[1] - B[1] // Visited 1st | |||
1812 | // Lane 2 : A[2] = B[2] + C[2] // Visited 2nd | |||
1813 | // Lane 3 : A[3] = C[3] - B[3] // Visited 4th | |||
1814 | // we will start at Lane 1, since the operands of the subtraction cannot | |||
1815 | // be reordered. Then we will visit the rest of the lanes in a circular | |||
1816 | // fashion. That is, Lanes 2, then Lane 0, and finally Lane 3. | |||
1817 | ||||
1818 | // Find the first lane that we will start our search from. | |||
1819 | unsigned FirstLane = getBestLaneToStartReordering(); | |||
1820 | ||||
1821 | // Initialize the modes. | |||
1822 | for (unsigned OpIdx = 0; OpIdx != NumOperands; ++OpIdx) { | |||
1823 | Value *OpLane0 = getValue(OpIdx, FirstLane); | |||
1824 | // Keep track if we have instructions with all the same opcode on one | |||
1825 | // side. | |||
1826 | if (isa<LoadInst>(OpLane0)) | |||
1827 | ReorderingModes[OpIdx] = ReorderingMode::Load; | |||
1828 | else if (isa<Instruction>(OpLane0)) { | |||
1829 | // Check if OpLane0 should be broadcast. | |||
1830 | if (shouldBroadcast(OpLane0, OpIdx, FirstLane)) | |||
1831 | ReorderingModes[OpIdx] = ReorderingMode::Splat; | |||
1832 | else | |||
1833 | ReorderingModes[OpIdx] = ReorderingMode::Opcode; | |||
1834 | } | |||
1835 | else if (isa<Constant>(OpLane0)) | |||
1836 | ReorderingModes[OpIdx] = ReorderingMode::Constant; | |||
1837 | else if (isa<Argument>(OpLane0)) | |||
1838 | // Our best hope is a Splat. It may save some cost in some cases. | |||
1839 | ReorderingModes[OpIdx] = ReorderingMode::Splat; | |||
1840 | else | |||
1841 | // NOTE: This should be unreachable. | |||
1842 | ReorderingModes[OpIdx] = ReorderingMode::Failed; | |||
1843 | } | |||
1844 | ||||
1845 | // Check that we don't have same operands. No need to reorder if operands | |||
1846 | // are just perfect diamond or shuffled diamond match. Do not do it only | |||
1847 | // for possible broadcasts or non-power of 2 number of scalars (just for | |||
1848 | // now). | |||
1849 | auto &&SkipReordering = [this]() { | |||
1850 | SmallPtrSet<Value *, 4> UniqueValues; | |||
1851 | ArrayRef<OperandData> Op0 = OpsVec.front(); | |||
1852 | for (const OperandData &Data : Op0) | |||
1853 | UniqueValues.insert(Data.V); | |||
1854 | for (ArrayRef<OperandData> Op : drop_begin(OpsVec, 1)) { | |||
1855 | if (any_of(Op, [&UniqueValues](const OperandData &Data) { | |||
1856 | return !UniqueValues.contains(Data.V); | |||
1857 | })) | |||
1858 | return false; | |||
1859 | } | |||
1860 | // TODO: Check if we can remove a check for non-power-2 number of | |||
1861 | // scalars after full support of non-power-2 vectorization. | |||
1862 | return UniqueValues.size() != 2 && isPowerOf2_32(UniqueValues.size()); | |||
1863 | }; | |||
1864 | ||||
1865 | // If the initial strategy fails for any of the operand indexes, then we | |||
1866 | // perform reordering again in a second pass. This helps avoid assigning | |||
1867 | // high priority to the failed strategy, and should improve reordering for | |||
1868 | // the non-failed operand indexes. | |||
1869 | for (int Pass = 0; Pass != 2; ++Pass) { | |||
1870 | // Check if no need to reorder operands since they're are perfect or | |||
1871 | // shuffled diamond match. | |||
1872 | // Need to to do it to avoid extra external use cost counting for | |||
1873 | // shuffled matches, which may cause regressions. | |||
1874 | if (SkipReordering()) | |||
1875 | break; | |||
1876 | // Skip the second pass if the first pass did not fail. | |||
1877 | bool StrategyFailed = false; | |||
1878 | // Mark all operand data as free to use. | |||
1879 | clearUsed(); | |||
1880 | // We keep the original operand order for the FirstLane, so reorder the | |||
1881 | // rest of the lanes. We are visiting the nodes in a circular fashion, | |||
1882 | // using FirstLane as the center point and increasing the radius | |||
1883 | // distance. | |||
1884 | SmallVector<SmallVector<Value *, 2>> MainAltOps(NumOperands); | |||
1885 | for (unsigned I = 0; I < NumOperands; ++I) | |||
1886 | MainAltOps[I].push_back(getData(I, FirstLane).V); | |||
1887 | ||||
1888 | for (unsigned Distance = 1; Distance != NumLanes; ++Distance) { | |||
1889 | // Visit the lane on the right and then the lane on the left. | |||
1890 | for (int Direction : {+1, -1}) { | |||
1891 | int Lane = FirstLane + Direction * Distance; | |||
1892 | if (Lane < 0 || Lane >= (int)NumLanes) | |||
1893 | continue; | |||
1894 | int LastLane = Lane - Direction; | |||
1895 | assert(LastLane >= 0 && LastLane < (int)NumLanes &&(static_cast <bool> (LastLane >= 0 && LastLane < (int)NumLanes && "Out of bounds") ? void (0) : __assert_fail ("LastLane >= 0 && LastLane < (int)NumLanes && \"Out of bounds\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 1896, __extension__ __PRETTY_FUNCTION__)) | |||
1896 | "Out of bounds")(static_cast <bool> (LastLane >= 0 && LastLane < (int)NumLanes && "Out of bounds") ? void (0) : __assert_fail ("LastLane >= 0 && LastLane < (int)NumLanes && \"Out of bounds\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 1896, __extension__ __PRETTY_FUNCTION__)); | |||
1897 | // Look for a good match for each operand. | |||
1898 | for (unsigned OpIdx = 0; OpIdx != NumOperands; ++OpIdx) { | |||
1899 | // Search for the operand that matches SortedOps[OpIdx][Lane-1]. | |||
1900 | Optional<unsigned> BestIdx = getBestOperand( | |||
1901 | OpIdx, Lane, LastLane, ReorderingModes, MainAltOps[OpIdx]); | |||
1902 | // By not selecting a value, we allow the operands that follow to | |||
1903 | // select a better matching value. We will get a non-null value in | |||
1904 | // the next run of getBestOperand(). | |||
1905 | if (BestIdx) { | |||
1906 | // Swap the current operand with the one returned by | |||
1907 | // getBestOperand(). | |||
1908 | swap(OpIdx, BestIdx.getValue(), Lane); | |||
1909 | } else { | |||
1910 | // We failed to find a best operand, set mode to 'Failed'. | |||
1911 | ReorderingModes[OpIdx] = ReorderingMode::Failed; | |||
1912 | // Enable the second pass. | |||
1913 | StrategyFailed = true; | |||
1914 | } | |||
1915 | // Try to get the alternate opcode and follow it during analysis. | |||
1916 | if (MainAltOps[OpIdx].size() != 2) { | |||
1917 | OperandData &AltOp = getData(OpIdx, Lane); | |||
1918 | InstructionsState OpS = | |||
1919 | getSameOpcode({MainAltOps[OpIdx].front(), AltOp.V}); | |||
1920 | if (OpS.getOpcode() && OpS.isAltShuffle()) | |||
1921 | MainAltOps[OpIdx].push_back(AltOp.V); | |||
1922 | } | |||
1923 | } | |||
1924 | } | |||
1925 | } | |||
1926 | // Skip second pass if the strategy did not fail. | |||
1927 | if (!StrategyFailed) | |||
1928 | break; | |||
1929 | } | |||
1930 | } | |||
1931 | ||||
1932 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | |||
1933 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) static StringRef getModeStr(ReorderingMode RMode) { | |||
1934 | switch (RMode) { | |||
1935 | case ReorderingMode::Load: | |||
1936 | return "Load"; | |||
1937 | case ReorderingMode::Opcode: | |||
1938 | return "Opcode"; | |||
1939 | case ReorderingMode::Constant: | |||
1940 | return "Constant"; | |||
1941 | case ReorderingMode::Splat: | |||
1942 | return "Splat"; | |||
1943 | case ReorderingMode::Failed: | |||
1944 | return "Failed"; | |||
1945 | } | |||
1946 | llvm_unreachable("Unimplemented Reordering Type")::llvm::llvm_unreachable_internal("Unimplemented Reordering Type" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 1946); | |||
1947 | } | |||
1948 | ||||
1949 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) static raw_ostream &printMode(ReorderingMode RMode, | |||
1950 | raw_ostream &OS) { | |||
1951 | return OS << getModeStr(RMode); | |||
1952 | } | |||
1953 | ||||
1954 | /// Debug print. | |||
1955 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) static void dumpMode(ReorderingMode RMode) { | |||
1956 | printMode(RMode, dbgs()); | |||
1957 | } | |||
1958 | ||||
1959 | friend raw_ostream &operator<<(raw_ostream &OS, ReorderingMode RMode) { | |||
1960 | return printMode(RMode, OS); | |||
1961 | } | |||
1962 | ||||
1963 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) raw_ostream &print(raw_ostream &OS) const { | |||
1964 | const unsigned Indent = 2; | |||
1965 | unsigned Cnt = 0; | |||
1966 | for (const OperandDataVec &OpDataVec : OpsVec) { | |||
1967 | OS << "Operand " << Cnt++ << "\n"; | |||
1968 | for (const OperandData &OpData : OpDataVec) { | |||
1969 | OS.indent(Indent) << "{"; | |||
1970 | if (Value *V = OpData.V) | |||
1971 | OS << *V; | |||
1972 | else | |||
1973 | OS << "null"; | |||
1974 | OS << ", APO:" << OpData.APO << "}\n"; | |||
1975 | } | |||
1976 | OS << "\n"; | |||
1977 | } | |||
1978 | return OS; | |||
1979 | } | |||
1980 | ||||
1981 | /// Debug print. | |||
1982 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void dump() const { print(dbgs()); } | |||
1983 | #endif | |||
1984 | }; | |||
1985 | ||||
1986 | /// Checks if the instruction is marked for deletion. | |||
1987 | bool isDeleted(Instruction *I) const { return DeletedInstructions.count(I); } | |||
1988 | ||||
1989 | /// Removes an instruction from its block and eventually deletes it. | |||
1990 | /// It's like Instruction::eraseFromParent() except that the actual deletion | |||
1991 | /// is delayed until BoUpSLP is destructed. | |||
1992 | void eraseInstruction(Instruction *I) { | |||
1993 | DeletedInstructions.insert(I); | |||
1994 | } | |||
1995 | ||||
1996 | ~BoUpSLP(); | |||
1997 | ||||
1998 | private: | |||
1999 | /// Check if the operands on the edges \p Edges of the \p UserTE allows | |||
2000 | /// reordering (i.e. the operands can be reordered because they have only one | |||
2001 | /// user and reordarable). | |||
2002 | /// \param ReorderableGathers List of all gather nodes that require reordering | |||
2003 | /// (e.g., gather of extractlements or partially vectorizable loads). | |||
2004 | /// \param GatherOps List of gather operand nodes for \p UserTE that require | |||
2005 | /// reordering, subset of \p NonVectorized. | |||
2006 | bool | |||
2007 | canReorderOperands(TreeEntry *UserTE, | |||
2008 | SmallVectorImpl<std::pair<unsigned, TreeEntry *>> &Edges, | |||
2009 | ArrayRef<TreeEntry *> ReorderableGathers, | |||
2010 | SmallVectorImpl<TreeEntry *> &GatherOps); | |||
2011 | ||||
2012 | /// Returns vectorized operand \p OpIdx of the node \p UserTE from the graph, | |||
2013 | /// if any. If it is not vectorized (gather node), returns nullptr. | |||
2014 | TreeEntry *getVectorizedOperand(TreeEntry *UserTE, unsigned OpIdx) { | |||
2015 | ArrayRef<Value *> VL = UserTE->getOperand(OpIdx); | |||
2016 | TreeEntry *TE = nullptr; | |||
2017 | const auto *It = find_if(VL, [this, &TE](Value *V) { | |||
2018 | TE = getTreeEntry(V); | |||
2019 | return TE; | |||
2020 | }); | |||
2021 | if (It != VL.end() && TE->isSame(VL)) | |||
2022 | return TE; | |||
2023 | return nullptr; | |||
2024 | } | |||
2025 | ||||
2026 | /// Returns vectorized operand \p OpIdx of the node \p UserTE from the graph, | |||
2027 | /// if any. If it is not vectorized (gather node), returns nullptr. | |||
2028 | const TreeEntry *getVectorizedOperand(const TreeEntry *UserTE, | |||
2029 | unsigned OpIdx) const { | |||
2030 | return const_cast<BoUpSLP *>(this)->getVectorizedOperand( | |||
2031 | const_cast<TreeEntry *>(UserTE), OpIdx); | |||
2032 | } | |||
2033 | ||||
2034 | /// Checks if all users of \p I are the part of the vectorization tree. | |||
2035 | bool areAllUsersVectorized(Instruction *I, | |||
2036 | ArrayRef<Value *> VectorizedVals) const; | |||
2037 | ||||
2038 | /// \returns the cost of the vectorizable entry. | |||
2039 | InstructionCost getEntryCost(const TreeEntry *E, | |||
2040 | ArrayRef<Value *> VectorizedVals); | |||
2041 | ||||
2042 | /// This is the recursive part of buildTree. | |||
2043 | void buildTree_rec(ArrayRef<Value *> Roots, unsigned Depth, | |||
2044 | const EdgeInfo &EI); | |||
2045 | ||||
2046 | /// \returns true if the ExtractElement/ExtractValue instructions in \p VL can | |||
2047 | /// be vectorized to use the original vector (or aggregate "bitcast" to a | |||
2048 | /// vector) and sets \p CurrentOrder to the identity permutation; otherwise | |||
2049 | /// returns false, setting \p CurrentOrder to either an empty vector or a | |||
2050 | /// non-identity permutation that allows to reuse extract instructions. | |||
2051 | bool canReuseExtract(ArrayRef<Value *> VL, Value *OpValue, | |||
2052 | SmallVectorImpl<unsigned> &CurrentOrder) const; | |||
2053 | ||||
2054 | /// Vectorize a single entry in the tree. | |||
2055 | Value *vectorizeTree(TreeEntry *E); | |||
2056 | ||||
2057 | /// Vectorize a single entry in the tree, starting in \p VL. | |||
2058 | Value *vectorizeTree(ArrayRef<Value *> VL); | |||
2059 | ||||
2060 | /// Create a new vector from a list of scalar values. Produces a sequence | |||
2061 | /// which exploits values reused across lanes, and arranges the inserts | |||
2062 | /// for ease of later optimization. | |||
2063 | Value *createBuildVector(ArrayRef<Value *> VL); | |||
2064 | ||||
2065 | /// \returns the scalarization cost for this type. Scalarization in this | |||
2066 | /// context means the creation of vectors from a group of scalars. If \p | |||
2067 | /// NeedToShuffle is true, need to add a cost of reshuffling some of the | |||
2068 | /// vector elements. | |||
2069 | InstructionCost getGatherCost(FixedVectorType *Ty, | |||
2070 | const APInt &ShuffledIndices, | |||
2071 | bool NeedToShuffle) const; | |||
2072 | ||||
2073 | /// Checks if the gathered \p VL can be represented as shuffle(s) of previous | |||
2074 | /// tree entries. | |||
2075 | /// \returns ShuffleKind, if gathered values can be represented as shuffles of | |||
2076 | /// previous tree entries. \p Mask is filled with the shuffle mask. | |||
2077 | Optional<TargetTransformInfo::ShuffleKind> | |||
2078 | isGatherShuffledEntry(const TreeEntry *TE, SmallVectorImpl<int> &Mask, | |||
2079 | SmallVectorImpl<const TreeEntry *> &Entries); | |||
2080 | ||||
2081 | /// \returns the scalarization cost for this list of values. Assuming that | |||
2082 | /// this subtree gets vectorized, we may need to extract the values from the | |||
2083 | /// roots. This method calculates the cost of extracting the values. | |||
2084 | InstructionCost getGatherCost(ArrayRef<Value *> VL) const; | |||
2085 | ||||
2086 | /// Set the Builder insert point to one after the last instruction in | |||
2087 | /// the bundle | |||
2088 | void setInsertPointAfterBundle(const TreeEntry *E); | |||
2089 | ||||
2090 | /// \returns a vector from a collection of scalars in \p VL. | |||
2091 | Value *gather(ArrayRef<Value *> VL); | |||
2092 | ||||
2093 | /// \returns whether the VectorizableTree is fully vectorizable and will | |||
2094 | /// be beneficial even the tree height is tiny. | |||
2095 | bool isFullyVectorizableTinyTree(bool ForReduction) const; | |||
2096 | ||||
2097 | /// Reorder commutative or alt operands to get better probability of | |||
2098 | /// generating vectorized code. | |||
2099 | static void reorderInputsAccordingToOpcode(ArrayRef<Value *> VL, | |||
2100 | SmallVectorImpl<Value *> &Left, | |||
2101 | SmallVectorImpl<Value *> &Right, | |||
2102 | const DataLayout &DL, | |||
2103 | ScalarEvolution &SE, | |||
2104 | const BoUpSLP &R); | |||
2105 | struct TreeEntry { | |||
2106 | using VecTreeTy = SmallVector<std::unique_ptr<TreeEntry>, 8>; | |||
2107 | TreeEntry(VecTreeTy &Container) : Container(Container) {} | |||
2108 | ||||
2109 | /// \returns true if the scalars in VL are equal to this entry. | |||
2110 | bool isSame(ArrayRef<Value *> VL) const { | |||
2111 | auto &&IsSame = [VL](ArrayRef<Value *> Scalars, ArrayRef<int> Mask) { | |||
2112 | if (Mask.size() != VL.size() && VL.size() == Scalars.size()) | |||
2113 | return std::equal(VL.begin(), VL.end(), Scalars.begin()); | |||
2114 | return VL.size() == Mask.size() && | |||
2115 | std::equal(VL.begin(), VL.end(), Mask.begin(), | |||
2116 | [Scalars](Value *V, int Idx) { | |||
2117 | return (isa<UndefValue>(V) && | |||
2118 | Idx == UndefMaskElem) || | |||
2119 | (Idx != UndefMaskElem && V == Scalars[Idx]); | |||
2120 | }); | |||
2121 | }; | |||
2122 | if (!ReorderIndices.empty()) { | |||
2123 | // TODO: implement matching if the nodes are just reordered, still can | |||
2124 | // treat the vector as the same if the list of scalars matches VL | |||
2125 | // directly, without reordering. | |||
2126 | SmallVector<int> Mask; | |||
2127 | inversePermutation(ReorderIndices, Mask); | |||
2128 | if (VL.size() == Scalars.size()) | |||
2129 | return IsSame(Scalars, Mask); | |||
2130 | if (VL.size() == ReuseShuffleIndices.size()) { | |||
2131 | ::addMask(Mask, ReuseShuffleIndices); | |||
2132 | return IsSame(Scalars, Mask); | |||
2133 | } | |||
2134 | return false; | |||
2135 | } | |||
2136 | return IsSame(Scalars, ReuseShuffleIndices); | |||
2137 | } | |||
2138 | ||||
2139 | /// \returns true if current entry has same operands as \p TE. | |||
2140 | bool hasEqualOperands(const TreeEntry &TE) const { | |||
2141 | if (TE.getNumOperands() != getNumOperands()) | |||
2142 | return false; | |||
2143 | SmallBitVector Used(getNumOperands()); | |||
2144 | for (unsigned I = 0, E = getNumOperands(); I < E; ++I) { | |||
2145 | unsigned PrevCount = Used.count(); | |||
2146 | for (unsigned K = 0; K < E; ++K) { | |||
2147 | if (Used.test(K)) | |||
2148 | continue; | |||
2149 | if (getOperand(K) == TE.getOperand(I)) { | |||
2150 | Used.set(K); | |||
2151 | break; | |||
2152 | } | |||
2153 | } | |||
2154 | // Check if we actually found the matching operand. | |||
2155 | if (PrevCount == Used.count()) | |||
2156 | return false; | |||
2157 | } | |||
2158 | return true; | |||
2159 | } | |||
2160 | ||||
2161 | /// \return Final vectorization factor for the node. Defined by the total | |||
2162 | /// number of vectorized scalars, including those, used several times in the | |||
2163 | /// entry and counted in the \a ReuseShuffleIndices, if any. | |||
2164 | unsigned getVectorFactor() const { | |||
2165 | if (!ReuseShuffleIndices.empty()) | |||
2166 | return ReuseShuffleIndices.size(); | |||
2167 | return Scalars.size(); | |||
2168 | }; | |||
2169 | ||||
2170 | /// A vector of scalars. | |||
2171 | ValueList Scalars; | |||
2172 | ||||
2173 | /// The Scalars are vectorized into this value. It is initialized to Null. | |||
2174 | Value *VectorizedValue = nullptr; | |||
2175 | ||||
2176 | /// Do we need to gather this sequence or vectorize it | |||
2177 | /// (either with vector instruction or with scatter/gather | |||
2178 | /// intrinsics for store/load)? | |||
2179 | enum EntryState { Vectorize, ScatterVectorize, NeedToGather }; | |||
2180 | EntryState State; | |||
2181 | ||||
2182 | /// Does this sequence require some shuffling? | |||
2183 | SmallVector<int, 4> ReuseShuffleIndices; | |||
2184 | ||||
2185 | /// Does this entry require reordering? | |||
2186 | SmallVector<unsigned, 4> ReorderIndices; | |||
2187 | ||||
2188 | /// Points back to the VectorizableTree. | |||
2189 | /// | |||
2190 | /// Only used for Graphviz right now. Unfortunately GraphTrait::NodeRef has | |||
2191 | /// to be a pointer and needs to be able to initialize the child iterator. | |||
2192 | /// Thus we need a reference back to the container to translate the indices | |||
2193 | /// to entries. | |||
2194 | VecTreeTy &Container; | |||
2195 | ||||
2196 | /// The TreeEntry index containing the user of this entry. We can actually | |||
2197 | /// have multiple users so the data structure is not truly a tree. | |||
2198 | SmallVector<EdgeInfo, 1> UserTreeIndices; | |||
2199 | ||||
2200 | /// The index of this treeEntry in VectorizableTree. | |||
2201 | int Idx = -1; | |||
2202 | ||||
2203 | private: | |||
2204 | /// The operands of each instruction in each lane Operands[op_index][lane]. | |||
2205 | /// Note: This helps avoid the replication of the code that performs the | |||
2206 | /// reordering of operands during buildTree_rec() and vectorizeTree(). | |||
2207 | SmallVector<ValueList, 2> Operands; | |||
2208 | ||||
2209 | /// The main/alternate instruction. | |||
2210 | Instruction *MainOp = nullptr; | |||
2211 | Instruction *AltOp = nullptr; | |||
2212 | ||||
2213 | public: | |||
2214 | /// Set this bundle's \p OpIdx'th operand to \p OpVL. | |||
2215 | void setOperand(unsigned OpIdx, ArrayRef<Value *> OpVL) { | |||
2216 | if (Operands.size() < OpIdx + 1) | |||
2217 | Operands.resize(OpIdx + 1); | |||
2218 | assert(Operands[OpIdx].empty() && "Already resized?")(static_cast <bool> (Operands[OpIdx].empty() && "Already resized?") ? void (0) : __assert_fail ("Operands[OpIdx].empty() && \"Already resized?\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2218, __extension__ __PRETTY_FUNCTION__)); | |||
2219 | assert(OpVL.size() <= Scalars.size() &&(static_cast <bool> (OpVL.size() <= Scalars.size() && "Number of operands is greater than the number of scalars.") ? void (0) : __assert_fail ("OpVL.size() <= Scalars.size() && \"Number of operands is greater than the number of scalars.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2220, __extension__ __PRETTY_FUNCTION__)) | |||
2220 | "Number of operands is greater than the number of scalars.")(static_cast <bool> (OpVL.size() <= Scalars.size() && "Number of operands is greater than the number of scalars.") ? void (0) : __assert_fail ("OpVL.size() <= Scalars.size() && \"Number of operands is greater than the number of scalars.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2220, __extension__ __PRETTY_FUNCTION__)); | |||
2221 | Operands[OpIdx].resize(OpVL.size()); | |||
2222 | copy(OpVL, Operands[OpIdx].begin()); | |||
2223 | } | |||
2224 | ||||
2225 | /// Set the operands of this bundle in their original order. | |||
2226 | void setOperandsInOrder() { | |||
2227 | assert(Operands.empty() && "Already initialized?")(static_cast <bool> (Operands.empty() && "Already initialized?" ) ? void (0) : __assert_fail ("Operands.empty() && \"Already initialized?\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2227, __extension__ __PRETTY_FUNCTION__)); | |||
2228 | auto *I0 = cast<Instruction>(Scalars[0]); | |||
2229 | Operands.resize(I0->getNumOperands()); | |||
2230 | unsigned NumLanes = Scalars.size(); | |||
2231 | for (unsigned OpIdx = 0, NumOperands = I0->getNumOperands(); | |||
2232 | OpIdx != NumOperands; ++OpIdx) { | |||
2233 | Operands[OpIdx].resize(NumLanes); | |||
2234 | for (unsigned Lane = 0; Lane != NumLanes; ++Lane) { | |||
2235 | auto *I = cast<Instruction>(Scalars[Lane]); | |||
2236 | assert(I->getNumOperands() == NumOperands &&(static_cast <bool> (I->getNumOperands() == NumOperands && "Expected same number of operands") ? void (0) : __assert_fail ("I->getNumOperands() == NumOperands && \"Expected same number of operands\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2237, __extension__ __PRETTY_FUNCTION__)) | |||
2237 | "Expected same number of operands")(static_cast <bool> (I->getNumOperands() == NumOperands && "Expected same number of operands") ? void (0) : __assert_fail ("I->getNumOperands() == NumOperands && \"Expected same number of operands\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2237, __extension__ __PRETTY_FUNCTION__)); | |||
2238 | Operands[OpIdx][Lane] = I->getOperand(OpIdx); | |||
2239 | } | |||
2240 | } | |||
2241 | } | |||
2242 | ||||
2243 | /// Reorders operands of the node to the given mask \p Mask. | |||
2244 | void reorderOperands(ArrayRef<int> Mask) { | |||
2245 | for (ValueList &Operand : Operands) | |||
2246 | reorderScalars(Operand, Mask); | |||
2247 | } | |||
2248 | ||||
2249 | /// \returns the \p OpIdx operand of this TreeEntry. | |||
2250 | ValueList &getOperand(unsigned OpIdx) { | |||
2251 | assert(OpIdx < Operands.size() && "Off bounds")(static_cast <bool> (OpIdx < Operands.size() && "Off bounds") ? void (0) : __assert_fail ("OpIdx < Operands.size() && \"Off bounds\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2251, __extension__ __PRETTY_FUNCTION__)); | |||
2252 | return Operands[OpIdx]; | |||
2253 | } | |||
2254 | ||||
2255 | /// \returns the \p OpIdx operand of this TreeEntry. | |||
2256 | ArrayRef<Value *> getOperand(unsigned OpIdx) const { | |||
2257 | assert(OpIdx < Operands.size() && "Off bounds")(static_cast <bool> (OpIdx < Operands.size() && "Off bounds") ? void (0) : __assert_fail ("OpIdx < Operands.size() && \"Off bounds\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2257, __extension__ __PRETTY_FUNCTION__)); | |||
2258 | return Operands[OpIdx]; | |||
2259 | } | |||
2260 | ||||
2261 | /// \returns the number of operands. | |||
2262 | unsigned getNumOperands() const { return Operands.size(); } | |||
2263 | ||||
2264 | /// \return the single \p OpIdx operand. | |||
2265 | Value *getSingleOperand(unsigned OpIdx) const { | |||
2266 | assert(OpIdx < Operands.size() && "Off bounds")(static_cast <bool> (OpIdx < Operands.size() && "Off bounds") ? void (0) : __assert_fail ("OpIdx < Operands.size() && \"Off bounds\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2266, __extension__ __PRETTY_FUNCTION__)); | |||
2267 | assert(!Operands[OpIdx].empty() && "No operand available")(static_cast <bool> (!Operands[OpIdx].empty() && "No operand available") ? void (0) : __assert_fail ("!Operands[OpIdx].empty() && \"No operand available\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2267, __extension__ __PRETTY_FUNCTION__)); | |||
2268 | return Operands[OpIdx][0]; | |||
2269 | } | |||
2270 | ||||
2271 | /// Some of the instructions in the list have alternate opcodes. | |||
2272 | bool isAltShuffle() const { return MainOp != AltOp; } | |||
2273 | ||||
2274 | bool isOpcodeOrAlt(Instruction *I) const { | |||
2275 | unsigned CheckedOpcode = I->getOpcode(); | |||
2276 | return (getOpcode() == CheckedOpcode || | |||
2277 | getAltOpcode() == CheckedOpcode); | |||
2278 | } | |||
2279 | ||||
2280 | /// Chooses the correct key for scheduling data. If \p Op has the same (or | |||
2281 | /// alternate) opcode as \p OpValue, the key is \p Op. Otherwise the key is | |||
2282 | /// \p OpValue. | |||
2283 | Value *isOneOf(Value *Op) const { | |||
2284 | auto *I = dyn_cast<Instruction>(Op); | |||
2285 | if (I && isOpcodeOrAlt(I)) | |||
2286 | return Op; | |||
2287 | return MainOp; | |||
2288 | } | |||
2289 | ||||
2290 | void setOperations(const InstructionsState &S) { | |||
2291 | MainOp = S.MainOp; | |||
2292 | AltOp = S.AltOp; | |||
2293 | } | |||
2294 | ||||
2295 | Instruction *getMainOp() const { | |||
2296 | return MainOp; | |||
2297 | } | |||
2298 | ||||
2299 | Instruction *getAltOp() const { | |||
2300 | return AltOp; | |||
2301 | } | |||
2302 | ||||
2303 | /// The main/alternate opcodes for the list of instructions. | |||
2304 | unsigned getOpcode() const { | |||
2305 | return MainOp ? MainOp->getOpcode() : 0; | |||
2306 | } | |||
2307 | ||||
2308 | unsigned getAltOpcode() const { | |||
2309 | return AltOp ? AltOp->getOpcode() : 0; | |||
2310 | } | |||
2311 | ||||
2312 | /// When ReuseReorderShuffleIndices is empty it just returns position of \p | |||
2313 | /// V within vector of Scalars. Otherwise, try to remap on its reuse index. | |||
2314 | int findLaneForValue(Value *V) const { | |||
2315 | unsigned FoundLane = std::distance(Scalars.begin(), find(Scalars, V)); | |||
2316 | assert(FoundLane < Scalars.size() && "Couldn't find extract lane")(static_cast <bool> (FoundLane < Scalars.size() && "Couldn't find extract lane") ? void (0) : __assert_fail ("FoundLane < Scalars.size() && \"Couldn't find extract lane\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2316, __extension__ __PRETTY_FUNCTION__)); | |||
2317 | if (!ReorderIndices.empty()) | |||
2318 | FoundLane = ReorderIndices[FoundLane]; | |||
2319 | assert(FoundLane < Scalars.size() && "Couldn't find extract lane")(static_cast <bool> (FoundLane < Scalars.size() && "Couldn't find extract lane") ? void (0) : __assert_fail ("FoundLane < Scalars.size() && \"Couldn't find extract lane\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2319, __extension__ __PRETTY_FUNCTION__)); | |||
2320 | if (!ReuseShuffleIndices.empty()) { | |||
2321 | FoundLane = std::distance(ReuseShuffleIndices.begin(), | |||
2322 | find(ReuseShuffleIndices, FoundLane)); | |||
2323 | } | |||
2324 | return FoundLane; | |||
2325 | } | |||
2326 | ||||
2327 | #ifndef NDEBUG | |||
2328 | /// Debug printer. | |||
2329 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void dump() const { | |||
2330 | dbgs() << Idx << ".\n"; | |||
2331 | for (unsigned OpI = 0, OpE = Operands.size(); OpI != OpE; ++OpI) { | |||
2332 | dbgs() << "Operand " << OpI << ":\n"; | |||
2333 | for (const Value *V : Operands[OpI]) | |||
2334 | dbgs().indent(2) << *V << "\n"; | |||
2335 | } | |||
2336 | dbgs() << "Scalars: \n"; | |||
2337 | for (Value *V : Scalars) | |||
2338 | dbgs().indent(2) << *V << "\n"; | |||
2339 | dbgs() << "State: "; | |||
2340 | switch (State) { | |||
2341 | case Vectorize: | |||
2342 | dbgs() << "Vectorize\n"; | |||
2343 | break; | |||
2344 | case ScatterVectorize: | |||
2345 | dbgs() << "ScatterVectorize\n"; | |||
2346 | break; | |||
2347 | case NeedToGather: | |||
2348 | dbgs() << "NeedToGather\n"; | |||
2349 | break; | |||
2350 | } | |||
2351 | dbgs() << "MainOp: "; | |||
2352 | if (MainOp) | |||
2353 | dbgs() << *MainOp << "\n"; | |||
2354 | else | |||
2355 | dbgs() << "NULL\n"; | |||
2356 | dbgs() << "AltOp: "; | |||
2357 | if (AltOp) | |||
2358 | dbgs() << *AltOp << "\n"; | |||
2359 | else | |||
2360 | dbgs() << "NULL\n"; | |||
2361 | dbgs() << "VectorizedValue: "; | |||
2362 | if (VectorizedValue) | |||
2363 | dbgs() << *VectorizedValue << "\n"; | |||
2364 | else | |||
2365 | dbgs() << "NULL\n"; | |||
2366 | dbgs() << "ReuseShuffleIndices: "; | |||
2367 | if (ReuseShuffleIndices.empty()) | |||
2368 | dbgs() << "Empty"; | |||
2369 | else | |||
2370 | for (int ReuseIdx : ReuseShuffleIndices) | |||
2371 | dbgs() << ReuseIdx << ", "; | |||
2372 | dbgs() << "\n"; | |||
2373 | dbgs() << "ReorderIndices: "; | |||
2374 | for (unsigned ReorderIdx : ReorderIndices) | |||
2375 | dbgs() << ReorderIdx << ", "; | |||
2376 | dbgs() << "\n"; | |||
2377 | dbgs() << "UserTreeIndices: "; | |||
2378 | for (const auto &EInfo : UserTreeIndices) | |||
2379 | dbgs() << EInfo << ", "; | |||
2380 | dbgs() << "\n"; | |||
2381 | } | |||
2382 | #endif | |||
2383 | }; | |||
2384 | ||||
2385 | #ifndef NDEBUG | |||
2386 | void dumpTreeCosts(const TreeEntry *E, InstructionCost ReuseShuffleCost, | |||
2387 | InstructionCost VecCost, | |||
2388 | InstructionCost ScalarCost) const { | |||
2389 | dbgs() << "SLP: Calculated costs for Tree:\n"; E->dump(); | |||
2390 | dbgs() << "SLP: Costs:\n"; | |||
2391 | dbgs() << "SLP: ReuseShuffleCost = " << ReuseShuffleCost << "\n"; | |||
2392 | dbgs() << "SLP: VectorCost = " << VecCost << "\n"; | |||
2393 | dbgs() << "SLP: ScalarCost = " << ScalarCost << "\n"; | |||
2394 | dbgs() << "SLP: ReuseShuffleCost + VecCost - ScalarCost = " << | |||
2395 | ReuseShuffleCost + VecCost - ScalarCost << "\n"; | |||
2396 | } | |||
2397 | #endif | |||
2398 | ||||
2399 | /// Create a new VectorizableTree entry. | |||
2400 | TreeEntry *newTreeEntry(ArrayRef<Value *> VL, Optional<ScheduleData *> Bundle, | |||
2401 | const InstructionsState &S, | |||
2402 | const EdgeInfo &UserTreeIdx, | |||
2403 | ArrayRef<int> ReuseShuffleIndices = None, | |||
2404 | ArrayRef<unsigned> ReorderIndices = None) { | |||
2405 | TreeEntry::EntryState EntryState = | |||
2406 | Bundle ? TreeEntry::Vectorize : TreeEntry::NeedToGather; | |||
2407 | return newTreeEntry(VL, EntryState, Bundle, S, UserTreeIdx, | |||
2408 | ReuseShuffleIndices, ReorderIndices); | |||
2409 | } | |||
2410 | ||||
2411 | TreeEntry *newTreeEntry(ArrayRef<Value *> VL, | |||
2412 | TreeEntry::EntryState EntryState, | |||
2413 | Optional<ScheduleData *> Bundle, | |||
2414 | const InstructionsState &S, | |||
2415 | const EdgeInfo &UserTreeIdx, | |||
2416 | ArrayRef<int> ReuseShuffleIndices = None, | |||
2417 | ArrayRef<unsigned> ReorderIndices = None) { | |||
2418 | assert(((!Bundle && EntryState == TreeEntry::NeedToGather) ||(static_cast <bool> (((!Bundle && EntryState == TreeEntry::NeedToGather) || (Bundle && EntryState != TreeEntry::NeedToGather)) && "Need to vectorize gather entry?" ) ? void (0) : __assert_fail ("((!Bundle && EntryState == TreeEntry::NeedToGather) || (Bundle && EntryState != TreeEntry::NeedToGather)) && \"Need to vectorize gather entry?\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2420, __extension__ __PRETTY_FUNCTION__)) | |||
2419 | (Bundle && EntryState != TreeEntry::NeedToGather)) &&(static_cast <bool> (((!Bundle && EntryState == TreeEntry::NeedToGather) || (Bundle && EntryState != TreeEntry::NeedToGather)) && "Need to vectorize gather entry?" ) ? void (0) : __assert_fail ("((!Bundle && EntryState == TreeEntry::NeedToGather) || (Bundle && EntryState != TreeEntry::NeedToGather)) && \"Need to vectorize gather entry?\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2420, __extension__ __PRETTY_FUNCTION__)) | |||
2420 | "Need to vectorize gather entry?")(static_cast <bool> (((!Bundle && EntryState == TreeEntry::NeedToGather) || (Bundle && EntryState != TreeEntry::NeedToGather)) && "Need to vectorize gather entry?" ) ? void (0) : __assert_fail ("((!Bundle && EntryState == TreeEntry::NeedToGather) || (Bundle && EntryState != TreeEntry::NeedToGather)) && \"Need to vectorize gather entry?\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2420, __extension__ __PRETTY_FUNCTION__)); | |||
2421 | VectorizableTree.push_back(std::make_unique<TreeEntry>(VectorizableTree)); | |||
2422 | TreeEntry *Last = VectorizableTree.back().get(); | |||
2423 | Last->Idx = VectorizableTree.size() - 1; | |||
2424 | Last->State = EntryState; | |||
2425 | Last->ReuseShuffleIndices.append(ReuseShuffleIndices.begin(), | |||
2426 | ReuseShuffleIndices.end()); | |||
2427 | if (ReorderIndices.empty()) { | |||
2428 | Last->Scalars.assign(VL.begin(), VL.end()); | |||
2429 | Last->setOperations(S); | |||
2430 | } else { | |||
2431 | // Reorder scalars and build final mask. | |||
2432 | Last->Scalars.assign(VL.size(), nullptr); | |||
2433 | transform(ReorderIndices, Last->Scalars.begin(), | |||
2434 | [VL](unsigned Idx) -> Value * { | |||
2435 | if (Idx >= VL.size()) | |||
2436 | return UndefValue::get(VL.front()->getType()); | |||
2437 | return VL[Idx]; | |||
2438 | }); | |||
2439 | InstructionsState S = getSameOpcode(Last->Scalars); | |||
2440 | Last->setOperations(S); | |||
2441 | Last->ReorderIndices.append(ReorderIndices.begin(), ReorderIndices.end()); | |||
2442 | } | |||
2443 | if (Last->State != TreeEntry::NeedToGather) { | |||
2444 | for (Value *V : VL) { | |||
2445 | assert(!getTreeEntry(V) && "Scalar already in tree!")(static_cast <bool> (!getTreeEntry(V) && "Scalar already in tree!" ) ? void (0) : __assert_fail ("!getTreeEntry(V) && \"Scalar already in tree!\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2445, __extension__ __PRETTY_FUNCTION__)); | |||
2446 | ScalarToTreeEntry[V] = Last; | |||
2447 | } | |||
2448 | // Update the scheduler bundle to point to this TreeEntry. | |||
2449 | ScheduleData *BundleMember = Bundle.getValue(); | |||
2450 | assert((BundleMember || isa<PHINode>(S.MainOp) ||(static_cast <bool> ((BundleMember || isa<PHINode> (S.MainOp) || isVectorLikeInstWithConstOps(S.MainOp) || doesNotNeedToSchedule (VL)) && "Bundle and VL out of sync") ? void (0) : __assert_fail ("(BundleMember || isa<PHINode>(S.MainOp) || isVectorLikeInstWithConstOps(S.MainOp) || doesNotNeedToSchedule(VL)) && \"Bundle and VL out of sync\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2453, __extension__ __PRETTY_FUNCTION__)) | |||
2451 | isVectorLikeInstWithConstOps(S.MainOp) ||(static_cast <bool> ((BundleMember || isa<PHINode> (S.MainOp) || isVectorLikeInstWithConstOps(S.MainOp) || doesNotNeedToSchedule (VL)) && "Bundle and VL out of sync") ? void (0) : __assert_fail ("(BundleMember || isa<PHINode>(S.MainOp) || isVectorLikeInstWithConstOps(S.MainOp) || doesNotNeedToSchedule(VL)) && \"Bundle and VL out of sync\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2453, __extension__ __PRETTY_FUNCTION__)) | |||
2452 | doesNotNeedToSchedule(VL)) &&(static_cast <bool> ((BundleMember || isa<PHINode> (S.MainOp) || isVectorLikeInstWithConstOps(S.MainOp) || doesNotNeedToSchedule (VL)) && "Bundle and VL out of sync") ? void (0) : __assert_fail ("(BundleMember || isa<PHINode>(S.MainOp) || isVectorLikeInstWithConstOps(S.MainOp) || doesNotNeedToSchedule(VL)) && \"Bundle and VL out of sync\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2453, __extension__ __PRETTY_FUNCTION__)) | |||
2453 | "Bundle and VL out of sync")(static_cast <bool> ((BundleMember || isa<PHINode> (S.MainOp) || isVectorLikeInstWithConstOps(S.MainOp) || doesNotNeedToSchedule (VL)) && "Bundle and VL out of sync") ? void (0) : __assert_fail ("(BundleMember || isa<PHINode>(S.MainOp) || isVectorLikeInstWithConstOps(S.MainOp) || doesNotNeedToSchedule(VL)) && \"Bundle and VL out of sync\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2453, __extension__ __PRETTY_FUNCTION__)); | |||
2454 | if (BundleMember) { | |||
2455 | for (Value *V : VL) { | |||
2456 | if (doesNotNeedToBeScheduled(V)) | |||
2457 | continue; | |||
2458 | assert(BundleMember && "Unexpected end of bundle.")(static_cast <bool> (BundleMember && "Unexpected end of bundle." ) ? void (0) : __assert_fail ("BundleMember && \"Unexpected end of bundle.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2458, __extension__ __PRETTY_FUNCTION__)); | |||
2459 | BundleMember->TE = Last; | |||
2460 | BundleMember = BundleMember->NextInBundle; | |||
2461 | } | |||
2462 | } | |||
2463 | assert(!BundleMember && "Bundle and VL out of sync")(static_cast <bool> (!BundleMember && "Bundle and VL out of sync" ) ? void (0) : __assert_fail ("!BundleMember && \"Bundle and VL out of sync\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2463, __extension__ __PRETTY_FUNCTION__)); | |||
2464 | } else { | |||
2465 | MustGather.insert(VL.begin(), VL.end()); | |||
2466 | } | |||
2467 | ||||
2468 | if (UserTreeIdx.UserTE) | |||
2469 | Last->UserTreeIndices.push_back(UserTreeIdx); | |||
2470 | ||||
2471 | return Last; | |||
2472 | } | |||
2473 | ||||
2474 | /// -- Vectorization State -- | |||
2475 | /// Holds all of the tree entries. | |||
2476 | TreeEntry::VecTreeTy VectorizableTree; | |||
2477 | ||||
2478 | #ifndef NDEBUG | |||
2479 | /// Debug printer. | |||
2480 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void dumpVectorizableTree() const { | |||
2481 | for (unsigned Id = 0, IdE = VectorizableTree.size(); Id != IdE; ++Id) { | |||
2482 | VectorizableTree[Id]->dump(); | |||
2483 | dbgs() << "\n"; | |||
2484 | } | |||
2485 | } | |||
2486 | #endif | |||
2487 | ||||
2488 | TreeEntry *getTreeEntry(Value *V) { return ScalarToTreeEntry.lookup(V); } | |||
2489 | ||||
2490 | const TreeEntry *getTreeEntry(Value *V) const { | |||
2491 | return ScalarToTreeEntry.lookup(V); | |||
2492 | } | |||
2493 | ||||
2494 | /// Maps a specific scalar to its tree entry. | |||
2495 | SmallDenseMap<Value*, TreeEntry *> ScalarToTreeEntry; | |||
2496 | ||||
2497 | /// Maps a value to the proposed vectorizable size. | |||
2498 | SmallDenseMap<Value *, unsigned> InstrElementSize; | |||
2499 | ||||
2500 | /// A list of scalars that we found that we need to keep as scalars. | |||
2501 | ValueSet MustGather; | |||
2502 | ||||
2503 | /// This POD struct describes one external user in the vectorized tree. | |||
2504 | struct ExternalUser { | |||
2505 | ExternalUser(Value *S, llvm::User *U, int L) | |||
2506 | : Scalar(S), User(U), Lane(L) {} | |||
2507 | ||||
2508 | // Which scalar in our function. | |||
2509 | Value *Scalar; | |||
2510 | ||||
2511 | // Which user that uses the scalar. | |||
2512 | llvm::User *User; | |||
2513 | ||||
2514 | // Which lane does the scalar belong to. | |||
2515 | int Lane; | |||
2516 | }; | |||
2517 | using UserList = SmallVector<ExternalUser, 16>; | |||
2518 | ||||
2519 | /// Checks if two instructions may access the same memory. | |||
2520 | /// | |||
2521 | /// \p Loc1 is the location of \p Inst1. It is passed explicitly because it | |||
2522 | /// is invariant in the calling loop. | |||
2523 | bool isAliased(const MemoryLocation &Loc1, Instruction *Inst1, | |||
2524 | Instruction *Inst2) { | |||
2525 | // First check if the result is already in the cache. | |||
2526 | AliasCacheKey key = std::make_pair(Inst1, Inst2); | |||
2527 | Optional<bool> &result = AliasCache[key]; | |||
2528 | if (result.hasValue()) { | |||
2529 | return result.getValue(); | |||
2530 | } | |||
2531 | bool aliased = true; | |||
2532 | if (Loc1.Ptr && isSimple(Inst1)) | |||
2533 | aliased = isModOrRefSet(BatchAA.getModRefInfo(Inst2, Loc1)); | |||
2534 | // Store the result in the cache. | |||
2535 | result = aliased; | |||
2536 | return aliased; | |||
2537 | } | |||
2538 | ||||
2539 | using AliasCacheKey = std::pair<Instruction *, Instruction *>; | |||
2540 | ||||
2541 | /// Cache for alias results. | |||
2542 | /// TODO: consider moving this to the AliasAnalysis itself. | |||
2543 | DenseMap<AliasCacheKey, Optional<bool>> AliasCache; | |||
2544 | ||||
2545 | // Cache for pointerMayBeCaptured calls inside AA. This is preserved | |||
2546 | // globally through SLP because we don't perform any action which | |||
2547 | // invalidates capture results. | |||
2548 | BatchAAResults BatchAA; | |||
2549 | ||||
2550 | /// Temporary store for deleted instructions. Instructions will be deleted | |||
2551 | /// eventually when the BoUpSLP is destructed. The deferral is required to | |||
2552 | /// ensure that there are no incorrect collisions in the AliasCache, which | |||
2553 | /// can happen if a new instruction is allocated at the same address as a | |||
2554 | /// previously deleted instruction. | |||
2555 | DenseSet<Instruction *> DeletedInstructions; | |||
2556 | ||||
2557 | /// A list of values that need to extracted out of the tree. | |||
2558 | /// This list holds pairs of (Internal Scalar : External User). External User | |||
2559 | /// can be nullptr, it means that this Internal Scalar will be used later, | |||
2560 | /// after vectorization. | |||
2561 | UserList ExternalUses; | |||
2562 | ||||
2563 | /// Values used only by @llvm.assume calls. | |||
2564 | SmallPtrSet<const Value *, 32> EphValues; | |||
2565 | ||||
2566 | /// Holds all of the instructions that we gathered. | |||
2567 | SetVector<Instruction *> GatherShuffleSeq; | |||
2568 | ||||
2569 | /// A list of blocks that we are going to CSE. | |||
2570 | SetVector<BasicBlock *> CSEBlocks; | |||
2571 | ||||
2572 | /// Contains all scheduling relevant data for an instruction. | |||
2573 | /// A ScheduleData either represents a single instruction or a member of an | |||
2574 | /// instruction bundle (= a group of instructions which is combined into a | |||
2575 | /// vector instruction). | |||
2576 | struct ScheduleData { | |||
2577 | // The initial value for the dependency counters. It means that the | |||
2578 | // dependencies are not calculated yet. | |||
2579 | enum { InvalidDeps = -1 }; | |||
2580 | ||||
2581 | ScheduleData() = default; | |||
2582 | ||||
2583 | void init(int BlockSchedulingRegionID, Value *OpVal) { | |||
2584 | FirstInBundle = this; | |||
2585 | NextInBundle = nullptr; | |||
2586 | NextLoadStore = nullptr; | |||
2587 | IsScheduled = false; | |||
2588 | SchedulingRegionID = BlockSchedulingRegionID; | |||
2589 | clearDependencies(); | |||
2590 | OpValue = OpVal; | |||
2591 | TE = nullptr; | |||
2592 | } | |||
2593 | ||||
2594 | /// Verify basic self consistency properties | |||
2595 | void verify() { | |||
2596 | if (hasValidDependencies()) { | |||
2597 | assert(UnscheduledDeps <= Dependencies && "invariant")(static_cast <bool> (UnscheduledDeps <= Dependencies && "invariant") ? void (0) : __assert_fail ("UnscheduledDeps <= Dependencies && \"invariant\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2597, __extension__ __PRETTY_FUNCTION__)); | |||
2598 | } else { | |||
2599 | assert(UnscheduledDeps == Dependencies && "invariant")(static_cast <bool> (UnscheduledDeps == Dependencies && "invariant") ? void (0) : __assert_fail ("UnscheduledDeps == Dependencies && \"invariant\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2599, __extension__ __PRETTY_FUNCTION__)); | |||
2600 | } | |||
2601 | ||||
2602 | if (IsScheduled) { | |||
2603 | assert(isSchedulingEntity() &&(static_cast <bool> (isSchedulingEntity() && "unexpected scheduled state" ) ? void (0) : __assert_fail ("isSchedulingEntity() && \"unexpected scheduled state\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2604, __extension__ __PRETTY_FUNCTION__)) | |||
2604 | "unexpected scheduled state")(static_cast <bool> (isSchedulingEntity() && "unexpected scheduled state" ) ? void (0) : __assert_fail ("isSchedulingEntity() && \"unexpected scheduled state\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2604, __extension__ __PRETTY_FUNCTION__)); | |||
2605 | for (const ScheduleData *BundleMember = this; BundleMember; | |||
2606 | BundleMember = BundleMember->NextInBundle) { | |||
2607 | assert(BundleMember->hasValidDependencies() &&(static_cast <bool> (BundleMember->hasValidDependencies () && BundleMember->UnscheduledDeps == 0 && "unexpected scheduled state") ? void (0) : __assert_fail ("BundleMember->hasValidDependencies() && BundleMember->UnscheduledDeps == 0 && \"unexpected scheduled state\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2609, __extension__ __PRETTY_FUNCTION__)) | |||
2608 | BundleMember->UnscheduledDeps == 0 &&(static_cast <bool> (BundleMember->hasValidDependencies () && BundleMember->UnscheduledDeps == 0 && "unexpected scheduled state") ? void (0) : __assert_fail ("BundleMember->hasValidDependencies() && BundleMember->UnscheduledDeps == 0 && \"unexpected scheduled state\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2609, __extension__ __PRETTY_FUNCTION__)) | |||
2609 | "unexpected scheduled state")(static_cast <bool> (BundleMember->hasValidDependencies () && BundleMember->UnscheduledDeps == 0 && "unexpected scheduled state") ? void (0) : __assert_fail ("BundleMember->hasValidDependencies() && BundleMember->UnscheduledDeps == 0 && \"unexpected scheduled state\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2609, __extension__ __PRETTY_FUNCTION__)); | |||
2610 | assert((BundleMember == this || !BundleMember->IsScheduled) &&(static_cast <bool> ((BundleMember == this || !BundleMember ->IsScheduled) && "only bundle is marked scheduled" ) ? void (0) : __assert_fail ("(BundleMember == this || !BundleMember->IsScheduled) && \"only bundle is marked scheduled\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2611, __extension__ __PRETTY_FUNCTION__)) | |||
2611 | "only bundle is marked scheduled")(static_cast <bool> ((BundleMember == this || !BundleMember ->IsScheduled) && "only bundle is marked scheduled" ) ? void (0) : __assert_fail ("(BundleMember == this || !BundleMember->IsScheduled) && \"only bundle is marked scheduled\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2611, __extension__ __PRETTY_FUNCTION__)); | |||
2612 | } | |||
2613 | } | |||
2614 | ||||
2615 | assert(Inst->getParent() == FirstInBundle->Inst->getParent() &&(static_cast <bool> (Inst->getParent() == FirstInBundle ->Inst->getParent() && "all bundle members must be in same basic block" ) ? void (0) : __assert_fail ("Inst->getParent() == FirstInBundle->Inst->getParent() && \"all bundle members must be in same basic block\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2616, __extension__ __PRETTY_FUNCTION__)) | |||
2616 | "all bundle members must be in same basic block")(static_cast <bool> (Inst->getParent() == FirstInBundle ->Inst->getParent() && "all bundle members must be in same basic block" ) ? void (0) : __assert_fail ("Inst->getParent() == FirstInBundle->Inst->getParent() && \"all bundle members must be in same basic block\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2616, __extension__ __PRETTY_FUNCTION__)); | |||
2617 | } | |||
2618 | ||||
2619 | /// Returns true if the dependency information has been calculated. | |||
2620 | /// Note that depenendency validity can vary between instructions within | |||
2621 | /// a single bundle. | |||
2622 | bool hasValidDependencies() const { return Dependencies != InvalidDeps; } | |||
2623 | ||||
2624 | /// Returns true for single instructions and for bundle representatives | |||
2625 | /// (= the head of a bundle). | |||
2626 | bool isSchedulingEntity() const { return FirstInBundle == this; } | |||
2627 | ||||
2628 | /// Returns true if it represents an instruction bundle and not only a | |||
2629 | /// single instruction. | |||
2630 | bool isPartOfBundle() const { | |||
2631 | return NextInBundle != nullptr || FirstInBundle != this || TE; | |||
2632 | } | |||
2633 | ||||
2634 | /// Returns true if it is ready for scheduling, i.e. it has no more | |||
2635 | /// unscheduled depending instructions/bundles. | |||
2636 | bool isReady() const { | |||
2637 | assert(isSchedulingEntity() &&(static_cast <bool> (isSchedulingEntity() && "can't consider non-scheduling entity for ready list" ) ? void (0) : __assert_fail ("isSchedulingEntity() && \"can't consider non-scheduling entity for ready list\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2638, __extension__ __PRETTY_FUNCTION__)) | |||
2638 | "can't consider non-scheduling entity for ready list")(static_cast <bool> (isSchedulingEntity() && "can't consider non-scheduling entity for ready list" ) ? void (0) : __assert_fail ("isSchedulingEntity() && \"can't consider non-scheduling entity for ready list\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2638, __extension__ __PRETTY_FUNCTION__)); | |||
2639 | return unscheduledDepsInBundle() == 0 && !IsScheduled; | |||
2640 | } | |||
2641 | ||||
2642 | /// Modifies the number of unscheduled dependencies for this instruction, | |||
2643 | /// and returns the number of remaining dependencies for the containing | |||
2644 | /// bundle. | |||
2645 | int incrementUnscheduledDeps(int Incr) { | |||
2646 | assert(hasValidDependencies() &&(static_cast <bool> (hasValidDependencies() && "increment of unscheduled deps would be meaningless" ) ? void (0) : __assert_fail ("hasValidDependencies() && \"increment of unscheduled deps would be meaningless\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2647, __extension__ __PRETTY_FUNCTION__)) | |||
2647 | "increment of unscheduled deps would be meaningless")(static_cast <bool> (hasValidDependencies() && "increment of unscheduled deps would be meaningless" ) ? void (0) : __assert_fail ("hasValidDependencies() && \"increment of unscheduled deps would be meaningless\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2647, __extension__ __PRETTY_FUNCTION__)); | |||
2648 | UnscheduledDeps += Incr; | |||
2649 | return FirstInBundle->unscheduledDepsInBundle(); | |||
2650 | } | |||
2651 | ||||
2652 | /// Sets the number of unscheduled dependencies to the number of | |||
2653 | /// dependencies. | |||
2654 | void resetUnscheduledDeps() { | |||
2655 | UnscheduledDeps = Dependencies; | |||
2656 | } | |||
2657 | ||||
2658 | /// Clears all dependency information. | |||
2659 | void clearDependencies() { | |||
2660 | Dependencies = InvalidDeps; | |||
2661 | resetUnscheduledDeps(); | |||
2662 | MemoryDependencies.clear(); | |||
2663 | ControlDependencies.clear(); | |||
2664 | } | |||
2665 | ||||
2666 | int unscheduledDepsInBundle() const { | |||
2667 | assert(isSchedulingEntity() && "only meaningful on the bundle")(static_cast <bool> (isSchedulingEntity() && "only meaningful on the bundle" ) ? void (0) : __assert_fail ("isSchedulingEntity() && \"only meaningful on the bundle\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2667, __extension__ __PRETTY_FUNCTION__)); | |||
2668 | int Sum = 0; | |||
2669 | for (const ScheduleData *BundleMember = this; BundleMember; | |||
2670 | BundleMember = BundleMember->NextInBundle) { | |||
2671 | if (BundleMember->UnscheduledDeps == InvalidDeps) | |||
2672 | return InvalidDeps; | |||
2673 | Sum += BundleMember->UnscheduledDeps; | |||
2674 | } | |||
2675 | return Sum; | |||
2676 | } | |||
2677 | ||||
2678 | void dump(raw_ostream &os) const { | |||
2679 | if (!isSchedulingEntity()) { | |||
2680 | os << "/ " << *Inst; | |||
2681 | } else if (NextInBundle) { | |||
2682 | os << '[' << *Inst; | |||
2683 | ScheduleData *SD = NextInBundle; | |||
2684 | while (SD) { | |||
2685 | os << ';' << *SD->Inst; | |||
2686 | SD = SD->NextInBundle; | |||
2687 | } | |||
2688 | os << ']'; | |||
2689 | } else { | |||
2690 | os << *Inst; | |||
2691 | } | |||
2692 | } | |||
2693 | ||||
2694 | Instruction *Inst = nullptr; | |||
2695 | ||||
2696 | /// Opcode of the current instruction in the schedule data. | |||
2697 | Value *OpValue = nullptr; | |||
2698 | ||||
2699 | /// The TreeEntry that this instruction corresponds to. | |||
2700 | TreeEntry *TE = nullptr; | |||
2701 | ||||
2702 | /// Points to the head in an instruction bundle (and always to this for | |||
2703 | /// single instructions). | |||
2704 | ScheduleData *FirstInBundle = nullptr; | |||
2705 | ||||
2706 | /// Single linked list of all instructions in a bundle. Null if it is a | |||
2707 | /// single instruction. | |||
2708 | ScheduleData *NextInBundle = nullptr; | |||
2709 | ||||
2710 | /// Single linked list of all memory instructions (e.g. load, store, call) | |||
2711 | /// in the block - until the end of the scheduling region. | |||
2712 | ScheduleData *NextLoadStore = nullptr; | |||
2713 | ||||
2714 | /// The dependent memory instructions. | |||
2715 | /// This list is derived on demand in calculateDependencies(). | |||
2716 | SmallVector<ScheduleData *, 4> MemoryDependencies; | |||
2717 | ||||
2718 | /// List of instructions which this instruction could be control dependent | |||
2719 | /// on. Allowing such nodes to be scheduled below this one could introduce | |||
2720 | /// a runtime fault which didn't exist in the original program. | |||
2721 | /// ex: this is a load or udiv following a readonly call which inf loops | |||
2722 | SmallVector<ScheduleData *, 4> ControlDependencies; | |||
2723 | ||||
2724 | /// This ScheduleData is in the current scheduling region if this matches | |||
2725 | /// the current SchedulingRegionID of BlockScheduling. | |||
2726 | int SchedulingRegionID = 0; | |||
2727 | ||||
2728 | /// Used for getting a "good" final ordering of instructions. | |||
2729 | int SchedulingPriority = 0; | |||
2730 | ||||
2731 | /// The number of dependencies. Constitutes of the number of users of the | |||
2732 | /// instruction plus the number of dependent memory instructions (if any). | |||
2733 | /// This value is calculated on demand. | |||
2734 | /// If InvalidDeps, the number of dependencies is not calculated yet. | |||
2735 | int Dependencies = InvalidDeps; | |||
2736 | ||||
2737 | /// The number of dependencies minus the number of dependencies of scheduled | |||
2738 | /// instructions. As soon as this is zero, the instruction/bundle gets ready | |||
2739 | /// for scheduling. | |||
2740 | /// Note that this is negative as long as Dependencies is not calculated. | |||
2741 | int UnscheduledDeps = InvalidDeps; | |||
2742 | ||||
2743 | /// True if this instruction is scheduled (or considered as scheduled in the | |||
2744 | /// dry-run). | |||
2745 | bool IsScheduled = false; | |||
2746 | }; | |||
2747 | ||||
2748 | #ifndef NDEBUG | |||
2749 | friend inline raw_ostream &operator<<(raw_ostream &os, | |||
2750 | const BoUpSLP::ScheduleData &SD) { | |||
2751 | SD.dump(os); | |||
2752 | return os; | |||
2753 | } | |||
2754 | #endif | |||
2755 | ||||
2756 | friend struct GraphTraits<BoUpSLP *>; | |||
2757 | friend struct DOTGraphTraits<BoUpSLP *>; | |||
2758 | ||||
2759 | /// Contains all scheduling data for a basic block. | |||
2760 | /// It does not schedules instructions, which are not memory read/write | |||
2761 | /// instructions and their operands are either constants, or arguments, or | |||
2762 | /// phis, or instructions from others blocks, or their users are phis or from | |||
2763 | /// the other blocks. The resulting vector instructions can be placed at the | |||
2764 | /// beginning of the basic block without scheduling (if operands does not need | |||
2765 | /// to be scheduled) or at the end of the block (if users are outside of the | |||
2766 | /// block). It allows to save some compile time and memory used by the | |||
2767 | /// compiler. | |||
2768 | /// ScheduleData is assigned for each instruction in between the boundaries of | |||
2769 | /// the tree entry, even for those, which are not part of the graph. It is | |||
2770 | /// required to correctly follow the dependencies between the instructions and | |||
2771 | /// their correct scheduling. The ScheduleData is not allocated for the | |||
2772 | /// instructions, which do not require scheduling, like phis, nodes with | |||
2773 | /// extractelements/insertelements only or nodes with instructions, with | |||
2774 | /// uses/operands outside of the block. | |||
2775 | struct BlockScheduling { | |||
2776 | BlockScheduling(BasicBlock *BB) | |||
2777 | : BB(BB), ChunkSize(BB->size()), ChunkPos(ChunkSize) {} | |||
2778 | ||||
2779 | void clear() { | |||
2780 | ReadyInsts.clear(); | |||
2781 | ScheduleStart = nullptr; | |||
2782 | ScheduleEnd = nullptr; | |||
2783 | FirstLoadStoreInRegion = nullptr; | |||
2784 | LastLoadStoreInRegion = nullptr; | |||
2785 | RegionHasStackSave = false; | |||
2786 | ||||
2787 | // Reduce the maximum schedule region size by the size of the | |||
2788 | // previous scheduling run. | |||
2789 | ScheduleRegionSizeLimit -= ScheduleRegionSize; | |||
2790 | if (ScheduleRegionSizeLimit < MinScheduleRegionSize) | |||
2791 | ScheduleRegionSizeLimit = MinScheduleRegionSize; | |||
2792 | ScheduleRegionSize = 0; | |||
2793 | ||||
2794 | // Make a new scheduling region, i.e. all existing ScheduleData is not | |||
2795 | // in the new region yet. | |||
2796 | ++SchedulingRegionID; | |||
2797 | } | |||
2798 | ||||
2799 | ScheduleData *getScheduleData(Instruction *I) { | |||
2800 | if (BB != I->getParent()) | |||
2801 | // Avoid lookup if can't possibly be in map. | |||
2802 | return nullptr; | |||
2803 | ScheduleData *SD = ScheduleDataMap.lookup(I); | |||
2804 | if (SD && isInSchedulingRegion(SD)) | |||
2805 | return SD; | |||
2806 | return nullptr; | |||
2807 | } | |||
2808 | ||||
2809 | ScheduleData *getScheduleData(Value *V) { | |||
2810 | if (auto *I = dyn_cast<Instruction>(V)) | |||
2811 | return getScheduleData(I); | |||
2812 | return nullptr; | |||
2813 | } | |||
2814 | ||||
2815 | ScheduleData *getScheduleData(Value *V, Value *Key) { | |||
2816 | if (V == Key) | |||
2817 | return getScheduleData(V); | |||
2818 | auto I = ExtraScheduleDataMap.find(V); | |||
2819 | if (I != ExtraScheduleDataMap.end()) { | |||
2820 | ScheduleData *SD = I->second.lookup(Key); | |||
2821 | if (SD && isInSchedulingRegion(SD)) | |||
2822 | return SD; | |||
2823 | } | |||
2824 | return nullptr; | |||
2825 | } | |||
2826 | ||||
2827 | bool isInSchedulingRegion(ScheduleData *SD) const { | |||
2828 | return SD->SchedulingRegionID == SchedulingRegionID; | |||
2829 | } | |||
2830 | ||||
2831 | /// Marks an instruction as scheduled and puts all dependent ready | |||
2832 | /// instructions into the ready-list. | |||
2833 | template <typename ReadyListType> | |||
2834 | void schedule(ScheduleData *SD, ReadyListType &ReadyList) { | |||
2835 | SD->IsScheduled = true; | |||
| ||||
2836 | LLVM_DEBUG(dbgs() << "SLP: schedule " << *SD << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: schedule " << *SD << "\n"; } } while (false); | |||
2837 | ||||
2838 | for (ScheduleData *BundleMember = SD; BundleMember; | |||
2839 | BundleMember = BundleMember->NextInBundle) { | |||
2840 | if (BundleMember->Inst != BundleMember->OpValue) | |||
2841 | continue; | |||
2842 | ||||
2843 | // Handle the def-use chain dependencies. | |||
2844 | ||||
2845 | // Decrement the unscheduled counter and insert to ready list if ready. | |||
2846 | auto &&DecrUnsched = [this, &ReadyList](Instruction *I) { | |||
2847 | doForAllOpcodes(I, [&ReadyList](ScheduleData *OpDef) { | |||
2848 | if (OpDef && OpDef->hasValidDependencies() && | |||
2849 | OpDef->incrementUnscheduledDeps(-1) == 0) { | |||
2850 | // There are no more unscheduled dependencies after | |||
2851 | // decrementing, so we can put the dependent instruction | |||
2852 | // into the ready list. | |||
2853 | ScheduleData *DepBundle = OpDef->FirstInBundle; | |||
2854 | assert(!DepBundle->IsScheduled &&(static_cast <bool> (!DepBundle->IsScheduled && "already scheduled bundle gets ready") ? void (0) : __assert_fail ("!DepBundle->IsScheduled && \"already scheduled bundle gets ready\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2855, __extension__ __PRETTY_FUNCTION__)) | |||
2855 | "already scheduled bundle gets ready")(static_cast <bool> (!DepBundle->IsScheduled && "already scheduled bundle gets ready") ? void (0) : __assert_fail ("!DepBundle->IsScheduled && \"already scheduled bundle gets ready\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2855, __extension__ __PRETTY_FUNCTION__)); | |||
2856 | ReadyList.insert(DepBundle); | |||
2857 | LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: gets ready (def): " << *DepBundle << "\n"; } } while (false) | |||
2858 | << "SLP: gets ready (def): " << *DepBundle << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: gets ready (def): " << *DepBundle << "\n"; } } while (false); | |||
2859 | } | |||
2860 | }); | |||
2861 | }; | |||
2862 | ||||
2863 | // If BundleMember is a vector bundle, its operands may have been | |||
2864 | // reordered during buildTree(). We therefore need to get its operands | |||
2865 | // through the TreeEntry. | |||
2866 | if (TreeEntry *TE = BundleMember->TE) { | |||
2867 | // Need to search for the lane since the tree entry can be reordered. | |||
2868 | int Lane = std::distance(TE->Scalars.begin(), | |||
2869 | find(TE->Scalars, BundleMember->Inst)); | |||
2870 | assert(Lane >= 0 && "Lane not set")(static_cast <bool> (Lane >= 0 && "Lane not set" ) ? void (0) : __assert_fail ("Lane >= 0 && \"Lane not set\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2870, __extension__ __PRETTY_FUNCTION__)); | |||
2871 | ||||
2872 | // Since vectorization tree is being built recursively this assertion | |||
2873 | // ensures that the tree entry has all operands set before reaching | |||
2874 | // this code. Couple of exceptions known at the moment are extracts | |||
2875 | // where their second (immediate) operand is not added. Since | |||
2876 | // immediates do not affect scheduler behavior this is considered | |||
2877 | // okay. | |||
2878 | auto *In = BundleMember->Inst; | |||
2879 | assert(In &&(static_cast <bool> (In && (isa<ExtractValueInst >(In) || isa<ExtractElementInst>(In) || In->getNumOperands () == TE->getNumOperands()) && "Missed TreeEntry operands?" ) ? void (0) : __assert_fail ("In && (isa<ExtractValueInst>(In) || isa<ExtractElementInst>(In) || In->getNumOperands() == TE->getNumOperands()) && \"Missed TreeEntry operands?\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2882, __extension__ __PRETTY_FUNCTION__)) | |||
2880 | (isa<ExtractValueInst>(In) || isa<ExtractElementInst>(In) ||(static_cast <bool> (In && (isa<ExtractValueInst >(In) || isa<ExtractElementInst>(In) || In->getNumOperands () == TE->getNumOperands()) && "Missed TreeEntry operands?" ) ? void (0) : __assert_fail ("In && (isa<ExtractValueInst>(In) || isa<ExtractElementInst>(In) || In->getNumOperands() == TE->getNumOperands()) && \"Missed TreeEntry operands?\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2882, __extension__ __PRETTY_FUNCTION__)) | |||
2881 | In->getNumOperands() == TE->getNumOperands()) &&(static_cast <bool> (In && (isa<ExtractValueInst >(In) || isa<ExtractElementInst>(In) || In->getNumOperands () == TE->getNumOperands()) && "Missed TreeEntry operands?" ) ? void (0) : __assert_fail ("In && (isa<ExtractValueInst>(In) || isa<ExtractElementInst>(In) || In->getNumOperands() == TE->getNumOperands()) && \"Missed TreeEntry operands?\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2882, __extension__ __PRETTY_FUNCTION__)) | |||
2882 | "Missed TreeEntry operands?")(static_cast <bool> (In && (isa<ExtractValueInst >(In) || isa<ExtractElementInst>(In) || In->getNumOperands () == TE->getNumOperands()) && "Missed TreeEntry operands?" ) ? void (0) : __assert_fail ("In && (isa<ExtractValueInst>(In) || isa<ExtractElementInst>(In) || In->getNumOperands() == TE->getNumOperands()) && \"Missed TreeEntry operands?\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2882, __extension__ __PRETTY_FUNCTION__)); | |||
2883 | (void)In; // fake use to avoid build failure when assertions disabled | |||
2884 | ||||
2885 | for (unsigned OpIdx = 0, NumOperands = TE->getNumOperands(); | |||
2886 | OpIdx != NumOperands; ++OpIdx) | |||
2887 | if (auto *I = dyn_cast<Instruction>(TE->getOperand(OpIdx)[Lane])) | |||
2888 | DecrUnsched(I); | |||
2889 | } else { | |||
2890 | // If BundleMember is a stand-alone instruction, no operand reordering | |||
2891 | // has taken place, so we directly access its operands. | |||
2892 | for (Use &U : BundleMember->Inst->operands()) | |||
2893 | if (auto *I = dyn_cast<Instruction>(U.get())) | |||
2894 | DecrUnsched(I); | |||
2895 | } | |||
2896 | // Handle the memory dependencies. | |||
2897 | for (ScheduleData *MemoryDepSD : BundleMember->MemoryDependencies) { | |||
2898 | if (MemoryDepSD->hasValidDependencies() && | |||
2899 | MemoryDepSD->incrementUnscheduledDeps(-1) == 0) { | |||
2900 | // There are no more unscheduled dependencies after decrementing, | |||
2901 | // so we can put the dependent instruction into the ready list. | |||
2902 | ScheduleData *DepBundle = MemoryDepSD->FirstInBundle; | |||
2903 | assert(!DepBundle->IsScheduled &&(static_cast <bool> (!DepBundle->IsScheduled && "already scheduled bundle gets ready") ? void (0) : __assert_fail ("!DepBundle->IsScheduled && \"already scheduled bundle gets ready\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2904, __extension__ __PRETTY_FUNCTION__)) | |||
2904 | "already scheduled bundle gets ready")(static_cast <bool> (!DepBundle->IsScheduled && "already scheduled bundle gets ready") ? void (0) : __assert_fail ("!DepBundle->IsScheduled && \"already scheduled bundle gets ready\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2904, __extension__ __PRETTY_FUNCTION__)); | |||
2905 | ReadyList.insert(DepBundle); | |||
2906 | LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: gets ready (mem): " << *DepBundle << "\n"; } } while (false) | |||
2907 | << "SLP: gets ready (mem): " << *DepBundle << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: gets ready (mem): " << *DepBundle << "\n"; } } while (false); | |||
2908 | } | |||
2909 | } | |||
2910 | // Handle the control dependencies. | |||
2911 | for (ScheduleData *DepSD : BundleMember->ControlDependencies) { | |||
2912 | if (DepSD->incrementUnscheduledDeps(-1) == 0) { | |||
2913 | // There are no more unscheduled dependencies after decrementing, | |||
2914 | // so we can put the dependent instruction into the ready list. | |||
2915 | ScheduleData *DepBundle = DepSD->FirstInBundle; | |||
2916 | assert(!DepBundle->IsScheduled &&(static_cast <bool> (!DepBundle->IsScheduled && "already scheduled bundle gets ready") ? void (0) : __assert_fail ("!DepBundle->IsScheduled && \"already scheduled bundle gets ready\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2917, __extension__ __PRETTY_FUNCTION__)) | |||
2917 | "already scheduled bundle gets ready")(static_cast <bool> (!DepBundle->IsScheduled && "already scheduled bundle gets ready") ? void (0) : __assert_fail ("!DepBundle->IsScheduled && \"already scheduled bundle gets ready\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2917, __extension__ __PRETTY_FUNCTION__)); | |||
2918 | ReadyList.insert(DepBundle); | |||
2919 | LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: gets ready (ctl): " << *DepBundle << "\n"; } } while (false) | |||
2920 | << "SLP: gets ready (ctl): " << *DepBundle << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: gets ready (ctl): " << *DepBundle << "\n"; } } while (false); | |||
2921 | } | |||
2922 | } | |||
2923 | ||||
2924 | } | |||
2925 | } | |||
2926 | ||||
2927 | /// Verify basic self consistency properties of the data structure. | |||
2928 | void verify() { | |||
2929 | if (!ScheduleStart) | |||
2930 | return; | |||
2931 | ||||
