File: | build/source/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp |
Warning: | line 11663, column 30 Called C++ object pointer is null |
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/PostOrderIterator.h" | |||
23 | #include "llvm/ADT/PriorityQueue.h" | |||
24 | #include "llvm/ADT/STLExtras.h" | |||
25 | #include "llvm/ADT/SetOperations.h" | |||
26 | #include "llvm/ADT/SetVector.h" | |||
27 | #include "llvm/ADT/SmallBitVector.h" | |||
28 | #include "llvm/ADT/SmallPtrSet.h" | |||
29 | #include "llvm/ADT/SmallSet.h" | |||
30 | #include "llvm/ADT/SmallString.h" | |||
31 | #include "llvm/ADT/Statistic.h" | |||
32 | #include "llvm/ADT/iterator.h" | |||
33 | #include "llvm/ADT/iterator_range.h" | |||
34 | #include "llvm/Analysis/AliasAnalysis.h" | |||
35 | #include "llvm/Analysis/AssumptionCache.h" | |||
36 | #include "llvm/Analysis/CodeMetrics.h" | |||
37 | #include "llvm/Analysis/DemandedBits.h" | |||
38 | #include "llvm/Analysis/GlobalsModRef.h" | |||
39 | #include "llvm/Analysis/IVDescriptors.h" | |||
40 | #include "llvm/Analysis/LoopAccessAnalysis.h" | |||
41 | #include "llvm/Analysis/LoopInfo.h" | |||
42 | #include "llvm/Analysis/MemoryLocation.h" | |||
43 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" | |||
44 | #include "llvm/Analysis/ScalarEvolution.h" | |||
45 | #include "llvm/Analysis/ScalarEvolutionExpressions.h" | |||
46 | #include "llvm/Analysis/TargetLibraryInfo.h" | |||
47 | #include "llvm/Analysis/TargetTransformInfo.h" | |||
48 | #include "llvm/Analysis/ValueTracking.h" | |||
49 | #include "llvm/Analysis/VectorUtils.h" | |||
50 | #include "llvm/IR/Attributes.h" | |||
51 | #include "llvm/IR/BasicBlock.h" | |||
52 | #include "llvm/IR/Constant.h" | |||
53 | #include "llvm/IR/Constants.h" | |||
54 | #include "llvm/IR/DataLayout.h" | |||
55 | #include "llvm/IR/DerivedTypes.h" | |||
56 | #include "llvm/IR/Dominators.h" | |||
57 | #include "llvm/IR/Function.h" | |||
58 | #include "llvm/IR/IRBuilder.h" | |||
59 | #include "llvm/IR/InstrTypes.h" | |||
60 | #include "llvm/IR/Instruction.h" | |||
61 | #include "llvm/IR/Instructions.h" | |||
62 | #include "llvm/IR/IntrinsicInst.h" | |||
63 | #include "llvm/IR/Intrinsics.h" | |||
64 | #include "llvm/IR/Module.h" | |||
65 | #include "llvm/IR/Operator.h" | |||
66 | #include "llvm/IR/PatternMatch.h" | |||
67 | #include "llvm/IR/Type.h" | |||
68 | #include "llvm/IR/Use.h" | |||
69 | #include "llvm/IR/User.h" | |||
70 | #include "llvm/IR/Value.h" | |||
71 | #include "llvm/IR/ValueHandle.h" | |||
72 | #ifdef EXPENSIVE_CHECKS | |||
73 | #include "llvm/IR/Verifier.h" | |||
74 | #endif | |||
75 | #include "llvm/Pass.h" | |||
76 | #include "llvm/Support/Casting.h" | |||
77 | #include "llvm/Support/CommandLine.h" | |||
78 | #include "llvm/Support/Compiler.h" | |||
79 | #include "llvm/Support/DOTGraphTraits.h" | |||
80 | #include "llvm/Support/Debug.h" | |||
81 | #include "llvm/Support/ErrorHandling.h" | |||
82 | #include "llvm/Support/GraphWriter.h" | |||
83 | #include "llvm/Support/InstructionCost.h" | |||
84 | #include "llvm/Support/KnownBits.h" | |||
85 | #include "llvm/Support/MathExtras.h" | |||
86 | #include "llvm/Support/raw_ostream.h" | |||
87 | #include "llvm/Transforms/Utils/InjectTLIMappings.h" | |||
88 | #include "llvm/Transforms/Utils/Local.h" | |||
89 | #include "llvm/Transforms/Utils/LoopUtils.h" | |||
90 | #include <algorithm> | |||
91 | #include <cassert> | |||
92 | #include <cstdint> | |||
93 | #include <iterator> | |||
94 | #include <memory> | |||
95 | #include <optional> | |||
96 | #include <set> | |||
97 | #include <string> | |||
98 | #include <tuple> | |||
99 | #include <utility> | |||
100 | #include <vector> | |||
101 | ||||
102 | using namespace llvm; | |||
103 | using namespace llvm::PatternMatch; | |||
104 | using namespace slpvectorizer; | |||
105 | ||||
106 | #define SV_NAME"slp-vectorizer" "slp-vectorizer" | |||
107 | #define DEBUG_TYPE"SLP" "SLP" | |||
108 | ||||
109 | STATISTIC(NumVectorInstructions, "Number of vector instructions generated")static llvm::Statistic NumVectorInstructions = {"SLP", "NumVectorInstructions" , "Number of vector instructions generated"}; | |||
110 | ||||
111 | cl::opt<bool> RunSLPVectorization("vectorize-slp", cl::init(true), cl::Hidden, | |||
112 | cl::desc("Run the SLP vectorization passes")); | |||
113 | ||||
114 | static cl::opt<int> | |||
115 | SLPCostThreshold("slp-threshold", cl::init(0), cl::Hidden, | |||
116 | cl::desc("Only vectorize if you gain more than this " | |||
117 | "number ")); | |||
118 | ||||
119 | static cl::opt<bool> | |||
120 | ShouldVectorizeHor("slp-vectorize-hor", cl::init(true), cl::Hidden, | |||
121 | cl::desc("Attempt to vectorize horizontal reductions")); | |||
122 | ||||
123 | static cl::opt<bool> ShouldStartVectorizeHorAtStore( | |||
124 | "slp-vectorize-hor-store", cl::init(false), cl::Hidden, | |||
125 | cl::desc( | |||
126 | "Attempt to vectorize horizontal reductions feeding into a store")); | |||
127 | ||||
128 | // NOTE: If AllowHorRdxIdenityOptimization is true, the optimization will run | |||
129 | // even if we match a reduction but do not vectorize in the end. | |||
130 | static cl::opt<bool> AllowHorRdxIdenityOptimization( | |||
131 | "slp-optimize-identity-hor-reduction-ops", cl::init(true), cl::Hidden, | |||
132 | cl::desc("Allow optimization of original scalar identity operations on " | |||
133 | "matched horizontal reductions.")); | |||
134 | ||||
135 | static cl::opt<int> | |||
136 | MaxVectorRegSizeOption("slp-max-reg-size", cl::init(128), cl::Hidden, | |||
137 | cl::desc("Attempt to vectorize for this register size in bits")); | |||
138 | ||||
139 | static cl::opt<unsigned> | |||
140 | MaxVFOption("slp-max-vf", cl::init(0), cl::Hidden, | |||
141 | cl::desc("Maximum SLP vectorization factor (0=unlimited)")); | |||
142 | ||||
143 | static cl::opt<int> | |||
144 | MaxStoreLookup("slp-max-store-lookup", cl::init(32), cl::Hidden, | |||
145 | cl::desc("Maximum depth of the lookup for consecutive stores.")); | |||
146 | ||||
147 | /// Limits the size of scheduling regions in a block. | |||
148 | /// It avoid long compile times for _very_ large blocks where vector | |||
149 | /// instructions are spread over a wide range. | |||
150 | /// This limit is way higher than needed by real-world functions. | |||
151 | static cl::opt<int> | |||
152 | ScheduleRegionSizeBudget("slp-schedule-budget", cl::init(100000), cl::Hidden, | |||
153 | cl::desc("Limit the size of the SLP scheduling region per block")); | |||
154 | ||||
155 | static cl::opt<int> MinVectorRegSizeOption( | |||
156 | "slp-min-reg-size", cl::init(128), cl::Hidden, | |||
157 | cl::desc("Attempt to vectorize for this register size in bits")); | |||
158 | ||||
159 | static cl::opt<unsigned> RecursionMaxDepth( | |||
160 | "slp-recursion-max-depth", cl::init(12), cl::Hidden, | |||
161 | cl::desc("Limit the recursion depth when building a vectorizable tree")); | |||
162 | ||||
163 | static cl::opt<unsigned> MinTreeSize( | |||
164 | "slp-min-tree-size", cl::init(3), cl::Hidden, | |||
165 | cl::desc("Only vectorize small trees if they are fully vectorizable")); | |||
166 | ||||
167 | // The maximum depth that the look-ahead score heuristic will explore. | |||
168 | // The higher this value, the higher the compilation time overhead. | |||
169 | static cl::opt<int> LookAheadMaxDepth( | |||
170 | "slp-max-look-ahead-depth", cl::init(2), cl::Hidden, | |||
171 | cl::desc("The maximum look-ahead depth for operand reordering scores")); | |||
172 | ||||
173 | // The maximum depth that the look-ahead score heuristic will explore | |||
174 | // when it probing among candidates for vectorization tree roots. | |||