2932 | assert(ScheduleStart->getParent() == ScheduleEnd->getParent() &&(static_cast <bool> (ScheduleStart->getParent() == ScheduleEnd ->getParent() && ScheduleStart->comesBefore(ScheduleEnd ) && "Not a valid scheduling region?") ? void (0) : __assert_fail ("ScheduleStart->getParent() == ScheduleEnd->getParent() && ScheduleStart->comesBefore(ScheduleEnd) && \"Not a valid scheduling region?\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2934, __extension__ __PRETTY_FUNCTION__)) | |||
2933 | ScheduleStart->comesBefore(ScheduleEnd) &&(static_cast <bool> (ScheduleStart->getParent() == ScheduleEnd ->getParent() && ScheduleStart->comesBefore(ScheduleEnd ) && "Not a valid scheduling region?") ? void (0) : __assert_fail ("ScheduleStart->getParent() == ScheduleEnd->getParent() && ScheduleStart->comesBefore(ScheduleEnd) && \"Not a valid scheduling region?\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2934, __extension__ __PRETTY_FUNCTION__)) | |||
2934 | "Not a valid scheduling region?")(static_cast <bool> (ScheduleStart->getParent() == ScheduleEnd ->getParent() && ScheduleStart->comesBefore(ScheduleEnd ) && "Not a valid scheduling region?") ? void (0) : __assert_fail ("ScheduleStart->getParent() == ScheduleEnd->getParent() && ScheduleStart->comesBefore(ScheduleEnd) && \"Not a valid scheduling region?\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2934, __extension__ __PRETTY_FUNCTION__)); | |||
2935 | ||||
2936 | for (auto *I = ScheduleStart; I != ScheduleEnd; I = I->getNextNode()) { | |||
2937 | auto *SD = getScheduleData(I); | |||
2938 | if (!SD) | |||
2939 | continue; | |||
2940 | assert(isInSchedulingRegion(SD) &&(static_cast <bool> (isInSchedulingRegion(SD) && "primary schedule data not in window?") ? void (0) : __assert_fail ("isInSchedulingRegion(SD) && \"primary schedule data not in window?\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2941, __extension__ __PRETTY_FUNCTION__)) | |||
2941 | "primary schedule data not in window?")(static_cast <bool> (isInSchedulingRegion(SD) && "primary schedule data not in window?") ? void (0) : __assert_fail ("isInSchedulingRegion(SD) && \"primary schedule data not in window?\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2941, __extension__ __PRETTY_FUNCTION__)); | |||
2942 | assert(isInSchedulingRegion(SD->FirstInBundle) &&(static_cast <bool> (isInSchedulingRegion(SD->FirstInBundle ) && "entire bundle in window!") ? void (0) : __assert_fail ("isInSchedulingRegion(SD->FirstInBundle) && \"entire bundle in window!\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2943, __extension__ __PRETTY_FUNCTION__)) | |||
2943 | "entire bundle in window!")(static_cast <bool> (isInSchedulingRegion(SD->FirstInBundle ) && "entire bundle in window!") ? void (0) : __assert_fail ("isInSchedulingRegion(SD->FirstInBundle) && \"entire bundle in window!\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2943, __extension__ __PRETTY_FUNCTION__)); | |||
2944 | (void)SD; | |||
2945 | doForAllOpcodes(I, [](ScheduleData *SD) { SD->verify(); }); | |||
2946 | } | |||
2947 | ||||
2948 | for (auto *SD : ReadyInsts) { | |||
2949 | assert(SD->isSchedulingEntity() && SD->isReady() &&(static_cast <bool> (SD->isSchedulingEntity() && SD->isReady() && "item in ready list not ready?") ? void (0) : __assert_fail ("SD->isSchedulingEntity() && SD->isReady() && \"item in ready list not ready?\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2950, __extension__ __PRETTY_FUNCTION__)) | |||
2950 | "item in ready list not ready?")(static_cast <bool> (SD->isSchedulingEntity() && SD->isReady() && "item in ready list not ready?") ? void (0) : __assert_fail ("SD->isSchedulingEntity() && SD->isReady() && \"item in ready list not ready?\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2950, __extension__ __PRETTY_FUNCTION__)); | |||
2951 | (void)SD; | |||
2952 | } | |||
2953 | } | |||
2954 | ||||
2955 | void doForAllOpcodes(Value *V, | |||
2956 | function_ref<void(ScheduleData *SD)> Action) { | |||
2957 | if (ScheduleData *SD = getScheduleData(V)) | |||
2958 | Action(SD); | |||
2959 | auto I = ExtraScheduleDataMap.find(V); | |||
2960 | if (I != ExtraScheduleDataMap.end()) | |||
2961 | for (auto &P : I->second) | |||
2962 | if (isInSchedulingRegion(P.second)) | |||
2963 | Action(P.second); | |||
2964 | } | |||
2965 | ||||
2966 | /// Put all instructions into the ReadyList which are ready for scheduling. | |||
2967 | template <typename ReadyListType> | |||
2968 | void initialFillReadyList(ReadyListType &ReadyList) { | |||
2969 | for (auto *I = ScheduleStart; I != ScheduleEnd; I = I->getNextNode()) { | |||
2970 | doForAllOpcodes(I, [&](ScheduleData *SD) { | |||
2971 | if (SD->isSchedulingEntity() && SD->hasValidDependencies() && | |||
2972 | SD->isReady()) { | |||
2973 | ReadyList.insert(SD); | |||
2974 | LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: initially in ready list: " << *SD << "\n"; } } while (false) | |||
2975 | << "SLP: initially in ready list: " << *SD << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: initially in ready list: " << *SD << "\n"; } } while (false); | |||
2976 | } | |||
2977 | }); | |||
2978 | } | |||
2979 | } | |||
2980 | ||||
2981 | /// Build a bundle from the ScheduleData nodes corresponding to the | |||
2982 | /// scalar instruction for each lane. | |||
2983 | ScheduleData *buildBundle(ArrayRef<Value *> VL); | |||
2984 | ||||
2985 | /// Checks if a bundle of instructions can be scheduled, i.e. has no | |||
2986 | /// cyclic dependencies. This is only a dry-run, no instructions are | |||
2987 | /// actually moved at this stage. | |||
2988 | /// \returns the scheduling bundle. The returned Optional value is non-None | |||
2989 | /// if \p VL is allowed to be scheduled. | |||
2990 | Optional<ScheduleData *> | |||
2991 | tryScheduleBundle(ArrayRef<Value *> VL, BoUpSLP *SLP, | |||
2992 | const InstructionsState &S); | |||
2993 | ||||
2994 | /// Un-bundles a group of instructions. | |||
2995 | void cancelScheduling(ArrayRef<Value *> VL, Value *OpValue); | |||
2996 | ||||
2997 | /// Allocates schedule data chunk. | |||
2998 | ScheduleData *allocateScheduleDataChunks(); | |||
2999 | ||||
3000 | /// Extends the scheduling region so that V is inside the region. | |||
3001 | /// \returns true if the region size is within the limit. | |||
3002 | bool extendSchedulingRegion(Value *V, const InstructionsState &S); | |||
3003 | ||||
3004 | /// Initialize the ScheduleData structures for new instructions in the | |||
3005 | /// scheduling region. | |||
3006 | void initScheduleData(Instruction *FromI, Instruction *ToI, | |||
3007 | ScheduleData *PrevLoadStore, | |||
3008 | ScheduleData *NextLoadStore); | |||
3009 | ||||
3010 | /// Updates the dependency information of a bundle and of all instructions/ | |||
3011 | /// bundles which depend on the original bundle. | |||
3012 | void calculateDependencies(ScheduleData *SD, bool InsertInReadyList, | |||
3013 | BoUpSLP *SLP); | |||
3014 | ||||
3015 | /// Sets all instruction in the scheduling region to un-scheduled. | |||
3016 | void resetSchedule(); | |||
3017 | ||||
3018 | BasicBlock *BB; | |||
3019 | ||||
3020 | /// Simple memory allocation for ScheduleData. | |||
3021 | std::vector<std::unique_ptr<ScheduleData[]>> ScheduleDataChunks; | |||
3022 | ||||
3023 | /// The size of a ScheduleData array in ScheduleDataChunks. | |||
3024 | int ChunkSize; | |||
3025 | ||||
3026 | /// The allocator position in the current chunk, which is the last entry | |||
3027 | /// of ScheduleDataChunks. | |||
3028 | int ChunkPos; | |||
3029 | ||||
3030 | /// Attaches ScheduleData to Instruction. | |||
3031 | /// Note that the mapping survives during all vectorization iterations, i.e. | |||
3032 | /// ScheduleData structures are recycled. | |||
3033 | DenseMap<Instruction *, ScheduleData *> ScheduleDataMap; | |||
3034 | ||||
3035 | /// Attaches ScheduleData to Instruction with the leading key. | |||
3036 | DenseMap<Value *, SmallDenseMap<Value *, ScheduleData *>> | |||
3037 | ExtraScheduleDataMap; | |||
3038 | ||||
3039 | /// The ready-list for scheduling (only used for the dry-run). | |||
3040 | SetVector<ScheduleData *> ReadyInsts; | |||
3041 | ||||
3042 | /// The first instruction of the scheduling region. | |||
3043 | Instruction *ScheduleStart = nullptr; | |||
3044 | ||||
3045 | /// The first instruction _after_ the scheduling region. | |||
3046 | Instruction *ScheduleEnd = nullptr; | |||
3047 | ||||
3048 | /// The first memory accessing instruction in the scheduling region | |||
3049 | /// (can be null). | |||
3050 | ScheduleData *FirstLoadStoreInRegion = nullptr; | |||
3051 | ||||
3052 | /// The last memory accessing instruction in the scheduling region | |||
3053 | /// (can be null). | |||
3054 | ScheduleData *LastLoadStoreInRegion = nullptr; | |||
3055 | ||||
3056 | /// Is there an llvm.stacksave or llvm.stackrestore in the scheduling | |||
3057 | /// region? Used to optimize the dependence calculation for the | |||
3058 | /// common case where there isn't. | |||
3059 | bool RegionHasStackSave = false; | |||
3060 | ||||
3061 | /// The current size of the scheduling region. | |||
3062 | int ScheduleRegionSize = 0; | |||
3063 | ||||
3064 | /// The maximum size allowed for the scheduling region. | |||
3065 | int ScheduleRegionSizeLimit = ScheduleRegionSizeBudget; | |||
3066 | ||||
3067 | /// The ID of the scheduling region. For a new vectorization iteration this | |||
3068 | /// is incremented which "removes" all ScheduleData from the region. | |||
3069 | /// Make sure that the initial SchedulingRegionID is greater than the | |||
3070 | /// initial SchedulingRegionID in ScheduleData (which is 0). | |||
3071 | int SchedulingRegionID = 1; | |||
3072 | }; | |||
3073 | ||||
3074 | /// Attaches the BlockScheduling structures to basic blocks. | |||
3075 | MapVector<BasicBlock *, std::unique_ptr<BlockScheduling>> BlocksSchedules; | |||
3076 | ||||
3077 | /// Performs the "real" scheduling. Done before vectorization is actually | |||
3078 | /// performed in a basic block. | |||
3079 | void scheduleBlock(BlockScheduling *BS); | |||
3080 | ||||
3081 | /// List of users to ignore during scheduling and that don't need extracting. | |||
3082 | ArrayRef<Value *> UserIgnoreList; | |||
3083 | ||||
3084 | /// A DenseMapInfo implementation for holding DenseMaps and DenseSets of | |||
3085 | /// sorted SmallVectors of unsigned. | |||
3086 | struct OrdersTypeDenseMapInfo { | |||
3087 | static OrdersType getEmptyKey() { | |||
3088 | OrdersType V; | |||
3089 | V.push_back(~1U); | |||
3090 | return V; | |||
3091 | } | |||
3092 | ||||
3093 | static OrdersType getTombstoneKey() { | |||
3094 | OrdersType V; | |||
3095 | V.push_back(~2U); | |||
3096 | return V; | |||
3097 | } | |||
3098 | ||||
3099 | static unsigned getHashValue(const OrdersType &V) { | |||
3100 | return static_cast<unsigned>(hash_combine_range(V.begin(), V.end())); | |||
3101 | } | |||
3102 | ||||
3103 | static bool isEqual(const OrdersType &LHS, const OrdersType &RHS) { | |||
3104 | return LHS == RHS; | |||
3105 | } | |||
3106 | }; | |||
3107 | ||||
3108 | // Analysis and block reference. | |||
3109 | Function *F; | |||
3110 | ScalarEvolution *SE; | |||
3111 | TargetTransformInfo *TTI; | |||
3112 | TargetLibraryInfo *TLI; | |||
3113 | LoopInfo *LI; | |||
3114 | DominatorTree *DT; | |||
3115 | AssumptionCache *AC; | |||
3116 | DemandedBits *DB; | |||
3117 | const DataLayout *DL; | |||
3118 | OptimizationRemarkEmitter *ORE; | |||
3119 | ||||
3120 | unsigned MaxVecRegSize; // This is set by TTI or overridden by cl::opt. | |||
3121 | unsigned MinVecRegSize; // Set by cl::opt (default: 128). | |||
3122 | ||||
3123 | /// Instruction builder to construct the vectorized tree. | |||
3124 | IRBuilder<> Builder; | |||
3125 | ||||
3126 | /// A map of scalar integer values to the smallest bit width with which they | |||
3127 | /// can legally be represented. The values map to (width, signed) pairs, | |||
3128 | /// where "width" indicates the minimum bit width and "signed" is True if the | |||
3129 | /// value must be signed-extended, rather than zero-extended, back to its | |||
3130 | /// original width. | |||
3131 | MapVector<Value *, std::pair<uint64_t, bool>> MinBWs; | |||
3132 | }; | |||
3133 | ||||
3134 | } // end namespace slpvectorizer | |||
3135 | ||||
3136 | template <> struct GraphTraits<BoUpSLP *> { | |||
3137 | using TreeEntry = BoUpSLP::TreeEntry; | |||
3138 | ||||
3139 | /// NodeRef has to be a pointer per the GraphWriter. | |||
3140 | using NodeRef = TreeEntry *; | |||
3141 | ||||
3142 | using ContainerTy = BoUpSLP::TreeEntry::VecTreeTy; | |||
3143 | ||||
3144 | /// Add the VectorizableTree to the index iterator to be able to return | |||
3145 | /// TreeEntry pointers. | |||
3146 | struct ChildIteratorType | |||
3147 | : public iterator_adaptor_base< | |||
3148 | ChildIteratorType, SmallVector<BoUpSLP::EdgeInfo, 1>::iterator> { | |||
3149 | ContainerTy &VectorizableTree; | |||
3150 | ||||
3151 | ChildIteratorType(SmallVector<BoUpSLP::EdgeInfo, 1>::iterator W, | |||
3152 | ContainerTy &VT) | |||
3153 | : ChildIteratorType::iterator_adaptor_base(W), VectorizableTree(VT) {} | |||
3154 | ||||
3155 | NodeRef operator*() { return I->UserTE; } | |||
3156 | }; | |||
3157 | ||||
3158 | static NodeRef getEntryNode(BoUpSLP &R) { | |||
3159 | return R.VectorizableTree[0].get(); | |||
3160 | } | |||
3161 | ||||
3162 | static ChildIteratorType child_begin(NodeRef N) { | |||
3163 | return {N->UserTreeIndices.begin(), N->Container}; | |||
3164 | } | |||
3165 | ||||
3166 | static ChildIteratorType child_end(NodeRef N) { | |||
3167 | return {N->UserTreeIndices.end(), N->Container}; | |||
3168 | } | |||
3169 | ||||
3170 | /// For the node iterator we just need to turn the TreeEntry iterator into a | |||
3171 | /// TreeEntry* iterator so that it dereferences to NodeRef. | |||
3172 | class nodes_iterator { | |||
3173 | using ItTy = ContainerTy::iterator; | |||
3174 | ItTy It; | |||
3175 | ||||
3176 | public: | |||
3177 | nodes_iterator(const ItTy &It2) : It(It2) {} | |||
3178 | NodeRef operator*() { return It->get(); } | |||
3179 | nodes_iterator operator++() { | |||
3180 | ++It; | |||
3181 | return *this; | |||
3182 | } | |||
3183 | bool operator!=(const nodes_iterator &N2) const { return N2.It != It; } | |||
3184 | }; | |||
3185 | ||||
3186 | static nodes_iterator nodes_begin(BoUpSLP *R) { | |||
3187 | return nodes_iterator(R->VectorizableTree.begin()); | |||
3188 | } | |||
3189 | ||||
3190 | static nodes_iterator nodes_end(BoUpSLP *R) { | |||
3191 | return nodes_iterator(R->VectorizableTree.end()); | |||
3192 | } | |||
3193 | ||||
3194 | static unsigned size(BoUpSLP *R) { return R->VectorizableTree.size(); } | |||
3195 | }; | |||
3196 | ||||
3197 | template <> struct DOTGraphTraits<BoUpSLP *> : public DefaultDOTGraphTraits { | |||
3198 | using TreeEntry = BoUpSLP::TreeEntry; | |||
3199 | ||||
3200 | DOTGraphTraits(bool isSimple = false) : DefaultDOTGraphTraits(isSimple) {} | |||
3201 | ||||
3202 | std::string getNodeLabel(const TreeEntry *Entry, const BoUpSLP *R) { | |||
3203 | std::string Str; | |||
3204 | raw_string_ostream OS(Str); | |||
3205 | if (isSplat(Entry->Scalars)) | |||
3206 | OS << "<splat> "; | |||
3207 | for (auto V : Entry->Scalars) { | |||
3208 | OS << *V; | |||
3209 | if (llvm::any_of(R->ExternalUses, [&](const BoUpSLP::ExternalUser &EU) { | |||
3210 | return EU.Scalar == V; | |||
3211 | })) | |||
3212 | OS << " <extract>"; | |||
3213 | OS << "\n"; | |||
3214 | } | |||
3215 | return Str; | |||
3216 | } | |||
3217 | ||||
3218 | static std::string getNodeAttributes(const TreeEntry *Entry, | |||
3219 | const BoUpSLP *) { | |||
3220 | if (Entry->State == TreeEntry::NeedToGather) | |||
3221 | return "color=red"; | |||
3222 | return ""; | |||
3223 | } | |||
3224 | }; | |||
3225 | ||||
3226 | } // end namespace llvm | |||
3227 | ||||
3228 | BoUpSLP::~BoUpSLP() { | |||
3229 | SmallVector<WeakTrackingVH> DeadInsts; | |||
3230 | for (auto *I : DeletedInstructions) { | |||
3231 | for (Use &U : I->operands()) { | |||
3232 | auto *Op = dyn_cast<Instruction>(U.get()); | |||
3233 | if (Op && !DeletedInstructions.count(Op) && Op->hasOneUser() && | |||
3234 | wouldInstructionBeTriviallyDead(Op, TLI)) | |||
3235 | DeadInsts.emplace_back(Op); | |||
3236 | } | |||
3237 | I->dropAllReferences(); | |||
3238 | } | |||
3239 | for (auto *I : DeletedInstructions) { | |||
3240 | assert(I->use_empty() &&(static_cast <bool> (I->use_empty() && "trying to erase instruction with users." ) ? void (0) : __assert_fail ("I->use_empty() && \"trying to erase instruction with users.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 3241, __extension__ __PRETTY_FUNCTION__)) | |||
3241 | "trying to erase instruction with users.")(static_cast <bool> (I->use_empty() && "trying to erase instruction with users." ) ? void (0) : __assert_fail ("I->use_empty() && \"trying to erase instruction with users.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 3241, __extension__ __PRETTY_FUNCTION__)); | |||
3242 | I->eraseFromParent(); | |||
3243 | } | |||
3244 | ||||
3245 | // Cleanup any dead scalar code feeding the vectorized instructions | |||
3246 | RecursivelyDeleteTriviallyDeadInstructions(DeadInsts, TLI); | |||
3247 | ||||
3248 | #ifdef EXPENSIVE_CHECKS | |||
3249 | // If we could guarantee that this call is not extremely slow, we could | |||
3250 | // remove the ifdef limitation (see PR47712). | |||
3251 | assert(!verifyFunction(*F, &dbgs()))(static_cast <bool> (!verifyFunction(*F, &dbgs())) ? void (0) : __assert_fail ("!verifyFunction(*F, &dbgs())" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 3251, __extension__ __PRETTY_FUNCTION__)); | |||
3252 | #endif | |||
3253 | } | |||
3254 | ||||
3255 | /// Reorders the given \p Reuses mask according to the given \p Mask. \p Reuses | |||
3256 | /// contains original mask for the scalars reused in the node. Procedure | |||
3257 | /// transform this mask in accordance with the given \p Mask. | |||
3258 | static void reorderReuses(SmallVectorImpl<int> &Reuses, ArrayRef<int> Mask) { | |||
3259 | assert(!Mask.empty() && Reuses.size() == Mask.size() &&(static_cast <bool> (!Mask.empty() && Reuses.size () == Mask.size() && "Expected non-empty mask.") ? void (0) : __assert_fail ("!Mask.empty() && Reuses.size() == Mask.size() && \"Expected non-empty mask.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 3260, __extension__ __PRETTY_FUNCTION__)) | |||
3260 | "Expected non-empty mask.")(static_cast <bool> (!Mask.empty() && Reuses.size () == Mask.size() && "Expected non-empty mask.") ? void (0) : __assert_fail ("!Mask.empty() && Reuses.size() == Mask.size() && \"Expected non-empty mask.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 3260, __extension__ __PRETTY_FUNCTION__)); | |||
3261 | SmallVector<int> Prev(Reuses.begin(), Reuses.end()); | |||
3262 | Prev.swap(Reuses); | |||
3263 | for (unsigned I = 0, E = Prev.size(); I < E; ++I) | |||
3264 | if (Mask[I] != UndefMaskElem) | |||
3265 | Reuses[Mask[I]] = Prev[I]; | |||
3266 | } | |||
3267 | ||||
3268 | /// Reorders the given \p Order according to the given \p Mask. \p Order - is | |||
3269 | /// the original order of the scalars. Procedure transforms the provided order | |||
3270 | /// in accordance with the given \p Mask. If the resulting \p Order is just an | |||
3271 | /// identity order, \p Order is cleared. | |||
3272 | static void reorderOrder(SmallVectorImpl<unsigned> &Order, ArrayRef<int> Mask) { | |||
3273 | assert(!Mask.empty() && "Expected non-empty mask.")(static_cast <bool> (!Mask.empty() && "Expected non-empty mask." ) ? void (0) : __assert_fail ("!Mask.empty() && \"Expected non-empty mask.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 3273, __extension__ __PRETTY_FUNCTION__)); | |||
3274 | SmallVector<int> MaskOrder; | |||
3275 | if (Order.empty()) { | |||
3276 | MaskOrder.resize(Mask.size()); | |||
3277 | std::iota(MaskOrder.begin(), MaskOrder.end(), 0); | |||
3278 | } else { | |||
3279 | inversePermutation(Order, MaskOrder); | |||
3280 | } | |||
3281 | reorderReuses(MaskOrder, Mask); | |||
3282 | if (ShuffleVectorInst::isIdentityMask(MaskOrder)) { | |||
3283 | Order.clear(); | |||
3284 | return; | |||
3285 | } | |||
3286 | Order.assign(Mask.size(), Mask.size()); | |||
3287 | for (unsigned I = 0, E = Mask.size(); I < E; ++I) | |||
3288 | if (MaskOrder[I] != UndefMaskElem) | |||
3289 | Order[MaskOrder[I]] = I; | |||
3290 | fixupOrderingIndices(Order); | |||
3291 | } | |||
3292 | ||||
3293 | Optional<BoUpSLP::OrdersType> | |||
3294 | BoUpSLP::findReusedOrderedScalars(const BoUpSLP::TreeEntry &TE) { | |||
3295 | assert(TE.State == TreeEntry::NeedToGather && "Expected gather node only.")(static_cast <bool> (TE.State == TreeEntry::NeedToGather && "Expected gather node only.") ? void (0) : __assert_fail ("TE.State == TreeEntry::NeedToGather && \"Expected gather node only.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 3295, __extension__ __PRETTY_FUNCTION__)); | |||
3296 | unsigned NumScalars = TE.Scalars.size(); | |||
3297 | OrdersType CurrentOrder(NumScalars, NumScalars); | |||
3298 | SmallVector<int> Positions; | |||
3299 | SmallBitVector UsedPositions(NumScalars); | |||
3300 | const TreeEntry *STE = nullptr; | |||
3301 | // Try to find all gathered scalars that are gets vectorized in other | |||
3302 | // vectorize node. Here we can have only one single tree vector node to | |||
3303 | // correctly identify order of the gathered scalars. | |||
3304 | for (unsigned I = 0; I < NumScalars; ++I) { | |||
3305 | Value *V = TE.Scalars[I]; | |||
3306 | if (!isa<LoadInst, ExtractElementInst, ExtractValueInst>(V)) | |||
3307 | continue; | |||
3308 | if (const auto *LocalSTE = getTreeEntry(V)) { | |||
3309 | if (!STE) | |||
3310 | STE = LocalSTE; | |||
3311 | else if (STE != LocalSTE) | |||
3312 | // Take the order only from the single vector node. | |||
3313 | return None; | |||
3314 | unsigned Lane = | |||
3315 | std::distance(STE->Scalars.begin(), find(STE->Scalars, V)); | |||
3316 | if (Lane >= NumScalars) | |||
3317 | return None; | |||
3318 | if (CurrentOrder[Lane] != NumScalars) { | |||
3319 | if (Lane != I) | |||
3320 | continue; | |||
3321 | UsedPositions.reset(CurrentOrder[Lane]); | |||
3322 | } | |||
3323 | // The partial identity (where only some elements of the gather node are | |||
3324 | // in the identity order) is good. | |||
3325 | CurrentOrder[Lane] = I; | |||
3326 | UsedPositions.set(I); | |||
3327 | } | |||
3328 | } | |||
3329 | // Need to keep the order if we have a vector entry and at least 2 scalars or | |||
3330 | // the vectorized entry has just 2 scalars. | |||
3331 | if (STE && (UsedPositions.count() > 1 || STE->Scalars.size() == 2)) { | |||
3332 | auto &&IsIdentityOrder = [NumScalars](ArrayRef<unsigned> CurrentOrder) { | |||
3333 | for (unsigned I = 0; I < NumScalars; ++I) | |||
3334 | if (CurrentOrder[I] != I && CurrentOrder[I] != NumScalars) | |||
3335 | return false; | |||
3336 | return true; | |||
3337 | }; | |||
3338 | if (IsIdentityOrder(CurrentOrder)) { | |||
3339 | CurrentOrder.clear(); | |||
3340 | return CurrentOrder; | |||
3341 | } | |||
3342 | auto *It = CurrentOrder.begin(); | |||
3343 | for (unsigned I = 0; I < NumScalars;) { | |||
3344 | if (UsedPositions.test(I)) { | |||
3345 | ++I; | |||
3346 | continue; | |||
3347 | } | |||
3348 | if (*It == NumScalars) { | |||
3349 | *It = I; | |||
3350 | ++I; | |||
3351 | } | |||
3352 | ++It; | |||
3353 | } | |||
3354 | return CurrentOrder; | |||
3355 | } | |||
3356 | return None; | |||
3357 | } | |||
3358 | ||||
3359 | Optional<BoUpSLP::OrdersType> BoUpSLP::getReorderingData(const TreeEntry &TE, | |||
3360 | bool TopToBottom) { | |||
3361 | // No need to reorder if need to shuffle reuses, still need to shuffle the | |||
3362 | // node. | |||
3363 | if (!TE.ReuseShuffleIndices.empty()) | |||
3364 | return None; | |||
3365 | if (TE.State == TreeEntry::Vectorize && | |||
3366 | (isa<LoadInst, ExtractElementInst, ExtractValueInst>(TE.getMainOp()) || | |||
3367 | (TopToBottom && isa<StoreInst, InsertElementInst>(TE.getMainOp()))) && | |||
3368 | !TE.isAltShuffle()) | |||
3369 | return TE.ReorderIndices; | |||
3370 | if (TE.State == TreeEntry::NeedToGather) { | |||
3371 | // TODO: add analysis of other gather nodes with extractelement | |||
3372 | // instructions and other values/instructions, not only undefs. | |||
3373 | if (((TE.getOpcode() == Instruction::ExtractElement && | |||
3374 | !TE.isAltShuffle()) || | |||
3375 | (all_of(TE.Scalars, | |||
3376 | [](Value *V) { | |||
3377 | return isa<UndefValue, ExtractElementInst>(V); | |||
3378 | }) && | |||
3379 | any_of(TE.Scalars, | |||
3380 | [](Value *V) { return isa<ExtractElementInst>(V); }))) && | |||
3381 | all_of(TE.Scalars, | |||
3382 | [](Value *V) { | |||
3383 | auto *EE = dyn_cast<ExtractElementInst>(V); | |||
3384 | return !EE || isa<FixedVectorType>(EE->getVectorOperandType()); | |||
3385 | }) && | |||
3386 | allSameType(TE.Scalars)) { | |||
3387 | // Check that gather of extractelements can be represented as | |||
3388 | // just a shuffle of a single vector. | |||
3389 | OrdersType CurrentOrder; | |||
3390 | bool Reuse = canReuseExtract(TE.Scalars, TE.getMainOp(), CurrentOrder); | |||
3391 | if (Reuse || !CurrentOrder.empty()) { | |||
3392 | if (!CurrentOrder.empty()) | |||
3393 | fixupOrderingIndices(CurrentOrder); | |||
3394 | return CurrentOrder; | |||
3395 | } | |||
3396 | } | |||
3397 | if (Optional<OrdersType> CurrentOrder = findReusedOrderedScalars(TE)) | |||
3398 | return CurrentOrder; | |||
3399 | } | |||
3400 | return None; | |||
3401 | } | |||
3402 | ||||
3403 | void BoUpSLP::reorderTopToBottom() { | |||
3404 | // Maps VF to the graph nodes. | |||
3405 | DenseMap<unsigned, SetVector<TreeEntry *>> VFToOrderedEntries; | |||
3406 | // ExtractElement gather nodes which can be vectorized and need to handle | |||
3407 | // their ordering. | |||
3408 | DenseMap<const TreeEntry *, OrdersType> GathersToOrders; | |||
3409 | // Find all reorderable nodes with the given VF. | |||
3410 | // Currently the are vectorized stores,loads,extracts + some gathering of | |||
3411 | // extracts. | |||
3412 | for_each(VectorizableTree, [this, &VFToOrderedEntries, &GathersToOrders]( | |||
3413 | const std::unique_ptr<TreeEntry> &TE) { | |||
3414 | if (Optional<OrdersType> CurrentOrder = | |||
3415 | getReorderingData(*TE, /*TopToBottom=*/true)) { | |||
3416 | // Do not include ordering for nodes used in the alt opcode vectorization, | |||
3417 | // better to reorder them during bottom-to-top stage. If follow the order | |||
3418 | // here, it causes reordering of the whole graph though actually it is | |||
3419 | // profitable just to reorder the subgraph that starts from the alternate | |||
3420 | // opcode vectorization node. Such nodes already end-up with the shuffle | |||
3421 | // instruction and it is just enough to change this shuffle rather than | |||
3422 | // rotate the scalars for the whole graph. | |||
3423 | unsigned Cnt = 0; | |||
3424 | const TreeEntry *UserTE = TE.get(); | |||
3425 | while (UserTE && Cnt < RecursionMaxDepth) { | |||
3426 | if (UserTE->UserTreeIndices.size() != 1) | |||
3427 | break; | |||
3428 | if (all_of(UserTE->UserTreeIndices, [](const EdgeInfo &EI) { | |||
3429 | return EI.UserTE->State == TreeEntry::Vectorize && | |||
3430 | EI.UserTE->isAltShuffle() && EI.UserTE->Idx != 0; | |||
3431 | })) | |||
3432 | return; | |||
3433 | if (UserTE->UserTreeIndices.empty()) | |||
3434 | UserTE = nullptr; | |||
3435 | else | |||
3436 | UserTE = UserTE->UserTreeIndices.back().UserTE; | |||
3437 | ++Cnt; | |||
3438 | } | |||
3439 | VFToOrderedEntries[TE->Scalars.size()].insert(TE.get()); | |||
3440 | if (TE->State != TreeEntry::Vectorize) | |||
3441 | GathersToOrders.try_emplace(TE.get(), *CurrentOrder); | |||
3442 | } | |||
3443 | }); | |||
3444 | ||||
3445 | // Reorder the graph nodes according to their vectorization factor. | |||
3446 | for (unsigned VF = VectorizableTree.front()->Scalars.size(); VF > 1; | |||
3447 | VF /= 2) { | |||
3448 | auto It = VFToOrderedEntries.find(VF); | |||
3449 | if (It == VFToOrderedEntries.end()) | |||
3450 | continue; | |||
3451 | // Try to find the most profitable order. We just are looking for the most | |||
3452 | // used order and reorder scalar elements in the nodes according to this | |||
3453 | // mostly used order. | |||
3454 | ArrayRef<TreeEntry *> OrderedEntries = It->second.getArrayRef(); | |||
3455 | // All operands are reordered and used only in this node - propagate the | |||
3456 | // most used order to the user node. | |||
3457 | MapVector<OrdersType, unsigned, | |||
3458 | DenseMap<OrdersType, unsigned, OrdersTypeDenseMapInfo>> | |||
3459 | OrdersUses; | |||
3460 | SmallPtrSet<const TreeEntry *, 4> VisitedOps; | |||
3461 | for (const TreeEntry *OpTE : OrderedEntries) { | |||
3462 | // No need to reorder this nodes, still need to extend and to use shuffle, | |||
3463 | // just need to merge reordering shuffle and the reuse shuffle. | |||
3464 | if (!OpTE->ReuseShuffleIndices.empty()) | |||
3465 | continue; | |||
3466 | // Count number of orders uses. | |||
3467 | const auto &Order = [OpTE, &GathersToOrders]() -> const OrdersType & { | |||
3468 | if (OpTE->State == TreeEntry::NeedToGather) | |||
3469 | return GathersToOrders.find(OpTE)->second; | |||
3470 | return OpTE->ReorderIndices; | |||
3471 | }(); | |||
3472 | // Stores actually store the mask, not the order, need to invert. | |||
3473 | if (OpTE->State == TreeEntry::Vectorize && !OpTE->isAltShuffle() && | |||
3474 | OpTE->getOpcode() == Instruction::Store && !Order.empty()) { | |||
3475 | SmallVector<int> Mask; | |||
3476 | inversePermutation(Order, Mask); | |||
3477 | unsigned E = Order.size(); | |||
3478 | OrdersType CurrentOrder(E, E); | |||
3479 | transform(Mask, CurrentOrder.begin(), [E](int Idx) { | |||
3480 | return Idx == UndefMaskElem ? E : static_cast<unsigned>(Idx); | |||
3481 | }); | |||
3482 | fixupOrderingIndices(CurrentOrder); | |||
3483 | ++OrdersUses.insert(std::make_pair(CurrentOrder, 0)).first->second; | |||
3484 | } else { | |||
3485 | ++OrdersUses.insert(std::make_pair(Order, 0)).first->second; | |||
3486 | } | |||
3487 | } | |||
3488 | // Set order of the user node. | |||
3489 | if (OrdersUses.empty()) | |||
3490 | continue; | |||
3491 | // Choose the most used order. | |||
3492 | ArrayRef<unsigned> BestOrder = OrdersUses.front().first; | |||
3493 | unsigned Cnt = OrdersUses.front().second; | |||
3494 | for (const auto &Pair : drop_begin(OrdersUses)) { | |||
3495 | if (Cnt < Pair.second || (Cnt == Pair.second && Pair.first.empty())) { | |||
3496 | BestOrder = Pair.first; | |||
3497 | Cnt = Pair.second; | |||
3498 | } | |||
3499 | } | |||
3500 | // Set order of the user node. | |||
3501 | if (BestOrder.empty()) | |||
3502 | continue; | |||
3503 | SmallVector<int> Mask; | |||
3504 | inversePermutation(BestOrder, Mask); | |||
3505 | SmallVector<int> MaskOrder(BestOrder.size(), UndefMaskElem); | |||
3506 | unsigned E = BestOrder.size(); | |||
3507 | transform(BestOrder, MaskOrder.begin(), [E](unsigned I) { | |||
3508 | return I < E ? static_cast<int>(I) : UndefMaskElem; | |||
3509 | }); | |||
3510 | // Do an actual reordering, if profitable. | |||
3511 | for (std::unique_ptr<TreeEntry> &TE : VectorizableTree) { | |||
3512 | // Just do the reordering for the nodes with the given VF. | |||
3513 | if (TE->Scalars.size() != VF) { | |||
3514 | if (TE->ReuseShuffleIndices.size() == VF) { | |||
3515 | // Need to reorder the reuses masks of the operands with smaller VF to | |||
3516 | // be able to find the match between the graph nodes and scalar | |||
3517 | // operands of the given node during vectorization/cost estimation. | |||
3518 | assert(all_of(TE->UserTreeIndices,(static_cast <bool> (all_of(TE->UserTreeIndices, [VF , &TE](const EdgeInfo &EI) { return EI.UserTE->Scalars .size() == VF || EI.UserTE->Scalars.size() == TE->Scalars .size(); }) && "All users must be of VF size.") ? void (0) : __assert_fail ("all_of(TE->UserTreeIndices, [VF, &TE](const EdgeInfo &EI) { return EI.UserTE->Scalars.size() == VF || EI.UserTE->Scalars.size() == TE->Scalars.size(); }) && \"All users must be of VF size.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 3524, __extension__ __PRETTY_FUNCTION__)) | |||
3519 | [VF, &TE](const EdgeInfo &EI) {(static_cast <bool> (all_of(TE->UserTreeIndices, [VF , &TE](const EdgeInfo &EI) { return EI.UserTE->Scalars .size() == VF || EI.UserTE->Scalars.size() == TE->Scalars .size(); }) && "All users must be of VF size.") ? void (0) : __assert_fail ("all_of(TE->UserTreeIndices, [VF, &TE](const EdgeInfo &EI) { return EI.UserTE->Scalars.size() == VF || EI.UserTE->Scalars.size() == TE->Scalars.size(); }) && \"All users must be of VF size.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 3524, __extension__ __PRETTY_FUNCTION__)) | |||
3520 | return EI.UserTE->Scalars.size() == VF ||(static_cast <bool> (all_of(TE->UserTreeIndices, [VF , &TE](const EdgeInfo &EI) { return EI.UserTE->Scalars .size() == VF || EI.UserTE->Scalars.size() == TE->Scalars .size(); }) && "All users must be of VF size.") ? void (0) : __assert_fail ("all_of(TE->UserTreeIndices, [VF, &TE](const EdgeInfo &EI) { return EI.UserTE->Scalars.size() == VF || EI.UserTE->Scalars.size() == TE->Scalars.size(); }) && \"All users must be of VF size.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 3524, __extension__ __PRETTY_FUNCTION__)) | |||
3521 | EI.UserTE->Scalars.size() ==(static_cast <bool> (all_of(TE->UserTreeIndices, [VF , &TE](const EdgeInfo &EI) { return EI.UserTE->Scalars .size() == VF || EI.UserTE->Scalars.size() == TE->Scalars .size(); }) && "All users must be of VF size.") ? void (0) : __assert_fail ("all_of(TE->UserTreeIndices, [VF, &TE](const EdgeInfo &EI) { return EI.UserTE->Scalars.size() == VF || EI.UserTE->Scalars.size() == TE->Scalars.size(); }) && \"All users must be of VF size.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 3524, __extension__ __PRETTY_FUNCTION__)) | |||
3522 | TE->Scalars.size();(static_cast <bool> (all_of(TE->UserTreeIndices, [VF , &TE](const EdgeInfo &EI) { return EI.UserTE->Scalars .size() == VF || EI.UserTE->Scalars.size() == TE->Scalars .size(); }) && "All users must be of VF size.") ? void (0) : __assert_fail ("all_of(TE->UserTreeIndices, [VF, &TE](const EdgeInfo &EI) { return EI.UserTE->Scalars.size() == VF || EI.UserTE->Scalars.size() == TE->Scalars.size(); }) && \"All users must be of VF size.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 3524, __extension__ __PRETTY_FUNCTION__)) | |||
3523 | }) &&(static_cast <bool> (all_of(TE->UserTreeIndices, [VF , &TE](const EdgeInfo &EI) { return EI.UserTE->Scalars .size() == VF || EI.UserTE->Scalars.size() == TE->Scalars .size(); }) && "All users must be of VF size.") ? void (0) : __assert_fail ("all_of(TE->UserTreeIndices, [VF, &TE](const EdgeInfo &EI) { return EI.UserTE->Scalars.size() == VF || EI.UserTE->Scalars.size() == TE->Scalars.size(); }) && \"All users must be of VF size.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 3524, __extension__ __PRETTY_FUNCTION__)) | |||
3524 | "All users must be of VF size.")(static_cast <bool> (all_of(TE->UserTreeIndices, [VF , &TE](const EdgeInfo &EI) { return EI.UserTE->Scalars .size() == VF || EI.UserTE->Scalars.size() == TE->Scalars .size(); }) && "All users must be of VF size.") ? void (0) : __assert_fail ("all_of(TE->UserTreeIndices, [VF, &TE](const EdgeInfo &EI) { return EI.UserTE->Scalars.size() == VF || EI.UserTE->Scalars.size() == TE->Scalars.size(); }) && \"All users must be of VF size.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 3524, __extension__ __PRETTY_FUNCTION__)); | |||
3525 | // Update ordering of the operands with the smaller VF than the given | |||
3526 | // one. | |||
3527 | reorderReuses(TE->ReuseShuffleIndices, Mask); | |||
3528 | } | |||
3529 | continue; | |||
3530 | } | |||
3531 | if (TE->State == TreeEntry::Vectorize && | |||
3532 | isa<ExtractElementInst, ExtractValueInst, LoadInst, StoreInst, | |||
3533 | InsertElementInst>(TE->getMainOp()) && | |||
3534 | !TE->isAltShuffle()) { | |||
3535 | // Build correct orders for extract{element,value}, loads and | |||
3536 | // stores. | |||
3537 | reorderOrder(TE->ReorderIndices, Mask); | |||
3538 | if (isa<InsertElementInst, StoreInst>(TE->getMainOp())) | |||
3539 | TE->reorderOperands(Mask); | |||
3540 | } else { | |||
3541 | // Reorder the node and its operands. | |||
3542 | TE->reorderOperands(Mask); | |||
3543 | assert(TE->ReorderIndices.empty() &&(static_cast <bool> (TE->ReorderIndices.empty() && "Expected empty reorder sequence.") ? void (0) : __assert_fail ("TE->ReorderIndices.empty() && \"Expected empty reorder sequence.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 3544, __extension__ __PRETTY_FUNCTION__)) | |||
3544 | "Expected empty reorder sequence.")(static_cast <bool> (TE->ReorderIndices.empty() && "Expected empty reorder sequence.") ? void (0) : __assert_fail ("TE->ReorderIndices.empty() && \"Expected empty reorder sequence.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 3544, __extension__ __PRETTY_FUNCTION__)); | |||
3545 | reorderScalars(TE->Scalars, Mask); | |||
3546 | } | |||
3547 | if (!TE->ReuseShuffleIndices.empty()) { | |||
3548 | // Apply reversed order to keep the original ordering of the reused | |||
3549 | // elements to avoid extra reorder indices shuffling. | |||
3550 | OrdersType CurrentOrder; | |||
3551 | reorderOrder(CurrentOrder, MaskOrder); | |||
3552 | SmallVector<int> NewReuses; | |||
3553 | inversePermutation(CurrentOrder, NewReuses); | |||
3554 | addMask(NewReuses, TE->ReuseShuffleIndices); | |||
3555 | TE->ReuseShuffleIndices.swap(NewReuses); | |||
3556 | } | |||
3557 | } | |||
3558 | } | |||
3559 | } | |||
3560 | ||||
3561 | bool BoUpSLP::canReorderOperands( | |||
3562 | TreeEntry *UserTE, SmallVectorImpl<std::pair<unsigned, TreeEntry *>> &Edges, | |||
3563 | ArrayRef<TreeEntry *> ReorderableGathers, | |||
3564 | SmallVectorImpl<TreeEntry *> &GatherOps) { | |||
3565 | for (unsigned I = 0, E = UserTE->getNumOperands(); I < E; ++I) { | |||
3566 | if (any_of(Edges, [I](const std::pair<unsigned, TreeEntry *> &OpData) { | |||
3567 | return OpData.first == I && | |||
3568 | OpData.second->State == TreeEntry::Vectorize; | |||
3569 | })) | |||
3570 | continue; | |||
3571 | if (TreeEntry *TE = getVectorizedOperand(UserTE, I)) { | |||
3572 | // Do not reorder if operand node is used by many user nodes. | |||
3573 | if (any_of(TE->UserTreeIndices, | |||
3574 | [UserTE](const EdgeInfo &EI) { return EI.UserTE != UserTE; })) | |||
3575 | return false; | |||
3576 | // Add the node to the list of the ordered nodes with the identity | |||
3577 | // order. | |||
3578 | Edges.emplace_back(I, TE); | |||
3579 | continue; | |||
3580 | } | |||
3581 | ArrayRef<Value *> VL = UserTE->getOperand(I); | |||
3582 | TreeEntry *Gather = nullptr; | |||
3583 | if (count_if(ReorderableGathers, [VL, &Gather](TreeEntry *TE) { | |||
3584 | assert(TE->State != TreeEntry::Vectorize &&(static_cast <bool> (TE->State != TreeEntry::Vectorize && "Only non-vectorized nodes are expected.") ? void (0) : __assert_fail ("TE->State != TreeEntry::Vectorize && \"Only non-vectorized nodes are expected.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 3585, __extension__ __PRETTY_FUNCTION__)) | |||
3585 | "Only non-vectorized nodes are expected.")(static_cast <bool> (TE->State != TreeEntry::Vectorize && "Only non-vectorized nodes are expected.") ? void (0) : __assert_fail ("TE->State != TreeEntry::Vectorize && \"Only non-vectorized nodes are expected.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 3585, __extension__ __PRETTY_FUNCTION__)); | |||
3586 | if (TE->isSame(VL)) { | |||
3587 | Gather = TE; | |||
3588 | return true; | |||
3589 | } | |||
3590 | return false; | |||
3591 | }) > 1) | |||
3592 | return false; | |||
3593 | if (Gather) | |||
3594 | GatherOps.push_back(Gather); | |||
3595 | } | |||
3596 | return true; | |||
3597 | } | |||
3598 | ||||
3599 | void BoUpSLP::reorderBottomToTop(bool IgnoreReorder) { | |||
3600 | SetVector<TreeEntry *> OrderedEntries; | |||
3601 | DenseMap<const TreeEntry *, OrdersType> GathersToOrders; | |||
3602 | // Find all reorderable leaf nodes with the given VF. | |||
3603 | // Currently the are vectorized loads,extracts without alternate operands + | |||
3604 | // some gathering of extracts. | |||
3605 | SmallVector<TreeEntry *> NonVectorized; | |||
3606 | for_each(VectorizableTree, [this, &OrderedEntries, &GathersToOrders, | |||
3607 | &NonVectorized]( | |||
3608 | const std::unique_ptr<TreeEntry> &TE) { | |||
3609 | if (TE->State != TreeEntry::Vectorize) | |||
3610 | NonVectorized.push_back(TE.get()); | |||
3611 | if (Optional<OrdersType> CurrentOrder = | |||
3612 | getReorderingData(*TE, /*TopToBottom=*/false)) { | |||
3613 | OrderedEntries.insert(TE.get()); | |||
3614 | if (TE->State != TreeEntry::Vectorize) | |||
3615 | GathersToOrders.try_emplace(TE.get(), *CurrentOrder); | |||
3616 | } | |||
3617 | }); | |||
3618 | ||||
3619 | // 1. Propagate order to the graph nodes, which use only reordered nodes. | |||
3620 | // I.e., if the node has operands, that are reordered, try to make at least | |||
3621 | // one operand order in the natural order and reorder others + reorder the | |||
3622 | // user node itself. | |||
3623 | SmallPtrSet<const TreeEntry *, 4> Visited; | |||
3624 | while (!OrderedEntries.empty()) { | |||
3625 | // 1. Filter out only reordered nodes. | |||
3626 | // 2. If the entry has multiple uses - skip it and jump to the next node. | |||
3627 | MapVector<TreeEntry *, SmallVector<std::pair<unsigned, TreeEntry *>>> Users; | |||
3628 | SmallVector<TreeEntry *> Filtered; | |||
3629 | for (TreeEntry *TE : OrderedEntries) { | |||
3630 | if (!(TE->State == TreeEntry::Vectorize || | |||
3631 | (TE->State == TreeEntry::NeedToGather && | |||
3632 | GathersToOrders.count(TE))) || | |||
3633 | TE->UserTreeIndices.empty() || !TE->ReuseShuffleIndices.empty() || | |||
3634 | !all_of(drop_begin(TE->UserTreeIndices), | |||
3635 | [TE](const EdgeInfo &EI) { | |||
3636 | return EI.UserTE == TE->UserTreeIndices.front().UserTE; | |||
3637 | }) || | |||
3638 | !Visited.insert(TE).second) { | |||
3639 | Filtered.push_back(TE); | |||
3640 | continue; | |||
3641 | } | |||
3642 | // Build a map between user nodes and their operands order to speedup | |||
3643 | // search. The graph currently does not provide this dependency directly. | |||
3644 | for (EdgeInfo &EI : TE->UserTreeIndices) { | |||
3645 | TreeEntry *UserTE = EI.UserTE; | |||
3646 | auto It = Users.find(UserTE); | |||
3647 | if (It == Users.end()) | |||
3648 | It = Users.insert({UserTE, {}}).first; | |||
3649 | It->second.emplace_back(EI.EdgeIdx, TE); | |||
3650 | } | |||
3651 | } | |||
3652 | // Erase filtered entries. | |||
3653 | for_each(Filtered, | |||
3654 | [&OrderedEntries](TreeEntry *TE) { OrderedEntries.remove(TE); }); | |||
3655 | for (auto &Data : Users) { | |||
3656 | // Check that operands are used only in the User node. | |||
3657 | SmallVector<TreeEntry *> GatherOps; | |||
3658 | if (!canReorderOperands(Data.first, Data.second, NonVectorized, | |||
3659 | GatherOps)) { | |||
3660 | for_each(Data.second, | |||
3661 | [&OrderedEntries](const std::pair<unsigned, TreeEntry *> &Op) { | |||
3662 | OrderedEntries.remove(Op.second); | |||
3663 | }); | |||
3664 | continue; | |||
3665 | } | |||
3666 | // All operands are reordered and used only in this node - propagate the | |||
3667 | // most used order to the user node. | |||
3668 | MapVector<OrdersType, unsigned, | |||
3669 | DenseMap<OrdersType, unsigned, OrdersTypeDenseMapInfo>> | |||
3670 | OrdersUses; | |||
3671 | // Do the analysis for each tree entry only once, otherwise the order of | |||
3672 | // the same node my be considered several times, though might be not | |||
3673 | // profitable. | |||
3674 | SmallPtrSet<const TreeEntry *, 4> VisitedOps; | |||
3675 | SmallPtrSet<const TreeEntry *, 4> VisitedUsers; | |||
3676 | for (const auto &Op : Data.second) { | |||
3677 | TreeEntry *OpTE = Op.second; | |||
3678 | if (!VisitedOps.insert(OpTE).second) | |||
3679 | continue; | |||
3680 | if (!OpTE->ReuseShuffleIndices.empty() || | |||
3681 | (IgnoreReorder && OpTE == VectorizableTree.front().get())) | |||
3682 | continue; | |||
3683 | const auto &Order = [OpTE, &GathersToOrders]() -> const OrdersType & { | |||
3684 | if (OpTE->State == TreeEntry::NeedToGather) | |||
3685 | return GathersToOrders.find(OpTE)->second; | |||
3686 | return OpTE->ReorderIndices; | |||
3687 | }(); | |||
3688 | unsigned NumOps = count_if( | |||
3689 | Data.second, [OpTE](const std::pair<unsigned, TreeEntry *> &P) { | |||
3690 | return P.second == OpTE; | |||
3691 | }); | |||
3692 | // Stores actually store the mask, not the order, need to invert. | |||
3693 | if (OpTE->State == TreeEntry::Vectorize && !OpTE->isAltShuffle() && | |||
3694 | OpTE->getOpcode() == Instruction::Store && !Order.empty()) { | |||
3695 | SmallVector<int> Mask; | |||
3696 | inversePermutation(Order, Mask); | |||
3697 | unsigned E = Order.size(); | |||
3698 | OrdersType CurrentOrder(E, E); | |||
3699 | transform(Mask, CurrentOrder.begin(), [E](int Idx) { | |||
3700 | return Idx == UndefMaskElem ? E : static_cast<unsigned>(Idx); | |||
3701 | }); | |||
3702 | fixupOrderingIndices(CurrentOrder); | |||
3703 | OrdersUses.insert(std::make_pair(CurrentOrder, 0)).first->second += | |||
3704 | NumOps; | |||
3705 | } else { | |||
3706 | OrdersUses.insert(std::make_pair(Order, 0)).first->second += NumOps; | |||
3707 | } | |||
3708 | auto Res = OrdersUses.insert(std::make_pair(OrdersType(), 0)); | |||
3709 | const auto &&AllowsReordering = [IgnoreReorder, &GathersToOrders]( | |||
3710 | const TreeEntry *TE) { | |||
3711 | if (!TE->ReorderIndices.empty() || !TE->ReuseShuffleIndices.empty() || | |||
3712 | (TE->State == TreeEntry::Vectorize && TE->isAltShuffle()) || | |||
3713 | (IgnoreReorder && TE->Idx == 0)) | |||
3714 | return true; | |||
3715 | if (TE->State == TreeEntry::NeedToGather) { | |||
3716 | auto It = GathersToOrders.find(TE); | |||
3717 | if (It != GathersToOrders.end()) | |||
3718 | return !It->second.empty(); | |||
3719 | return true; | |||
3720 | } | |||
3721 | return false; | |||
3722 | }; | |||
3723 | for (const EdgeInfo &EI : OpTE->UserTreeIndices) { | |||
3724 | TreeEntry *UserTE = EI.UserTE; | |||
3725 | if (!VisitedUsers.insert(UserTE).second) | |||
3726 | continue; | |||
3727 | // May reorder user node if it requires reordering, has reused | |||
3728 | // scalars, is an alternate op vectorize node or its op nodes require | |||
3729 | // reordering. | |||
3730 | if (AllowsReordering(UserTE)) | |||
3731 | continue; | |||
3732 | // Check if users allow reordering. | |||
3733 | // Currently look up just 1 level of operands to avoid increase of | |||
3734 | // the compile time. | |||
3735 | // Profitable to reorder if definitely more operands allow | |||
3736 | // reordering rather than those with natural order. | |||
3737 | ArrayRef<std::pair<unsigned, TreeEntry *>> Ops = Users[UserTE]; | |||
3738 | if (static_cast<unsigned>(count_if( | |||
3739 | Ops, [UserTE, &AllowsReordering]( | |||
3740 | const std::pair<unsigned, TreeEntry *> &Op) { | |||
3741 | return AllowsReordering(Op.second) && | |||
3742 | all_of(Op.second->UserTreeIndices, | |||
3743 | [UserTE](const EdgeInfo &EI) { | |||
3744 | return EI.UserTE == UserTE; | |||
3745 | }); | |||
3746 | })) <= Ops.size() / 2) | |||
3747 | ++Res.first->second; | |||
3748 | } | |||
3749 | } | |||
3750 | // If no orders - skip current nodes and jump to the next one, if any. | |||
3751 | if (OrdersUses.empty()) { | |||
3752 | for_each(Data.second, | |||
3753 | [&OrderedEntries](const std::pair<unsigned, TreeEntry *> &Op) { | |||
3754 | OrderedEntries.remove(Op.second); | |||
3755 | }); | |||
3756 | continue; | |||
3757 | } | |||
3758 | // Choose the best order. | |||
3759 | ArrayRef<unsigned> BestOrder = OrdersUses.front().first; | |||
3760 | unsigned Cnt = OrdersUses.front().second; | |||
3761 | for (const auto &Pair : drop_begin(OrdersUses)) { | |||
3762 | if (Cnt < Pair.second || (Cnt == Pair.second && Pair.first.empty())) { | |||
3763 | BestOrder = Pair.first; | |||
3764 | Cnt = Pair.second; | |||
3765 | } | |||
3766 | } | |||
3767 | // Set order of the user node (reordering of operands and user nodes). | |||
3768 | if (BestOrder.empty()) { | |||
3769 | for_each(Data.second, | |||
3770 | [&OrderedEntries](const std::pair<unsigned, TreeEntry *> &Op) { | |||
3771 | OrderedEntries.remove(Op.second); | |||
3772 | }); | |||
3773 | continue; | |||
3774 | } | |||
3775 | // Erase operands from OrderedEntries list and adjust their orders. | |||
3776 | VisitedOps.clear(); | |||
3777 | SmallVector<int> Mask; | |||
3778 | inversePermutation(BestOrder, Mask); | |||
3779 | SmallVector<int> MaskOrder(BestOrder.size(), UndefMaskElem); | |||
3780 | unsigned E = BestOrder.size(); | |||
3781 | transform(BestOrder, MaskOrder.begin(), [E](unsigned I) { | |||
3782 | return I < E ? static_cast<int>(I) : UndefMaskElem; | |||
3783 | }); | |||
3784 | for (const std::pair<unsigned, TreeEntry *> &Op : Data.second) { | |||
3785 | TreeEntry *TE = Op.second; | |||
3786 | OrderedEntries.remove(TE); | |||
3787 | if (!VisitedOps.insert(TE).second) | |||
3788 | continue; | |||
3789 | if (TE->ReuseShuffleIndices.size() == BestOrder.size()) { | |||
3790 | // Just reorder reuses indices. | |||
3791 | reorderReuses(TE->ReuseShuffleIndices, Mask); | |||
3792 | continue; | |||
3793 | } | |||
3794 | // Gathers are processed separately. | |||
3795 | if (TE->State != TreeEntry::Vectorize) | |||
3796 | continue; | |||
3797 | assert((BestOrder.size() == TE->ReorderIndices.size() ||(static_cast <bool> ((BestOrder.size() == TE->ReorderIndices .size() || TE->ReorderIndices.empty()) && "Non-matching sizes of user/operand entries." ) ? void (0) : __assert_fail ("(BestOrder.size() == TE->ReorderIndices.size() || TE->ReorderIndices.empty()) && \"Non-matching sizes of user/operand entries.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 3799, __extension__ __PRETTY_FUNCTION__)) | |||
3798 | TE->ReorderIndices.empty()) &&(static_cast <bool> ((BestOrder.size() == TE->ReorderIndices .size() || TE->ReorderIndices.empty()) && "Non-matching sizes of user/operand entries." ) ? void (0) : __assert_fail ("(BestOrder.size() == TE->ReorderIndices.size() || TE->ReorderIndices.empty()) && \"Non-matching sizes of user/operand entries.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 3799, __extension__ __PRETTY_FUNCTION__)) | |||
3799 | "Non-matching sizes of user/operand entries.")(static_cast <bool> ((BestOrder.size() == TE->ReorderIndices .size() || TE->ReorderIndices.empty()) && "Non-matching sizes of user/operand entries." ) ? void (0) : __assert_fail ("(BestOrder.size() == TE->ReorderIndices.size() || TE->ReorderIndices.empty()) && \"Non-matching sizes of user/operand entries.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 3799, __extension__ __PRETTY_FUNCTION__)); | |||
3800 | reorderOrder(TE->ReorderIndices, Mask); | |||
3801 | } | |||
3802 | // For gathers just need to reorder its scalars. | |||
3803 | for (TreeEntry *Gather : GatherOps) { | |||
3804 | assert(Gather->ReorderIndices.empty() &&(static_cast <bool> (Gather->ReorderIndices.empty() && "Unexpected reordering of gathers.") ? void (0) : __assert_fail ("Gather->ReorderIndices.empty() && \"Unexpected reordering of gathers.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 3805, __extension__ __PRETTY_FUNCTION__)) | |||
3805 | "Unexpected reordering of gathers.")(static_cast <bool> (Gather->ReorderIndices.empty() && "Unexpected reordering of gathers.") ? void (0) : __assert_fail ("Gather->ReorderIndices.empty() && \"Unexpected reordering of gathers.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 3805, __extension__ __PRETTY_FUNCTION__)); | |||
3806 | if (!Gather->ReuseShuffleIndices.empty()) { | |||
3807 | // Just reorder reuses indices. | |||
3808 | reorderReuses(Gather->ReuseShuffleIndices, Mask); | |||
3809 | continue; | |||
3810 | } | |||
3811 | reorderScalars(Gather->Scalars, Mask); | |||
3812 | OrderedEntries.remove(Gather); | |||
3813 | } | |||
3814 | // Reorder operands of the user node and set the ordering for the user | |||
3815 | // node itself. | |||
3816 | if (Data.first->State != TreeEntry::Vectorize || | |||
3817 | !isa<ExtractElementInst, ExtractValueInst, LoadInst>( | |||
3818 | Data.first->getMainOp()) || | |||
3819 | Data.first->isAltShuffle()) | |||
3820 | Data.first->reorderOperands(Mask); | |||
3821 | if (!isa<InsertElementInst, StoreInst>(Data.first->getMainOp()) || | |||
3822 | Data.first->isAltShuffle()) { | |||
3823 | reorderScalars(Data.first->Scalars, Mask); | |||
3824 | reorderOrder(Data.first->ReorderIndices, MaskOrder); | |||
3825 | if (Data.first->ReuseShuffleIndices.empty() && | |||
3826 | !Data.first->ReorderIndices.empty() && | |||
3827 | !Data.first->isAltShuffle()) { | |||
3828 | // Insert user node to the list to try to sink reordering deeper in | |||
3829 | // the graph. | |||
3830 | OrderedEntries.insert(Data.first); | |||
3831 | } | |||
3832 | } else { | |||
3833 | reorderOrder(Data.first->ReorderIndices, Mask); | |||
3834 | } | |||
3835 | } | |||
3836 | } | |||
3837 | // If the reordering is unnecessary, just remove the reorder. | |||
3838 | if (IgnoreReorder && !VectorizableTree.front()->ReorderIndices.empty() && | |||
3839 | VectorizableTree.front()->ReuseShuffleIndices.empty()) | |||
3840 | VectorizableTree.front()->ReorderIndices.clear(); | |||
3841 | } | |||
3842 | ||||
3843 | void BoUpSLP::buildExternalUses( | |||
3844 | const ExtraValueToDebugLocsMap &ExternallyUsedValues) { | |||
3845 | // Collect the values that we need to extract from the tree. | |||
3846 | for (auto &TEPtr : VectorizableTree) { | |||
3847 | TreeEntry *Entry = TEPtr.get(); | |||
3848 | ||||
3849 | // No need to handle users of gathered values. | |||
3850 | if (Entry->State == TreeEntry::NeedToGather) | |||
3851 | continue; | |||
3852 | ||||
3853 | // For each lane: | |||
3854 | for (int Lane = 0, LE = Entry->Scalars.size(); Lane != LE; ++Lane) { | |||
3855 | Value *Scalar = Entry->Scalars[Lane]; | |||
3856 | int FoundLane = Entry->findLaneForValue(Scalar); | |||
3857 | ||||
3858 | // Check if the scalar is externally used as an extra arg. | |||
3859 | auto ExtI = ExternallyUsedValues.find(Scalar); | |||
3860 | if (ExtI != ExternallyUsedValues.end()) { | |||
3861 | LLVM_DEBUG(dbgs() << "SLP: Need to extract: Extra arg from lane "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Need to extract: Extra arg from lane " << Lane << " from " << *Scalar << ".\n" ; } } while (false) | |||
3862 | << Lane << " from " << *Scalar << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Need to extract: Extra arg from lane " << Lane << " from " << *Scalar << ".\n" ; } } while (false); | |||
3863 | ExternalUses.emplace_back(Scalar, nullptr, FoundLane); | |||
3864 | } | |||
3865 | for (User *U : Scalar->users()) { | |||
3866 | LLVM_DEBUG(dbgs() << "SLP: Checking user:" << *U << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Checking user:" << *U << ".\n"; } } while (false); | |||
3867 | ||||
3868 | Instruction *UserInst = dyn_cast<Instruction>(U); | |||
3869 | if (!UserInst) | |||
3870 | continue; | |||
3871 | ||||
3872 | if (isDeleted(UserInst)) | |||
3873 | continue; | |||
3874 | ||||
3875 | // Skip in-tree scalars that become vectors | |||
3876 | if (TreeEntry *UseEntry = getTreeEntry(U)) { | |||
3877 | Value *UseScalar = UseEntry->Scalars[0]; | |||
3878 | // Some in-tree scalars will remain as scalar in vectorized | |||
3879 | // instructions. If that is the case, the one in Lane 0 will | |||
3880 | // be used. | |||
3881 | if (UseScalar != U || | |||
3882 | UseEntry->State == TreeEntry::ScatterVectorize || | |||
3883 | !InTreeUserNeedToExtract(Scalar, UserInst, TLI)) { | |||
3884 | LLVM_DEBUG(dbgs() << "SLP: \tInternal user will be removed:" << *Udo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: \tInternal user will be removed:" << *U << ".\n"; } } while (false) | |||
3885 | << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: \tInternal user will be removed:" << *U << ".\n"; } } while (false); | |||
3886 | assert(UseEntry->State != TreeEntry::NeedToGather && "Bad state")(static_cast <bool> (UseEntry->State != TreeEntry::NeedToGather && "Bad state") ? void (0) : __assert_fail ("UseEntry->State != TreeEntry::NeedToGather && \"Bad state\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 3886, __extension__ __PRETTY_FUNCTION__)); | |||
3887 | continue; | |||
3888 | } | |||
3889 | } | |||
3890 | ||||
3891 | // Ignore users in the user ignore list. | |||
3892 | if (is_contained(UserIgnoreList, UserInst)) | |||
3893 | continue; | |||
3894 | ||||
3895 | LLVM_DEBUG(dbgs() << "SLP: Need to extract:" << *U << " from lane "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Need to extract:" << * U << " from lane " << Lane << " from " << *Scalar << ".\n"; } } while (false) | |||
3896 | << Lane << " from " << *Scalar << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Need to extract:" << * U << " from lane " << Lane << " from " << *Scalar << ".\n"; } } while (false); | |||
3897 | ExternalUses.push_back(ExternalUser(Scalar, U, FoundLane)); | |||
3898 | } | |||
3899 | } | |||
3900 | } | |||
3901 | } | |||
3902 | ||||
3903 | void BoUpSLP::buildTree(ArrayRef<Value *> Roots, | |||
3904 | ArrayRef<Value *> UserIgnoreLst) { | |||
3905 | deleteTree(); | |||
3906 | UserIgnoreList = UserIgnoreLst; | |||
3907 | if (!allSameType(Roots)) | |||
3908 | return; | |||
3909 | buildTree_rec(Roots, 0, EdgeInfo()); | |||
3910 | } | |||
3911 | ||||
3912 | namespace { | |||
3913 | /// Tracks the state we can represent the loads in the given sequence. | |||
3914 | enum class LoadsState { Gather, Vectorize, ScatterVectorize }; | |||
3915 | } // anonymous namespace | |||
3916 | ||||
3917 | /// Checks if the given array of loads can be represented as a vectorized, | |||
3918 | /// scatter or just simple gather. | |||
3919 | static LoadsState canVectorizeLoads(ArrayRef<Value *> VL, const Value *VL0, | |||
3920 | const TargetTransformInfo &TTI, | |||
3921 | const DataLayout &DL, ScalarEvolution &SE, | |||
3922 | SmallVectorImpl<unsigned> &Order, | |||
3923 | SmallVectorImpl<Value *> &PointerOps) { | |||
3924 | // Check that a vectorized load would load the same memory as a scalar | |||
3925 | // load. For example, we don't want to vectorize loads that are smaller | |||
3926 | // than 8-bit. Even though we have a packed struct {<i2, i2, i2, i2>} LLVM | |||
3927 | // treats loading/storing it as an i8 struct. If we vectorize loads/stores | |||
3928 | // from such a struct, we read/write packed bits disagreeing with the | |||
3929 | // unvectorized version. | |||
3930 | Type *ScalarTy = VL0->getType(); | |||
3931 | ||||
3932 | if (DL.getTypeSizeInBits(ScalarTy) != DL.getTypeAllocSizeInBits(ScalarTy)) | |||
3933 | return LoadsState::Gather; | |||
3934 | ||||
3935 | // Make sure all loads in the bundle are simple - we can't vectorize | |||
3936 | // atomic or volatile loads. | |||
3937 | PointerOps.clear(); | |||
3938 | PointerOps.resize(VL.size()); | |||
3939 | auto *POIter = PointerOps.begin(); | |||
3940 | for (Value *V : VL) { | |||
3941 | auto *L = cast<LoadInst>(V); | |||
3942 | if (!L->isSimple()) | |||
3943 | return LoadsState::Gather; | |||
3944 | *POIter = L->getPointerOperand(); | |||
3945 | ++POIter; | |||
3946 | } | |||
3947 | ||||
3948 | Order.clear(); | |||
3949 | // Check the order of pointer operands. | |||
3950 | if (llvm::sortPtrAccesses(PointerOps, ScalarTy, DL, SE, Order)) { | |||
3951 | Value *Ptr0; | |||
3952 | Value *PtrN; | |||
3953 | if (Order.empty()) { | |||
3954 | Ptr0 = PointerOps.front(); | |||
3955 | PtrN = PointerOps.back(); | |||
3956 | } else { | |||
3957 | Ptr0 = PointerOps[Order.front()]; | |||
3958 | PtrN = PointerOps[Order.back()]; | |||
3959 | } | |||
3960 | Optional<int> Diff = | |||
3961 | getPointersDiff(ScalarTy, Ptr0, ScalarTy, PtrN, DL, SE); | |||
3962 | // Check that the sorted loads are consecutive. | |||
3963 | if (static_cast<unsigned>(*Diff) == VL.size() - 1) | |||
3964 | return LoadsState::Vectorize; | |||
3965 | Align CommonAlignment = cast<LoadInst>(VL0)->getAlign(); | |||
3966 | for (Value *V : VL) | |||
3967 | CommonAlignment = | |||
3968 | commonAlignment(CommonAlignment, cast<LoadInst>(V)->getAlign()); | |||
3969 | if (TTI.isLegalMaskedGather(FixedVectorType::get(ScalarTy, VL.size()), | |||
3970 | CommonAlignment)) | |||
3971 | return LoadsState::ScatterVectorize; | |||
3972 | } | |||
3973 | ||||
3974 | return LoadsState::Gather; | |||
3975 | } | |||
3976 | ||||
3977 | /// \return true if the specified list of values has only one instruction that | |||
3978 | /// requires scheduling, false otherwise. | |||
3979 | #ifndef NDEBUG | |||
3980 | static bool needToScheduleSingleInstruction(ArrayRef<Value *> VL) { | |||
3981 | Value *NeedsScheduling = nullptr; | |||
3982 | for (Value *V : VL) { | |||
3983 | if (doesNotNeedToBeScheduled(V)) | |||
3984 | continue; | |||
3985 | if (!NeedsScheduling) { | |||
3986 | NeedsScheduling = V; | |||
3987 | continue; | |||
3988 | } | |||
3989 | return false; | |||
3990 | } | |||
3991 | return NeedsScheduling; | |||
3992 | } | |||
3993 | #endif | |||
3994 | ||||
3995 | void BoUpSLP::buildTree_rec(ArrayRef<Value *> VL, unsigned Depth, | |||
3996 | const EdgeInfo &UserTreeIdx) { | |||
3997 | assert((allConstant(VL) || allSameType(VL)) && "Invalid types!")(static_cast <bool> ((allConstant(VL) || allSameType(VL )) && "Invalid types!") ? void (0) : __assert_fail ("(allConstant(VL) || allSameType(VL)) && \"Invalid types!\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 3997, __extension__ __PRETTY_FUNCTION__)); | |||
3998 | ||||
3999 | SmallVector<int> ReuseShuffleIndicies; | |||
4000 | SmallVector<Value *> UniqueValues; | |||
4001 | auto &&TryToFindDuplicates = [&VL, &ReuseShuffleIndicies, &UniqueValues, | |||
4002 | &UserTreeIdx, | |||
4003 | this](const InstructionsState &S) { | |||
4004 | // Check that every instruction appears once in this bundle. | |||
4005 | DenseMap<Value *, unsigned> UniquePositions; | |||
4006 | for (Value *V : VL) { | |||
4007 | if (isConstant(V)) { | |||
4008 | ReuseShuffleIndicies.emplace_back( | |||
4009 | isa<UndefValue>(V) ? UndefMaskElem : UniqueValues.size()); | |||
4010 | UniqueValues.emplace_back(V); | |||
4011 | continue; | |||
4012 | } | |||
4013 | auto Res = UniquePositions.try_emplace(V, UniqueValues.size()); | |||
4014 | ReuseShuffleIndicies.emplace_back(Res.first->second); | |||
4015 | if (Res.second) | |||
4016 | UniqueValues.emplace_back(V); | |||
4017 | } | |||
4018 | size_t NumUniqueScalarValues = UniqueValues.size(); | |||
4019 | if (NumUniqueScalarValues == VL.size()) { | |||
4020 | ReuseShuffleIndicies.clear(); | |||
4021 | } else { | |||
4022 | LLVM_DEBUG(dbgs() << "SLP: Shuffle for reused scalars.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Shuffle for reused scalars.\n" ; } } while (false); | |||
4023 | if (NumUniqueScalarValues <= 1 || | |||
4024 | (UniquePositions.size() == 1 && all_of(UniqueValues, | |||
4025 | [](Value *V) { | |||
4026 | return isa<UndefValue>(V) || | |||
4027 | !isConstant(V); | |||
4028 | })) || | |||
4029 | !llvm::isPowerOf2_32(NumUniqueScalarValues)) { | |||
4030 | LLVM_DEBUG(dbgs() << "SLP: Scalar used twice in bundle.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Scalar used twice in bundle.\n" ; } } while (false); | |||
4031 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx); | |||
4032 | return false; | |||
4033 | } | |||
4034 | VL = UniqueValues; | |||
4035 | } | |||
4036 | return true; | |||
4037 | }; | |||
4038 | ||||
4039 | InstructionsState S = getSameOpcode(VL); | |||
4040 | if (Depth == RecursionMaxDepth) { | |||
4041 | LLVM_DEBUG(dbgs() << "SLP: Gathering due to max recursion depth.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering due to max recursion depth.\n" ; } } while (false); | |||
4042 | if (TryToFindDuplicates(S)) | |||
4043 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | |||
4044 | ReuseShuffleIndicies); | |||
4045 | return; | |||
4046 | } | |||
4047 | ||||
4048 | // Don't handle scalable vectors | |||
4049 | if (S.getOpcode() == Instruction::ExtractElement && | |||
4050 | isa<ScalableVectorType>( | |||
4051 | cast<ExtractElementInst>(S.OpValue)->getVectorOperandType())) { | |||
4052 | LLVM_DEBUG(dbgs() << "SLP: Gathering due to scalable vector type.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering due to scalable vector type.\n" ; } } while (false); | |||
4053 | if (TryToFindDuplicates(S)) | |||
4054 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | |||
4055 | ReuseShuffleIndicies); | |||
4056 | return; | |||
4057 | } | |||
4058 | ||||
4059 | // Don't handle vectors. | |||
4060 | if (S.OpValue->getType()->isVectorTy() && | |||
4061 | !isa<InsertElementInst>(S.OpValue)) { | |||
4062 | LLVM_DEBUG(dbgs() << "SLP: Gathering due to vector type.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering due to vector type.\n" ; } } while (false); | |||
4063 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx); | |||
4064 | return; | |||
4065 | } | |||
4066 | ||||
4067 | if (StoreInst *SI = dyn_cast<StoreInst>(S.OpValue)) | |||
4068 | if (SI->getValueOperand()->getType()->isVectorTy()) { | |||
4069 | LLVM_DEBUG(dbgs() << "SLP: Gathering due to store vector type.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering due to store vector type.\n" ; } } while (false); | |||
4070 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx); | |||
4071 | return; | |||
4072 | } | |||
4073 | ||||
4074 | // If all of the operands are identical or constant we have a simple solution. | |||
4075 | // If we deal with insert/extract instructions, they all must have constant | |||
4076 | // indices, otherwise we should gather them, not try to vectorize. | |||
4077 | if (allConstant(VL) || isSplat(VL) || !allSameBlock(VL) || !S.getOpcode() || | |||
4078 | (isa<InsertElementInst, ExtractValueInst, ExtractElementInst>(S.MainOp) && | |||
4079 | !all_of(VL, isVectorLikeInstWithConstOps))) { | |||
4080 | LLVM_DEBUG(dbgs() << "SLP: Gathering due to C,S,B,O. \n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering due to C,S,B,O. \n" ; } } while (false); | |||
4081 | if (TryToFindDuplicates(S)) | |||
4082 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | |||
4083 | ReuseShuffleIndicies); | |||
4084 | return; | |||
4085 | } | |||
4086 | ||||
4087 | // We now know that this is a vector of instructions of the same type from | |||
4088 | // the same block. | |||
4089 | ||||
4090 | // Don't vectorize ephemeral values. | |||
4091 | for (Value *V : VL) { | |||
4092 | if (EphValues.count(V)) { | |||
4093 | LLVM_DEBUG(dbgs() << "SLP: The instruction (" << *Vdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: The instruction (" << * V << ") is ephemeral.\n"; } } while (false) | |||
4094 | << ") is ephemeral.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: The instruction (" << * V << ") is ephemeral.\n"; } } while (false); | |||
4095 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx); | |||
4096 | return; | |||
4097 | } | |||
4098 | } | |||
4099 | ||||
4100 | // Check if this is a duplicate of another entry. | |||
4101 | if (TreeEntry *E = getTreeEntry(S.OpValue)) { | |||
4102 | LLVM_DEBUG(dbgs() << "SLP: \tChecking bundle: " << *S.OpValue << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: \tChecking bundle: " << *S.OpValue << ".\n"; } } while (false); | |||
4103 | if (!E->isSame(VL)) { | |||
4104 | LLVM_DEBUG(dbgs() << "SLP: Gathering due to partial overlap.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering due to partial overlap.\n" ; } } while (false); | |||
4105 | if (TryToFindDuplicates(S)) | |||
4106 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | |||
4107 | ReuseShuffleIndicies); | |||
4108 | return; | |||
4109 | } | |||
4110 | // Record the reuse of the tree node. FIXME, currently this is only used to | |||
4111 | // properly draw the graph rather than for the actual vectorization. | |||
4112 | E->UserTreeIndices.push_back(UserTreeIdx); | |||
4113 | LLVM_DEBUG(dbgs() << "SLP: Perfect diamond merge at " << *S.OpValuedo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Perfect diamond merge at " << *S.OpValue << ".\n"; } } while (false) | |||
4114 | << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Perfect diamond merge at " << *S.OpValue << ".\n"; } } while (false); | |||
4115 | return; | |||
4116 | } | |||
4117 | ||||
4118 | // Check that none of the instructions in the bundle are already in the tree. | |||
4119 | for (Value *V : VL) { | |||
4120 | auto *I = dyn_cast<Instruction>(V); | |||
4121 | if (!I) | |||
4122 | continue; | |||
4123 | if (getTreeEntry(I)) { | |||
4124 | LLVM_DEBUG(dbgs() << "SLP: The instruction (" << *Vdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: The instruction (" << * V << ") is already in tree.\n"; } } while (false) | |||
4125 | << ") is already in tree.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: The instruction (" << * V << ") is already in tree.\n"; } } while (false); | |||
4126 | if (TryToFindDuplicates(S)) | |||
4127 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | |||
4128 | ReuseShuffleIndicies); | |||
4129 | return; | |||
4130 | } | |||
4131 | } | |||
4132 | ||||
4133 | // The reduction nodes (stored in UserIgnoreList) also should stay scalar. | |||
4134 | for (Value *V : VL) { | |||
4135 | if (is_contained(UserIgnoreList, V)) { | |||
4136 | LLVM_DEBUG(dbgs() << "SLP: Gathering due to gathered scalar.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering due to gathered scalar.\n" ; } } while (false); | |||
4137 | if (TryToFindDuplicates(S)) | |||
4138 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | |||
4139 | ReuseShuffleIndicies); | |||
4140 | return; | |||
4141 | } | |||
4142 | } | |||
4143 | ||||
4144 | // Check that all of the users of the scalars that we want to vectorize are | |||
4145 | // schedulable. | |||
4146 | auto *VL0 = cast<Instruction>(S.OpValue); | |||
4147 | BasicBlock *BB = VL0->getParent(); | |||
4148 | ||||
4149 | if (!DT->isReachableFromEntry(BB)) { | |||
4150 | // Don't go into unreachable blocks. They may contain instructions with | |||
4151 | // dependency cycles which confuse the final scheduling. | |||
4152 | LLVM_DEBUG(dbgs() << "SLP: bundle in unreachable block.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: bundle in unreachable block.\n" ; } } while (false); | |||
4153 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx); | |||
4154 | return; | |||
4155 | } | |||
4156 | ||||
4157 | // Check that every instruction appears once in this bundle. | |||
4158 | if (!TryToFindDuplicates(S)) | |||
4159 | return; | |||
4160 | ||||
4161 | auto &BSRef = BlocksSchedules[BB]; | |||
4162 | if (!BSRef) | |||
4163 | BSRef = std::make_unique<BlockScheduling>(BB); | |||
4164 | ||||
4165 | BlockScheduling &BS = *BSRef; | |||
4166 | ||||
4167 | Optional<ScheduleData *> Bundle = BS.tryScheduleBundle(VL, this, S); | |||
4168 | #ifdef EXPENSIVE_CHECKS | |||
4169 | // Make sure we didn't break any internal invariants | |||
4170 | BS.verify(); | |||
4171 | #endif | |||
4172 | if (!Bundle) { | |||
4173 | LLVM_DEBUG(dbgs() << "SLP: We are not able to schedule this bundle!\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: We are not able to schedule this bundle!\n" ; } } while (false); | |||
4174 | assert((!BS.getScheduleData(VL0) ||(static_cast <bool> ((!BS.getScheduleData(VL0) || !BS.getScheduleData (VL0)->isPartOfBundle()) && "tryScheduleBundle should cancelScheduling on failure" ) ? void (0) : __assert_fail ("(!BS.getScheduleData(VL0) || !BS.getScheduleData(VL0)->isPartOfBundle()) && \"tryScheduleBundle should cancelScheduling on failure\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 4176, __extension__ __PRETTY_FUNCTION__)) | |||
4175 | !BS.getScheduleData(VL0)->isPartOfBundle()) &&(static_cast <bool> ((!BS.getScheduleData(VL0) || !BS.getScheduleData (VL0)->isPartOfBundle()) && "tryScheduleBundle should cancelScheduling on failure" ) ? void (0) : __assert_fail ("(!BS.getScheduleData(VL0) || !BS.getScheduleData(VL0)->isPartOfBundle()) && \"tryScheduleBundle should cancelScheduling on failure\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 4176, __extension__ __PRETTY_FUNCTION__)) | |||
4176 | "tryScheduleBundle should cancelScheduling on failure")(static_cast <bool> ((!BS.getScheduleData(VL0) || !BS.getScheduleData (VL0)->isPartOfBundle()) && "tryScheduleBundle should cancelScheduling on failure" ) ? void (0) : __assert_fail ("(!BS.getScheduleData(VL0) || !BS.getScheduleData(VL0)->isPartOfBundle()) && \"tryScheduleBundle should cancelScheduling on failure\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 4176, __extension__ __PRETTY_FUNCTION__)); | |||
4177 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | |||
4178 | ReuseShuffleIndicies); | |||
4179 | return; | |||
4180 | } | |||
4181 | LLVM_DEBUG(dbgs() << "SLP: We are able to schedule this bundle.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: We are able to schedule this bundle.\n" ; } } while (false); | |||
4182 | ||||
4183 | unsigned ShuffleOrOp = S.isAltShuffle() ? | |||
4184 | (unsigned) Instruction::ShuffleVector : S.getOpcode(); | |||
4185 | switch (ShuffleOrOp) { | |||
4186 | case Instruction::PHI: { | |||
4187 | auto *PH = cast<PHINode>(VL0); | |||
4188 | ||||
4189 | // Check for terminator values (e.g. invoke). | |||
4190 | for (Value *V : VL) | |||
4191 | for (Value *Incoming : cast<PHINode>(V)->incoming_values()) { | |||
4192 | Instruction *Term = dyn_cast<Instruction>(Incoming); | |||
4193 | if (Term && Term->isTerminator()) { | |||
4194 | LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Need to swizzle PHINodes (terminator use).\n" ; } } while (false) | |||
4195 | << "SLP: Need to swizzle PHINodes (terminator use).\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Need to swizzle PHINodes (terminator use).\n" ; } } while (false); | |||
4196 | BS.cancelScheduling(VL, VL0); | |||
4197 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | |||
4198 | ReuseShuffleIndicies); | |||
4199 | return; | |||
4200 | } | |||
4201 | } | |||
4202 | ||||
4203 | TreeEntry *TE = | |||
4204 | newTreeEntry(VL, Bundle, S, UserTreeIdx, ReuseShuffleIndicies); | |||
4205 | LLVM_DEBUG(dbgs() << "SLP: added a vector of PHINodes.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: added a vector of PHINodes.\n" ; } } while (false); | |||
4206 | ||||
4207 | // Keeps the reordered operands to avoid code duplication. | |||
4208 | SmallVector<ValueList, 2> OperandsVec; | |||
4209 | for (unsigned I = 0, E = PH->getNumIncomingValues(); I < E; ++I) { | |||
4210 | if (!DT->isReachableFromEntry(PH->getIncomingBlock(I))) { | |||
4211 | ValueList Operands(VL.size(), PoisonValue::get(PH->getType())); | |||
4212 | TE->setOperand(I, Operands); | |||
4213 | OperandsVec.push_back(Operands); | |||
4214 | continue; | |||
4215 | } | |||
4216 | ValueList Operands; | |||
4217 | // Prepare the operand vector. | |||
4218 | for (Value *V : VL) | |||
4219 | Operands.push_back(cast<PHINode>(V)->getIncomingValueForBlock( | |||
4220 | PH->getIncomingBlock(I))); | |||
4221 | TE->setOperand(I, Operands); | |||
4222 | OperandsVec.push_back(Operands); | |||
4223 | } | |||
4224 | for (unsigned OpIdx = 0, OpE = OperandsVec.size(); OpIdx != OpE; ++OpIdx) | |||
4225 | buildTree_rec(OperandsVec[OpIdx], Depth + 1, {TE, OpIdx}); | |||
4226 | return; | |||
4227 | } | |||
4228 | case Instruction::ExtractValue: | |||
4229 | case Instruction::ExtractElement: { | |||
4230 | OrdersType CurrentOrder; | |||
4231 | bool Reuse = canReuseExtract(VL, VL0, CurrentOrder); | |||
4232 | if (Reuse) { | |||
4233 | LLVM_DEBUG(dbgs() << "SLP: Reusing or shuffling extract sequence.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Reusing or shuffling extract sequence.\n" ; } } while (false); | |||
4234 | newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, | |||
4235 | ReuseShuffleIndicies); | |||
4236 | // This is a special case, as it does not gather, but at the same time | |||
4237 | // we are not extending buildTree_rec() towards the operands. | |||
4238 | ValueList Op0; | |||
4239 | Op0.assign(VL.size(), VL0->getOperand(0)); | |||
4240 | VectorizableTree.back()->setOperand(0, Op0); | |||
4241 | return; | |||
4242 | } | |||
4243 | if (!CurrentOrder.empty()) { | |||
4244 | LLVM_DEBUG({do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { { dbgs() << "SLP: Reusing or shuffling of reordered extract sequence " "with order"; for (unsigned Idx : CurrentOrder) dbgs() << " " << Idx; dbgs() << "\n"; }; } } while (false) | |||
4245 | dbgs() << "SLP: Reusing or shuffling of reordered extract sequence "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { { dbgs() << "SLP: Reusing or shuffling of reordered extract sequence " "with order"; for (unsigned Idx : CurrentOrder) dbgs() << " " << Idx; dbgs() << "\n"; }; } } while (false) | |||
4246 | "with order";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { { dbgs() << "SLP: Reusing or shuffling of reordered extract sequence " "with order"; for (unsigned Idx : CurrentOrder) dbgs() << " " << Idx; dbgs() << "\n"; }; } } while (false) | |||
4247 | for (unsigned Idx : CurrentOrder)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { { dbgs() << "SLP: Reusing or shuffling of reordered extract sequence " "with order"; for (unsigned Idx : CurrentOrder) dbgs() << " " << Idx; dbgs() << "\n"; }; } } while (false) | |||
4248 | dbgs() << " " << Idx;do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { { dbgs() << "SLP: Reusing or shuffling of reordered extract sequence " "with order"; for (unsigned Idx : CurrentOrder) dbgs() << " " << Idx; dbgs() << "\n"; }; } } while (false) | |||
4249 | dbgs() << "\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { { dbgs() << "SLP: Reusing or shuffling of reordered extract sequence " "with order"; for (unsigned Idx : CurrentOrder) dbgs() << " " << Idx; dbgs() << "\n"; }; } } while (false) | |||
4250 | })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { { dbgs() << "SLP: Reusing or shuffling of reordered extract sequence " "with order"; for (unsigned Idx : CurrentOrder) dbgs() << " " << Idx; dbgs() << "\n"; }; } } while (false); | |||
4251 | fixupOrderingIndices(CurrentOrder); | |||
4252 | // Insert new order with initial value 0, if it does not exist, | |||
4253 | // otherwise return the iterator to the existing one. | |||
4254 | newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, | |||
4255 | ReuseShuffleIndicies, CurrentOrder); | |||
4256 | // This is a special case, as it does not gather, but at the same time | |||
4257 | // we are not extending buildTree_rec() towards the operands. | |||
4258 | ValueList Op0; | |||
4259 | Op0.assign(VL.size(), VL0->getOperand(0)); | |||
4260 | VectorizableTree.back()->setOperand(0, Op0); | |||
4261 | return; | |||
4262 | } | |||
4263 | LLVM_DEBUG(dbgs() << "SLP: Gather extract sequence.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gather extract sequence.\n"; } } while (false); | |||
4264 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | |||
4265 | ReuseShuffleIndicies); | |||
4266 | BS.cancelScheduling(VL, VL0); | |||
4267 | return; | |||
4268 | } | |||
4269 | case Instruction::InsertElement: { | |||
4270 | assert(ReuseShuffleIndicies.empty() && "All inserts should be unique")(static_cast <bool> (ReuseShuffleIndicies.empty() && "All inserts should be unique") ? void (0) : __assert_fail ( "ReuseShuffleIndicies.empty() && \"All inserts should be unique\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 4270, __extension__ __PRETTY_FUNCTION__)); | |||
4271 | ||||
4272 | // Check that we have a buildvector and not a shuffle of 2 or more | |||
4273 | // different vectors. | |||
4274 | ValueSet SourceVectors; | |||
4275 | for (Value *V : VL) { | |||
4276 | SourceVectors.insert(cast<Instruction>(V)->getOperand(0)); | |||
4277 | assert(getInsertIndex(V) != None && "Non-constant or undef index?")(static_cast <bool> (getInsertIndex(V) != None && "Non-constant or undef index?") ? void (0) : __assert_fail ( "getInsertIndex(V) != None && \"Non-constant or undef index?\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 4277, __extension__ __PRETTY_FUNCTION__)); | |||
4278 | } | |||
4279 | ||||
4280 | if (count_if(VL, [&SourceVectors](Value *V) { | |||
4281 | return !SourceVectors.contains(V); | |||
4282 | }) >= 2) { | |||
4283 | // Found 2nd source vector - cancel. | |||
4284 | LLVM_DEBUG(dbgs() << "SLP: Gather of insertelement vectors with "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gather of insertelement vectors with " "different source vectors.\n"; } } while (false) | |||
4285 | "different source vectors.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gather of insertelement vectors with " "different source vectors.\n"; } } while (false); | |||
4286 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx); | |||
4287 | BS.cancelScheduling(VL, VL0); | |||
4288 | return; | |||
4289 | } | |||
4290 | ||||
4291 | auto OrdCompare = [](const std::pair<int, int> &P1, | |||
4292 | const std::pair<int, int> &P2) { | |||
4293 | return P1.first > P2.first; | |||
4294 | }; | |||
4295 | PriorityQueue<std::pair<int, int>, SmallVector<std::pair<int, int>>, | |||
4296 | decltype(OrdCompare)> | |||
4297 | Indices(OrdCompare); | |||
4298 | for (int I = 0, E = VL.size(); I < E; ++I) { | |||
4299 | unsigned Idx = *getInsertIndex(VL[I]); | |||
4300 | Indices.emplace(Idx, I); | |||
4301 | } | |||
4302 | OrdersType CurrentOrder(VL.size(), VL.size()); | |||
4303 | bool IsIdentity = true; | |||
4304 | for (int I = 0, E = VL.size(); I < E; ++I) { | |||
4305 | CurrentOrder[Indices.top().second] = I; | |||
4306 | IsIdentity &= Indices.top().second == I; | |||
4307 | Indices.pop(); | |||
4308 | } | |||
4309 | if (IsIdentity) | |||
4310 | CurrentOrder.clear(); | |||
4311 | TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, | |||
4312 | None, CurrentOrder); | |||
4313 | LLVM_DEBUG(dbgs() << "SLP: added inserts bundle.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: added inserts bundle.\n"; } } while (false); | |||
4314 | ||||
4315 | constexpr int NumOps = 2; | |||
4316 | ValueList VectorOperands[NumOps]; | |||
4317 | for (int I = 0; I < NumOps; ++I) { | |||
4318 | for (Value *V : VL) | |||
4319 | VectorOperands[I].push_back(cast<Instruction>(V)->getOperand(I)); | |||
4320 | ||||
4321 | TE->setOperand(I, VectorOperands[I]); | |||
4322 | } | |||
4323 | buildTree_rec(VectorOperands[NumOps - 1], Depth + 1, {TE, NumOps - 1}); | |||
4324 | return; | |||
4325 | } | |||
4326 | case Instruction::Load: { | |||
4327 | // Check that a vectorized load would load the same memory as a scalar | |||
4328 | // load. For example, we don't want to vectorize loads that are smaller | |||
4329 | // than 8-bit. Even though we have a packed struct {<i2, i2, i2, i2>} LLVM | |||
4330 | // treats loading/storing it as an i8 struct. If we vectorize loads/stores | |||
4331 | // from such a struct, we read/write packed bits disagreeing with the | |||
4332 | // unvectorized version. | |||
4333 | SmallVector<Value *> PointerOps; | |||
4334 | OrdersType CurrentOrder; | |||
4335 | TreeEntry *TE = nullptr; | |||
4336 | switch (canVectorizeLoads(VL, VL0, *TTI, *DL, *SE, CurrentOrder, | |||
4337 | PointerOps)) { | |||
4338 | case LoadsState::Vectorize: | |||
4339 | if (CurrentOrder.empty()) { | |||
4340 | // Original loads are consecutive and does not require reordering. | |||
4341 | TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, | |||
4342 | ReuseShuffleIndicies); | |||
4343 | LLVM_DEBUG(dbgs() << "SLP: added a vector of loads.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: added a vector of loads.\n"; } } while (false); | |||
4344 | } else { | |||
4345 | fixupOrderingIndices(CurrentOrder); | |||
4346 | // Need to reorder. | |||
4347 | TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, | |||
4348 | ReuseShuffleIndicies, CurrentOrder); | |||
4349 | LLVM_DEBUG(dbgs() << "SLP: added a vector of jumbled loads.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: added a vector of jumbled loads.\n" ; } } while (false); | |||
4350 | } | |||
4351 | TE->setOperandsInOrder(); | |||
4352 | break; | |||
4353 | case LoadsState::ScatterVectorize: | |||
4354 | // Vectorizing non-consecutive loads with `llvm.masked.gather`. | |||
4355 | TE = newTreeEntry(VL, TreeEntry::ScatterVectorize, Bundle, S, | |||
4356 | UserTreeIdx, ReuseShuffleIndicies); | |||
4357 | TE->setOperandsInOrder(); | |||
4358 | buildTree_rec(PointerOps, Depth + 1, {TE, 0}); | |||
4359 | LLVM_DEBUG(dbgs() << "SLP: added a vector of non-consecutive loads.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: added a vector of non-consecutive loads.\n" ; } } while (false); | |||
4360 | break; | |||
4361 | case LoadsState::Gather: | |||
4362 | BS.cancelScheduling(VL, VL0); | |||
4363 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | |||
4364 | ReuseShuffleIndicies); | |||
4365 | #ifndef NDEBUG | |||
4366 | Type *ScalarTy = VL0->getType(); | |||
4367 | if (DL->getTypeSizeInBits(ScalarTy) != | |||
4368 | DL->getTypeAllocSizeInBits(ScalarTy)) | |||
4369 | LLVM_DEBUG(dbgs() << "SLP: Gathering loads of non-packed type.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering loads of non-packed type.\n" ; } } while (false); | |||
4370 | else if (any_of(VL, [](Value *V) { | |||
4371 | return !cast<LoadInst>(V)->isSimple(); | |||
4372 | })) | |||
4373 | LLVM_DEBUG(dbgs() << "SLP: Gathering non-simple loads.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering non-simple loads.\n" ; } } while (false); | |||
4374 | else | |||
4375 | LLVM_DEBUG(dbgs() << "SLP: Gathering non-consecutive loads.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering non-consecutive loads.\n" ; } } while (false); | |||
4376 | #endif // NDEBUG | |||
4377 | break; | |||
4378 | } | |||
4379 | return; | |||
4380 | } | |||
4381 | case Instruction::ZExt: | |||
4382 | case Instruction::SExt: | |||
4383 | case Instruction::FPToUI: | |||
4384 | case Instruction::FPToSI: | |||
4385 | case Instruction::FPExt: | |||
4386 | case Instruction::PtrToInt: | |||
4387 | case Instruction::IntToPtr: | |||
4388 | case Instruction::SIToFP: | |||
4389 | case Instruction::UIToFP: | |||
4390 | case Instruction::Trunc: | |||
4391 | case Instruction::FPTrunc: | |||
4392 | case Instruction::BitCast: { | |||
4393 | Type *SrcTy = VL0->getOperand(0)->getType(); | |||
4394 | for (Value *V : VL) { | |||
4395 | Type *Ty = cast<Instruction>(V)->getOperand(0)->getType(); | |||
4396 | if (Ty != SrcTy || !isValidElementType(Ty)) { | |||
4397 | BS.cancelScheduling(VL, VL0); | |||
4398 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | |||
4399 | ReuseShuffleIndicies); | |||
4400 | LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering casts with different src types.\n" ; } } while (false) | |||
4401 | << "SLP: Gathering casts with different src types.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering casts with different src types.\n" ; } } while (false); | |||
4402 | return; | |||
4403 | } | |||
4404 | } | |||
4405 | TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, | |||
4406 | ReuseShuffleIndicies); | |||
4407 | LLVM_DEBUG(dbgs() << "SLP: added a vector of casts.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: added a vector of casts.\n"; } } while (false); | |||
4408 | ||||
4409 | TE->setOperandsInOrder(); | |||
4410 | for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) { | |||
4411 | ValueList Operands; | |||
4412 | // Prepare the operand vector. | |||
4413 | for (Value *V : VL) | |||
4414 | Operands.push_back(cast<Instruction>(V)->getOperand(i)); | |||
4415 | ||||
4416 | buildTree_rec(Operands, Depth + 1, {TE, i}); | |||
4417 | } | |||
4418 | return; | |||
4419 | } | |||
4420 | case Instruction::ICmp: | |||
4421 | case Instruction::FCmp: { | |||
4422 | // Check that all of the compares have the same predicate. | |||
4423 | CmpInst::Predicate P0 = cast<CmpInst>(VL0)->getPredicate(); | |||
4424 | CmpInst::Predicate SwapP0 = CmpInst::getSwappedPredicate(P0); | |||
4425 | Type *ComparedTy = VL0->getOperand(0)->getType(); | |||
4426 | for (Value *V : VL) { | |||
4427 | CmpInst *Cmp = cast<CmpInst>(V); | |||
4428 | if ((Cmp->getPredicate() != P0 && Cmp->getPredicate() != SwapP0) || | |||
4429 | Cmp->getOperand(0)->getType() != ComparedTy) { | |||
4430 | BS.cancelScheduling(VL, VL0); | |||
4431 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | |||
4432 | ReuseShuffleIndicies); | |||
4433 | LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering cmp with different predicate.\n" ; } } while (false) | |||
4434 | << "SLP: Gathering cmp with different predicate.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering cmp with different predicate.\n" ; } } while (false); | |||
4435 | return; | |||
4436 | } | |||
4437 | } | |||
4438 | ||||
4439 | TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, | |||
4440 | ReuseShuffleIndicies); | |||
4441 | LLVM_DEBUG(dbgs() << "SLP: added a vector of compares.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: added a vector of compares.\n" ; } } while (false); | |||
4442 | ||||
4443 | ValueList Left, Right; | |||
4444 | if (cast<CmpInst>(VL0)->isCommutative()) { | |||
4445 | // Commutative predicate - collect + sort operands of the instructions | |||
4446 | // so that each side is more likely to have the same opcode. | |||
4447 | assert(P0 == SwapP0 && "Commutative Predicate mismatch")(static_cast <bool> (P0 == SwapP0 && "Commutative Predicate mismatch" ) ? void (0) : __assert_fail ("P0 == SwapP0 && \"Commutative Predicate mismatch\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 4447, __extension__ __PRETTY_FUNCTION__)); | |||
4448 | reorderInputsAccordingToOpcode(VL, Left, Right, *DL, *SE, *this); | |||
4449 | } else { | |||
4450 | // Collect operands - commute if it uses the swapped predicate. | |||
4451 | for (Value *V : VL) { | |||
4452 | auto *Cmp = cast<CmpInst>(V); | |||
4453 | Value *LHS = Cmp->getOperand(0); | |||
4454 | Value *RHS = Cmp->getOperand(1); | |||
4455 | if (Cmp->getPredicate() != P0) | |||
4456 | std::swap(LHS, RHS); | |||
4457 | Left.push_back(LHS); | |||
4458 | Right.push_back(RHS); | |||
4459 | } | |||
4460 | } | |||
4461 | TE->setOperand(0, Left); | |||
4462 | TE->setOperand(1, Right); | |||
4463 | buildTree_rec(Left, Depth + 1, {TE, 0}); | |||
4464 | buildTree_rec(Right, Depth + 1, {TE, 1}); | |||
4465 | return; | |||
4466 | } | |||
4467 | case Instruction::Select: | |||
4468 | case Instruction::FNeg: | |||
4469 | case Instruction::Add: | |||
4470 | case Instruction::FAdd: | |||
4471 | case Instruction::Sub: | |||
4472 | case Instruction::FSub: | |||
4473 | case Instruction::Mul: | |||
4474 | case Instruction::FMul: | |||
4475 | case Instruction::UDiv: | |||
4476 | case Instruction::SDiv: | |||
4477 | case Instruction::FDiv: | |||
4478 | case Instruction::URem: | |||
4479 | case Instruction::SRem: | |||
4480 | case Instruction::FRem: | |||
4481 | case Instruction::Shl: | |||
4482 | case Instruction::LShr: | |||
4483 | case Instruction::AShr: | |||
4484 | case Instruction::And: | |||
4485 | case Instruction::Or: | |||
4486 | case Instruction::Xor: { | |||
4487 | TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, | |||
4488 | ReuseShuffleIndicies); | |||
4489 | LLVM_DEBUG(dbgs() << "SLP: added a vector of un/bin op.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: added a vector of un/bin op.\n" ; } } while (false); | |||
4490 | ||||
4491 | // Sort operands of the instructions so that each side is more likely to | |||
4492 | // have the same opcode. | |||
4493 | if (isa<BinaryOperator>(VL0) && VL0->isCommutative()) { | |||
4494 | ValueList Left, Right; | |||
4495 | reorderInputsAccordingToOpcode(VL, Left, Right, *DL, *SE, *this); | |||
4496 | TE->setOperand(0, Left); | |||
4497 | TE->setOperand(1, Right); | |||
4498 | buildTree_rec(Left, Depth + 1, {TE, 0}); | |||
4499 | buildTree_rec(Right, Depth + 1, {TE, 1}); | |||
4500 | return; | |||
4501 | } | |||
4502 | ||||
4503 | TE->setOperandsInOrder(); | |||
4504 | for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) { | |||
4505 | ValueList Operands; | |||
4506 | // Prepare the operand vector. | |||
4507 | for (Value *V : VL) | |||
4508 | Operands.push_back(cast<Instruction>(V)->getOperand(i)); | |||
4509 | ||||
4510 | buildTree_rec(Operands, Depth + 1, {TE, i}); | |||
4511 | } | |||
4512 | return; | |||
4513 | } | |||
4514 | case Instruction::GetElementPtr: { | |||
4515 | // We don't combine GEPs with complicated (nested) indexing. | |||
4516 | for (Value *V : VL) { | |||
4517 | if (cast<Instruction>(V)->getNumOperands() != 2) { | |||
4518 | LLVM_DEBUG(dbgs() << "SLP: not-vectorizable GEP (nested indexes).\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: not-vectorizable GEP (nested indexes).\n" ; } } while (false); | |||
4519 | BS.cancelScheduling(VL, VL0); | |||
4520 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | |||
4521 | ReuseShuffleIndicies); | |||
4522 | return; | |||
4523 | } | |||
4524 | } | |||
4525 | ||||
4526 | // We can't combine several GEPs into one vector if they operate on | |||
4527 | // different types. | |||
4528 | Type *Ty0 = cast<GEPOperator>(VL0)->getSourceElementType(); | |||
4529 | for (Value *V : VL) { | |||
4530 | Type *CurTy = cast<GEPOperator>(V)->getSourceElementType(); | |||
4531 | if (Ty0 != CurTy) { | |||
4532 | LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: not-vectorizable GEP (different types).\n" ; } } while (false) | |||
4533 | << "SLP: not-vectorizable GEP (different types).\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: not-vectorizable GEP (different types).\n" ; } } while (false); | |||
4534 | BS.cancelScheduling(VL, VL0); | |||
4535 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | |||
4536 | ReuseShuffleIndicies); | |||
4537 | return; | |||
4538 | } | |||
4539 | } | |||
4540 | ||||
4541 | // We don't combine GEPs with non-constant indexes. | |||
4542 | Type *Ty1 = VL0->getOperand(1)->getType(); | |||
4543 | for (Value *V : VL) { | |||
4544 | auto Op = cast<Instruction>(V)->getOperand(1); | |||
4545 | if (!isa<ConstantInt>(Op) || | |||
4546 | (Op->getType() != Ty1 && | |||
4547 | Op->getType()->getScalarSizeInBits() > | |||
4548 | DL->getIndexSizeInBits( | |||
4549 | V->getType()->getPointerAddressSpace()))) { | |||
4550 | LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: not-vectorizable GEP (non-constant indexes).\n" ; } } while (false) | |||
4551 | << "SLP: not-vectorizable GEP (non-constant indexes).\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: not-vectorizable GEP (non-constant indexes).\n" ; } } while (false); | |||
4552 | BS.cancelScheduling(VL, VL0); | |||
4553 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | |||
4554 | ReuseShuffleIndicies); | |||
4555 | return; | |||
4556 | } | |||
4557 | } | |||
4558 | ||||
4559 | TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, | |||
4560 | ReuseShuffleIndicies); | |||
4561 | LLVM_DEBUG(dbgs() << "SLP: added a vector of GEPs.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: added a vector of GEPs.\n"; } } while (false); | |||
4562 | SmallVector<ValueList, 2> Operands(2); | |||
4563 | // Prepare the operand vector for pointer operands. | |||
4564 | for (Value *V : VL) | |||
4565 | Operands.front().push_back( | |||
4566 | cast<GetElementPtrInst>(V)->getPointerOperand()); | |||
4567 | TE->setOperand(0, Operands.front()); | |||
4568 | // Need to cast all indices to the same type before vectorization to | |||
4569 | // avoid crash. | |||
4570 | // Required to be able to find correct matches between different gather | |||
4571 | // nodes and reuse the vectorized values rather than trying to gather them | |||
4572 | // again. | |||
4573 | int IndexIdx = 1; | |||
4574 | Type *VL0Ty = VL0->getOperand(IndexIdx)->getType(); | |||
4575 | Type *Ty = all_of(VL, | |||
4576 | [VL0Ty, IndexIdx](Value *V) { | |||
4577 | return VL0Ty == cast<GetElementPtrInst>(V) | |||
4578 | ->getOperand(IndexIdx) | |||
4579 | ->getType(); | |||
4580 | }) | |||
4581 | ? VL0Ty | |||
4582 | : DL->getIndexType(cast<GetElementPtrInst>(VL0) | |||
4583 | ->getPointerOperandType() | |||
4584 | ->getScalarType()); | |||
4585 | // Prepare the operand vector. | |||
4586 | for (Value *V : VL) { | |||
4587 | auto *Op = cast<Instruction>(V)->getOperand(IndexIdx); | |||
4588 | auto *CI = cast<ConstantInt>(Op); | |||
4589 | Operands.back().push_back(ConstantExpr::getIntegerCast( | |||
4590 | CI, Ty, CI->getValue().isSignBitSet())); | |||
4591 | } | |||
4592 | TE->setOperand(IndexIdx, Operands.back()); | |||
4593 | ||||
4594 | for (unsigned I = 0, Ops = Operands.size(); I < Ops; ++I) | |||
4595 | buildTree_rec(Operands[I], Depth + 1, {TE, I}); | |||
4596 | return; | |||
4597 | } | |||
4598 | case Instruction::Store: { | |||
4599 | // Check if the stores are consecutive or if we need to swizzle them. | |||
4600 | llvm::Type *ScalarTy = cast<StoreInst>(VL0)->getValueOperand()->getType(); | |||
4601 | // Avoid types that are padded when being allocated as scalars, while | |||
4602 | // being packed together in a vector (such as i1). | |||
4603 | if (DL->getTypeSizeInBits(ScalarTy) != | |||
4604 | DL->getTypeAllocSizeInBits(ScalarTy)) { | |||
4605 | BS.cancelScheduling(VL, VL0); | |||
4606 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | |||
4607 | ReuseShuffleIndicies); | |||
4608 | LLVM_DEBUG(dbgs() << "SLP: Gathering stores of non-packed type.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering stores of non-packed type.\n" ; } } while (false); | |||
4609 | return; | |||
4610 | } | |||
4611 | // Make sure all stores in the bundle are simple - we can't vectorize | |||
4612 | // atomic or volatile stores. | |||
4613 | SmallVector<Value *, 4> PointerOps(VL.size()); | |||
4614 | ValueList Operands(VL.size()); | |||
4615 | auto POIter = PointerOps.begin(); | |||
4616 | auto OIter = Operands.begin(); | |||
4617 | for (Value *V : VL) { | |||
4618 | auto *SI = cast<StoreInst>(V); | |||
4619 | if (!SI->isSimple()) { | |||
4620 | BS.cancelScheduling(VL, VL0); | |||
4621 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | |||
4622 | ReuseShuffleIndicies); | |||
4623 | LLVM_DEBUG(dbgs() << "SLP: Gathering non-simple stores.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering non-simple stores.\n" ; } } while (false); | |||
4624 | return; | |||
4625 | } | |||
4626 | *POIter = SI->getPointerOperand(); | |||
4627 | *OIter = SI->getValueOperand(); | |||
4628 | ++POIter; | |||
4629 | ++OIter; | |||
4630 | } | |||
4631 | ||||
4632 | OrdersType CurrentOrder; | |||
4633 | // Check the order of pointer operands. | |||
4634 | if (llvm::sortPtrAccesses(PointerOps, ScalarTy, *DL, *SE, CurrentOrder)) { | |||
4635 | Value *Ptr0; | |||
4636 | Value *PtrN; | |||
4637 | if (CurrentOrder.empty()) { | |||
4638 | Ptr0 = PointerOps.front(); | |||
4639 | PtrN = PointerOps.back(); | |||
4640 | } else { | |||
4641 | Ptr0 = PointerOps[CurrentOrder.front()]; | |||
4642 | PtrN = PointerOps[CurrentOrder.back()]; | |||
4643 | } | |||
4644 | Optional<int> Dist = | |||
4645 | getPointersDiff(ScalarTy, Ptr0, ScalarTy, PtrN, *DL, *SE); | |||
4646 | // Check that the sorted pointer operands are consecutive. | |||
4647 | if (static_cast<unsigned>(*Dist) == VL.size() - 1) { | |||
4648 | if (CurrentOrder.empty()) { | |||
4649 | // Original stores are consecutive and does not require reordering. | |||
4650 | TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, | |||
4651 | UserTreeIdx, ReuseShuffleIndicies); | |||
4652 | TE->setOperandsInOrder(); | |||
4653 | buildTree_rec(Operands, Depth + 1, {TE, 0}); | |||
4654 | LLVM_DEBUG(dbgs() << "SLP: added a vector of stores.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: added a vector of stores.\n" ; } } while (false); | |||
4655 | } else { | |||
4656 | fixupOrderingIndices(CurrentOrder); | |||
4657 | TreeEntry *TE = | |||
4658 | newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, | |||
4659 | ReuseShuffleIndicies, CurrentOrder); | |||
4660 | TE->setOperandsInOrder(); | |||
4661 | buildTree_rec(Operands, Depth + 1, {TE, 0}); | |||
4662 | LLVM_DEBUG(dbgs() << "SLP: added a vector of jumbled stores.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: added a vector of jumbled stores.\n" ; } } while (false); | |||
4663 | } | |||
4664 | return; | |||
4665 | } | |||
4666 | } | |||
4667 | ||||
4668 | BS.cancelScheduling(VL, VL0); | |||
4669 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | |||
4670 | ReuseShuffleIndicies); | |||
4671 | LLVM_DEBUG(dbgs() << "SLP: Non-consecutive store.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Non-consecutive store.\n"; } } while (false); | |||
4672 | return; | |||
4673 | } | |||
4674 | case Instruction::Call: { | |||
4675 | // Check if the calls are all to the same vectorizable intrinsic or | |||
4676 | // library function. | |||
4677 | CallInst *CI = cast<CallInst>(VL0); | |||
4678 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); | |||
4679 | ||||
4680 | VFShape Shape = VFShape::get( | |||
4681 | *CI, ElementCount::getFixed(static_cast<unsigned int>(VL.size())), | |||
4682 | false /*HasGlobalPred*/); | |||
4683 | Function *VecFunc = VFDatabase(*CI).getVectorizedFunction(Shape); | |||
4684 | ||||
4685 | if (!VecFunc && !isTriviallyVectorizable(ID)) { | |||
4686 | BS.cancelScheduling(VL, VL0); | |||
4687 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | |||
4688 | ReuseShuffleIndicies); | |||
4689 | LLVM_DEBUG(dbgs() << "SLP: Non-vectorizable call.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Non-vectorizable call.\n"; } } while (false); | |||
4690 | return; | |||
4691 | } | |||
4692 | Function *F = CI->getCalledFunction(); | |||
4693 | unsigned NumArgs = CI->arg_size(); | |||
4694 | SmallVector<Value*, 4> ScalarArgs(NumArgs, nullptr); | |||
4695 | for (unsigned j = 0; j != NumArgs; ++j) | |||
4696 | if (hasVectorInstrinsicScalarOpd(ID, j)) | |||
4697 | ScalarArgs[j] = CI->getArgOperand(j); | |||
4698 | for (Value *V : VL) { | |||
4699 | CallInst *CI2 = dyn_cast<CallInst>(V); | |||
4700 | if (!CI2 || CI2->getCalledFunction() != F || | |||
4701 | getVectorIntrinsicIDForCall(CI2, TLI) != ID || | |||
4702 | (VecFunc && | |||
4703 | VecFunc != VFDatabase(*CI2).getVectorizedFunction(Shape)) || | |||
4704 | !CI->hasIdenticalOperandBundleSchema(*CI2)) { | |||
4705 | BS.cancelScheduling(VL, VL0); | |||
4706 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | |||
4707 | ReuseShuffleIndicies); | |||
4708 | LLVM_DEBUG(dbgs() << "SLP: mismatched calls:" << *CI << "!=" << *Vdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: mismatched calls:" << * CI << "!=" << *V << "\n"; } } while (false) | |||
4709 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: mismatched calls:" << * CI << "!=" << *V << "\n"; } } while (false); | |||
4710 | return; | |||
4711 | } | |||
4712 | // Some intrinsics have scalar arguments and should be same in order for | |||
4713 | // them to be vectorized. | |||
4714 | for (unsigned j = 0; j != NumArgs; ++j) { | |||
4715 | if (hasVectorInstrinsicScalarOpd(ID, j)) { | |||
4716 | Value *A1J = CI2->getArgOperand(j); | |||
4717 | if (ScalarArgs[j] != A1J) { | |||
4718 | BS.cancelScheduling(VL, VL0); | |||
4719 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | |||
4720 | ReuseShuffleIndicies); | |||
4721 | LLVM_DEBUG(dbgs() << "SLP: mismatched arguments in call:" << *CIdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: mismatched arguments in call:" << *CI << " argument " << ScalarArgs[j] << "!=" << A1J << "\n"; } } while (false) | |||
4722 | << " argument " << ScalarArgs[j] << "!=" << A1Jdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: mismatched arguments in call:" << *CI << " argument " << ScalarArgs[j] << "!=" << A1J << "\n"; } } while (false) | |||
4723 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: mismatched arguments in call:" << *CI << " argument " << ScalarArgs[j] << "!=" << A1J << "\n"; } } while (false); | |||
4724 | return; | |||
4725 | } | |||
4726 | } | |||
4727 | } | |||
4728 | // Verify that the bundle operands are identical between the two calls. | |||
4729 | if (CI->hasOperandBundles() && | |||
4730 | !std::equal(CI->op_begin() + CI->getBundleOperandsStartIndex(), | |||
4731 | CI->op_begin() + CI->getBundleOperandsEndIndex(), | |||
4732 | CI2->op_begin() + CI2->getBundleOperandsStartIndex())) { | |||
4733 | BS.cancelScheduling(VL, VL0); | |||
4734 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | |||
4735 | ReuseShuffleIndicies); | |||
4736 | LLVM_DEBUG(dbgs() << "SLP: mismatched bundle operands in calls:"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: mismatched bundle operands in calls:" << *CI << "!=" << *V << '\n'; } } while (false) | |||
4737 | << *CI << "!=" << *V << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: mismatched bundle operands in calls:" << *CI << "!=" << *V << '\n'; } } while (false); | |||
4738 | return; | |||
4739 | } | |||
4740 | } | |||
4741 | ||||
4742 | TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, | |||
4743 | ReuseShuffleIndicies); | |||
4744 | TE->setOperandsInOrder(); | |||
4745 | for (unsigned i = 0, e = CI->arg_size(); i != e; ++i) { | |||
4746 | // For scalar operands no need to to create an entry since no need to | |||
4747 | // vectorize it. | |||
4748 | if (hasVectorInstrinsicScalarOpd(ID, i)) | |||
4749 | continue; | |||
4750 | ValueList Operands; | |||
4751 | // Prepare the operand vector. | |||
4752 | for (Value *V : VL) { | |||
4753 | auto *CI2 = cast<CallInst>(V); | |||
4754 | Operands.push_back(CI2->getArgOperand(i)); | |||
4755 | } | |||
4756 | buildTree_rec(Operands, Depth + 1, {TE, i}); | |||
4757 | } | |||
4758 | return; | |||
4759 | } | |||
4760 | case Instruction::ShuffleVector: { | |||
4761 | // If this is not an alternate sequence of opcode like add-sub | |||
4762 | // then do not vectorize this instruction. | |||
4763 | if (!S.isAltShuffle()) { | |||
4764 | BS.cancelScheduling(VL, VL0); | |||
4765 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | |||
4766 | ReuseShuffleIndicies); | |||
4767 | LLVM_DEBUG(dbgs() << "SLP: ShuffleVector are not vectorized.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: ShuffleVector are not vectorized.\n" ; } } while (false); | |||
4768 | return; | |||
4769 | } | |||
4770 | TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, | |||
4771 | ReuseShuffleIndicies); | |||
4772 | LLVM_DEBUG(dbgs() << "SLP: added a ShuffleVector op.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: added a ShuffleVector op.\n" ; } } while (false); | |||
4773 | ||||
4774 | // Reorder operands if reordering would enable vectorization. | |||
4775 | auto *CI = dyn_cast<CmpInst>(VL0); | |||
4776 | if (isa<BinaryOperator>(VL0) || CI) { | |||
4777 | ValueList Left, Right; | |||
4778 | if (!CI || all_of(VL, [](Value *V) { | |||
4779 | return cast<CmpInst>(V)->isCommutative(); | |||
4780 | })) { | |||
4781 | reorderInputsAccordingToOpcode(VL, Left, Right, *DL, *SE, *this); | |||
4782 | } else { | |||
4783 | CmpInst::Predicate P0 = CI->getPredicate(); | |||
4784 | CmpInst::Predicate AltP0 = cast<CmpInst>(S.AltOp)->getPredicate(); | |||
4785 | assert(P0 != AltP0 &&(static_cast <bool> (P0 != AltP0 && "Expected different main/alternate predicates." ) ? void (0) : __assert_fail ("P0 != AltP0 && \"Expected different main/alternate predicates.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 4786, __extension__ __PRETTY_FUNCTION__)) | |||
4786 | "Expected different main/alternate predicates.")(static_cast <bool> (P0 != AltP0 && "Expected different main/alternate predicates." ) ? void (0) : __assert_fail ("P0 != AltP0 && \"Expected different main/alternate predicates.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 4786, __extension__ __PRETTY_FUNCTION__)); | |||
4787 | CmpInst::Predicate AltP0Swapped = CmpInst::getSwappedPredicate(AltP0); | |||
4788 | Value *BaseOp0 = VL0->getOperand(0); | |||
4789 | Value *BaseOp1 = VL0->getOperand(1); | |||
4790 | // Collect operands - commute if it uses the swapped predicate or | |||
4791 | // alternate operation. | |||
4792 | for (Value *V : VL) { | |||
4793 | auto *Cmp = cast<CmpInst>(V); | |||
4794 | Value *LHS = Cmp->getOperand(0); | |||
4795 | Value *RHS = Cmp->getOperand(1); | |||
4796 | CmpInst::Predicate CurrentPred = Cmp->getPredicate(); | |||
4797 | if (P0 == AltP0Swapped) { | |||
4798 | if (CI != Cmp && S.AltOp != Cmp && | |||
4799 | ((P0 == CurrentPred && | |||
4800 | !areCompatibleCmpOps(BaseOp0, BaseOp1, LHS, RHS)) || | |||
4801 | (AltP0 == CurrentPred && | |||
4802 | areCompatibleCmpOps(BaseOp0, BaseOp1, LHS, RHS)))) | |||
4803 | std::swap(LHS, RHS); | |||
4804 | } else if (P0 != CurrentPred && AltP0 != CurrentPred) { | |||
4805 | std::swap(LHS, RHS); | |||
4806 | } | |||
4807 | Left.push_back(LHS); | |||
4808 | Right.push_back(RHS); | |||
4809 | } | |||
4810 | } | |||
4811 | TE->setOperand(0, Left); | |||
4812 | TE->setOperand(1, Right); | |||
4813 | buildTree_rec(Left, Depth + 1, {TE, 0}); | |||
4814 | buildTree_rec(Right, Depth + 1, {TE, 1}); | |||
4815 | return; | |||
4816 | } | |||
4817 | ||||
4818 | TE->setOperandsInOrder(); | |||
4819 | for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) { | |||
4820 | ValueList Operands; | |||
4821 | // Prepare the operand vector. | |||
4822 | for (Value *V : VL) | |||
4823 | Operands.push_back(cast<Instruction>(V)->getOperand(i)); | |||
4824 | ||||
4825 | buildTree_rec(Operands, Depth + 1, {TE, i}); | |||
4826 | } | |||
4827 | return; | |||
4828 | } | |||
4829 | default: | |||
4830 | BS.cancelScheduling(VL, VL0); | |||
4831 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | |||
4832 | ReuseShuffleIndicies); | |||
4833 | LLVM_DEBUG(dbgs() << "SLP: Gathering unknown instruction.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering unknown instruction.\n" ; } } while (false); | |||
4834 | return; | |||
4835 | } | |||
4836 | } | |||
4837 | ||||
4838 | unsigned BoUpSLP::canMapToVector(Type *T, const DataLayout &DL) const { | |||
4839 | unsigned N = 1; | |||
4840 | Type *EltTy = T; | |||
4841 | ||||
4842 | while (isa<StructType>(EltTy) || isa<ArrayType>(EltTy) || | |||
4843 | isa<VectorType>(EltTy)) { | |||
4844 | if (auto *ST = dyn_cast<StructType>(EltTy)) { | |||
4845 | // Check that struct is homogeneous. | |||
4846 | for (const auto *Ty : ST->elements()) | |||
4847 | if (Ty != *ST->element_begin()) | |||
4848 | return 0; | |||
4849 | N *= ST->getNumElements(); | |||
4850 | EltTy = *ST->element_begin(); | |||
4851 | } else if (auto *AT = dyn_cast<ArrayType>(EltTy)) { | |||
4852 | N *= AT->getNumElements(); | |||
4853 | EltTy = AT->getElementType(); | |||
4854 | } else { | |||
4855 | auto *VT = cast<FixedVectorType>(EltTy); | |||
4856 | N *= VT->getNumElements(); | |||
4857 | EltTy = VT->getElementType(); | |||
4858 | } | |||
4859 | } | |||
4860 | ||||
4861 | if (!isValidElementType(EltTy)) | |||
4862 | return 0; | |||
4863 | uint64_t VTSize = DL.getTypeStoreSizeInBits(FixedVectorType::get(EltTy, N)); | |||
4864 | if (VTSize < MinVecRegSize || VTSize > MaxVecRegSize || VTSize != DL.getTypeStoreSizeInBits(T)) | |||
4865 | return 0; | |||
4866 | return N; | |||
4867 | } | |||
4868 | ||||
4869 | bool BoUpSLP::canReuseExtract(ArrayRef<Value *> VL, Value *OpValue, | |||
4870 | SmallVectorImpl<unsigned> &CurrentOrder) const { | |||
4871 | const auto *It = find_if(VL, [](Value *V) { | |||
4872 | return isa<ExtractElementInst, ExtractValueInst>(V); | |||
4873 | }); | |||
4874 | assert(It != VL.end() && "Expected at least one extract instruction.")(static_cast <bool> (It != VL.end() && "Expected at least one extract instruction." ) ? void (0) : __assert_fail ("It != VL.end() && \"Expected at least one extract instruction.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 4874, __extension__ __PRETTY_FUNCTION__)); | |||
4875 | auto *E0 = cast<Instruction>(*It); | |||
4876 | assert(all_of(VL,(static_cast <bool> (all_of(VL, [](Value *V) { return isa <UndefValue, ExtractElementInst, ExtractValueInst>( V); }) && "Invalid opcode") ? void (0) : __assert_fail ( "all_of(VL, [](Value *V) { return isa<UndefValue, ExtractElementInst, ExtractValueInst>( V); }) && \"Invalid opcode\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 4881, __extension__ __PRETTY_FUNCTION__)) | |||
4877 | [](Value *V) {(static_cast <bool> (all_of(VL, [](Value *V) { return isa <UndefValue, ExtractElementInst, ExtractValueInst>( V); }) && "Invalid opcode") ? void (0) : __assert_fail ( "all_of(VL, [](Value *V) { return isa<UndefValue, ExtractElementInst, ExtractValueInst>( V); }) && \"Invalid opcode\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 4881, __extension__ __PRETTY_FUNCTION__)) | |||
4878 | return isa<UndefValue, ExtractElementInst, ExtractValueInst>((static_cast <bool> (all_of(VL, [](Value *V) { return isa <UndefValue, ExtractElementInst, ExtractValueInst>( V); }) && "Invalid opcode") ? void (0) : __assert_fail ( "all_of(VL, [](Value *V) { return isa<UndefValue, ExtractElementInst, ExtractValueInst>( V); }) && \"Invalid opcode\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 4881, __extension__ __PRETTY_FUNCTION__)) | |||
4879 | V);(static_cast <bool> (all_of(VL, [](Value *V) { return isa <UndefValue, ExtractElementInst, ExtractValueInst>( V); }) && "Invalid opcode") ? void (0) : __assert_fail ( "all_of(VL, [](Value *V) { return isa<UndefValue, ExtractElementInst, ExtractValueInst>( V); }) && \"Invalid opcode\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 4881, __extension__ __PRETTY_FUNCTION__)) | |||
4880 | }) &&(static_cast <bool> (all_of(VL, [](Value *V) { return isa <UndefValue, ExtractElementInst, ExtractValueInst>( V); }) && "Invalid opcode") ? void (0) : __assert_fail ( "all_of(VL, [](Value *V) { return isa<UndefValue, ExtractElementInst, ExtractValueInst>( V); }) && \"Invalid opcode\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 4881, __extension__ __PRETTY_FUNCTION__)) | |||
4881 | "Invalid opcode")(static_cast <bool> (all_of(VL, [](Value *V) { return isa <UndefValue, ExtractElementInst, ExtractValueInst>( V); }) && "Invalid opcode") ? void (0) : __assert_fail ( "all_of(VL, [](Value *V) { return isa<UndefValue, ExtractElementInst, ExtractValueInst>( V); }) && \"Invalid opcode\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 4881, __extension__ __PRETTY_FUNCTION__)); | |||
4882 | // Check if all of the extracts come from the same vector and from the | |||
4883 | // correct offset. | |||
4884 | Value *Vec = E0->getOperand(0); | |||
4885 | ||||
4886 | CurrentOrder.clear(); | |||
4887 | ||||
4888 | // We have to extract from a vector/aggregate with the same number of elements. | |||
4889 | unsigned NElts; | |||
4890 | if (E0->getOpcode() == Instruction::ExtractValue) { | |||
4891 | const DataLayout &DL = E0->getModule()->getDataLayout(); | |||
4892 | NElts = canMapToVector(Vec->getType(), DL); | |||
4893 | if (!NElts) | |||
4894 | return false; | |||
4895 | // Check if load can be rewritten as load of vector. | |||
4896 | LoadInst *LI = dyn_cast<LoadInst>(Vec); | |||
4897 | if (!LI || !LI->isSimple() || !LI->hasNUses(VL.size())) | |||
4898 | return false; | |||
4899 | } else { | |||
4900 | NElts = cast<FixedVectorType>(Vec->getType())->getNumElements(); | |||
4901 | } | |||
4902 | ||||
4903 | if (NElts != VL.size()) | |||
4904 | return false; | |||
4905 | ||||
4906 | // Check that all of the indices extract from the correct offset. | |||
4907 | bool ShouldKeepOrder = true; | |||
4908 | unsigned E = VL.size(); | |||
4909 | // Assign to all items the initial value E + 1 so we can check if the extract | |||
4910 | // instruction index was used already. | |||
4911 | // Also, later we can check that all the indices are used and we have a | |||
4912 | // consecutive access in the extract instructions, by checking that no | |||
4913 | // element of CurrentOrder still has value E + 1. | |||
4914 | CurrentOrder.assign(E, E); | |||
4915 | unsigned I = 0; | |||
4916 | for (; I < E; ++I) { | |||
4917 | auto *Inst = dyn_cast<Instruction>(VL[I]); | |||
4918 | if (!Inst) | |||
4919 | continue; | |||
4920 | if (Inst->getOperand(0) != Vec) | |||
4921 | break; | |||
4922 | if (auto *EE = dyn_cast<ExtractElementInst>(Inst)) | |||
4923 | if (isa<UndefValue>(EE->getIndexOperand())) | |||
4924 | continue; | |||
4925 | Optional<unsigned> Idx = getExtractIndex(Inst); | |||
4926 | if (!Idx) | |||
4927 | break; | |||
4928 | const unsigned ExtIdx = *Idx; | |||
4929 | if (ExtIdx != I) { | |||
4930 | if (ExtIdx >= E || CurrentOrder[ExtIdx] != E) | |||
4931 | break; | |||
4932 | ShouldKeepOrder = false; | |||
4933 | CurrentOrder[ExtIdx] = I; | |||
4934 | } else { | |||
4935 | if (CurrentOrder[I] != E) | |||
4936 | break; | |||
4937 | CurrentOrder[I] = I; | |||
4938 | } | |||
4939 | } | |||
4940 | if (I < E) { | |||
4941 | CurrentOrder.clear(); | |||
4942 | return false; | |||
4943 | } | |||
4944 | if (ShouldKeepOrder) | |||
4945 | CurrentOrder.clear(); | |||
4946 | ||||
4947 | return ShouldKeepOrder; | |||
4948 | } | |||
4949 | ||||
4950 | bool BoUpSLP::areAllUsersVectorized(Instruction *I, | |||
4951 | ArrayRef<Value *> VectorizedVals) const { | |||
4952 | return (I->hasOneUse() && is_contained(VectorizedVals, I)) || | |||
4953 | all_of(I->users(), [this](User *U) { | |||
4954 | return ScalarToTreeEntry.count(U) > 0 || | |||
4955 | isVectorLikeInstWithConstOps(U) || | |||
4956 | (isa<ExtractElementInst>(U) && MustGather.contains(U)); | |||
4957 | }); | |||
4958 | } | |||
4959 | ||||
4960 | static std::pair<InstructionCost, InstructionCost> | |||
4961 | getVectorCallCosts(CallInst *CI, FixedVectorType *VecTy, | |||
4962 | TargetTransformInfo *TTI, TargetLibraryInfo *TLI) { | |||
4963 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); | |||
4964 | ||||
4965 | // Calculate the cost of the scalar and vector calls. | |||
4966 | SmallVector<Type *, 4> VecTys; | |||
4967 | for (Use &Arg : CI->args()) | |||
4968 | VecTys.push_back( | |||
4969 | FixedVectorType::get(Arg->getType(), VecTy->getNumElements())); | |||
4970 | FastMathFlags FMF; | |||
4971 | if (auto *FPCI = dyn_cast<FPMathOperator>(CI)) | |||
4972 | FMF = FPCI->getFastMathFlags(); | |||
4973 | SmallVector<const Value *> Arguments(CI->args()); | |||
4974 | IntrinsicCostAttributes CostAttrs(ID, VecTy, Arguments, VecTys, FMF, | |||
4975 | dyn_cast<IntrinsicInst>(CI)); | |||
4976 | auto IntrinsicCost = | |||
4977 | TTI->getIntrinsicInstrCost(CostAttrs, TTI::TCK_RecipThroughput); | |||
4978 | ||||
4979 | auto Shape = VFShape::get(*CI, ElementCount::getFixed(static_cast<unsigned>( | |||
4980 | VecTy->getNumElements())), | |||
4981 | false /*HasGlobalPred*/); | |||
4982 | Function *VecFunc = VFDatabase(*CI).getVectorizedFunction(Shape); | |||
4983 | auto LibCost = IntrinsicCost; | |||
4984 | if (!CI->isNoBuiltin() && VecFunc) { | |||
4985 | // Calculate the cost of the vector library call. | |||
4986 | // If the corresponding vector call is cheaper, return its cost. | |||
4987 | LibCost = TTI->getCallInstrCost(nullptr, VecTy, VecTys, | |||
4988 | TTI::TCK_RecipThroughput); | |||
4989 | } | |||
4990 | return {IntrinsicCost, LibCost}; | |||
4991 | } | |||
4992 | ||||
4993 | /// Compute the cost of creating a vector of type \p VecTy containing the | |||
4994 | /// extracted values from \p VL. | |||
4995 | static InstructionCost | |||
4996 | computeExtractCost(ArrayRef<Value *> VL, FixedVectorType *VecTy, | |||
4997 | TargetTransformInfo::ShuffleKind ShuffleKind, | |||
4998 | ArrayRef<int> Mask, TargetTransformInfo &TTI) { | |||
4999 | unsigned NumOfParts = TTI.getNumberOfParts(VecTy); | |||
5000 | ||||
5001 | if (ShuffleKind != TargetTransformInfo::SK_PermuteSingleSrc || !NumOfParts || | |||
5002 | VecTy->getNumElements() < NumOfParts) | |||
5003 | return TTI.getShuffleCost(ShuffleKind, VecTy, Mask); | |||
5004 | ||||
5005 | bool AllConsecutive = true; | |||
5006 | unsigned EltsPerVector = VecTy->getNumElements() / NumOfParts; | |||
5007 | unsigned Idx = -1; | |||
5008 | InstructionCost Cost = 0; | |||
5009 | ||||
5010 | // Process extracts in blocks of EltsPerVector to check if the source vector | |||
5011 | // operand can be re-used directly. If not, add the cost of creating a shuffle | |||
5012 | // to extract the values into a vector register. | |||
5013 | for (auto *V : VL) { | |||
5014 | ++Idx; | |||
5015 | ||||
5016 | // Need to exclude undefs from analysis. | |||
5017 | if (isa<UndefValue>(V) || Mask[Idx] == UndefMaskElem) | |||
5018 | continue; | |||
5019 | ||||
5020 | // Reached the start of a new vector registers. | |||
5021 | if (Idx % EltsPerVector == 0) { | |||
5022 | AllConsecutive = true; | |||
5023 | continue; | |||
5024 | } | |||
5025 | ||||
5026 | // Check all extracts for a vector register on the target directly | |||
5027 | // extract values in order. | |||
5028 | unsigned CurrentIdx = *getExtractIndex(cast<Instruction>(V)); | |||
5029 | if (!isa<UndefValue>(VL[Idx - 1]) && Mask[Idx - 1] != UndefMaskElem) { | |||
5030 | unsigned PrevIdx = *getExtractIndex(cast<Instruction>(VL[Idx - 1])); | |||
5031 | AllConsecutive &= PrevIdx + 1 == CurrentIdx && | |||
5032 | CurrentIdx % EltsPerVector == Idx % EltsPerVector; | |||
5033 | } | |||
5034 | ||||
5035 | if (AllConsecutive) | |||
5036 | continue; | |||
5037 | ||||
5038 | // Skip all indices, except for the last index per vector block. | |||
5039 | if ((Idx + 1) % EltsPerVector != 0 && Idx + 1 != VL.size()) | |||
5040 | continue; | |||
5041 | ||||
5042 | // If we have a series of extracts which are not consecutive and hence | |||
5043 | // cannot re-use the source vector register directly, compute the shuffle | |||
5044 | // cost to extract the a vector with EltsPerVector elements. | |||
5045 | Cost += TTI.getShuffleCost( | |||
5046 | TargetTransformInfo::SK_PermuteSingleSrc, | |||
5047 | FixedVectorType::get(VecTy->getElementType(), EltsPerVector)); | |||
5048 | } | |||
5049 | return Cost; | |||
5050 | } | |||
5051 | ||||
5052 | /// Build shuffle mask for shuffle graph entries and lists of main and alternate | |||
5053 | /// operations operands. | |||
5054 | static void | |||
5055 | buildShuffleEntryMask(ArrayRef<Value *> VL, ArrayRef<unsigned> ReorderIndices, | |||
5056 | ArrayRef<int> ReusesIndices, | |||
5057 | const function_ref<bool(Instruction *)> IsAltOp, | |||
5058 | SmallVectorImpl<int> &Mask, | |||
5059 | SmallVectorImpl<Value *> *OpScalars = nullptr, | |||
5060 | SmallVectorImpl<Value *> *AltScalars = nullptr) { | |||
5061 | unsigned Sz = VL.size(); | |||
5062 | Mask.assign(Sz, UndefMaskElem); | |||
5063 | SmallVector<int> OrderMask; | |||
5064 | if (!ReorderIndices.empty()) | |||
5065 | inversePermutation(ReorderIndices, OrderMask); | |||
5066 | for (unsigned I = 0; I < Sz; ++I) { | |||
5067 | unsigned Idx = I; | |||
5068 | if (!ReorderIndices.empty()) | |||
5069 | Idx = OrderMask[I]; | |||
5070 | auto *OpInst = cast<Instruction>(VL[Idx]); | |||
5071 | if (IsAltOp(OpInst)) { | |||
5072 | Mask[I] = Sz + Idx; | |||
5073 | if (AltScalars) | |||
5074 | AltScalars->push_back(OpInst); | |||
5075 | } else { | |||
5076 | Mask[I] = Idx; | |||
5077 | if (OpScalars) | |||
5078 | OpScalars->push_back(OpInst); | |||
5079 | } | |||
5080 | } | |||
5081 | if (!ReusesIndices.empty()) { | |||
5082 | SmallVector<int> NewMask(ReusesIndices.size(), UndefMaskElem); | |||
5083 | transform(ReusesIndices, NewMask.begin(), [&Mask](int Idx) { | |||
5084 | return Idx != UndefMaskElem ? Mask[Idx] : UndefMaskElem; | |||
5085 | }); | |||
5086 | Mask.swap(NewMask); | |||
5087 | } | |||
5088 | } | |||
5089 | ||||
5090 | /// Checks if the specified instruction \p I is an alternate operation for the | |||
5091 | /// given \p MainOp and \p AltOp instructions. | |||
5092 | static bool isAlternateInstruction(const Instruction *I, | |||
5093 | const Instruction *MainOp, | |||
5094 | const Instruction *AltOp) { | |||
5095 | if (auto *CI0 = dyn_cast<CmpInst>(MainOp)) { | |||
5096 | auto *AltCI0 = cast<CmpInst>(AltOp); | |||
5097 | auto *CI = cast<CmpInst>(I); | |||
5098 | CmpInst::Predicate P0 = CI0->getPredicate(); | |||
5099 | CmpInst::Predicate AltP0 = AltCI0->getPredicate(); | |||
5100 | assert(P0 != AltP0 && "Expected different main/alternate predicates.")(static_cast <bool> (P0 != AltP0 && "Expected different main/alternate predicates." ) ? void (0) : __assert_fail ("P0 != AltP0 && \"Expected different main/alternate predicates.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 5100, __extension__ __PRETTY_FUNCTION__)); | |||
5101 | CmpInst::Predicate AltP0Swapped = CmpInst::getSwappedPredicate(AltP0); | |||
5102 | CmpInst::Predicate CurrentPred = CI->getPredicate(); | |||
5103 | if (P0 == AltP0Swapped) | |||
5104 | return I == AltCI0 || | |||
5105 | (I != MainOp && | |||
5106 | !areCompatibleCmpOps(CI0->getOperand(0), CI0->getOperand(1), | |||
5107 | CI->getOperand(0), CI->getOperand(1))); | |||
5108 | return AltP0 == CurrentPred || AltP0Swapped == CurrentPred; | |||
5109 | } | |||
5110 | return I->getOpcode() == AltOp->getOpcode(); | |||
5111 | } | |||
5112 | ||||
5113 | InstructionCost BoUpSLP::getEntryCost(const TreeEntry *E, | |||
5114 | ArrayRef<Value *> VectorizedVals) { | |||
5115 | ArrayRef<Value*> VL = E->Scalars; | |||
5116 | ||||
5117 | Type *ScalarTy = VL[0]->getType(); | |||
5118 | if (StoreInst *SI = dyn_cast<StoreInst>(VL[0])) | |||
5119 | ScalarTy = SI->getValueOperand()->getType(); | |||
5120 | else if (CmpInst *CI = dyn_cast<CmpInst>(VL[0])) | |||
5121 | ScalarTy = CI->getOperand(0)->getType(); | |||
5122 | else if (auto *IE = dyn_cast<InsertElementInst>(VL[0])) | |||
5123 | ScalarTy = IE->getOperand(1)->getType(); | |||
5124 | auto *VecTy = FixedVectorType::get(ScalarTy, VL.size()); | |||
5125 | TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput; | |||
5126 | ||||
5127 | // If we have computed a smaller type for the expression, update VecTy so | |||
5128 | // that the costs will be accurate. | |||
5129 | if (MinBWs.count(VL[0])) | |||
5130 | VecTy = FixedVectorType::get( | |||
5131 | IntegerType::get(F->getContext(), MinBWs[VL[0]].first), VL.size()); | |||
5132 | unsigned EntryVF = E->getVectorFactor(); | |||
5133 | auto *FinalVecTy = FixedVectorType::get(VecTy->getElementType(), EntryVF); | |||
5134 | ||||
5135 | bool NeedToShuffleReuses = !E->ReuseShuffleIndices.empty(); | |||
5136 | // FIXME: it tries to fix a problem with MSVC buildbots. | |||
5137 | TargetTransformInfo &TTIRef = *TTI; | |||
5138 | auto &&AdjustExtractsCost = [this, &TTIRef, CostKind, VL, VecTy, | |||
5139 | VectorizedVals, E](InstructionCost &Cost) { | |||
5140 | DenseMap<Value *, int> ExtractVectorsTys; | |||
5141 | SmallPtrSet<Value *, 4> CheckedExtracts; | |||
5142 | for (auto *V : VL) { | |||
5143 | if (isa<UndefValue>(V)) | |||
5144 | continue; | |||
5145 | // If all users of instruction are going to be vectorized and this | |||
5146 | // instruction itself is not going to be vectorized, consider this | |||
5147 | // instruction as dead and remove its cost from the final cost of the | |||
5148 | // vectorized tree. | |||
5149 | // Also, avoid adjusting the cost for extractelements with multiple uses | |||
5150 | // in different graph entries. | |||
5151 | const TreeEntry *VE = getTreeEntry(V); | |||
5152 | if (!CheckedExtracts.insert(V).second || | |||
5153 | !areAllUsersVectorized(cast<Instruction>(V), VectorizedVals) || | |||
5154 | (VE && VE != E)) | |||
5155 | continue; | |||
5156 | auto *EE = cast<ExtractElementInst>(V); | |||
5157 | Optional<unsigned> EEIdx = getExtractIndex(EE); | |||
5158 | if (!EEIdx) | |||
5159 | continue; | |||
5160 | unsigned Idx = *EEIdx; | |||
5161 | if (TTIRef.getNumberOfParts(VecTy) != | |||
5162 | TTIRef.getNumberOfParts(EE->getVectorOperandType())) { | |||
5163 | auto It = | |||
5164 | ExtractVectorsTys.try_emplace(EE->getVectorOperand(), Idx).first; | |||
5165 | It->getSecond() = std::min<int>(It->second, Idx); | |||
5166 | } | |||
5167 | // Take credit for instruction that will become dead. | |||
5168 | if (EE->hasOneUse()) { | |||
5169 | Instruction *Ext = EE->user_back(); | |||
5170 | if ((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) && | |||
5171 | all_of(Ext->users(), | |||
5172 | [](User *U) { return isa<GetElementPtrInst>(U); })) { | |||
5173 | // Use getExtractWithExtendCost() to calculate the cost of | |||
5174 | // extractelement/ext pair. | |||
5175 | Cost -= | |||
5176 | TTIRef.getExtractWithExtendCost(Ext->getOpcode(), Ext->getType(), | |||
5177 | EE->getVectorOperandType(), Idx); | |||
5178 | // Add back the cost of s|zext which is subtracted separately. | |||
5179 | Cost += TTIRef.getCastInstrCost( | |||
5180 | Ext->getOpcode(), Ext->getType(), EE->getType(), | |||
5181 | TTI::getCastContextHint(Ext), CostKind, Ext); | |||
5182 | continue; | |||
5183 | } | |||
5184 | } | |||
5185 | Cost -= TTIRef.getVectorInstrCost(Instruction::ExtractElement, | |||
5186 | EE->getVectorOperandType(), Idx); | |||
5187 | } | |||
5188 | // Add a cost for subvector extracts/inserts if required. | |||
5189 | for (const auto &Data : ExtractVectorsTys) { | |||
5190 | auto *EEVTy = cast<FixedVectorType>(Data.first->getType()); | |||
5191 | unsigned NumElts = VecTy->getNumElements(); | |||
5192 | if (Data.second % NumElts == 0) | |||
5193 | continue; | |||
5194 | if (TTIRef.getNumberOfParts(EEVTy) > TTIRef.getNumberOfParts(VecTy)) { | |||
5195 | unsigned Idx = (Data.second / NumElts) * NumElts; | |||
5196 | unsigned EENumElts = EEVTy->getNumElements(); | |||
5197 | if (Idx + NumElts <= EENumElts) { | |||
5198 | Cost += | |||
5199 | TTIRef.getShuffleCost(TargetTransformInfo::SK_ExtractSubvector, | |||
5200 | EEVTy, None, Idx, VecTy); | |||
5201 | } else { | |||
5202 | // Need to round up the subvector type vectorization factor to avoid a | |||
5203 | // crash in cost model functions. Make SubVT so that Idx + VF of SubVT | |||
5204 | // <= EENumElts. | |||
5205 | auto *SubVT = | |||
5206 | FixedVectorType::get(VecTy->getElementType(), EENumElts - Idx); | |||
5207 | Cost += | |||
5208 | TTIRef.getShuffleCost(TargetTransformInfo::SK_ExtractSubvector, | |||
5209 | EEVTy, None, Idx, SubVT); | |||
5210 | } | |||
5211 | } else { | |||
5212 | Cost += TTIRef.getShuffleCost(TargetTransformInfo::SK_InsertSubvector, | |||
5213 | VecTy, None, 0, EEVTy); | |||
5214 | } | |||
5215 | } | |||
5216 | }; | |||
5217 | if (E->State == TreeEntry::NeedToGather) { | |||
5218 | if (allConstant(VL)) | |||
5219 | return 0; | |||
5220 | if (isa<InsertElementInst>(VL[0])) | |||
5221 | return InstructionCost::getInvalid(); | |||
5222 | SmallVector<int> Mask; | |||
5223 | SmallVector<const TreeEntry *> Entries; | |||
5224 | Optional<TargetTransformInfo::ShuffleKind> Shuffle = | |||
5225 | isGatherShuffledEntry(E, Mask, Entries); | |||
5226 | if (Shuffle.hasValue()) { | |||
5227 | InstructionCost GatherCost = 0; | |||
5228 | if (ShuffleVectorInst::isIdentityMask(Mask)) { | |||
5229 | // Perfect match in the graph, will reuse the previously vectorized | |||
5230 | // node. Cost is 0. | |||
5231 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: perfect diamond match for gather bundle that starts with " << *VL.front() << ".\n"; } } while (false) | |||
5232 | dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: perfect diamond match for gather bundle that starts with " << *VL.front() << ".\n"; } } while (false) | |||
5233 | << "SLP: perfect diamond match for gather bundle that starts with "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: perfect diamond match for gather bundle that starts with " << *VL.front() << ".\n"; } } while (false) | |||
5234 | << *VL.front() << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: perfect diamond match for gather bundle that starts with " << *VL.front() << ".\n"; } } while (false); | |||
5235 | if (NeedToShuffleReuses) | |||
5236 | GatherCost = | |||
5237 | TTI->getShuffleCost(TargetTransformInfo::SK_PermuteSingleSrc, | |||
5238 | FinalVecTy, E->ReuseShuffleIndices); | |||
5239 | } else { | |||
5240 | LLVM_DEBUG(dbgs() << "SLP: shuffled " << Entries.size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: shuffled " << Entries. size() << " entries for bundle that starts with " << *VL.front() << ".\n"; } } while (false) | |||
5241 | << " entries for bundle that starts with "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: shuffled " << Entries. size() << " entries for bundle that starts with " << *VL.front() << ".\n"; } } while (false) | |||
5242 | << *VL.front() << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: shuffled " << Entries. size() << " entries for bundle that starts with " << *VL.front() << ".\n"; } } while (false); | |||
5243 | // Detected that instead of gather we can emit a shuffle of single/two | |||
5244 | // previously vectorized nodes. Add the cost of the permutation rather | |||
5245 | // than gather. | |||
5246 | ::addMask(Mask, E->ReuseShuffleIndices); | |||
5247 | GatherCost = TTI->getShuffleCost(*Shuffle, FinalVecTy, Mask); | |||
5248 | } | |||
5249 | return GatherCost; | |||
5250 | } | |||
5251 | if ((E->getOpcode() == Instruction::ExtractElement || | |||
5252 | all_of(E->Scalars, | |||
5253 | [](Value *V) { | |||
5254 | return isa<ExtractElementInst, UndefValue>(V); | |||
5255 | })) && | |||
5256 | allSameType(VL)) { | |||
5257 | // Check that gather of extractelements can be represented as just a | |||
5258 | // shuffle of a single/two vectors the scalars are extracted from. | |||
5259 | SmallVector<int> Mask; | |||
5260 | Optional<TargetTransformInfo::ShuffleKind> ShuffleKind = | |||
5261 | isFixedVectorShuffle(VL, Mask); | |||
5262 | if (ShuffleKind.hasValue()) { | |||
5263 | // Found the bunch of extractelement instructions that must be gathered | |||
5264 | // into a vector and can be represented as a permutation elements in a | |||
5265 | // single input vector or of 2 input vectors. | |||
5266 | InstructionCost Cost = | |||
5267 | computeExtractCost(VL, VecTy, *ShuffleKind, Mask, *TTI); | |||
5268 | AdjustExtractsCost(Cost); | |||
5269 | if (NeedToShuffleReuses) | |||
5270 | Cost += TTI->getShuffleCost(TargetTransformInfo::SK_PermuteSingleSrc, | |||
5271 | FinalVecTy, E->ReuseShuffleIndices); | |||
5272 | return Cost; | |||
5273 | } | |||
5274 | } | |||
5275 | if (isSplat(VL)) { | |||
5276 | // Found the broadcasting of the single scalar, calculate the cost as the | |||
5277 | // broadcast. | |||
5278 | assert(VecTy == FinalVecTy &&(static_cast <bool> (VecTy == FinalVecTy && "No reused scalars expected for broadcast." ) ? void (0) : __assert_fail ("VecTy == FinalVecTy && \"No reused scalars expected for broadcast.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 5279, __extension__ __PRETTY_FUNCTION__)) | |||
5279 | "No reused scalars expected for broadcast.")(static_cast <bool> (VecTy == FinalVecTy && "No reused scalars expected for broadcast." ) ? void (0) : __assert_fail ("VecTy == FinalVecTy && \"No reused scalars expected for broadcast.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 5279, __extension__ __PRETTY_FUNCTION__)); | |||
5280 | return TTI->getShuffleCost(TargetTransformInfo::SK_Broadcast, VecTy, | |||
5281 | /*Mask=*/None, /*Index=*/0, | |||
5282 | /*SubTp=*/nullptr, /*Args=*/VL); | |||
5283 | } | |||
5284 | InstructionCost ReuseShuffleCost = 0; | |||
5285 | if (NeedToShuffleReuses) | |||
5286 | ReuseShuffleCost = TTI->getShuffleCost( | |||
5287 | TTI::SK_PermuteSingleSrc, FinalVecTy, E->ReuseShuffleIndices); | |||
5288 | // Improve gather cost for gather of loads, if we can group some of the | |||
5289 | // loads into vector loads. | |||
5290 | if (VL.size() > 2 && E->getOpcode() == Instruction::Load && | |||
5291 | !E->isAltShuffle()) { | |||
5292 | BoUpSLP::ValueSet VectorizedLoads; | |||
5293 | unsigned StartIdx = 0; | |||
5294 | unsigned VF = VL.size() / 2; | |||
5295 | unsigned VectorizedCnt = 0; | |||
5296 | unsigned ScatterVectorizeCnt = 0; | |||
5297 | const unsigned Sz = DL->getTypeSizeInBits(E->getMainOp()->getType()); | |||
5298 | for (unsigned MinVF = getMinVF(2 * Sz); VF >= MinVF; VF /= 2) { | |||
5299 | for (unsigned Cnt = StartIdx, End = VL.size(); Cnt + VF <= End; | |||
5300 | Cnt += VF) { | |||
5301 | ArrayRef<Value *> Slice = VL.slice(Cnt, VF); | |||
5302 | if (!VectorizedLoads.count(Slice.front()) && | |||
5303 | !VectorizedLoads.count(Slice.back()) && allSameBlock(Slice)) { | |||
5304 | SmallVector<Value *> PointerOps; | |||
5305 | OrdersType CurrentOrder; | |||
5306 | LoadsState LS = canVectorizeLoads(Slice, Slice.front(), *TTI, *DL, | |||
5307 | *SE, CurrentOrder, PointerOps); | |||
5308 | switch (LS) { | |||
5309 | case LoadsState::Vectorize: | |||
5310 | case LoadsState::ScatterVectorize: | |||
5311 | // Mark the vectorized loads so that we don't vectorize them | |||
5312 | // again. | |||
5313 | if (LS == LoadsState::Vectorize) | |||
5314 | ++VectorizedCnt; | |||
5315 | else | |||
5316 | ++ScatterVectorizeCnt; | |||
5317 | VectorizedLoads.insert(Slice.begin(), Slice.end()); | |||
5318 | // If we vectorized initial block, no need to try to vectorize it | |||
5319 | // again. | |||
5320 | if (Cnt == StartIdx) | |||
5321 | StartIdx += VF; | |||
5322 | break; | |||