175 | // The higher this value, the higher the compilation time overhead but unlike | |||
176 | // similar limit for operands ordering this is less frequently used, hence | |||
177 | // impact of higher value is less noticeable. | |||
178 | static cl::opt<int> RootLookAheadMaxDepth( | |||
179 | "slp-max-root-look-ahead-depth", cl::init(2), cl::Hidden, | |||
180 | cl::desc("The maximum look-ahead depth for searching best rooting option")); | |||
181 | ||||
182 | static cl::opt<bool> | |||
183 | ViewSLPTree("view-slp-tree", cl::Hidden, | |||
184 | cl::desc("Display the SLP trees with Graphviz")); | |||
185 | ||||
186 | // Limit the number of alias checks. The limit is chosen so that | |||
187 | // it has no negative effect on the llvm benchmarks. | |||
188 | static const unsigned AliasedCheckLimit = 10; | |||
189 | ||||
190 | // Another limit for the alias checks: The maximum distance between load/store | |||
191 | // instructions where alias checks are done. | |||
192 | // This limit is useful for very large basic blocks. | |||
193 | static const unsigned MaxMemDepDistance = 160; | |||
194 | ||||
195 | /// If the ScheduleRegionSizeBudget is exhausted, we allow small scheduling | |||
196 | /// regions to be handled. | |||
197 | static const int MinScheduleRegionSize = 16; | |||
198 | ||||
199 | /// Predicate for the element types that the SLP vectorizer supports. | |||
200 | /// | |||
201 | /// The most important thing to filter here are types which are invalid in LLVM | |||
202 | /// vectors. We also filter target specific types which have absolutely no | |||
203 | /// meaningful vectorization path such as x86_fp80 and ppc_f128. This just | |||
204 | /// avoids spending time checking the cost model and realizing that they will | |||
205 | /// be inevitably scalarized. | |||
206 | static bool isValidElementType(Type *Ty) { | |||
207 | return VectorType::isValidElementType(Ty) && !Ty->isX86_FP80Ty() && | |||
208 | !Ty->isPPC_FP128Ty(); | |||
209 | } | |||
210 | ||||
211 | /// \returns True if the value is a constant (but not globals/constant | |||
212 | /// expressions). | |||
213 | static bool isConstant(Value *V) { | |||
214 | return isa<Constant>(V) && !isa<ConstantExpr, GlobalValue>(V); | |||
215 | } | |||
216 | ||||
217 | /// Checks if \p V is one of vector-like instructions, i.e. undef, | |||
218 | /// insertelement/extractelement with constant indices for fixed vector type or | |||
219 | /// extractvalue instruction. | |||
220 | static bool isVectorLikeInstWithConstOps(Value *V) { | |||
221 | if (!isa<InsertElementInst, ExtractElementInst>(V) && | |||
222 | !isa<ExtractValueInst, UndefValue>(V)) | |||
223 | return false; | |||
224 | auto *I = dyn_cast<Instruction>(V); | |||
225 | if (!I || isa<ExtractValueInst>(I)) | |||
226 | return true; | |||
227 | if (!isa<FixedVectorType>(I->getOperand(0)->getType())) | |||
228 | return false; | |||
229 | if (isa<ExtractElementInst>(I)) | |||
230 | return isConstant(I->getOperand(1)); | |||
231 | 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", 231, __extension__ __PRETTY_FUNCTION__)); | |||
232 | return isConstant(I->getOperand(2)); | |||
233 | } | |||
234 | ||||
235 | /// \returns true if all of the instructions in \p VL are in the same block or | |||
236 | /// false otherwise. | |||
237 | static bool allSameBlock(ArrayRef<Value *> VL) { | |||
238 | Instruction *I0 = dyn_cast<Instruction>(VL[0]); | |||
239 | if (!I0) | |||
240 | return false; | |||
241 | if (all_of(VL, isVectorLikeInstWithConstOps)) | |||
242 | return true; | |||
243 | ||||
244 | BasicBlock *BB = I0->getParent(); | |||
245 | for (int I = 1, E = VL.size(); I < E; I++) { | |||
246 | auto *II = dyn_cast<Instruction>(VL[I]); | |||
247 | if (!II) | |||
248 | return false; | |||
249 | ||||
250 | if (BB != II->getParent()) | |||
251 | return false; | |||
252 | } | |||
253 | return true; | |||
254 | } | |||
255 | ||||
256 | /// \returns True if all of the values in \p VL are constants (but not | |||
257 | /// globals/constant expressions). | |||
258 | static bool allConstant(ArrayRef<Value *> VL) { | |||
259 | // Constant expressions and globals can't be vectorized like normal integer/FP | |||
260 | // constants. | |||
261 | return all_of(VL, isConstant); | |||
262 | } | |||
263 | ||||
264 | /// \returns True if all of the values in \p VL are identical or some of them | |||
265 | /// are UndefValue. | |||
266 | static bool isSplat(ArrayRef<Value *> VL) { | |||
267 | Value *FirstNonUndef = nullptr; | |||
268 | for (Value *V : VL) { | |||
269 | if (isa<UndefValue>(V)) | |||
270 | continue; | |||
271 | if (!FirstNonUndef) { | |||
272 | FirstNonUndef = V; | |||
273 | continue; | |||
274 | } | |||
275 | if (V != FirstNonUndef) | |||
276 | return false; | |||
277 | } | |||
278 | return FirstNonUndef != nullptr; | |||
279 | } | |||
280 | ||||
281 | /// \returns True if \p I is commutative, handles CmpInst and BinaryOperator. | |||
282 | static bool isCommutative(Instruction *I) { | |||
283 | if (auto *Cmp = dyn_cast<CmpInst>(I)) | |||
284 | return Cmp->isCommutative(); | |||
285 | if (auto *BO = dyn_cast<BinaryOperator>(I)) | |||
286 | return BO->isCommutative(); | |||
287 | // TODO: This should check for generic Instruction::isCommutative(), but | |||
288 | // we need to confirm that the caller code correctly handles Intrinsics | |||
289 | // for example (does not have 2 operands). | |||
290 | return false; | |||
291 | } | |||
292 | ||||
293 | /// \returns inserting index of InsertElement or InsertValue instruction, | |||
294 | /// using Offset as base offset for index. | |||
295 | static std::optional<unsigned> getInsertIndex(const Value *InsertInst, | |||
296 | unsigned Offset = 0) { | |||
297 | int Index = Offset; | |||
298 | if (const auto *IE = dyn_cast<InsertElementInst>(InsertInst)) { | |||
299 | const auto *VT = dyn_cast<FixedVectorType>(IE->getType()); | |||
300 | if (!VT) | |||
301 | return std::nullopt; | |||
302 | const auto *CI = dyn_cast<ConstantInt>(IE->getOperand(2)); | |||
303 | if (!CI) | |||
304 | return std::nullopt; | |||
305 | if (CI->getValue().uge(VT->getNumElements())) | |||
306 | return std::nullopt; | |||
307 | Index *= VT->getNumElements(); | |||
308 | Index += CI->getZExtValue(); | |||
309 | return Index; | |||
310 | } | |||
311 | ||||
312 | const auto *IV = cast<InsertValueInst>(InsertInst); | |||
313 | Type *CurrentType = IV->getType(); | |||
314 | for (unsigned I : IV->indices()) { | |||
315 | if (const auto *ST = dyn_cast<StructType>(CurrentType)) { | |||
316 | Index *= ST->getNumElements(); | |||
317 | CurrentType = ST->getElementType(I); | |||
318 | } else if (const auto *AT = dyn_cast<ArrayType>(CurrentType)) { | |||
319 | Index *= AT->getNumElements(); | |||
320 | CurrentType = AT->getElementType(); | |||
321 | } else { | |||
322 | return std::nullopt; | |||
323 | } | |||
324 | Index += I; | |||
325 | } | |||
326 | return Index; | |||
327 | } | |||
328 | ||||
329 | namespace { | |||
330 | /// Specifies the way the mask should be analyzed for undefs/poisonous elements | |||
331 | /// in the shuffle mask. | |||
332 | enum class UseMask { | |||
333 | FirstArg, ///< The mask is expected to be for permutation of 1-2 vectors, | |||
334 | ///< check for the mask elements for the first argument (mask | |||
335 | ///< indices are in range [0:VF)). | |||
336 | SecondArg, ///< The mask is expected to be for permutation of 2 vectors, check | |||
337 | ///< for the mask elements for the second argument (mask indices | |||
338 | ///< are in range [VF:2*VF)) | |||
339 | UndefsAsMask ///< Consider undef mask elements (-1) as placeholders for | |||