5323 | case LoadsState::Gather: | |||
5324 | break; | |||
5325 | } | |||
5326 | } | |||
5327 | } | |||
5328 | // Check if the whole array was vectorized already - exit. | |||
5329 | if (StartIdx >= VL.size()) | |||
5330 | break; | |||
5331 | // Found vectorizable parts - exit. | |||
5332 | if (!VectorizedLoads.empty()) | |||
5333 | break; | |||
5334 | } | |||
5335 | if (!VectorizedLoads.empty()) { | |||
5336 | InstructionCost GatherCost = 0; | |||
5337 | unsigned NumParts = TTI->getNumberOfParts(VecTy); | |||
5338 | bool NeedInsertSubvectorAnalysis = | |||
5339 | !NumParts || (VL.size() / VF) > NumParts; | |||
5340 | // Get the cost for gathered loads. | |||
5341 | for (unsigned I = 0, End = VL.size(); I < End; I += VF) { | |||
5342 | if (VectorizedLoads.contains(VL[I])) | |||
5343 | continue; | |||
5344 | GatherCost += getGatherCost(VL.slice(I, VF)); | |||
5345 | } | |||
5346 | // The cost for vectorized loads. | |||
5347 | InstructionCost ScalarsCost = 0; | |||
5348 | for (Value *V : VectorizedLoads) { | |||
5349 | auto *LI = cast<LoadInst>(V); | |||
5350 | ScalarsCost += TTI->getMemoryOpCost( | |||
5351 | Instruction::Load, LI->getType(), LI->getAlign(), | |||
5352 | LI->getPointerAddressSpace(), CostKind, LI); | |||
5353 | } | |||
5354 | auto *LI = cast<LoadInst>(E->getMainOp()); | |||
5355 | auto *LoadTy = FixedVectorType::get(LI->getType(), VF); | |||
5356 | Align Alignment = LI->getAlign(); | |||
5357 | GatherCost += | |||
5358 | VectorizedCnt * | |||
5359 | TTI->getMemoryOpCost(Instruction::Load, LoadTy, Alignment, | |||
5360 | LI->getPointerAddressSpace(), CostKind, LI); | |||
5361 | GatherCost += ScatterVectorizeCnt * | |||
5362 | TTI->getGatherScatterOpCost( | |||
5363 | Instruction::Load, LoadTy, LI->getPointerOperand(), | |||
5364 | /*VariableMask=*/false, Alignment, CostKind, LI); | |||
5365 | if (NeedInsertSubvectorAnalysis) { | |||
5366 | // Add the cost for the subvectors insert. | |||
5367 | for (int I = VF, E = VL.size(); I < E; I += VF) | |||
5368 | GatherCost += TTI->getShuffleCost(TTI::SK_InsertSubvector, VecTy, | |||
5369 | None, I, LoadTy); | |||
5370 | } | |||
5371 | return ReuseShuffleCost + GatherCost - ScalarsCost; | |||
5372 | } | |||
5373 | } | |||
5374 | return ReuseShuffleCost + getGatherCost(VL); | |||
5375 | } | |||
5376 | InstructionCost CommonCost = 0; | |||
5377 | SmallVector<int> Mask; | |||
5378 | if (!E->ReorderIndices.empty()) { | |||
5379 | SmallVector<int> NewMask; | |||
5380 | if (E->getOpcode() == Instruction::Store) { | |||
5381 | // For stores the order is actually a mask. | |||
5382 | NewMask.resize(E->ReorderIndices.size()); | |||
5383 | copy(E->ReorderIndices, NewMask.begin()); | |||
5384 | } else { | |||
5385 | inversePermutation(E->ReorderIndices, NewMask); | |||
5386 | } | |||
5387 | ::addMask(Mask, NewMask); | |||
5388 | } | |||
5389 | if (NeedToShuffleReuses) | |||
5390 | ::addMask(Mask, E->ReuseShuffleIndices); | |||
5391 | if (!Mask.empty() && !ShuffleVectorInst::isIdentityMask(Mask)) | |||
5392 | CommonCost = | |||
5393 | TTI->getShuffleCost(TTI::SK_PermuteSingleSrc, FinalVecTy, Mask); | |||
5394 | assert((E->State == TreeEntry::Vectorize ||(static_cast <bool> ((E->State == TreeEntry::Vectorize || E->State == TreeEntry::ScatterVectorize) && "Unhandled state" ) ? void (0) : __assert_fail ("(E->State == TreeEntry::Vectorize || E->State == TreeEntry::ScatterVectorize) && \"Unhandled state\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 5396, __extension__ __PRETTY_FUNCTION__)) | |||
5395 | E->State == TreeEntry::ScatterVectorize) &&(static_cast <bool> ((E->State == TreeEntry::Vectorize || E->State == TreeEntry::ScatterVectorize) && "Unhandled state" ) ? void (0) : __assert_fail ("(E->State == TreeEntry::Vectorize || E->State == TreeEntry::ScatterVectorize) && \"Unhandled state\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 5396, __extension__ __PRETTY_FUNCTION__)) | |||
5396 | "Unhandled state")(static_cast <bool> ((E->State == TreeEntry::Vectorize || E->State == TreeEntry::ScatterVectorize) && "Unhandled state" ) ? void (0) : __assert_fail ("(E->State == TreeEntry::Vectorize || E->State == TreeEntry::ScatterVectorize) && \"Unhandled state\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 5396, __extension__ __PRETTY_FUNCTION__)); | |||
5397 | assert(E->getOpcode() && allSameType(VL) && allSameBlock(VL) && "Invalid VL")(static_cast <bool> (E->getOpcode() && allSameType (VL) && allSameBlock(VL) && "Invalid VL") ? void (0) : __assert_fail ("E->getOpcode() && allSameType(VL) && allSameBlock(VL) && \"Invalid VL\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 5397, __extension__ __PRETTY_FUNCTION__)); | |||
5398 | Instruction *VL0 = E->getMainOp(); | |||
5399 | unsigned ShuffleOrOp = | |||
5400 | E->isAltShuffle() ? (unsigned)Instruction::ShuffleVector : E->getOpcode(); | |||
5401 | switch (ShuffleOrOp) { | |||
5402 | case Instruction::PHI: | |||
5403 | return 0; | |||
5404 | ||||
5405 | case Instruction::ExtractValue: | |||
5406 | case Instruction::ExtractElement: { | |||
5407 | // The common cost of removal ExtractElement/ExtractValue instructions + | |||
5408 | // the cost of shuffles, if required to resuffle the original vector. | |||
5409 | if (NeedToShuffleReuses) { | |||
5410 | unsigned Idx = 0; | |||
5411 | for (unsigned I : E->ReuseShuffleIndices) { | |||
5412 | if (ShuffleOrOp == Instruction::ExtractElement) { | |||
5413 | auto *EE = cast<ExtractElementInst>(VL[I]); | |||
5414 | CommonCost -= TTI->getVectorInstrCost(Instruction::ExtractElement, | |||
5415 | EE->getVectorOperandType(), | |||
5416 | *getExtractIndex(EE)); | |||
5417 | } else { | |||
5418 | CommonCost -= TTI->getVectorInstrCost(Instruction::ExtractElement, | |||
5419 | VecTy, Idx); | |||
5420 | ++Idx; | |||
5421 | } | |||
5422 | } | |||
5423 | Idx = EntryVF; | |||
5424 | for (Value *V : VL) { | |||
5425 | if (ShuffleOrOp == Instruction::ExtractElement) { | |||
5426 | auto *EE = cast<ExtractElementInst>(V); | |||
5427 | CommonCost += TTI->getVectorInstrCost(Instruction::ExtractElement, | |||
5428 | EE->getVectorOperandType(), | |||
5429 | *getExtractIndex(EE)); | |||
5430 | } else { | |||
5431 | --Idx; | |||
5432 | CommonCost += TTI->getVectorInstrCost(Instruction::ExtractElement, | |||
5433 | VecTy, Idx); | |||
5434 | } | |||
5435 | } | |||
5436 | } | |||
5437 | if (ShuffleOrOp == Instruction::ExtractValue) { | |||
5438 | for (unsigned I = 0, E = VL.size(); I < E; ++I) { | |||
5439 | auto *EI = cast<Instruction>(VL[I]); | |||
5440 | // Take credit for instruction that will become dead. | |||
5441 | if (EI->hasOneUse()) { | |||
5442 | Instruction *Ext = EI->user_back(); | |||
5443 | if ((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) && | |||
5444 | all_of(Ext->users(), | |||
5445 | [](User *U) { return isa<GetElementPtrInst>(U); })) { | |||
5446 | // Use getExtractWithExtendCost() to calculate the cost of | |||
5447 | // extractelement/ext pair. | |||
5448 | CommonCost -= TTI->getExtractWithExtendCost( | |||
5449 | Ext->getOpcode(), Ext->getType(), VecTy, I); | |||
5450 | // Add back the cost of s|zext which is subtracted separately. | |||
5451 | CommonCost += TTI->getCastInstrCost( | |||
5452 | Ext->getOpcode(), Ext->getType(), EI->getType(), | |||
5453 | TTI::getCastContextHint(Ext), CostKind, Ext); | |||
5454 | continue; | |||
5455 | } | |||
5456 | } | |||
5457 | CommonCost -= | |||
5458 | TTI->getVectorInstrCost(Instruction::ExtractElement, VecTy, I); | |||
5459 | } | |||
5460 | } else { | |||
5461 | AdjustExtractsCost(CommonCost); | |||
5462 | } | |||
5463 | return CommonCost; | |||
5464 | } | |||
5465 | case Instruction::InsertElement: { | |||
5466 | assert(E->ReuseShuffleIndices.empty() &&(static_cast <bool> (E->ReuseShuffleIndices.empty() && "Unique insertelements only are expected.") ? void (0) : __assert_fail ("E->ReuseShuffleIndices.empty() && \"Unique insertelements only are expected.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 5467, __extension__ __PRETTY_FUNCTION__)) | |||
5467 | "Unique insertelements only are expected.")(static_cast <bool> (E->ReuseShuffleIndices.empty() && "Unique insertelements only are expected.") ? void (0) : __assert_fail ("E->ReuseShuffleIndices.empty() && \"Unique insertelements only are expected.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 5467, __extension__ __PRETTY_FUNCTION__)); | |||
5468 | auto *SrcVecTy = cast<FixedVectorType>(VL0->getType()); | |||
5469 | ||||
5470 | unsigned const NumElts = SrcVecTy->getNumElements(); | |||
5471 | unsigned const NumScalars = VL.size(); | |||
5472 | APInt DemandedElts = APInt::getZero(NumElts); | |||
5473 | // TODO: Add support for Instruction::InsertValue. | |||
5474 | SmallVector<int> Mask; | |||
5475 | if (!E->ReorderIndices.empty()) { | |||
5476 | inversePermutation(E->ReorderIndices, Mask); | |||
5477 | Mask.append(NumElts - NumScalars, UndefMaskElem); | |||
5478 | } else { | |||
5479 | Mask.assign(NumElts, UndefMaskElem); | |||
5480 | std::iota(Mask.begin(), std::next(Mask.begin(), NumScalars), 0); | |||
5481 | } | |||
5482 | unsigned Offset = *getInsertIndex(VL0); | |||
5483 | bool IsIdentity = true; | |||
5484 | SmallVector<int> PrevMask(NumElts, UndefMaskElem); | |||
5485 | Mask.swap(PrevMask); | |||
5486 | for (unsigned I = 0; I < NumScalars; ++I) { | |||
5487 | unsigned InsertIdx = *getInsertIndex(VL[PrevMask[I]]); | |||
5488 | DemandedElts.setBit(InsertIdx); | |||
5489 | IsIdentity &= InsertIdx - Offset == I; | |||
5490 | Mask[InsertIdx - Offset] = I; | |||
5491 | } | |||
5492 | assert(Offset < NumElts && "Failed to find vector index offset")(static_cast <bool> (Offset < NumElts && "Failed to find vector index offset" ) ? void (0) : __assert_fail ("Offset < NumElts && \"Failed to find vector index offset\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 5492, __extension__ __PRETTY_FUNCTION__)); | |||
5493 | ||||
5494 | InstructionCost Cost = 0; | |||
5495 | Cost -= TTI->getScalarizationOverhead(SrcVecTy, DemandedElts, | |||
5496 | /*Insert*/ true, /*Extract*/ false); | |||
5497 | ||||
5498 | if (IsIdentity && NumElts != NumScalars && Offset % NumScalars != 0) { | |||
5499 | // FIXME: Replace with SK_InsertSubvector once it is properly supported. | |||
5500 | unsigned Sz = PowerOf2Ceil(Offset + NumScalars); | |||
5501 | Cost += TTI->getShuffleCost( | |||
5502 | TargetTransformInfo::SK_PermuteSingleSrc, | |||
5503 | FixedVectorType::get(SrcVecTy->getElementType(), Sz)); | |||
5504 | } else if (!IsIdentity) { | |||
5505 | auto *FirstInsert = | |||
5506 | cast<Instruction>(*find_if(E->Scalars, [E](Value *V) { | |||
5507 | return !is_contained(E->Scalars, | |||
5508 | cast<Instruction>(V)->getOperand(0)); | |||
5509 | })); | |||
5510 | if (isUndefVector(FirstInsert->getOperand(0))) { | |||
5511 | Cost += TTI->getShuffleCost(TTI::SK_PermuteSingleSrc, SrcVecTy, Mask); | |||
5512 | } else { | |||
5513 | SmallVector<int> InsertMask(NumElts); | |||
5514 | std::iota(InsertMask.begin(), InsertMask.end(), 0); | |||
5515 | for (unsigned I = 0; I < NumElts; I++) { | |||
5516 | if (Mask[I] != UndefMaskElem) | |||
5517 | InsertMask[Offset + I] = NumElts + I; | |||
5518 | } | |||
5519 | Cost += | |||
5520 | TTI->getShuffleCost(TTI::SK_PermuteTwoSrc, SrcVecTy, InsertMask); | |||
5521 | } | |||
5522 | } | |||
5523 | ||||
5524 | return Cost; | |||
5525 | } | |||
5526 | case Instruction::ZExt: | |||
5527 | case Instruction::SExt: | |||
5528 | case Instruction::FPToUI: | |||
5529 | case Instruction::FPToSI: | |||
5530 | case Instruction::FPExt: | |||
5531 | case Instruction::PtrToInt: | |||
5532 | case Instruction::IntToPtr: | |||
5533 | case Instruction::SIToFP: | |||
5534 | case Instruction::UIToFP: | |||
5535 | case Instruction::Trunc: | |||
5536 | case Instruction::FPTrunc: | |||
5537 | case Instruction::BitCast: { | |||
5538 | Type *SrcTy = VL0->getOperand(0)->getType(); | |||
5539 | InstructionCost ScalarEltCost = | |||
5540 | TTI->getCastInstrCost(E->getOpcode(), ScalarTy, SrcTy, | |||
5541 | TTI::getCastContextHint(VL0), CostKind, VL0); | |||
5542 | if (NeedToShuffleReuses) { | |||
5543 | CommonCost -= (EntryVF - VL.size()) * ScalarEltCost; | |||
5544 | } | |||
5545 | ||||
5546 | // Calculate the cost of this instruction. | |||
5547 | InstructionCost ScalarCost = VL.size() * ScalarEltCost; | |||
5548 | ||||
5549 | auto *SrcVecTy = FixedVectorType::get(SrcTy, VL.size()); | |||
5550 | InstructionCost VecCost = 0; | |||
5551 | // Check if the values are candidates to demote. | |||
5552 | if (!MinBWs.count(VL0) || VecTy != SrcVecTy) { | |||
5553 | VecCost = CommonCost + TTI->getCastInstrCost( | |||
5554 | E->getOpcode(), VecTy, SrcVecTy, | |||
5555 | TTI::getCastContextHint(VL0), CostKind, VL0); | |||
5556 | } | |||
5557 | LLVM_DEBUG(dumpTreeCosts(E, CommonCost, VecCost, ScalarCost))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dumpTreeCosts(E, CommonCost, VecCost, ScalarCost); } } while (false); | |||
5558 | return VecCost - ScalarCost; | |||
5559 | } | |||
5560 | case Instruction::FCmp: | |||
5561 | case Instruction::ICmp: | |||
5562 | case Instruction::Select: { | |||
5563 | // Calculate the cost of this instruction. | |||
5564 | InstructionCost ScalarEltCost = | |||
5565 | TTI->getCmpSelInstrCost(E->getOpcode(), ScalarTy, Builder.getInt1Ty(), | |||
5566 | CmpInst::BAD_ICMP_PREDICATE, CostKind, VL0); | |||
5567 | if (NeedToShuffleReuses) { | |||
5568 | CommonCost -= (EntryVF - VL.size()) * ScalarEltCost; | |||
5569 | } | |||
5570 | auto *MaskTy = FixedVectorType::get(Builder.getInt1Ty(), VL.size()); | |||
5571 | InstructionCost ScalarCost = VecTy->getNumElements() * ScalarEltCost; | |||
5572 | ||||
5573 | // Check if all entries in VL are either compares or selects with compares | |||
5574 | // as condition that have the same predicates. | |||
5575 | CmpInst::Predicate VecPred = CmpInst::BAD_ICMP_PREDICATE; | |||
5576 | bool First = true; | |||
5577 | for (auto *V : VL) { | |||
5578 | CmpInst::Predicate CurrentPred; | |||
5579 | auto MatchCmp = m_Cmp(CurrentPred, m_Value(), m_Value()); | |||
5580 | if ((!match(V, m_Select(MatchCmp, m_Value(), m_Value())) && | |||
5581 | !match(V, MatchCmp)) || | |||
5582 | (!First && VecPred != CurrentPred)) { | |||
5583 | VecPred = CmpInst::BAD_ICMP_PREDICATE; | |||
5584 | break; | |||
5585 | } | |||
5586 | First = false; | |||
5587 | VecPred = CurrentPred; | |||
5588 | } | |||
5589 | ||||
5590 | InstructionCost VecCost = TTI->getCmpSelInstrCost( | |||
5591 | E->getOpcode(), VecTy, MaskTy, VecPred, CostKind, VL0); | |||
5592 | // Check if it is possible and profitable to use min/max for selects in | |||
5593 | // VL. | |||
5594 | // | |||
5595 | auto IntrinsicAndUse = canConvertToMinOrMaxIntrinsic(VL); | |||
5596 | if (IntrinsicAndUse.first != Intrinsic::not_intrinsic) { | |||
5597 | IntrinsicCostAttributes CostAttrs(IntrinsicAndUse.first, VecTy, | |||
5598 | {VecTy, VecTy}); | |||
5599 | InstructionCost IntrinsicCost = | |||
5600 | TTI->getIntrinsicInstrCost(CostAttrs, CostKind); | |||
5601 | // If the selects are the only uses of the compares, they will be dead | |||
5602 | // and we can adjust the cost by removing their cost. | |||
5603 | if (IntrinsicAndUse.second) | |||
5604 | IntrinsicCost -= TTI->getCmpSelInstrCost(Instruction::ICmp, VecTy, | |||
5605 | MaskTy, VecPred, CostKind); | |||
5606 | VecCost = std::min(VecCost, IntrinsicCost); | |||
5607 | } | |||
5608 | LLVM_DEBUG(dumpTreeCosts(E, CommonCost, VecCost, ScalarCost))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dumpTreeCosts(E, CommonCost, VecCost, ScalarCost); } } while (false); | |||
5609 | return CommonCost + VecCost - ScalarCost; | |||
5610 | } | |||
5611 | case Instruction::FNeg: | |||
5612 | case Instruction::Add: | |||
5613 | case Instruction::FAdd: | |||
5614 | case Instruction::Sub: | |||
5615 | case Instruction::FSub: | |||
5616 | case Instruction::Mul: | |||
5617 | case Instruction::FMul: | |||
5618 | case Instruction::UDiv: | |||
5619 | case Instruction::SDiv: | |||
5620 | case Instruction::FDiv: | |||
5621 | case Instruction::URem: | |||
5622 | case Instruction::SRem: | |||
5623 | case Instruction::FRem: | |||
5624 | case Instruction::Shl: | |||
5625 | case Instruction::LShr: | |||
5626 | case Instruction::AShr: | |||
5627 | case Instruction::And: | |||
5628 | case Instruction::Or: | |||
5629 | case Instruction::Xor: { | |||
5630 | // Certain instructions can be cheaper to vectorize if they have a | |||
5631 | // constant second vector operand. | |||
5632 | TargetTransformInfo::OperandValueKind Op1VK = | |||
5633 | TargetTransformInfo::OK_AnyValue; | |||
5634 | TargetTransformInfo::OperandValueKind Op2VK = | |||
5635 | TargetTransformInfo::OK_UniformConstantValue; | |||
5636 | TargetTransformInfo::OperandValueProperties Op1VP = | |||
5637 | TargetTransformInfo::OP_None; | |||
5638 | TargetTransformInfo::OperandValueProperties Op2VP = | |||
5639 | TargetTransformInfo::OP_PowerOf2; | |||
5640 | ||||
5641 | // If all operands are exactly the same ConstantInt then set the | |||
5642 | // operand kind to OK_UniformConstantValue. | |||
5643 | // If instead not all operands are constants, then set the operand kind | |||
5644 | // to OK_AnyValue. If all operands are constants but not the same, | |||
5645 | // then set the operand kind to OK_NonUniformConstantValue. | |||
5646 | ConstantInt *CInt0 = nullptr; | |||
5647 | for (unsigned i = 0, e = VL.size(); i < e; ++i) { | |||
5648 | const Instruction *I = cast<Instruction>(VL[i]); | |||
5649 | unsigned OpIdx = isa<BinaryOperator>(I) ? 1 : 0; | |||
5650 | ConstantInt *CInt = dyn_cast<ConstantInt>(I->getOperand(OpIdx)); | |||
5651 | if (!CInt) { | |||
5652 | Op2VK = TargetTransformInfo::OK_AnyValue; | |||
5653 | Op2VP = TargetTransformInfo::OP_None; | |||
5654 | break; | |||
5655 | } | |||
5656 | if (Op2VP == TargetTransformInfo::OP_PowerOf2 && | |||
5657 | !CInt->getValue().isPowerOf2()) | |||
5658 | Op2VP = TargetTransformInfo::OP_None; | |||
5659 | if (i == 0) { | |||
5660 | CInt0 = CInt; | |||
5661 | continue; | |||
5662 | } | |||
5663 | if (CInt0 != CInt) | |||
5664 | Op2VK = TargetTransformInfo::OK_NonUniformConstantValue; | |||
5665 | } | |||
5666 | ||||
5667 | SmallVector<const Value *, 4> Operands(VL0->operand_values()); | |||
5668 | InstructionCost ScalarEltCost = | |||
5669 | TTI->getArithmeticInstrCost(E->getOpcode(), ScalarTy, CostKind, Op1VK, | |||
5670 | Op2VK, Op1VP, Op2VP, Operands, VL0); | |||
5671 | if (NeedToShuffleReuses) { | |||
5672 | CommonCost -= (EntryVF - VL.size()) * ScalarEltCost; | |||
5673 | } | |||
5674 | InstructionCost ScalarCost = VecTy->getNumElements() * ScalarEltCost; | |||
5675 | InstructionCost VecCost = | |||
5676 | TTI->getArithmeticInstrCost(E->getOpcode(), VecTy, CostKind, Op1VK, | |||
5677 | Op2VK, Op1VP, Op2VP, Operands, VL0); | |||
5678 | LLVM_DEBUG(dumpTreeCosts(E, CommonCost, VecCost, ScalarCost))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dumpTreeCosts(E, CommonCost, VecCost, ScalarCost); } } while (false); | |||
5679 | return CommonCost + VecCost - ScalarCost; | |||
5680 | } | |||
5681 | case Instruction::GetElementPtr: { | |||
5682 | TargetTransformInfo::OperandValueKind Op1VK = | |||
5683 | TargetTransformInfo::OK_AnyValue; | |||
5684 | TargetTransformInfo::OperandValueKind Op2VK = | |||
5685 | TargetTransformInfo::OK_UniformConstantValue; | |||
5686 | ||||
5687 | InstructionCost ScalarEltCost = TTI->getArithmeticInstrCost( | |||
5688 | Instruction::Add, ScalarTy, CostKind, Op1VK, Op2VK); | |||
5689 | if (NeedToShuffleReuses) { | |||
5690 | CommonCost -= (EntryVF - VL.size()) * ScalarEltCost; | |||
5691 | } | |||
5692 | InstructionCost ScalarCost = VecTy->getNumElements() * ScalarEltCost; | |||
5693 | InstructionCost VecCost = TTI->getArithmeticInstrCost( | |||
5694 | Instruction::Add, VecTy, CostKind, Op1VK, Op2VK); | |||
5695 | LLVM_DEBUG(dumpTreeCosts(E, CommonCost, VecCost, ScalarCost))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dumpTreeCosts(E, CommonCost, VecCost, ScalarCost); } } while (false); | |||
5696 | return CommonCost + VecCost - ScalarCost; | |||
5697 | } | |||
5698 | case Instruction::Load: { | |||
5699 | // Cost of wide load - cost of scalar loads. | |||
5700 | Align Alignment = cast<LoadInst>(VL0)->getAlign(); | |||
5701 | InstructionCost ScalarEltCost = TTI->getMemoryOpCost( | |||
5702 | Instruction::Load, ScalarTy, Alignment, 0, CostKind, VL0); | |||
5703 | if (NeedToShuffleReuses) { | |||
5704 | CommonCost -= (EntryVF - VL.size()) * ScalarEltCost; | |||
5705 | } | |||
5706 | InstructionCost ScalarLdCost = VecTy->getNumElements() * ScalarEltCost; | |||
5707 | InstructionCost VecLdCost; | |||
5708 | if (E->State == TreeEntry::Vectorize) { | |||
5709 | VecLdCost = TTI->getMemoryOpCost(Instruction::Load, VecTy, Alignment, 0, | |||
5710 | CostKind, VL0); | |||
5711 | } else { | |||
5712 | assert(E->State == TreeEntry::ScatterVectorize && "Unknown EntryState")(static_cast <bool> (E->State == TreeEntry::ScatterVectorize && "Unknown EntryState") ? void (0) : __assert_fail ( "E->State == TreeEntry::ScatterVectorize && \"Unknown EntryState\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 5712, __extension__ __PRETTY_FUNCTION__)); | |||
5713 | Align CommonAlignment = Alignment; | |||
5714 | for (Value *V : VL) | |||
5715 | CommonAlignment = | |||
5716 | commonAlignment(CommonAlignment, cast<LoadInst>(V)->getAlign()); | |||
5717 | VecLdCost = TTI->getGatherScatterOpCost( | |||
5718 | Instruction::Load, VecTy, cast<LoadInst>(VL0)->getPointerOperand(), | |||
5719 | /*VariableMask=*/false, CommonAlignment, CostKind, VL0); | |||
5720 | } | |||
5721 | LLVM_DEBUG(dumpTreeCosts(E, CommonCost, VecLdCost, ScalarLdCost))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dumpTreeCosts(E, CommonCost, VecLdCost, ScalarLdCost ); } } while (false); | |||
5722 | return CommonCost + VecLdCost - ScalarLdCost; | |||
5723 | } | |||
5724 | case Instruction::Store: { | |||
5725 | // We know that we can merge the stores. Calculate the cost. | |||
5726 | bool IsReorder = !E->ReorderIndices.empty(); | |||
5727 | auto *SI = | |||
5728 | cast<StoreInst>(IsReorder ? VL[E->ReorderIndices.front()] : VL0); | |||
5729 | Align Alignment = SI->getAlign(); | |||
5730 | InstructionCost ScalarEltCost = TTI->getMemoryOpCost( | |||
5731 | Instruction::Store, ScalarTy, Alignment, 0, CostKind, VL0); | |||
5732 | InstructionCost ScalarStCost = VecTy->getNumElements() * ScalarEltCost; | |||
5733 | InstructionCost VecStCost = TTI->getMemoryOpCost( | |||
5734 | Instruction::Store, VecTy, Alignment, 0, CostKind, VL0); | |||
5735 | LLVM_DEBUG(dumpTreeCosts(E, CommonCost, VecStCost, ScalarStCost))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dumpTreeCosts(E, CommonCost, VecStCost, ScalarStCost ); } } while (false); | |||
5736 | return CommonCost + VecStCost - ScalarStCost; | |||
5737 | } | |||
5738 | case Instruction::Call: { | |||
5739 | CallInst *CI = cast<CallInst>(VL0); | |||
5740 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); | |||
5741 | ||||
5742 | // Calculate the cost of the scalar and vector calls. | |||
5743 | IntrinsicCostAttributes CostAttrs(ID, *CI, 1); | |||
5744 | InstructionCost ScalarEltCost = | |||
5745 | TTI->getIntrinsicInstrCost(CostAttrs, CostKind); | |||
5746 | if (NeedToShuffleReuses) { | |||
5747 | CommonCost -= (EntryVF - VL.size()) * ScalarEltCost; | |||
5748 | } | |||
5749 | InstructionCost ScalarCallCost = VecTy->getNumElements() * ScalarEltCost; | |||
5750 | ||||
5751 | auto VecCallCosts = getVectorCallCosts(CI, VecTy, TTI, TLI); | |||
5752 | InstructionCost VecCallCost = | |||
5753 | std::min(VecCallCosts.first, VecCallCosts.second); | |||
5754 | ||||
5755 | LLVM_DEBUG(dbgs() << "SLP: Call cost " << VecCallCost - ScalarCallCostdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Call cost " << VecCallCost - ScalarCallCost << " (" << VecCallCost << "-" << ScalarCallCost << ")" << " for " << *CI << "\n"; } } while (false) | |||
5756 | << " (" << VecCallCost << "-" << ScalarCallCost << ")"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Call cost " << VecCallCost - ScalarCallCost << " (" << VecCallCost << "-" << ScalarCallCost << ")" << " for " << *CI << "\n"; } } while (false) | |||
5757 | << " for " << *CI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Call cost " << VecCallCost - ScalarCallCost << " (" << VecCallCost << "-" << ScalarCallCost << ")" << " for " << *CI << "\n"; } } while (false); | |||
5758 | ||||
5759 | return CommonCost + VecCallCost - ScalarCallCost; | |||
5760 | } | |||
5761 | case Instruction::ShuffleVector: { | |||
5762 | assert(E->isAltShuffle() &&(static_cast <bool> (E->isAltShuffle() && (( Instruction::isBinaryOp(E->getOpcode()) && Instruction ::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E ->getOpcode()) && Instruction::isCast(E->getAltOpcode ())) || (isa<CmpInst>(VL0) && isa<CmpInst> (E->getAltOp()))) && "Invalid Shuffle Vector Operand" ) ? void (0) : __assert_fail ("E->isAltShuffle() && ((Instruction::isBinaryOp(E->getOpcode()) && Instruction::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E->getOpcode()) && Instruction::isCast(E->getAltOpcode())) || (isa<CmpInst>(VL0) && isa<CmpInst>(E->getAltOp()))) && \"Invalid Shuffle Vector Operand\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 5768, __extension__ __PRETTY_FUNCTION__)) | |||
5763 | ((Instruction::isBinaryOp(E->getOpcode()) &&(static_cast <bool> (E->isAltShuffle() && (( Instruction::isBinaryOp(E->getOpcode()) && Instruction ::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E ->getOpcode()) && Instruction::isCast(E->getAltOpcode ())) || (isa<CmpInst>(VL0) && isa<CmpInst> (E->getAltOp()))) && "Invalid Shuffle Vector Operand" ) ? void (0) : __assert_fail ("E->isAltShuffle() && ((Instruction::isBinaryOp(E->getOpcode()) && Instruction::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E->getOpcode()) && Instruction::isCast(E->getAltOpcode())) || (isa<CmpInst>(VL0) && isa<CmpInst>(E->getAltOp()))) && \"Invalid Shuffle Vector Operand\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 5768, __extension__ __PRETTY_FUNCTION__)) | |||
5764 | Instruction::isBinaryOp(E->getAltOpcode())) ||(static_cast <bool> (E->isAltShuffle() && (( Instruction::isBinaryOp(E->getOpcode()) && Instruction ::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E ->getOpcode()) && Instruction::isCast(E->getAltOpcode ())) || (isa<CmpInst>(VL0) && isa<CmpInst> (E->getAltOp()))) && "Invalid Shuffle Vector Operand" ) ? void (0) : __assert_fail ("E->isAltShuffle() && ((Instruction::isBinaryOp(E->getOpcode()) && Instruction::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E->getOpcode()) && Instruction::isCast(E->getAltOpcode())) || (isa<CmpInst>(VL0) && isa<CmpInst>(E->getAltOp()))) && \"Invalid Shuffle Vector Operand\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 5768, __extension__ __PRETTY_FUNCTION__)) | |||
5765 | (Instruction::isCast(E->getOpcode()) &&(static_cast <bool> (E->isAltShuffle() && (( Instruction::isBinaryOp(E->getOpcode()) && Instruction ::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E ->getOpcode()) && Instruction::isCast(E->getAltOpcode ())) || (isa<CmpInst>(VL0) && isa<CmpInst> (E->getAltOp()))) && "Invalid Shuffle Vector Operand" ) ? void (0) : __assert_fail ("E->isAltShuffle() && ((Instruction::isBinaryOp(E->getOpcode()) && Instruction::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E->getOpcode()) && Instruction::isCast(E->getAltOpcode())) || (isa<CmpInst>(VL0) && isa<CmpInst>(E->getAltOp()))) && \"Invalid Shuffle Vector Operand\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 5768, __extension__ __PRETTY_FUNCTION__)) | |||
5766 | Instruction::isCast(E->getAltOpcode())) ||(static_cast <bool> (E->isAltShuffle() && (( Instruction::isBinaryOp(E->getOpcode()) && Instruction ::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E ->getOpcode()) && Instruction::isCast(E->getAltOpcode ())) || (isa<CmpInst>(VL0) && isa<CmpInst> (E->getAltOp()))) && "Invalid Shuffle Vector Operand" ) ? void (0) : __assert_fail ("E->isAltShuffle() && ((Instruction::isBinaryOp(E->getOpcode()) && Instruction::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E->getOpcode()) && Instruction::isCast(E->getAltOpcode())) || (isa<CmpInst>(VL0) && isa<CmpInst>(E->getAltOp()))) && \"Invalid Shuffle Vector Operand\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 5768, __extension__ __PRETTY_FUNCTION__)) | |||
5767 | (isa<CmpInst>(VL0) && isa<CmpInst>(E->getAltOp()))) &&(static_cast <bool> (E->isAltShuffle() && (( Instruction::isBinaryOp(E->getOpcode()) && Instruction ::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E ->getOpcode()) && Instruction::isCast(E->getAltOpcode ())) || (isa<CmpInst>(VL0) && isa<CmpInst> (E->getAltOp()))) && "Invalid Shuffle Vector Operand" ) ? void (0) : __assert_fail ("E->isAltShuffle() && ((Instruction::isBinaryOp(E->getOpcode()) && Instruction::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E->getOpcode()) && Instruction::isCast(E->getAltOpcode())) || (isa<CmpInst>(VL0) && isa<CmpInst>(E->getAltOp()))) && \"Invalid Shuffle Vector Operand\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 5768, __extension__ __PRETTY_FUNCTION__)) | |||
5768 | "Invalid Shuffle Vector Operand")(static_cast <bool> (E->isAltShuffle() && (( Instruction::isBinaryOp(E->getOpcode()) && Instruction ::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E ->getOpcode()) && Instruction::isCast(E->getAltOpcode ())) || (isa<CmpInst>(VL0) && isa<CmpInst> (E->getAltOp()))) && "Invalid Shuffle Vector Operand" ) ? void (0) : __assert_fail ("E->isAltShuffle() && ((Instruction::isBinaryOp(E->getOpcode()) && Instruction::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E->getOpcode()) && Instruction::isCast(E->getAltOpcode())) || (isa<CmpInst>(VL0) && isa<CmpInst>(E->getAltOp()))) && \"Invalid Shuffle Vector Operand\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 5768, __extension__ __PRETTY_FUNCTION__)); | |||
5769 | InstructionCost ScalarCost = 0; | |||
5770 | if (NeedToShuffleReuses) { | |||
5771 | for (unsigned Idx : E->ReuseShuffleIndices) { | |||
5772 | Instruction *I = cast<Instruction>(VL[Idx]); | |||
5773 | CommonCost -= TTI->getInstructionCost(I, CostKind); | |||
5774 | } | |||
5775 | for (Value *V : VL) { | |||
5776 | Instruction *I = cast<Instruction>(V); | |||
5777 | CommonCost += TTI->getInstructionCost(I, CostKind); | |||
5778 | } | |||
5779 | } | |||
5780 | for (Value *V : VL) { | |||
5781 | Instruction *I = cast<Instruction>(V); | |||
5782 | assert(E->isOpcodeOrAlt(I) && "Unexpected main/alternate opcode")(static_cast <bool> (E->isOpcodeOrAlt(I) && "Unexpected main/alternate opcode" ) ? void (0) : __assert_fail ("E->isOpcodeOrAlt(I) && \"Unexpected main/alternate opcode\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 5782, __extension__ __PRETTY_FUNCTION__)); | |||
5783 | ScalarCost += TTI->getInstructionCost(I, CostKind); | |||
5784 | } | |||
5785 | // VecCost is equal to sum of the cost of creating 2 vectors | |||
5786 | // and the cost of creating shuffle. | |||
5787 | InstructionCost VecCost = 0; | |||
5788 | // Try to find the previous shuffle node with the same operands and same | |||
5789 | // main/alternate ops. | |||
5790 | auto &&TryFindNodeWithEqualOperands = [this, E]() { | |||
5791 | for (const std::unique_ptr<TreeEntry> &TE : VectorizableTree) { | |||
5792 | if (TE.get() == E) | |||
5793 | break; | |||
5794 | if (TE->isAltShuffle() && | |||
5795 | ((TE->getOpcode() == E->getOpcode() && | |||
5796 | TE->getAltOpcode() == E->getAltOpcode()) || | |||
5797 | (TE->getOpcode() == E->getAltOpcode() && | |||
5798 | TE->getAltOpcode() == E->getOpcode())) && | |||
5799 | TE->hasEqualOperands(*E)) | |||
5800 | return true; | |||
5801 | } | |||
5802 | return false; | |||
5803 | }; | |||
5804 | if (TryFindNodeWithEqualOperands()) { | |||
5805 | LLVM_DEBUG({do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { { dbgs() << "SLP: diamond match for alternate node found.\n" ; E->dump(); }; } } while (false) | |||
5806 | dbgs() << "SLP: diamond match for alternate node found.\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { { dbgs() << "SLP: diamond match for alternate node found.\n" ; E->dump(); }; } } while (false) | |||
5807 | E->dump();do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { { dbgs() << "SLP: diamond match for alternate node found.\n" ; E->dump(); }; } } while (false) | |||
5808 | })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { { dbgs() << "SLP: diamond match for alternate node found.\n" ; E->dump(); }; } } while (false); | |||
5809 | // No need to add new vector costs here since we're going to reuse | |||
5810 | // same main/alternate vector ops, just do different shuffling. | |||
5811 | } else if (Instruction::isBinaryOp(E->getOpcode())) { | |||
5812 | VecCost = TTI->getArithmeticInstrCost(E->getOpcode(), VecTy, CostKind); | |||
5813 | VecCost += TTI->getArithmeticInstrCost(E->getAltOpcode(), VecTy, | |||
5814 | CostKind); | |||
5815 | } else if (auto *CI0 = dyn_cast<CmpInst>(VL0)) { | |||
5816 | VecCost = TTI->getCmpSelInstrCost(E->getOpcode(), ScalarTy, | |||
5817 | Builder.getInt1Ty(), | |||
5818 | CI0->getPredicate(), CostKind, VL0); | |||
5819 | VecCost += TTI->getCmpSelInstrCost( | |||
5820 | E->getOpcode(), ScalarTy, Builder.getInt1Ty(), | |||
5821 | cast<CmpInst>(E->getAltOp())->getPredicate(), CostKind, | |||
5822 | E->getAltOp()); | |||
5823 | } else { | |||
5824 | Type *Src0SclTy = E->getMainOp()->getOperand(0)->getType(); | |||
5825 | Type *Src1SclTy = E->getAltOp()->getOperand(0)->getType(); | |||
5826 | auto *Src0Ty = FixedVectorType::get(Src0SclTy, VL.size()); | |||
5827 | auto *Src1Ty = FixedVectorType::get(Src1SclTy, VL.size()); | |||
5828 | VecCost = TTI->getCastInstrCost(E->getOpcode(), VecTy, Src0Ty, | |||
5829 | TTI::CastContextHint::None, CostKind); | |||
5830 | VecCost += TTI->getCastInstrCost(E->getAltOpcode(), VecTy, Src1Ty, | |||
5831 | TTI::CastContextHint::None, CostKind); | |||
5832 | } | |||
5833 | ||||
5834 | SmallVector<int> Mask; | |||
5835 | buildShuffleEntryMask( | |||
5836 | E->Scalars, E->ReorderIndices, E->ReuseShuffleIndices, | |||
5837 | [E](Instruction *I) { | |||
5838 | assert(E->isOpcodeOrAlt(I) && "Unexpected main/alternate opcode")(static_cast <bool> (E->isOpcodeOrAlt(I) && "Unexpected main/alternate opcode" ) ? void (0) : __assert_fail ("E->isOpcodeOrAlt(I) && \"Unexpected main/alternate opcode\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 5838, __extension__ __PRETTY_FUNCTION__)); | |||
5839 | return isAlternateInstruction(I, E->getMainOp(), E->getAltOp()); | |||
5840 | }, | |||
5841 | Mask); | |||
5842 | CommonCost = | |||
5843 | TTI->getShuffleCost(TargetTransformInfo::SK_Select, FinalVecTy, Mask); | |||
5844 | LLVM_DEBUG(dumpTreeCosts(E, CommonCost, VecCost, ScalarCost))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dumpTreeCosts(E, CommonCost, VecCost, ScalarCost); } } while (false); | |||
5845 | return CommonCost + VecCost - ScalarCost; | |||
5846 | } | |||
5847 | default: | |||
5848 | llvm_unreachable("Unknown instruction")::llvm::llvm_unreachable_internal("Unknown instruction", "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5848); | |||
5849 | } | |||
5850 | } | |||
5851 | ||||
5852 | bool BoUpSLP::isFullyVectorizableTinyTree(bool ForReduction) const { | |||
5853 | LLVM_DEBUG(dbgs() << "SLP: Check whether the tree with height "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Check whether the tree with height " << VectorizableTree.size() << " is fully vectorizable .\n" ; } } while (false) | |||
5854 | << VectorizableTree.size() << " is fully vectorizable .\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Check whether the tree with height " << VectorizableTree.size() << " is fully vectorizable .\n" ; } } while (false); | |||
5855 | ||||
5856 | auto &&AreVectorizableGathers = [this](const TreeEntry *TE, unsigned Limit) { | |||
5857 | SmallVector<int> Mask; | |||
5858 | return TE->State == TreeEntry::NeedToGather && | |||
5859 | !any_of(TE->Scalars, | |||
5860 | [this](Value *V) { return EphValues.contains(V); }) && | |||
5861 | (allConstant(TE->Scalars) || isSplat(TE->Scalars) || | |||
5862 | TE->Scalars.size() < Limit || | |||
5863 | ((TE->getOpcode() == Instruction::ExtractElement || | |||
5864 | all_of(TE->Scalars, | |||
5865 | [](Value *V) { | |||
5866 | return isa<ExtractElementInst, UndefValue>(V); | |||
5867 | })) && | |||
5868 | isFixedVectorShuffle(TE->Scalars, Mask)) || | |||
5869 | (TE->State == TreeEntry::NeedToGather && | |||
5870 | TE->getOpcode() == Instruction::Load && !TE->isAltShuffle())); | |||
5871 | }; | |||
5872 | ||||
5873 | // We only handle trees of heights 1 and 2. | |||
5874 | if (VectorizableTree.size() == 1 && | |||
5875 | (VectorizableTree[0]->State == TreeEntry::Vectorize || | |||
5876 | (ForReduction && | |||
5877 | AreVectorizableGathers(VectorizableTree[0].get(), | |||
5878 | VectorizableTree[0]->Scalars.size()) && | |||
5879 | VectorizableTree[0]->getVectorFactor() > 2))) | |||
5880 | return true; | |||
5881 | ||||
5882 | if (VectorizableTree.size() != 2) | |||
5883 | return false; | |||
5884 | ||||
5885 | // Handle splat and all-constants stores. Also try to vectorize tiny trees | |||
5886 | // with the second gather nodes if they have less scalar operands rather than | |||
5887 | // the initial tree element (may be profitable to shuffle the second gather) | |||
5888 | // or they are extractelements, which form shuffle. | |||
5889 | SmallVector<int> Mask; | |||
5890 | if (VectorizableTree[0]->State == TreeEntry::Vectorize && | |||
5891 | AreVectorizableGathers(VectorizableTree[1].get(), | |||
5892 | VectorizableTree[0]->Scalars.size())) | |||
5893 | return true; | |||
5894 | ||||
5895 | // Gathering cost would be too much for tiny trees. | |||
5896 | if (VectorizableTree[0]->State == TreeEntry::NeedToGather || | |||
5897 | (VectorizableTree[1]->State == TreeEntry::NeedToGather && | |||
5898 | VectorizableTree[0]->State != TreeEntry::ScatterVectorize)) | |||
5899 | return false; | |||
5900 | ||||
5901 | return true; | |||
5902 | } | |||
5903 | ||||
5904 | static bool isLoadCombineCandidateImpl(Value *Root, unsigned NumElts, | |||
5905 | TargetTransformInfo *TTI, | |||
5906 | bool MustMatchOrInst) { | |||
5907 | // Look past the root to find a source value. Arbitrarily follow the | |||
5908 | // path through operand 0 of any 'or'. Also, peek through optional | |||
5909 | // shift-left-by-multiple-of-8-bits. | |||
5910 | Value *ZextLoad = Root; | |||
5911 | const APInt *ShAmtC; | |||
5912 | bool FoundOr = false; | |||
5913 | while (!isa<ConstantExpr>(ZextLoad) && | |||
5914 | (match(ZextLoad, m_Or(m_Value(), m_Value())) || | |||
5915 | (match(ZextLoad, m_Shl(m_Value(), m_APInt(ShAmtC))) && | |||
5916 | ShAmtC->urem(8) == 0))) { | |||
5917 | auto *BinOp = cast<BinaryOperator>(ZextLoad); | |||
5918 | ZextLoad = BinOp->getOperand(0); | |||
5919 | if (BinOp->getOpcode() == Instruction::Or) | |||
5920 | FoundOr = true; | |||
5921 | } | |||
5922 | // Check if the input is an extended load of the required or/shift expression. | |||
5923 | Value *Load; | |||
5924 | if ((MustMatchOrInst && !FoundOr) || ZextLoad == Root || | |||
5925 | !match(ZextLoad, m_ZExt(m_Value(Load))) || !isa<LoadInst>(Load)) | |||
5926 | return false; | |||
5927 | ||||
5928 | // Require that the total load bit width is a legal integer type. | |||
5929 | // For example, <8 x i8> --> i64 is a legal integer on a 64-bit target. | |||
5930 | // But <16 x i8> --> i128 is not, so the backend probably can't reduce it. | |||
5931 | Type *SrcTy = Load->getType(); | |||
5932 | unsigned LoadBitWidth = SrcTy->getIntegerBitWidth() * NumElts; | |||
5933 | if (!TTI->isTypeLegal(IntegerType::get(Root->getContext(), LoadBitWidth))) | |||
5934 | return false; | |||
5935 | ||||
5936 | // Everything matched - assume that we can fold the whole sequence using | |||
5937 | // load combining. | |||
5938 | LLVM_DEBUG(dbgs() << "SLP: Assume load combining for tree starting at "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Assume load combining for tree starting at " << *(cast<Instruction>(Root)) << "\n"; } } while (false) | |||
5939 | << *(cast<Instruction>(Root)) << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Assume load combining for tree starting at " << *(cast<Instruction>(Root)) << "\n"; } } while (false); | |||
5940 | ||||
5941 | return true; | |||
5942 | } | |||
5943 | ||||
5944 | bool BoUpSLP::isLoadCombineReductionCandidate(RecurKind RdxKind) const { | |||
5945 | if (RdxKind != RecurKind::Or) | |||
5946 | return false; | |||
5947 | ||||
5948 | unsigned NumElts = VectorizableTree[0]->Scalars.size(); | |||
5949 | Value *FirstReduced = VectorizableTree[0]->Scalars[0]; | |||
5950 | return isLoadCombineCandidateImpl(FirstReduced, NumElts, TTI, | |||
5951 | /* MatchOr */ false); | |||
5952 | } | |||
5953 | ||||
5954 | bool BoUpSLP::isLoadCombineCandidate() const { | |||
5955 | // Peek through a final sequence of stores and check if all operations are | |||
5956 | // likely to be load-combined. | |||
5957 | unsigned NumElts = VectorizableTree[0]->Scalars.size(); | |||
5958 | for (Value *Scalar : VectorizableTree[0]->Scalars) { | |||
5959 | Value *X; | |||
5960 | if (!match(Scalar, m_Store(m_Value(X), m_Value())) || | |||
5961 | !isLoadCombineCandidateImpl(X, NumElts, TTI, /* MatchOr */ true)) | |||
5962 | return false; | |||
5963 | } | |||
5964 | return true; | |||
5965 | } | |||
5966 | ||||
5967 | bool BoUpSLP::isTreeTinyAndNotFullyVectorizable(bool ForReduction) const { | |||
5968 | // No need to vectorize inserts of gathered values. | |||
5969 | if (VectorizableTree.size() == 2 && | |||
5970 | isa<InsertElementInst>(VectorizableTree[0]->Scalars[0]) && | |||
5971 | VectorizableTree[1]->State == TreeEntry::NeedToGather) | |||
5972 | return true; | |||
5973 | ||||
5974 | // We can vectorize the tree if its size is greater than or equal to the | |||
5975 | // minimum size specified by the MinTreeSize command line option. | |||
5976 | if (VectorizableTree.size() >= MinTreeSize) | |||
5977 | return false; | |||
5978 | ||||
5979 | // If we have a tiny tree (a tree whose size is less than MinTreeSize), we | |||
5980 | // can vectorize it if we can prove it fully vectorizable. | |||
5981 | if (isFullyVectorizableTinyTree(ForReduction)) | |||
5982 | return false; | |||
5983 | ||||
5984 | assert(VectorizableTree.empty()(static_cast <bool> (VectorizableTree.empty() ? ExternalUses .empty() : true && "We shouldn't have any external users" ) ? void (0) : __assert_fail ("VectorizableTree.empty() ? ExternalUses.empty() : true && \"We shouldn't have any external users\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 5986, __extension__ __PRETTY_FUNCTION__)) | |||
5985 | ? ExternalUses.empty()(static_cast <bool> (VectorizableTree.empty() ? ExternalUses .empty() : true && "We shouldn't have any external users" ) ? void (0) : __assert_fail ("VectorizableTree.empty() ? ExternalUses.empty() : true && \"We shouldn't have any external users\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 5986, __extension__ __PRETTY_FUNCTION__)) | |||
5986 | : true && "We shouldn't have any external users")(static_cast <bool> (VectorizableTree.empty() ? ExternalUses .empty() : true && "We shouldn't have any external users" ) ? void (0) : __assert_fail ("VectorizableTree.empty() ? ExternalUses.empty() : true && \"We shouldn't have any external users\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 5986, __extension__ __PRETTY_FUNCTION__)); | |||
5987 | ||||
5988 | // Otherwise, we can't vectorize the tree. It is both tiny and not fully | |||
5989 | // vectorizable. | |||
5990 | return true; | |||
5991 | } | |||
5992 | ||||
5993 | InstructionCost BoUpSLP::getSpillCost() const { | |||
5994 | // Walk from the bottom of the tree to the top, tracking which values are | |||
5995 | // live. When we see a call instruction that is not part of our tree, | |||
5996 | // query TTI to see if there is a cost to keeping values live over it | |||
5997 | // (for example, if spills and fills are required). | |||
5998 | unsigned BundleWidth = VectorizableTree.front()->Scalars.size(); | |||
5999 | InstructionCost Cost = 0; | |||
6000 | ||||
6001 | SmallPtrSet<Instruction*, 4> LiveValues; | |||
6002 | Instruction *PrevInst = nullptr; | |||
6003 | ||||
6004 | // The entries in VectorizableTree are not necessarily ordered by their | |||
6005 | // position in basic blocks. Collect them and order them by dominance so later | |||
6006 | // instructions are guaranteed to be visited first. For instructions in | |||
6007 | // different basic blocks, we only scan to the beginning of the block, so | |||
6008 | // their order does not matter, as long as all instructions in a basic block | |||
6009 | // are grouped together. Using dominance ensures a deterministic order. | |||
6010 | SmallVector<Instruction *, 16> OrderedScalars; | |||
6011 | for (const auto &TEPtr : VectorizableTree) { | |||
6012 | Instruction *Inst = dyn_cast<Instruction>(TEPtr->Scalars[0]); | |||
6013 | if (!Inst) | |||
6014 | continue; | |||
6015 | OrderedScalars.push_back(Inst); | |||
6016 | } | |||
6017 | llvm::sort(OrderedScalars, [&](Instruction *A, Instruction *B) { | |||
6018 | auto *NodeA = DT->getNode(A->getParent()); | |||
6019 | auto *NodeB = DT->getNode(B->getParent()); | |||
6020 | assert(NodeA && "Should only process reachable instructions")(static_cast <bool> (NodeA && "Should only process reachable instructions" ) ? void (0) : __assert_fail ("NodeA && \"Should only process reachable instructions\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 6020, __extension__ __PRETTY_FUNCTION__)); | |||
6021 | assert(NodeB && "Should only process reachable instructions")(static_cast <bool> (NodeB && "Should only process reachable instructions" ) ? void (0) : __assert_fail ("NodeB && \"Should only process reachable instructions\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 6021, __extension__ __PRETTY_FUNCTION__)); | |||
6022 | assert((NodeA == NodeB) == (NodeA->getDFSNumIn() == NodeB->getDFSNumIn()) &&(static_cast <bool> ((NodeA == NodeB) == (NodeA->getDFSNumIn () == NodeB->getDFSNumIn()) && "Different nodes should have different DFS numbers" ) ? void (0) : __assert_fail ("(NodeA == NodeB) == (NodeA->getDFSNumIn() == NodeB->getDFSNumIn()) && \"Different nodes should have different DFS numbers\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 6023, __extension__ __PRETTY_FUNCTION__)) | |||
6023 | "Different nodes should have different DFS numbers")(static_cast <bool> ((NodeA == NodeB) == (NodeA->getDFSNumIn () == NodeB->getDFSNumIn()) && "Different nodes should have different DFS numbers" ) ? void (0) : __assert_fail ("(NodeA == NodeB) == (NodeA->getDFSNumIn() == NodeB->getDFSNumIn()) && \"Different nodes should have different DFS numbers\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 6023, __extension__ __PRETTY_FUNCTION__)); | |||
6024 | if (NodeA != NodeB) | |||
6025 | return NodeA->getDFSNumIn() < NodeB->getDFSNumIn(); | |||
6026 | return B->comesBefore(A); | |||
6027 | }); | |||
6028 | ||||
6029 | for (Instruction *Inst : OrderedScalars) { | |||
6030 | if (!PrevInst) { | |||
6031 | PrevInst = Inst; | |||
6032 | continue; | |||
6033 | } | |||
6034 | ||||
6035 | // Update LiveValues. | |||
6036 | LiveValues.erase(PrevInst); | |||
6037 | for (auto &J : PrevInst->operands()) { | |||
6038 | if (isa<Instruction>(&*J) && getTreeEntry(&*J)) | |||
6039 | LiveValues.insert(cast<Instruction>(&*J)); | |||
6040 | } | |||
6041 | ||||
6042 | LLVM_DEBUG({do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { { dbgs() << "SLP: #LV: " << LiveValues .size(); for (auto *X : LiveValues) dbgs() << " " << X->getName(); dbgs() << ", Looking at "; Inst->dump (); }; } } while (false) | |||
6043 | dbgs() << "SLP: #LV: " << LiveValues.size();do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { { dbgs() << "SLP: #LV: " << LiveValues .size(); for (auto *X : LiveValues) dbgs() << " " << X->getName(); dbgs() << ", Looking at "; Inst->dump (); }; } } while (false) | |||
6044 | for (auto *X : LiveValues)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { { dbgs() << "SLP: #LV: " << LiveValues .size(); for (auto *X : LiveValues) dbgs() << " " << X->getName(); dbgs() << ", Looking at "; Inst->dump (); }; } } while (false) | |||
6045 | dbgs() << " " << X->getName();do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { { dbgs() << "SLP: #LV: " << LiveValues .size(); for (auto *X : LiveValues) dbgs() << " " << X->getName(); dbgs() << ", Looking at "; Inst->dump (); }; } } while (false) | |||
6046 | dbgs() << ", Looking at ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { { dbgs() << "SLP: #LV: " << LiveValues .size(); for (auto *X : LiveValues) dbgs() << " " << X->getName(); dbgs() << ", Looking at "; Inst->dump (); }; } } while (false) | |||
6047 | Inst->dump();do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { { dbgs() << "SLP: #LV: " << LiveValues .size(); for (auto *X : LiveValues) dbgs() << " " << X->getName(); dbgs() << ", Looking at "; Inst->dump (); }; } } while (false) | |||
6048 | })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { { dbgs() << "SLP: #LV: " << LiveValues .size(); for (auto *X : LiveValues) dbgs() << " " << X->getName(); dbgs() << ", Looking at "; Inst->dump (); }; } } while (false); | |||
6049 | ||||
6050 | // Now find the sequence of instructions between PrevInst and Inst. | |||
6051 | unsigned NumCalls = 0; | |||
6052 | BasicBlock::reverse_iterator InstIt = ++Inst->getIterator().getReverse(), | |||
6053 | PrevInstIt = | |||
6054 | PrevInst->getIterator().getReverse(); | |||
6055 | while (InstIt != PrevInstIt) { | |||
6056 | if (PrevInstIt == PrevInst->getParent()->rend()) { | |||
6057 | PrevInstIt = Inst->getParent()->rbegin(); | |||
6058 | continue; | |||
6059 | } | |||
6060 | ||||
6061 | // Debug information does not impact spill cost. | |||
6062 | if ((isa<CallInst>(&*PrevInstIt) && | |||
6063 | !isa<DbgInfoIntrinsic>(&*PrevInstIt)) && | |||
6064 | &*PrevInstIt != PrevInst) | |||
6065 | NumCalls++; | |||
6066 | ||||
6067 | ++PrevInstIt; | |||
6068 | } | |||
6069 | ||||
6070 | if (NumCalls) { | |||
6071 | SmallVector<Type*, 4> V; | |||
6072 | for (auto *II : LiveValues) { | |||
6073 | auto *ScalarTy = II->getType(); | |||
6074 | if (auto *VectorTy = dyn_cast<FixedVectorType>(ScalarTy)) | |||
6075 | ScalarTy = VectorTy->getElementType(); | |||
6076 | V.push_back(FixedVectorType::get(ScalarTy, BundleWidth)); | |||
6077 | } | |||
6078 | Cost += NumCalls * TTI->getCostOfKeepingLiveOverCall(V); | |||
6079 | } | |||
6080 | ||||
6081 | PrevInst = Inst; | |||
6082 | } | |||
6083 | ||||
6084 | return Cost; | |||
6085 | } | |||
6086 | ||||
6087 | /// Check if two insertelement instructions are from the same buildvector. | |||
6088 | static bool areTwoInsertFromSameBuildVector(InsertElementInst *VU, | |||
6089 | InsertElementInst *V) { | |||
6090 | // Instructions must be from the same basic blocks. | |||
6091 | if (VU->getParent() != V->getParent()) | |||
6092 | return false; | |||
6093 | // Checks if 2 insertelements are from the same buildvector. | |||
6094 | if (VU->getType() != V->getType()) | |||
6095 | return false; | |||
6096 | // Multiple used inserts are separate nodes. | |||
6097 | if (!VU->hasOneUse() && !V->hasOneUse()) | |||
6098 | return false; | |||
6099 | auto *IE1 = VU; | |||
6100 | auto *IE2 = V; | |||
6101 | // Go through the vector operand of insertelement instructions trying to find | |||
6102 | // either VU as the original vector for IE2 or V as the original vector for | |||
6103 | // IE1. | |||
6104 | do { | |||
6105 | if (IE2 == VU || IE1 == V) | |||
6106 | return true; | |||
6107 | if (IE1) { | |||
6108 | if (IE1 != VU && !IE1->hasOneUse()) | |||
6109 | IE1 = nullptr; | |||
6110 | else | |||
6111 | IE1 = dyn_cast<InsertElementInst>(IE1->getOperand(0)); | |||
6112 | } | |||
6113 | if (IE2) { | |||
6114 | if (IE2 != V && !IE2->hasOneUse()) | |||
6115 | IE2 = nullptr; | |||
6116 | else | |||
6117 | IE2 = dyn_cast<InsertElementInst>(IE2->getOperand(0)); | |||
6118 | } | |||
6119 | } while (IE1 || IE2); | |||
6120 | return false; | |||
6121 | } | |||
6122 | ||||
6123 | InstructionCost BoUpSLP::getTreeCost(ArrayRef<Value *> VectorizedVals) { | |||
6124 | InstructionCost Cost = 0; | |||
6125 | LLVM_DEBUG(dbgs() << "SLP: Calculating cost for tree of size "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Calculating cost for tree of size " << VectorizableTree.size() << ".\n"; } } while ( false) | |||
6126 | << VectorizableTree.size() << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Calculating cost for tree of size " << VectorizableTree.size() << ".\n"; } } while ( false); | |||
6127 | ||||
6128 | unsigned BundleWidth = VectorizableTree[0]->Scalars.size(); | |||
6129 | ||||
6130 | for (unsigned I = 0, E = VectorizableTree.size(); I < E; ++I) { | |||
6131 | TreeEntry &TE = *VectorizableTree[I]; | |||
6132 | ||||
6133 | InstructionCost C = getEntryCost(&TE, VectorizedVals); | |||
6134 | Cost += C; | |||
6135 | LLVM_DEBUG(dbgs() << "SLP: Adding cost " << Cdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Adding cost " << C << " for bundle that starts with " << *TE.Scalars[0] << ".\n" << "SLP: Current total cost = " << Cost << "\n"; } } while (false) | |||
6136 | << " for bundle that starts with " << *TE.Scalars[0]do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Adding cost " << C << " for bundle that starts with " << *TE.Scalars[0] << ".\n" << "SLP: Current total cost = " << Cost << "\n"; } } while (false) | |||
6137 | << ".\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Adding cost " << C << " for bundle that starts with " << *TE.Scalars[0] << ".\n" << "SLP: Current total cost = " << Cost << "\n"; } } while (false) | |||
6138 | << "SLP: Current total cost = " << Cost << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Adding cost " << C << " for bundle that starts with " << *TE.Scalars[0] << ".\n" << "SLP: Current total cost = " << Cost << "\n"; } } while (false); | |||
6139 | } | |||
6140 | ||||
6141 | SmallPtrSet<Value *, 16> ExtractCostCalculated; | |||
6142 | InstructionCost ExtractCost = 0; | |||
6143 | SmallVector<unsigned> VF; | |||
6144 | SmallVector<SmallVector<int>> ShuffleMask; | |||
6145 | SmallVector<Value *> FirstUsers; | |||
6146 | SmallVector<APInt> DemandedElts; | |||
6147 | for (ExternalUser &EU : ExternalUses) { | |||
6148 | // We only add extract cost once for the same scalar. | |||
6149 | if (!isa_and_nonnull<InsertElementInst>(EU.User) && | |||
6150 | !ExtractCostCalculated.insert(EU.Scalar).second) | |||
6151 | continue; | |||
6152 | ||||
6153 | // Uses by ephemeral values are free (because the ephemeral value will be | |||
6154 | // removed prior to code generation, and so the extraction will be | |||
6155 | // removed as well). | |||
6156 | if (EphValues.count(EU.User)) | |||
6157 | continue; | |||
6158 | ||||
6159 | // No extract cost for vector "scalar" | |||
6160 | if (isa<FixedVectorType>(EU.Scalar->getType())) | |||
6161 | continue; | |||
6162 | ||||
6163 | // Already counted the cost for external uses when tried to adjust the cost | |||
6164 | // for extractelements, no need to add it again. | |||
6165 | if (isa<ExtractElementInst>(EU.Scalar)) | |||
6166 | continue; | |||
6167 | ||||
6168 | // If found user is an insertelement, do not calculate extract cost but try | |||
6169 | // to detect it as a final shuffled/identity match. | |||
6170 | if (auto *VU = dyn_cast_or_null<InsertElementInst>(EU.User)) { | |||
6171 | if (auto *FTy = dyn_cast<FixedVectorType>(VU->getType())) { | |||
6172 | Optional<unsigned> InsertIdx = getInsertIndex(VU); | |||
6173 | if (InsertIdx) { | |||
6174 | auto *It = find_if(FirstUsers, [VU](Value *V) { | |||
6175 | return areTwoInsertFromSameBuildVector(VU, | |||
6176 | cast<InsertElementInst>(V)); | |||
6177 | }); | |||
6178 | int VecId = -1; | |||
6179 | if (It == FirstUsers.end()) { | |||
6180 | VF.push_back(FTy->getNumElements()); | |||
6181 | ShuffleMask.emplace_back(VF.back(), UndefMaskElem); | |||
6182 | // Find the insertvector, vectorized in tree, if any. | |||
6183 | Value *Base = VU; | |||
6184 | while (isa<InsertElementInst>(Base)) { | |||
6185 | // Build the mask for the vectorized insertelement instructions. | |||
6186 | if (const TreeEntry *E = getTreeEntry(Base)) { | |||
6187 | VU = cast<InsertElementInst>(Base); | |||
6188 | do { | |||
6189 | int Idx = E->findLaneForValue(Base); | |||
6190 | ShuffleMask.back()[Idx] = Idx; | |||
6191 | Base = cast<InsertElementInst>(Base)->getOperand(0); | |||
6192 | } while (E == getTreeEntry(Base)); | |||
6193 | break; | |||
6194 | } | |||
6195 | Base = cast<InsertElementInst>(Base)->getOperand(0); | |||
6196 | } | |||
6197 | FirstUsers.push_back(VU); | |||
6198 | DemandedElts.push_back(APInt::getZero(VF.back())); | |||
6199 | VecId = FirstUsers.size() - 1; | |||
6200 | } else { | |||
6201 | VecId = std::distance(FirstUsers.begin(), It); | |||
6202 | } | |||
6203 | ShuffleMask[VecId][*InsertIdx] = EU.Lane; | |||
6204 | DemandedElts[VecId].setBit(*InsertIdx); | |||
6205 | continue; | |||
6206 | } | |||
6207 | } | |||
6208 | } | |||
6209 | ||||
6210 | // If we plan to rewrite the tree in a smaller type, we will need to sign | |||
6211 | // extend the extracted value back to the original type. Here, we account | |||
6212 | // for the extract and the added cost of the sign extend if needed. | |||
6213 | auto *VecTy = FixedVectorType::get(EU.Scalar->getType(), BundleWidth); | |||
6214 | auto *ScalarRoot = VectorizableTree[0]->Scalars[0]; | |||
6215 | if (MinBWs.count(ScalarRoot)) { | |||
6216 | auto *MinTy = IntegerType::get(F->getContext(), MinBWs[ScalarRoot].first); | |||
6217 | auto Extend = | |||
6218 | MinBWs[ScalarRoot].second ? Instruction::SExt : Instruction::ZExt; | |||
6219 | VecTy = FixedVectorType::get(MinTy, BundleWidth); | |||
6220 | ExtractCost += TTI->getExtractWithExtendCost(Extend, EU.Scalar->getType(), | |||
6221 | VecTy, EU.Lane); | |||
6222 | } else { | |||
6223 | ExtractCost += | |||
6224 | TTI->getVectorInstrCost(Instruction::ExtractElement, VecTy, EU.Lane); | |||
6225 | } | |||
6226 | } | |||
6227 | ||||
6228 | InstructionCost SpillCost = getSpillCost(); | |||
6229 | Cost += SpillCost + ExtractCost; | |||
6230 | if (FirstUsers.size() == 1) { | |||
6231 | int Limit = ShuffleMask.front().size() * 2; | |||
6232 | if (all_of(ShuffleMask.front(), [Limit](int Idx) { return Idx < Limit; }) && | |||
6233 | !ShuffleVectorInst::isIdentityMask(ShuffleMask.front())) { | |||
6234 | InstructionCost C = TTI->getShuffleCost( | |||
6235 | TTI::SK_PermuteSingleSrc, | |||
6236 | cast<FixedVectorType>(FirstUsers.front()->getType()), | |||
6237 | ShuffleMask.front()); | |||
6238 | LLVM_DEBUG(dbgs() << "SLP: Adding cost " << Cdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Adding cost " << C << " for final shuffle of insertelement external users " << *VectorizableTree.front()->Scalars.front() << ".\n" << "SLP: Current total cost = " << Cost << "\n"; } } while (false) | |||
6239 | << " for final shuffle of insertelement external users "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Adding cost " << C << " for final shuffle of insertelement external users " << *VectorizableTree.front()->Scalars.front() << ".\n" << "SLP: Current total cost = " << Cost << "\n"; } } while (false) | |||
6240 | << *VectorizableTree.front()->Scalars.front() << ".\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Adding cost " << C << " for final shuffle of insertelement external users " << *VectorizableTree.front()->Scalars.front() << ".\n" << "SLP: Current total cost = " << Cost << "\n"; } } while (false) | |||
6241 | << "SLP: Current total cost = " << Cost << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Adding cost " << C << " for final shuffle of insertelement external users " << *VectorizableTree.front()->Scalars.front() << ".\n" << "SLP: Current total cost = " << Cost << "\n"; } } while (false); | |||
6242 | Cost += C; | |||
6243 | } | |||
6244 | InstructionCost InsertCost = TTI->getScalarizationOverhead( | |||
6245 | cast<FixedVectorType>(FirstUsers.front()->getType()), | |||
6246 | DemandedElts.front(), /*Insert*/ true, /*Extract*/ false); | |||
6247 | LLVM_DEBUG(dbgs() << "SLP: subtracting the cost " << InsertCostdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: subtracting the cost " << InsertCost << " for insertelements gather.\n" << "SLP: Current total cost = " << Cost << "\n"; } } while (false) | |||
6248 | << " for insertelements gather.\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: subtracting the cost " << InsertCost << " for insertelements gather.\n" << "SLP: Current total cost = " << Cost << "\n"; } } while (false) | |||
6249 | << "SLP: Current total cost = " << Cost << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: subtracting the cost " << InsertCost << " for insertelements gather.\n" << "SLP: Current total cost = " << Cost << "\n"; } } while (false); | |||
6250 | Cost -= InsertCost; | |||
6251 | } else if (FirstUsers.size() >= 2) { | |||
6252 | unsigned MaxVF = *std::max_element(VF.begin(), VF.end()); | |||
6253 | // Combined masks of the first 2 vectors. | |||
6254 | SmallVector<int> CombinedMask(MaxVF, UndefMaskElem); | |||
6255 | copy(ShuffleMask.front(), CombinedMask.begin()); | |||
6256 | APInt CombinedDemandedElts = DemandedElts.front().zextOrSelf(MaxVF); | |||
6257 | auto *VecTy = FixedVectorType::get( | |||
6258 | cast<VectorType>(FirstUsers.front()->getType())->getElementType(), | |||
6259 | MaxVF); | |||
6260 | for (int I = 0, E = ShuffleMask[1].size(); I < E; ++I) { | |||
6261 | if (ShuffleMask[1][I] != UndefMaskElem) { | |||
6262 | CombinedMask[I] = ShuffleMask[1][I] + MaxVF; | |||
6263 | CombinedDemandedElts.setBit(I); | |||
6264 | } | |||
6265 | } | |||
6266 | InstructionCost C = | |||
6267 | TTI->getShuffleCost(TTI::SK_PermuteTwoSrc, VecTy, CombinedMask); | |||
6268 | LLVM_DEBUG(dbgs() << "SLP: Adding cost " << Cdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Adding cost " << C << " for final shuffle of vector node and external " "insertelement users " << *VectorizableTree.front()->Scalars.front() << ".\n" << "SLP: Current total cost = " << Cost << "\n"; } } while (false) | |||
6269 | << " for final shuffle of vector node and external "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Adding cost " << C << " for final shuffle of vector node and external " "insertelement users " << *VectorizableTree.front()->Scalars.front() << ".\n" << "SLP: Current total cost = " << Cost << "\n"; } } while (false) | |||
6270 | "insertelement users "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Adding cost " << C << " for final shuffle of vector node and external " "insertelement users " << *VectorizableTree.front()->Scalars.front() << ".\n" << "SLP: Current total cost = " << Cost << "\n"; } } while (false) | |||
6271 | << *VectorizableTree.front()->Scalars.front() << ".\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Adding cost " << C << " for final shuffle of vector node and external " "insertelement users " << *VectorizableTree.front()->Scalars.front() << ".\n" << "SLP: Current total cost = " << Cost << "\n"; } } while (false) | |||
6272 | << "SLP: Current total cost = " << Cost << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Adding cost " << C << " for final shuffle of vector node and external " "insertelement users " << *VectorizableTree.front()->Scalars.front() << ".\n" << "SLP: Current total cost = " << Cost << "\n"; } } while (false); | |||
6273 | Cost += C; | |||
6274 | InstructionCost InsertCost = TTI->getScalarizationOverhead( | |||
6275 | VecTy, CombinedDemandedElts, /*Insert*/ true, /*Extract*/ false); | |||
6276 | LLVM_DEBUG(dbgs() << "SLP: subtracting the cost " << InsertCostdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: subtracting the cost " << InsertCost << " for insertelements gather.\n" << "SLP: Current total cost = " << Cost << "\n"; } } while (false) | |||
6277 | << " for insertelements gather.\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: subtracting the cost " << InsertCost << " for insertelements gather.\n" << "SLP: Current total cost = " << Cost << "\n"; } } while (false) | |||
6278 | << "SLP: Current total cost = " << Cost << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: subtracting the cost " << InsertCost << " for insertelements gather.\n" << "SLP: Current total cost = " << Cost << "\n"; } } while (false); | |||
6279 | Cost -= InsertCost; | |||
6280 | for (int I = 2, E = FirstUsers.size(); I < E; ++I) { | |||
6281 | // Other elements - permutation of 2 vectors (the initial one and the | |||
6282 | // next Ith incoming vector). | |||
6283 | unsigned VF = ShuffleMask[I].size(); | |||
6284 | for (unsigned Idx = 0; Idx < VF; ++Idx) { | |||
6285 | int Mask = ShuffleMask[I][Idx]; | |||
6286 | if (Mask != UndefMaskElem) | |||
6287 | CombinedMask[Idx] = MaxVF + Mask; | |||
6288 | else if (CombinedMask[Idx] != UndefMaskElem) | |||
6289 | CombinedMask[Idx] = Idx; | |||
6290 | } | |||
6291 | for (unsigned Idx = VF; Idx < MaxVF; ++Idx) | |||
6292 | if (CombinedMask[Idx] != UndefMaskElem) | |||
6293 | CombinedMask[Idx] = Idx; | |||
6294 | InstructionCost C = | |||
6295 | TTI->getShuffleCost(TTI::SK_PermuteTwoSrc, VecTy, CombinedMask); | |||
6296 | LLVM_DEBUG(dbgs() << "SLP: Adding cost " << Cdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Adding cost " << C << " for final shuffle of vector node and external " "insertelement users " << *VectorizableTree.front()->Scalars.front() << ".\n" << "SLP: Current total cost = " << Cost << "\n"; } } while (false) | |||
6297 | << " for final shuffle of vector node and external "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Adding cost " << C << " for final shuffle of vector node and external " "insertelement users " << *VectorizableTree.front()->Scalars.front() << ".\n" << "SLP: Current total cost = " << Cost << "\n"; } } while (false) | |||
6298 | "insertelement users "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Adding cost " << C << " for final shuffle of vector node and external " "insertelement users " << *VectorizableTree.front()->Scalars.front() << ".\n" << "SLP: Current total cost = " << Cost << "\n"; } } while (false) | |||
6299 | << *VectorizableTree.front()->Scalars.front() << ".\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Adding cost " << C << " for final shuffle of vector node and external " "insertelement users " << *VectorizableTree.front()->Scalars.front() << ".\n" << "SLP: Current total cost = " << Cost << "\n"; } } while (false) | |||
6300 | << "SLP: Current total cost = " << Cost << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Adding cost " << C << " for final shuffle of vector node and external " "insertelement users " << *VectorizableTree.front()->Scalars.front() << ".\n" << "SLP: Current total cost = " << Cost << "\n"; } } while (false); | |||
6301 | Cost += C; | |||
6302 | InstructionCost InsertCost = TTI->getScalarizationOverhead( | |||
6303 | cast<FixedVectorType>(FirstUsers[I]->getType()), DemandedElts[I], | |||
6304 | /*Insert*/ true, /*Extract*/ false); | |||
6305 | LLVM_DEBUG(dbgs() << "SLP: subtracting the cost " << InsertCostdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: subtracting the cost " << InsertCost << " for insertelements gather.\n" << "SLP: Current total cost = " << Cost << "\n"; } } while (false) | |||
6306 | << " for insertelements gather.\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: subtracting the cost " << InsertCost << " for insertelements gather.\n" << "SLP: Current total cost = " << Cost << "\n"; } } while (false) | |||
6307 | << "SLP: Current total cost = " << Cost << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: subtracting the cost " << InsertCost << " for insertelements gather.\n" << "SLP: Current total cost = " << Cost << "\n"; } } while (false); | |||
6308 | Cost -= InsertCost; | |||
6309 | } | |||
6310 | } | |||
6311 | ||||
6312 | #ifndef NDEBUG | |||
6313 | SmallString<256> Str; | |||
6314 | { | |||
6315 | raw_svector_ostream OS(Str); | |||
6316 | OS << "SLP: Spill Cost = " << SpillCost << ".\n" | |||
6317 | << "SLP: Extract Cost = " << ExtractCost << ".\n" | |||
6318 | << "SLP: Total Cost = " << Cost << ".\n"; | |||
6319 | } | |||
6320 | LLVM_DEBUG(dbgs() << Str)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << Str; } } while (false); | |||
6321 | if (ViewSLPTree) | |||
6322 | ViewGraph(this, "SLP" + F->getName(), false, Str); | |||
6323 | #endif | |||
6324 | ||||
6325 | return Cost; | |||
6326 | } | |||
6327 | ||||