340 | ///< future shuffle elements and mark them as ones as being used | |||
341 | ///< in future. Non-undef elements are considered as unused since | |||
342 | ///< they're already marked as used in the mask. | |||
343 | }; | |||
344 | } // namespace | |||
345 | ||||
346 | /// Prepares a use bitset for the given mask either for the first argument or | |||
347 | /// for the second. | |||
348 | static SmallBitVector buildUseMask(int VF, ArrayRef<int> Mask, | |||
349 | UseMask MaskArg) { | |||
350 | SmallBitVector UseMask(VF, true); | |||
351 | for (auto [Idx, Value] : enumerate(Mask)) { | |||
352 | if (Value == PoisonMaskElem) { | |||
353 | if (MaskArg == UseMask::UndefsAsMask) | |||
354 | UseMask.reset(Idx); | |||
355 | continue; | |||
356 | } | |||
357 | if (MaskArg == UseMask::FirstArg && Value < VF) | |||
358 | UseMask.reset(Value); | |||
359 | else if (MaskArg == UseMask::SecondArg && Value >= VF) | |||
360 | UseMask.reset(Value - VF); | |||
361 | } | |||
362 | return UseMask; | |||
363 | } | |||
364 | ||||
365 | /// Checks if the given value is actually an undefined constant vector. | |||
366 | /// Also, if the \p UseMask is not empty, tries to check if the non-masked | |||
367 | /// elements actually mask the insertelement buildvector, if any. | |||
368 | template <bool IsPoisonOnly = false> | |||
369 | static SmallBitVector isUndefVector(const Value *V, | |||
370 | const SmallBitVector &UseMask = {}) { | |||
371 | SmallBitVector Res(UseMask.empty() ? 1 : UseMask.size(), true); | |||
372 | using T = std::conditional_t<IsPoisonOnly, PoisonValue, UndefValue>; | |||
373 | if (isa<T>(V)) | |||
374 | return Res; | |||
375 | auto *VecTy = dyn_cast<FixedVectorType>(V->getType()); | |||
376 | if (!VecTy) | |||
377 | return Res.reset(); | |||
378 | auto *C = dyn_cast<Constant>(V); | |||
379 | if (!C) { | |||
380 | if (!UseMask.empty()) { | |||
381 | const Value *Base = V; | |||
382 | while (auto *II = dyn_cast<InsertElementInst>(Base)) { | |||
383 | Base = II->getOperand(0); | |||
384 | if (isa<T>(II->getOperand(1))) | |||
385 | continue; | |||
386 | std::optional<unsigned> Idx = getInsertIndex(II); | |||
387 | if (!Idx) | |||
388 | continue; | |||
389 | if (*Idx < UseMask.size() && !UseMask.test(*Idx)) | |||
390 | Res.reset(*Idx); | |||
391 | } | |||
392 | // TODO: Add analysis for shuffles here too. | |||
393 | if (V == Base) { | |||
394 | Res.reset(); | |||
395 | } else { | |||
396 | SmallBitVector SubMask(UseMask.size(), false); | |||
397 | Res &= isUndefVector<IsPoisonOnly>(Base, SubMask); | |||
398 | } | |||
399 | } else { | |||
400 | Res.reset(); | |||
401 | } | |||
402 | return Res; | |||
403 | } | |||
404 | for (unsigned I = 0, E = VecTy->getNumElements(); I != E; ++I) { | |||
405 | if (Constant *Elem = C->getAggregateElement(I)) | |||
406 | if (!isa<T>(Elem) && | |||
407 | (UseMask.empty() || (I < UseMask.size() && !UseMask.test(I)))) | |||
408 | Res.reset(I); | |||
409 | } | |||
410 | return Res; | |||
411 | } | |||
412 | ||||
413 | /// Checks if the vector of instructions can be represented as a shuffle, like: | |||
414 | /// %x0 = extractelement <4 x i8> %x, i32 0 | |||
415 | /// %x3 = extractelement <4 x i8> %x, i32 3 | |||
416 | /// %y1 = extractelement <4 x i8> %y, i32 1 | |||
417 | /// %y2 = extractelement <4 x i8> %y, i32 2 | |||
418 | /// %x0x0 = mul i8 %x0, %x0 | |||
419 | /// %x3x3 = mul i8 %x3, %x3 | |||
420 | /// %y1y1 = mul i8 %y1, %y1 | |||
421 | /// %y2y2 = mul i8 %y2, %y2 | |||
422 | /// %ins1 = insertelement <4 x i8> poison, i8 %x0x0, i32 0 | |||
423 | /// %ins2 = insertelement <4 x i8> %ins1, i8 %x3x3, i32 1 | |||
424 | /// %ins3 = insertelement <4 x i8> %ins2, i8 %y1y1, i32 2 | |||
425 | /// %ins4 = insertelement <4 x i8> %ins3, i8 %y2y2, i32 3 | |||
426 | /// ret <4 x i8> %ins4 | |||
427 | /// can be transformed into: | |||
428 | /// %1 = shufflevector <4 x i8> %x, <4 x i8> %y, <4 x i32> <i32 0, i32 3, i32 5, | |||
429 | /// i32 6> | |||
430 | /// %2 = mul <4 x i8> %1, %1 | |||
431 | /// ret <4 x i8> %2 | |||
432 | /// We convert this initially to something like: | |||
433 | /// %x0 = extractelement <4 x i8> %x, i32 0 | |||
434 | /// %x3 = extractelement <4 x i8> %x, i32 3 | |||
435 | /// %y1 = extractelement <4 x i8> %y, i32 1 | |||
436 | /// %y2 = extractelement <4 x i8> %y, i32 2 | |||
437 | /// %1 = insertelement <4 x i8> poison, i8 %x0, i32 0 | |||
438 | /// %2 = insertelement <4 x i8> %1, i8 %x3, i32 1 | |||
439 | /// %3 = insertelement <4 x i8> %2, i8 %y1, i32 2 | |||
440 | /// %4 = insertelement <4 x i8> %3, i8 %y2, i32 3 | |||
441 | /// %5 = mul <4 x i8> %4, %4 | |||
442 | /// %6 = extractelement <4 x i8> %5, i32 0 | |||
443 | /// %ins1 = insertelement <4 x i8> poison, i8 %6, i32 0 | |||
444 | /// %7 = extractelement <4 x i8> %5, i32 1 | |||
445 | /// %ins2 = insertelement <4 x i8> %ins1, i8 %7, i32 1 | |||
446 | /// %8 = extractelement <4 x i8> %5, i32 2 | |||
447 | /// %ins3 = insertelement <4 x i8> %ins2, i8 %8, i32 2 | |||
448 | /// %9 = extractelement <4 x i8> %5, i32 3 | |||
449 | /// %ins4 = insertelement <4 x i8> %ins3, i8 %9, i32 3 | |||
450 | /// ret <4 x i8> %ins4 | |||
451 | /// InstCombiner transforms this into a shuffle and vector mul | |||
452 | /// Mask will return the Shuffle Mask equivalent to the extracted elements. | |||
453 | /// TODO: Can we split off and reuse the shuffle mask detection from | |||
454 | /// ShuffleVectorInst/getShuffleCost? | |||
455 | static std::optional<TargetTransformInfo::ShuffleKind> | |||
456 | isFixedVectorShuffle(ArrayRef<Value *> VL, SmallVectorImpl<int> &Mask) { | |||
457 | const auto *It = | |||
458 | find_if(VL, [](Value *V) { return isa<ExtractElementInst>(V); }); | |||
459 | if (It == VL.end()) | |||
460 | return std::nullopt; | |||
461 | auto *EI0 = cast<ExtractElementInst>(*It); | |||
462 | if (isa<ScalableVectorType>(EI0->getVectorOperandType())) | |||
463 | return std::nullopt; | |||
464 | unsigned Size = | |||
465 | cast<FixedVectorType>(EI0->getVectorOperandType())->getNumElements(); | |||
466 | Value *Vec1 = nullptr; | |||
467 | Value *Vec2 = nullptr; | |||
468 | enum ShuffleMode { Unknown, Select, Permute }; | |||
469 | ShuffleMode CommonShuffleMode = Unknown; | |||
470 | Mask.assign(VL.size(), PoisonMaskElem); | |||
471 | for (unsigned I = 0, E = VL.size(); I < E; ++I) { | |||
472 | // Undef can be represented as an undef element in a vector. | |||
473 | if (isa<UndefValue>(VL[I])) | |||
474 | continue; | |||
475 | auto *EI = cast<ExtractElementInst>(VL[I]); | |||
476 | if (isa<ScalableVectorType>(EI->getVectorOperandType())) | |||
477 | return std::nullopt; | |||
478 | auto *Vec = EI->getVectorOperand(); | |||
479 | // We can extractelement from undef or poison vector. | |||
480 | if (isUndefVector(Vec).all()) | |||
481 | continue; | |||
482 | // All vector operands must have the same number of vector elements. | |||
483 | if (cast<FixedVectorType>(Vec->getType())->getNumElements() != Size) | |||
484 | return std::nullopt; | |||
485 | if (isa<UndefValue>(EI->getIndexOperand())) | |||
486 | continue; | |||
487 | auto *Idx = dyn_cast<ConstantInt>(EI->getIndexOperand()); | |||
488 | if (!Idx) | |||
489 | return std::nullopt; | |||
490 | // Undefined behavior if Idx is negative or >= Size. | |||
491 | if (Idx->getValue().uge(Size)) | |||
492 | continue; | |||
493 | unsigned IntIdx = Idx->getValue().getZExtValue(); | |||
494 | Mask[I] = IntIdx; | |||
495 | // For correct shuffling we have to have at most 2 different vector operands | |||
496 | // in all extractelement instructions. | |||
497 | if (!Vec1 || Vec1 == Vec) { | |||
498 | Vec1 = Vec; | |||
499 | } else if (!Vec2 || Vec2 == Vec) { | |||
500 | Vec2 = Vec; | |||
501 | Mask[I] += Size; | |||
502 | } else { | |||
503 | return std::nullopt; | |||
504 | } | |||
505 | if (CommonShuffleMode == Permute) | |||
506 | continue; | |||
507 | // If the extract index is not the same as the operation number, it is a | |||
508 | // permutation. | |||
509 | if (IntIdx != I) { | |||
510 | CommonShuffleMode = Permute; | |||
511 | continue; | |||
512 | } | |||
513 | CommonShuffleMode = Select; | |||
514 | } | |||
515 | // If we're not crossing lanes in different vectors, consider it as blending. | |||
516 | if (CommonShuffleMode == Select && Vec2) | |||
517 | return TargetTransformInfo::SK_Select; | |||
518 | // If Vec2 was never used, we have a permutation of a single vector, otherwise | |||
519 | // we have permutation of 2 vectors. | |||
520 | return Vec2 ? TargetTransformInfo::SK_PermuteTwoSrc | |||
521 | : TargetTransformInfo::SK_PermuteSingleSrc; | |||
522 | } | |||
523 | ||||
524 | /// \returns True if Extract{Value,Element} instruction extracts element Idx. | |||
525 | static std::optional<unsigned> getExtractIndex(Instruction *E) { | |||
526 | unsigned Opcode = E->getOpcode(); | |||
527 | 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", 529, __extension__ __PRETTY_FUNCTION__)) | |||
528 | 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", 529, __extension__ __PRETTY_FUNCTION__)) | |||
529 | "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", 529, __extension__ __PRETTY_FUNCTION__)); | |||
530 | if (Opcode == Instruction::ExtractElement) { | |||
531 | auto *CI = dyn_cast<ConstantInt>(E->getOperand(1)); | |||
532 | if (!CI) | |||
533 | return std::nullopt; | |||
534 | return CI->getZExtValue(); | |||
535 | } | |||
536 | auto *EI = cast<ExtractValueInst>(E); | |||
537 | if (EI->getNumIndices() != 1) | |||
538 | return std::nullopt; | |||
539 | return *EI->idx_begin(); | |||
540 | } | |||
541 | ||||
542 | /// Tries to find extractelement instructions with constant indices from fixed | |||
543 | /// vector type and gather such instructions into a bunch, which highly likely | |||
544 | /// might be detected as a shuffle of 1 or 2 input vectors. If this attempt was | |||
545 | /// successful, the matched scalars are replaced by poison values in \p VL for | |||
546 | /// future analysis. | |||
547 | static std::optional<TTI::ShuffleKind> | |||
548 | tryToGatherExtractElements(SmallVectorImpl<Value *> &VL, | |||
549 | SmallVectorImpl<int> &Mask) { | |||
550 | // Scan list of gathered scalars for extractelements that can be represented | |||
551 | // as shuffles. | |||
552 | MapVector<Value *, SmallVector<int>> VectorOpToIdx; | |||
553 | SmallVector<int> UndefVectorExtracts; | |||
554 | for (int I = 0, E = VL.size(); I < E; ++I) { | |||
555 | auto *EI = dyn_cast<ExtractElementInst>(VL[I]); | |||
556 | if (!EI) { | |||
557 | if (isa<UndefValue>(VL[I])) | |||
558 | UndefVectorExtracts.push_back(I); | |||
559 | continue; | |||
560 | } | |||
561 | auto *VecTy = dyn_cast<FixedVectorType>(EI->getVectorOperandType()); | |||
562 | if (!VecTy || !isa<ConstantInt, UndefValue>(EI->getIndexOperand())) | |||
563 | continue; | |||
564 | std::optional<unsigned> Idx = getExtractIndex(EI); | |||
565 | // Undefined index. | |||
566 | if (!Idx) { | |||
567 | UndefVectorExtracts.push_back(I); | |||
568 | continue; | |||
569 | } | |||
570 | SmallBitVector ExtractMask(VecTy->getNumElements(), true); | |||
571 | ExtractMask.reset(*Idx); | |||
572 | if (isUndefVector(EI->getVectorOperand(), ExtractMask).all()) { | |||
573 | UndefVectorExtracts.push_back(I); | |||
574 | continue; | |||
575 | } | |||
576 | VectorOpToIdx[EI->getVectorOperand()].push_back(I); | |||
577 | } | |||
578 | // Sort the vector operands by the maximum number of uses in extractelements. | |||
579 | MapVector<unsigned, SmallVector<Value *>> VFToVector; | |||
580 | for (const auto &Data : VectorOpToIdx) | |||
581 | VFToVector[cast<FixedVectorType>(Data.first->getType())->getNumElements()] | |||
582 | .push_back(Data.first); | |||
583 | for (auto &Data : VFToVector) { | |||
584 | stable_sort(Data.second, [&VectorOpToIdx](Value *V1, Value *V2) { | |||
585 | return VectorOpToIdx.find(V1)->second.size() > | |||
586 | VectorOpToIdx.find(V2)->second.size(); | |||
587 | }); | |||
588 | } | |||
589 | // Find the best pair of the vectors with the same number of elements or a | |||
590 | // single vector. | |||
591 | const int UndefSz = UndefVectorExtracts.size(); | |||
592 | unsigned SingleMax = 0; | |||
593 | Value *SingleVec = nullptr; | |||
594 | unsigned PairMax = 0; | |||
595 | std::pair<Value *, Value *> PairVec(nullptr, nullptr); | |||
596 | for (auto &Data : VFToVector) { | |||
597 | Value *V1 = Data.second.front(); | |||
598 | if (SingleMax < VectorOpToIdx[V1].size() + UndefSz) { | |||
599 | SingleMax = VectorOpToIdx[V1].size() + UndefSz; | |||
600 | SingleVec = V1; | |||
601 | } | |||
602 | Value *V2 = nullptr; | |||
603 | if (Data.second.size() > 1) | |||
604 | V2 = *std::next(Data.second.begin()); | |||
605 | if (V2 && PairMax < VectorOpToIdx[V1].size() + VectorOpToIdx[V2].size() + | |||
606 | UndefSz) { | |||
607 | PairMax = VectorOpToIdx[V1].size() + VectorOpToIdx[V2].size() + UndefSz; | |||
608 | PairVec = std::make_pair(V1, V2); | |||
609 | } | |||
610 | } | |||
611 | if (SingleMax == 0 && PairMax == 0 && UndefSz == 0) | |||
612 | return std::nullopt; | |||
613 | // Check if better to perform a shuffle of 2 vectors or just of a single | |||
614 | // vector. | |||
615 | SmallVector<Value *> SavedVL(VL.begin(), VL.end()); | |||
616 | SmallVector<Value *> GatheredExtracts( | |||
617 | VL.size(), PoisonValue::get(VL.front()->getType())); | |||
618 | if (SingleMax >= PairMax && SingleMax) { | |||
619 | for (int Idx : VectorOpToIdx[SingleVec]) | |||
620 | std::swap(GatheredExtracts[Idx], VL[Idx]); | |||
621 | } else { | |||
622 | for (Value *V : {PairVec.first, PairVec.second}) | |||
623 | for (int Idx : VectorOpToIdx[V]) | |||
624 | std::swap(GatheredExtracts[Idx], VL[Idx]); | |||
625 | } | |||
626 | // Add extracts from undefs too. | |||
627 | for (int Idx : UndefVectorExtracts) | |||
628 | std::swap(GatheredExtracts[Idx], VL[Idx]); | |||
629 | // Check that gather of extractelements can be represented as just a | |||
630 | // shuffle of a single/two vectors the scalars are extracted from. | |||
631 | std::optional<TTI::ShuffleKind> Res = | |||
632 | isFixedVectorShuffle(GatheredExtracts, Mask); | |||
633 | if (!Res) { | |||
634 | // TODO: try to check other subsets if possible. | |||
635 | // Restore the original VL if attempt was not successful. | |||
636 | VL.swap(SavedVL); | |||
637 | return std::nullopt; | |||
638 | } | |||
639 | // Restore unused scalars from mask, if some of the extractelements were not | |||
640 | // selected for shuffle. | |||
641 | for (int I = 0, E = GatheredExtracts.size(); I < E; ++I) { | |||
642 | auto *EI = dyn_cast<ExtractElementInst>(VL[I]); | |||
643 | if (!EI || !isa<FixedVectorType>(EI->getVectorOperandType()) || | |||
644 | !isa<ConstantInt, UndefValue>(EI->getIndexOperand()) || | |||
645 | is_contained(UndefVectorExtracts, I)) | |||
646 | continue; | |||
647 | if (Mask[I] == PoisonMaskElem && !isa<PoisonValue>(GatheredExtracts[I])) | |||
648 | std::swap(VL[I], GatheredExtracts[I]); | |||
649 | } | |||
650 | return Res; | |||
651 | } | |||
652 | ||||
653 | namespace { | |||
654 | ||||
655 | /// Main data required for vectorization of instructions. | |||
656 | struct InstructionsState { | |||
657 | /// The very first instruction in the list with the main opcode. | |||
658 | Value *OpValue = nullptr; | |||
659 | ||||
660 | /// The main/alternate instruction. | |||