6328 | Optional<TargetTransformInfo::ShuffleKind> | |||
6329 | BoUpSLP::isGatherShuffledEntry(const TreeEntry *TE, SmallVectorImpl<int> &Mask, | |||
6330 | SmallVectorImpl<const TreeEntry *> &Entries) { | |||
6331 | // TODO: currently checking only for Scalars in the tree entry, need to count | |||
6332 | // reused elements too for better cost estimation. | |||
6333 | Mask.assign(TE->Scalars.size(), UndefMaskElem); | |||
6334 | Entries.clear(); | |||
6335 | // Build a lists of values to tree entries. | |||
6336 | DenseMap<Value *, SmallPtrSet<const TreeEntry *, 4>> ValueToTEs; | |||
6337 | for (const std::unique_ptr<TreeEntry> &EntryPtr : VectorizableTree) { | |||
6338 | if (EntryPtr.get() == TE) | |||
6339 | break; | |||
6340 | if (EntryPtr->State != TreeEntry::NeedToGather) | |||
6341 | continue; | |||
6342 | for (Value *V : EntryPtr->Scalars) | |||
6343 | ValueToTEs.try_emplace(V).first->getSecond().insert(EntryPtr.get()); | |||
6344 | } | |||
6345 | // Find all tree entries used by the gathered values. If no common entries | |||
6346 | // found - not a shuffle. | |||
6347 | // Here we build a set of tree nodes for each gathered value and trying to | |||
6348 | // find the intersection between these sets. If we have at least one common | |||
6349 | // tree node for each gathered value - we have just a permutation of the | |||
6350 | // single vector. If we have 2 different sets, we're in situation where we | |||
6351 | // have a permutation of 2 input vectors. | |||
6352 | SmallVector<SmallPtrSet<const TreeEntry *, 4>> UsedTEs; | |||
6353 | DenseMap<Value *, int> UsedValuesEntry; | |||
6354 | for (Value *V : TE->Scalars) { | |||
6355 | if (isa<UndefValue>(V)) | |||
6356 | continue; | |||
6357 | // Build a list of tree entries where V is used. | |||
6358 | SmallPtrSet<const TreeEntry *, 4> VToTEs; | |||
6359 | auto It = ValueToTEs.find(V); | |||
6360 | if (It != ValueToTEs.end()) | |||
6361 | VToTEs = It->second; | |||
6362 | if (const TreeEntry *VTE = getTreeEntry(V)) | |||
6363 | VToTEs.insert(VTE); | |||
6364 | if (VToTEs.empty()) | |||
6365 | return None; | |||
6366 | if (UsedTEs.empty()) { | |||
6367 | // The first iteration, just insert the list of nodes to vector. | |||
6368 | UsedTEs.push_back(VToTEs); | |||
6369 | } else { | |||
6370 | // Need to check if there are any previously used tree nodes which use V. | |||
6371 | // If there are no such nodes, consider that we have another one input | |||
6372 | // vector. | |||
6373 | SmallPtrSet<const TreeEntry *, 4> SavedVToTEs(VToTEs); | |||
6374 | unsigned Idx = 0; | |||
6375 | for (SmallPtrSet<const TreeEntry *, 4> &Set : UsedTEs) { | |||
6376 | // Do we have a non-empty intersection of previously listed tree entries | |||
6377 | // and tree entries using current V? | |||
6378 | set_intersect(VToTEs, Set); | |||
6379 | if (!VToTEs.empty()) { | |||
6380 | // Yes, write the new subset and continue analysis for the next | |||
6381 | // scalar. | |||
6382 | Set.swap(VToTEs); | |||
6383 | break; | |||
6384 | } | |||
6385 | VToTEs = SavedVToTEs; | |||
6386 | ++Idx; | |||
6387 | } | |||
6388 | // No non-empty intersection found - need to add a second set of possible | |||
6389 | // source vectors. | |||
6390 | if (Idx == UsedTEs.size()) { | |||
6391 | // If the number of input vectors is greater than 2 - not a permutation, | |||
6392 | // fallback to the regular gather. | |||
6393 | if (UsedTEs.size() == 2) | |||
6394 | return None; | |||
6395 | UsedTEs.push_back(SavedVToTEs); | |||
6396 | Idx = UsedTEs.size() - 1; | |||
6397 | } | |||
6398 | UsedValuesEntry.try_emplace(V, Idx); | |||
6399 | } | |||
6400 | } | |||
6401 | ||||
6402 | unsigned VF = 0; | |||
6403 | if (UsedTEs.size() == 1) { | |||
6404 | // Try to find the perfect match in another gather node at first. | |||
6405 | auto It = find_if(UsedTEs.front(), [TE](const TreeEntry *EntryPtr) { | |||
6406 | return EntryPtr->isSame(TE->Scalars); | |||
6407 | }); | |||
6408 | if (It != UsedTEs.front().end()) { | |||
6409 | Entries.push_back(*It); | |||
6410 | std::iota(Mask.begin(), Mask.end(), 0); | |||
6411 | return TargetTransformInfo::SK_PermuteSingleSrc; | |||
6412 | } | |||
6413 | // No perfect match, just shuffle, so choose the first tree node. | |||
6414 | Entries.push_back(*UsedTEs.front().begin()); | |||
6415 | } else { | |||
6416 | // Try to find nodes with the same vector factor. | |||
6417 | assert(UsedTEs.size() == 2 && "Expected at max 2 permuted entries.")(static_cast <bool> (UsedTEs.size() == 2 && "Expected at max 2 permuted entries." ) ? void (0) : __assert_fail ("UsedTEs.size() == 2 && \"Expected at max 2 permuted entries.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 6417, __extension__ __PRETTY_FUNCTION__)); | |||
6418 | DenseMap<int, const TreeEntry *> VFToTE; | |||
6419 | for (const TreeEntry *TE : UsedTEs.front()) | |||
6420 | VFToTE.try_emplace(TE->getVectorFactor(), TE); | |||
6421 | for (const TreeEntry *TE : UsedTEs.back()) { | |||
6422 | auto It = VFToTE.find(TE->getVectorFactor()); | |||
6423 | if (It != VFToTE.end()) { | |||
6424 | VF = It->first; | |||
6425 | Entries.push_back(It->second); | |||
6426 | Entries.push_back(TE); | |||
6427 | break; | |||
6428 | } | |||
6429 | } | |||
6430 | // No 2 source vectors with the same vector factor - give up and do regular | |||
6431 | // gather. | |||
6432 | if (Entries.empty()) | |||
6433 | return None; | |||
6434 | } | |||
6435 | ||||
6436 | // Build a shuffle mask for better cost estimation and vector emission. | |||
6437 | for (int I = 0, E = TE->Scalars.size(); I < E; ++I) { | |||
6438 | Value *V = TE->Scalars[I]; | |||
6439 | if (isa<UndefValue>(V)) | |||
6440 | continue; | |||
6441 | unsigned Idx = UsedValuesEntry.lookup(V); | |||
6442 | const TreeEntry *VTE = Entries[Idx]; | |||
6443 | int FoundLane = VTE->findLaneForValue(V); | |||
6444 | Mask[I] = Idx * VF + FoundLane; | |||
6445 | // Extra check required by isSingleSourceMaskImpl function (called by | |||
6446 | // ShuffleVectorInst::isSingleSourceMask). | |||
6447 | if (Mask[I] >= 2 * E) | |||
6448 | return None; | |||
6449 | } | |||
6450 | switch (Entries.size()) { | |||
6451 | case 1: | |||
6452 | return TargetTransformInfo::SK_PermuteSingleSrc; | |||
6453 | case 2: | |||
6454 | return TargetTransformInfo::SK_PermuteTwoSrc; | |||
6455 | default: | |||
6456 | break; | |||
6457 | } | |||
6458 | return None; | |||
6459 | } | |||
6460 | ||||
6461 | InstructionCost BoUpSLP::getGatherCost(FixedVectorType *Ty, | |||
6462 | const APInt &ShuffledIndices, | |||
6463 | bool NeedToShuffle) const { | |||
6464 | InstructionCost Cost = | |||
6465 | TTI->getScalarizationOverhead(Ty, ~ShuffledIndices, /*Insert*/ true, | |||
6466 | /*Extract*/ false); | |||
6467 | if (NeedToShuffle) | |||
6468 | Cost += TTI->getShuffleCost(TargetTransformInfo::SK_PermuteSingleSrc, Ty); | |||
6469 | return Cost; | |||
6470 | } | |||
6471 | ||||
6472 | InstructionCost BoUpSLP::getGatherCost(ArrayRef<Value *> VL) const { | |||
6473 | // Find the type of the operands in VL. | |||
6474 | Type *ScalarTy = VL[0]->getType(); | |||
6475 | if (StoreInst *SI = dyn_cast<StoreInst>(VL[0])) | |||
6476 | ScalarTy = SI->getValueOperand()->getType(); | |||
6477 | auto *VecTy = FixedVectorType::get(ScalarTy, VL.size()); | |||
6478 | bool DuplicateNonConst = false; | |||
6479 | // Find the cost of inserting/extracting values from the vector. | |||
6480 | // Check if the same elements are inserted several times and count them as | |||
6481 | // shuffle candidates. | |||
6482 | APInt ShuffledElements = APInt::getZero(VL.size()); | |||
6483 | DenseSet<Value *> UniqueElements; | |||
6484 | // Iterate in reverse order to consider insert elements with the high cost. | |||
6485 | for (unsigned I = VL.size(); I > 0; --I) { | |||
6486 | unsigned Idx = I - 1; | |||
6487 | // No need to shuffle duplicates for constants. | |||
6488 | if (isConstant(VL[Idx])) { | |||
6489 | ShuffledElements.setBit(Idx); | |||
6490 | continue; | |||
6491 | } | |||
6492 | if (!UniqueElements.insert(VL[Idx]).second) { | |||
6493 | DuplicateNonConst = true; | |||
6494 | ShuffledElements.setBit(Idx); | |||
6495 | } | |||
6496 | } | |||
6497 | return getGatherCost(VecTy, ShuffledElements, DuplicateNonConst); | |||
6498 | } | |||
6499 | ||||
6500 | // Perform operand reordering on the instructions in VL and return the reordered | |||
6501 | // operands in Left and Right. | |||
6502 | void BoUpSLP::reorderInputsAccordingToOpcode(ArrayRef<Value *> VL, | |||
6503 | SmallVectorImpl<Value *> &Left, | |||
6504 | SmallVectorImpl<Value *> &Right, | |||
6505 | const DataLayout &DL, | |||
6506 | ScalarEvolution &SE, | |||
6507 | const BoUpSLP &R) { | |||
6508 | if (VL.empty()) | |||
6509 | return; | |||
6510 | VLOperands Ops(VL, DL, SE, R); | |||
6511 | // Reorder the operands in place. | |||
6512 | Ops.reorder(); | |||
6513 | Left = Ops.getVL(0); | |||
6514 | Right = Ops.getVL(1); | |||
6515 | } | |||
6516 | ||||
6517 | void BoUpSLP::setInsertPointAfterBundle(const TreeEntry *E) { | |||
6518 | // Get the basic block this bundle is in. All instructions in the bundle | |||
6519 | // should be in this block. | |||
6520 | auto *Front = E->getMainOp(); | |||
6521 | auto *BB = Front->getParent(); | |||
6522 | assert(llvm::all_of(E->Scalars, [=](Value *V) -> bool {(static_cast <bool> (llvm::all_of(E->Scalars, [=](Value *V) -> bool { auto *I = cast<Instruction>(V); return !E->isOpcodeOrAlt(I) || I->getParent() == BB; })) ? void (0) : __assert_fail ("llvm::all_of(E->Scalars, [=](Value *V) -> bool { auto *I = cast<Instruction>(V); return !E->isOpcodeOrAlt(I) || I->getParent() == BB; })" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 6525, __extension__ __PRETTY_FUNCTION__)) | |||
6523 | auto *I = cast<Instruction>(V);(static_cast <bool> (llvm::all_of(E->Scalars, [=](Value *V) -> bool { auto *I = cast<Instruction>(V); return !E->isOpcodeOrAlt(I) || I->getParent() == BB; })) ? void (0) : __assert_fail ("llvm::all_of(E->Scalars, [=](Value *V) -> bool { auto *I = cast<Instruction>(V); return !E->isOpcodeOrAlt(I) || I->getParent() == BB; })" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 6525, __extension__ __PRETTY_FUNCTION__)) | |||
6524 | return !E->isOpcodeOrAlt(I) || I->getParent() == BB;(static_cast <bool> (llvm::all_of(E->Scalars, [=](Value *V) -> bool { auto *I = cast<Instruction>(V); return !E->isOpcodeOrAlt(I) || I->getParent() == BB; })) ? void (0) : __assert_fail ("llvm::all_of(E->Scalars, [=](Value *V) -> bool { auto *I = cast<Instruction>(V); return !E->isOpcodeOrAlt(I) || I->getParent() == BB; })" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 6525, __extension__ __PRETTY_FUNCTION__)) | |||
6525 | }))(static_cast <bool> (llvm::all_of(E->Scalars, [=](Value *V) -> bool { auto *I = cast<Instruction>(V); return !E->isOpcodeOrAlt(I) || I->getParent() == BB; })) ? void (0) : __assert_fail ("llvm::all_of(E->Scalars, [=](Value *V) -> bool { auto *I = cast<Instruction>(V); return !E->isOpcodeOrAlt(I) || I->getParent() == BB; })" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 6525, __extension__ __PRETTY_FUNCTION__)); | |||
6526 | ||||
6527 | auto &&FindLastInst = [E, Front]() { | |||
6528 | Instruction *LastInst = Front; | |||
6529 | for (Value *V : E->Scalars) { | |||
6530 | auto *I = dyn_cast<Instruction>(V); | |||
6531 | if (!I) | |||
6532 | continue; | |||
6533 | if (LastInst->comesBefore(I)) | |||
6534 | LastInst = I; | |||
6535 | } | |||
6536 | return LastInst; | |||
6537 | }; | |||
6538 | ||||
6539 | auto &&FindFirstInst = [E, Front]() { | |||
6540 | Instruction *FirstInst = Front; | |||
6541 | for (Value *V : E->Scalars) { | |||
6542 | auto *I = dyn_cast<Instruction>(V); | |||
6543 | if (!I) | |||
6544 | continue; | |||
6545 | if (I->comesBefore(FirstInst)) | |||
6546 | FirstInst = I; | |||
6547 | } | |||
6548 | return FirstInst; | |||
6549 | }; | |||
6550 | ||||
6551 | // Set the insert point to the beginning of the basic block if the entry | |||
6552 | // should not be scheduled. | |||
6553 | if (E->State != TreeEntry::NeedToGather && | |||
6554 | doesNotNeedToSchedule(E->Scalars)) { | |||
6555 | BasicBlock::iterator InsertPt; | |||
6556 | if (all_of(E->Scalars, isUsedOutsideBlock)) | |||
6557 | InsertPt = FindLastInst()->getIterator(); | |||
6558 | else | |||
6559 | InsertPt = FindFirstInst()->getIterator(); | |||
6560 | Builder.SetInsertPoint(BB, InsertPt); | |||
6561 | Builder.SetCurrentDebugLocation(Front->getDebugLoc()); | |||
6562 | return; | |||
6563 | } | |||
6564 | ||||
6565 | // The last instruction in the bundle in program order. | |||
6566 | Instruction *LastInst = nullptr; | |||
6567 | ||||
6568 | // Find the last instruction. The common case should be that BB has been | |||
6569 | // scheduled, and the last instruction is VL.back(). So we start with | |||
6570 | // VL.back() and iterate over schedule data until we reach the end of the | |||
6571 | // bundle. The end of the bundle is marked by null ScheduleData. | |||
6572 | if (BlocksSchedules.count(BB)) { | |||
6573 | Value *V = E->isOneOf(E->Scalars.back()); | |||
6574 | if (doesNotNeedToBeScheduled(V)) | |||
6575 | V = *find_if_not(E->Scalars, doesNotNeedToBeScheduled); | |||
6576 | auto *Bundle = BlocksSchedules[BB]->getScheduleData(V); | |||
6577 | if (Bundle && Bundle->isPartOfBundle()) | |||
6578 | for (; Bundle; Bundle = Bundle->NextInBundle) | |||
6579 | if (Bundle->OpValue == Bundle->Inst) | |||
6580 | LastInst = Bundle->Inst; | |||
6581 | } | |||
6582 | ||||
6583 | // LastInst can still be null at this point if there's either not an entry | |||
6584 | // for BB in BlocksSchedules or there's no ScheduleData available for | |||
6585 | // VL.back(). This can be the case if buildTree_rec aborts for various | |||
6586 | // reasons (e.g., the maximum recursion depth is reached, the maximum region | |||
6587 | // size is reached, etc.). ScheduleData is initialized in the scheduling | |||
6588 | // "dry-run". | |||
6589 | // | |||
6590 | // If this happens, we can still find the last instruction by brute force. We | |||
6591 | // iterate forwards from Front (inclusive) until we either see all | |||
6592 | // instructions in the bundle or reach the end of the block. If Front is the | |||
6593 | // last instruction in program order, LastInst will be set to Front, and we | |||
6594 | // will visit all the remaining instructions in the block. | |||
6595 | // | |||
6596 | // One of the reasons we exit early from buildTree_rec is to place an upper | |||
6597 | // bound on compile-time. Thus, taking an additional compile-time hit here is | |||
6598 | // not ideal. However, this should be exceedingly rare since it requires that | |||
6599 | // we both exit early from buildTree_rec and that the bundle be out-of-order | |||
6600 | // (causing us to iterate all the way to the end of the block). | |||
6601 | if (!LastInst) | |||
6602 | LastInst = FindLastInst(); | |||
6603 | assert(LastInst && "Failed to find last instruction in bundle")(static_cast <bool> (LastInst && "Failed to find last instruction in bundle" ) ? void (0) : __assert_fail ("LastInst && \"Failed to find last instruction in bundle\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 6603, __extension__ __PRETTY_FUNCTION__)); | |||
6604 | ||||
6605 | // Set the insertion point after the last instruction in the bundle. Set the | |||
6606 | // debug location to Front. | |||
6607 | Builder.SetInsertPoint(BB, ++LastInst->getIterator()); | |||
6608 | Builder.SetCurrentDebugLocation(Front->getDebugLoc()); | |||
6609 | } | |||
6610 | ||||
6611 | Value *BoUpSLP::gather(ArrayRef<Value *> VL) { | |||
6612 | // List of instructions/lanes from current block and/or the blocks which are | |||
6613 | // part of the current loop. These instructions will be inserted at the end to | |||
6614 | // make it possible to optimize loops and hoist invariant instructions out of | |||
6615 | // the loops body with better chances for success. | |||
6616 | SmallVector<std::pair<Value *, unsigned>, 4> PostponedInsts; | |||
6617 | SmallSet<int, 4> PostponedIndices; | |||
6618 | Loop *L = LI->getLoopFor(Builder.GetInsertBlock()); | |||
6619 | auto &&CheckPredecessor = [](BasicBlock *InstBB, BasicBlock *InsertBB) { | |||
6620 | SmallPtrSet<BasicBlock *, 4> Visited; | |||
6621 | while (InsertBB && InsertBB != InstBB && Visited.insert(InsertBB).second) | |||
6622 | InsertBB = InsertBB->getSinglePredecessor(); | |||
6623 | return InsertBB && InsertBB == InstBB; | |||
6624 | }; | |||
6625 | for (int I = 0, E = VL.size(); I < E; ++I) { | |||
6626 | if (auto *Inst = dyn_cast<Instruction>(VL[I])) | |||
6627 | if ((CheckPredecessor(Inst->getParent(), Builder.GetInsertBlock()) || | |||
6628 | getTreeEntry(Inst) || (L && (L->contains(Inst)))) && | |||
6629 | PostponedIndices.insert(I).second) | |||
6630 | PostponedInsts.emplace_back(Inst, I); | |||
6631 | } | |||
6632 | ||||
6633 | auto &&CreateInsertElement = [this](Value *Vec, Value *V, unsigned Pos) { | |||
6634 | Vec = Builder.CreateInsertElement(Vec, V, Builder.getInt32(Pos)); | |||
6635 | auto *InsElt = dyn_cast<InsertElementInst>(Vec); | |||
6636 | if (!InsElt) | |||
6637 | return Vec; | |||
6638 | GatherShuffleSeq.insert(InsElt); | |||
6639 | CSEBlocks.insert(InsElt->getParent()); | |||
6640 | // Add to our 'need-to-extract' list. | |||
6641 | if (TreeEntry *Entry = getTreeEntry(V)) { | |||
6642 | // Find which lane we need to extract. | |||
6643 | unsigned FoundLane = Entry->findLaneForValue(V); | |||
6644 | ExternalUses.emplace_back(V, InsElt, FoundLane); | |||
6645 | } | |||
6646 | return Vec; | |||
6647 | }; | |||
6648 | Value *Val0 = | |||
6649 | isa<StoreInst>(VL[0]) ? cast<StoreInst>(VL[0])->getValueOperand() : VL[0]; | |||
6650 | FixedVectorType *VecTy = FixedVectorType::get(Val0->getType(), VL.size()); | |||
6651 | Value *Vec = PoisonValue::get(VecTy); | |||
6652 | SmallVector<int> NonConsts; | |||
6653 | // Insert constant values at first. | |||
6654 | for (int I = 0, E = VL.size(); I < E; ++I) { | |||
6655 | if (PostponedIndices.contains(I)) | |||
6656 | continue; | |||
6657 | if (!isConstant(VL[I])) { | |||
6658 | NonConsts.push_back(I); | |||
6659 | continue; | |||
6660 | } | |||
6661 | Vec = CreateInsertElement(Vec, VL[I], I); | |||
6662 | } | |||
6663 | // Insert non-constant values. | |||
6664 | for (int I : NonConsts) | |||
6665 | Vec = CreateInsertElement(Vec, VL[I], I); | |||
6666 | // Append instructions, which are/may be part of the loop, in the end to make | |||
6667 | // it possible to hoist non-loop-based instructions. | |||
6668 | for (const std::pair<Value *, unsigned> &Pair : PostponedInsts) | |||
6669 | Vec = CreateInsertElement(Vec, Pair.first, Pair.second); | |||
6670 | ||||
6671 | return Vec; | |||
6672 | } | |||
6673 | ||||
6674 | namespace { | |||
6675 | /// Merges shuffle masks and emits final shuffle instruction, if required. | |||
6676 | class ShuffleInstructionBuilder { | |||
6677 | IRBuilderBase &Builder; | |||
6678 | const unsigned VF = 0; | |||
6679 | bool IsFinalized = false; | |||
6680 | SmallVector<int, 4> Mask; | |||
6681 | /// Holds all of the instructions that we gathered. | |||
6682 | SetVector<Instruction *> &GatherShuffleSeq; | |||
6683 | /// A list of blocks that we are going to CSE. | |||
6684 | SetVector<BasicBlock *> &CSEBlocks; | |||
6685 | ||||
6686 | public: | |||
6687 | ShuffleInstructionBuilder(IRBuilderBase &Builder, unsigned VF, | |||
6688 | SetVector<Instruction *> &GatherShuffleSeq, | |||
6689 | SetVector<BasicBlock *> &CSEBlocks) | |||
6690 | : Builder(Builder), VF(VF), GatherShuffleSeq(GatherShuffleSeq), | |||
6691 | CSEBlocks(CSEBlocks) {} | |||
6692 | ||||
6693 | /// Adds a mask, inverting it before applying. | |||
6694 | void addInversedMask(ArrayRef<unsigned> SubMask) { | |||
6695 | if (SubMask.empty()) | |||
6696 | return; | |||
6697 | SmallVector<int, 4> NewMask; | |||
6698 | inversePermutation(SubMask, NewMask); | |||
6699 | addMask(NewMask); | |||
6700 | } | |||
6701 | ||||
6702 | /// Functions adds masks, merging them into single one. | |||
6703 | void addMask(ArrayRef<unsigned> SubMask) { | |||
6704 | SmallVector<int, 4> NewMask(SubMask.begin(), SubMask.end()); | |||
6705 | addMask(NewMask); | |||
6706 | } | |||
6707 | ||||
6708 | void addMask(ArrayRef<int> SubMask) { ::addMask(Mask, SubMask); } | |||
6709 | ||||
6710 | Value *finalize(Value *V) { | |||
6711 | IsFinalized = true; | |||
6712 | unsigned ValueVF = cast<FixedVectorType>(V->getType())->getNumElements(); | |||
6713 | if (VF == ValueVF && Mask.empty()) | |||
6714 | return V; | |||
6715 | SmallVector<int, 4> NormalizedMask(VF, UndefMaskElem); | |||
6716 | std::iota(NormalizedMask.begin(), NormalizedMask.end(), 0); | |||
6717 | addMask(NormalizedMask); | |||
6718 | ||||
6719 | if (VF == ValueVF && ShuffleVectorInst::isIdentityMask(Mask)) | |||
6720 | return V; | |||
6721 | Value *Vec = Builder.CreateShuffleVector(V, Mask, "shuffle"); | |||
6722 | if (auto *I = dyn_cast<Instruction>(Vec)) { | |||
6723 | GatherShuffleSeq.insert(I); | |||
6724 | CSEBlocks.insert(I->getParent()); | |||
6725 | } | |||
6726 | return Vec; | |||
6727 | } | |||
6728 | ||||
6729 | ~ShuffleInstructionBuilder() { | |||
6730 | assert((IsFinalized || Mask.empty()) &&(static_cast <bool> ((IsFinalized || Mask.empty()) && "Shuffle construction must be finalized.") ? void (0) : __assert_fail ("(IsFinalized || Mask.empty()) && \"Shuffle construction must be finalized.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 6731, __extension__ __PRETTY_FUNCTION__)) | |||
6731 | "Shuffle construction must be finalized.")(static_cast <bool> ((IsFinalized || Mask.empty()) && "Shuffle construction must be finalized.") ? void (0) : __assert_fail ("(IsFinalized || Mask.empty()) && \"Shuffle construction must be finalized.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 6731, __extension__ __PRETTY_FUNCTION__)); | |||
6732 | } | |||
6733 | }; | |||
6734 | } // namespace | |||
6735 | ||||
6736 | Value *BoUpSLP::vectorizeTree(ArrayRef<Value *> VL) { | |||
6737 | const unsigned VF = VL.size(); | |||
6738 | InstructionsState S = getSameOpcode(VL); | |||
6739 | if (S.getOpcode()) { | |||
6740 | if (TreeEntry *E = getTreeEntry(S.OpValue)) | |||
6741 | if (E->isSame(VL)) { | |||
6742 | Value *V = vectorizeTree(E); | |||
6743 | if (VF != cast<FixedVectorType>(V->getType())->getNumElements()) { | |||
6744 | if (!E->ReuseShuffleIndices.empty()) { | |||
6745 | // Reshuffle to get only unique values. | |||
6746 | // If some of the scalars are duplicated in the vectorization tree | |||
6747 | // entry, we do not vectorize them but instead generate a mask for | |||
6748 | // the reuses. But if there are several users of the same entry, | |||
6749 | // they may have different vectorization factors. This is especially | |||
6750 | // important for PHI nodes. In this case, we need to adapt the | |||
6751 | // resulting instruction for the user vectorization factor and have | |||
6752 | // to reshuffle it again to take only unique elements of the vector. | |||
6753 | // Without this code the function incorrectly returns reduced vector | |||
6754 | // instruction with the same elements, not with the unique ones. | |||
6755 | ||||
6756 | // block: | |||
6757 | // %phi = phi <2 x > { .., %entry} {%shuffle, %block} | |||
6758 | // %2 = shuffle <2 x > %phi, poison, <4 x > <1, 1, 0, 0> | |||
6759 | // ... (use %2) | |||
6760 | // %shuffle = shuffle <2 x> %2, poison, <2 x> {2, 0} | |||
6761 | // br %block | |||
6762 | SmallVector<int> UniqueIdxs(VF, UndefMaskElem); | |||
6763 | SmallSet<int, 4> UsedIdxs; | |||
6764 | int Pos = 0; | |||
6765 | int Sz = VL.size(); | |||
6766 | for (int Idx : E->ReuseShuffleIndices) { | |||
6767 | if (Idx != Sz && Idx != UndefMaskElem && | |||
6768 | UsedIdxs.insert(Idx).second) | |||
6769 | UniqueIdxs[Idx] = Pos; | |||
6770 | ++Pos; | |||
6771 | } | |||
6772 | assert(VF >= UsedIdxs.size() && "Expected vectorization factor "(static_cast <bool> (VF >= UsedIdxs.size() && "Expected vectorization factor " "less than original vector size." ) ? void (0) : __assert_fail ("VF >= UsedIdxs.size() && \"Expected vectorization factor \" \"less than original vector size.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 6773, __extension__ __PRETTY_FUNCTION__)) | |||
6773 | "less than original vector size.")(static_cast <bool> (VF >= UsedIdxs.size() && "Expected vectorization factor " "less than original vector size." ) ? void (0) : __assert_fail ("VF >= UsedIdxs.size() && \"Expected vectorization factor \" \"less than original vector size.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 6773, __extension__ __PRETTY_FUNCTION__)); | |||
6774 | UniqueIdxs.append(VF - UsedIdxs.size(), UndefMaskElem); | |||
6775 | V = Builder.CreateShuffleVector(V, UniqueIdxs, "shrink.shuffle"); | |||
6776 | } else { | |||
6777 | assert(VF < cast<FixedVectorType>(V->getType())->getNumElements() &&(static_cast <bool> (VF < cast<FixedVectorType> (V->getType())->getNumElements() && "Expected vectorization factor less " "than original vector size.") ? void (0) : __assert_fail ("VF < cast<FixedVectorType>(V->getType())->getNumElements() && \"Expected vectorization factor less \" \"than original vector size.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 6779, __extension__ __PRETTY_FUNCTION__)) | |||
6778 | "Expected vectorization factor less "(static_cast <bool> (VF < cast<FixedVectorType> (V->getType())->getNumElements() && "Expected vectorization factor less " "than original vector size.") ? void (0) : __assert_fail ("VF < cast<FixedVectorType>(V->getType())->getNumElements() && \"Expected vectorization factor less \" \"than original vector size.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 6779, __extension__ __PRETTY_FUNCTION__)) | |||
6779 | "than original vector size.")(static_cast <bool> (VF < cast<FixedVectorType> (V->getType())->getNumElements() && "Expected vectorization factor less " "than original vector size.") ? void (0) : __assert_fail ("VF < cast<FixedVectorType>(V->getType())->getNumElements() && \"Expected vectorization factor less \" \"than original vector size.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 6779, __extension__ __PRETTY_FUNCTION__)); | |||
6780 | SmallVector<int> UniformMask(VF, 0); | |||
6781 | std::iota(UniformMask.begin(), UniformMask.end(), 0); | |||
6782 | V = Builder.CreateShuffleVector(V, UniformMask, "shrink.shuffle"); | |||
6783 | } | |||
6784 | if (auto *I = dyn_cast<Instruction>(V)) { | |||
6785 | GatherShuffleSeq.insert(I); | |||
6786 | CSEBlocks.insert(I->getParent()); | |||
6787 | } | |||
6788 | } | |||
6789 | return V; | |||
6790 | } | |||
6791 | } | |||
6792 | ||||
6793 | // Can't vectorize this, so simply build a new vector with each lane | |||
6794 | // corresponding to the requested value. | |||
6795 | return createBuildVector(VL); | |||
6796 | } | |||
6797 | Value *BoUpSLP::createBuildVector(ArrayRef<Value *> VL) { | |||
6798 | unsigned VF = VL.size(); | |||
6799 | // Exploit possible reuse of values across lanes. | |||
6800 | SmallVector<int> ReuseShuffleIndicies; | |||
6801 | SmallVector<Value *> UniqueValues; | |||
6802 | if (VL.size() > 2) { | |||
6803 | DenseMap<Value *, unsigned> UniquePositions; | |||
6804 | unsigned NumValues = | |||
6805 | std::distance(VL.begin(), find_if(reverse(VL), [](Value *V) { | |||
6806 | return !isa<UndefValue>(V); | |||
6807 | }).base()); | |||
6808 | VF = std::max<unsigned>(VF, PowerOf2Ceil(NumValues)); | |||
6809 | int UniqueVals = 0; | |||
6810 | for (Value *V : VL.drop_back(VL.size() - VF)) { | |||
6811 | if (isa<UndefValue>(V)) { | |||
6812 | ReuseShuffleIndicies.emplace_back(UndefMaskElem); | |||
6813 | continue; | |||
6814 | } | |||
6815 | if (isConstant(V)) { | |||
6816 | ReuseShuffleIndicies.emplace_back(UniqueValues.size()); | |||
6817 | UniqueValues.emplace_back(V); | |||
6818 | continue; | |||
6819 | } | |||
6820 | auto Res = UniquePositions.try_emplace(V, UniqueValues.size()); | |||
6821 | ReuseShuffleIndicies.emplace_back(Res.first->second); | |||
6822 | if (Res.second) { | |||
6823 | UniqueValues.emplace_back(V); | |||
6824 | ++UniqueVals; | |||
6825 | } | |||
6826 | } | |||
6827 | if (UniqueVals == 1 && UniqueValues.size() == 1) { | |||
6828 | // Emit pure splat vector. | |||
6829 | ReuseShuffleIndicies.append(VF - ReuseShuffleIndicies.size(), | |||
6830 | UndefMaskElem); | |||
6831 | } else if (UniqueValues.size() >= VF - 1 || UniqueValues.size() <= 1) { | |||
6832 | ReuseShuffleIndicies.clear(); | |||
6833 | UniqueValues.clear(); | |||
6834 | UniqueValues.append(VL.begin(), std::next(VL.begin(), NumValues)); | |||
6835 | } | |||
6836 | UniqueValues.append(VF - UniqueValues.size(), | |||
6837 | PoisonValue::get(VL[0]->getType())); | |||
6838 | VL = UniqueValues; | |||
6839 | } | |||
6840 | ||||
6841 | ShuffleInstructionBuilder ShuffleBuilder(Builder, VF, GatherShuffleSeq, | |||
6842 | CSEBlocks); | |||
6843 | Value *Vec = gather(VL); | |||
6844 | if (!ReuseShuffleIndicies.empty()) { | |||
6845 | ShuffleBuilder.addMask(ReuseShuffleIndicies); | |||
6846 | Vec = ShuffleBuilder.finalize(Vec); | |||
6847 | } | |||
6848 | return Vec; | |||
6849 | } | |||
6850 | ||||
6851 | Value *BoUpSLP::vectorizeTree(TreeEntry *E) { | |||
6852 | IRBuilder<>::InsertPointGuard Guard(Builder); | |||
6853 | ||||
6854 | if (E->VectorizedValue) { | |||
6855 | LLVM_DEBUG(dbgs() << "SLP: Diamond merged for " << *E->Scalars[0] << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Diamond merged for " << *E->Scalars[0] << ".\n"; } } while (false); | |||
6856 | return E->VectorizedValue; | |||
6857 | } | |||
6858 | ||||
6859 | bool NeedToShuffleReuses = !E->ReuseShuffleIndices.empty(); | |||
6860 | unsigned VF = E->getVectorFactor(); | |||
6861 | ShuffleInstructionBuilder ShuffleBuilder(Builder, VF, GatherShuffleSeq, | |||
6862 | CSEBlocks); | |||
6863 | if (E->State == TreeEntry::NeedToGather) { | |||
6864 | if (E->getMainOp()) | |||
6865 | setInsertPointAfterBundle(E); | |||
6866 | Value *Vec; | |||
6867 | SmallVector<int> Mask; | |||
6868 | SmallVector<const TreeEntry *> Entries; | |||
6869 | Optional<TargetTransformInfo::ShuffleKind> Shuffle = | |||
6870 | isGatherShuffledEntry(E, Mask, Entries); | |||
6871 | if (Shuffle.hasValue()) { | |||
6872 | assert((Entries.size() == 1 || Entries.size() == 2) &&(static_cast <bool> ((Entries.size() == 1 || Entries.size () == 2) && "Expected shuffle of 1 or 2 entries.") ? void (0) : __assert_fail ("(Entries.size() == 1 || Entries.size() == 2) && \"Expected shuffle of 1 or 2 entries.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 6873, __extension__ __PRETTY_FUNCTION__)) | |||
6873 | "Expected shuffle of 1 or 2 entries.")(static_cast <bool> ((Entries.size() == 1 || Entries.size () == 2) && "Expected shuffle of 1 or 2 entries.") ? void (0) : __assert_fail ("(Entries.size() == 1 || Entries.size() == 2) && \"Expected shuffle of 1 or 2 entries.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 6873, __extension__ __PRETTY_FUNCTION__)); | |||
6874 | Vec = Builder.CreateShuffleVector(Entries.front()->VectorizedValue, | |||
6875 | Entries.back()->VectorizedValue, Mask); | |||
6876 | if (auto *I = dyn_cast<Instruction>(Vec)) { | |||
6877 | GatherShuffleSeq.insert(I); | |||
6878 | CSEBlocks.insert(I->getParent()); | |||
6879 | } | |||
6880 | } else { | |||
6881 | Vec = gather(E->Scalars); | |||
6882 | } | |||
6883 | if (NeedToShuffleReuses) { | |||
6884 | ShuffleBuilder.addMask(E->ReuseShuffleIndices); | |||
6885 | Vec = ShuffleBuilder.finalize(Vec); | |||
6886 | } | |||
6887 | E->VectorizedValue = Vec; | |||
6888 | return Vec; | |||
6889 | } | |||
6890 | ||||
6891 | assert((E->State == TreeEntry::Vectorize ||(static_cast <bool> ((E->State == TreeEntry::Vectorize || E->State == TreeEntry::ScatterVectorize) && "Unhandled state" ) ? void (0) : __assert_fail ("(E->State == TreeEntry::Vectorize || E->State == TreeEntry::ScatterVectorize) && \"Unhandled state\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 6893, __extension__ __PRETTY_FUNCTION__)) | |||
6892 | E->State == TreeEntry::ScatterVectorize) &&(static_cast <bool> ((E->State == TreeEntry::Vectorize || E->State == TreeEntry::ScatterVectorize) && "Unhandled state" ) ? void (0) : __assert_fail ("(E->State == TreeEntry::Vectorize || E->State == TreeEntry::ScatterVectorize) && \"Unhandled state\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 6893, __extension__ __PRETTY_FUNCTION__)) | |||
6893 | "Unhandled state")(static_cast <bool> ((E->State == TreeEntry::Vectorize || E->State == TreeEntry::ScatterVectorize) && "Unhandled state" ) ? void (0) : __assert_fail ("(E->State == TreeEntry::Vectorize || E->State == TreeEntry::ScatterVectorize) && \"Unhandled state\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 6893, __extension__ __PRETTY_FUNCTION__)); | |||
6894 | unsigned ShuffleOrOp = | |||
6895 | E->isAltShuffle() ? (unsigned)Instruction::ShuffleVector : E->getOpcode(); | |||
6896 | Instruction *VL0 = E->getMainOp(); | |||
6897 | Type *ScalarTy = VL0->getType(); | |||
6898 | if (auto *Store = dyn_cast<StoreInst>(VL0)) | |||
6899 | ScalarTy = Store->getValueOperand()->getType(); | |||
6900 | else if (auto *IE = dyn_cast<InsertElementInst>(VL0)) | |||
6901 | ScalarTy = IE->getOperand(1)->getType(); | |||
6902 | auto *VecTy = FixedVectorType::get(ScalarTy, E->Scalars.size()); | |||
6903 | switch (ShuffleOrOp) { | |||
6904 | case Instruction::PHI: { | |||
6905 | assert((static_cast <bool> ((E->ReorderIndices.empty() || E != VectorizableTree.front().get()) && "PHI reordering is free." ) ? void (0) : __assert_fail ("(E->ReorderIndices.empty() || E != VectorizableTree.front().get()) && \"PHI reordering is free.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 6907, __extension__ __PRETTY_FUNCTION__)) | |||
6906 | (E->ReorderIndices.empty() || E != VectorizableTree.front().get()) &&(static_cast <bool> ((E->ReorderIndices.empty() || E != VectorizableTree.front().get()) && "PHI reordering is free." ) ? void (0) : __assert_fail ("(E->ReorderIndices.empty() || E != VectorizableTree.front().get()) && \"PHI reordering is free.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 6907, __extension__ __PRETTY_FUNCTION__)) | |||
6907 | "PHI reordering is free.")(static_cast <bool> ((E->ReorderIndices.empty() || E != VectorizableTree.front().get()) && "PHI reordering is free." ) ? void (0) : __assert_fail ("(E->ReorderIndices.empty() || E != VectorizableTree.front().get()) && \"PHI reordering is free.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 6907, __extension__ __PRETTY_FUNCTION__)); | |||
6908 | auto *PH = cast<PHINode>(VL0); | |||
6909 | Builder.SetInsertPoint(PH->getParent()->getFirstNonPHI()); | |||
6910 | Builder.SetCurrentDebugLocation(PH->getDebugLoc()); | |||
6911 | PHINode *NewPhi = Builder.CreatePHI(VecTy, PH->getNumIncomingValues()); | |||
6912 | Value *V = NewPhi; | |||
6913 | ||||
6914 | // Adjust insertion point once all PHI's have been generated. | |||
6915 | Builder.SetInsertPoint(&*PH->getParent()->getFirstInsertionPt()); | |||
6916 | Builder.SetCurrentDebugLocation(PH->getDebugLoc()); | |||
6917 | ||||
6918 | ShuffleBuilder.addInversedMask(E->ReorderIndices); | |||
6919 | ShuffleBuilder.addMask(E->ReuseShuffleIndices); | |||
6920 | V = ShuffleBuilder.finalize(V); | |||
6921 | ||||
6922 | E->VectorizedValue = V; | |||
6923 | ||||
6924 | // PHINodes may have multiple entries from the same block. We want to | |||
6925 | // visit every block once. | |||
6926 | SmallPtrSet<BasicBlock*, 4> VisitedBBs; | |||
6927 | ||||
6928 | for (unsigned i = 0, e = PH->getNumIncomingValues(); i < e; ++i) { | |||
6929 | ValueList Operands; | |||
6930 | BasicBlock *IBB = PH->getIncomingBlock(i); | |||
6931 | ||||
6932 | if (!VisitedBBs.insert(IBB).second) { | |||
6933 | NewPhi->addIncoming(NewPhi->getIncomingValueForBlock(IBB), IBB); | |||
6934 | continue; | |||
6935 | } | |||
6936 | ||||
6937 | Builder.SetInsertPoint(IBB->getTerminator()); | |||
6938 | Builder.SetCurrentDebugLocation(PH->getDebugLoc()); | |||
6939 | Value *Vec = vectorizeTree(E->getOperand(i)); | |||
6940 | NewPhi->addIncoming(Vec, IBB); | |||
6941 | } | |||
6942 | ||||
6943 | assert(NewPhi->getNumIncomingValues() == PH->getNumIncomingValues() &&(static_cast <bool> (NewPhi->getNumIncomingValues() == PH->getNumIncomingValues() && "Invalid number of incoming values" ) ? void (0) : __assert_fail ("NewPhi->getNumIncomingValues() == PH->getNumIncomingValues() && \"Invalid number of incoming values\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 6944, __extension__ __PRETTY_FUNCTION__)) | |||
6944 | "Invalid number of incoming values")(static_cast <bool> (NewPhi->getNumIncomingValues() == PH->getNumIncomingValues() && "Invalid number of incoming values" ) ? void (0) : __assert_fail ("NewPhi->getNumIncomingValues() == PH->getNumIncomingValues() && \"Invalid number of incoming values\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 6944, __extension__ __PRETTY_FUNCTION__)); | |||
6945 | return V; | |||
6946 | } | |||
6947 | ||||
6948 | case Instruction::ExtractElement: { | |||
6949 | Value *V = E->getSingleOperand(0); | |||
6950 | Builder.SetInsertPoint(VL0); | |||
6951 | ShuffleBuilder.addInversedMask(E->ReorderIndices); | |||
6952 | ShuffleBuilder.addMask(E->ReuseShuffleIndices); | |||
6953 | V = ShuffleBuilder.finalize(V); | |||
6954 | E->VectorizedValue = V; | |||
6955 | return V; | |||
6956 | } | |||
6957 | case Instruction::ExtractValue: { | |||
6958 | auto *LI = cast<LoadInst>(E->getSingleOperand(0)); | |||
6959 | Builder.SetInsertPoint(LI); | |||
6960 | auto *PtrTy = PointerType::get(VecTy, LI->getPointerAddressSpace()); | |||
6961 | Value *Ptr = Builder.CreateBitCast(LI->getOperand(0), PtrTy); | |||
6962 | LoadInst *V = Builder.CreateAlignedLoad(VecTy, Ptr, LI->getAlign()); | |||
6963 | Value *NewV = propagateMetadata(V, E->Scalars); | |||
6964 | ShuffleBuilder.addInversedMask(E->ReorderIndices); | |||
6965 | ShuffleBuilder.addMask(E->ReuseShuffleIndices); | |||
6966 | NewV = ShuffleBuilder.finalize(NewV); | |||
6967 | E->VectorizedValue = NewV; | |||
6968 | return NewV; | |||
6969 | } | |||
6970 | case Instruction::InsertElement: { | |||
6971 | assert(E->ReuseShuffleIndices.empty() && "All inserts should be unique")(static_cast <bool> (E->ReuseShuffleIndices.empty() && "All inserts should be unique") ? void (0) : __assert_fail ( "E->ReuseShuffleIndices.empty() && \"All inserts should be unique\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 6971, __extension__ __PRETTY_FUNCTION__)); | |||
6972 | Builder.SetInsertPoint(cast<Instruction>(E->Scalars.back())); | |||
6973 | Value *V = vectorizeTree(E->getOperand(1)); | |||
6974 | ||||
6975 | // Create InsertVector shuffle if necessary | |||
6976 | auto *FirstInsert = cast<Instruction>(*find_if(E->Scalars, [E](Value *V) { | |||
6977 | return !is_contained(E->Scalars, cast<Instruction>(V)->getOperand(0)); | |||
6978 | })); | |||
6979 | const unsigned NumElts = | |||
6980 | cast<FixedVectorType>(FirstInsert->getType())->getNumElements(); | |||
6981 | const unsigned NumScalars = E->Scalars.size(); | |||
6982 | ||||
6983 | unsigned Offset = *getInsertIndex(VL0); | |||
6984 | assert(Offset < NumElts && "Failed to find vector index offset")(static_cast <bool> (Offset < NumElts && "Failed to find vector index offset" ) ? void (0) : __assert_fail ("Offset < NumElts && \"Failed to find vector index offset\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 6984, __extension__ __PRETTY_FUNCTION__)); | |||
6985 | ||||
6986 | // Create shuffle to resize vector | |||
6987 | SmallVector<int> Mask; | |||
6988 | if (!E->ReorderIndices.empty()) { | |||
6989 | inversePermutation(E->ReorderIndices, Mask); | |||
6990 | Mask.append(NumElts - NumScalars, UndefMaskElem); | |||
6991 | } else { | |||
6992 | Mask.assign(NumElts, UndefMaskElem); | |||
6993 | std::iota(Mask.begin(), std::next(Mask.begin(), NumScalars), 0); | |||
6994 | } | |||
6995 | // Create InsertVector shuffle if necessary | |||
6996 | bool IsIdentity = true; | |||
6997 | SmallVector<int> PrevMask(NumElts, UndefMaskElem); | |||
6998 | Mask.swap(PrevMask); | |||
6999 | for (unsigned I = 0; I < NumScalars; ++I) { | |||
7000 | Value *Scalar = E->Scalars[PrevMask[I]]; | |||
7001 | unsigned InsertIdx = *getInsertIndex(Scalar); | |||
7002 | IsIdentity &= InsertIdx - Offset == I; | |||
7003 | Mask[InsertIdx - Offset] = I; | |||
7004 | } | |||
7005 | if (!IsIdentity || NumElts != NumScalars) { | |||
7006 | V = Builder.CreateShuffleVector(V, Mask); | |||
7007 | if (auto *I = dyn_cast<Instruction>(V)) { | |||
7008 | GatherShuffleSeq.insert(I); | |||
7009 | CSEBlocks.insert(I->getParent()); | |||
7010 | } | |||
7011 | } | |||
7012 | ||||
7013 | if ((!IsIdentity || Offset != 0 || | |||
7014 | !isUndefVector(FirstInsert->getOperand(0))) && | |||
7015 | NumElts != NumScalars) { | |||
7016 | SmallVector<int> InsertMask(NumElts); | |||
7017 | std::iota(InsertMask.begin(), InsertMask.end(), 0); | |||
7018 | for (unsigned I = 0; I < NumElts; I++) { | |||
7019 | if (Mask[I] != UndefMaskElem) | |||
7020 | InsertMask[Offset + I] = NumElts + I; | |||
7021 | } | |||
7022 | ||||
7023 | V = Builder.CreateShuffleVector( | |||
7024 | FirstInsert->getOperand(0), V, InsertMask, | |||
7025 | cast<Instruction>(E->Scalars.back())->getName()); | |||
7026 | if (auto *I = dyn_cast<Instruction>(V)) { | |||
7027 | GatherShuffleSeq.insert(I); | |||
7028 | CSEBlocks.insert(I->getParent()); | |||
7029 | } | |||
7030 | } | |||
7031 | ||||
7032 | ++NumVectorInstructions; | |||
7033 | E->VectorizedValue = V; | |||
7034 | return V; | |||
7035 | } | |||
7036 | case Instruction::ZExt: | |||
7037 | case Instruction::SExt: | |||
7038 | case Instruction::FPToUI: | |||
7039 | case Instruction::FPToSI: | |||
7040 | case Instruction::FPExt: | |||
7041 | case Instruction::PtrToInt: | |||
7042 | case Instruction::IntToPtr: | |||
7043 | case Instruction::SIToFP: | |||
7044 | case Instruction::UIToFP: | |||
7045 | case Instruction::Trunc: | |||
7046 | case Instruction::FPTrunc: | |||
7047 | case Instruction::BitCast: { | |||
7048 | setInsertPointAfterBundle(E); | |||
7049 | ||||
7050 | Value *InVec = vectorizeTree(E->getOperand(0)); | |||
7051 | ||||
7052 | if (E->VectorizedValue) { | |||
7053 | LLVM_DEBUG(dbgs() << "SLP: Diamond merged for " << *VL0 << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Diamond merged for " << *VL0 << ".\n"; } } while (false); | |||
7054 | return E->VectorizedValue; | |||
7055 | } | |||
7056 | ||||
7057 | auto *CI = cast<CastInst>(VL0); | |||
7058 | Value *V = Builder.CreateCast(CI->getOpcode(), InVec, VecTy); | |||
7059 | ShuffleBuilder.addInversedMask(E->ReorderIndices); | |||
7060 | ShuffleBuilder.addMask(E->ReuseShuffleIndices); | |||
7061 | V = ShuffleBuilder.finalize(V); | |||
7062 | ||||
7063 | E->VectorizedValue = V; | |||
7064 | ++NumVectorInstructions; | |||
7065 | return V; | |||
7066 | } | |||
7067 | case Instruction::FCmp: | |||
7068 | case Instruction::ICmp: { | |||
7069 | setInsertPointAfterBundle(E); | |||
7070 | ||||
7071 | Value *L = vectorizeTree(E->getOperand(0)); | |||
7072 | Value *R = vectorizeTree(E->getOperand(1)); | |||
7073 | ||||
7074 | if (E->VectorizedValue) { | |||
7075 | LLVM_DEBUG(dbgs() << "SLP: Diamond merged for " << *VL0 << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Diamond merged for " << *VL0 << ".\n"; } } while (false); | |||
7076 | return E->VectorizedValue; | |||
7077 | } | |||
7078 | ||||
7079 | CmpInst::Predicate P0 = cast<CmpInst>(VL0)->getPredicate(); | |||
7080 | Value *V = Builder.CreateCmp(P0, L, R); | |||
7081 | propagateIRFlags(V, E->Scalars, VL0); | |||
7082 | ShuffleBuilder.addInversedMask(E->ReorderIndices); | |||
7083 | ShuffleBuilder.addMask(E->ReuseShuffleIndices); | |||
7084 | V = ShuffleBuilder.finalize(V); | |||
7085 | ||||
7086 | E->VectorizedValue = V; | |||
7087 | ++NumVectorInstructions; | |||
7088 | return V; | |||
7089 | } | |||
7090 | case Instruction::Select: { | |||
7091 | setInsertPointAfterBundle(E); | |||
7092 | ||||
7093 | Value *Cond = vectorizeTree(E->getOperand(0)); | |||
7094 | Value *True = vectorizeTree(E->getOperand(1)); | |||
7095 | Value *False = vectorizeTree(E->getOperand(2)); | |||
7096 | ||||
7097 | if (E->VectorizedValue) { | |||
7098 | LLVM_DEBUG(dbgs() << "SLP: Diamond merged for " << *VL0 << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Diamond merged for " << *VL0 << ".\n"; } } while (false); | |||
7099 | return E->VectorizedValue; | |||
7100 | } | |||
7101 | ||||
7102 | Value *V = Builder.CreateSelect(Cond, True, False); | |||
7103 | ShuffleBuilder.addInversedMask(E->ReorderIndices); | |||
7104 | ShuffleBuilder.addMask(E->ReuseShuffleIndices); | |||
7105 | V = ShuffleBuilder.finalize(V); | |||
7106 | ||||
7107 | E->VectorizedValue = V; | |||
7108 | ++NumVectorInstructions; | |||
7109 | return V; | |||
7110 | } | |||
7111 | case Instruction::FNeg: { | |||
7112 | setInsertPointAfterBundle(E); | |||
7113 | ||||
7114 | Value *Op = vectorizeTree(E->getOperand(0)); | |||
7115 | ||||
7116 | if (E->VectorizedValue) { | |||
7117 | LLVM_DEBUG(dbgs() << "SLP: Diamond merged for " << *VL0 << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Diamond merged for " << *VL0 << ".\n"; } } while (false); | |||
7118 | return E->VectorizedValue; | |||
7119 | } | |||
7120 | ||||
7121 | Value *V = Builder.CreateUnOp( | |||
7122 | static_cast<Instruction::UnaryOps>(E->getOpcode()), Op); | |||
7123 | propagateIRFlags(V, E->Scalars, VL0); | |||
7124 | if (auto *I = dyn_cast<Instruction>(V)) | |||
7125 | V = propagateMetadata(I, E->Scalars); | |||
7126 | ||||
7127 | ShuffleBuilder.addInversedMask(E->ReorderIndices); | |||
7128 | ShuffleBuilder.addMask(E->ReuseShuffleIndices); | |||
7129 | V = ShuffleBuilder.finalize(V); | |||
7130 | ||||
7131 | E->VectorizedValue = V; | |||
7132 | ++NumVectorInstructions; | |||
7133 | ||||
7134 | return V; | |||
7135 | } | |||
7136 | case Instruction::Add: | |||
7137 | case Instruction::FAdd: | |||
7138 | case Instruction::Sub: | |||
7139 | case Instruction::FSub: | |||
7140 | case Instruction::Mul: | |||
7141 | case Instruction::FMul: | |||
7142 | case Instruction::UDiv: | |||
7143 | case Instruction::SDiv: | |||
7144 | case Instruction::FDiv: | |||
7145 | case Instruction::URem: | |||
7146 | case Instruction::SRem: | |||
7147 | case Instruction::FRem: | |||
7148 | case Instruction::Shl: | |||
7149 | case Instruction::LShr: | |||
7150 | case Instruction::AShr: | |||
7151 | case Instruction::And: | |||
7152 | case Instruction::Or: | |||
7153 | case Instruction::Xor: { | |||
7154 | setInsertPointAfterBundle(E); | |||
7155 | ||||
7156 | Value *LHS = vectorizeTree(E->getOperand(0)); | |||
7157 | Value *RHS = vectorizeTree(E->getOperand(1)); | |||
7158 | ||||
7159 | if (E->VectorizedValue) { | |||
7160 | LLVM_DEBUG(dbgs() << "SLP: Diamond merged for " << *VL0 << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Diamond merged for " << *VL0 << ".\n"; } } while (false); | |||
7161 | return E->VectorizedValue; | |||
7162 | } | |||
7163 | ||||
7164 | Value *V = Builder.CreateBinOp( | |||
7165 | static_cast<Instruction::BinaryOps>(E->getOpcode()), LHS, | |||
7166 | RHS); | |||
7167 | propagateIRFlags(V, E->Scalars, VL0); | |||
7168 | if (auto *I = dyn_cast<Instruction>(V)) | |||
7169 | V = propagateMetadata(I, E->Scalars); | |||
7170 | ||||
7171 | ShuffleBuilder.addInversedMask(E->ReorderIndices); | |||
7172 | ShuffleBuilder.addMask(E->ReuseShuffleIndices); | |||
7173 | V = ShuffleBuilder.finalize(V); | |||
7174 | ||||
7175 | E->VectorizedValue = V; | |||
7176 | ++NumVectorInstructions; | |||
7177 | ||||
7178 | return V; | |||
7179 | } | |||
7180 | case Instruction::Load: { | |||
7181 | // Loads are inserted at the head of the tree because we don't want to | |||
7182 | // sink them all the way down past store instructions. | |||
7183 | setInsertPointAfterBundle(E); | |||
7184 | ||||
7185 | LoadInst *LI = cast<LoadInst>(VL0); | |||
7186 | Instruction *NewLI; | |||
7187 | unsigned AS = LI->getPointerAddressSpace(); | |||
7188 | Value *PO = LI->getPointerOperand(); | |||
7189 | if (E->State == TreeEntry::Vectorize) { | |||
7190 | Value *VecPtr = Builder.CreateBitCast(PO, VecTy->getPointerTo(AS)); | |||
7191 | NewLI = Builder.CreateAlignedLoad(VecTy, VecPtr, LI->getAlign()); | |||
7192 | ||||
7193 | // The pointer operand uses an in-tree scalar so we add the new BitCast | |||
7194 | // or LoadInst to ExternalUses list to make sure that an extract will | |||
7195 | // be generated in the future. | |||
7196 | if (TreeEntry *Entry = getTreeEntry(PO)) { | |||
7197 | // Find which lane we need to extract. | |||
7198 | unsigned FoundLane = Entry->findLaneForValue(PO); | |||
7199 | ExternalUses.emplace_back( | |||
7200 | PO, PO != VecPtr ? cast<User>(VecPtr) : NewLI, FoundLane); | |||
7201 | } | |||
7202 | } else { | |||
7203 | assert(E->State == TreeEntry::ScatterVectorize && "Unhandled state")(static_cast <bool> (E->State == TreeEntry::ScatterVectorize && "Unhandled state") ? void (0) : __assert_fail ("E->State == TreeEntry::ScatterVectorize && \"Unhandled state\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7203, __extension__ __PRETTY_FUNCTION__)); | |||
7204 | Value *VecPtr = vectorizeTree(E->getOperand(0)); | |||
7205 | // Use the minimum alignment of the gathered loads. | |||
7206 | Align CommonAlignment = LI->getAlign(); | |||
7207 | for (Value *V : E->Scalars) | |||
7208 | CommonAlignment = | |||
7209 | commonAlignment(CommonAlignment, cast<LoadInst>(V)->getAlign()); | |||
7210 | NewLI = Builder.CreateMaskedGather(VecTy, VecPtr, CommonAlignment); | |||
7211 | } | |||
7212 | Value *V = propagateMetadata(NewLI, E->Scalars); | |||
7213 | ||||
7214 | ShuffleBuilder.addInversedMask(E->ReorderIndices); | |||
7215 | ShuffleBuilder.addMask(E->ReuseShuffleIndices); | |||
7216 | V = ShuffleBuilder.finalize(V); | |||
7217 | E->VectorizedValue = V; | |||
7218 | ++NumVectorInstructions; | |||
7219 | return V; | |||
7220 | } | |||
7221 | case Instruction::Store: { | |||
7222 | auto *SI = cast<StoreInst>(VL0); | |||
7223 | unsigned AS = SI->getPointerAddressSpace(); | |||
7224 | ||||
7225 | setInsertPointAfterBundle(E); | |||
7226 | ||||
7227 | Value *VecValue = vectorizeTree(E->getOperand(0)); | |||
7228 | ShuffleBuilder.addMask(E->ReorderIndices); | |||
7229 | VecValue = ShuffleBuilder.finalize(VecValue); | |||
7230 | ||||
7231 | Value *ScalarPtr = SI->getPointerOperand(); | |||
7232 | Value *VecPtr = Builder.CreateBitCast( | |||
7233 | ScalarPtr, VecValue->getType()->getPointerTo(AS)); | |||
7234 | StoreInst *ST = | |||
7235 | Builder.CreateAlignedStore(VecValue, VecPtr, SI->getAlign()); | |||
7236 | ||||
7237 | // The pointer operand uses an in-tree scalar, so add the new BitCast or | |||
7238 | // StoreInst to ExternalUses to make sure that an extract will be | |||
7239 | // generated in the future. | |||
7240 | if (TreeEntry *Entry = getTreeEntry(ScalarPtr)) { | |||
7241 | // Find which lane we need to extract. | |||
7242 | unsigned FoundLane = Entry->findLaneForValue(ScalarPtr); | |||
7243 | ExternalUses.push_back(ExternalUser( | |||
7244 | ScalarPtr, ScalarPtr != VecPtr ? cast<User>(VecPtr) : ST, | |||
7245 | FoundLane)); | |||
7246 | } | |||
7247 | ||||
7248 | Value *V = propagateMetadata(ST, E->Scalars); | |||
7249 | ||||
7250 | E->VectorizedValue = V; | |||
7251 | ++NumVectorInstructions; | |||
7252 | return V; | |||
7253 | } | |||
7254 | case Instruction::GetElementPtr: { | |||
7255 | auto *GEP0 = cast<GetElementPtrInst>(VL0); | |||
7256 | setInsertPointAfterBundle(E); | |||
7257 | ||||
7258 | Value *Op0 = vectorizeTree(E->getOperand(0)); | |||
7259 | ||||
7260 | SmallVector<Value *> OpVecs; | |||
7261 | for (int J = 1, N = GEP0->getNumOperands(); J < N; ++J) { | |||
7262 | Value *OpVec = vectorizeTree(E->getOperand(J)); | |||
7263 | OpVecs.push_back(OpVec); | |||
7264 | } | |||
7265 | ||||
7266 | Value *V = Builder.CreateGEP(GEP0->getSourceElementType(), Op0, OpVecs); | |||
7267 | if (Instruction *I = dyn_cast<Instruction>(V)) | |||
7268 | V = propagateMetadata(I, E->Scalars); | |||
7269 | ||||
7270 | ShuffleBuilder.addInversedMask(E->ReorderIndices); | |||
7271 | ShuffleBuilder.addMask(E->ReuseShuffleIndices); | |||
7272 | V = ShuffleBuilder.finalize(V); | |||
7273 | ||||
7274 | E->VectorizedValue = V; | |||
7275 | ++NumVectorInstructions; | |||
7276 | ||||
7277 | return V; | |||
7278 | } | |||
7279 | case Instruction::Call: { | |||
7280 | CallInst *CI = cast<CallInst>(VL0); | |||
7281 | setInsertPointAfterBundle(E); | |||
7282 | ||||
7283 | Intrinsic::ID IID = Intrinsic::not_intrinsic; | |||
7284 | if (Function *FI = CI->getCalledFunction()) | |||
7285 | IID = FI->getIntrinsicID(); | |||
7286 | ||||
7287 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); | |||
7288 | ||||
7289 | auto VecCallCosts = getVectorCallCosts(CI, VecTy, TTI, TLI); | |||
7290 | bool UseIntrinsic = ID != Intrinsic::not_intrinsic && | |||
7291 | VecCallCosts.first <= VecCallCosts.second; | |||
7292 | ||||
7293 | Value *ScalarArg = nullptr; | |||
7294 | std::vector<Value *> OpVecs; | |||
7295 | SmallVector<Type *, 2> TysForDecl = | |||
7296 | {FixedVectorType::get(CI->getType(), E->Scalars.size())}; | |||
7297 | for (int j = 0, e = CI->arg_size(); j < e; ++j) { | |||
7298 | ValueList OpVL; | |||
7299 | // Some intrinsics have scalar arguments. This argument should not be | |||
7300 | // vectorized. | |||
7301 | if (UseIntrinsic && hasVectorInstrinsicScalarOpd(IID, j)) { | |||
7302 | CallInst *CEI = cast<CallInst>(VL0); | |||
7303 | ScalarArg = CEI->getArgOperand(j); | |||
7304 | OpVecs.push_back(CEI->getArgOperand(j)); | |||
7305 | if (hasVectorInstrinsicOverloadedScalarOpd(IID, j)) | |||
7306 | TysForDecl.push_back(ScalarArg->getType()); | |||
7307 | continue; | |||
7308 | } | |||
7309 | ||||
7310 | Value *OpVec = vectorizeTree(E->getOperand(j)); | |||
7311 | LLVM_DEBUG(dbgs() << "SLP: OpVec[" << j << "]: " << *OpVec << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: OpVec[" << j << "]: " << *OpVec << "\n"; } } while (false); | |||
7312 | OpVecs.push_back(OpVec); | |||
7313 | } | |||
7314 | ||||
7315 | Function *CF; | |||
7316 | if (!UseIntrinsic) { | |||
7317 | VFShape Shape = | |||
7318 | VFShape::get(*CI, ElementCount::getFixed(static_cast<unsigned>( | |||
7319 | VecTy->getNumElements())), | |||
7320 | false /*HasGlobalPred*/); | |||
7321 | CF = VFDatabase(*CI).getVectorizedFunction(Shape); | |||
7322 | } else { | |||
7323 | CF = Intrinsic::getDeclaration(F->getParent(), ID, TysForDecl); | |||
7324 | } | |||
7325 | ||||
7326 | SmallVector<OperandBundleDef, 1> OpBundles; | |||
7327 | CI->getOperandBundlesAsDefs(OpBundles); | |||
7328 | Value *V = Builder.CreateCall(CF, OpVecs, OpBundles); | |||
7329 | ||||
7330 | // The scalar argument uses an in-tree scalar so we add the new vectorized | |||
7331 | // call to ExternalUses list to make sure that an extract will be | |||
7332 | // generated in the future. | |||
7333 | if (ScalarArg) { | |||
7334 | if (TreeEntry *Entry = getTreeEntry(ScalarArg)) { | |||
7335 | // Find which lane we need to extract. | |||
7336 | unsigned FoundLane = Entry->findLaneForValue(ScalarArg); | |||
7337 | ExternalUses.push_back( | |||
7338 | ExternalUser(ScalarArg, cast<User>(V), FoundLane)); | |||
7339 | } | |||
7340 | } | |||
7341 | ||||
7342 | propagateIRFlags(V, E->Scalars, VL0); | |||
7343 | ShuffleBuilder.addInversedMask(E->ReorderIndices); | |||
7344 | ShuffleBuilder.addMask(E->ReuseShuffleIndices); | |||
7345 | V = ShuffleBuilder.finalize(V); | |||
7346 | ||||
7347 | E->VectorizedValue = V; | |||
7348 | ++NumVectorInstructions; | |||
7349 | return V; | |||
7350 | } | |||
7351 | case Instruction::ShuffleVector: { | |||
7352 | assert(E->isAltShuffle() &&(static_cast <bool> (E->isAltShuffle() && (( Instruction::isBinaryOp(E->getOpcode()) && Instruction ::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E ->getOpcode()) && Instruction::isCast(E->getAltOpcode ())) || (isa<CmpInst>(VL0) && isa<CmpInst> (E->getAltOp()))) && "Invalid Shuffle Vector Operand" ) ? void (0) : __assert_fail ("E->isAltShuffle() && ((Instruction::isBinaryOp(E->getOpcode()) && Instruction::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E->getOpcode()) && Instruction::isCast(E->getAltOpcode())) || (isa<CmpInst>(VL0) && isa<CmpInst>(E->getAltOp()))) && \"Invalid Shuffle Vector Operand\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7358, __extension__ __PRETTY_FUNCTION__)) | |||
7353 | ((Instruction::isBinaryOp(E->getOpcode()) &&(static_cast <bool> (E->isAltShuffle() && (( Instruction::isBinaryOp(E->getOpcode()) && Instruction ::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E ->getOpcode()) && Instruction::isCast(E->getAltOpcode ())) || (isa<CmpInst>(VL0) && isa<CmpInst> (E->getAltOp()))) && "Invalid Shuffle Vector Operand" ) ? void (0) : __assert_fail ("E->isAltShuffle() && ((Instruction::isBinaryOp(E->getOpcode()) && Instruction::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E->getOpcode()) && Instruction::isCast(E->getAltOpcode())) || (isa<CmpInst>(VL0) && isa<CmpInst>(E->getAltOp()))) && \"Invalid Shuffle Vector Operand\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7358, __extension__ __PRETTY_FUNCTION__)) | |||
7354 | Instruction::isBinaryOp(E->getAltOpcode())) ||(static_cast <bool> (E->isAltShuffle() && (( Instruction::isBinaryOp(E->getOpcode()) && Instruction ::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E ->getOpcode()) && Instruction::isCast(E->getAltOpcode ())) || (isa<CmpInst>(VL0) && isa<CmpInst> (E->getAltOp()))) && "Invalid Shuffle Vector Operand" ) ? void (0) : __assert_fail ("E->isAltShuffle() && ((Instruction::isBinaryOp(E->getOpcode()) && Instruction::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E->getOpcode()) && Instruction::isCast(E->getAltOpcode())) || (isa<CmpInst>(VL0) && isa<CmpInst>(E->getAltOp()))) && \"Invalid Shuffle Vector Operand\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7358, __extension__ __PRETTY_FUNCTION__)) | |||
7355 | (Instruction::isCast(E->getOpcode()) &&(static_cast <bool> (E->isAltShuffle() && (( Instruction::isBinaryOp(E->getOpcode()) && Instruction ::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E ->getOpcode()) && Instruction::isCast(E->getAltOpcode ())) || (isa<CmpInst>(VL0) && isa<CmpInst> (E->getAltOp()))) && "Invalid Shuffle Vector Operand" ) ? void (0) : __assert_fail ("E->isAltShuffle() && ((Instruction::isBinaryOp(E->getOpcode()) && Instruction::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E->getOpcode()) && Instruction::isCast(E->getAltOpcode())) || (isa<CmpInst>(VL0) && isa<CmpInst>(E->getAltOp()))) && \"Invalid Shuffle Vector Operand\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7358, __extension__ __PRETTY_FUNCTION__)) | |||
7356 | Instruction::isCast(E->getAltOpcode())) ||(static_cast <bool> (E->isAltShuffle() && (( Instruction::isBinaryOp(E->getOpcode()) && Instruction ::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E ->getOpcode()) && Instruction::isCast(E->getAltOpcode ())) || (isa<CmpInst>(VL0) && isa<CmpInst> (E->getAltOp()))) && "Invalid Shuffle Vector Operand" ) ? void (0) : __assert_fail ("E->isAltShuffle() && ((Instruction::isBinaryOp(E->getOpcode()) && Instruction::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E->getOpcode()) && Instruction::isCast(E->getAltOpcode())) || (isa<CmpInst>(VL0) && isa<CmpInst>(E->getAltOp()))) && \"Invalid Shuffle Vector Operand\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7358, __extension__ __PRETTY_FUNCTION__)) | |||
7357 | (isa<CmpInst>(VL0) && isa<CmpInst>(E->getAltOp()))) &&(static_cast <bool> (E->isAltShuffle() && (( Instruction::isBinaryOp(E->getOpcode()) && Instruction ::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E ->getOpcode()) && Instruction::isCast(E->getAltOpcode ())) || (isa<CmpInst>(VL0) && isa<CmpInst> (E->getAltOp()))) && "Invalid Shuffle Vector Operand" ) ? void (0) : __assert_fail ("E->isAltShuffle() && ((Instruction::isBinaryOp(E->getOpcode()) && Instruction::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E->getOpcode()) && Instruction::isCast(E->getAltOpcode())) || (isa<CmpInst>(VL0) && isa<CmpInst>(E->getAltOp()))) && \"Invalid Shuffle Vector Operand\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7358, __extension__ __PRETTY_FUNCTION__)) | |||
7358 | "Invalid Shuffle Vector Operand")(static_cast <bool> (E->isAltShuffle() && (( Instruction::isBinaryOp(E->getOpcode()) && Instruction ::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E ->getOpcode()) && Instruction::isCast(E->getAltOpcode ())) || (isa<CmpInst>(VL0) && isa<CmpInst> (E->getAltOp()))) && "Invalid Shuffle Vector Operand" ) ? void (0) : __assert_fail ("E->isAltShuffle() && ((Instruction::isBinaryOp(E->getOpcode()) && Instruction::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E->getOpcode()) && Instruction::isCast(E->getAltOpcode())) || (isa<CmpInst>(VL0) && isa<CmpInst>(E->getAltOp()))) && \"Invalid Shuffle Vector Operand\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7358, __extension__ __PRETTY_FUNCTION__)); | |||
7359 | ||||
7360 | Value *LHS = nullptr, *RHS = nullptr; | |||
7361 | if (Instruction::isBinaryOp(E->getOpcode()) || isa<CmpInst>(VL0)) { | |||
7362 | setInsertPointAfterBundle(E); | |||
7363 | LHS = vectorizeTree(E->getOperand(0)); | |||
7364 | RHS = vectorizeTree(E->getOperand(1)); | |||
7365 | } else { | |||
7366 | setInsertPointAfterBundle(E); | |||
7367 | LHS = vectorizeTree(E->getOperand(0)); | |||
7368 | } | |||
7369 | ||||
7370 | if (E->VectorizedValue) { | |||
7371 | LLVM_DEBUG(dbgs() << "SLP: Diamond merged for " << *VL0 << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Diamond merged for " << *VL0 << ".\n"; } } while (false); | |||
7372 | return E->VectorizedValue; | |||
7373 | } | |||
7374 | ||||
7375 | Value *V0, *V1; | |||
7376 | if (Instruction::isBinaryOp(E->getOpcode())) { | |||
7377 | V0 = Builder.CreateBinOp( | |||
7378 | static_cast<Instruction::BinaryOps>(E->getOpcode()), LHS, RHS); | |||
7379 | V1 = Builder.CreateBinOp( | |||
7380 | static_cast<Instruction::BinaryOps>(E->getAltOpcode()), LHS, RHS); | |||
7381 | } else if (auto *CI0 = dyn_cast<CmpInst>(VL0)) { | |||
7382 | V0 = Builder.CreateCmp(CI0->getPredicate(), LHS, RHS); | |||
7383 | auto *AltCI = cast<CmpInst>(E->getAltOp()); | |||
7384 | CmpInst::Predicate AltPred = AltCI->getPredicate(); | |||
7385 | V1 = Builder.CreateCmp(AltPred, LHS, RHS); | |||
7386 | } else { | |||
7387 | V0 = Builder.CreateCast( | |||
7388 | static_cast<Instruction::CastOps>(E->getOpcode()), LHS, VecTy); | |||
7389 | V1 = Builder.CreateCast( | |||
7390 | static_cast<Instruction::CastOps>(E->getAltOpcode()), LHS, VecTy); | |||
7391 | } | |||
7392 | // Add V0 and V1 to later analysis to try to find and remove matching | |||
7393 | // instruction, if any. | |||
7394 | for (Value *V : {V0, V1}) { | |||
7395 | if (auto *I = dyn_cast<Instruction>(V)) { | |||
7396 | GatherShuffleSeq.insert(I); | |||
7397 | CSEBlocks.insert(I->getParent()); | |||
7398 | } | |||
7399 | } | |||
7400 | ||||
7401 | // Create shuffle to take alternate operations from the vector. | |||
7402 | // Also, gather up main and alt scalar ops to propagate IR flags to | |||
7403 | // each vector operation. | |||
7404 | ValueList OpScalars, AltScalars; | |||
7405 | SmallVector<int> Mask; | |||
7406 | buildShuffleEntryMask( | |||
7407 | E->Scalars, E->ReorderIndices, E->ReuseShuffleIndices, | |||
7408 | [E](Instruction *I) { | |||
7409 | assert(E->isOpcodeOrAlt(I) && "Unexpected main/alternate opcode")(static_cast <bool> (E->isOpcodeOrAlt(I) && "Unexpected main/alternate opcode" ) ? void (0) : __assert_fail ("E->isOpcodeOrAlt(I) && \"Unexpected main/alternate opcode\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7409, __extension__ __PRETTY_FUNCTION__)); | |||
7410 | return isAlternateInstruction(I, E->getMainOp(), E->getAltOp()); | |||
7411 | }, | |||
7412 | Mask, &OpScalars, &AltScalars); | |||
7413 | ||||
7414 | propagateIRFlags(V0, OpScalars); | |||
7415 | propagateIRFlags(V1, AltScalars); | |||
7416 | ||||
7417 | Value *V = Builder.CreateShuffleVector(V0, V1, Mask); | |||
7418 | if (auto *I = dyn_cast<Instruction>(V)) { | |||
7419 | V = propagateMetadata(I, E->Scalars); | |||
7420 | GatherShuffleSeq.insert(I); | |||
7421 | CSEBlocks.insert(I->getParent()); | |||
7422 | } | |||
7423 | V = ShuffleBuilder.finalize(V); | |||
7424 | ||||
7425 | E->VectorizedValue = V; | |||
7426 | ++NumVectorInstructions; | |||
7427 | ||||
7428 | return V; | |||
7429 | } | |||
7430 | default: | |||
7431 | llvm_unreachable("unknown inst")::llvm::llvm_unreachable_internal("unknown inst", "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 7431); | |||
7432 | } | |||
7433 | return nullptr; | |||
7434 | } | |||
7435 | ||||
7436 | Value *BoUpSLP::vectorizeTree() { | |||
7437 | ExtraValueToDebugLocsMap ExternallyUsedValues; | |||
7438 | return vectorizeTree(ExternallyUsedValues); | |||
7439 | } | |||
7440 | ||||
7441 | Value * | |||
7442 | BoUpSLP::vectorizeTree(ExtraValueToDebugLocsMap &ExternallyUsedValues) { | |||
7443 | // All blocks must be scheduled before any instructions are inserted. | |||
7444 | for (auto &BSIter : BlocksSchedules) { | |||
7445 | scheduleBlock(BSIter.second.get()); | |||
| ||||
7446 | } | |||
7447 | ||||
7448 | Builder.SetInsertPoint(&F->getEntryBlock().front()); | |||
7449 | auto *VectorRoot = vectorizeTree(VectorizableTree[0].get()); | |||
7450 | ||||
7451 | // If the vectorized tree can be rewritten in a smaller type, we truncate the | |||
7452 | // vectorized root. InstCombine will then rewrite the entire expression. We | |||
7453 | // sign extend the extracted values below. | |||
7454 | auto *ScalarRoot = VectorizableTree[0]->Scalars[0]; | |||
7455 | if (MinBWs.count(ScalarRoot)) { | |||
7456 | if (auto *I = dyn_cast<Instruction>(VectorRoot)) { | |||
7457 | // If current instr is a phi and not the last phi, insert it after the | |||
7458 | // last phi node. | |||
7459 | if (isa<PHINode>(I)) | |||
7460 | Builder.SetInsertPoint(&*I->getParent()->getFirstInsertionPt()); | |||
7461 | else | |||
7462 | Builder.SetInsertPoint(&*++BasicBlock::iterator(I)); | |||
7463 | } | |||
7464 | auto BundleWidth = VectorizableTree[0]->Scalars.size(); | |||
7465 | auto *MinTy = IntegerType::get(F->getContext(), MinBWs[ScalarRoot].first); | |||
7466 | auto *VecTy = FixedVectorType::get(MinTy, BundleWidth); | |||
7467 | auto *Trunc = Builder.CreateTrunc(VectorRoot, VecTy); | |||
7468 | VectorizableTree[0]->VectorizedValue = Trunc; | |||
7469 | } | |||
7470 | ||||
7471 | LLVM_DEBUG(dbgs() << "SLP: Extracting " << ExternalUses.size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Extracting " << ExternalUses .size() << " values .\n"; } } while (false) | |||
7472 | << " values .\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Extracting " << ExternalUses .size() << " values .\n"; } } while (false); | |||
7473 | ||||
7474 | // Extract all of the elements with the external uses. | |||
7475 | for (const auto &ExternalUse : ExternalUses) { | |||
7476 | Value *Scalar = ExternalUse.Scalar; | |||
7477 | llvm::User *User = ExternalUse.User; | |||
7478 | ||||
7479 | // Skip users that we already RAUW. This happens when one instruction | |||
7480 | // has multiple uses of the same value. | |||
7481 | if (User && !is_contained(Scalar->users(), User)) | |||
7482 | continue; | |||
7483 | TreeEntry *E = getTreeEntry(Scalar); | |||
7484 | assert(E && "Invalid scalar")(static_cast <bool> (E && "Invalid scalar") ? void (0) : __assert_fail ("E && \"Invalid scalar\"", "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 7484, __extension__ __PRETTY_FUNCTION__)); | |||
7485 | assert(E->State != TreeEntry::NeedToGather &&(static_cast <bool> (E->State != TreeEntry::NeedToGather && "Extracting from a gather list") ? void (0) : __assert_fail ("E->State != TreeEntry::NeedToGather && \"Extracting from a gather list\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7486, __extension__ __PRETTY_FUNCTION__)) | |||
7486 | "Extracting from a gather list")(static_cast <bool> (E->State != TreeEntry::NeedToGather && "Extracting from a gather list") ? void (0) : __assert_fail ("E->State != TreeEntry::NeedToGather && \"Extracting from a gather list\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7486, __extension__ __PRETTY_FUNCTION__)); | |||
7487 | ||||
7488 | Value *Vec = E->VectorizedValue; | |||
7489 | assert(Vec && "Can't find vectorizable value")(static_cast <bool> (Vec && "Can't find vectorizable value" ) ? void (0) : __assert_fail ("Vec && \"Can't find vectorizable value\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7489, __extension__ __PRETTY_FUNCTION__)); | |||
7490 | ||||