661 | Instruction *MainOp = nullptr; | |||
662 | Instruction *AltOp = nullptr; | |||
663 | ||||
664 | /// The main/alternate opcodes for the list of instructions. | |||
665 | unsigned getOpcode() const { | |||
666 | return MainOp ? MainOp->getOpcode() : 0; | |||
667 | } | |||
668 | ||||
669 | unsigned getAltOpcode() const { | |||
670 | return AltOp ? AltOp->getOpcode() : 0; | |||
671 | } | |||
672 | ||||
673 | /// Some of the instructions in the list have alternate opcodes. | |||
674 | bool isAltShuffle() const { return AltOp != MainOp; } | |||
675 | ||||
676 | bool isOpcodeOrAlt(Instruction *I) const { | |||
677 | unsigned CheckedOpcode = I->getOpcode(); | |||
678 | return getOpcode() == CheckedOpcode || getAltOpcode() == CheckedOpcode; | |||
679 | } | |||
680 | ||||
681 | InstructionsState() = delete; | |||
682 | InstructionsState(Value *OpValue, Instruction *MainOp, Instruction *AltOp) | |||
683 | : OpValue(OpValue), MainOp(MainOp), AltOp(AltOp) {} | |||
684 | }; | |||
685 | ||||
686 | } // end anonymous namespace | |||
687 | ||||
688 | /// Chooses the correct key for scheduling data. If \p Op has the same (or | |||
689 | /// alternate) opcode as \p OpValue, the key is \p Op. Otherwise the key is \p | |||
690 | /// OpValue. | |||
691 | static Value *isOneOf(const InstructionsState &S, Value *Op) { | |||
692 | auto *I = dyn_cast<Instruction>(Op); | |||
693 | if (I && S.isOpcodeOrAlt(I)) | |||
694 | return Op; | |||
695 | return S.OpValue; | |||
696 | } | |||
697 | ||||
698 | /// \returns true if \p Opcode is allowed as part of of the main/alternate | |||
699 | /// instruction for SLP vectorization. | |||
700 | /// | |||
701 | /// Example of unsupported opcode is SDIV that can potentially cause UB if the | |||
702 | /// "shuffled out" lane would result in division by zero. | |||
703 | static bool isValidForAlternation(unsigned Opcode) { | |||
704 | if (Instruction::isIntDivRem(Opcode)) | |||
705 | return false; | |||
706 | ||||
707 | return true; | |||
708 | } | |||
709 | ||||
710 | static InstructionsState getSameOpcode(ArrayRef<Value *> VL, | |||
711 | const TargetLibraryInfo &TLI, | |||
712 | unsigned BaseIndex = 0); | |||
713 | ||||
714 | /// Checks if the provided operands of 2 cmp instructions are compatible, i.e. | |||
715 | /// compatible instructions or constants, or just some other regular values. | |||
716 | static bool areCompatibleCmpOps(Value *BaseOp0, Value *BaseOp1, Value *Op0, | |||
717 | Value *Op1, const TargetLibraryInfo &TLI) { | |||
718 | return (isConstant(BaseOp0) && isConstant(Op0)) || | |||
719 | (isConstant(BaseOp1) && isConstant(Op1)) || | |||
720 | (!isa<Instruction>(BaseOp0) && !isa<Instruction>(Op0) && | |||
721 | !isa<Instruction>(BaseOp1) && !isa<Instruction>(Op1)) || | |||
722 | BaseOp0 == Op0 || BaseOp1 == Op1 || | |||
723 | getSameOpcode({BaseOp0, Op0}, TLI).getOpcode() || | |||
724 | getSameOpcode({BaseOp1, Op1}, TLI).getOpcode(); | |||
725 | } | |||
726 | ||||
727 | /// \returns true if a compare instruction \p CI has similar "look" and | |||
728 | /// same predicate as \p BaseCI, "as is" or with its operands and predicate | |||
729 | /// swapped, false otherwise. | |||
730 | static bool isCmpSameOrSwapped(const CmpInst *BaseCI, const CmpInst *CI, | |||
731 | const TargetLibraryInfo &TLI) { | |||
732 | assert(BaseCI->getOperand(0)->getType() == CI->getOperand(0)->getType() &&(static_cast <bool> (BaseCI->getOperand(0)->getType () == CI->getOperand(0)->getType() && "Assessing comparisons of different types?" ) ? void (0) : __assert_fail ("BaseCI->getOperand(0)->getType() == CI->getOperand(0)->getType() && \"Assessing comparisons of different types?\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 733, __extension__ __PRETTY_FUNCTION__)) | |||
733 | "Assessing comparisons of different types?")(static_cast <bool> (BaseCI->getOperand(0)->getType () == CI->getOperand(0)->getType() && "Assessing comparisons of different types?" ) ? void (0) : __assert_fail ("BaseCI->getOperand(0)->getType() == CI->getOperand(0)->getType() && \"Assessing comparisons of different types?\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 733, __extension__ __PRETTY_FUNCTION__)); | |||
734 | CmpInst::Predicate BasePred = BaseCI->getPredicate(); | |||
735 | CmpInst::Predicate Pred = CI->getPredicate(); | |||
736 | CmpInst::Predicate SwappedPred = CmpInst::getSwappedPredicate(Pred); | |||
737 | ||||
738 | Value *BaseOp0 = BaseCI->getOperand(0); | |||
739 | Value *BaseOp1 = BaseCI->getOperand(1); | |||
740 | Value *Op0 = CI->getOperand(0); | |||
741 | Value *Op1 = CI->getOperand(1); | |||
742 | ||||
743 | return (BasePred == Pred && | |||
744 | areCompatibleCmpOps(BaseOp0, BaseOp1, Op0, Op1, TLI)) || | |||
745 | (BasePred == SwappedPred && | |||
746 | areCompatibleCmpOps(BaseOp0, BaseOp1, Op1, Op0, TLI)); | |||
747 | } | |||
748 | ||||
749 | /// \returns analysis of the Instructions in \p VL described in | |||
750 | /// InstructionsState, the Opcode that we suppose the whole list | |||
751 | /// could be vectorized even if its structure is diverse. | |||
752 | static InstructionsState getSameOpcode(ArrayRef<Value *> VL, | |||
753 | const TargetLibraryInfo &TLI, | |||
754 | unsigned BaseIndex) { | |||
755 | // Make sure these are all Instructions. | |||
756 | if (llvm::any_of(VL, [](Value *V) { return !isa<Instruction>(V); })) | |||
757 | return InstructionsState(VL[BaseIndex], nullptr, nullptr); | |||
758 | ||||
759 | bool IsCastOp = isa<CastInst>(VL[BaseIndex]); | |||
760 | bool IsBinOp = isa<BinaryOperator>(VL[BaseIndex]); | |||
761 | bool IsCmpOp = isa<CmpInst>(VL[BaseIndex]); | |||
762 | CmpInst::Predicate BasePred = | |||
763 | IsCmpOp ? cast<CmpInst>(VL[BaseIndex])->getPredicate() | |||
764 | : CmpInst::BAD_ICMP_PREDICATE; | |||
765 | unsigned Opcode = cast<Instruction>(VL[BaseIndex])->getOpcode(); | |||
766 | unsigned AltOpcode = Opcode; | |||
767 | unsigned AltIndex = BaseIndex; | |||
768 | ||||
769 | // Check for one alternate opcode from another BinaryOperator. | |||
770 | // TODO - generalize to support all operators (types, calls etc.). | |||
771 | auto *IBase = cast<Instruction>(VL[BaseIndex]); | |||
772 | Intrinsic::ID BaseID = 0; | |||
773 | SmallVector<VFInfo> BaseMappings; | |||
774 | if (auto *CallBase = dyn_cast<CallInst>(IBase)) { | |||
775 | BaseID = getVectorIntrinsicIDForCall(CallBase, &TLI); | |||
776 | BaseMappings = VFDatabase(*CallBase).getMappings(*CallBase); | |||
777 | if (!isTriviallyVectorizable(BaseID) && BaseMappings.empty()) | |||
778 | return InstructionsState(VL[BaseIndex], nullptr, nullptr); | |||
779 | } | |||
780 | for (int Cnt = 0, E = VL.size(); Cnt < E; Cnt++) { | |||
781 | auto *I = cast<Instruction>(VL[Cnt]); | |||
782 | unsigned InstOpcode = I->getOpcode(); | |||
783 | if (IsBinOp && isa<BinaryOperator>(I)) { | |||
784 | if (InstOpcode == Opcode || InstOpcode == AltOpcode) | |||
785 | continue; | |||
786 | if (Opcode == AltOpcode && isValidForAlternation(InstOpcode) && | |||
787 | isValidForAlternation(Opcode)) { | |||
788 | AltOpcode = InstOpcode; | |||
789 | AltIndex = Cnt; | |||
790 | continue; | |||
791 | } | |||
792 | } else if (IsCastOp && isa<CastInst>(I)) { | |||
793 | Value *Op0 = IBase->getOperand(0); | |||
794 | Type *Ty0 = Op0->getType(); | |||
795 | Value *Op1 = I->getOperand(0); | |||
796 | Type *Ty1 = Op1->getType(); | |||
797 | if (Ty0 == Ty1) { | |||
798 | if (InstOpcode == Opcode || InstOpcode == AltOpcode) | |||
799 | continue; | |||
800 | if (Opcode == AltOpcode) { | |||
801 | 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", 803, __extension__ __PRETTY_FUNCTION__)) | |||
802 | 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", 803, __extension__ __PRETTY_FUNCTION__)) | |||
803 | "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", 803, __extension__ __PRETTY_FUNCTION__)); | |||
804 | AltOpcode = InstOpcode; | |||
805 | AltIndex = Cnt; | |||
806 | continue; | |||
807 | } | |||
808 | } | |||
809 | } else if (auto *Inst = dyn_cast<CmpInst>(VL[Cnt]); Inst && IsCmpOp) { | |||
810 | auto *BaseInst = cast<CmpInst>(VL[BaseIndex]); | |||
811 | Type *Ty0 = BaseInst->getOperand(0)->getType(); | |||
812 | Type *Ty1 = Inst->getOperand(0)->getType(); | |||
813 | if (Ty0 == Ty1) { | |||
814 | assert(InstOpcode == Opcode && "Expected same CmpInst opcode.")(static_cast <bool> (InstOpcode == Opcode && "Expected same CmpInst opcode." ) ? void (0) : __assert_fail ("InstOpcode == Opcode && \"Expected same CmpInst opcode.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 814, __extension__ __PRETTY_FUNCTION__)); | |||
815 | // Check for compatible operands. If the corresponding operands are not | |||
816 | // compatible - need to perform alternate vectorization. | |||
817 | CmpInst::Predicate CurrentPred = Inst->getPredicate(); | |||
818 | CmpInst::Predicate SwappedCurrentPred = | |||
819 | CmpInst::getSwappedPredicate(CurrentPred); | |||
820 | ||||
821 | if (E == 2 && | |||
822 | (BasePred == CurrentPred || BasePred == SwappedCurrentPred)) | |||
823 | continue; | |||
824 | ||||
825 | if (isCmpSameOrSwapped(BaseInst, Inst, TLI)) | |||
826 | continue; | |||
827 | auto *AltInst = cast<CmpInst>(VL[AltIndex]); | |||
828 | if (AltIndex != BaseIndex) { | |||
829 | if (isCmpSameOrSwapped(AltInst, Inst, TLI)) | |||
830 | continue; | |||
831 | } else if (BasePred != CurrentPred) { | |||
832 | assert((static_cast <bool> (isValidForAlternation(InstOpcode) && "CmpInst isn't safe for alternation, logic needs to be updated!" ) ? void (0) : __assert_fail ("isValidForAlternation(InstOpcode) && \"CmpInst isn't safe for alternation, logic needs to be updated!\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 834, __extension__ __PRETTY_FUNCTION__)) | |||
833 | isValidForAlternation(InstOpcode) &&(static_cast <bool> (isValidForAlternation(InstOpcode) && "CmpInst isn't safe for alternation, logic needs to be updated!" ) ? void (0) : __assert_fail ("isValidForAlternation(InstOpcode) && \"CmpInst isn't safe for alternation, logic needs to be updated!\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 834, __extension__ __PRETTY_FUNCTION__)) | |||
834 | "CmpInst isn't safe for alternation, logic needs to be updated!")(static_cast <bool> (isValidForAlternation(InstOpcode) && "CmpInst isn't safe for alternation, logic needs to be updated!" ) ? void (0) : __assert_fail ("isValidForAlternation(InstOpcode) && \"CmpInst isn't safe for alternation, logic needs to be updated!\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 834, __extension__ __PRETTY_FUNCTION__)); | |||
835 | AltIndex = Cnt; | |||
836 | continue; | |||
837 | } | |||
838 | CmpInst::Predicate AltPred = AltInst->getPredicate(); | |||
839 | if (BasePred == CurrentPred || BasePred == SwappedCurrentPred || | |||
840 | AltPred == CurrentPred || AltPred == SwappedCurrentPred) | |||
841 | continue; | |||
842 | } | |||
843 | } else if (InstOpcode == Opcode || InstOpcode == AltOpcode) { | |||
844 | if (auto *Gep = dyn_cast<GetElementPtrInst>(I)) { | |||
845 | if (Gep->getNumOperands() != 2 || | |||
846 | Gep->getOperand(0)->getType() != IBase->getOperand(0)->getType()) | |||
847 | return InstructionsState(VL[BaseIndex], nullptr, nullptr); | |||
848 | } else if (auto *EI = dyn_cast<ExtractElementInst>(I)) { | |||
849 | if (!isVectorLikeInstWithConstOps(EI)) | |||
850 | return InstructionsState(VL[BaseIndex], nullptr, nullptr); | |||
851 | } else if (auto *LI = dyn_cast<LoadInst>(I)) { | |||
852 | auto *BaseLI = cast<LoadInst>(IBase); | |||
853 | if (!LI->isSimple() || !BaseLI->isSimple()) | |||
854 | return InstructionsState(VL[BaseIndex], nullptr, nullptr); | |||
855 | } else if (auto *Call = dyn_cast<CallInst>(I)) { | |||
856 | auto *CallBase = cast<CallInst>(IBase); | |||
857 | if (Call->getCalledFunction() != CallBase->getCalledFunction()) | |||
858 | return InstructionsState(VL[BaseIndex], nullptr, nullptr); | |||
859 | if (Call->hasOperandBundles() && | |||
860 | !std::equal(Call->op_begin() + Call->getBundleOperandsStartIndex(), | |||
861 | Call->op_begin() + Call->getBundleOperandsEndIndex(), | |||
862 | CallBase->op_begin() + | |||
863 | CallBase->getBundleOperandsStartIndex())) | |||
864 | return InstructionsState(VL[BaseIndex], nullptr, nullptr); | |||
865 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(Call, &TLI); | |||
866 | if (ID != BaseID) | |||
867 | return InstructionsState(VL[BaseIndex], nullptr, nullptr); | |||
868 | if (!ID) { | |||
869 | SmallVector<VFInfo> Mappings = VFDatabase(*Call).getMappings(*Call); | |||
870 | if (Mappings.size() != BaseMappings.size() || | |||
871 | Mappings.front().ISA != BaseMappings.front().ISA || | |||
872 | Mappings.front().ScalarName != BaseMappings.front().ScalarName || | |||
873 | Mappings.front().VectorName != BaseMappings.front().VectorName || | |||
874 | Mappings.front().Shape.VF != BaseMappings.front().Shape.VF || | |||
875 | Mappings.front().Shape.Parameters != | |||
876 | BaseMappings.front().Shape.Parameters) | |||
877 | return InstructionsState(VL[BaseIndex], nullptr, nullptr); | |||
878 | } | |||
879 | } | |||
880 | continue; | |||
881 | } | |||
882 | return InstructionsState(VL[BaseIndex], nullptr, nullptr); | |||
883 | } | |||
884 | ||||
885 | return InstructionsState(VL[BaseIndex], cast<Instruction>(VL[BaseIndex]), | |||
886 | cast<Instruction>(VL[AltIndex])); | |||
887 | } | |||
888 | ||||
889 | /// \returns true if all of the values in \p VL have the same type or false | |||
890 | /// otherwise. | |||
891 | static bool allSameType(ArrayRef<Value *> VL) { | |||
892 | Type *Ty = VL[0]->getType(); | |||
893 | for (int i = 1, e = VL.size(); i < e; i++) | |||
894 | if (VL[i]->getType() != Ty) | |||
895 | return false; | |||
896 | ||||
897 | return true; | |||
898 | } | |||
899 | ||||
900 | /// \returns True if in-tree use also needs extract. This refers to | |||
901 | /// possible scalar operand in vectorized instruction. | |||
902 | static bool InTreeUserNeedToExtract(Value *Scalar, Instruction *UserInst, | |||
903 | TargetLibraryInfo *TLI) { | |||
904 | unsigned Opcode = UserInst->getOpcode(); | |||
905 | switch (Opcode) { | |||
906 | case Instruction::Load: { | |||
907 | LoadInst *LI = cast<LoadInst>(UserInst); | |||
908 | return (LI->getPointerOperand() == Scalar); | |||
909 | } | |||
910 | case Instruction::Store: { | |||
911 | StoreInst *SI = cast<StoreInst>(UserInst); | |||
912 | return (SI->getPointerOperand() == Scalar); | |||
913 | } | |||
914 | case Instruction::Call: { | |||
915 | CallInst *CI = cast<CallInst>(UserInst); | |||
916 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); | |||
917 | for (unsigned i = 0, e = CI->arg_size(); i != e; ++i) { | |||
918 | if (isVectorIntrinsicWithScalarOpAtArg(ID, i)) | |||
919 | return (CI->getArgOperand(i) == Scalar); | |||
920 | } | |||
921 | [[fallthrough]]; | |||
922 | } | |||
923 | default: | |||
924 | return false; | |||
925 | } | |||
926 | } | |||
927 | ||||
928 | /// \returns the AA location that is being access by the instruction. | |||
929 | static MemoryLocation getLocation(Instruction *I) { | |||
930 | if (StoreInst *SI = dyn_cast<StoreInst>(I)) | |||
931 | return MemoryLocation::get(SI); | |||
932 | if (LoadInst *LI = dyn_cast<LoadInst>(I)) | |||
933 | return MemoryLocation::get(LI); | |||
934 | return MemoryLocation(); | |||