7491 | Value *Lane = Builder.getInt32(ExternalUse.Lane); | |||
7492 | auto ExtractAndExtendIfNeeded = [&](Value *Vec) { | |||
7493 | if (Scalar->getType() != Vec->getType()) { | |||
7494 | Value *Ex; | |||
7495 | // "Reuse" the existing extract to improve final codegen. | |||
7496 | if (auto *ES = dyn_cast<ExtractElementInst>(Scalar)) { | |||
7497 | Ex = Builder.CreateExtractElement(ES->getOperand(0), | |||
7498 | ES->getOperand(1)); | |||
7499 | } else { | |||
7500 | Ex = Builder.CreateExtractElement(Vec, Lane); | |||
7501 | } | |||
7502 | // If necessary, sign-extend or zero-extend ScalarRoot | |||
7503 | // to the larger type. | |||
7504 | if (!MinBWs.count(ScalarRoot)) | |||
7505 | return Ex; | |||
7506 | if (MinBWs[ScalarRoot].second) | |||
7507 | return Builder.CreateSExt(Ex, Scalar->getType()); | |||
7508 | return Builder.CreateZExt(Ex, Scalar->getType()); | |||
7509 | } | |||
7510 | assert(isa<FixedVectorType>(Scalar->getType()) &&(static_cast <bool> (isa<FixedVectorType>(Scalar-> getType()) && isa<InsertElementInst>(Scalar) && "In-tree scalar of vector type is not insertelement?") ? void (0) : __assert_fail ("isa<FixedVectorType>(Scalar->getType()) && isa<InsertElementInst>(Scalar) && \"In-tree scalar of vector type is not insertelement?\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7512, __extension__ __PRETTY_FUNCTION__)) | |||
7511 | isa<InsertElementInst>(Scalar) &&(static_cast <bool> (isa<FixedVectorType>(Scalar-> getType()) && isa<InsertElementInst>(Scalar) && "In-tree scalar of vector type is not insertelement?") ? void (0) : __assert_fail ("isa<FixedVectorType>(Scalar->getType()) && isa<InsertElementInst>(Scalar) && \"In-tree scalar of vector type is not insertelement?\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7512, __extension__ __PRETTY_FUNCTION__)) | |||
7512 | "In-tree scalar of vector type is not insertelement?")(static_cast <bool> (isa<FixedVectorType>(Scalar-> getType()) && isa<InsertElementInst>(Scalar) && "In-tree scalar of vector type is not insertelement?") ? void (0) : __assert_fail ("isa<FixedVectorType>(Scalar->getType()) && isa<InsertElementInst>(Scalar) && \"In-tree scalar of vector type is not insertelement?\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7512, __extension__ __PRETTY_FUNCTION__)); | |||
7513 | return Vec; | |||
7514 | }; | |||
7515 | // If User == nullptr, the Scalar is used as extra arg. Generate | |||
7516 | // ExtractElement instruction and update the record for this scalar in | |||
7517 | // ExternallyUsedValues. | |||
7518 | if (!User) { | |||
7519 | assert(ExternallyUsedValues.count(Scalar) &&(static_cast <bool> (ExternallyUsedValues.count(Scalar) && "Scalar with nullptr as an external user must be registered in " "ExternallyUsedValues map") ? void (0) : __assert_fail ("ExternallyUsedValues.count(Scalar) && \"Scalar with nullptr as an external user must be registered in \" \"ExternallyUsedValues map\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7521, __extension__ __PRETTY_FUNCTION__)) | |||
7520 | "Scalar with nullptr as an external user must be registered in "(static_cast <bool> (ExternallyUsedValues.count(Scalar) && "Scalar with nullptr as an external user must be registered in " "ExternallyUsedValues map") ? void (0) : __assert_fail ("ExternallyUsedValues.count(Scalar) && \"Scalar with nullptr as an external user must be registered in \" \"ExternallyUsedValues map\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7521, __extension__ __PRETTY_FUNCTION__)) | |||
7521 | "ExternallyUsedValues map")(static_cast <bool> (ExternallyUsedValues.count(Scalar) && "Scalar with nullptr as an external user must be registered in " "ExternallyUsedValues map") ? void (0) : __assert_fail ("ExternallyUsedValues.count(Scalar) && \"Scalar with nullptr as an external user must be registered in \" \"ExternallyUsedValues map\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7521, __extension__ __PRETTY_FUNCTION__)); | |||
7522 | if (auto *VecI = dyn_cast<Instruction>(Vec)) { | |||
7523 | Builder.SetInsertPoint(VecI->getParent(), | |||
7524 | std::next(VecI->getIterator())); | |||
7525 | } else { | |||
7526 | Builder.SetInsertPoint(&F->getEntryBlock().front()); | |||
7527 | } | |||
7528 | Value *NewInst = ExtractAndExtendIfNeeded(Vec); | |||
7529 | CSEBlocks.insert(cast<Instruction>(Scalar)->getParent()); | |||
7530 | auto &NewInstLocs = ExternallyUsedValues[NewInst]; | |||
7531 | auto It = ExternallyUsedValues.find(Scalar); | |||
7532 | assert(It != ExternallyUsedValues.end() &&(static_cast <bool> (It != ExternallyUsedValues.end() && "Externally used scalar is not found in ExternallyUsedValues" ) ? void (0) : __assert_fail ("It != ExternallyUsedValues.end() && \"Externally used scalar is not found in ExternallyUsedValues\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7533, __extension__ __PRETTY_FUNCTION__)) | |||
7533 | "Externally used scalar is not found in ExternallyUsedValues")(static_cast <bool> (It != ExternallyUsedValues.end() && "Externally used scalar is not found in ExternallyUsedValues" ) ? void (0) : __assert_fail ("It != ExternallyUsedValues.end() && \"Externally used scalar is not found in ExternallyUsedValues\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7533, __extension__ __PRETTY_FUNCTION__)); | |||
7534 | NewInstLocs.append(It->second); | |||
7535 | ExternallyUsedValues.erase(Scalar); | |||
7536 | // Required to update internally referenced instructions. | |||
7537 | Scalar->replaceAllUsesWith(NewInst); | |||
7538 | continue; | |||
7539 | } | |||
7540 | ||||
7541 | // Generate extracts for out-of-tree users. | |||
7542 | // Find the insertion point for the extractelement lane. | |||
7543 | if (auto *VecI = dyn_cast<Instruction>(Vec)) { | |||
7544 | if (PHINode *PH = dyn_cast<PHINode>(User)) { | |||
7545 | for (int i = 0, e = PH->getNumIncomingValues(); i != e; ++i) { | |||
7546 | if (PH->getIncomingValue(i) == Scalar) { | |||
7547 | Instruction *IncomingTerminator = | |||
7548 | PH->getIncomingBlock(i)->getTerminator(); | |||
7549 | if (isa<CatchSwitchInst>(IncomingTerminator)) { | |||
7550 | Builder.SetInsertPoint(VecI->getParent(), | |||
7551 | std::next(VecI->getIterator())); | |||
7552 | } else { | |||
7553 | Builder.SetInsertPoint(PH->getIncomingBlock(i)->getTerminator()); | |||
7554 | } | |||
7555 | Value *NewInst = ExtractAndExtendIfNeeded(Vec); | |||
7556 | CSEBlocks.insert(PH->getIncomingBlock(i)); | |||
7557 | PH->setOperand(i, NewInst); | |||
7558 | } | |||
7559 | } | |||
7560 | } else { | |||
7561 | Builder.SetInsertPoint(cast<Instruction>(User)); | |||
7562 | Value *NewInst = ExtractAndExtendIfNeeded(Vec); | |||
7563 | CSEBlocks.insert(cast<Instruction>(User)->getParent()); | |||
7564 | User->replaceUsesOfWith(Scalar, NewInst); | |||
7565 | } | |||
7566 | } else { | |||
7567 | Builder.SetInsertPoint(&F->getEntryBlock().front()); | |||
7568 | Value *NewInst = ExtractAndExtendIfNeeded(Vec); | |||
7569 | CSEBlocks.insert(&F->getEntryBlock()); | |||
7570 | User->replaceUsesOfWith(Scalar, NewInst); | |||
7571 | } | |||
7572 | ||||
7573 | LLVM_DEBUG(dbgs() << "SLP: Replaced:" << *User << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Replaced:" << *User << ".\n"; } } while (false); | |||
7574 | } | |||
7575 | ||||
7576 | // For each vectorized value: | |||
7577 | for (auto &TEPtr : VectorizableTree) { | |||
7578 | TreeEntry *Entry = TEPtr.get(); | |||
7579 | ||||
7580 | // No need to handle users of gathered values. | |||
7581 | if (Entry->State == TreeEntry::NeedToGather) | |||
7582 | continue; | |||
7583 | ||||
7584 | assert(Entry->VectorizedValue && "Can't find vectorizable value")(static_cast <bool> (Entry->VectorizedValue && "Can't find vectorizable value") ? void (0) : __assert_fail ( "Entry->VectorizedValue && \"Can't find vectorizable value\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7584, __extension__ __PRETTY_FUNCTION__)); | |||
7585 | ||||
7586 | // For each lane: | |||
7587 | for (int Lane = 0, LE = Entry->Scalars.size(); Lane != LE; ++Lane) { | |||
7588 | Value *Scalar = Entry->Scalars[Lane]; | |||
7589 | ||||
7590 | #ifndef NDEBUG | |||
7591 | Type *Ty = Scalar->getType(); | |||
7592 | if (!Ty->isVoidTy()) { | |||
7593 | for (User *U : Scalar->users()) { | |||
7594 | LLVM_DEBUG(dbgs() << "SLP: \tvalidating user:" << *U << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: \tvalidating user:" << *U << ".\n"; } } while (false); | |||
7595 | ||||
7596 | // It is legal to delete users in the ignorelist. | |||
7597 | assert((getTreeEntry(U) || is_contained(UserIgnoreList, U) ||(static_cast <bool> ((getTreeEntry(U) || is_contained(UserIgnoreList , U) || (isa_and_nonnull<Instruction>(U) && isDeleted (cast<Instruction>(U)))) && "Deleting out-of-tree value" ) ? void (0) : __assert_fail ("(getTreeEntry(U) || is_contained(UserIgnoreList, U) || (isa_and_nonnull<Instruction>(U) && isDeleted(cast<Instruction>(U)))) && \"Deleting out-of-tree value\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7600, __extension__ __PRETTY_FUNCTION__)) | |||
7598 | (isa_and_nonnull<Instruction>(U) &&(static_cast <bool> ((getTreeEntry(U) || is_contained(UserIgnoreList , U) || (isa_and_nonnull<Instruction>(U) && isDeleted (cast<Instruction>(U)))) && "Deleting out-of-tree value" ) ? void (0) : __assert_fail ("(getTreeEntry(U) || is_contained(UserIgnoreList, U) || (isa_and_nonnull<Instruction>(U) && isDeleted(cast<Instruction>(U)))) && \"Deleting out-of-tree value\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7600, __extension__ __PRETTY_FUNCTION__)) | |||
7599 | isDeleted(cast<Instruction>(U)))) &&(static_cast <bool> ((getTreeEntry(U) || is_contained(UserIgnoreList , U) || (isa_and_nonnull<Instruction>(U) && isDeleted (cast<Instruction>(U)))) && "Deleting out-of-tree value" ) ? void (0) : __assert_fail ("(getTreeEntry(U) || is_contained(UserIgnoreList, U) || (isa_and_nonnull<Instruction>(U) && isDeleted(cast<Instruction>(U)))) && \"Deleting out-of-tree value\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7600, __extension__ __PRETTY_FUNCTION__)) | |||
7600 | "Deleting out-of-tree value")(static_cast <bool> ((getTreeEntry(U) || is_contained(UserIgnoreList , U) || (isa_and_nonnull<Instruction>(U) && isDeleted (cast<Instruction>(U)))) && "Deleting out-of-tree value" ) ? void (0) : __assert_fail ("(getTreeEntry(U) || is_contained(UserIgnoreList, U) || (isa_and_nonnull<Instruction>(U) && isDeleted(cast<Instruction>(U)))) && \"Deleting out-of-tree value\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7600, __extension__ __PRETTY_FUNCTION__)); | |||
7601 | } | |||
7602 | } | |||
7603 | #endif | |||
7604 | LLVM_DEBUG(dbgs() << "SLP: \tErasing scalar:" << *Scalar << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: \tErasing scalar:" << * Scalar << ".\n"; } } while (false); | |||
7605 | eraseInstruction(cast<Instruction>(Scalar)); | |||
7606 | } | |||
7607 | } | |||
7608 | ||||
7609 | Builder.ClearInsertionPoint(); | |||
7610 | InstrElementSize.clear(); | |||
7611 | ||||
7612 | return VectorizableTree[0]->VectorizedValue; | |||
7613 | } | |||
7614 | ||||
7615 | void BoUpSLP::optimizeGatherSequence() { | |||
7616 | LLVM_DEBUG(dbgs() << "SLP: Optimizing " << GatherShuffleSeq.size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Optimizing " << GatherShuffleSeq .size() << " gather sequences instructions.\n"; } } while (false) | |||
7617 | << " gather sequences instructions.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Optimizing " << GatherShuffleSeq .size() << " gather sequences instructions.\n"; } } while (false); | |||
7618 | // LICM InsertElementInst sequences. | |||
7619 | for (Instruction *I : GatherShuffleSeq) { | |||
7620 | if (isDeleted(I)) | |||
7621 | continue; | |||
7622 | ||||
7623 | // Check if this block is inside a loop. | |||
7624 | Loop *L = LI->getLoopFor(I->getParent()); | |||
7625 | if (!L) | |||
7626 | continue; | |||
7627 | ||||
7628 | // Check if it has a preheader. | |||
7629 | BasicBlock *PreHeader = L->getLoopPreheader(); | |||
7630 | if (!PreHeader) | |||
7631 | continue; | |||
7632 | ||||
7633 | // If the vector or the element that we insert into it are | |||
7634 | // instructions that are defined in this basic block then we can't | |||
7635 | // hoist this instruction. | |||
7636 | if (any_of(I->operands(), [L](Value *V) { | |||
7637 | auto *OpI = dyn_cast<Instruction>(V); | |||
7638 | return OpI && L->contains(OpI); | |||
7639 | })) | |||
7640 | continue; | |||
7641 | ||||
7642 | // We can hoist this instruction. Move it to the pre-header. | |||
7643 | I->moveBefore(PreHeader->getTerminator()); | |||
7644 | } | |||
7645 | ||||
7646 | // Make a list of all reachable blocks in our CSE queue. | |||
7647 | SmallVector<const DomTreeNode *, 8> CSEWorkList; | |||
7648 | CSEWorkList.reserve(CSEBlocks.size()); | |||
7649 | for (BasicBlock *BB : CSEBlocks) | |||
7650 | if (DomTreeNode *N = DT->getNode(BB)) { | |||
7651 | assert(DT->isReachableFromEntry(N))(static_cast <bool> (DT->isReachableFromEntry(N)) ? void (0) : __assert_fail ("DT->isReachableFromEntry(N)", "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 7651, __extension__ __PRETTY_FUNCTION__)); | |||
7652 | CSEWorkList.push_back(N); | |||
7653 | } | |||
7654 | ||||
7655 | // Sort blocks by domination. This ensures we visit a block after all blocks | |||
7656 | // dominating it are visited. | |||
7657 | llvm::sort(CSEWorkList, [](const DomTreeNode *A, const DomTreeNode *B) { | |||
7658 | assert((A == B) == (A->getDFSNumIn() == B->getDFSNumIn()) &&(static_cast <bool> ((A == B) == (A->getDFSNumIn() == B->getDFSNumIn()) && "Different nodes should have different DFS numbers" ) ? void (0) : __assert_fail ("(A == B) == (A->getDFSNumIn() == B->getDFSNumIn()) && \"Different nodes should have different DFS numbers\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7659, __extension__ __PRETTY_FUNCTION__)) | |||
7659 | "Different nodes should have different DFS numbers")(static_cast <bool> ((A == B) == (A->getDFSNumIn() == B->getDFSNumIn()) && "Different nodes should have different DFS numbers" ) ? void (0) : __assert_fail ("(A == B) == (A->getDFSNumIn() == B->getDFSNumIn()) && \"Different nodes should have different DFS numbers\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7659, __extension__ __PRETTY_FUNCTION__)); | |||
7660 | return A->getDFSNumIn() < B->getDFSNumIn(); | |||
7661 | }); | |||
7662 | ||||
7663 | // Less defined shuffles can be replaced by the more defined copies. | |||
7664 | // Between two shuffles one is less defined if it has the same vector operands | |||
7665 | // and its mask indeces are the same as in the first one or undefs. E.g. | |||
7666 | // shuffle %0, poison, <0, 0, 0, undef> is less defined than shuffle %0, | |||
7667 | // poison, <0, 0, 0, 0>. | |||
7668 | auto &&IsIdenticalOrLessDefined = [this](Instruction *I1, Instruction *I2, | |||
7669 | SmallVectorImpl<int> &NewMask) { | |||
7670 | if (I1->getType() != I2->getType()) | |||
7671 | return false; | |||
7672 | auto *SI1 = dyn_cast<ShuffleVectorInst>(I1); | |||
7673 | auto *SI2 = dyn_cast<ShuffleVectorInst>(I2); | |||
7674 | if (!SI1 || !SI2) | |||
7675 | return I1->isIdenticalTo(I2); | |||
7676 | if (SI1->isIdenticalTo(SI2)) | |||
7677 | return true; | |||
7678 | for (int I = 0, E = SI1->getNumOperands(); I < E; ++I) | |||
7679 | if (SI1->getOperand(I) != SI2->getOperand(I)) | |||
7680 | return false; | |||
7681 | // Check if the second instruction is more defined than the first one. | |||
7682 | NewMask.assign(SI2->getShuffleMask().begin(), SI2->getShuffleMask().end()); | |||
7683 | ArrayRef<int> SM1 = SI1->getShuffleMask(); | |||
7684 | // Count trailing undefs in the mask to check the final number of used | |||
7685 | // registers. | |||
7686 | unsigned LastUndefsCnt = 0; | |||
7687 | for (int I = 0, E = NewMask.size(); I < E; ++I) { | |||
7688 | if (SM1[I] == UndefMaskElem) | |||
7689 | ++LastUndefsCnt; | |||
7690 | else | |||
7691 | LastUndefsCnt = 0; | |||
7692 | if (NewMask[I] != UndefMaskElem && SM1[I] != UndefMaskElem && | |||
7693 | NewMask[I] != SM1[I]) | |||
7694 | return false; | |||
7695 | if (NewMask[I] == UndefMaskElem) | |||
7696 | NewMask[I] = SM1[I]; | |||
7697 | } | |||
7698 | // Check if the last undefs actually change the final number of used vector | |||
7699 | // registers. | |||
7700 | return SM1.size() - LastUndefsCnt > 1 && | |||
7701 | TTI->getNumberOfParts(SI1->getType()) == | |||
7702 | TTI->getNumberOfParts( | |||
7703 | FixedVectorType::get(SI1->getType()->getElementType(), | |||
7704 | SM1.size() - LastUndefsCnt)); | |||
7705 | }; | |||
7706 | // Perform O(N^2) search over the gather/shuffle sequences and merge identical | |||
7707 | // instructions. TODO: We can further optimize this scan if we split the | |||
7708 | // instructions into different buckets based on the insert lane. | |||
7709 | SmallVector<Instruction *, 16> Visited; | |||
7710 | for (auto I = CSEWorkList.begin(), E = CSEWorkList.end(); I != E; ++I) { | |||
7711 | assert(*I &&(static_cast <bool> (*I && (I == CSEWorkList.begin () || !DT->dominates(*I, *std::prev(I))) && "Worklist not sorted properly!" ) ? void (0) : __assert_fail ("*I && (I == CSEWorkList.begin() || !DT->dominates(*I, *std::prev(I))) && \"Worklist not sorted properly!\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7713, __extension__ __PRETTY_FUNCTION__)) | |||
7712 | (I == CSEWorkList.begin() || !DT->dominates(*I, *std::prev(I))) &&(static_cast <bool> (*I && (I == CSEWorkList.begin () || !DT->dominates(*I, *std::prev(I))) && "Worklist not sorted properly!" ) ? void (0) : __assert_fail ("*I && (I == CSEWorkList.begin() || !DT->dominates(*I, *std::prev(I))) && \"Worklist not sorted properly!\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7713, __extension__ __PRETTY_FUNCTION__)) | |||
7713 | "Worklist not sorted properly!")(static_cast <bool> (*I && (I == CSEWorkList.begin () || !DT->dominates(*I, *std::prev(I))) && "Worklist not sorted properly!" ) ? void (0) : __assert_fail ("*I && (I == CSEWorkList.begin() || !DT->dominates(*I, *std::prev(I))) && \"Worklist not sorted properly!\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7713, __extension__ __PRETTY_FUNCTION__)); | |||
7714 | BasicBlock *BB = (*I)->getBlock(); | |||
7715 | // For all instructions in blocks containing gather sequences: | |||
7716 | for (Instruction &In : llvm::make_early_inc_range(*BB)) { | |||
7717 | if (isDeleted(&In)) | |||
7718 | continue; | |||
7719 | if (!isa<InsertElementInst>(&In) && !isa<ExtractElementInst>(&In) && | |||
7720 | !isa<ShuffleVectorInst>(&In) && !GatherShuffleSeq.contains(&In)) | |||
7721 | continue; | |||
7722 | ||||
7723 | // Check if we can replace this instruction with any of the | |||
7724 | // visited instructions. | |||
7725 | bool Replaced = false; | |||
7726 | for (Instruction *&V : Visited) { | |||
7727 | SmallVector<int> NewMask; | |||
7728 | if (IsIdenticalOrLessDefined(&In, V, NewMask) && | |||
7729 | DT->dominates(V->getParent(), In.getParent())) { | |||
7730 | In.replaceAllUsesWith(V); | |||
7731 | eraseInstruction(&In); | |||
7732 | if (auto *SI = dyn_cast<ShuffleVectorInst>(V)) | |||
7733 | if (!NewMask.empty()) | |||
7734 | SI->setShuffleMask(NewMask); | |||
7735 | Replaced = true; | |||
7736 | break; | |||
7737 | } | |||
7738 | if (isa<ShuffleVectorInst>(In) && isa<ShuffleVectorInst>(V) && | |||
7739 | GatherShuffleSeq.contains(V) && | |||
7740 | IsIdenticalOrLessDefined(V, &In, NewMask) && | |||
7741 | DT->dominates(In.getParent(), V->getParent())) { | |||
7742 | In.moveAfter(V); | |||
7743 | V->replaceAllUsesWith(&In); | |||
7744 | eraseInstruction(V); | |||
7745 | if (auto *SI = dyn_cast<ShuffleVectorInst>(&In)) | |||
7746 | if (!NewMask.empty()) | |||
7747 | SI->setShuffleMask(NewMask); | |||
7748 | V = &In; | |||
7749 | Replaced = true; | |||
7750 | break; | |||
7751 | } | |||
7752 | } | |||
7753 | if (!Replaced) { | |||
7754 | assert(!is_contained(Visited, &In))(static_cast <bool> (!is_contained(Visited, &In)) ? void (0) : __assert_fail ("!is_contained(Visited, &In)", "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7754, __extension__ __PRETTY_FUNCTION__)); | |||
7755 | Visited.push_back(&In); | |||
7756 | } | |||
7757 | } | |||
7758 | } | |||
7759 | CSEBlocks.clear(); | |||
7760 | GatherShuffleSeq.clear(); | |||
7761 | } | |||
7762 | ||||
7763 | BoUpSLP::ScheduleData * | |||
7764 | BoUpSLP::BlockScheduling::buildBundle(ArrayRef<Value *> VL) { | |||
7765 | ScheduleData *Bundle = nullptr; | |||
7766 | ScheduleData *PrevInBundle = nullptr; | |||
7767 | for (Value *V : VL) { | |||
7768 | if (doesNotNeedToBeScheduled(V)) | |||
7769 | continue; | |||
7770 | ScheduleData *BundleMember = getScheduleData(V); | |||
7771 | assert(BundleMember &&(static_cast <bool> (BundleMember && "no ScheduleData for bundle member " "(maybe not in same basic block)") ? void (0) : __assert_fail ("BundleMember && \"no ScheduleData for bundle member \" \"(maybe not in same basic block)\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7773, __extension__ __PRETTY_FUNCTION__)) | |||
7772 | "no ScheduleData for bundle member "(static_cast <bool> (BundleMember && "no ScheduleData for bundle member " "(maybe not in same basic block)") ? void (0) : __assert_fail ("BundleMember && \"no ScheduleData for bundle member \" \"(maybe not in same basic block)\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7773, __extension__ __PRETTY_FUNCTION__)) | |||
7773 | "(maybe not in same basic block)")(static_cast <bool> (BundleMember && "no ScheduleData for bundle member " "(maybe not in same basic block)") ? void (0) : __assert_fail ("BundleMember && \"no ScheduleData for bundle member \" \"(maybe not in same basic block)\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7773, __extension__ __PRETTY_FUNCTION__)); | |||
7774 | assert(BundleMember->isSchedulingEntity() &&(static_cast <bool> (BundleMember->isSchedulingEntity () && "bundle member already part of other bundle") ? void (0) : __assert_fail ("BundleMember->isSchedulingEntity() && \"bundle member already part of other bundle\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7775, __extension__ __PRETTY_FUNCTION__)) | |||
7775 | "bundle member already part of other bundle")(static_cast <bool> (BundleMember->isSchedulingEntity () && "bundle member already part of other bundle") ? void (0) : __assert_fail ("BundleMember->isSchedulingEntity() && \"bundle member already part of other bundle\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7775, __extension__ __PRETTY_FUNCTION__)); | |||
7776 | if (PrevInBundle) { | |||
7777 | PrevInBundle->NextInBundle = BundleMember; | |||
7778 | } else { | |||
7779 | Bundle = BundleMember; | |||
7780 | } | |||
7781 | ||||
7782 | // Group the instructions to a bundle. | |||
7783 | BundleMember->FirstInBundle = Bundle; | |||
7784 | PrevInBundle = BundleMember; | |||
7785 | } | |||
7786 | assert(Bundle && "Failed to find schedule bundle")(static_cast <bool> (Bundle && "Failed to find schedule bundle" ) ? void (0) : __assert_fail ("Bundle && \"Failed to find schedule bundle\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7786, __extension__ __PRETTY_FUNCTION__)); | |||
7787 | return Bundle; | |||
7788 | } | |||
7789 | ||||
7790 | // Groups the instructions to a bundle (which is then a single scheduling entity) | |||
7791 | // and schedules instructions until the bundle gets ready. | |||
7792 | Optional<BoUpSLP::ScheduleData *> | |||
7793 | BoUpSLP::BlockScheduling::tryScheduleBundle(ArrayRef<Value *> VL, BoUpSLP *SLP, | |||
7794 | const InstructionsState &S) { | |||
7795 | // No need to schedule PHIs, insertelement, extractelement and extractvalue | |||
7796 | // instructions. | |||
7797 | if (isa<PHINode>(S.OpValue) || isVectorLikeInstWithConstOps(S.OpValue) || | |||
7798 | doesNotNeedToSchedule(VL)) | |||
7799 | return nullptr; | |||
7800 | ||||
7801 | // Initialize the instruction bundle. | |||
7802 | Instruction *OldScheduleEnd = ScheduleEnd; | |||
7803 | LLVM_DEBUG(dbgs() << "SLP: bundle: " << *S.OpValue << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: bundle: " << *S.OpValue << "\n"; } } while (false); | |||
7804 | ||||
7805 | auto TryScheduleBundleImpl = [this, OldScheduleEnd, SLP](bool ReSchedule, | |||
7806 | ScheduleData *Bundle) { | |||
7807 | // The scheduling region got new instructions at the lower end (or it is a | |||
7808 | // new region for the first bundle). This makes it necessary to | |||
7809 | // recalculate all dependencies. | |||
7810 | // It is seldom that this needs to be done a second time after adding the | |||
7811 | // initial bundle to the region. | |||
7812 | if (ScheduleEnd != OldScheduleEnd) { | |||
7813 | for (auto *I = ScheduleStart; I != ScheduleEnd; I = I->getNextNode()) | |||
7814 | doForAllOpcodes(I, [](ScheduleData *SD) { SD->clearDependencies(); }); | |||
7815 | ReSchedule = true; | |||
7816 | } | |||
7817 | if (Bundle) { | |||
7818 | LLVM_DEBUG(dbgs() << "SLP: try schedule bundle " << *Bundledo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: try schedule bundle " << *Bundle << " in block " << BB->getName() << "\n"; } } while (false) | |||
7819 | << " in block " << BB->getName() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: try schedule bundle " << *Bundle << " in block " << BB->getName() << "\n"; } } while (false); | |||
7820 | calculateDependencies(Bundle, /*InsertInReadyList=*/true, SLP); | |||
7821 | } | |||
7822 | ||||
7823 | if (ReSchedule) { | |||
7824 | resetSchedule(); | |||
7825 | initialFillReadyList(ReadyInsts); | |||
7826 | } | |||
7827 | ||||
7828 | // Now try to schedule the new bundle or (if no bundle) just calculate | |||
7829 | // dependencies. As soon as the bundle is "ready" it means that there are no | |||
7830 | // cyclic dependencies and we can schedule it. Note that's important that we | |||
7831 | // don't "schedule" the bundle yet (see cancelScheduling). | |||
7832 | while (((!Bundle && ReSchedule) || (Bundle && !Bundle->isReady())) && | |||
7833 | !ReadyInsts.empty()) { | |||
7834 | ScheduleData *Picked = ReadyInsts.pop_back_val(); | |||
7835 | assert(Picked->isSchedulingEntity() && Picked->isReady() &&(static_cast <bool> (Picked->isSchedulingEntity() && Picked->isReady() && "must be ready to schedule") ? void (0) : __assert_fail ("Picked->isSchedulingEntity() && Picked->isReady() && \"must be ready to schedule\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7836, __extension__ __PRETTY_FUNCTION__)) | |||
7836 | "must be ready to schedule")(static_cast <bool> (Picked->isSchedulingEntity() && Picked->isReady() && "must be ready to schedule") ? void (0) : __assert_fail ("Picked->isSchedulingEntity() && Picked->isReady() && \"must be ready to schedule\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7836, __extension__ __PRETTY_FUNCTION__)); | |||
7837 | schedule(Picked, ReadyInsts); | |||
7838 | } | |||
7839 | }; | |||
7840 | ||||
7841 | // Make sure that the scheduling region contains all | |||
7842 | // instructions of the bundle. | |||
7843 | for (Value *V : VL) { | |||
7844 | if (doesNotNeedToBeScheduled(V)) | |||
7845 | continue; | |||
7846 | if (!extendSchedulingRegion(V, S)) { | |||
7847 | // If the scheduling region got new instructions at the lower end (or it | |||
7848 | // is a new region for the first bundle). This makes it necessary to | |||
7849 | // recalculate all dependencies. | |||
7850 | // Otherwise the compiler may crash trying to incorrectly calculate | |||
7851 | // dependencies and emit instruction in the wrong order at the actual | |||
7852 | // scheduling. | |||
7853 | TryScheduleBundleImpl(/*ReSchedule=*/false, nullptr); | |||
7854 | return None; | |||
7855 | } | |||
7856 | } | |||
7857 | ||||
7858 | bool ReSchedule = false; | |||
7859 | for (Value *V : VL) { | |||
7860 | if (doesNotNeedToBeScheduled(V)) | |||
7861 | continue; | |||
7862 | ScheduleData *BundleMember = getScheduleData(V); | |||
7863 | assert(BundleMember &&(static_cast <bool> (BundleMember && "no ScheduleData for bundle member (maybe not in same basic block)" ) ? void (0) : __assert_fail ("BundleMember && \"no ScheduleData for bundle member (maybe not in same basic block)\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7864, __extension__ __PRETTY_FUNCTION__)) | |||
7864 | "no ScheduleData for bundle member (maybe not in same basic block)")(static_cast <bool> (BundleMember && "no ScheduleData for bundle member (maybe not in same basic block)" ) ? void (0) : __assert_fail ("BundleMember && \"no ScheduleData for bundle member (maybe not in same basic block)\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7864, __extension__ __PRETTY_FUNCTION__)); | |||
7865 | ||||
7866 | // Make sure we don't leave the pieces of the bundle in the ready list when | |||
7867 | // whole bundle might not be ready. | |||
7868 | ReadyInsts.remove(BundleMember); | |||
7869 | ||||
7870 | if (!BundleMember->IsScheduled) | |||
7871 | continue; | |||
7872 | // A bundle member was scheduled as single instruction before and now | |||
7873 | // needs to be scheduled as part of the bundle. We just get rid of the | |||
7874 | // existing schedule. | |||
7875 | LLVM_DEBUG(dbgs() << "SLP: reset schedule because " << *BundleMemberdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: reset schedule because " << *BundleMember << " was already scheduled\n"; } } while (false) | |||
7876 | << " was already scheduled\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: reset schedule because " << *BundleMember << " was already scheduled\n"; } } while (false); | |||
7877 | ReSchedule = true; | |||
7878 | } | |||
7879 | ||||
7880 | auto *Bundle = buildBundle(VL); | |||
7881 | TryScheduleBundleImpl(ReSchedule, Bundle); | |||
7882 | if (!Bundle->isReady()) { | |||
7883 | cancelScheduling(VL, S.OpValue); | |||
7884 | return None; | |||
7885 | } | |||
7886 | return Bundle; | |||
7887 | } | |||
7888 | ||||
7889 | void BoUpSLP::BlockScheduling::cancelScheduling(ArrayRef<Value *> VL, | |||
7890 | Value *OpValue) { | |||
7891 | if (isa<PHINode>(OpValue) || isVectorLikeInstWithConstOps(OpValue) || | |||
7892 | doesNotNeedToSchedule(VL)) | |||
7893 | return; | |||
7894 | ||||
7895 | if (doesNotNeedToBeScheduled(OpValue)) | |||
7896 | OpValue = *find_if_not(VL, doesNotNeedToBeScheduled); | |||
7897 | ScheduleData *Bundle = getScheduleData(OpValue); | |||
7898 | LLVM_DEBUG(dbgs() << "SLP: cancel scheduling of " << *Bundle << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: cancel scheduling of " << *Bundle << "\n"; } } while (false); | |||
7899 | assert(!Bundle->IsScheduled &&(static_cast <bool> (!Bundle->IsScheduled && "Can't cancel bundle which is already scheduled") ? void (0) : __assert_fail ("!Bundle->IsScheduled && \"Can't cancel bundle which is already scheduled\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7900, __extension__ __PRETTY_FUNCTION__)) | |||
7900 | "Can't cancel bundle which is already scheduled")(static_cast <bool> (!Bundle->IsScheduled && "Can't cancel bundle which is already scheduled") ? void (0) : __assert_fail ("!Bundle->IsScheduled && \"Can't cancel bundle which is already scheduled\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7900, __extension__ __PRETTY_FUNCTION__)); | |||
7901 | assert(Bundle->isSchedulingEntity() &&(static_cast <bool> (Bundle->isSchedulingEntity() && (Bundle->isPartOfBundle() || needToScheduleSingleInstruction (VL)) && "tried to unbundle something which is not a bundle" ) ? void (0) : __assert_fail ("Bundle->isSchedulingEntity() && (Bundle->isPartOfBundle() || needToScheduleSingleInstruction(VL)) && \"tried to unbundle something which is not a bundle\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7903, __extension__ __PRETTY_FUNCTION__)) | |||
7902 | (Bundle->isPartOfBundle() || needToScheduleSingleInstruction(VL)) &&(static_cast <bool> (Bundle->isSchedulingEntity() && (Bundle->isPartOfBundle() || needToScheduleSingleInstruction (VL)) && "tried to unbundle something which is not a bundle" ) ? void (0) : __assert_fail ("Bundle->isSchedulingEntity() && (Bundle->isPartOfBundle() || needToScheduleSingleInstruction(VL)) && \"tried to unbundle something which is not a bundle\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7903, __extension__ __PRETTY_FUNCTION__)) | |||
7903 | "tried to unbundle something which is not a bundle")(static_cast <bool> (Bundle->isSchedulingEntity() && (Bundle->isPartOfBundle() || needToScheduleSingleInstruction (VL)) && "tried to unbundle something which is not a bundle" ) ? void (0) : __assert_fail ("Bundle->isSchedulingEntity() && (Bundle->isPartOfBundle() || needToScheduleSingleInstruction(VL)) && \"tried to unbundle something which is not a bundle\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7903, __extension__ __PRETTY_FUNCTION__)); | |||
7904 | ||||
7905 | // Remove the bundle from the ready list. | |||
7906 | if (Bundle->isReady()) | |||
7907 | ReadyInsts.remove(Bundle); | |||
7908 | ||||
7909 | // Un-bundle: make single instructions out of the bundle. | |||
7910 | ScheduleData *BundleMember = Bundle; | |||
7911 | while (BundleMember) { | |||
7912 | assert(BundleMember->FirstInBundle == Bundle && "corrupt bundle links")(static_cast <bool> (BundleMember->FirstInBundle == Bundle && "corrupt bundle links") ? void (0) : __assert_fail ("BundleMember->FirstInBundle == Bundle && \"corrupt bundle links\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7912, __extension__ __PRETTY_FUNCTION__)); | |||
7913 | BundleMember->FirstInBundle = BundleMember; | |||
7914 | ScheduleData *Next = BundleMember->NextInBundle; | |||
7915 | BundleMember->NextInBundle = nullptr; | |||
7916 | BundleMember->TE = nullptr; | |||
7917 | if (BundleMember->unscheduledDepsInBundle() == 0) { | |||
7918 | ReadyInsts.insert(BundleMember); | |||
7919 | } | |||
7920 | BundleMember = Next; | |||
7921 | } | |||
7922 | } | |||
7923 | ||||
7924 | BoUpSLP::ScheduleData *BoUpSLP::BlockScheduling::allocateScheduleDataChunks() { | |||
7925 | // Allocate a new ScheduleData for the instruction. | |||
7926 | if (ChunkPos >= ChunkSize) { | |||
7927 | ScheduleDataChunks.push_back(std::make_unique<ScheduleData[]>(ChunkSize)); | |||
7928 | ChunkPos = 0; | |||
7929 | } | |||
7930 | return &(ScheduleDataChunks.back()[ChunkPos++]); | |||
7931 | } | |||
7932 | ||||
7933 | bool BoUpSLP::BlockScheduling::extendSchedulingRegion(Value *V, | |||
7934 | const InstructionsState &S) { | |||
7935 | if (getScheduleData(V, isOneOf(S, V))) | |||
7936 | return true; | |||
7937 | Instruction *I = dyn_cast<Instruction>(V); | |||
7938 | assert(I && "bundle member must be an instruction")(static_cast <bool> (I && "bundle member must be an instruction" ) ? void (0) : __assert_fail ("I && \"bundle member must be an instruction\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7938, __extension__ __PRETTY_FUNCTION__)); | |||
7939 | assert(!isa<PHINode>(I) && !isVectorLikeInstWithConstOps(I) &&(static_cast <bool> (!isa<PHINode>(I) && ! isVectorLikeInstWithConstOps(I) && !doesNotNeedToBeScheduled (I) && "phi nodes/insertelements/extractelements/extractvalues don't need to " "be scheduled") ? void (0) : __assert_fail ("!isa<PHINode>(I) && !isVectorLikeInstWithConstOps(I) && !doesNotNeedToBeScheduled(I) && \"phi nodes/insertelements/extractelements/extractvalues don't need to \" \"be scheduled\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7942, __extension__ __PRETTY_FUNCTION__)) | |||
7940 | !doesNotNeedToBeScheduled(I) &&(static_cast <bool> (!isa<PHINode>(I) && ! isVectorLikeInstWithConstOps(I) && !doesNotNeedToBeScheduled (I) && "phi nodes/insertelements/extractelements/extractvalues don't need to " "be scheduled") ? void (0) : __assert_fail ("!isa<PHINode>(I) && !isVectorLikeInstWithConstOps(I) && !doesNotNeedToBeScheduled(I) && \"phi nodes/insertelements/extractelements/extractvalues don't need to \" \"be scheduled\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7942, __extension__ __PRETTY_FUNCTION__)) | |||
7941 | "phi nodes/insertelements/extractelements/extractvalues don't need to "(static_cast <bool> (!isa<PHINode>(I) && ! isVectorLikeInstWithConstOps(I) && !doesNotNeedToBeScheduled (I) && "phi nodes/insertelements/extractelements/extractvalues don't need to " "be scheduled") ? void (0) : __assert_fail ("!isa<PHINode>(I) && !isVectorLikeInstWithConstOps(I) && !doesNotNeedToBeScheduled(I) && \"phi nodes/insertelements/extractelements/extractvalues don't need to \" \"be scheduled\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7942, __extension__ __PRETTY_FUNCTION__)) | |||
7942 | "be scheduled")(static_cast <bool> (!isa<PHINode>(I) && ! isVectorLikeInstWithConstOps(I) && !doesNotNeedToBeScheduled (I) && "phi nodes/insertelements/extractelements/extractvalues don't need to " "be scheduled") ? void (0) : __assert_fail ("!isa<PHINode>(I) && !isVectorLikeInstWithConstOps(I) && !doesNotNeedToBeScheduled(I) && \"phi nodes/insertelements/extractelements/extractvalues don't need to \" \"be scheduled\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7942, __extension__ __PRETTY_FUNCTION__)); | |||
7943 | auto &&CheckScheduleForI = [this, &S](Instruction *I) -> bool { | |||
7944 | ScheduleData *ISD = getScheduleData(I); | |||
7945 | if (!ISD) | |||
7946 | return false; | |||
7947 | assert(isInSchedulingRegion(ISD) &&(static_cast <bool> (isInSchedulingRegion(ISD) && "ScheduleData not in scheduling region") ? void (0) : __assert_fail ("isInSchedulingRegion(ISD) && \"ScheduleData not in scheduling region\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7948, __extension__ __PRETTY_FUNCTION__)) | |||
7948 | "ScheduleData not in scheduling region")(static_cast <bool> (isInSchedulingRegion(ISD) && "ScheduleData not in scheduling region") ? void (0) : __assert_fail ("isInSchedulingRegion(ISD) && \"ScheduleData not in scheduling region\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7948, __extension__ __PRETTY_FUNCTION__)); | |||
7949 | ScheduleData *SD = allocateScheduleDataChunks(); | |||
7950 | SD->Inst = I; | |||
7951 | SD->init(SchedulingRegionID, S.OpValue); | |||
7952 | ExtraScheduleDataMap[I][S.OpValue] = SD; | |||
7953 | return true; | |||
7954 | }; | |||
7955 | if (CheckScheduleForI(I)) | |||
7956 | return true; | |||
7957 | if (!ScheduleStart) { | |||
7958 | // It's the first instruction in the new region. | |||
7959 | initScheduleData(I, I->getNextNode(), nullptr, nullptr); | |||
7960 | ScheduleStart = I; | |||
7961 | ScheduleEnd = I->getNextNode(); | |||
7962 | if (isOneOf(S, I) != I) | |||
7963 | CheckScheduleForI(I); | |||
7964 | assert(ScheduleEnd && "tried to vectorize a terminator?")(static_cast <bool> (ScheduleEnd && "tried to vectorize a terminator?" ) ? void (0) : __assert_fail ("ScheduleEnd && \"tried to vectorize a terminator?\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7964, __extension__ __PRETTY_FUNCTION__)); | |||
7965 | LLVM_DEBUG(dbgs() << "SLP: initialize schedule region to " << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: initialize schedule region to " << *I << "\n"; } } while (false); | |||
7966 | return true; | |||
7967 | } | |||
7968 | // Search up and down at the same time, because we don't know if the new | |||
7969 | // instruction is above or below the existing scheduling region. | |||
7970 | BasicBlock::reverse_iterator UpIter = | |||
7971 | ++ScheduleStart->getIterator().getReverse(); | |||
7972 | BasicBlock::reverse_iterator UpperEnd = BB->rend(); | |||
7973 | BasicBlock::iterator DownIter = ScheduleEnd->getIterator(); | |||
7974 | BasicBlock::iterator LowerEnd = BB->end(); | |||
7975 | while (UpIter != UpperEnd && DownIter != LowerEnd && &*UpIter != I && | |||
7976 | &*DownIter != I) { | |||
7977 | if (++ScheduleRegionSize > ScheduleRegionSizeLimit) { | |||
7978 | LLVM_DEBUG(dbgs() << "SLP: exceeded schedule region size limit\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: exceeded schedule region size limit\n" ; } } while (false); | |||
7979 | return false; | |||
7980 | } | |||
7981 | ||||
7982 | ++UpIter; | |||
7983 | ++DownIter; | |||
7984 | } | |||
7985 | if (DownIter == LowerEnd || (UpIter != UpperEnd && &*UpIter == I)) { | |||
7986 | assert(I->getParent() == ScheduleStart->getParent() &&(static_cast <bool> (I->getParent() == ScheduleStart ->getParent() && "Instruction is in wrong basic block." ) ? void (0) : __assert_fail ("I->getParent() == ScheduleStart->getParent() && \"Instruction is in wrong basic block.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7987, __extension__ __PRETTY_FUNCTION__)) | |||
7987 | "Instruction is in wrong basic block.")(static_cast <bool> (I->getParent() == ScheduleStart ->getParent() && "Instruction is in wrong basic block." ) ? void (0) : __assert_fail ("I->getParent() == ScheduleStart->getParent() && \"Instruction is in wrong basic block.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7987, __extension__ __PRETTY_FUNCTION__)); | |||
7988 | initScheduleData(I, ScheduleStart, nullptr, FirstLoadStoreInRegion); | |||
7989 | ScheduleStart = I; | |||
7990 | if (isOneOf(S, I) != I) | |||
7991 | CheckScheduleForI(I); | |||
7992 | LLVM_DEBUG(dbgs() << "SLP: extend schedule region start to " << *Ido { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: extend schedule region start to " << *I << "\n"; } } while (false) | |||
7993 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: extend schedule region start to " << *I << "\n"; } } while (false); | |||
7994 | return true; | |||
7995 | } | |||
7996 | assert((UpIter == UpperEnd || (DownIter != LowerEnd && &*DownIter == I)) &&(static_cast <bool> ((UpIter == UpperEnd || (DownIter != LowerEnd && &*DownIter == I)) && "Expected to reach top of the basic block or instruction down the " "lower end.") ? void (0) : __assert_fail ("(UpIter == UpperEnd || (DownIter != LowerEnd && &*DownIter == I)) && \"Expected to reach top of the basic block or instruction down the \" \"lower end.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7998, __extension__ __PRETTY_FUNCTION__)) | |||
7997 | "Expected to reach top of the basic block or instruction down the "(static_cast <bool> ((UpIter == UpperEnd || (DownIter != LowerEnd && &*DownIter == I)) && "Expected to reach top of the basic block or instruction down the " "lower end.") ? void (0) : __assert_fail ("(UpIter == UpperEnd || (DownIter != LowerEnd && &*DownIter == I)) && \"Expected to reach top of the basic block or instruction down the \" \"lower end.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7998, __extension__ __PRETTY_FUNCTION__)) | |||
7998 | "lower end.")(static_cast <bool> ((UpIter == UpperEnd || (DownIter != LowerEnd && &*DownIter == I)) && "Expected to reach top of the basic block or instruction down the " "lower end.") ? void (0) : __assert_fail ("(UpIter == UpperEnd || (DownIter != LowerEnd && &*DownIter == I)) && \"Expected to reach top of the basic block or instruction down the \" \"lower end.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7998, __extension__ __PRETTY_FUNCTION__)); | |||
7999 | assert(I->getParent() == ScheduleEnd->getParent() &&(static_cast <bool> (I->getParent() == ScheduleEnd-> getParent() && "Instruction is in wrong basic block." ) ? void (0) : __assert_fail ("I->getParent() == ScheduleEnd->getParent() && \"Instruction is in wrong basic block.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8000, __extension__ __PRETTY_FUNCTION__)) | |||
8000 | "Instruction is in wrong basic block.")(static_cast <bool> (I->getParent() == ScheduleEnd-> getParent() && "Instruction is in wrong basic block." ) ? void (0) : __assert_fail ("I->getParent() == ScheduleEnd->getParent() && \"Instruction is in wrong basic block.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8000, __extension__ __PRETTY_FUNCTION__)); | |||
8001 | initScheduleData(ScheduleEnd, I->getNextNode(), LastLoadStoreInRegion, | |||
8002 | nullptr); | |||
8003 | ScheduleEnd = I->getNextNode(); | |||
8004 | if (isOneOf(S, I) != I) | |||
8005 | CheckScheduleForI(I); | |||
8006 | assert(ScheduleEnd && "tried to vectorize a terminator?")(static_cast <bool> (ScheduleEnd && "tried to vectorize a terminator?" ) ? void (0) : __assert_fail ("ScheduleEnd && \"tried to vectorize a terminator?\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8006, __extension__ __PRETTY_FUNCTION__)); | |||
8007 | LLVM_DEBUG(dbgs() << "SLP: extend schedule region end to " << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: extend schedule region end to " << *I << "\n"; } } while (false); | |||
8008 | return true; | |||
8009 | } | |||
8010 | ||||
8011 | void BoUpSLP::BlockScheduling::initScheduleData(Instruction *FromI, | |||
8012 | Instruction *ToI, | |||
8013 | ScheduleData *PrevLoadStore, | |||
8014 | ScheduleData *NextLoadStore) { | |||
8015 | ScheduleData *CurrentLoadStore = PrevLoadStore; | |||
8016 | for (Instruction *I = FromI; I != ToI; I = I->getNextNode()) { | |||
8017 | // No need to allocate data for non-schedulable instructions. | |||
8018 | if (doesNotNeedToBeScheduled(I)) | |||
8019 | continue; | |||
8020 | ScheduleData *SD = ScheduleDataMap.lookup(I); | |||
8021 | if (!SD) { | |||
8022 | SD = allocateScheduleDataChunks(); | |||
8023 | ScheduleDataMap[I] = SD; | |||
8024 | SD->Inst = I; | |||
8025 | } | |||
8026 | assert(!isInSchedulingRegion(SD) &&(static_cast <bool> (!isInSchedulingRegion(SD) && "new ScheduleData already in scheduling region") ? void (0) : __assert_fail ("!isInSchedulingRegion(SD) && \"new ScheduleData already in scheduling region\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8027, __extension__ __PRETTY_FUNCTION__)) | |||
8027 | "new ScheduleData already in scheduling region")(static_cast <bool> (!isInSchedulingRegion(SD) && "new ScheduleData already in scheduling region") ? void (0) : __assert_fail ("!isInSchedulingRegion(SD) && \"new ScheduleData already in scheduling region\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8027, __extension__ __PRETTY_FUNCTION__)); | |||
8028 | SD->init(SchedulingRegionID, I); | |||
8029 | ||||
8030 | if (I->mayReadOrWriteMemory() && | |||
8031 | (!isa<IntrinsicInst>(I) || | |||
8032 | (cast<IntrinsicInst>(I)->getIntrinsicID() != Intrinsic::sideeffect && | |||
8033 | cast<IntrinsicInst>(I)->getIntrinsicID() != | |||
8034 | Intrinsic::pseudoprobe))) { | |||
8035 | // Update the linked list of memory accessing instructions. | |||
8036 | if (CurrentLoadStore) { | |||
8037 | CurrentLoadStore->NextLoadStore = SD; | |||
8038 | } else { | |||
8039 | FirstLoadStoreInRegion = SD; | |||
8040 | } | |||
8041 | CurrentLoadStore = SD; | |||
8042 | } | |||
8043 | ||||
8044 | if (match(I, m_Intrinsic<Intrinsic::stacksave>()) || | |||
8045 | match(I, m_Intrinsic<Intrinsic::stackrestore>())) | |||
8046 | RegionHasStackSave = true; | |||
8047 | } | |||
8048 | if (NextLoadStore) { | |||
8049 | if (CurrentLoadStore) | |||
8050 | CurrentLoadStore->NextLoadStore = NextLoadStore; | |||
8051 | } else { | |||
8052 | LastLoadStoreInRegion = CurrentLoadStore; | |||
8053 | } | |||
8054 | } | |||
8055 | ||||
8056 | void BoUpSLP::BlockScheduling::calculateDependencies(ScheduleData *SD, | |||
8057 | bool InsertInReadyList, | |||
8058 | BoUpSLP *SLP) { | |||
8059 | assert(SD->isSchedulingEntity())(static_cast <bool> (SD->isSchedulingEntity()) ? void (0) : __assert_fail ("SD->isSchedulingEntity()", "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 8059, __extension__ __PRETTY_FUNCTION__)); | |||
8060 | ||||
8061 | SmallVector<ScheduleData *, 10> WorkList; | |||
8062 | WorkList.push_back(SD); | |||
8063 | ||||
8064 | while (!WorkList.empty()) { | |||
8065 | ScheduleData *SD = WorkList.pop_back_val(); | |||
8066 | for (ScheduleData *BundleMember = SD; BundleMember; | |||
8067 | BundleMember = BundleMember->NextInBundle) { | |||
8068 | assert(isInSchedulingRegion(BundleMember))(static_cast <bool> (isInSchedulingRegion(BundleMember) ) ? void (0) : __assert_fail ("isInSchedulingRegion(BundleMember)" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8068, __extension__ __PRETTY_FUNCTION__)); | |||
8069 | if (BundleMember->hasValidDependencies()) | |||
8070 | continue; | |||
8071 | ||||
8072 | LLVM_DEBUG(dbgs() << "SLP: update deps of " << *BundleMemberdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: update deps of " << *BundleMember << "\n"; } } while (false) | |||
8073 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: update deps of " << *BundleMember << "\n"; } } while (false); | |||
8074 | BundleMember->Dependencies = 0; | |||
8075 | BundleMember->resetUnscheduledDeps(); | |||
8076 | ||||
8077 | // Handle def-use chain dependencies. | |||
8078 | if (BundleMember->OpValue != BundleMember->Inst) { | |||
8079 | if (ScheduleData *UseSD = getScheduleData(BundleMember->Inst)) { | |||
8080 | BundleMember->Dependencies++; | |||
8081 | ScheduleData *DestBundle = UseSD->FirstInBundle; | |||
8082 | if (!DestBundle->IsScheduled) | |||
8083 | BundleMember->incrementUnscheduledDeps(1); | |||
8084 | if (!DestBundle->hasValidDependencies()) | |||
8085 | WorkList.push_back(DestBundle); | |||
8086 | } | |||
8087 | } else { | |||
8088 | for (User *U : BundleMember->Inst->users()) { | |||
8089 | if (ScheduleData *UseSD = getScheduleData(cast<Instruction>(U))) { | |||
8090 | BundleMember->Dependencies++; | |||
8091 | ScheduleData *DestBundle = UseSD->FirstInBundle; | |||
8092 | if (!DestBundle->IsScheduled) | |||
8093 | BundleMember->incrementUnscheduledDeps(1); | |||
8094 | if (!DestBundle->hasValidDependencies()) | |||
8095 | WorkList.push_back(DestBundle); | |||
8096 | } | |||
8097 | } | |||
8098 | } | |||
8099 | ||||
8100 | auto makeControlDependent = [&](Instruction *I) { | |||
8101 | auto *DepDest = getScheduleData(I); | |||
8102 | assert(DepDest && "must be in schedule window")(static_cast <bool> (DepDest && "must be in schedule window" ) ? void (0) : __assert_fail ("DepDest && \"must be in schedule window\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8102, __extension__ __PRETTY_FUNCTION__)); | |||
8103 | DepDest->ControlDependencies.push_back(BundleMember); | |||
8104 | BundleMember->Dependencies++; | |||
8105 | ScheduleData *DestBundle = DepDest->FirstInBundle; | |||
8106 | if (!DestBundle->IsScheduled) | |||
8107 | BundleMember->incrementUnscheduledDeps(1); | |||
8108 | if (!DestBundle->hasValidDependencies()) | |||
8109 | WorkList.push_back(DestBundle); | |||
8110 | }; | |||
8111 | ||||
8112 | // Any instruction which isn't safe to speculate at the begining of the | |||
8113 | // block is control dependend on any early exit or non-willreturn call | |||
8114 | // which proceeds it. | |||
8115 | if (!isGuaranteedToTransferExecutionToSuccessor(BundleMember->Inst)) { | |||
8116 | for (Instruction *I = BundleMember->Inst->getNextNode(); | |||
8117 | I != ScheduleEnd; I = I->getNextNode()) { | |||
8118 | if (isSafeToSpeculativelyExecute(I, &*BB->begin())) | |||
8119 | continue; | |||
8120 | ||||
8121 | // Add the dependency | |||
8122 | makeControlDependent(I); | |||
8123 | ||||
8124 | if (!isGuaranteedToTransferExecutionToSuccessor(I)) | |||
8125 | // Everything past here must be control dependent on I. | |||
8126 | break; | |||
8127 | } | |||
8128 | } | |||
8129 | ||||
8130 | if (RegionHasStackSave) { | |||
8131 | // If we have an inalloc alloca instruction, it needs to be scheduled | |||
8132 | // after any preceeding stacksave. We also need to prevent any alloca | |||
8133 | // from reordering above a preceeding stackrestore. | |||
8134 | if (match(BundleMember->Inst, m_Intrinsic<Intrinsic::stacksave>()) || | |||
8135 | match(BundleMember->Inst, m_Intrinsic<Intrinsic::stackrestore>())) { | |||
8136 | for (Instruction *I = BundleMember->Inst->getNextNode(); | |||
8137 | I != ScheduleEnd; I = I->getNextNode()) { | |||
8138 | if (match(I, m_Intrinsic<Intrinsic::stacksave>()) || | |||
8139 | match(I, m_Intrinsic<Intrinsic::stackrestore>())) | |||
8140 | // Any allocas past here must be control dependent on I, and I | |||
8141 | // must be memory dependend on BundleMember->Inst. | |||
8142 | break; | |||
8143 | ||||
8144 | if (!isa<AllocaInst>(I)) | |||
8145 | continue; | |||
8146 | ||||
8147 | // Add the dependency | |||
8148 | makeControlDependent(I); | |||
8149 | } | |||
8150 | } | |||
8151 | ||||
8152 | // In addition to the cases handle just above, we need to prevent | |||
8153 | // allocas from moving below a stacksave. The stackrestore case | |||
8154 | // is currently thought to be conservatism. | |||
8155 | if (isa<AllocaInst>(BundleMember->Inst)) { | |||
8156 | for (Instruction *I = BundleMember->Inst->getNextNode(); | |||
8157 | I != ScheduleEnd; I = I->getNextNode()) { | |||
8158 | if (!match(I, m_Intrinsic<Intrinsic::stacksave>()) && | |||
8159 | !match(I, m_Intrinsic<Intrinsic::stackrestore>())) | |||
8160 | continue; | |||
8161 | ||||
8162 | // Add the dependency | |||
8163 | makeControlDependent(I); | |||
8164 | break; | |||
8165 | } | |||
8166 | } | |||
8167 | } | |||
8168 | ||||
8169 | // Handle the memory dependencies (if any). | |||
8170 | ScheduleData *DepDest = BundleMember->NextLoadStore; | |||
8171 | if (!DepDest) | |||
8172 | continue; | |||
8173 | Instruction *SrcInst = BundleMember->Inst; | |||
8174 | assert(SrcInst->mayReadOrWriteMemory() &&(static_cast <bool> (SrcInst->mayReadOrWriteMemory() && "NextLoadStore list for non memory effecting bundle?" ) ? void (0) : __assert_fail ("SrcInst->mayReadOrWriteMemory() && \"NextLoadStore list for non memory effecting bundle?\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8175, __extension__ __PRETTY_FUNCTION__)) | |||
8175 | "NextLoadStore list for non memory effecting bundle?")(static_cast <bool> (SrcInst->mayReadOrWriteMemory() && "NextLoadStore list for non memory effecting bundle?" ) ? void (0) : __assert_fail ("SrcInst->mayReadOrWriteMemory() && \"NextLoadStore list for non memory effecting bundle?\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8175, __extension__ __PRETTY_FUNCTION__)); | |||
8176 | MemoryLocation SrcLoc = getLocation(SrcInst); | |||
8177 | bool SrcMayWrite = BundleMember->Inst->mayWriteToMemory(); | |||
8178 | unsigned numAliased = 0; | |||
8179 | unsigned DistToSrc = 1; | |||
8180 | ||||
8181 | for ( ; DepDest; DepDest = DepDest->NextLoadStore) { | |||
8182 | assert(isInSchedulingRegion(DepDest))(static_cast <bool> (isInSchedulingRegion(DepDest)) ? void (0) : __assert_fail ("isInSchedulingRegion(DepDest)", "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 8182, __extension__ __PRETTY_FUNCTION__)); | |||
8183 | ||||
8184 | // We have two limits to reduce the complexity: | |||
8185 | // 1) AliasedCheckLimit: It's a small limit to reduce calls to | |||
8186 | // SLP->isAliased (which is the expensive part in this loop). | |||
8187 | // 2) MaxMemDepDistance: It's for very large blocks and it aborts | |||
8188 | // the whole loop (even if the loop is fast, it's quadratic). | |||
8189 | // It's important for the loop break condition (see below) to | |||
8190 | // check this limit even between two read-only instructions. | |||
8191 | if (DistToSrc >= MaxMemDepDistance || | |||
8192 | ((SrcMayWrite || DepDest->Inst->mayWriteToMemory()) && | |||
8193 | (numAliased >= AliasedCheckLimit || | |||
8194 | SLP->isAliased(SrcLoc, SrcInst, DepDest->Inst)))) { | |||
8195 | ||||
8196 | // We increment the counter only if the locations are aliased | |||
8197 | // (instead of counting all alias checks). This gives a better | |||
8198 | // balance between reduced runtime and accurate dependencies. | |||
8199 | numAliased++; | |||
8200 | ||||
8201 | DepDest->MemoryDependencies.push_back(BundleMember); | |||
8202 | BundleMember->Dependencies++; | |||
8203 | ScheduleData *DestBundle = DepDest->FirstInBundle; | |||
8204 | if (!DestBundle->IsScheduled) { | |||
8205 | BundleMember->incrementUnscheduledDeps(1); | |||
8206 | } | |||
8207 | if (!DestBundle->hasValidDependencies()) { | |||
8208 | WorkList.push_back(DestBundle); | |||
8209 | } | |||
8210 | } | |||
8211 | ||||
8212 | // Example, explaining the loop break condition: Let's assume our | |||
8213 | // starting instruction is i0 and MaxMemDepDistance = 3. | |||
8214 | // | |||
8215 | // +--------v--v--v | |||
8216 | // i0,i1,i2,i3,i4,i5,i6,i7,i8 | |||
8217 | // +--------^--^--^ | |||
8218 | // | |||
8219 | // MaxMemDepDistance let us stop alias-checking at i3 and we add | |||
8220 | // dependencies from i0 to i3,i4,.. (even if they are not aliased). | |||
8221 | // Previously we already added dependencies from i3 to i6,i7,i8 | |||
8222 | // (because of MaxMemDepDistance). As we added a dependency from | |||
8223 | // i0 to i3, we have transitive dependencies from i0 to i6,i7,i8 | |||
8224 | // and we can abort this loop at i6. | |||
8225 | if (DistToSrc >= 2 * MaxMemDepDistance) | |||
8226 | break; | |||
8227 | DistToSrc++; | |||
8228 | } | |||
8229 | } | |||
8230 | if (InsertInReadyList && SD->isReady()) { | |||
8231 | ReadyInsts.insert(SD); | |||
8232 | LLVM_DEBUG(dbgs() << "SLP: gets ready on update: " << *SD->Instdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: gets ready on update: " << *SD->Inst << "\n"; } } while (false) | |||
8233 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: gets ready on update: " << *SD->Inst << "\n"; } } while (false); | |||
8234 | } | |||
8235 | } | |||
8236 | } | |||
8237 | ||||
8238 | void BoUpSLP::BlockScheduling::resetSchedule() { | |||
8239 | assert(ScheduleStart &&(static_cast <bool> (ScheduleStart && "tried to reset schedule on block which has not been scheduled" ) ? void (0) : __assert_fail ("ScheduleStart && \"tried to reset schedule on block which has not been scheduled\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8240, __extension__ __PRETTY_FUNCTION__)) | |||
8240 | "tried to reset schedule on block which has not been scheduled")(static_cast <bool> (ScheduleStart && "tried to reset schedule on block which has not been scheduled" ) ? void (0) : __assert_fail ("ScheduleStart && \"tried to reset schedule on block which has not been scheduled\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8240, __extension__ __PRETTY_FUNCTION__)); | |||
8241 | for (Instruction *I = ScheduleStart; I != ScheduleEnd; I = I->getNextNode()) { | |||
8242 | doForAllOpcodes(I, [&](ScheduleData *SD) { | |||
8243 | assert(isInSchedulingRegion(SD) &&(static_cast <bool> (isInSchedulingRegion(SD) && "ScheduleData not in scheduling region") ? void (0) : __assert_fail ("isInSchedulingRegion(SD) && \"ScheduleData not in scheduling region\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8244, __extension__ __PRETTY_FUNCTION__)) | |||
8244 | "ScheduleData not in scheduling region")(static_cast <bool> (isInSchedulingRegion(SD) && "ScheduleData not in scheduling region") ? void (0) : __assert_fail ("isInSchedulingRegion(SD) && \"ScheduleData not in scheduling region\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8244, __extension__ __PRETTY_FUNCTION__)); | |||
8245 | SD->IsScheduled = false; | |||
8246 | SD->resetUnscheduledDeps(); | |||
8247 | }); | |||
8248 | } | |||
8249 | ReadyInsts.clear(); | |||
8250 | } | |||
8251 | ||||
8252 | void BoUpSLP::scheduleBlock(BlockScheduling *BS) { | |||
8253 | if (!BS->ScheduleStart) | |||
8254 | return; | |||
8255 | ||||
8256 | LLVM_DEBUG(dbgs() << "SLP: schedule block " << BS->BB->getName() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: schedule block " << BS ->BB->getName() << "\n"; } } while (false); | |||
8257 | ||||
8258 | // A key point - if we got here, pre-scheduling was able to find a valid | |||
8259 | // scheduling of the sub-graph of the scheduling window which consists | |||
8260 | // of all vector bundles and their transitive users. As such, we do not | |||
8261 | // need to reschedule anything *outside of* that subgraph. | |||
8262 | ||||
8263 | BS->resetSchedule(); | |||
8264 | ||||
8265 | // For the real scheduling we use a more sophisticated ready-list: it is | |||
8266 | // sorted by the original instruction location. This lets the final schedule | |||
8267 | // be as close as possible to the original instruction order. | |||
8268 | // WARNING: If changing this order causes a correctness issue, that means | |||
8269 | // there is some missing dependence edge in the schedule data graph. | |||
8270 | struct ScheduleDataCompare { | |||
8271 | bool operator()(ScheduleData *SD1, ScheduleData *SD2) const { | |||
8272 | return SD2->SchedulingPriority < SD1->SchedulingPriority; | |||
8273 | } | |||
8274 | }; | |||
8275 | std::set<ScheduleData *, ScheduleDataCompare> ReadyInsts; | |||
8276 | ||||
8277 | // Ensure that all dependency data is updated (for nodes in the sub-graph) | |||
8278 | // and fill the ready-list with initial instructions. | |||
8279 | int Idx = 0; | |||
8280 | for (auto *I = BS->ScheduleStart; I != BS->ScheduleEnd; | |||
8281 | I = I->getNextNode()) { | |||
8282 | BS->doForAllOpcodes(I, [this, &Idx, BS](ScheduleData *SD) { | |||
8283 | TreeEntry *SDTE = getTreeEntry(SD->Inst); | |||
8284 | (void)SDTE; | |||
8285 | assert((isVectorLikeInstWithConstOps(SD->Inst) ||(static_cast <bool> ((isVectorLikeInstWithConstOps(SD-> Inst) || SD->isPartOfBundle() == (SDTE && !doesNotNeedToSchedule (SDTE->Scalars))) && "scheduler and vectorizer bundle mismatch" ) ? void (0) : __assert_fail ("(isVectorLikeInstWithConstOps(SD->Inst) || SD->isPartOfBundle() == (SDTE && !doesNotNeedToSchedule(SDTE->Scalars))) && \"scheduler and vectorizer bundle mismatch\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8288, __extension__ __PRETTY_FUNCTION__)) | |||
8286 | SD->isPartOfBundle() ==(static_cast <bool> ((isVectorLikeInstWithConstOps(SD-> Inst) || SD->isPartOfBundle() == (SDTE && !doesNotNeedToSchedule (SDTE->Scalars))) && "scheduler and vectorizer bundle mismatch" ) ? void (0) : __assert_fail ("(isVectorLikeInstWithConstOps(SD->Inst) || SD->isPartOfBundle() == (SDTE && !doesNotNeedToSchedule(SDTE->Scalars))) && \"scheduler and vectorizer bundle mismatch\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8288, __extension__ __PRETTY_FUNCTION__)) | |||
8287 | (SDTE && !doesNotNeedToSchedule(SDTE->Scalars))) &&(static_cast <bool> ((isVectorLikeInstWithConstOps(SD-> Inst) || SD->isPartOfBundle() == (SDTE && !doesNotNeedToSchedule (SDTE->Scalars))) && "scheduler and vectorizer bundle mismatch" ) ? void (0) : __assert_fail ("(isVectorLikeInstWithConstOps(SD->Inst) || SD->isPartOfBundle() == (SDTE && !doesNotNeedToSchedule(SDTE->Scalars))) && \"scheduler and vectorizer bundle mismatch\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8288, __extension__ __PRETTY_FUNCTION__)) | |||
8288 | "scheduler and vectorizer bundle mismatch")(static_cast <bool> ((isVectorLikeInstWithConstOps(SD-> Inst) || SD->isPartOfBundle() == (SDTE && !doesNotNeedToSchedule (SDTE->Scalars))) && "scheduler and vectorizer bundle mismatch" ) ? void (0) : __assert_fail ("(isVectorLikeInstWithConstOps(SD->Inst) || SD->isPartOfBundle() == (SDTE && !doesNotNeedToSchedule(SDTE->Scalars))) && \"scheduler and vectorizer bundle mismatch\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8288, __extension__ __PRETTY_FUNCTION__)); | |||
8289 | SD->FirstInBundle->SchedulingPriority = Idx++; | |||
8290 | ||||
8291 | if (SD->isSchedulingEntity() && SD->isPartOfBundle()) | |||
8292 | BS->calculateDependencies(SD, false, this); | |||
8293 | }); | |||
8294 | } | |||
8295 | BS->initialFillReadyList(ReadyInsts); | |||
8296 | ||||
8297 | Instruction *LastScheduledInst = BS->ScheduleEnd; | |||
8298 | ||||
8299 | // Do the "real" scheduling. | |||
8300 | while (!ReadyInsts.empty()) { | |||
8301 | ScheduleData *picked = *ReadyInsts.begin(); | |||
8302 | ReadyInsts.erase(ReadyInsts.begin()); | |||
8303 | ||||
8304 | // Move the scheduled instruction(s) to their dedicated places, if not | |||
8305 | // there yet. | |||
8306 | for (ScheduleData *BundleMember = picked; BundleMember; | |||
8307 | BundleMember = BundleMember->NextInBundle) { | |||
8308 | Instruction *pickedInst = BundleMember->Inst; | |||
8309 | if (pickedInst->getNextNode() != LastScheduledInst) | |||
8310 | pickedInst->moveBefore(LastScheduledInst); | |||
8311 | LastScheduledInst = pickedInst; | |||
8312 | } | |||
8313 | ||||
8314 | BS->schedule(picked, ReadyInsts); | |||
8315 | } | |||
8316 | ||||
8317 | // Check that we didn't break any of our invariants. | |||
8318 | #ifdef EXPENSIVE_CHECKS | |||
8319 | BS->verify(); | |||
8320 | #endif | |||
8321 | ||||
8322 | #if !defined(NDEBUG) || defined(EXPENSIVE_CHECKS) | |||
8323 | // Check that all schedulable entities got scheduled | |||
8324 | for (auto *I = BS->ScheduleStart; I != BS->ScheduleEnd; I = I->getNextNode()) { | |||
8325 | BS->doForAllOpcodes(I, [&](ScheduleData *SD) { | |||
8326 | if (SD->isSchedulingEntity() && SD->hasValidDependencies()) { | |||
8327 | assert(SD->IsScheduled && "must be scheduled at this point")(static_cast <bool> (SD->IsScheduled && "must be scheduled at this point" ) ? void (0) : __assert_fail ("SD->IsScheduled && \"must be scheduled at this point\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8327, __extension__ __PRETTY_FUNCTION__)); | |||
8328 | } | |||
8329 | }); | |||
8330 | } | |||
8331 | #endif | |||
8332 | ||||
8333 | // Avoid duplicate scheduling of the block. | |||
8334 | BS->ScheduleStart = nullptr; | |||
8335 | } | |||
8336 | ||||
8337 | unsigned BoUpSLP::getVectorElementSize(Value *V) { | |||
8338 | // If V is a store, just return the width of the stored value (or value | |||
8339 | // truncated just before storing) without traversing the expression tree. | |||
8340 | // This is the common case. | |||
8341 | if (auto *Store = dyn_cast<StoreInst>(V)) { | |||
8342 | if (auto *Trunc = dyn_cast<TruncInst>(Store->getValueOperand())) | |||
8343 | return DL->getTypeSizeInBits(Trunc->getSrcTy()); | |||
8344 | return DL->getTypeSizeInBits(Store->getValueOperand()->getType()); | |||
8345 | } | |||
8346 | ||||
8347 | if (auto *IEI = dyn_cast<InsertElementInst>(V)) | |||
8348 | return getVectorElementSize(IEI->getOperand(1)); | |||
8349 | ||||
8350 | auto E = InstrElementSize.find(V); | |||
8351 | if (E != InstrElementSize.end()) | |||
8352 | return E->second; | |||
8353 | ||||
8354 | // If V is not a store, we can traverse the expression tree to find loads | |||
8355 | // that feed it. The type of the loaded value may indicate a more suitable | |||
8356 | // width than V's type. We want to base the vector element size on the width | |||
8357 | // of memory operations where possible. | |||
8358 | SmallVector<std::pair<Instruction *, BasicBlock *>, 16> Worklist; | |||
8359 | SmallPtrSet<Instruction *, 16> Visited; | |||
8360 | if (auto *I = dyn_cast<Instruction>(V)) { | |||
8361 | Worklist.emplace_back(I, I->getParent()); | |||
8362 | Visited.insert(I); | |||
8363 | } | |||
8364 | ||||
8365 | // Traverse the expression tree in bottom-up order looking for loads. If we | |||
8366 | // encounter an instruction we don't yet handle, we give up. | |||
8367 | auto Width = 0u; | |||
8368 | while (!Worklist.empty()) { | |||
8369 | Instruction *I; | |||
8370 | BasicBlock *Parent; | |||
8371 | std::tie(I, Parent) = Worklist.pop_back_val(); | |||
8372 | ||||
8373 | // We should only be looking at scalar instructions here. If the current | |||
8374 | // instruction has a vector type, skip. | |||
8375 | auto *Ty = I->getType(); | |||
8376 | if (isa<VectorType>(Ty)) | |||
8377 | continue; | |||
8378 | ||||
8379 | // If the current instruction is a load, update MaxWidth to reflect the | |||
8380 | // width of the loaded value. | |||
8381 | if (isa<LoadInst>(I) || isa<ExtractElementInst>(I) || | |||
8382 | isa<ExtractValueInst>(I)) | |||
8383 | Width = std::max<unsigned>(Width, DL->getTypeSizeInBits(Ty)); | |||
8384 | ||||
8385 | // Otherwise, we need to visit the operands of the instruction. We only | |||
8386 | // handle the interesting cases from buildTree here. If an operand is an | |||
8387 | // instruction we haven't yet visited and from the same basic block as the | |||
8388 | // user or the use is a PHI node, we add it to the worklist. | |||
8389 | else if (isa<PHINode>(I) || isa<CastInst>(I) || isa<GetElementPtrInst>(I) || | |||
8390 | isa<CmpInst>(I) || isa<SelectInst>(I) || isa<BinaryOperator>(I) || | |||
8391 | isa<UnaryOperator>(I)) { | |||
8392 | for (Use &U : I->operands()) | |||
8393 | if (auto *J = dyn_cast<Instruction>(U.get())) | |||
8394 | if (Visited.insert(J).second && | |||
8395 | (isa<PHINode>(I) || J->getParent() == Parent)) | |||
8396 | Worklist.emplace_back(J, J->getParent()); | |||
8397 | } else { | |||
8398 | break; | |||
8399 | } | |||
8400 | } | |||
8401 | ||||
8402 | // If we didn't encounter a memory access in the expression tree, or if we | |||
8403 | // gave up for some reason, just return the width of V. Otherwise, return the | |||