935 | } | |||
936 | ||||
937 | /// \returns True if the instruction is not a volatile or atomic load/store. | |||
938 | static bool isSimple(Instruction *I) { | |||
939 | if (LoadInst *LI = dyn_cast<LoadInst>(I)) | |||
940 | return LI->isSimple(); | |||
941 | if (StoreInst *SI = dyn_cast<StoreInst>(I)) | |||
942 | return SI->isSimple(); | |||
943 | if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I)) | |||
944 | return !MI->isVolatile(); | |||
945 | return true; | |||
946 | } | |||
947 | ||||
948 | /// Shuffles \p Mask in accordance with the given \p SubMask. | |||
949 | /// \param ExtendingManyInputs Supports reshuffling of the mask with not only | |||
950 | /// one but two input vectors. | |||
951 | static void addMask(SmallVectorImpl<int> &Mask, ArrayRef<int> SubMask, | |||
952 | bool ExtendingManyInputs = false) { | |||
953 | if (SubMask.empty()) | |||
954 | return; | |||
955 | assert((!ExtendingManyInputs || SubMask.size() > Mask.size()) &&(static_cast <bool> ((!ExtendingManyInputs || SubMask.size () > Mask.size()) && "SubMask with many inputs support must be larger than the mask." ) ? void (0) : __assert_fail ("(!ExtendingManyInputs || SubMask.size() > Mask.size()) && \"SubMask with many inputs support must be larger than the mask.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 956, __extension__ __PRETTY_FUNCTION__)) | |||
956 | "SubMask with many inputs support must be larger than the mask.")(static_cast <bool> ((!ExtendingManyInputs || SubMask.size () > Mask.size()) && "SubMask with many inputs support must be larger than the mask." ) ? void (0) : __assert_fail ("(!ExtendingManyInputs || SubMask.size() > Mask.size()) && \"SubMask with many inputs support must be larger than the mask.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 956, __extension__ __PRETTY_FUNCTION__)); | |||
957 | if (Mask.empty()) { | |||
958 | Mask.append(SubMask.begin(), SubMask.end()); | |||
959 | return; | |||
960 | } | |||
961 | SmallVector<int> NewMask(SubMask.size(), PoisonMaskElem); | |||
962 | int TermValue = std::min(Mask.size(), SubMask.size()); | |||
963 | for (int I = 0, E = SubMask.size(); I < E; ++I) { | |||
964 | if ((!ExtendingManyInputs && | |||
965 | (SubMask[I] >= TermValue || Mask[SubMask[I]] >= TermValue)) || | |||
966 | SubMask[I] == PoisonMaskElem) | |||
967 | continue; | |||
968 | NewMask[I] = Mask[SubMask[I]]; | |||
969 | } | |||
970 | Mask.swap(NewMask); | |||
971 | } | |||
972 | ||||
973 | /// Order may have elements assigned special value (size) which is out of | |||
974 | /// bounds. Such indices only appear on places which correspond to undef values | |||
975 | /// (see canReuseExtract for details) and used in order to avoid undef values | |||
976 | /// have effect on operands ordering. | |||
977 | /// The first loop below simply finds all unused indices and then the next loop | |||
978 | /// nest assigns these indices for undef values positions. | |||
979 | /// As an example below Order has two undef positions and they have assigned | |||
980 | /// values 3 and 7 respectively: | |||
981 | /// before: 6 9 5 4 9 2 1 0 | |||
982 | /// after: 6 3 5 4 7 2 1 0 | |||
983 | static void fixupOrderingIndices(SmallVectorImpl<unsigned> &Order) { | |||
984 | const unsigned Sz = Order.size(); | |||
985 | SmallBitVector UnusedIndices(Sz, /*t=*/true); | |||
986 | SmallBitVector MaskedIndices(Sz); | |||
987 | for (unsigned I = 0; I < Sz; ++I) { | |||
988 | if (Order[I] < Sz) | |||
989 | UnusedIndices.reset(Order[I]); | |||
990 | else | |||
991 | MaskedIndices.set(I); | |||
992 | } | |||
993 | if (MaskedIndices.none()) | |||
994 | return; | |||
995 | 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", 996, __extension__ __PRETTY_FUNCTION__)) | |||
996 | "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", 996, __extension__ __PRETTY_FUNCTION__)); | |||
997 | int Idx = UnusedIndices.find_first(); | |||
998 | int MIdx = MaskedIndices.find_first(); | |||
999 | while (MIdx >= 0) { | |||
1000 | 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", 1000, __extension__ __PRETTY_FUNCTION__)); | |||
1001 | Order[MIdx] = Idx; | |||
1002 | Idx = UnusedIndices.find_next(Idx); | |||
1003 | MIdx = MaskedIndices.find_next(MIdx); | |||
1004 | } | |||
1005 | } | |||
1006 | ||||
1007 | namespace llvm { | |||
1008 | ||||
1009 | static void inversePermutation(ArrayRef<unsigned> Indices, | |||
1010 | SmallVectorImpl<int> &Mask) { | |||
1011 | Mask.clear(); | |||
1012 | const unsigned E = Indices.size(); | |||
1013 | Mask.resize(E, PoisonMaskElem); | |||
1014 | for (unsigned I = 0; I < E; ++I) | |||
1015 | Mask[Indices[I]] = I; | |||
1016 | } | |||
1017 | ||||
1018 | /// Reorders the list of scalars in accordance with the given \p Mask. | |||
1019 | static void reorderScalars(SmallVectorImpl<Value *> &Scalars, | |||
1020 | ArrayRef<int> Mask) { | |||
1021 | 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", 1021, __extension__ __PRETTY_FUNCTION__)); | |||
1022 | SmallVector<Value *> Prev(Scalars.size(), | |||
1023 | UndefValue::get(Scalars.front()->getType())); | |||
1024 | Prev.swap(Scalars); | |||
1025 | for (unsigned I = 0, E = Prev.size(); I < E; ++I) | |||
1026 | if (Mask[I] != PoisonMaskElem) | |||
1027 | Scalars[Mask[I]] = Prev[I]; | |||
1028 | } | |||
1029 | ||||
1030 | /// Checks if the provided value does not require scheduling. It does not | |||
1031 | /// require scheduling if this is not an instruction or it is an instruction | |||
1032 | /// that does not read/write memory and all operands are either not instructions | |||
1033 | /// or phi nodes or instructions from different blocks. | |||
1034 | static bool areAllOperandsNonInsts(Value *V) { | |||
1035 | auto *I = dyn_cast<Instruction>(V); | |||
1036 | if (!I) | |||
1037 | return true; | |||
1038 | return !mayHaveNonDefUseDependency(*I) && | |||
1039 | all_of(I->operands(), [I](Value *V) { | |||
1040 | auto *IO = dyn_cast<Instruction>(V); | |||
1041 | if (!IO) | |||
1042 | return true; | |||
1043 | return isa<PHINode>(IO) || IO->getParent() != I->getParent(); | |||
1044 | }); | |||
1045 | } | |||
1046 | ||||
1047 | /// Checks if the provided value does not require scheduling. It does not | |||
1048 | /// require scheduling if this is not an instruction or it is an instruction | |||
1049 | /// that does not read/write memory and all users are phi nodes or instructions | |||
1050 | /// from the different blocks. | |||
1051 | static bool isUsedOutsideBlock(Value *V) { | |||
1052 | auto *I = dyn_cast<Instruction>(V); | |||
1053 | if (!I) | |||
1054 | return true; | |||
1055 | // Limits the number of uses to save compile time. | |||
1056 | constexpr int UsesLimit = 8; | |||
1057 | return !I->mayReadOrWriteMemory() && !I->hasNUsesOrMore(UsesLimit) && | |||
1058 | all_of(I->users(), [I](User *U) { | |||
1059 | auto *IU = dyn_cast<Instruction>(U); | |||
1060 | if (!IU) | |||
1061 | return true; | |||
1062 | return IU->getParent() != I->getParent() || isa<PHINode>(IU); | |||
1063 | }); | |||
1064 | } | |||
1065 | ||||
1066 | /// Checks if the specified value does not require scheduling. It does not | |||
1067 | /// require scheduling if all operands and all users do not need to be scheduled | |||
1068 | /// in the current basic block. | |||
1069 | static bool doesNotNeedToBeScheduled(Value *V) { | |||
1070 | return areAllOperandsNonInsts(V) && isUsedOutsideBlock(V); | |||
1071 | } | |||
1072 | ||||
1073 | /// Checks if the specified array of instructions does not require scheduling. | |||
1074 | /// It is so if all either instructions have operands that do not require | |||