8404 | // maximum width we found. | |||
8405 | if (!Width) { | |||
8406 | if (auto *CI = dyn_cast<CmpInst>(V)) | |||
8407 | V = CI->getOperand(0); | |||
8408 | Width = DL->getTypeSizeInBits(V->getType()); | |||
8409 | } | |||
8410 | ||||
8411 | for (Instruction *I : Visited) | |||
8412 | InstrElementSize[I] = Width; | |||
8413 | ||||
8414 | return Width; | |||
8415 | } | |||
8416 | ||||
8417 | // Determine if a value V in a vectorizable expression Expr can be demoted to a | |||
8418 | // smaller type with a truncation. We collect the values that will be demoted | |||
8419 | // in ToDemote and additional roots that require investigating in Roots. | |||
8420 | static bool collectValuesToDemote(Value *V, SmallPtrSetImpl<Value *> &Expr, | |||
8421 | SmallVectorImpl<Value *> &ToDemote, | |||
8422 | SmallVectorImpl<Value *> &Roots) { | |||
8423 | // We can always demote constants. | |||
8424 | if (isa<Constant>(V)) { | |||
8425 | ToDemote.push_back(V); | |||
8426 | return true; | |||
8427 | } | |||
8428 | ||||
8429 | // If the value is not an instruction in the expression with only one use, it | |||
8430 | // cannot be demoted. | |||
8431 | auto *I = dyn_cast<Instruction>(V); | |||
8432 | if (!I || !I->hasOneUse() || !Expr.count(I)) | |||
8433 | return false; | |||
8434 | ||||
8435 | switch (I->getOpcode()) { | |||
8436 | ||||
8437 | // We can always demote truncations and extensions. Since truncations can | |||
8438 | // seed additional demotion, we save the truncated value. | |||
8439 | case Instruction::Trunc: | |||
8440 | Roots.push_back(I->getOperand(0)); | |||
8441 | break; | |||
8442 | case Instruction::ZExt: | |||
8443 | case Instruction::SExt: | |||
8444 | if (isa<ExtractElementInst>(I->getOperand(0)) || | |||
8445 | isa<InsertElementInst>(I->getOperand(0))) | |||
8446 | return false; | |||
8447 | break; | |||
8448 | ||||
8449 | // We can demote certain binary operations if we can demote both of their | |||
8450 | // operands. | |||
8451 | case Instruction::Add: | |||
8452 | case Instruction::Sub: | |||
8453 | case Instruction::Mul: | |||
8454 | case Instruction::And: | |||
8455 | case Instruction::Or: | |||
8456 | case Instruction::Xor: | |||
8457 | if (!collectValuesToDemote(I->getOperand(0), Expr, ToDemote, Roots) || | |||
8458 | !collectValuesToDemote(I->getOperand(1), Expr, ToDemote, Roots)) | |||
8459 | return false; | |||
8460 | break; | |||
8461 | ||||
8462 | // We can demote selects if we can demote their true and false values. | |||
8463 | case Instruction::Select: { | |||
8464 | SelectInst *SI = cast<SelectInst>(I); | |||
8465 | if (!collectValuesToDemote(SI->getTrueValue(), Expr, ToDemote, Roots) || | |||
8466 | !collectValuesToDemote(SI->getFalseValue(), Expr, ToDemote, Roots)) | |||
8467 | return false; | |||
8468 | break; | |||
8469 | } | |||
8470 | ||||
8471 | // We can demote phis if we can demote all their incoming operands. Note that | |||
8472 | // we don't need to worry about cycles since we ensure single use above. | |||
8473 | case Instruction::PHI: { | |||
8474 | PHINode *PN = cast<PHINode>(I); | |||
8475 | for (Value *IncValue : PN->incoming_values()) | |||
8476 | if (!collectValuesToDemote(IncValue, Expr, ToDemote, Roots)) | |||
8477 | return false; | |||
8478 | break; | |||
8479 | } | |||
8480 | ||||
8481 | // Otherwise, conservatively give up. | |||
8482 | default: | |||
8483 | return false; | |||
8484 | } | |||
8485 | ||||
8486 | // Record the value that we can demote. | |||
8487 | ToDemote.push_back(V); | |||
8488 | return true; | |||
8489 | } | |||
8490 | ||||
8491 | void BoUpSLP::computeMinimumValueSizes() { | |||
8492 | // If there are no external uses, the expression tree must be rooted by a | |||
8493 | // store. We can't demote in-memory values, so there is nothing to do here. | |||
8494 | if (ExternalUses.empty()) | |||
8495 | return; | |||
8496 | ||||
8497 | // We only attempt to truncate integer expressions. | |||
8498 | auto &TreeRoot = VectorizableTree[0]->Scalars; | |||
8499 | auto *TreeRootIT = dyn_cast<IntegerType>(TreeRoot[0]->getType()); | |||
8500 | if (!TreeRootIT) | |||
8501 | return; | |||
8502 | ||||
8503 | // If the expression is not rooted by a store, these roots should have | |||
8504 | // external uses. We will rely on InstCombine to rewrite the expression in | |||
8505 | // the narrower type. However, InstCombine only rewrites single-use values. | |||
8506 | // This means that if a tree entry other than a root is used externally, it | |||
8507 | // must have multiple uses and InstCombine will not rewrite it. The code | |||
8508 | // below ensures that only the roots are used externally. | |||
8509 | SmallPtrSet<Value *, 32> Expr(TreeRoot.begin(), TreeRoot.end()); | |||
8510 | for (auto &EU : ExternalUses) | |||
8511 | if (!Expr.erase(EU.Scalar)) | |||
8512 | return; | |||
8513 | if (!Expr.empty()) | |||
8514 | return; | |||
8515 | ||||
8516 | // Collect the scalar values of the vectorizable expression. We will use this | |||
8517 | // context to determine which values can be demoted. If we see a truncation, | |||
8518 | // we mark it as seeding another demotion. | |||
8519 | for (auto &EntryPtr : VectorizableTree) | |||
8520 | Expr.insert(EntryPtr->Scalars.begin(), EntryPtr->Scalars.end()); | |||
8521 | ||||
8522 | // Ensure the roots of the vectorizable tree don't form a cycle. They must | |||
8523 | // have a single external user that is not in the vectorizable tree. | |||
8524 | for (auto *Root : TreeRoot) | |||
8525 | if (!Root->hasOneUse() || Expr.count(*Root->user_begin())) | |||
8526 | return; | |||
8527 | ||||
8528 | // Conservatively determine if we can actually truncate the roots of the | |||
8529 | // expression. Collect the values that can be demoted in ToDemote and | |||
8530 | // additional roots that require investigating in Roots. | |||
8531 | SmallVector<Value *, 32> ToDemote; | |||
8532 | SmallVector<Value *, 4> Roots; | |||
8533 | for (auto *Root : TreeRoot) | |||
8534 | if (!collectValuesToDemote(Root, Expr, ToDemote, Roots)) | |||
8535 | return; | |||
8536 | ||||
8537 | // The maximum bit width required to represent all the values that can be | |||
8538 | // demoted without loss of precision. It would be safe to truncate the roots | |||
8539 | // of the expression to this width. | |||
8540 | auto MaxBitWidth = 8u; | |||
8541 | ||||
8542 | // We first check if all the bits of the roots are demanded. If they're not, | |||
8543 | // we can truncate the roots to this narrower type. | |||
8544 | for (auto *Root : TreeRoot) { | |||
8545 | auto Mask = DB->getDemandedBits(cast<Instruction>(Root)); | |||
8546 | MaxBitWidth = std::max<unsigned>( | |||
8547 | Mask.getBitWidth() - Mask.countLeadingZeros(), MaxBitWidth); | |||
8548 | } | |||
8549 | ||||
8550 | // True if the roots can be zero-extended back to their original type, rather | |||
8551 | // than sign-extended. We know that if the leading bits are not demanded, we | |||
8552 | // can safely zero-extend. So we initialize IsKnownPositive to True. | |||
8553 | bool IsKnownPositive = true; | |||
8554 | ||||
8555 | // If all the bits of the roots are demanded, we can try a little harder to | |||
8556 | // compute a narrower type. This can happen, for example, if the roots are | |||
8557 | // getelementptr indices. InstCombine promotes these indices to the pointer | |||
8558 | // width. Thus, all their bits are technically demanded even though the | |||
8559 | // address computation might be vectorized in a smaller type. | |||
8560 | // | |||
8561 | // We start by looking at each entry that can be demoted. We compute the | |||
8562 | // maximum bit width required to store the scalar by using ValueTracking to | |||
8563 | // compute the number of high-order bits we can truncate. | |||
8564 | if (MaxBitWidth == DL->getTypeSizeInBits(TreeRoot[0]->getType()) && | |||
8565 | llvm::all_of(TreeRoot, [](Value *R) { | |||
8566 | assert(R->hasOneUse() && "Root should have only one use!")(static_cast <bool> (R->hasOneUse() && "Root should have only one use!" ) ? void (0) : __assert_fail ("R->hasOneUse() && \"Root should have only one use!\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8566, __extension__ __PRETTY_FUNCTION__)); | |||
8567 | return isa<GetElementPtrInst>(R->user_back()); | |||
8568 | })) { | |||
8569 | MaxBitWidth = 8u; | |||
8570 | ||||
8571 | // Determine if the sign bit of all the roots is known to be zero. If not, | |||
8572 | // IsKnownPositive is set to False. | |||
8573 | IsKnownPositive = llvm::all_of(TreeRoot, [&](Value *R) { | |||
8574 | KnownBits Known = computeKnownBits(R, *DL); | |||
8575 | return Known.isNonNegative(); | |||
8576 | }); | |||
8577 | ||||
8578 | // Determine the maximum number of bits required to store the scalar | |||
8579 | // values. | |||
8580 | for (auto *Scalar : ToDemote) { | |||
8581 | auto NumSignBits = ComputeNumSignBits(Scalar, *DL, 0, AC, nullptr, DT); | |||
8582 | auto NumTypeBits = DL->getTypeSizeInBits(Scalar->getType()); | |||
8583 | MaxBitWidth = std::max<unsigned>(NumTypeBits - NumSignBits, MaxBitWidth); | |||
8584 | } | |||
8585 | ||||
8586 | // If we can't prove that the sign bit is zero, we must add one to the | |||
8587 | // maximum bit width to account for the unknown sign bit. This preserves | |||
8588 | // the existing sign bit so we can safely sign-extend the root back to the | |||
8589 | // original type. Otherwise, if we know the sign bit is zero, we will | |||
8590 | // zero-extend the root instead. | |||
8591 | // | |||
8592 | // FIXME: This is somewhat suboptimal, as there will be cases where adding | |||
8593 | // one to the maximum bit width will yield a larger-than-necessary | |||
8594 | // type. In general, we need to add an extra bit only if we can't | |||
8595 | // prove that the upper bit of the original type is equal to the | |||
8596 | // upper bit of the proposed smaller type. If these two bits are the | |||
8597 | // same (either zero or one) we know that sign-extending from the | |||
8598 | // smaller type will result in the same value. Here, since we can't | |||
8599 | // yet prove this, we are just making the proposed smaller type | |||
8600 | // larger to ensure correctness. | |||
8601 | if (!IsKnownPositive) | |||
8602 | ++MaxBitWidth; | |||
8603 | } | |||
8604 | ||||
8605 | // Round MaxBitWidth up to the next power-of-two. | |||
8606 | if (!isPowerOf2_64(MaxBitWidth)) | |||
8607 | MaxBitWidth = NextPowerOf2(MaxBitWidth); | |||
8608 | ||||
8609 | // If the maximum bit width we compute is less than the with of the roots' | |||
8610 | // type, we can proceed with the narrowing. Otherwise, do nothing. | |||
8611 | if (MaxBitWidth >= TreeRootIT->getBitWidth()) | |||
8612 | return; | |||
8613 | ||||
8614 | // If we can truncate the root, we must collect additional values that might | |||
8615 | // be demoted as a result. That is, those seeded by truncations we will | |||
8616 | // modify. | |||
8617 | while (!Roots.empty()) | |||
8618 | collectValuesToDemote(Roots.pop_back_val(), Expr, ToDemote, Roots); | |||
8619 | ||||
8620 | // Finally, map the values we can demote to the maximum bit with we computed. | |||
8621 | for (auto *Scalar : ToDemote) | |||
8622 | MinBWs[Scalar] = std::make_pair(MaxBitWidth, !IsKnownPositive); | |||
8623 | } | |||
8624 | ||||
8625 | namespace { | |||
8626 | ||||
8627 | /// The SLPVectorizer Pass. | |||
8628 | struct SLPVectorizer : public FunctionPass { | |||
8629 | SLPVectorizerPass Impl; | |||
8630 | ||||
8631 | /// Pass identification, replacement for typeid | |||
8632 | static char ID; | |||
8633 | ||||
8634 | explicit SLPVectorizer() : FunctionPass(ID) { | |||
8635 | initializeSLPVectorizerPass(*PassRegistry::getPassRegistry()); | |||
8636 | } | |||
8637 | ||||
8638 | bool doInitialization(Module &M) override { return false; } | |||
8639 | ||||
8640 | bool runOnFunction(Function &F) override { | |||
8641 | if (skipFunction(F)) | |||
8642 | return false; | |||
8643 | ||||
8644 | auto *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE(); | |||
8645 | auto *TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F); | |||
8646 | auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>(); | |||
8647 | auto *TLI = TLIP ? &TLIP->getTLI(F) : nullptr; | |||
8648 | auto *AA = &getAnalysis<AAResultsWrapperPass>().getAAResults(); | |||
8649 | auto *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); | |||
8650 | auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); | |||
8651 | auto *AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F); | |||
8652 | auto *DB = &getAnalysis<DemandedBitsWrapperPass>().getDemandedBits(); | |||
8653 | auto *ORE = &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE(); | |||
8654 | ||||
8655 | return Impl.runImpl(F, SE, TTI, TLI, AA, LI, DT, AC, DB, ORE); | |||
8656 | } | |||
8657 | ||||
8658 | void getAnalysisUsage(AnalysisUsage &AU) const override { | |||
8659 | FunctionPass::getAnalysisUsage(AU); | |||
8660 | AU.addRequired<AssumptionCacheTracker>(); | |||
8661 | AU.addRequired<ScalarEvolutionWrapperPass>(); | |||
8662 | AU.addRequired<AAResultsWrapperPass>(); | |||
8663 | AU.addRequired<TargetTransformInfoWrapperPass>(); | |||
8664 | AU.addRequired<LoopInfoWrapperPass>(); | |||
8665 | AU.addRequired<DominatorTreeWrapperPass>(); | |||
8666 | AU.addRequired<DemandedBitsWrapperPass>(); | |||
8667 | AU.addRequired<OptimizationRemarkEmitterWrapperPass>(); | |||
8668 | AU.addRequired<InjectTLIMappingsLegacy>(); | |||
8669 | AU.addPreserved<LoopInfoWrapperPass>(); | |||
8670 | AU.addPreserved<DominatorTreeWrapperPass>(); | |||
8671 | AU.addPreserved<AAResultsWrapperPass>(); | |||
8672 | AU.addPreserved<GlobalsAAWrapperPass>(); | |||
8673 | AU.setPreservesCFG(); | |||
8674 | } | |||
8675 | }; | |||
8676 | ||||
8677 | } // end anonymous namespace | |||
8678 | ||||
8679 | PreservedAnalyses SLPVectorizerPass::run(Function &F, FunctionAnalysisManager &AM) { | |||
8680 | auto *SE = &AM.getResult<ScalarEvolutionAnalysis>(F); | |||
8681 | auto *TTI = &AM.getResult<TargetIRAnalysis>(F); | |||
8682 | auto *TLI = AM.getCachedResult<TargetLibraryAnalysis>(F); | |||
8683 | auto *AA = &AM.getResult<AAManager>(F); | |||
8684 | auto *LI = &AM.getResult<LoopAnalysis>(F); | |||
8685 | auto *DT = &AM.getResult<DominatorTreeAnalysis>(F); | |||
8686 | auto *AC = &AM.getResult<AssumptionAnalysis>(F); | |||
8687 | auto *DB = &AM.getResult<DemandedBitsAnalysis>(F); | |||
8688 | auto *ORE = &AM.getResult<OptimizationRemarkEmitterAnalysis>(F); | |||
8689 | ||||
8690 | bool Changed = runImpl(F, SE, TTI, TLI, AA, LI, DT, AC, DB, ORE); | |||
8691 | if (!Changed) | |||
8692 | return PreservedAnalyses::all(); | |||
8693 | ||||
8694 | PreservedAnalyses PA; | |||
8695 | PA.preserveSet<CFGAnalyses>(); | |||
8696 | return PA; | |||
8697 | } | |||
8698 | ||||
8699 | bool SLPVectorizerPass::runImpl(Function &F, ScalarEvolution *SE_, | |||
8700 | TargetTransformInfo *TTI_, | |||
8701 | TargetLibraryInfo *TLI_, AAResults *AA_, | |||
8702 | LoopInfo *LI_, DominatorTree *DT_, | |||
8703 | AssumptionCache *AC_, DemandedBits *DB_, | |||
8704 | OptimizationRemarkEmitter *ORE_) { | |||
8705 | if (!RunSLPVectorization) | |||
8706 | return false; | |||
8707 | SE = SE_; | |||
8708 | TTI = TTI_; | |||
8709 | TLI = TLI_; | |||
8710 | AA = AA_; | |||
8711 | LI = LI_; | |||
8712 | DT = DT_; | |||
8713 | AC = AC_; | |||
8714 | DB = DB_; | |||
8715 | DL = &F.getParent()->getDataLayout(); | |||
8716 | ||||
8717 | Stores.clear(); | |||
8718 | GEPs.clear(); | |||
8719 | bool Changed = false; | |||
8720 | ||||
8721 | // If the target claims to have no vector registers don't attempt | |||
8722 | // vectorization. | |||
8723 | if (!TTI->getNumberOfRegisters(TTI->getRegisterClassForType(true))) { | |||
8724 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Didn't find any vector registers for target, abort.\n" ; } } while (false) | |||
8725 | dbgs() << "SLP: Didn't find any vector registers for target, abort.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Didn't find any vector registers for target, abort.\n" ; } } while (false); | |||
8726 | return false; | |||
8727 | } | |||
8728 | ||||
8729 | // Don't vectorize when the attribute NoImplicitFloat is used. | |||
8730 | if (F.hasFnAttribute(Attribute::NoImplicitFloat)) | |||
8731 | return false; | |||
8732 | ||||
8733 | LLVM_DEBUG(dbgs() << "SLP: Analyzing blocks in " << F.getName() << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing blocks in " << F.getName() << ".\n"; } } while (false); | |||
8734 | ||||
8735 | // Use the bottom up slp vectorizer to construct chains that start with | |||
8736 | // store instructions. | |||
8737 | BoUpSLP R(&F, SE, TTI, TLI, AA, LI, DT, AC, DB, DL, ORE_); | |||
8738 | ||||
8739 | // A general note: the vectorizer must use BoUpSLP::eraseInstruction() to | |||
8740 | // delete instructions. | |||
8741 | ||||
8742 | // Update DFS numbers now so that we can use them for ordering. | |||
8743 | DT->updateDFSNumbers(); | |||
8744 | ||||
8745 | // Scan the blocks in the function in post order. | |||
8746 | for (auto BB : post_order(&F.getEntryBlock())) { | |||
8747 | collectSeedInstructions(BB); | |||
8748 | ||||
8749 | // Vectorize trees that end at stores. | |||
8750 | if (!Stores.empty()) { | |||
8751 | LLVM_DEBUG(dbgs() << "SLP: Found stores for " << Stores.size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Found stores for " << Stores .size() << " underlying objects.\n"; } } while (false) | |||
8752 | << " underlying objects.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Found stores for " << Stores .size() << " underlying objects.\n"; } } while (false); | |||
8753 | Changed |= vectorizeStoreChains(R); | |||
8754 | } | |||
8755 | ||||
8756 | // Vectorize trees that end at reductions. | |||
8757 | Changed |= vectorizeChainsInBlock(BB, R); | |||
8758 | ||||
8759 | // Vectorize the index computations of getelementptr instructions. This | |||
8760 | // is primarily intended to catch gather-like idioms ending at | |||
8761 | // non-consecutive loads. | |||
8762 | if (!GEPs.empty()) { | |||
8763 | LLVM_DEBUG(dbgs() << "SLP: Found GEPs for " << GEPs.size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Found GEPs for " << GEPs .size() << " underlying objects.\n"; } } while (false) | |||
8764 | << " underlying objects.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Found GEPs for " << GEPs .size() << " underlying objects.\n"; } } while (false); | |||
8765 | Changed |= vectorizeGEPIndices(BB, R); | |||
8766 | } | |||
8767 | } | |||
8768 | ||||
8769 | if (Changed) { | |||
8770 | R.optimizeGatherSequence(); | |||
8771 | LLVM_DEBUG(dbgs() << "SLP: vectorized \"" << F.getName() << "\"\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: vectorized \"" << F.getName () << "\"\n"; } } while (false); | |||
8772 | } | |||
8773 | return Changed; | |||
8774 | } | |||
8775 | ||||
8776 | bool SLPVectorizerPass::vectorizeStoreChain(ArrayRef<Value *> Chain, BoUpSLP &R, | |||
8777 | unsigned Idx) { | |||
8778 | LLVM_DEBUG(dbgs() << "SLP: Analyzing a store chain of length " << Chain.size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing a store chain of length " << Chain.size() << "\n"; } } while (false) | |||
8779 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing a store chain of length " << Chain.size() << "\n"; } } while (false); | |||
8780 | const unsigned Sz = R.getVectorElementSize(Chain[0]); | |||
8781 | const unsigned MinVF = R.getMinVecRegSize() / Sz; | |||
8782 | unsigned VF = Chain.size(); | |||
8783 | ||||
8784 | if (!isPowerOf2_32(Sz) || !isPowerOf2_32(VF) || VF < 2 || VF < MinVF) | |||
8785 | return false; | |||
8786 | ||||
8787 | LLVM_DEBUG(dbgs() << "SLP: Analyzing " << VF << " stores at offset " << Idxdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing " << VF << " stores at offset " << Idx << "\n"; } } while ( false) | |||
8788 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing " << VF << " stores at offset " << Idx << "\n"; } } while ( false); | |||
8789 | ||||
8790 | R.buildTree(Chain); | |||
8791 | if (R.isTreeTinyAndNotFullyVectorizable()) | |||
8792 | return false; | |||
8793 | if (R.isLoadCombineCandidate()) | |||
8794 | return false; | |||
8795 | R.reorderTopToBottom(); | |||
8796 | R.reorderBottomToTop(); | |||
8797 | R.buildExternalUses(); | |||
8798 | ||||
8799 | R.computeMinimumValueSizes(); | |||
8800 | ||||
8801 | InstructionCost Cost = R.getTreeCost(); | |||
8802 | ||||
8803 | LLVM_DEBUG(dbgs() << "SLP: Found cost = " << Cost << " for VF =" << VF << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Found cost = " << Cost << " for VF =" << VF << "\n"; } } while (false ); | |||
8804 | if (Cost < -SLPCostThreshold) { | |||
8805 | LLVM_DEBUG(dbgs() << "SLP: Decided to vectorize cost = " << Cost << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Decided to vectorize cost = " << Cost << "\n"; } } while (false); | |||
8806 | ||||
8807 | using namespace ore; | |||
8808 | ||||
8809 | R.getORE()->emit(OptimizationRemark(SV_NAME"slp-vectorizer", "StoresVectorized", | |||
8810 | cast<StoreInst>(Chain[0])) | |||
8811 | << "Stores SLP vectorized with cost " << NV("Cost", Cost) | |||
8812 | << " and with tree size " | |||
8813 | << NV("TreeSize", R.getTreeSize())); | |||
8814 | ||||
8815 | R.vectorizeTree(); | |||
8816 | return true; | |||
8817 | } | |||
8818 | ||||
8819 | return false; | |||
8820 | } | |||
8821 | ||||
8822 | bool SLPVectorizerPass::vectorizeStores(ArrayRef<StoreInst *> Stores, | |||
8823 | BoUpSLP &R) { | |||
8824 | // We may run into multiple chains that merge into a single chain. We mark the | |||
8825 | // stores that we vectorized so that we don't visit the same store twice. | |||
8826 | BoUpSLP::ValueSet VectorizedStores; | |||
8827 | bool Changed = false; | |||
8828 | ||||
8829 | int E = Stores.size(); | |||
8830 | SmallBitVector Tails(E, false); | |||
8831 | int MaxIter = MaxStoreLookup.getValue(); | |||
8832 | SmallVector<std::pair<int, int>, 16> ConsecutiveChain( | |||
8833 | E, std::make_pair(E, INT_MAX2147483647)); | |||
8834 | SmallVector<SmallBitVector, 4> CheckedPairs(E, SmallBitVector(E, false)); | |||
8835 | int IterCnt; | |||
8836 | auto &&FindConsecutiveAccess = [this, &Stores, &Tails, &IterCnt, MaxIter, | |||
8837 | &CheckedPairs, | |||
8838 | &ConsecutiveChain](int K, int Idx) { | |||
8839 | if (IterCnt >= MaxIter) | |||
8840 | return true; | |||
8841 | if (CheckedPairs[Idx].test(K)) | |||
8842 | return ConsecutiveChain[K].second == 1 && | |||
8843 | ConsecutiveChain[K].first == Idx; | |||
8844 | ++IterCnt; | |||
8845 | CheckedPairs[Idx].set(K); | |||
8846 | CheckedPairs[K].set(Idx); | |||
8847 | Optional<int> Diff = getPointersDiff( | |||
8848 | Stores[K]->getValueOperand()->getType(), Stores[K]->getPointerOperand(), | |||
8849 | Stores[Idx]->getValueOperand()->getType(), | |||
8850 | Stores[Idx]->getPointerOperand(), *DL, *SE, /*StrictCheck=*/true); | |||
8851 | if (!Diff || *Diff == 0) | |||
8852 | return false; | |||
8853 | int Val = *Diff; | |||
8854 | if (Val < 0) { | |||
8855 | if (ConsecutiveChain[Idx].second > -Val) { | |||
8856 | Tails.set(K); | |||
8857 | ConsecutiveChain[Idx] = std::make_pair(K, -Val); | |||
8858 | } | |||
8859 | return false; | |||
8860 | } | |||
8861 | if (ConsecutiveChain[K].second <= Val) | |||
8862 | return false; | |||
8863 | ||||
8864 | Tails.set(Idx); | |||
8865 | ConsecutiveChain[K] = std::make_pair(Idx, Val); | |||
8866 | return Val == 1; | |||
8867 | }; | |||
8868 | // Do a quadratic search on all of the given stores in reverse order and find | |||
8869 | // all of the pairs of stores that follow each other. | |||
8870 | for (int Idx = E - 1; Idx >= 0; --Idx) { | |||
8871 | // If a store has multiple consecutive store candidates, search according | |||
8872 | // to the sequence: Idx-1, Idx+1, Idx-2, Idx+2, ... | |||
8873 | // This is because usually pairing with immediate succeeding or preceding | |||
8874 | // candidate create the best chance to find slp vectorization opportunity. | |||
8875 | const int MaxLookDepth = std::max(E - Idx, Idx + 1); | |||
8876 | IterCnt = 0; | |||
8877 | for (int Offset = 1, F = MaxLookDepth; Offset < F; ++Offset) | |||
8878 | if ((Idx >= Offset && FindConsecutiveAccess(Idx - Offset, Idx)) || | |||
8879 | (Idx + Offset < E && FindConsecutiveAccess(Idx + Offset, Idx))) | |||
8880 | break; | |||
8881 | } | |||
8882 | ||||
8883 | // Tracks if we tried to vectorize stores starting from the given tail | |||
8884 | // already. | |||
8885 | SmallBitVector TriedTails(E, false); | |||
8886 | // For stores that start but don't end a link in the chain: | |||
8887 | for (int Cnt = E; Cnt > 0; --Cnt) { | |||
8888 | int I = Cnt - 1; | |||
8889 | if (ConsecutiveChain[I].first == E || Tails.test(I)) | |||
8890 | continue; | |||
8891 | // We found a store instr that starts a chain. Now follow the chain and try | |||
8892 | // to vectorize it. | |||
8893 | BoUpSLP::ValueList Operands; | |||
8894 | // Collect the chain into a list. | |||
8895 | while (I != E && !VectorizedStores.count(Stores[I])) { | |||
8896 | Operands.push_back(Stores[I]); | |||
8897 | Tails.set(I); | |||
8898 | if (ConsecutiveChain[I].second != 1) { | |||
8899 | // Mark the new end in the chain and go back, if required. It might be | |||
8900 | // required if the original stores come in reversed order, for example. | |||
8901 | if (ConsecutiveChain[I].first != E && | |||
8902 | Tails.test(ConsecutiveChain[I].first) && !TriedTails.test(I) && | |||
8903 | !VectorizedStores.count(Stores[ConsecutiveChain[I].first])) { | |||
8904 | TriedTails.set(I); | |||
8905 | Tails.reset(ConsecutiveChain[I].first); | |||
8906 | if (Cnt < ConsecutiveChain[I].first + 2) | |||
8907 | Cnt = ConsecutiveChain[I].first + 2; | |||
8908 | } | |||
8909 | break; | |||
8910 | } | |||
8911 | // Move to the next value in the chain. | |||
8912 | I = ConsecutiveChain[I].first; | |||
8913 | } | |||
8914 | assert(!Operands.empty() && "Expected non-empty list of stores.")(static_cast <bool> (!Operands.empty() && "Expected non-empty list of stores." ) ? void (0) : __assert_fail ("!Operands.empty() && \"Expected non-empty list of stores.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8914, __extension__ __PRETTY_FUNCTION__)); | |||
8915 | ||||
8916 | unsigned MaxVecRegSize = R.getMaxVecRegSize(); | |||
8917 | unsigned EltSize = R.getVectorElementSize(Operands[0]); | |||
8918 | unsigned MaxElts = llvm::PowerOf2Floor(MaxVecRegSize / EltSize); | |||
8919 | ||||
8920 | unsigned MinVF = R.getMinVF(EltSize); | |||
8921 | unsigned MaxVF = std::min(R.getMaximumVF(EltSize, Instruction::Store), | |||
8922 | MaxElts); | |||
8923 | ||||
8924 | // FIXME: Is division-by-2 the correct step? Should we assert that the | |||
8925 | // register size is a power-of-2? | |||
8926 | unsigned StartIdx = 0; | |||
8927 | for (unsigned Size = MaxVF; Size >= MinVF; Size /= 2) { | |||
8928 | for (unsigned Cnt = StartIdx, E = Operands.size(); Cnt + Size <= E;) { | |||
8929 | ArrayRef<Value *> Slice = makeArrayRef(Operands).slice(Cnt, Size); | |||
8930 | if (!VectorizedStores.count(Slice.front()) && | |||
8931 | !VectorizedStores.count(Slice.back()) && | |||
8932 | vectorizeStoreChain(Slice, R, Cnt)) { | |||
8933 | // Mark the vectorized stores so that we don't vectorize them again. | |||
8934 | VectorizedStores.insert(Slice.begin(), Slice.end()); | |||
8935 | Changed = true; | |||
8936 | // If we vectorized initial block, no need to try to vectorize it | |||
8937 | // again. | |||
8938 | if (Cnt == StartIdx) | |||
8939 | StartIdx += Size; | |||
8940 | Cnt += Size; | |||
8941 | continue; | |||
8942 | } | |||
8943 | ++Cnt; | |||
8944 | } | |||
8945 | // Check if the whole array was vectorized already - exit. | |||
8946 | if (StartIdx >= Operands.size()) | |||
8947 | break; | |||
8948 | } | |||
8949 | } | |||
8950 | ||||
8951 | return Changed; | |||
8952 | } | |||
8953 | ||||
8954 | void SLPVectorizerPass::collectSeedInstructions(BasicBlock *BB) { | |||
8955 | // Initialize the collections. We will make a single pass over the block. | |||
8956 | Stores.clear(); | |||
8957 | GEPs.clear(); | |||
8958 | ||||
8959 | // Visit the store and getelementptr instructions in BB and organize them in | |||
8960 | // Stores and GEPs according to the underlying objects of their pointer | |||
8961 | // operands. | |||
8962 | for (Instruction &I : *BB) { | |||
8963 | // Ignore store instructions that are volatile or have a pointer operand | |||
8964 | // that doesn't point to a scalar type. | |||
8965 | if (auto *SI = dyn_cast<StoreInst>(&I)) { | |||
8966 | if (!SI->isSimple()) | |||
8967 | continue; | |||
8968 | if (!isValidElementType(SI->getValueOperand()->getType())) | |||
8969 | continue; | |||
8970 | Stores[getUnderlyingObject(SI->getPointerOperand())].push_back(SI); | |||
8971 | } | |||
8972 | ||||
8973 | // Ignore getelementptr instructions that have more than one index, a | |||
8974 | // constant index, or a pointer operand that doesn't point to a scalar | |||
8975 | // type. | |||
8976 | else if (auto *GEP = dyn_cast<GetElementPtrInst>(&I)) { | |||
8977 | auto Idx = GEP->idx_begin()->get(); | |||
8978 | if (GEP->getNumIndices() > 1 || isa<Constant>(Idx)) | |||
8979 | continue; | |||
8980 | if (!isValidElementType(Idx->getType())) | |||
8981 | continue; | |||
8982 | if (GEP->getType()->isVectorTy()) | |||
8983 | continue; | |||
8984 | GEPs[GEP->getPointerOperand()].push_back(GEP); | |||
8985 | } | |||
8986 | } | |||
8987 | } | |||
8988 | ||||
8989 | bool SLPVectorizerPass::tryToVectorizePair(Value *A, Value *B, BoUpSLP &R) { | |||
8990 | if (!A || !B) | |||
8991 | return false; | |||
8992 | if (isa<InsertElementInst>(A) || isa<InsertElementInst>(B)) | |||
8993 | return false; | |||
8994 | Value *VL[] = {A, B}; | |||
8995 | return tryToVectorizeList(VL, R); | |||
8996 | } | |||
8997 | ||||
8998 | bool SLPVectorizerPass::tryToVectorizeList(ArrayRef<Value *> VL, BoUpSLP &R, | |||
8999 | bool LimitForRegisterSize) { | |||
9000 | if (VL.size() < 2) | |||
9001 | return false; | |||
9002 | ||||
9003 | LLVM_DEBUG(dbgs() << "SLP: Trying to vectorize a list of length = "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Trying to vectorize a list of length = " << VL.size() << ".\n"; } } while (false) | |||
9004 | << VL.size() << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Trying to vectorize a list of length = " << VL.size() << ".\n"; } } while (false); | |||
9005 | ||||
9006 | // Check that all of the parts are instructions of the same type, | |||
9007 | // we permit an alternate opcode via InstructionsState. | |||
9008 | InstructionsState S = getSameOpcode(VL); | |||
9009 | if (!S.getOpcode()) | |||
9010 | return false; | |||
9011 | ||||
9012 | Instruction *I0 = cast<Instruction>(S.OpValue); | |||
9013 | // Make sure invalid types (including vector type) are rejected before | |||
9014 | // determining vectorization factor for scalar instructions. | |||
9015 | for (Value *V : VL) { | |||
9016 | Type *Ty = V->getType(); | |||
9017 | if (!isa<InsertElementInst>(V) && !isValidElementType(Ty)) { | |||
9018 | // NOTE: the following will give user internal llvm type name, which may | |||
9019 | // not be useful. | |||
9020 | R.getORE()->emit([&]() { | |||
9021 | std::string type_str; | |||
9022 | llvm::raw_string_ostream rso(type_str); | |||
9023 | Ty->print(rso); | |||
9024 | return OptimizationRemarkMissed(SV_NAME"slp-vectorizer", "UnsupportedType", I0) | |||
9025 | << "Cannot SLP vectorize list: type " | |||
9026 | << rso.str() + " is unsupported by vectorizer"; | |||
9027 | }); | |||
9028 | return false; | |||
9029 | } | |||
9030 | } | |||
9031 | ||||
9032 | unsigned Sz = R.getVectorElementSize(I0); | |||
9033 | unsigned MinVF = R.getMinVF(Sz); | |||
9034 | unsigned MaxVF = std::max<unsigned>(PowerOf2Floor(VL.size()), MinVF); | |||
9035 | MaxVF = std::min(R.getMaximumVF(Sz, S.getOpcode()), MaxVF); | |||
9036 | if (MaxVF < 2) { | |||
9037 | R.getORE()->emit([&]() { | |||
9038 | return OptimizationRemarkMissed(SV_NAME"slp-vectorizer", "SmallVF", I0) | |||
9039 | << "Cannot SLP vectorize list: vectorization factor " | |||
9040 | << "less than 2 is not supported"; | |||
9041 | }); | |||
9042 | return false; | |||
9043 | } | |||
9044 | ||||
9045 | bool Changed = false; | |||
9046 | bool CandidateFound = false; | |||
9047 | InstructionCost MinCost = SLPCostThreshold.getValue(); | |||
9048 | Type *ScalarTy = VL[0]->getType(); | |||
9049 | if (auto *IE = dyn_cast<InsertElementInst>(VL[0])) | |||
9050 | ScalarTy = IE->getOperand(1)->getType(); | |||
9051 | ||||
9052 | unsigned NextInst = 0, MaxInst = VL.size(); | |||
9053 | for (unsigned VF = MaxVF; NextInst + 1 < MaxInst && VF >= MinVF; VF /= 2) { | |||
9054 | // No actual vectorization should happen, if number of parts is the same as | |||
9055 | // provided vectorization factor (i.e. the scalar type is used for vector | |||
9056 | // code during codegen). | |||
9057 | auto *VecTy = FixedVectorType::get(ScalarTy, VF); | |||
9058 | if (TTI->getNumberOfParts(VecTy) == VF) | |||
9059 | continue; | |||
9060 | for (unsigned I = NextInst; I < MaxInst; ++I) { | |||
9061 | unsigned OpsWidth = 0; | |||
9062 | ||||
9063 | if (I + VF > MaxInst) | |||
9064 | OpsWidth = MaxInst - I; | |||
9065 | else | |||
9066 | OpsWidth = VF; | |||
9067 | ||||
9068 | if (!isPowerOf2_32(OpsWidth)) | |||
9069 | continue; | |||
9070 | ||||
9071 | if ((LimitForRegisterSize && OpsWidth < MaxVF) || | |||
9072 | (VF > MinVF && OpsWidth <= VF / 2) || (VF == MinVF && OpsWidth < 2)) | |||
9073 | break; | |||
9074 | ||||
9075 | ArrayRef<Value *> Ops = VL.slice(I, OpsWidth); | |||
9076 | // Check that a previous iteration of this loop did not delete the Value. | |||
9077 | if (llvm::any_of(Ops, [&R](Value *V) { | |||
9078 | auto *I = dyn_cast<Instruction>(V); | |||
9079 | return I && R.isDeleted(I); | |||
9080 | })) | |||
9081 | continue; | |||
9082 | ||||
9083 | LLVM_DEBUG(dbgs() << "SLP: Analyzing " << OpsWidth << " operations "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing " << OpsWidth << " operations " << "\n"; } } while (false) | |||
9084 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing " << OpsWidth << " operations " << "\n"; } } while (false); | |||
9085 | ||||
9086 | R.buildTree(Ops); | |||
9087 | if (R.isTreeTinyAndNotFullyVectorizable()) | |||
9088 | continue; | |||
9089 | R.reorderTopToBottom(); | |||
9090 | R.reorderBottomToTop(!isa<InsertElementInst>(Ops.front())); | |||
9091 | R.buildExternalUses(); | |||
9092 | ||||
9093 | R.computeMinimumValueSizes(); | |||
9094 | InstructionCost Cost = R.getTreeCost(); | |||
9095 | CandidateFound = true; | |||
9096 | MinCost = std::min(MinCost, Cost); | |||
9097 | ||||
9098 | if (Cost < -SLPCostThreshold) { | |||
9099 | LLVM_DEBUG(dbgs() << "SLP: Vectorizing list at cost:" << Cost << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Vectorizing list at cost:" << Cost << ".\n"; } } while (false); | |||
9100 | R.getORE()->emit(OptimizationRemark(SV_NAME"slp-vectorizer", "VectorizedList", | |||
9101 | cast<Instruction>(Ops[0])) | |||
9102 | << "SLP vectorized with cost " << ore::NV("Cost", Cost) | |||
9103 | << " and with tree size " | |||
9104 | << ore::NV("TreeSize", R.getTreeSize())); | |||
9105 | ||||
9106 | R.vectorizeTree(); | |||
9107 | // Move to the next bundle. | |||
9108 | I += VF - 1; | |||
9109 | NextInst = I + 1; | |||
9110 | Changed = true; | |||
9111 | } | |||
9112 | } | |||
9113 | } | |||
9114 | ||||
9115 | if (!Changed && CandidateFound) { | |||
9116 | R.getORE()->emit([&]() { | |||
9117 | return OptimizationRemarkMissed(SV_NAME"slp-vectorizer", "NotBeneficial", I0) | |||
9118 | << "List vectorization was possible but not beneficial with cost " | |||
9119 | << ore::NV("Cost", MinCost) << " >= " | |||
9120 | << ore::NV("Treshold", -SLPCostThreshold); | |||
9121 | }); | |||
9122 | } else if (!Changed) { | |||
9123 | R.getORE()->emit([&]() { | |||
9124 | return OptimizationRemarkMissed(SV_NAME"slp-vectorizer", "NotPossible", I0) | |||
9125 | << "Cannot SLP vectorize list: vectorization was impossible" | |||
9126 | << " with available vectorization factors"; | |||
9127 | }); | |||
9128 | } | |||
9129 | return Changed; | |||
9130 | } | |||
9131 | ||||
9132 | bool SLPVectorizerPass::tryToVectorize(Instruction *I, BoUpSLP &R) { | |||
9133 | if (!I) | |||
9134 | return false; | |||
9135 | ||||
9136 | if (!isa<BinaryOperator>(I) && !isa<CmpInst>(I)) | |||
9137 | return false; | |||
9138 | ||||
9139 | Value *P = I->getParent(); | |||
9140 | ||||
9141 | // Vectorize in current basic block only. | |||
9142 | auto *Op0 = dyn_cast<Instruction>(I->getOperand(0)); | |||
9143 | auto *Op1 = dyn_cast<Instruction>(I->getOperand(1)); | |||
9144 | if (!Op0 || !Op1 || Op0->getParent() != P || Op1->getParent() != P) | |||
9145 | return false; | |||
9146 | ||||
9147 | // Try to vectorize V. | |||
9148 | if (tryToVectorizePair(Op0, Op1, R)) | |||
9149 | return true; | |||
9150 | ||||
9151 | auto *A = dyn_cast<BinaryOperator>(Op0); | |||
9152 | auto *B = dyn_cast<BinaryOperator>(Op1); | |||
9153 | // Try to skip B. | |||
9154 | if (B && B->hasOneUse()) { | |||
9155 | auto *B0 = dyn_cast<BinaryOperator>(B->getOperand(0)); | |||
9156 | auto *B1 = dyn_cast<BinaryOperator>(B->getOperand(1)); | |||
9157 | if (B0 && B0->getParent() == P && tryToVectorizePair(A, B0, R)) | |||
9158 | return true; | |||
9159 | if (B1 && B1->getParent() == P && tryToVectorizePair(A, B1, R)) | |||
9160 | return true; | |||
9161 | } | |||
9162 | ||||
9163 | // Try to skip A. | |||
9164 | if (A && A->hasOneUse()) { | |||
9165 | auto *A0 = dyn_cast<BinaryOperator>(A->getOperand(0)); | |||
9166 | auto *A1 = dyn_cast<BinaryOperator>(A->getOperand(1)); | |||
9167 | if (A0 && A0->getParent() == P && tryToVectorizePair(A0, B, R)) | |||
9168 | return true; | |||
9169 | if (A1 && A1->getParent() == P && tryToVectorizePair(A1, B, R)) | |||
9170 | return true; | |||
9171 | } | |||
9172 | return false; | |||
9173 | } | |||
9174 | ||||
9175 | namespace { | |||
9176 | ||||
9177 | /// Model horizontal reductions. | |||
9178 | /// | |||
9179 | /// A horizontal reduction is a tree of reduction instructions that has values | |||
9180 | /// that can be put into a vector as its leaves. For example: | |||
9181 | /// | |||
9182 | /// mul mul mul mul | |||
9183 | /// \ / \ / | |||
9184 | /// + + | |||
9185 | /// \ / | |||
9186 | /// + | |||
9187 | /// This tree has "mul" as its leaf values and "+" as its reduction | |||
9188 | /// instructions. A reduction can feed into a store or a binary operation | |||
9189 | /// feeding a phi. | |||
9190 | /// ... | |||
9191 | /// \ / | |||
9192 | /// + | |||
9193 | /// | | |||
9194 | /// phi += | |||
9195 | /// | |||
9196 | /// Or: | |||
9197 | /// ... | |||
9198 | /// \ / | |||
9199 | /// + | |||
9200 | /// | | |||
9201 | /// *p = | |||
9202 | /// | |||
9203 | class HorizontalReduction { | |||
9204 | using ReductionOpsType = SmallVector<Value *, 16>; | |||
9205 | using ReductionOpsListType = SmallVector<ReductionOpsType, 2>; | |||
9206 | ReductionOpsListType ReductionOps; | |||
9207 | SmallVector<Value *, 32> ReducedVals; | |||
9208 | // Use map vector to make stable output. | |||
9209 | MapVector<Instruction *, Value *> ExtraArgs; | |||
9210 | WeakTrackingVH ReductionRoot; | |||
9211 | /// The type of reduction operation. | |||
9212 | RecurKind RdxKind; | |||
9213 | ||||
9214 | const unsigned INVALID_OPERAND_INDEX = std::numeric_limits<unsigned>::max(); | |||
9215 | ||||
9216 | static bool isCmpSelMinMax(Instruction *I) { | |||
9217 | return match(I, m_Select(m_Cmp(), m_Value(), m_Value())) && | |||
9218 | RecurrenceDescriptor::isMinMaxRecurrenceKind(getRdxKind(I)); | |||
9219 | } | |||
9220 | ||||
9221 | // And/or are potentially poison-safe logical patterns like: | |||
9222 | // select x, y, false | |||
9223 | // select x, true, y | |||
9224 | static bool isBoolLogicOp(Instruction *I) { | |||
9225 | return match(I, m_LogicalAnd(m_Value(), m_Value())) || | |||
9226 | match(I, m_LogicalOr(m_Value(), m_Value())); | |||
9227 | } | |||
9228 | ||||
9229 | /// Checks if instruction is associative and can be vectorized. | |||
9230 | static bool isVectorizable(RecurKind Kind, Instruction *I) { | |||
9231 | if (Kind == RecurKind::None) | |||
9232 | return false; | |||
9233 | ||||
9234 | // Integer ops that map to select instructions or intrinsics are fine. | |||
9235 | if (RecurrenceDescriptor::isIntMinMaxRecurrenceKind(Kind) || | |||
9236 | isBoolLogicOp(I)) | |||
9237 | return true; | |||
9238 | ||||
9239 | if (Kind == RecurKind::FMax || Kind == RecurKind::FMin) { | |||
9240 | // FP min/max are associative except for NaN and -0.0. We do not | |||
9241 | // have to rule out -0.0 here because the intrinsic semantics do not | |||
9242 | // specify a fixed result for it. | |||
9243 | return I->getFastMathFlags().noNaNs(); | |||
9244 | } | |||
9245 | ||||
9246 | return I->isAssociative(); | |||
9247 | } | |||
9248 | ||||
9249 | static Value *getRdxOperand(Instruction *I, unsigned Index) { | |||
9250 | // Poison-safe 'or' takes the form: select X, true, Y | |||
9251 | // To make that work with the normal operand processing, we skip the | |||
9252 | // true value operand. | |||
9253 | // TODO: Change the code and data structures to handle this without a hack. | |||
9254 | if (getRdxKind(I) == RecurKind::Or && isa<SelectInst>(I) && Index == 1) | |||
9255 | return I->getOperand(2); | |||
9256 | return I->getOperand(Index); | |||
9257 | } | |||
9258 | ||||
9259 | /// Checks if the ParentStackElem.first should be marked as a reduction | |||
9260 | /// operation with an extra argument or as extra argument itself. | |||
9261 | void markExtraArg(std::pair<Instruction *, unsigned> &ParentStackElem, | |||
9262 | Value *ExtraArg) { | |||
9263 | if (ExtraArgs.count(ParentStackElem.first)) { | |||
9264 | ExtraArgs[ParentStackElem.first] = nullptr; | |||
9265 | // We ran into something like: | |||
9266 | // ParentStackElem.first = ExtraArgs[ParentStackElem.first] + ExtraArg. | |||
9267 | // The whole ParentStackElem.first should be considered as an extra value | |||
9268 | // in this case. | |||
9269 | // Do not perform analysis of remaining operands of ParentStackElem.first | |||
9270 | // instruction, this whole instruction is an extra argument. | |||
9271 | ParentStackElem.second = INVALID_OPERAND_INDEX; | |||
9272 | } else { | |||
9273 | // We ran into something like: | |||
9274 | // ParentStackElem.first += ... + ExtraArg + ... | |||
9275 | ExtraArgs[ParentStackElem.first] = ExtraArg; | |||
9276 | } | |||
9277 | } | |||
9278 | ||||
9279 | /// Creates reduction operation with the current opcode. | |||
9280 | static Value *createOp(IRBuilder<> &Builder, RecurKind Kind, Value *LHS, | |||
9281 | Value *RHS, const Twine &Name, bool UseSelect) { | |||
9282 | unsigned RdxOpcode = RecurrenceDescriptor::getOpcode(Kind); | |||
9283 | switch (Kind) { | |||
9284 | case RecurKind::Or: | |||
9285 | if (UseSelect && | |||
9286 | LHS->getType() == CmpInst::makeCmpResultType(LHS->getType())) | |||
9287 | return Builder.CreateSelect(LHS, Builder.getTrue(), RHS, Name); | |||
9288 | return Builder.CreateBinOp((Instruction::BinaryOps)RdxOpcode, LHS, RHS, | |||
9289 | Name); | |||
9290 | case RecurKind::And: | |||
9291 | if (UseSelect && | |||
9292 | LHS->getType() == CmpInst::makeCmpResultType(LHS->getType())) | |||
9293 | return Builder.CreateSelect(LHS, RHS, Builder.getFalse(), Name); | |||
9294 | return Builder.CreateBinOp((Instruction::BinaryOps)RdxOpcode, LHS, RHS, | |||
9295 | Name); | |||
9296 | case RecurKind::Add: | |||
9297 | case RecurKind::Mul: | |||
9298 | case RecurKind::Xor: | |||
9299 | case RecurKind::FAdd: | |||
9300 | case RecurKind::FMul: | |||
9301 | return Builder.CreateBinOp((Instruction::BinaryOps)RdxOpcode, LHS, RHS, | |||
9302 | Name); | |||
9303 | case RecurKind::FMax: | |||
9304 | return Builder.CreateBinaryIntrinsic(Intrinsic::maxnum, LHS, RHS); | |||
9305 | case RecurKind::FMin: | |||
9306 | return Builder.CreateBinaryIntrinsic(Intrinsic::minnum, LHS, RHS); | |||
9307 | case RecurKind::SMax: | |||
9308 | if (UseSelect) { | |||
9309 | Value *Cmp = Builder.CreateICmpSGT(LHS, RHS, Name); | |||
9310 | return Builder.CreateSelect(Cmp, LHS, RHS, Name); | |||
9311 | } | |||
9312 | return Builder.CreateBinaryIntrinsic(Intrinsic::smax, LHS, RHS); | |||
9313 | case RecurKind::SMin: | |||
9314 | if (UseSelect) { | |||
9315 | Value *Cmp = Builder.CreateICmpSLT(LHS, RHS, Name); | |||
9316 | return Builder.CreateSelect(Cmp, LHS, RHS, Name); | |||
9317 | } | |||
9318 | return Builder.CreateBinaryIntrinsic(Intrinsic::smin, LHS, RHS); | |||
9319 | case RecurKind::UMax: | |||
9320 | if (UseSelect) { | |||
9321 | Value *Cmp = Builder.CreateICmpUGT(LHS, RHS, Name); | |||
9322 | return Builder.CreateSelect(Cmp, LHS, RHS, Name); | |||
9323 | } | |||
9324 | return Builder.CreateBinaryIntrinsic(Intrinsic::umax, LHS, RHS); | |||
9325 | case RecurKind::UMin: | |||
9326 | if (UseSelect) { | |||
9327 | Value *Cmp = Builder.CreateICmpULT(LHS, RHS, Name); | |||
9328 | return Builder.CreateSelect(Cmp, LHS, RHS, Name); | |||
9329 | } | |||
9330 | return Builder.CreateBinaryIntrinsic(Intrinsic::umin, LHS, RHS); | |||
9331 | default: | |||
9332 | llvm_unreachable("Unknown reduction operation.")::llvm::llvm_unreachable_internal("Unknown reduction operation." , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 9332); | |||
9333 | } | |||
9334 | } | |||
9335 | ||||
9336 | /// Creates reduction operation with the current opcode with the IR flags | |||
9337 | /// from \p ReductionOps. | |||
9338 | static Value *createOp(IRBuilder<> &Builder, RecurKind RdxKind, Value *LHS, | |||
9339 | Value *RHS, const Twine &Name, | |||
9340 | const ReductionOpsListType &ReductionOps) { | |||
9341 | bool UseSelect = ReductionOps.size() == 2 || | |||
9342 | // Logical or/and. | |||
9343 | (ReductionOps.size() == 1 && | |||
9344 | isa<SelectInst>(ReductionOps.front().front())); | |||
9345 | assert((!UseSelect || ReductionOps.size() != 2 ||(static_cast <bool> ((!UseSelect || ReductionOps.size() != 2 || isa<SelectInst>(ReductionOps[1][0])) && "Expected cmp + select pairs for reduction") ? void (0) : __assert_fail ("(!UseSelect || ReductionOps.size() != 2 || isa<SelectInst>(ReductionOps[1][0])) && \"Expected cmp + select pairs for reduction\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 9347, __extension__ __PRETTY_FUNCTION__)) | |||
9346 | isa<SelectInst>(ReductionOps[1][0])) &&(static_cast <bool> ((!UseSelect || ReductionOps.size() != 2 || isa<SelectInst>(ReductionOps[1][0])) && "Expected cmp + select pairs for reduction") ? void (0) : __assert_fail ("(!UseSelect || ReductionOps.size() != 2 || isa<SelectInst>(ReductionOps[1][0])) && \"Expected cmp + select pairs for reduction\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 9347, __extension__ __PRETTY_FUNCTION__)) | |||
9347 | "Expected cmp + select pairs for reduction")(static_cast <bool> ((!UseSelect || ReductionOps.size() != 2 || isa<SelectInst>(ReductionOps[1][0])) && "Expected cmp + select pairs for reduction") ? void (0) : __assert_fail ("(!UseSelect || ReductionOps.size() != 2 || isa<SelectInst>(ReductionOps[1][0])) && \"Expected cmp + select pairs for reduction\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 9347, __extension__ __PRETTY_FUNCTION__)); | |||
9348 | Value *Op = createOp(Builder, RdxKind, LHS, RHS, Name, UseSelect); | |||
9349 | if (RecurrenceDescriptor::isIntMinMaxRecurrenceKind(RdxKind)) { | |||
9350 | if (auto *Sel = dyn_cast<SelectInst>(Op)) { | |||
9351 | propagateIRFlags(Sel->getCondition(), ReductionOps[0]); | |||
9352 | propagateIRFlags(Op, ReductionOps[1]); | |||
9353 | return Op; | |||
9354 | } | |||
9355 | } | |||
9356 | propagateIRFlags(Op, ReductionOps[0]); | |||
9357 | return Op; | |||
9358 | } | |||
9359 | ||||
9360 | /// Creates reduction operation with the current opcode with the IR flags | |||
9361 | /// from \p I. | |||
9362 | static Value *createOp(IRBuilder<> &Builder, RecurKind RdxKind, Value *LHS, | |||
9363 | Value *RHS, const Twine &Name, Instruction *I) { | |||
9364 | auto *SelI = dyn_cast<SelectInst>(I); | |||
9365 | Value *Op = createOp(Builder, RdxKind, LHS, RHS, Name, SelI != nullptr); | |||
9366 | if (SelI && RecurrenceDescriptor::isIntMinMaxRecurrenceKind(RdxKind)) { | |||
9367 | if (auto *Sel = dyn_cast<SelectInst>(Op)) | |||
9368 | propagateIRFlags(Sel->getCondition(), SelI->getCondition()); | |||
9369 | } | |||
9370 | propagateIRFlags(Op, I); | |||
9371 | return Op; | |||
9372 | } | |||
9373 | ||||
9374 | static RecurKind getRdxKind(Instruction *I) { | |||
9375 | assert(I && "Expected instruction for reduction matching")(static_cast <bool> (I && "Expected instruction for reduction matching" ) ? void (0) : __assert_fail ("I && \"Expected instruction for reduction matching\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 9375, __extension__ __PRETTY_FUNCTION__)); | |||
9376 | if (match(I, m_Add(m_Value(), m_Value()))) | |||
9377 | return RecurKind::Add; | |||
9378 | if (match(I, m_Mul(m_Value(), m_Value()))) | |||
9379 | return RecurKind::Mul; | |||
9380 | if (match(I, m_And(m_Value(), m_Value())) || | |||
9381 | match(I, m_LogicalAnd(m_Value(), m_Value()))) | |||
9382 | return RecurKind::And; | |||
9383 | if (match(I, m_Or(m_Value(), m_Value())) || | |||
9384 | match(I, m_LogicalOr(m_Value(), m_Value()))) | |||
9385 | return RecurKind::Or; | |||
9386 | if (match(I, m_Xor(m_Value(), m_Value()))) | |||
9387 | return RecurKind::Xor; | |||
9388 | if (match(I, m_FAdd(m_Value(), m_Value()))) | |||
9389 | return RecurKind::FAdd; | |||
9390 | if (match(I, m_FMul(m_Value(), m_Value()))) | |||
9391 | return RecurKind::FMul; | |||
9392 | ||||
9393 | if (match(I, m_Intrinsic<Intrinsic::maxnum>(m_Value(), m_Value()))) | |||
9394 | return RecurKind::FMax; | |||
9395 | if (match(I, m_Intrinsic<Intrinsic::minnum>(m_Value(), m_Value()))) | |||
9396 | return RecurKind::FMin; | |||
9397 | ||||
9398 | // This matches either cmp+select or intrinsics. SLP is expected to handle | |||
9399 | // either form. | |||
9400 | // TODO: If we are canonicalizing to intrinsics, we can remove several | |||
9401 | // special-case paths that deal with selects. | |||
9402 | if (match(I, m_SMax(m_Value(), m_Value()))) | |||
9403 | return RecurKind::SMax; | |||
9404 | if (match(I, m_SMin(m_Value(), m_Value()))) | |||
9405 | return RecurKind::SMin; | |||
9406 | if (match(I, m_UMax(m_Value(), m_Value()))) | |||
9407 | return RecurKind::UMax; | |||
9408 | if (match(I, m_UMin(m_Value(), m_Value()))) | |||
9409 | return RecurKind::UMin; | |||
9410 | ||||
9411 | if (auto *Select = dyn_cast<SelectInst>(I)) { | |||
9412 | // Try harder: look for min/max pattern based on instructions producing | |||
9413 | // same values such as: select ((cmp Inst1, Inst2), Inst1, Inst2). | |||
9414 | // During the intermediate stages of SLP, it's very common to have | |||
9415 | // pattern like this (since optimizeGatherSequence is run only once | |||
9416 | // at the end): | |||
9417 | // %1 = extractelement <2 x i32> %a, i32 0 | |||
9418 | // %2 = extractelement <2 x i32> %a, i32 1 | |||
9419 | // %cond = icmp sgt i32 %1, %2 | |||
9420 | // %3 = extractelement <2 x i32> %a, i32 0 | |||
9421 | // %4 = extractelement <2 x i32> %a, i32 1 | |||
9422 | // %select = select i1 %cond, i32 %3, i32 %4 | |||
9423 | CmpInst::Predicate Pred; | |||
9424 | Instruction *L1; | |||
9425 | Instruction *L2; | |||
9426 | ||||
9427 | Value *LHS = Select->getTrueValue(); | |||
9428 | Value *RHS = Select->getFalseValue(); | |||
9429 | Value *Cond = Select->getCondition(); | |||
9430 | ||||
9431 | // TODO: Support inverse predicates. | |||
9432 | if (match(Cond, m_Cmp(Pred, m_Specific(LHS), m_Instruction(L2)))) { | |||
9433 | if (!isa<ExtractElementInst>(RHS) || | |||
9434 | !L2->isIdenticalTo(cast<Instruction>(RHS))) | |||
9435 | return RecurKind::None; | |||
9436 | } else if (match(Cond, m_Cmp(Pred, m_Instruction(L1), m_Specific(RHS)))) { | |||
9437 | if (!isa<ExtractElementInst>(LHS) || | |||
9438 | !L1->isIdenticalTo(cast<Instruction>(LHS))) | |||
9439 | return RecurKind::None; | |||
9440 | } else { | |||
9441 | if (!isa<ExtractElementInst>(LHS) || !isa<ExtractElementInst>(RHS)) | |||
9442 | return RecurKind::None; | |||
9443 | if (!match(Cond, m_Cmp(Pred, m_Instruction(L1), m_Instruction(L2))) || | |||
9444 | !L1->isIdenticalTo(cast<Instruction>(LHS)) || | |||
9445 | !L2->isIdenticalTo(cast<Instruction>(RHS))) | |||
9446 | return RecurKind::None; | |||
9447 | } | |||
9448 | ||||
9449 | switch (Pred) { | |||
9450 | default: | |||
9451 | return RecurKind::None; | |||
9452 | case CmpInst::ICMP_SGT: | |||
9453 | case CmpInst::ICMP_SGE: | |||
9454 | return RecurKind::SMax; | |||
9455 | case CmpInst::ICMP_SLT: | |||
9456 | case CmpInst::ICMP_SLE: | |||
9457 | return RecurKind::SMin; | |||
9458 | case CmpInst::ICMP_UGT: | |||
9459 | case CmpInst::ICMP_UGE: | |||
9460 | return RecurKind::UMax; | |||
9461 | case CmpInst::ICMP_ULT: | |||
9462 | case CmpInst::ICMP_ULE: | |||
9463 | return RecurKind::UMin; | |||
9464 | } | |||
9465 | } | |||
9466 | return RecurKind::None; | |||
9467 | } | |||
9468 | ||||
9469 | /// Get the index of the first operand. | |||
9470 | static unsigned getFirstOperandIndex(Instruction *I) { | |||
9471 | return isCmpSelMinMax(I) ? 1 : 0; | |||
9472 | } | |||
9473 | ||||
9474 | /// Total number of operands in the reduction operation. | |||
9475 | static unsigned getNumberOfOperands(Instruction *I) { | |||
9476 | return isCmpSelMinMax(I) ? 3 : 2; | |||
9477 | } | |||
9478 | ||||
9479 | /// Checks if the instruction is in basic block \p BB. | |||
9480 | /// For a cmp+sel min/max reduction check that both ops are in \p BB. | |||
9481 | static bool hasSameParent(Instruction *I, BasicBlock *BB) { | |||
9482 | if (isCmpSelMinMax(I) || (isBoolLogicOp(I) && isa<SelectInst>(I))) { | |||
9483 | auto *Sel = cast<SelectInst>(I); | |||
9484 | auto *Cmp = dyn_cast<Instruction>(Sel->getCondition()); | |||
9485 | return Sel->getParent() == BB && Cmp && Cmp->getParent() == BB; | |||
9486 | } | |||
9487 | return I->getParent() == BB; | |||
9488 | } | |||
9489 | ||||
9490 | /// Expected number of uses for reduction operations/reduced values. | |||
9491 | static bool hasRequiredNumberOfUses(bool IsCmpSelMinMax, Instruction *I) { | |||
9492 | if (IsCmpSelMinMax) { | |||
9493 | // SelectInst must be used twice while the condition op must have single | |||
9494 | // use only. | |||
9495 | if (auto *Sel = dyn_cast<SelectInst>(I)) | |||
9496 | return Sel->hasNUses(2) && Sel->getCondition()->hasOneUse(); | |||
9497 | return I->hasNUses(2); | |||
9498 | } | |||
9499 | ||||
9500 | // Arithmetic reduction operation must be used once only. | |||
9501 | return I->hasOneUse(); | |||
9502 | } | |||
9503 | ||||
9504 | /// Initializes the list of reduction operations. | |||
9505 | void initReductionOps(Instruction *I) { | |||
9506 | if (isCmpSelMinMax(I)) | |||
9507 | ReductionOps.assign(2, ReductionOpsType()); | |||
9508 | else | |||
9509 | ReductionOps.assign(1, ReductionOpsType()); | |||
9510 | } | |||
9511 | ||||
9512 | /// Add all reduction operations for the reduction instruction \p I. | |||
9513 | void addReductionOps(Instruction *I) { | |||
9514 | if (isCmpSelMinMax(I)) { | |||
9515 | ReductionOps[0].emplace_back(cast<SelectInst>(I)->getCondition()); | |||
9516 | ReductionOps[1].emplace_back(I); | |||
9517 | } else { | |||
9518 | ReductionOps[0].emplace_back(I); | |||
9519 | } | |||
9520 | } | |||
9521 | ||||
9522 | static Value *getLHS(RecurKind Kind, Instruction *I) { | |||
9523 | if (Kind == RecurKind::None) | |||
9524 | return nullptr; | |||
9525 | return I->getOperand(getFirstOperandIndex(I)); | |||
9526 | } | |||
9527 | static Value *getRHS(RecurKind Kind, Instruction *I) { | |||
9528 | if (Kind == RecurKind::None) | |||
9529 | return nullptr; | |||
9530 | return I->getOperand(getFirstOperandIndex(I) + 1); | |||
9531 | } | |||
9532 | ||||
9533 | public: | |||
9534 | HorizontalReduction() = default; | |||
9535 | ||||
9536 | /// Try to find a reduction tree. | |||
9537 | bool matchAssociativeReduction(PHINode *Phi, Instruction *Inst) { | |||
9538 | assert((!Phi || is_contained(Phi->operands(), Inst)) &&(static_cast <bool> ((!Phi || is_contained(Phi->operands (), Inst)) && "Phi needs to use the binary operator") ? void (0) : __assert_fail ("(!Phi || is_contained(Phi->operands(), Inst)) && \"Phi needs to use the binary operator\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 9539, __extension__ __PRETTY_FUNCTION__)) | |||
9539 | "Phi needs to use the binary operator")(static_cast <bool> ((!Phi || is_contained(Phi->operands (), Inst)) && "Phi needs to use the binary operator") ? void (0) : __assert_fail ("(!Phi || is_contained(Phi->operands(), Inst)) && \"Phi needs to use the binary operator\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 9539, __extension__ __PRETTY_FUNCTION__)); | |||
9540 | assert((isa<BinaryOperator>(Inst) || isa<SelectInst>(Inst) ||(static_cast <bool> ((isa<BinaryOperator>(Inst) || isa<SelectInst>(Inst) || isa<IntrinsicInst>(Inst )) && "Expected binop, select, or intrinsic for reduction matching" ) ? void (0) : __assert_fail ("(isa<BinaryOperator>(Inst) || isa<SelectInst>(Inst) || isa<IntrinsicInst>(Inst)) && \"Expected binop, select, or intrinsic for reduction matching\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 9542, __extension__ __PRETTY_FUNCTION__)) | |||
9541 | isa<IntrinsicInst>(Inst)) &&(static_cast <bool> ((isa<BinaryOperator>(Inst) || isa<SelectInst>(Inst) || isa<IntrinsicInst>(Inst )) && "Expected binop, select, or intrinsic for reduction matching" ) ? void (0) : __assert_fail ("(isa<BinaryOperator>(Inst) || isa<SelectInst>(Inst) || isa<IntrinsicInst>(Inst)) && \"Expected binop, select, or intrinsic for reduction matching\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 9542, __extension__ __PRETTY_FUNCTION__)) | |||
9542 | "Expected binop, select, or intrinsic for reduction matching")(static_cast <bool> ((isa<BinaryOperator>(Inst) || isa<SelectInst>(Inst) || isa<IntrinsicInst>(Inst )) && "Expected binop, select, or intrinsic for reduction matching" ) ? void (0) : __assert_fail ("(isa<BinaryOperator>(Inst) || isa<SelectInst>(Inst) || isa<IntrinsicInst>(Inst)) && \"Expected binop, select, or intrinsic for reduction matching\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 9542, __extension__ __PRETTY_FUNCTION__)); | |||
9543 | RdxKind = getRdxKind(Inst); | |||
9544 | ||||
9545 | // We could have a initial reductions that is not an add. | |||
9546 | // r *= v1 + v2 + v3 + v4 | |||
9547 | // In such a case start looking for a tree rooted in the first '+'. | |||
9548 | if (Phi) { | |||
9549 | if (getLHS(RdxKind, Inst) == Phi) { | |||
9550 | Phi = nullptr; | |||
9551 | Inst = dyn_cast<Instruction>(getRHS(RdxKind, Inst)); | |||
9552 | if (!Inst) | |||
9553 | return false; | |||
9554 | RdxKind = getRdxKind(Inst); | |||
9555 | } else if (getRHS(RdxKind, Inst) == Phi) { | |||
9556 | Phi = nullptr; | |||
9557 | Inst = dyn_cast<Instruction>(getLHS(RdxKind, Inst)); | |||
9558 | if (!Inst) | |||
9559 | return false; | |||
9560 | RdxKind = getRdxKind(Inst); | |||
9561 | } | |||
9562 | } | |||
9563 | ||||
9564 | if (!isVectorizable(RdxKind, Inst)) | |||
9565 | return false; | |||
9566 | ||||
9567 | // Analyze "regular" integer/FP types for reductions - no target-specific | |||
9568 | // types or pointers. | |||
9569 | Type *Ty = Inst->getType(); | |||
9570 | if (!isValidElementType(Ty) || Ty->isPointerTy()) | |||
9571 | return false; | |||
9572 | ||||
9573 | // Though the ultimate reduction may have multiple uses, its condition must | |||
9574 | // have only single use. | |||
9575 | if (auto *Sel = dyn_cast<SelectInst>(Inst)) | |||
9576 | if (!Sel->getCondition()->hasOneUse()) | |||
9577 | return false; | |||
9578 | ||||
9579 | ReductionRoot = Inst; | |||
9580 | ||||
9581 | // The opcode for leaf values that we perform a reduction on. | |||
9582 | // For example: load(x) + load(y) + load(z) + fptoui(w) | |||
9583 | // The leaf opcode for 'w' does not match, so we don't include it as a | |||
9584 | // potential candidate for the reduction. | |||
9585 | unsigned LeafOpcode = 0; | |||
9586 | ||||
9587 | // Post-order traverse the reduction tree starting at Inst. We only handle | |||
9588 | // true trees containing binary operators or selects. | |||
9589 | SmallVector<std::pair<Instruction *, unsigned>, 32> Stack; | |||
9590 | Stack.push_back(std::make_pair(Inst, getFirstOperandIndex(Inst))); | |||
9591 | initReductionOps(Inst); | |||
9592 | while (!Stack.empty()) { | |||
9593 | Instruction *TreeN = Stack.back().first; | |||
9594 | unsigned EdgeToVisit = Stack.back().second++; | |||
9595 | const RecurKind TreeRdxKind = getRdxKind(TreeN); | |||
9596 | bool IsReducedValue = TreeRdxKind != RdxKind; | |||
9597 | ||||
9598 | // Postorder visit. | |||
9599 | if (IsReducedValue || EdgeToVisit >= getNumberOfOperands(TreeN)) { | |||
9600 | if (IsReducedValue) | |||
9601 | ReducedVals.push_back(TreeN); | |||
9602 | else { | |||
9603 | auto ExtraArgsIter = ExtraArgs.find(TreeN); | |||
9604 | if (ExtraArgsIter != ExtraArgs.end() && !ExtraArgsIter->second) { | |||
9605 | // Check if TreeN is an extra argument of its parent operation. | |||
9606 | if (Stack.size() <= 1) { | |||
9607 | // TreeN can't be an extra argument as it is a root reduction | |||
9608 | // operation. | |||
9609 | return false; | |||
9610 | } | |||
9611 | // Yes, TreeN is an extra argument, do not add it to a list of | |||
9612 | // reduction operations. | |||
9613 | // Stack[Stack.size() - 2] always points to the parent operation. | |||
9614 | markExtraArg(Stack[Stack.size() - 2], TreeN); | |||
9615 | ExtraArgs.erase(TreeN); | |||
9616 | } else | |||
9617 | addReductionOps(TreeN); | |||
9618 | } | |||
9619 | // Retract. | |||
9620 | Stack.pop_back(); | |||
9621 | continue; | |||
9622 | } | |||
9623 | ||||
9624 | // Visit operands. | |||
9625 | Value *EdgeVal = getRdxOperand(TreeN, EdgeToVisit); | |||
9626 | auto *EdgeInst = dyn_cast<Instruction>(EdgeVal); | |||
9627 | if (!EdgeInst) { | |||
9628 | // Edge value is not a reduction instruction or a leaf instruction. | |||
9629 | // (It may be a constant, function argument, or something else.) | |||
9630 | markExtraArg(Stack.back(), EdgeVal); | |||
9631 | continue; | |||
9632 | } | |||
9633 | RecurKind EdgeRdxKind = getRdxKind(EdgeInst); | |||
9634 | // Continue analysis if the next operand is a reduction operation or | |||
9635 | // (possibly) a leaf value. If the leaf value opcode is not set, | |||
9636 | // the first met operation != reduction operation is considered as the | |||
9637 | // leaf opcode. | |||
9638 | // Only handle trees in the current basic block. | |||
9639 | // Each tree node needs to have minimal number of users except for the | |||
9640 | // ultimate reduction. | |||
9641 | const bool IsRdxInst = EdgeRdxKind == RdxKind; | |||
9642 | if (EdgeInst != Phi && EdgeInst != Inst && | |||
9643 | hasSameParent(EdgeInst, Inst->getParent()) && | |||
9644 | hasRequiredNumberOfUses(isCmpSelMinMax(Inst), EdgeInst) && | |||
9645 | (!LeafOpcode || LeafOpcode == EdgeInst->getOpcode() || IsRdxInst)) { | |||
9646 | if (IsRdxInst) { | |||
9647 | // We need to be able to reassociate the reduction operations. | |||
9648 | if (!isVectorizable(EdgeRdxKind, EdgeInst)) { | |||
9649 | // I is an extra argument for TreeN (its parent operation). | |||
9650 | markExtraArg(Stack.back(), EdgeInst); | |||
9651 | continue; | |||
9652 | } | |||
9653 | } else if (!LeafOpcode) { | |||
9654 | LeafOpcode = EdgeInst->getOpcode(); | |||
9655 | } | |||
9656 | Stack.push_back( | |||
9657 | std::make_pair(EdgeInst, getFirstOperandIndex(EdgeInst))); | |||
9658 | continue; | |||
9659 | } | |||
9660 | // I is an extra argument for TreeN (its parent operation). | |||
9661 | markExtraArg(Stack.back(), EdgeInst); | |||
9662 | } | |||
9663 | return true; | |||
9664 | } | |||
9665 | ||||
9666 | /// Attempt to vectorize the tree found by matchAssociativeReduction. | |||
9667 | Value *tryToReduce(BoUpSLP &V, TargetTransformInfo *TTI) { | |||
9668 | // If there are a sufficient number of reduction values, reduce | |||
9669 | // to a nearby power-of-2. We can safely generate oversized | |||
9670 | // vectors and rely on the backend to split them to legal sizes. | |||
9671 | unsigned NumReducedVals = ReducedVals.size(); | |||
9672 | if (NumReducedVals < 4) | |||
9673 | return nullptr; | |||
9674 | ||||
9675 | // Intersect the fast-math-flags from all reduction operations. | |||
9676 | FastMathFlags RdxFMF; | |||
9677 | RdxFMF.set(); | |||
9678 | for (ReductionOpsType &RdxOp : ReductionOps) { | |||
9679 | for (Value *RdxVal : RdxOp) { | |||
9680 | if (auto *FPMO = dyn_cast<FPMathOperator>(RdxVal)) | |||
9681 | RdxFMF &= FPMO->getFastMathFlags(); | |||
9682 | } | |||
9683 | } | |||
9684 | ||||
9685 | IRBuilder<> Builder(cast<Instruction>(ReductionRoot)); | |||
9686 | Builder.setFastMathFlags(RdxFMF); | |||
9687 | ||||
9688 | BoUpSLP::ExtraValueToDebugLocsMap ExternallyUsedValues; | |||
9689 | // The same extra argument may be used several times, so log each attempt | |||
9690 | // to use it. | |||
9691 | for (const std::pair<Instruction *, Value *> &Pair : ExtraArgs) { | |||
9692 | assert(Pair.first && "DebugLoc must be set.")(static_cast <bool> (Pair.first && "DebugLoc must be set." ) ? void (0) : __assert_fail ("Pair.first && \"DebugLoc must be set.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 9692, __extension__ __PRETTY_FUNCTION__)); | |||
9693 | ExternallyUsedValues[Pair.second].push_back(Pair.first); | |||
9694 | } | |||
9695 | ||||
9696 | // The compare instruction of a min/max is the insertion point for new | |||
9697 | // instructions and may be replaced with a new compare instruction. | |||
9698 | auto getCmpForMinMaxReduction = [](Instruction *RdxRootInst) { | |||
9699 | assert(isa<SelectInst>(RdxRootInst) &&(static_cast <bool> (isa<SelectInst>(RdxRootInst) && "Expected min/max reduction to have select root instruction" ) ? void (0) : __assert_fail ("isa<SelectInst>(RdxRootInst) && \"Expected min/max reduction to have select root instruction\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 9700, __extension__ __PRETTY_FUNCTION__)) | |||
9700 | "Expected min/max reduction to have select root instruction")(static_cast <bool> (isa<SelectInst>(RdxRootInst) && "Expected min/max reduction to have select root instruction" ) ? void (0) : __assert_fail ("isa<SelectInst>(RdxRootInst) && \"Expected min/max reduction to have select root instruction\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 9700, __extension__ __PRETTY_FUNCTION__)); | |||
9701 | Value *ScalarCond = cast<SelectInst>(RdxRootInst)->getCondition(); | |||
9702 | assert(isa<Instruction>(ScalarCond) &&(static_cast <bool> (isa<Instruction>(ScalarCond) && "Expected min/max reduction to have compare condition" ) ? void (0) : __assert_fail ("isa<Instruction>(ScalarCond) && \"Expected min/max reduction to have compare condition\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 9703, __extension__ __PRETTY_FUNCTION__)) | |||
9703 | "Expected min/max reduction to have compare condition")(static_cast <bool> (isa<Instruction>(ScalarCond) && "Expected min/max reduction to have compare condition" ) ? void (0) : __assert_fail ("isa<Instruction>(ScalarCond) && \"Expected min/max reduction to have compare condition\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 9703, __extension__ __PRETTY_FUNCTION__)); | |||
9704 | return cast<Instruction>(ScalarCond); | |||
9705 | }; | |||
9706 | ||||
9707 | // The reduction root is used as the insertion point for new instructions, | |||
9708 | // so set it as externally used to prevent it from being deleted. | |||
9709 | ExternallyUsedValues[ReductionRoot]; | |||
9710 | SmallVector<Value *, 16> IgnoreList; | |||
9711 | for (ReductionOpsType &RdxOp : ReductionOps) | |||
9712 | IgnoreList.append(RdxOp.begin(), RdxOp.end()); | |||
9713 | ||||
9714 | unsigned ReduxWidth = PowerOf2Floor(NumReducedVals); | |||
9715 | if (NumReducedVals > ReduxWidth) { | |||
9716 | // In the loop below, we are building a tree based on a window of | |||
9717 | // 'ReduxWidth' values. | |||
9718 | // If the operands of those values have common traits (compare predicate, | |||
9719 | // constant operand, etc), then we want to group those together to | |||
9720 | // minimize the cost of the reduction. | |||
9721 | ||||
9722 | // TODO: This should be extended to count common operands for | |||
9723 | // compares and binops. | |||
9724 | ||||
9725 | // Step 1: Count the number of times each compare predicate occurs. | |||
9726 | SmallDenseMap<unsigned, unsigned> PredCountMap; | |||
9727 | for (Value *RdxVal : ReducedVals) { | |||
9728 | CmpInst::Predicate Pred; | |||
9729 | if (match(RdxVal, m_Cmp(Pred, m_Value(), m_Value()))) | |||
9730 | ++PredCountMap[Pred]; | |||
9731 | } | |||
9732 | // Step 2: Sort the values so the most common predicates come first. | |||
9733 | stable_sort(ReducedVals, [&PredCountMap](Value *A, Value *B) { | |||
9734 | CmpInst::Predicate PredA, PredB; | |||
9735 | if (match(A, m_Cmp(PredA, m_Value(), m_Value())) && | |||
9736 | match(B, m_Cmp(PredB, m_Value(), m_Value()))) { | |||
9737 | return PredCountMap[PredA] > PredCountMap[PredB]; | |||
9738 | } | |||
9739 | return false; | |||
9740 | }); | |||
9741 | } | |||
9742 | ||||
9743 | Value *VectorizedTree = nullptr; | |||
9744 | unsigned i = 0; | |||
9745 | while (i < NumReducedVals - ReduxWidth + 1 && ReduxWidth > 2) { | |||
9746 | ArrayRef<Value *> VL(&ReducedVals[i], ReduxWidth); | |||
9747 | V.buildTree(VL, IgnoreList); | |||
9748 | if (V.isTreeTinyAndNotFullyVectorizable(/*ForReduction=*/true)) | |||
9749 | break; | |||
9750 | if (V.isLoadCombineReductionCandidate(RdxKind)) | |||
9751 | break; | |||
9752 | V.reorderTopToBottom(); | |||
9753 | V.reorderBottomToTop(/*IgnoreReorder=*/true); | |||
9754 | V.buildExternalUses(ExternallyUsedValues); | |||
9755 | ||||
9756 | // For a poison-safe boolean logic reduction, do not replace select | |||
9757 | // instructions with logic ops. All reduced values will be frozen (see | |||
9758 | // below) to prevent leaking poison. | |||
9759 | if (isa<SelectInst>(ReductionRoot) && | |||
9760 | isBoolLogicOp(cast<Instruction>(ReductionRoot)) && | |||
9761 | NumReducedVals != ReduxWidth) | |||
9762 | break; | |||
9763 | ||||
9764 | V.computeMinimumValueSizes(); | |||
9765 | ||||
9766 | // Estimate cost. | |||
9767 | InstructionCost TreeCost = | |||
9768 | V.getTreeCost(makeArrayRef(&ReducedVals[i], ReduxWidth)); | |||
9769 | InstructionCost ReductionCost = | |||
9770 | getReductionCost(TTI, ReducedVals[i], ReduxWidth, RdxFMF); | |||
9771 | InstructionCost Cost = TreeCost + ReductionCost; | |||
9772 | if (!Cost.isValid()) { | |||
9773 | LLVM_DEBUG(dbgs() << "Encountered invalid baseline cost.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "Encountered invalid baseline cost.\n" ; } } while (false); | |||
9774 | return nullptr; | |||
9775 | } | |||
9776 | if (Cost >= -SLPCostThreshold) { | |||
9777 | V.getORE()->emit([&]() { | |||
9778 | return OptimizationRemarkMissed(SV_NAME"slp-vectorizer", "HorSLPNotBeneficial", | |||
9779 | cast<Instruction>(VL[0])) | |||
9780 | << "Vectorizing horizontal reduction is possible" | |||
9781 | << "but not beneficial with cost " << ore::NV("Cost", Cost) | |||
9782 | << " and threshold " | |||
9783 | << ore::NV("Threshold", -SLPCostThreshold); | |||
9784 | }); | |||
9785 | break; | |||
9786 | } | |||
9787 | ||||
9788 | LLVM_DEBUG(dbgs() << "SLP: Vectorizing horizontal reduction at cost:"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Vectorizing horizontal reduction at cost:" << Cost << ". (HorRdx)\n"; } } while (false) | |||
9789 | << Cost << ". (HorRdx)\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Vectorizing horizontal reduction at cost:" << Cost << ". (HorRdx)\n"; } } while (false); | |||
9790 | V.getORE()->emit([&]() { | |||
9791 | return OptimizationRemark(SV_NAME"slp-vectorizer", "VectorizedHorizontalReduction", | |||
9792 | cast<Instruction>(VL[0])) | |||
9793 | << "Vectorized horizontal reduction with cost " | |||
9794 | << ore::NV("Cost", Cost) << " and with tree size " | |||
9795 | << ore::NV("TreeSize", V.getTreeSize()); | |||
9796 | }); | |||
9797 | ||||
9798 | // Vectorize a tree. | |||
9799 | DebugLoc Loc = cast<Instruction>(ReducedVals[i])->getDebugLoc(); | |||
9800 | Value *VectorizedRoot = V.vectorizeTree(ExternallyUsedValues); | |||
9801 | ||||
9802 | // Emit a reduction. If the root is a select (min/max idiom), the insert | |||
9803 | // point is the compare condition of that select. | |||
9804 | Instruction *RdxRootInst = cast<Instruction>(ReductionRoot); | |||
9805 | if (isCmpSelMinMax(RdxRootInst)) | |||
9806 | Builder.SetInsertPoint(getCmpForMinMaxReduction(RdxRootInst)); | |||
9807 | else | |||
9808 | Builder.SetInsertPoint(RdxRootInst); | |||
9809 | ||||
9810 | // To prevent poison from leaking across what used to be sequential, safe, | |||
9811 | // scalar boolean logic operations, the reduction operand must be frozen. | |||
9812 | if (isa<SelectInst>(RdxRootInst) && isBoolLogicOp(RdxRootInst)) | |||
9813 | VectorizedRoot = Builder.CreateFreeze(VectorizedRoot); | |||
9814 | ||||
9815 | Value *ReducedSubTree = | |||
9816 | emitReduction(VectorizedRoot, Builder, ReduxWidth, TTI); | |||
9817 | ||||
9818 | if (!VectorizedTree) { | |||
9819 | // Initialize the final value in the reduction. | |||
9820 | VectorizedTree = ReducedSubTree; | |||
9821 | } else { | |||
9822 | // Update the final value in the reduction. | |||
9823 | Builder.SetCurrentDebugLocation(Loc); | |||
9824 | VectorizedTree = createOp(Builder, RdxKind, VectorizedTree, | |||
9825 | ReducedSubTree, "op.rdx", ReductionOps); | |||
9826 | } | |||
9827 | i += ReduxWidth; | |||
9828 | ReduxWidth = PowerOf2Floor(NumReducedVals - i); | |||
9829 | } | |||
9830 | ||||
9831 | if (VectorizedTree) { | |||
9832 | // Finish the reduction. | |||
9833 | for (; i < NumReducedVals; ++i) { | |||
9834 | auto *I = cast<Instruction>(ReducedVals[i]); | |||
9835 | Builder.SetCurrentDebugLocation(I->getDebugLoc()); | |||
9836 | VectorizedTree = | |||
9837 | createOp(Builder, RdxKind, VectorizedTree, I, "", ReductionOps); | |||
9838 | } | |||
9839 | for (auto &Pair : ExternallyUsedValues) { | |||
9840 | // Add each externally used value to the final reduction. | |||
9841 | for (auto *I : Pair.second) { | |||
9842 | Builder.SetCurrentDebugLocation(I->getDebugLoc()); | |||
9843 | VectorizedTree = createOp(Builder, RdxKind, VectorizedTree, | |||
9844 | Pair.first, "op.extra", I); | |||
9845 | } | |||
9846 | } | |||
9847 | ||||
9848 | ReductionRoot->replaceAllUsesWith(VectorizedTree); | |||
9849 | ||||
9850 | // The original scalar reduction is expected to have no remaining | |||
9851 | // uses outside the reduction tree itself. Assert that we got this | |||
9852 | // correct, replace internal uses with undef, and mark for eventual | |||
9853 | // deletion. | |||
9854 | #ifndef NDEBUG | |||
9855 | SmallSet<Value *, 4> IgnoreSet; | |||
9856 | IgnoreSet.insert(IgnoreList.begin(), IgnoreList.end()); | |||
9857 | #endif | |||
9858 | for (auto *Ignore : IgnoreList) { | |||
9859 | #ifndef NDEBUG | |||
9860 | for (auto *U : Ignore->users()) { | |||
9861 | assert(IgnoreSet.count(U))(static_cast <bool> (IgnoreSet.count(U)) ? void (0) : __assert_fail ("IgnoreSet.count(U)", "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 9861, __extension__ __PRETTY_FUNCTION__)); | |||
9862 | } | |||
9863 | #endif | |||
9864 | if (!Ignore->use_empty()) { | |||
9865 | Value *Undef = UndefValue::get(Ignore->getType()); | |||
9866 | Ignore->replaceAllUsesWith(Undef); | |||
9867 | } | |||
9868 | V.eraseInstruction(cast<Instruction>(Ignore)); | |||
9869 | } | |||
9870 | } | |||
9871 | return VectorizedTree; | |||
9872 | } | |||
9873 | ||||
9874 | unsigned numReductionValues() const { return ReducedVals.size(); } | |||
9875 | ||||
9876 | private: | |||
9877 | /// Calculate the cost of a reduction. | |||
9878 | InstructionCost getReductionCost(TargetTransformInfo *TTI, | |||
9879 | Value *FirstReducedVal, unsigned ReduxWidth, | |||
9880 | FastMathFlags FMF) { | |||
9881 | TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput; | |||
9882 | Type *ScalarTy = FirstReducedVal->getType(); | |||
9883 | FixedVectorType *VectorTy = FixedVectorType::get(ScalarTy, ReduxWidth); | |||
9884 | InstructionCost VectorCost, ScalarCost; | |||
9885 | switch (RdxKind) { | |||
9886 | case RecurKind::Add: | |||
9887 | case RecurKind::Mul: | |||
9888 | case RecurKind::Or: | |||
9889 | case RecurKind::And: | |||
9890 | case RecurKind::Xor: | |||
9891 | case RecurKind::FAdd: | |||
9892 | case RecurKind::FMul: { | |||
9893 | unsigned RdxOpcode = RecurrenceDescriptor::getOpcode(RdxKind); | |||
9894 | VectorCost = | |||
9895 | TTI->getArithmeticReductionCost(RdxOpcode, VectorTy, FMF, CostKind); | |||
9896 | ScalarCost = TTI->getArithmeticInstrCost(RdxOpcode, ScalarTy, CostKind); | |||
9897 | break; | |||
9898 | } | |||
9899 | case RecurKind::FMax: | |||
9900 | case RecurKind::FMin: { | |||
9901 | auto *SclCondTy = CmpInst::makeCmpResultType(ScalarTy); | |||
9902 | auto *VecCondTy = cast<VectorType>(CmpInst::makeCmpResultType(VectorTy)); | |||
9903 | VectorCost = TTI->getMinMaxReductionCost(VectorTy, VecCondTy, | |||
9904 | /*IsUnsigned=*/false, CostKind); | |||
9905 | CmpInst::Predicate RdxPred = getMinMaxReductionPredicate(RdxKind); | |||
9906 | ScalarCost = TTI->getCmpSelInstrCost(Instruction::FCmp, ScalarTy, | |||
9907 | SclCondTy, RdxPred, CostKind) + | |||
9908 | TTI->getCmpSelInstrCost(Instruction::Select, ScalarTy, | |||
9909 | SclCondTy, RdxPred, CostKind); | |||
9910 | break; | |||
9911 | } | |||
9912 | case RecurKind::SMax: | |||
9913 | case RecurKind::SMin: | |||
9914 | case RecurKind::UMax: | |||
9915 | case RecurKind::UMin: { | |||
9916 | auto *SclCondTy = CmpInst::makeCmpResultType(ScalarTy); | |||
9917 | auto *VecCondTy = cast<VectorType>(CmpInst::makeCmpResultType(VectorTy)); | |||
9918 | bool IsUnsigned = | |||
9919 | RdxKind == RecurKind::UMax || RdxKind == RecurKind::UMin; | |||
9920 | VectorCost = TTI->getMinMaxReductionCost(VectorTy, VecCondTy, IsUnsigned, | |||
9921 | CostKind); | |||
9922 | CmpInst::Predicate RdxPred = getMinMaxReductionPredicate(RdxKind); | |||
9923 | ScalarCost = TTI->getCmpSelInstrCost(Instruction::ICmp, ScalarTy, | |||
9924 | SclCondTy, RdxPred, CostKind) + | |||
9925 | TTI->getCmpSelInstrCost(Instruction::Select, ScalarTy, | |||
9926 | SclCondTy, RdxPred, CostKind); | |||
9927 | break; | |||
9928 | } | |||
9929 | default: | |||
9930 | llvm_unreachable("Expected arithmetic or min/max reduction operation")::llvm::llvm_unreachable_internal("Expected arithmetic or min/max reduction operation" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 9930); | |||
9931 | } | |||
9932 | ||||
9933 | // Scalar cost is repeated for N-1 elements. | |||
9934 | ScalarCost *= (ReduxWidth - 1); | |||
9935 | LLVM_DEBUG(dbgs() << "SLP: Adding cost " << VectorCost - ScalarCostdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Adding cost " << VectorCost - ScalarCost << " for reduction that starts with " << *FirstReducedVal << " (It is a splitting reduction)\n" ; } } while (false) | |||
9936 | << " for reduction that starts with " << *FirstReducedValdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Adding cost " << VectorCost - ScalarCost << " for reduction that starts with " << *FirstReducedVal << " (It is a splitting reduction)\n" ; } } while (false) | |||
9937 | << " (It is a splitting reduction)\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Adding cost " << VectorCost - ScalarCost << " for reduction that starts with " << *FirstReducedVal << " (It is a splitting reduction)\n" ; } } while (false); | |||
9938 | return VectorCost - ScalarCost; | |||
9939 | } | |||
9940 | ||||
9941 | /// Emit a horizontal reduction of the vectorized value. | |||
9942 | Value *emitReduction(Value *VectorizedValue, IRBuilder<> &Builder, | |||
9943 | unsigned ReduxWidth, const TargetTransformInfo *TTI) { | |||
9944 | assert(VectorizedValue && "Need to have a vectorized tree node")(static_cast <bool> (VectorizedValue && "Need to have a vectorized tree node" ) ? void (0) : __assert_fail ("VectorizedValue && \"Need to have a vectorized tree node\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 9944, __extension__ __PRETTY_FUNCTION__)); | |||
9945 | assert(isPowerOf2_32(ReduxWidth) &&(static_cast <bool> (isPowerOf2_32(ReduxWidth) && "We only handle power-of-two reductions for now") ? void (0) : __assert_fail ("isPowerOf2_32(ReduxWidth) && \"We only handle power-of-two reductions for now\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 9946, __extension__ __PRETTY_FUNCTION__)) | |||
9946 | "We only handle power-of-two reductions for now")(static_cast <bool> (isPowerOf2_32(ReduxWidth) && "We only handle power-of-two reductions for now") ? void (0) : __assert_fail ("isPowerOf2_32(ReduxWidth) && \"We only handle power-of-two reductions for now\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 9946, __extension__ __PRETTY_FUNCTION__)); | |||
9947 | assert(RdxKind != RecurKind::FMulAdd &&(static_cast <bool> (RdxKind != RecurKind::FMulAdd && "A call to the llvm.fmuladd intrinsic is not handled yet") ? void (0) : __assert_fail ("RdxKind != RecurKind::FMulAdd && \"A call to the llvm.fmuladd intrinsic is not handled yet\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 9948, __extension__ __PRETTY_FUNCTION__)) | |||
9948 | "A call to the llvm.fmuladd intrinsic is not handled yet")(static_cast <bool> (RdxKind != RecurKind::FMulAdd && "A call to the llvm.fmuladd intrinsic is not handled yet") ? void (0) : __assert_fail ("RdxKind != RecurKind::FMulAdd && \"A call to the llvm.fmuladd intrinsic is not handled yet\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 9948, __extension__ __PRETTY_FUNCTION__)); | |||
9949 | ||||
9950 | ++NumVectorInstructions; | |||
9951 | return createSimpleTargetReduction(Builder, TTI, VectorizedValue, RdxKind); | |||
9952 | } | |||
9953 | }; | |||
9954 | ||||
9955 | } // end anonymous namespace | |||
9956 | ||||
9957 | static Optional<unsigned> getAggregateSize(Instruction *InsertInst) { | |||
9958 | if (auto *IE = dyn_cast<InsertElementInst>(InsertInst)) | |||
9959 | return cast<FixedVectorType>(IE->getType())->getNumElements(); | |||
9960 | ||||
9961 | unsigned AggregateSize = 1; | |||
9962 | auto *IV = cast<InsertValueInst>(InsertInst); | |||
9963 | Type *CurrentType = IV->getType(); | |||
9964 | do { | |||
9965 | if (auto *ST = dyn_cast<StructType>(CurrentType)) { | |||
9966 | for (auto *Elt : ST->elements()) | |||
9967 | if (Elt != ST->getElementType(0)) // check homogeneity | |||
9968 | return None; | |||
9969 | AggregateSize *= ST->getNumElements(); | |||
9970 | CurrentType = ST->getElementType(0); | |||
9971 | } else if (auto *AT = dyn_cast<ArrayType>(CurrentType)) { | |||
9972 | AggregateSize *= AT->getNumElements(); | |||
9973 | CurrentType = AT->getElementType(); | |||
9974 | } else if (auto *VT = dyn_cast<FixedVectorType>(CurrentType)) { | |||
9975 | AggregateSize *= VT->getNumElements(); | |||
9976 | return AggregateSize; | |||
9977 | } else if (CurrentType->isSingleValueType()) { | |||
9978 | return AggregateSize; | |||
9979 | } else { | |||
9980 | return None; | |||
9981 | } | |||
9982 | } while (true); | |||
9983 | } | |||
9984 | ||||
9985 | static void findBuildAggregate_rec(Instruction *LastInsertInst, | |||
9986 | TargetTransformInfo *TTI, | |||
9987 | SmallVectorImpl<Value *> &BuildVectorOpds, | |||
9988 | SmallVectorImpl<Value *> &InsertElts, | |||
9989 | unsigned OperandOffset) { | |||
9990 | do { | |||
9991 | Value *InsertedOperand = LastInsertInst->getOperand(1); | |||
9992 | Optional<unsigned> OperandIndex = | |||
9993 | getInsertIndex(LastInsertInst, OperandOffset); | |||
9994 | if (!OperandIndex) | |||
9995 | return; | |||
9996 | if (isa<InsertElementInst>(InsertedOperand) || | |||
9997 | isa<InsertValueInst>(InsertedOperand)) { | |||
9998 | findBuildAggregate_rec(cast<Instruction>(InsertedOperand), TTI, | |||
9999 | BuildVectorOpds, InsertElts, *OperandIndex); | |||
10000 | ||||
10001 | } else { | |||
10002 | BuildVectorOpds[*OperandIndex] = InsertedOperand; | |||
10003 | InsertElts[*OperandIndex] = LastInsertInst; | |||
10004 | } | |||
10005 | LastInsertInst = dyn_cast<Instruction>(LastInsertInst->getOperand(0)); | |||
10006 | } while (LastInsertInst != nullptr && | |||
10007 | (isa<InsertValueInst>(LastInsertInst) || | |||
10008 | isa<InsertElementInst>(LastInsertInst)) && | |||
10009 | LastInsertInst->hasOneUse()); | |||
10010 | } | |||
10011 | ||||
10012 | /// Recognize construction of vectors like | |||
10013 | /// %ra = insertelement <4 x float> poison, float %s0, i32 0 | |||
10014 | /// %rb = insertelement <4 x float> %ra, float %s1, i32 1 | |||
10015 | /// %rc = insertelement <4 x float> %rb, float %s2, i32 2 | |||
10016 | /// %rd = insertelement <4 x float> %rc, float %s3, i32 3 | |||
10017 | /// starting from the last insertelement or insertvalue instruction. | |||
10018 | /// | |||
10019 | /// Also recognize homogeneous aggregates like {<2 x float>, <2 x float>}, | |||
10020 | /// {{float, float}, {float, float}}, [2 x {float, float}] and so on. | |||
10021 | /// See llvm/test/Transforms/SLPVectorizer/X86/pr42022.ll for examples. | |||
10022 | /// | |||
10023 | /// Assume LastInsertInst is of InsertElementInst or InsertValueInst type. | |||
10024 | /// | |||
10025 | /// \return true if it matches. | |||
10026 | static bool findBuildAggregate(Instruction *LastInsertInst, | |||
10027 | TargetTransformInfo *TTI, | |||
10028 | SmallVectorImpl<Value *> &BuildVectorOpds, | |||
10029 | SmallVectorImpl<Value *> &InsertElts) { | |||
10030 | ||||
10031 | assert((isa<InsertElementInst>(LastInsertInst) ||(static_cast <bool> ((isa<InsertElementInst>(LastInsertInst ) || isa<InsertValueInst>(LastInsertInst)) && "Expected insertelement or insertvalue instruction!" ) ? void (0) : __assert_fail ("(isa<InsertElementInst>(LastInsertInst) || isa<InsertValueInst>(LastInsertInst)) && \"Expected insertelement or insertvalue instruction!\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 10033, __extension__ __PRETTY_FUNCTION__)) | |||
10032 | isa<InsertValueInst>(LastInsertInst)) &&(static_cast <bool> ((isa<InsertElementInst>(LastInsertInst ) || isa<InsertValueInst>(LastInsertInst)) && "Expected insertelement or insertvalue instruction!" ) ? void (0) : __assert_fail ("(isa<InsertElementInst>(LastInsertInst) || isa<InsertValueInst>(LastInsertInst)) && \"Expected insertelement or insertvalue instruction!\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 10033, __extension__ __PRETTY_FUNCTION__)) | |||
10033 | "Expected insertelement or insertvalue instruction!")(static_cast <bool> ((isa<InsertElementInst>(LastInsertInst ) || isa<InsertValueInst>(LastInsertInst)) && "Expected insertelement or insertvalue instruction!" ) ? void (0) : __assert_fail ("(isa<InsertElementInst>(LastInsertInst) || isa<InsertValueInst>(LastInsertInst)) && \"Expected insertelement or insertvalue instruction!\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 10033, __extension__ __PRETTY_FUNCTION__)); | |||
10034 | ||||
10035 | assert((BuildVectorOpds.empty() && InsertElts.empty()) &&(static_cast <bool> ((BuildVectorOpds.empty() && InsertElts.empty()) && "Expected empty result vectors!" ) ? void (0) : __assert_fail ("(BuildVectorOpds.empty() && InsertElts.empty()) && \"Expected empty result vectors!\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 10036, __extension__ __PRETTY_FUNCTION__)) | |||
10036 | "Expected empty result vectors!")(static_cast <bool> ((BuildVectorOpds.empty() && InsertElts.empty()) && "Expected empty result vectors!" ) ? void (0) : __assert_fail ("(BuildVectorOpds.empty() && InsertElts.empty()) && \"Expected empty result vectors!\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 10036, __extension__ __PRETTY_FUNCTION__)); | |||
10037 | ||||
10038 | Optional<unsigned> AggregateSize = getAggregateSize(LastInsertInst); | |||
10039 | if (!AggregateSize) | |||
10040 | return false; | |||
10041 | BuildVectorOpds.resize(*AggregateSize); | |||
10042 | InsertElts.resize(*AggregateSize); | |||
10043 | ||||
10044 | findBuildAggregate_rec(LastInsertInst, TTI, BuildVectorOpds, InsertElts, 0); | |||
10045 | llvm::erase_value(BuildVectorOpds, nullptr); | |||
10046 | llvm::erase_value(InsertElts, nullptr); | |||
10047 | if (BuildVectorOpds.size() >= 2) | |||
10048 | return true; | |||
10049 | ||||
10050 | return false; | |||
10051 | } | |||
10052 | ||||
10053 | /// Try and get a reduction value from a phi node. | |||
10054 | /// | |||
10055 | /// Given a phi node \p P in a block \p ParentBB, consider possible reductions | |||
10056 | /// if they come from either \p ParentBB or a containing loop latch. | |||
10057 | /// | |||
10058 | /// \returns A candidate reduction value if possible, or \code nullptr \endcode | |||
10059 | /// if not possible. | |||
10060 | static Value *getReductionValue(const DominatorTree *DT, PHINode *P, | |||
10061 | BasicBlock *ParentBB, LoopInfo *LI) { | |||
10062 | // There are situations where the reduction value is not dominated by the | |||
10063 | // reduction phi. Vectorizing such cases has been reported to cause | |||
10064 | // miscompiles. See PR25787. | |||
10065 | auto DominatedReduxValue = [&](Value *R) { | |||
10066 | return isa<Instruction>(R) && | |||
10067 | DT->dominates(P->getParent(), cast<Instruction>(R)->getParent()); | |||
10068 | }; | |||
10069 | ||||
10070 | Value *Rdx = nullptr; | |||
10071 | ||||
10072 | // Return the incoming value if it comes from the same BB as the phi node. | |||
10073 | if (P->getIncomingBlock(0) == ParentBB) { | |||
10074 | Rdx = P->getIncomingValue(0); | |||
10075 | } else if (P->getIncomingBlock(1) == ParentBB) { | |||
10076 | Rdx = P->getIncomingValue(1); | |||
10077 | } | |||
10078 | ||||
10079 | if (Rdx && DominatedReduxValue(Rdx)) | |||
10080 | return Rdx; | |||
10081 | ||||
10082 | // Otherwise, check whether we have a loop latch to look at. | |||
10083 | Loop *BBL = LI->getLoopFor(ParentBB); | |||
10084 | if (!BBL) | |||
10085 | return nullptr; | |||
10086 | BasicBlock *BBLatch = BBL->getLoopLatch(); | |||
10087 | if (!BBLatch) | |||
10088 | return nullptr; | |||
10089 | ||||
10090 | // There is a loop latch, return the incoming value if it comes from | |||
10091 | // that. This reduction pattern occasionally turns up. | |||
10092 | if (P->getIncomingBlock(0) == BBLatch) { | |||
10093 | Rdx = P->getIncomingValue(0); | |||
10094 | } else if (P->getIncomingBlock(1) == BBLatch) { | |||
10095 | Rdx = P->getIncomingValue(1); | |||
10096 | } | |||
10097 | ||||
10098 | if (Rdx && DominatedReduxValue(Rdx)) | |||
10099 | return Rdx; | |||
10100 | ||||
10101 | return nullptr; | |||
10102 | } | |||
10103 | ||||
10104 | static bool matchRdxBop(Instruction *I, Value *&V0, Value *&V1) { | |||
10105 | if (match(I, m_BinOp(m_Value(V0), m_Value(V1)))) | |||
10106 | return true; | |||
10107 | if (match(I, m_Intrinsic<Intrinsic::maxnum>(m_Value(V0), m_Value(V1)))) | |||
10108 | return true; | |||
10109 | if (match(I, m_Intrinsic<Intrinsic::minnum>(m_Value(V0), m_Value(V1)))) | |||
10110 | return true; | |||
10111 | if (match(I, m_Intrinsic<Intrinsic::smax>(m_Value(V0), m_Value(V1)))) | |||
10112 | return true; | |||
10113 | if (match(I, m_Intrinsic<Intrinsic::smin>(m_Value(V0), m_Value(V1)))) | |||
10114 | return true; | |||
10115 | if (match(I, m_Intrinsic<Intrinsic::umax>(m_Value(V0), m_Value(V1)))) | |||
10116 | return true; | |||
10117 | if (match(I, m_Intrinsic<Intrinsic::umin>(m_Value(V0), m_Value(V1)))) | |||
10118 | return true; | |||
10119 | return false; | |||
10120 | } | |||
10121 | ||||
10122 | /// Attempt to reduce a horizontal reduction. | |||
10123 | /// If it is legal to match a horizontal reduction feeding the phi node \a P | |||
10124 | /// with reduction operators \a Root (or one of its operands) in a basic block | |||
10125 | /// \a BB, then check if it can be done. If horizontal reduction is not found | |||
10126 | /// and root instruction is a binary operation, vectorization of the operands is | |||
10127 | /// attempted. | |||
10128 | /// \returns true if a horizontal reduction was matched and reduced or operands | |||
10129 | /// of one of the binary instruction were vectorized. | |||
10130 | /// \returns false if a horizontal reduction was not matched (or not possible) | |||
10131 | /// or no vectorization of any binary operation feeding \a Root instruction was | |||
10132 | /// performed. | |||
10133 | static bool tryToVectorizeHorReductionOrInstOperands( | |||
10134 | PHINode *P, Instruction *Root, BasicBlock *BB, BoUpSLP &R, | |||
10135 | TargetTransformInfo *TTI, | |||
10136 | const function_ref<bool(Instruction *, BoUpSLP &)> Vectorize) { | |||
10137 | if (!ShouldVectorizeHor) | |||
10138 | return false; | |||
10139 | ||||
10140 | if (!Root) | |||
10141 | return false; | |||
10142 | ||||
10143 | if (Root->getParent() != BB || isa<PHINode>(Root)) | |||
10144 | return false; | |||
10145 | // Start analysis starting from Root instruction. If horizontal reduction is | |||
10146 | // found, try to vectorize it. If it is not a horizontal reduction or | |||
10147 | // vectorization is not possible or not effective, and currently analyzed | |||
10148 | // instruction is a binary operation, try to vectorize the operands, using | |||
10149 | // pre-order DFS traversal order. If the operands were not vectorized, repeat | |||
10150 | // the same procedure considering each operand as a possible root of the | |||
10151 | // horizontal reduction. | |||
10152 | // Interrupt the process if the Root instruction itself was vectorized or all | |||
10153 | // sub-trees not higher that RecursionMaxDepth were analyzed/vectorized. | |||
10154 | // Skip the analysis of CmpInsts.Compiler implements postanalysis of the | |||
10155 | // CmpInsts so we can skip extra attempts in | |||
10156 | // tryToVectorizeHorReductionOrInstOperands and save compile time. | |||
10157 | std::queue<std::pair<Instruction *, unsigned>> Stack; | |||
10158 | Stack.emplace(Root, 0); | |||
10159 | SmallPtrSet<Value *, 8> VisitedInstrs; | |||
10160 | SmallVector<WeakTrackingVH> PostponedInsts; | |||
10161 | bool Res = false; | |||
10162 | auto &&TryToReduce = [TTI, &P, &R](Instruction *Inst, Value *&B0, | |||
10163 | Value *&B1) -> Value * { | |||
10164 | bool IsBinop = matchRdxBop(Inst, B0, B1); | |||
10165 | bool IsSelect = match(Inst, m_Select(m_Value(), m_Value(), m_Value())); | |||
10166 | if (IsBinop || IsSelect) { | |||
10167 | HorizontalReduction HorRdx; | |||
10168 | if (HorRdx.matchAssociativeReduction(P, Inst)) | |||
10169 | return HorRdx.tryToReduce(R, TTI); | |||
10170 | } | |||
10171 | return nullptr; | |||
10172 | }; | |||
10173 | while (!Stack.empty()) { | |||
10174 | Instruction *Inst; | |||
10175 | unsigned Level; | |||
10176 | std::tie(Inst, Level) = Stack.front(); | |||
10177 | Stack.pop(); | |||
10178 | // Do not try to analyze instruction that has already been vectorized. | |||
10179 | // This may happen when we vectorize instruction operands on a previous | |||
10180 | // iteration while stack was populated before that happened. | |||
10181 | if (R.isDeleted(Inst)) | |||
10182 | continue; | |||
10183 | Value *B0 = nullptr, *B1 = nullptr; | |||
10184 | if (Value *V = TryToReduce(Inst, B0, B1)) { | |||
10185 | Res = true; | |||
10186 | // Set P to nullptr to avoid re-analysis of phi node in | |||
10187 | // matchAssociativeReduction function unless this is the root node. | |||
10188 | P = nullptr; | |||
10189 | if (auto *I = dyn_cast<Instruction>(V)) { | |||
10190 | // Try to find another reduction. | |||
10191 | Stack.emplace(I, Level); | |||
10192 | continue; | |||
10193 | } | |||
10194 | } else { | |||
10195 | bool IsBinop = B0 && B1; | |||
10196 | if (P && IsBinop) { | |||
10197 | Inst = dyn_cast<Instruction>(B0); | |||
10198 | if (Inst == P) | |||
10199 | Inst = dyn_cast<Instruction>(B1); | |||
10200 | if (!Inst) { | |||
10201 | // Set P to nullptr to avoid re-analysis of phi node in | |||
10202 | // matchAssociativeReduction function unless this is the root node. | |||
10203 | P = nullptr; | |||
10204 | continue; | |||
10205 | } | |||
10206 | } | |||
10207 | // Set P to nullptr to avoid re-analysis of phi node in | |||
10208 | // matchAssociativeReduction function unless this is the root node. | |||
10209 | P = nullptr; | |||
10210 | // Do not try to vectorize CmpInst operands, this is done separately. | |||
10211 | // Final attempt for binop args vectorization should happen after the loop | |||
10212 | // to try to find reductions. | |||
10213 | if (!isa<CmpInst>(Inst)) | |||
10214 | PostponedInsts.push_back(Inst); | |||
10215 | } | |||
10216 | ||||
10217 | // Try to vectorize operands. | |||
10218 | // Continue analysis for the instruction from the same basic block only to | |||
10219 | // save compile time. | |||
10220 | if (++Level < RecursionMaxDepth) | |||
10221 | for (auto *Op : Inst->operand_values()) | |||
10222 | if (VisitedInstrs.insert(Op).second) | |||
10223 | if (auto *I = dyn_cast<Instruction>(Op)) | |||
10224 | // Do not try to vectorize CmpInst operands, this is done | |||
10225 | // separately. | |||
10226 | if (!isa<PHINode>(I) && !isa<CmpInst>(I) && !R.isDeleted(I) && | |||
10227 | I->getParent() == BB) | |||
10228 | Stack.emplace(I, Level); | |||
10229 | } | |||
10230 | // Try to vectorized binops where reductions were not found. | |||
10231 | for (Value *V : PostponedInsts) | |||
10232 | if (auto *Inst = dyn_cast<Instruction>(V)) | |||
10233 | if (!R.isDeleted(Inst)) | |||
10234 | Res |= Vectorize(Inst, R); | |||
10235 | return Res; | |||
10236 | } | |||
10237 | ||||
10238 | bool SLPVectorizerPass::vectorizeRootInstruction(PHINode *P, Value *V, | |||
10239 | BasicBlock *BB, BoUpSLP &R, | |||
10240 | TargetTransformInfo *TTI) { | |||
10241 | auto *I = dyn_cast_or_null<Instruction>(V); | |||
10242 | if (!I) | |||
10243 | return false; | |||
10244 | ||||
10245 | if (!isa<BinaryOperator>(I)) | |||
10246 | P = nullptr; | |||
10247 | // Try to match and vectorize a horizontal reduction. | |||
10248 | auto &&ExtraVectorization = [this](Instruction *I, BoUpSLP &R) -> bool { | |||
10249 | return tryToVectorize(I, R); | |||
10250 | }; | |||
10251 | return tryToVectorizeHorReductionOrInstOperands(P, I, BB, R, TTI, | |||
10252 | ExtraVectorization); | |||
10253 | } | |||
10254 | ||||
10255 | bool SLPVectorizerPass::vectorizeInsertValueInst(InsertValueInst *IVI, | |||
10256 | BasicBlock *BB, BoUpSLP &R) { | |||
10257 | const DataLayout &DL = BB->getModule()->getDataLayout(); | |||
10258 | if (!R.canMapToVector(IVI->getType(), DL)) | |||
10259 | return false; | |||
10260 | ||||
10261 | SmallVector<Value *, 16> BuildVectorOpds; | |||
10262 | SmallVector<Value *, 16> BuildVectorInsts; | |||
10263 | if (!findBuildAggregate(IVI, TTI, BuildVectorOpds, BuildVectorInsts)) | |||
10264 | return false; | |||
10265 | ||||
10266 | LLVM_DEBUG(dbgs() << "SLP: array mappable to vector: " << *IVI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: array mappable to vector: " << *IVI << "\n"; } } while (false); | |||
10267 | // Aggregate value is unlikely to be processed in vector register. | |||
10268 | return tryToVectorizeList(BuildVectorOpds, R); | |||
10269 | } | |||
10270 | ||||
10271 | bool SLPVectorizerPass::vectorizeInsertElementInst(InsertElementInst *IEI, | |||
10272 | BasicBlock *BB, BoUpSLP &R) { | |||
10273 | SmallVector<Value *, 16> BuildVectorInsts; | |||
10274 | SmallVector<Value *, 16> BuildVectorOpds; | |||
10275 | SmallVector<int> Mask; | |||
10276 | if (!findBuildAggregate(IEI, TTI, BuildVectorOpds, BuildVectorInsts) || | |||
10277 | (llvm::all_of( | |||
10278 | BuildVectorOpds, | |||
10279 | [](Value *V) { return isa<ExtractElementInst, UndefValue>(V); }) && | |||
10280 | isFixedVectorShuffle(BuildVectorOpds, Mask))) | |||
10281 | return false; | |||
10282 | ||||
10283 | LLVM_DEBUG(dbgs() << "SLP: array mappable to vector: " << *IEI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: array mappable to vector: " << *IEI << "\n"; } } while (false); | |||
10284 | return tryToVectorizeList(BuildVectorInsts, R); | |||
10285 | } | |||
10286 | ||||
10287 | template <typename T> | |||
10288 | static bool | |||
10289 | tryToVectorizeSequence(SmallVectorImpl<T *> &Incoming, | |||
10290 | function_ref<unsigned(T *)> Limit, | |||
10291 | function_ref<bool(T *, T *)> Comparator, | |||
10292 | function_ref<bool(T *, T *)> AreCompatible, | |||
10293 | function_ref<bool(ArrayRef<T *>, bool)> TryToVectorizeHelper, | |||
10294 | bool LimitForRegisterSize) { | |||
10295 | bool Changed = false; | |||
10296 | // Sort by type, parent, operands. | |||
10297 | stable_sort(Incoming, Comparator); | |||
10298 | ||||
10299 | // Try to vectorize elements base on their type. | |||
10300 | SmallVector<T *> Candidates; | |||
10301 | for (auto *IncIt = Incoming.begin(), *E = Incoming.end(); IncIt != E;) { | |||
10302 | // Look for the next elements with the same type, parent and operand | |||
10303 | // kinds. | |||
10304 | auto *SameTypeIt = IncIt; | |||
10305 | while (SameTypeIt != E && AreCompatible(*SameTypeIt, *IncIt)) | |||
10306 | ++SameTypeIt; | |||
10307 | ||||
10308 | // Try to vectorize them. | |||
10309 | unsigned NumElts = (SameTypeIt - IncIt); | |||
10310 | LLVM_DEBUG(dbgs() << "SLP: Trying to vectorize starting at nodes ("do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Trying to vectorize starting at nodes (" << NumElts << ")\n"; } } while (false) | |||
10311 | << NumElts << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Trying to vectorize starting at nodes (" << NumElts << ")\n"; } } while (false); | |||
10312 | // The vectorization is a 3-state attempt: | |||
10313 | // 1. Try to vectorize instructions with the same/alternate opcodes with the | |||
10314 | // size of maximal register at first. | |||
10315 | // 2. Try to vectorize remaining instructions with the same type, if | |||
10316 | // possible. This may result in the better vectorization results rather than | |||
10317 | // if we try just to vectorize instructions with the same/alternate opcodes. | |||
10318 | // 3. Final attempt to try to vectorize all instructions with the | |||
10319 | // same/alternate ops only, this may result in some extra final | |||
10320 | // vectorization. | |||
10321 | if (NumElts > 1 && | |||
10322 | TryToVectorizeHelper(makeArrayRef(IncIt, NumElts), LimitForRegisterSize)) { | |||
10323 | // Success start over because instructions might have been changed. | |||
10324 | Changed = true; | |||
10325 | } else if (NumElts < Limit(*IncIt) && | |||
10326 | (Candidates.empty() || | |||
10327 | Candidates.front()->getType() == (*IncIt)->getType())) { | |||
10328 | Candidates.append(IncIt, std::next(IncIt, NumElts)); | |||
10329 | } | |||
10330 | // Final attempt to vectorize instructions with the same types. | |||
10331 | if (Candidates.size() > 1 && | |||
10332 | (SameTypeIt == E || (*SameTypeIt)->getType() != (*IncIt)->getType())) { | |||
10333 | if (TryToVectorizeHelper(Candidates, /*LimitForRegisterSize=*/false)) { | |||
10334 | // Success start over because instructions might have been changed. | |||
10335 | Changed = true; | |||
10336 | } else if (LimitForRegisterSize) { | |||
10337 | // Try to vectorize using small vectors. | |||
10338 | for (auto *It = Candidates.begin(), *End = Candidates.end(); | |||
10339 | It != End;) { | |||
10340 | auto *SameTypeIt = It; | |||
10341 | while (SameTypeIt != End && AreCompatible(*SameTypeIt, *It)) | |||
10342 | ++SameTypeIt; | |||
10343 | unsigned NumElts = (SameTypeIt - It); | |||
10344 | if (NumElts > 1 && TryToVectorizeHelper(makeArrayRef(It, NumElts), | |||
10345 | /*LimitForRegisterSize=*/false)) | |||
10346 | Changed = true; | |||
10347 | It = SameTypeIt; | |||
10348 | } | |||
10349 | } | |||
10350 | Candidates.clear(); | |||
10351 | } | |||
10352 | ||||
10353 | // Start over at the next instruction of a different type (or the end). | |||
10354 | IncIt = SameTypeIt; | |||
10355 | } | |||
10356 | return Changed; | |||
10357 | } | |||
10358 | ||||
10359 | /// Compare two cmp instructions. If IsCompatibility is true, function returns | |||
10360 | /// true if 2 cmps have same/swapped predicates and mos compatible corresponding | |||
10361 | /// operands. If IsCompatibility is false, function implements strict weak | |||
10362 | /// ordering relation between two cmp instructions, returning true if the first | |||
10363 | /// instruction is "less" than the second, i.e. its predicate is less than the | |||
10364 | /// predicate of the second or the operands IDs are less than the operands IDs | |||
10365 | /// of the second cmp instruction. | |||
10366 | template <bool IsCompatibility> | |||
10367 | static bool compareCmp(Value *V, Value *V2, | |||
10368 | function_ref<bool(Instruction *)> IsDeleted) { | |||
10369 | auto *CI1 = cast<CmpInst>(V); | |||
10370 | auto *CI2 = cast<CmpInst>(V2); | |||
10371 | if (IsDeleted(CI2) || !isValidElementType(CI2->getType())) | |||
10372 | return false; | |||
10373 | if (CI1->getOperand(0)->getType()->getTypeID() < | |||
10374 | CI2->getOperand(0)->getType()->getTypeID()) | |||
10375 | return !IsCompatibility; | |||
10376 | if (CI1->getOperand(0)->getType()->getTypeID() > | |||
10377 | CI2->getOperand(0)->getType()->getTypeID()) | |||
10378 | return false; | |||
10379 | CmpInst::Predicate Pred1 = CI1->getPredicate(); | |||
10380 | CmpInst::Predicate Pred2 = CI2->getPredicate(); | |||
10381 | CmpInst::Predicate SwapPred1 = CmpInst::getSwappedPredicate(Pred1); | |||
10382 | CmpInst::Predicate SwapPred2 = CmpInst::getSwappedPredicate(Pred2); | |||
10383 | CmpInst::Predicate BasePred1 = std::min(Pred1, SwapPred1); | |||
10384 | CmpInst::Predicate BasePred2 = std::min(Pred2, SwapPred2); | |||
10385 | if (BasePred1 < BasePred2) | |||
10386 | return !IsCompatibility; | |||
10387 | if (BasePred1 > BasePred2) | |||
10388 | return false; | |||
10389 | // Compare operands. | |||
10390 | bool LEPreds = Pred1 <= Pred2; | |||
10391 | bool GEPreds = Pred1 >= Pred2; | |||
10392 | for (int I = 0, E = CI1->getNumOperands(); I < E; ++I) { | |||
10393 | auto *Op1 = CI1->getOperand(LEPreds ? I : E - I - 1); | |||
10394 | auto *Op2 = CI2->getOperand(GEPreds ? I : E - I - 1); | |||
10395 | if (Op1->getValueID() < Op2->getValueID()) | |||
10396 | return !IsCompatibility; | |||
10397 | if (Op1->getValueID() > Op2->getValueID()) | |||
10398 | return false; | |||
10399 | if (auto *I1 = dyn_cast<Instruction>(Op1)) | |||
10400 | if (auto *I2 = dyn_cast<Instruction>(Op2)) { | |||
10401 | if (I1->getParent() != I2->getParent()) | |||
10402 | return false; | |||
10403 | InstructionsState S = getSameOpcode({I1, I2}); | |||
10404 | if (S.getOpcode()) | |||
10405 | continue; | |||
10406 | return false; | |||
10407 | } | |||
10408 | } | |||
10409 | return IsCompatibility; | |||
10410 | } | |||
10411 | ||||
10412 | bool SLPVectorizerPass::vectorizeSimpleInstructions( | |||
10413 | SmallVectorImpl<Instruction *> &Instructions, BasicBlock *BB, BoUpSLP &R, | |||
10414 | bool AtTerminator) { | |||
10415 | bool OpsChanged = false; | |||
10416 | SmallVector<Instruction *, 4> PostponedCmps; | |||
10417 | for (auto *I : reverse(Instructions)) { | |||
10418 | if (R.isDeleted(I)) | |||
10419 | continue; | |||
10420 | if (auto *LastInsertValue = dyn_cast<InsertValueInst>(I)) | |||
10421 | OpsChanged |= vectorizeInsertValueInst(LastInsertValue, BB, R); | |||
10422 | else if (auto *LastInsertElem = dyn_cast<InsertElementInst>(I)) | |||
10423 | OpsChanged |= vectorizeInsertElementInst(LastInsertElem, BB, R); | |||
10424 | else if (isa<CmpInst>(I)) | |||
10425 | PostponedCmps.push_back(I); | |||
10426 | } | |||
10427 | if (AtTerminator) { | |||
10428 | // Try to find reductions first. | |||
10429 | for (Instruction *I : PostponedCmps) { | |||
10430 | if (R.isDeleted(I)) | |||
10431 | continue; | |||
10432 | for (Value *Op : I->operands()) | |||
10433 | OpsChanged |= vectorizeRootInstruction(nullptr, Op, BB, R, TTI); | |||
10434 | } | |||
10435 | // Try to vectorize operands as vector bundles. | |||
10436 | for (Instruction *I : PostponedCmps) { | |||
10437 | if (R.isDeleted(I)) | |||
10438 | continue; | |||
10439 | OpsChanged |= tryToVectorize(I, R); | |||
10440 | } | |||
10441 | // Try to vectorize list of compares. | |||
10442 | // Sort by type, compare predicate, etc. | |||
10443 | auto &&CompareSorter = [&R](Value *V, Value *V2) { | |||
10444 | return compareCmp<false>(V, V2, | |||
10445 | [&R](Instruction *I) { return R.isDeleted(I); }); | |||
10446 | }; | |||
10447 | ||||
10448 | auto &&AreCompatibleCompares = [&R](Value *V1, Value *V2) { | |||
10449 | if (V1 == V2) | |||
10450 | return true; | |||
10451 | return compareCmp<true>(V1, V2, | |||
10452 | [&R](Instruction *I) { return R.isDeleted(I); }); | |||
10453 | }; | |||
10454 | auto Limit = [&R](Value *V) { | |||
10455 | unsigned EltSize = R.getVectorElementSize(V); | |||
10456 | return std::max(2U, R.getMaxVecRegSize() / EltSize); | |||
10457 | }; | |||
10458 | ||||
10459 | SmallVector<Value *> Vals(PostponedCmps.begin(), PostponedCmps.end()); | |||
10460 | OpsChanged |= tryToVectorizeSequence<Value>( | |||
10461 | Vals, Limit, CompareSorter, AreCompatibleCompares, | |||
10462 | [this, &R](ArrayRef<Value *> Candidates, bool LimitForRegisterSize) { | |||
10463 | // Exclude possible reductions from other blocks. | |||
10464 | bool ArePossiblyReducedInOtherBlock = | |||
10465 | any_of(Candidates, [](Value *V) { | |||
10466 | return any_of(V->users(), [V](User *U) { | |||
10467 | return isa<SelectInst>(U) && | |||
10468 | cast<SelectInst>(U)->getParent() != | |||
10469 | cast<Instruction>(V)->getParent(); | |||
10470 | }); | |||
10471 | }); | |||
10472 | if (ArePossiblyReducedInOtherBlock) | |||
10473 | return false; | |||
10474 | return tryToVectorizeList(Candidates, R, LimitForRegisterSize); | |||
10475 | }, | |||
10476 | /*LimitForRegisterSize=*/true); | |||
10477 | Instructions.clear(); | |||
10478 | } else { | |||
10479 | // Insert in reverse order since the PostponedCmps vector was filled in | |||
10480 | // reverse order. | |||
10481 | Instructions.assign(PostponedCmps.rbegin(), PostponedCmps.rend()); | |||
10482 | } | |||
10483 | return OpsChanged; | |||
10484 | } | |||
10485 | ||||
10486 | bool SLPVectorizerPass::vectorizeChainsInBlock(BasicBlock *BB, BoUpSLP &R) { | |||
10487 | bool Changed = false; | |||
10488 | SmallVector<Value *, 4> Incoming; | |||
10489 | SmallPtrSet<Value *, 16> VisitedInstrs; | |||
10490 | // Maps phi nodes to the non-phi nodes found in the use tree for each phi | |||
10491 | // node. Allows better to identify the chains that can be vectorized in the | |||
10492 | // better way. | |||
10493 | DenseMap<Value *, SmallVector<Value *, 4>> PHIToOpcodes; | |||
10494 | auto PHICompare = [this, &PHIToOpcodes](Value *V1, Value *V2) { | |||
10495 | assert(isValidElementType(V1->getType()) &&(static_cast <bool> (isValidElementType(V1->getType( )) && isValidElementType(V2->getType()) && "Expected vectorizable types only.") ? void (0) : __assert_fail ("isValidElementType(V1->getType()) && isValidElementType(V2->getType()) && \"Expected vectorizable types only.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 10497, __extension__ __PRETTY_FUNCTION__)) | |||
10496 | isValidElementType(V2->getType()) &&(static_cast <bool> (isValidElementType(V1->getType( )) && isValidElementType(V2->getType()) && "Expected vectorizable types only.") ? void (0) : __assert_fail ("isValidElementType(V1->getType()) && isValidElementType(V2->getType()) && \"Expected vectorizable types only.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 10497, __extension__ __PRETTY_FUNCTION__)) | |||
10497 | "Expected vectorizable types only.")(static_cast <bool> (isValidElementType(V1->getType( )) && isValidElementType(V2->getType()) && "Expected vectorizable types only.") ? void (0) : __assert_fail ("isValidElementType(V1->getType()) && isValidElementType(V2->getType()) && \"Expected vectorizable types only.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 10497, __extension__ __PRETTY_FUNCTION__)); | |||
10498 | // It is fine to compare type IDs here, since we expect only vectorizable | |||
10499 | // types, like ints, floats and pointers, we don't care about other type. | |||
10500 | if (V1->getType()->getTypeID() < V2->getType()->getTypeID()) | |||
10501 | return true; | |||
10502 | if (V1->getType()->getTypeID() > V2->getType()->getTypeID()) | |||
10503 | return false; | |||
10504 | ArrayRef<Value *> Opcodes1 = PHIToOpcodes[V1]; | |||
10505 | ArrayRef<Value *> Opcodes2 = PHIToOpcodes[V2]; | |||
10506 | if (Opcodes1.size() < Opcodes2.size()) | |||
10507 | return true; | |||
10508 | if (Opcodes1.size() > Opcodes2.size()) | |||
10509 | return false; | |||
10510 | Optional<bool> ConstOrder; | |||
10511 | for (int I = 0, E = Opcodes1.size(); I < E; ++I) { | |||
10512 | // Undefs are compatible with any other value. | |||
10513 | if (isa<UndefValue>(Opcodes1[I]) || isa<UndefValue>(Opcodes2[I])) { | |||
10514 | if (!ConstOrder) | |||
10515 | ConstOrder = | |||
10516 | !isa<UndefValue>(Opcodes1[I]) && isa<UndefValue>(Opcodes2[I]); | |||
10517 | continue; | |||
10518 | } | |||
10519 | if (auto *I1 = dyn_cast<Instruction>(Opcodes1[I])) | |||
10520 | if (auto *I2 = dyn_cast<Instruction>(Opcodes2[I])) { | |||
10521 | DomTreeNodeBase<BasicBlock> *NodeI1 = DT->getNode(I1->getParent()); | |||
10522 | DomTreeNodeBase<BasicBlock> *NodeI2 = DT->getNode(I2->getParent()); | |||
10523 | if (!NodeI1) | |||
10524 | return NodeI2 != nullptr; | |||
10525 | if (!NodeI2) | |||
10526 | return false; | |||
10527 | assert((NodeI1 == NodeI2) ==(static_cast <bool> ((NodeI1 == NodeI2) == (NodeI1-> getDFSNumIn() == NodeI2->getDFSNumIn()) && "Different nodes should have different DFS numbers" ) ? void (0) : __assert_fail ("(NodeI1 == NodeI2) == (NodeI1->getDFSNumIn() == NodeI2->getDFSNumIn()) && \"Different nodes should have different DFS numbers\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 10529, __extension__ __PRETTY_FUNCTION__)) | |||
10528 | (NodeI1->getDFSNumIn() == NodeI2->getDFSNumIn()) &&(static_cast <bool> ((NodeI1 == NodeI2) == (NodeI1-> getDFSNumIn() == NodeI2->getDFSNumIn()) && "Different nodes should have different DFS numbers" ) ? void (0) : __assert_fail ("(NodeI1 == NodeI2) == (NodeI1->getDFSNumIn() == NodeI2->getDFSNumIn()) && \"Different nodes should have different DFS numbers\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 10529, __extension__ __PRETTY_FUNCTION__)) | |||
10529 | "Different nodes should have different DFS numbers")(static_cast <bool> ((NodeI1 == NodeI2) == (NodeI1-> getDFSNumIn() == NodeI2->getDFSNumIn()) && "Different nodes should have different DFS numbers" ) ? void (0) : __assert_fail ("(NodeI1 == NodeI2) == (NodeI1->getDFSNumIn() == NodeI2->getDFSNumIn()) && \"Different nodes should have different DFS numbers\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 10529, __extension__ __PRETTY_FUNCTION__)); | |||
10530 | if (NodeI1 != NodeI2) | |||
10531 | return NodeI1->getDFSNumIn() < NodeI2->getDFSNumIn(); | |||
10532 | InstructionsState S = getSameOpcode({I1, I2}); | |||
10533 | if (S.getOpcode()) | |||
10534 | continue; | |||
10535 | return I1->getOpcode() < I2->getOpcode(); | |||
10536 | } | |||
10537 | if (isa<Constant>(Opcodes1[I]) && isa<Constant>(Opcodes2[I])) { | |||
10538 | if (!ConstOrder) | |||
10539 | ConstOrder = Opcodes1[I]->getValueID() < Opcodes2[I]->getValueID(); | |||
10540 | continue; | |||
10541 | } | |||
10542 | if (Opcodes1[I]->getValueID() < Opcodes2[I]->getValueID()) | |||
10543 | return true; | |||
10544 | if (Opcodes1[I]->getValueID() > Opcodes2[I]->getValueID()) | |||
10545 | return false; | |||
10546 | } | |||
10547 | return ConstOrder && *ConstOrder; | |||
10548 | }; | |||
10549 | auto AreCompatiblePHIs = [&PHIToOpcodes](Value *V1, Value *V2) { | |||
10550 | if (V1 == V2) | |||
10551 | return true; | |||
10552 | if (V1->getType() != V2->getType()) | |||
10553 | return false; | |||
10554 | ArrayRef<Value *> Opcodes1 = PHIToOpcodes[V1]; | |||
10555 | ArrayRef<Value *> Opcodes2 = PHIToOpcodes[V2]; | |||
10556 | if (Opcodes1.size() != Opcodes2.size()) | |||
10557 | return false; | |||
10558 | for (int I = 0, E = Opcodes1.size(); I < E; ++I) { | |||
10559 | // Undefs are compatible with any other value. | |||
10560 | if (isa<UndefValue>(Opcodes1[I]) || isa<UndefValue>(Opcodes2[I])) | |||
10561 | continue; | |||
10562 | if (auto *I1 = dyn_cast<Instruction>(Opcodes1[I])) | |||
10563 | if (auto *I2 = dyn_cast<Instruction>(Opcodes2[I])) { | |||
10564 | if (I1->getParent() != I2->getParent()) | |||
10565 | return false; | |||
10566 | InstructionsState S = getSameOpcode({I1, I2}); | |||
10567 | if (S.getOpcode()) | |||
10568 | continue; | |||
10569 | return false; | |||
10570 | } | |||
10571 | if (isa<Constant>(Opcodes1[I]) && isa<Constant>(Opcodes2[I])) | |||
10572 | continue; | |||
10573 | if (Opcodes1[I]->getValueID() != Opcodes2[I]->getValueID()) | |||
10574 | return false; | |||
10575 | } | |||
10576 | return true; | |||
10577 | }; | |||
10578 | auto Limit = [&R](Value *V) { | |||
10579 | unsigned EltSize = R.getVectorElementSize(V); | |||
10580 | return std::max(2U, R.getMaxVecRegSize() / EltSize); | |||
10581 | }; | |||
10582 | ||||
10583 | bool HaveVectorizedPhiNodes = false; | |||
10584 | do { | |||
10585 | // Collect the incoming values from the PHIs. | |||
10586 | Incoming.clear(); | |||
10587 | for (Instruction &I : *BB) { | |||
10588 | PHINode *P = dyn_cast<PHINode>(&I); | |||
10589 | if (!P) | |||
10590 | break; | |||
10591 | ||||
10592 | // No need to analyze deleted, vectorized and non-vectorizable | |||
10593 | // instructions. | |||
10594 | if (!VisitedInstrs.count(P) && !R.isDeleted(P) && | |||
10595 | isValidElementType(P->getType())) | |||
10596 | Incoming.push_back(P); | |||
10597 | } | |||
10598 | ||||
10599 | // Find the corresponding non-phi nodes for better matching when trying to | |||
10600 | // build the tree. | |||
10601 | for (Value *V : Incoming) { | |||
10602 | SmallVectorImpl<Value *> &Opcodes = | |||
10603 | PHIToOpcodes.try_emplace(V).first->getSecond(); | |||
10604 | if (!Opcodes.empty()) | |||
10605 | continue; | |||
10606 | SmallVector<Value *, 4> Nodes(1, V); | |||
10607 | SmallPtrSet<Value *, 4> Visited; | |||
10608 | while (!Nodes.empty()) { | |||
10609 | auto *PHI = cast<PHINode>(Nodes.pop_back_val()); | |||
10610 | if (!Visited.insert(PHI).second) | |||
10611 | continue; | |||
10612 | for (Value *V : PHI->incoming_values()) { | |||
10613 | if (auto *PHI1 = dyn_cast<PHINode>((V))) { | |||
10614 | Nodes.push_back(PHI1); | |||
10615 | continue; | |||
10616 | } | |||
10617 | Opcodes.emplace_back(V); | |||
10618 | } | |||
10619 | } | |||
10620 | } | |||
10621 | ||||
10622 | HaveVectorizedPhiNodes = tryToVectorizeSequence<Value>( | |||
10623 | Incoming, Limit, PHICompare, AreCompatiblePHIs, | |||
10624 | [this, &R](ArrayRef<Value *> Candidates, bool LimitForRegisterSize) { | |||
10625 | return tryToVectorizeList(Candidates, R, LimitForRegisterSize); | |||
10626 | }, | |||
10627 | /*LimitForRegisterSize=*/true); | |||
10628 | Changed |= HaveVectorizedPhiNodes; | |||
10629 | VisitedInstrs.insert(Incoming.begin(), Incoming.end()); | |||
10630 | } while (HaveVectorizedPhiNodes); | |||
10631 | ||||
10632 | VisitedInstrs.clear(); | |||
10633 | ||||
10634 | SmallVector<Instruction *, 8> PostProcessInstructions; | |||
10635 | SmallDenseSet<Instruction *, 4> KeyNodes; | |||
10636 | for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) { | |||
10637 | // Skip instructions with scalable type. The num of elements is unknown at | |||
10638 | // compile-time for scalable type. | |||
10639 | if (isa<ScalableVectorType>(it->getType())) | |||
10640 | continue; | |||
10641 | ||||
10642 | // Skip instructions marked for the deletion. | |||
10643 | if (R.isDeleted(&*it)) | |||
10644 | continue; | |||
10645 | // We may go through BB multiple times so skip the one we have checked. | |||
10646 | if (!VisitedInstrs.insert(&*it).second) { | |||
10647 | if (it->use_empty() && KeyNodes.contains(&*it) && | |||
10648 | vectorizeSimpleInstructions(PostProcessInstructions, BB, R, | |||
10649 | it->isTerminator())) { | |||
10650 | // We would like to start over since some instructions are deleted | |||
10651 | // and the iterator may become invalid value. | |||
10652 | Changed = true; | |||
10653 | it = BB->begin(); | |||
10654 | e = BB->end(); | |||
10655 | } | |||
10656 | continue; | |||
10657 | } | |||
10658 | ||||
10659 | if (isa<DbgInfoIntrinsic>(it)) | |||
10660 | continue; | |||
10661 | ||||
10662 | // Try to vectorize reductions that use PHINodes. | |||
10663 | if (PHINode *P = dyn_cast<PHINode>(it)) { | |||
10664 | // Check that the PHI is a reduction PHI. | |||
10665 | if (P->getNumIncomingValues() == 2) { | |||
10666 | // Try to match and vectorize a horizontal reduction. | |||
10667 | if (vectorizeRootInstruction(P, getReductionValue(DT, P, BB, LI), BB, R, | |||
10668 | TTI)) { | |||
10669 | Changed = true; | |||
10670 | it = BB->begin(); | |||
10671 | e = BB->end(); | |||
10672 | continue; | |||
10673 | } | |||
10674 | } | |||
10675 | // Try to vectorize the incoming values of the PHI, to catch reductions | |||
10676 | // that feed into PHIs. | |||
10677 | for (unsigned I = 0, E = P->getNumIncomingValues(); I != E; I++) { | |||
10678 | // Skip if the incoming block is the current BB for now. Also, bypass | |||
10679 | // unreachable IR for efficiency and to avoid crashing. | |||
10680 | // TODO: Collect the skipped incoming values and try to vectorize them | |||
10681 | // after processing BB. | |||
10682 | if (BB == P->getIncomingBlock(I) || | |||
10683 | !DT->isReachableFromEntry(P->getIncomingBlock(I))) | |||
10684 | continue; | |||
10685 | ||||
10686 | Changed |= vectorizeRootInstruction(nullptr, P->getIncomingValue(I), | |||
10687 | P->getIncomingBlock(I), R, TTI); | |||
10688 | } | |||
10689 | continue; | |||
10690 | } | |||
10691 | ||||
10692 | // Ran into an instruction without users, like terminator, or function call | |||
10693 | // with ignored return value, store. Ignore unused instructions (basing on | |||
10694 | // instruction type, except for CallInst and InvokeInst). | |||
10695 | if (it->use_empty() && (it->getType()->isVoidTy() || isa<CallInst>(it) || | |||
10696 | isa<InvokeInst>(it))) { | |||
10697 | KeyNodes.insert(&*it); | |||
10698 | bool OpsChanged = false; | |||
10699 | if (ShouldStartVectorizeHorAtStore || !isa<StoreInst>(it)) { | |||
10700 | for (auto *V : it->operand_values()) { | |||
10701 | // Try to match and vectorize a horizontal reduction. | |||
10702 | OpsChanged |= vectorizeRootInstruction(nullptr, V, BB, R, TTI); | |||
10703 | } | |||
10704 | } | |||
10705 | // Start vectorization of post-process list of instructions from the | |||
10706 | // top-tree instructions to try to vectorize as many instructions as | |||
10707 | // possible. | |||
10708 | OpsChanged |= vectorizeSimpleInstructions(PostProcessInstructions, BB, R, | |||
10709 | it->isTerminator()); | |||
10710 | if (OpsChanged) { | |||
10711 | // We would like to start over since some instructions are deleted | |||
10712 | // and the iterator may become invalid value. | |||
10713 | Changed = true; | |||
10714 | it = BB->begin(); | |||
10715 | e = BB->end(); | |||
10716 | continue; | |||
10717 | } | |||
10718 | } | |||
10719 | ||||
10720 | if (isa<InsertElementInst>(it) || isa<CmpInst>(it) || | |||
10721 | isa<InsertValueInst>(it)) | |||
10722 | PostProcessInstructions.push_back(&*it); | |||
10723 | } | |||
10724 | ||||
10725 | return Changed; | |||
10726 | } | |||
10727 | ||||
10728 | bool SLPVectorizerPass::vectorizeGEPIndices(BasicBlock *BB, BoUpSLP &R) { | |||
10729 | auto Changed = false; | |||
10730 | for (auto &Entry : GEPs) { | |||
10731 | // If the getelementptr list has fewer than two elements, there's nothing | |||
10732 | // to do. | |||
10733 | if (Entry.second.size() < 2) | |||
10734 | continue; | |||
10735 | ||||
10736 | LLVM_DEBUG(dbgs() << "SLP: Analyzing a getelementptr list of length "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing a getelementptr list of length " << Entry.second.size() << ".\n"; } } while (false ) | |||
10737 | << Entry.second.size() << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing a getelementptr list of length " << Entry.second.size() << ".\n"; } } while (false ); | |||
10738 | ||||
10739 | // Process the GEP list in chunks suitable for the target's supported | |||
10740 | // vector size. If a vector register can't hold 1 element, we are done. We | |||
10741 | // are trying to vectorize the index computations, so the maximum number of | |||
10742 | // elements is based on the size of the index expression, rather than the | |||
10743 | // size of the GEP itself (the target's pointer size). | |||
10744 | unsigned MaxVecRegSize = R.getMaxVecRegSize(); | |||
10745 | unsigned EltSize = R.getVectorElementSize(*Entry.second[0]->idx_begin()); | |||
10746 | if (MaxVecRegSize < EltSize) | |||
10747 | continue; | |||
10748 | ||||
10749 | unsigned MaxElts = MaxVecRegSize / EltSize; | |||
10750 | for (unsigned BI = 0, BE = Entry.second.size(); BI < BE; BI += MaxElts) { | |||
10751 | auto Len = std::min<unsigned>(BE - BI, MaxElts); | |||
10752 | ArrayRef<GetElementPtrInst *> GEPList(&Entry.second[BI], Len); | |||
10753 | ||||
10754 | // Initialize a set a candidate getelementptrs. Note that we use a | |||
10755 | // SetVector here to preserve program order. If the index computations | |||
10756 | // are vectorizable and begin with loads, we want to minimize the chance | |||
10757 | // of having to reorder them later. | |||
10758 | SetVector<Value *> Candidates(GEPList.begin(), GEPList.end()); | |||
10759 | ||||
10760 | // Some of the candidates may have already been vectorized after we | |||
10761 | // initially collected them. If so, they are marked as deleted, so remove | |||
10762 | // them from the set of candidates. | |||
10763 | Candidates.remove_if( | |||
10764 | [&R](Value *I) { return R.isDeleted(cast<Instruction>(I)); }); | |||
10765 | ||||
10766 | // Remove from the set of candidates all pairs of getelementptrs with | |||
10767 | // constant differences. Such getelementptrs are likely not good | |||
10768 | // candidates for vectorization in a bottom-up phase since one can be | |||
10769 | // computed from the other. We also ensure all candidate getelementptr | |||
10770 | // indices are unique. | |||
10771 | for (int I = 0, E = GEPList.size(); I < E && Candidates.size() > 1; ++I) { | |||
10772 | auto *GEPI = GEPList[I]; | |||
10773 | if (!Candidates.count(GEPI)) | |||
10774 | continue; | |||
10775 | auto *SCEVI = SE->getSCEV(GEPList[I]); | |||
10776 | for (int J = I + 1; J < E && Candidates.size() > 1; ++J) { | |||
10777 | auto *GEPJ = GEPList[J]; | |||
10778 | auto *SCEVJ = SE->getSCEV(GEPList[J]); | |||
10779 | if (isa<SCEVConstant>(SE->getMinusSCEV(SCEVI, SCEVJ))) { | |||
10780 | Candidates.remove(GEPI); | |||
10781 | Candidates.remove(GEPJ); | |||
10782 | } else if (GEPI->idx_begin()->get() == GEPJ->idx_begin()->get()) { | |||
10783 | Candidates.remove(GEPJ); | |||
10784 | } | |||
10785 | } | |||
10786 | } | |||
10787 | ||||
10788 | // We break out of the above computation as soon as we know there are | |||
10789 | // fewer than two candidates remaining. | |||
10790 | if (Candidates.size() < 2) | |||
10791 | continue; | |||
10792 | ||||
10793 | // Add the single, non-constant index of each candidate to the bundle. We | |||
10794 | // ensured the indices met these constraints when we originally collected | |||
10795 | // the getelementptrs. | |||
10796 | SmallVector<Value *, 16> Bundle(Candidates.size()); | |||
10797 | auto BundleIndex = 0u; | |||
10798 | for (auto *V : Candidates) { | |||
10799 | auto *GEP = cast<GetElementPtrInst>(V); | |||
10800 | auto *GEPIdx = GEP->idx_begin()->get(); | |||
10801 | assert(GEP->getNumIndices() == 1 || !isa<Constant>(GEPIdx))(static_cast <bool> (GEP->getNumIndices() == 1 || !isa <Constant>(GEPIdx)) ? void (0) : __assert_fail ("GEP->getNumIndices() == 1 || !isa<Constant>(GEPIdx)" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 10801, __extension__ __PRETTY_FUNCTION__)); | |||
10802 | Bundle[BundleIndex++] = GEPIdx; | |||
10803 | } | |||
10804 | ||||
10805 | // Try and vectorize the indices. We are currently only interested in | |||
10806 | // gather-like cases of the form: | |||
10807 | // | |||
10808 | // ... = g[a[0] - b[0]] + g[a[1] - b[1]] + ... | |||
10809 | // | |||
10810 | // where the loads of "a", the loads of "b", and the subtractions can be | |||
10811 | // performed in parallel. It's likely that detecting this pattern in a | |||
10812 | // bottom-up phase will be simpler and less costly than building a | |||
10813 | // full-blown top-down phase beginning at the consecutive loads. | |||
10814 | Changed |= tryToVectorizeList(Bundle, R); | |||
10815 | } | |||
10816 | } | |||
10817 | return Changed; | |||
10818 | } | |||
10819 | ||||
10820 | bool SLPVectorizerPass::vectorizeStoreChains(BoUpSLP &R) { | |||
10821 | bool Changed = false; | |||
10822 | // Sort by type, base pointers and values operand. Value operands must be | |||
10823 | // compatible (have the same opcode, same parent), otherwise it is | |||
10824 | // definitely not profitable to try to vectorize them. | |||
10825 | auto &&StoreSorter = [this](StoreInst *V, StoreInst *V2) { | |||
10826 | if (V->getPointerOperandType()->getTypeID() < | |||
10827 | V2->getPointerOperandType()->getTypeID()) | |||
10828 | return true; | |||
10829 | if (V->getPointerOperandType()->getTypeID() > | |||
10830 | V2->getPointerOperandType()->getTypeID()) | |||
10831 | return false; | |||
10832 | // UndefValues are compatible with all other values. | |||
10833 | if (isa<UndefValue>(V->getValueOperand()) || | |||
10834 | isa<UndefValue>(V2->getValueOperand())) | |||
10835 | return false; | |||
10836 | if (auto *I1 = dyn_cast<Instruction>(V->getValueOperand())) | |||
10837 | if (auto *I2 = dyn_cast<Instruction>(V2->getValueOperand())) { | |||
10838 | DomTreeNodeBase<llvm::BasicBlock> *NodeI1 = | |||
10839 | DT->getNode(I1->getParent()); | |||
10840 | DomTreeNodeBase<llvm::BasicBlock> *NodeI2 = | |||
10841 | DT->getNode(I2->getParent()); | |||
10842 | assert(NodeI1 && "Should only process reachable instructions")(static_cast <bool> (NodeI1 && "Should only process reachable instructions" ) ? void (0) : __assert_fail ("NodeI1 && \"Should only process reachable instructions\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 10842, __extension__ __PRETTY_FUNCTION__)); | |||
10843 | assert(NodeI1 && "Should only process reachable instructions")(static_cast <bool> (NodeI1 && "Should only process reachable instructions" ) ? void (0) : __assert_fail ("NodeI1 && \"Should only process reachable instructions\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 10843, __extension__ __PRETTY_FUNCTION__)); | |||
10844 | assert((NodeI1 == NodeI2) ==(static_cast <bool> ((NodeI1 == NodeI2) == (NodeI1-> getDFSNumIn() == NodeI2->getDFSNumIn()) && "Different nodes should have different DFS numbers" ) ? void (0) : __assert_fail ("(NodeI1 == NodeI2) == (NodeI1->getDFSNumIn() == NodeI2->getDFSNumIn()) && \"Different nodes should have different DFS numbers\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 10846, __extension__ __PRETTY_FUNCTION__)) | |||
10845 | (NodeI1->getDFSNumIn() == NodeI2->getDFSNumIn()) &&(static_cast <bool> ((NodeI1 == NodeI2) == (NodeI1-> getDFSNumIn() == NodeI2->getDFSNumIn()) && "Different nodes should have different DFS numbers" ) ? void (0) : __assert_fail ("(NodeI1 == NodeI2) == (NodeI1->getDFSNumIn() == NodeI2->getDFSNumIn()) && \"Different nodes should have different DFS numbers\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 10846, __extension__ __PRETTY_FUNCTION__)) | |||
10846 | "Different nodes should have different DFS numbers")(static_cast <bool> ((NodeI1 == NodeI2) == (NodeI1-> getDFSNumIn() == NodeI2->getDFSNumIn()) && "Different nodes should have different DFS numbers" ) ? void (0) : __assert_fail ("(NodeI1 == NodeI2) == (NodeI1->getDFSNumIn() == NodeI2->getDFSNumIn()) && \"Different nodes should have different DFS numbers\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 10846, __extension__ __PRETTY_FUNCTION__)); | |||
10847 | if (NodeI1 != NodeI2) | |||
10848 | return NodeI1->getDFSNumIn() < NodeI2->getDFSNumIn(); | |||
10849 | InstructionsState S = getSameOpcode({I1, I2}); | |||
10850 | if (S.getOpcode()) | |||
10851 | return false; | |||
10852 | return I1->getOpcode() < I2->getOpcode(); | |||
10853 | } | |||
10854 | if (isa<Constant>(V->getValueOperand()) && | |||
10855 | isa<Constant>(V2->getValueOperand())) | |||
10856 | return false; | |||
10857 | return V->getValueOperand()->getValueID() < | |||
10858 | V2->getValueOperand()->getValueID(); | |||
10859 | }; | |||
10860 | ||||
10861 | auto &&AreCompatibleStores = [](StoreInst *V1, StoreInst *V2) { | |||
10862 | if (V1 == V2) | |||
10863 | return true; | |||
10864 | if (V1->getPointerOperandType() != V2->getPointerOperandType()) | |||
10865 | return false; | |||
10866 | // Undefs are compatible with any other value. | |||
10867 | if (isa<UndefValue>(V1->getValueOperand()) || | |||
10868 | isa<UndefValue>(V2->getValueOperand())) | |||
10869 | return true; | |||
10870 | if (auto *I1 = dyn_cast<Instruction>(V1->getValueOperand())) | |||
10871 | if (auto *I2 = dyn_cast<Instruction>(V2->getValueOperand())) { | |||
10872 | if (I1->getParent() != I2->getParent()) | |||
10873 | return false; | |||
10874 | InstructionsState S = getSameOpcode({I1, I2}); | |||
10875 | return S.getOpcode() > 0; | |||
10876 | } | |||
10877 | if (isa<Constant>(V1->getValueOperand()) && | |||
10878 | isa<Constant>(V2->getValueOperand())) | |||
10879 | return true; | |||
10880 | return V1->getValueOperand()->getValueID() == | |||
10881 | V2->getValueOperand()->getValueID(); | |||
10882 | }; | |||
10883 | auto Limit = [&R, this](StoreInst *SI) { | |||
10884 | unsigned EltSize = DL->getTypeSizeInBits(SI->getValueOperand()->getType()); | |||
10885 | return R.getMinVF(EltSize); | |||
10886 | }; | |||
10887 | ||||
10888 | // Attempt to sort and vectorize each of the store-groups. | |||
10889 | for (auto &Pair : Stores) { | |||
10890 | if (Pair.second.size() < 2) | |||
10891 | continue; | |||
10892 | ||||
10893 | LLVM_DEBUG(dbgs() << "SLP: Analyzing a store chain of length "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing a store chain of length " << Pair.second.size() << ".\n"; } } while (false ) | |||
10894 | << Pair.second.size() << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing a store chain of length " << Pair.second.size() << ".\n"; } } while (false ); | |||
10895 | ||||
10896 | if (!isValidElementType(Pair.second.front()->getValueOperand()->getType())) | |||
10897 | continue; | |||
10898 | ||||
10899 | Changed |= tryToVectorizeSequence<StoreInst>( | |||
10900 | Pair.second, Limit, StoreSorter, AreCompatibleStores, | |||
10901 | [this, &R](ArrayRef<StoreInst *> Candidates, bool) { | |||
10902 | return vectorizeStores(Candidates, R); | |||
10903 | }, | |||
10904 | /*LimitForRegisterSize=*/false); | |||
10905 | } | |||
10906 | return Changed; | |||
10907 | } | |||
10908 | ||||
10909 | char SLPVectorizer::ID = 0; | |||
10910 | ||||
10911 | static const char lv_name[] = "SLP Vectorizer"; | |||
10912 | ||||
10913 | INITIALIZE_PASS_BEGIN(SLPVectorizer, SV_NAME, lv_name, false, false)static void *initializeSLPVectorizerPassOnce(PassRegistry & Registry) { | |||
10914 | INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)initializeAAResultsWrapperPassPass(Registry); | |||
10915 | INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)initializeTargetTransformInfoWrapperPassPass(Registry); | |||
10916 | INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)initializeAssumptionCacheTrackerPass(Registry); | |||
10917 | INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)initializeScalarEvolutionWrapperPassPass(Registry); | |||
10918 | INITIALIZE_PASS_DEPENDENCY(LoopSimplify)initializeLoopSimplifyPass(Registry); | |||
10919 | INITIALIZE_PASS_DEPENDENCY(DemandedBitsWrapperPass)initializeDemandedBitsWrapperPassPass(Registry); | |||
10920 | INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass)initializeOptimizationRemarkEmitterWrapperPassPass(Registry); | |||
10921 | INITIALIZE_PASS_DEPENDENCY(InjectTLIMappingsLegacy)initializeInjectTLIMappingsLegacyPass(Registry); | |||
10922 | INITIALIZE_PASS_END(SLPVectorizer, SV_NAME, lv_name, false, false)PassInfo *PI = new PassInfo( lv_name, "slp-vectorizer", & SLPVectorizer::ID, PassInfo::NormalCtor_t(callDefaultCtor< SLPVectorizer>), false, false); Registry.registerPass(*PI, true); return PI; } static llvm::once_flag InitializeSLPVectorizerPassFlag ; void llvm::initializeSLPVectorizerPass(PassRegistry &Registry ) { llvm::call_once(InitializeSLPVectorizerPassFlag, initializeSLPVectorizerPassOnce , std::ref(Registry)); } | |||
10923 | ||||
10924 | Pass *llvm::createSLPVectorizerPass() { return new SLPVectorizer(); } |