1075 | /// scheduling or their users do not require scheduling since they are phis or | |||
1076 | /// in other basic blocks. | |||
1077 | static bool doesNotNeedToSchedule(ArrayRef<Value *> VL) { | |||
1078 | return !VL.empty() && | |||
1079 | (all_of(VL, isUsedOutsideBlock) || all_of(VL, areAllOperandsNonInsts)); | |||
1080 | } | |||
1081 | ||||
1082 | namespace slpvectorizer { | |||
1083 | ||||
1084 | /// Bottom Up SLP Vectorizer. | |||
1085 | class BoUpSLP { | |||
1086 | struct TreeEntry; | |||
1087 | struct ScheduleData; | |||
1088 | class ShuffleCostEstimator; | |||
1089 | class ShuffleInstructionBuilder; | |||
1090 | ||||
1091 | public: | |||
1092 | using ValueList = SmallVector<Value *, 8>; | |||
1093 | using InstrList = SmallVector<Instruction *, 16>; | |||
1094 | using ValueSet = SmallPtrSet<Value *, 16>; | |||
1095 | using StoreList = SmallVector<StoreInst *, 8>; | |||
1096 | using ExtraValueToDebugLocsMap = | |||
1097 | MapVector<Value *, SmallVector<Instruction *, 2>>; | |||
1098 | using OrdersType = SmallVector<unsigned, 4>; | |||
1099 | ||||
1100 | BoUpSLP(Function *Func, ScalarEvolution *Se, TargetTransformInfo *Tti, | |||
1101 | TargetLibraryInfo *TLi, AAResults *Aa, LoopInfo *Li, | |||
1102 | DominatorTree *Dt, AssumptionCache *AC, DemandedBits *DB, | |||
1103 | const DataLayout *DL, OptimizationRemarkEmitter *ORE) | |||
1104 | : BatchAA(*Aa), F(Func), SE(Se), TTI(Tti), TLI(TLi), LI(Li), | |||
1105 | DT(Dt), AC(AC), DB(DB), DL(DL), ORE(ORE), Builder(Se->getContext()) { | |||
1106 | CodeMetrics::collectEphemeralValues(F, AC, EphValues); | |||
1107 | // Use the vector register size specified by the target unless overridden | |||
1108 | // by a command-line option. | |||
1109 | // TODO: It would be better to limit the vectorization factor based on | |||
1110 | // data type rather than just register size. For example, x86 AVX has | |||
1111 | // 256-bit registers, but it does not support integer operations | |||
1112 | // at that width (that requires AVX2). | |||
1113 | if (MaxVectorRegSizeOption.getNumOccurrences()) | |||
1114 | MaxVecRegSize = MaxVectorRegSizeOption; | |||
1115 | else | |||
1116 | MaxVecRegSize = | |||
1117 | TTI->getRegisterBitWidth(TargetTransformInfo::RGK_FixedWidthVector) | |||
1118 | .getFixedValue(); | |||
1119 | ||||
1120 | if (MinVectorRegSizeOption.getNumOccurrences()) | |||
1121 | MinVecRegSize = MinVectorRegSizeOption; | |||
1122 | else | |||
1123 | MinVecRegSize = TTI->getMinVectorRegisterBitWidth(); | |||
1124 | } | |||
1125 | ||||
1126 | /// Vectorize the tree that starts with the elements in \p VL. | |||
1127 | /// Returns the vectorized root. | |||
1128 | Value *vectorizeTree(); | |||
1129 | ||||
1130 | /// Vectorize the tree but with the list of externally used values \p | |||
1131 | /// ExternallyUsedValues. Values in this MapVector can be replaced but the | |||
1132 | /// generated extractvalue instructions. | |||
1133 | /// \param ReplacedExternals containd list of replaced external values | |||
1134 | /// {scalar, replace} after emitting extractelement for external uses. | |||
1135 | Value * | |||
1136 | vectorizeTree(const ExtraValueToDebugLocsMap &ExternallyUsedValues, | |||
1137 | SmallVectorImpl<std::pair<Value *, Value *>> &ReplacedExternals, | |||
1138 | Instruction *ReductionRoot = nullptr); | |||
1139 | ||||
1140 | /// \returns the cost incurred by unwanted spills and fills, caused by | |||
1141 | /// holding live values over call sites. | |||
1142 | InstructionCost getSpillCost() const; | |||
1143 | ||||
1144 | /// \returns the vectorization cost of the subtree that starts at \p VL. | |||
1145 | /// A negative number means that this is profitable. | |||
1146 | InstructionCost getTreeCost(ArrayRef<Value *> VectorizedVals = std::nullopt); | |||
1147 | ||||
1148 | /// Construct a vectorizable tree that starts at \p Roots, ignoring users for | |||
1149 | /// the purpose of scheduling and extraction in the \p UserIgnoreLst. | |||
1150 | void buildTree(ArrayRef<Value *> Roots, | |||
1151 | const SmallDenseSet<Value *> &UserIgnoreLst); | |||
1152 | ||||
1153 | /// Construct a vectorizable tree that starts at \p Roots. | |||
1154 | void buildTree(ArrayRef<Value *> Roots); | |||
1155 | ||||
1156 | /// Returns whether the root node has in-tree uses. | |||
1157 | bool doesRootHaveInTreeUses() const { | |||
1158 | return !VectorizableTree.empty() && | |||
1159 | !VectorizableTree.front()->UserTreeIndices.empty(); | |||
1160 | } | |||
1161 | ||||
1162 | /// Return the scalars of the root node. | |||
1163 | ArrayRef<Value *> getRootNodeScalars() const { | |||
1164 | assert(!VectorizableTree.empty() && "No graph to get the first node from")(static_cast <bool> (!VectorizableTree.empty() && "No graph to get the first node from") ? void (0) : __assert_fail ("!VectorizableTree.empty() && \"No graph to get the first node from\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 1164, __extension__ __PRETTY_FUNCTION__)); | |||
1165 | return VectorizableTree.front()->Scalars; | |||
1166 | } | |||
1167 | ||||
1168 | /// Builds external uses of the vectorized scalars, i.e. the list of | |||
1169 | /// vectorized scalars to be extracted, their lanes and their scalar users. \p | |||
1170 | /// ExternallyUsedValues contains additional list of external uses to handle | |||
1171 | /// vectorization of reductions. | |||
1172 | void | |||
1173 | buildExternalUses(const ExtraValueToDebugLocsMap &ExternallyUsedValues = {}); | |||
1174 | ||||
1175 | /// Clear the internal data structures that are created by 'buildTree'. | |||
1176 | void deleteTree() { | |||
1177 | VectorizableTree.clear(); | |||
1178 | ScalarToTreeEntry.clear(); | |||
1179 | MustGather.clear(); | |||
1180 | EntryToLastInstruction.clear(); | |||
1181 | ExternalUses.clear(); | |||
1182 | for (auto &Iter : BlocksSchedules) { | |||
1183 | BlockScheduling *BS = Iter.second.get(); | |||
1184 | BS->clear(); | |||
1185 | } | |||
1186 | MinBWs.clear(); | |||
1187 | InstrElementSize.clear(); | |||
1188 | UserIgnoreList = nullptr; | |||
1189 | PostponedGathers.clear(); | |||
1190 | ValueToGatherNodes.clear(); | |||
1191 | } | |||
1192 | ||||
1193 | unsigned getTreeSize() const { return VectorizableTree.size(); } | |||
1194 | ||||
1195 | /// Perform LICM and CSE on the newly generated gather sequences. | |||
1196 | void optimizeGatherSequence(); | |||
1197 | ||||
1198 | /// Checks if the specified gather tree entry \p TE can be represented as a | |||
1199 | /// shuffled vector entry + (possibly) permutation with other gathers. It | |||
1200 | /// implements the checks only for possibly ordered scalars (Loads, | |||
1201 | /// ExtractElement, ExtractValue), which can be part of the graph. | |||
1202 | std::optional<OrdersType> findReusedOrderedScalars(const TreeEntry &TE); | |||
1203 | ||||
1204 | /// Sort loads into increasing pointers offsets to allow greater clustering. | |||
1205 | std::optional<OrdersType> findPartiallyOrderedLoads(const TreeEntry &TE); | |||
1206 | ||||
1207 | /// Gets reordering data for the given tree entry. If the entry is vectorized | |||
1208 | /// - just return ReorderIndices, otherwise check if the scalars can be | |||
1209 | /// reordered and return the most optimal order. | |||
1210 | /// \return std::nullopt if ordering is not important, empty order, if | |||
1211 | /// identity order is important, or the actual order. | |||
1212 | /// \param TopToBottom If true, include the order of vectorized stores and | |||
1213 | /// insertelement nodes, otherwise skip them. | |||
1214 | std::optional<OrdersType> getReorderingData(const TreeEntry &TE, | |||
1215 | bool TopToBottom); | |||
1216 | ||||
1217 | /// Reorders the current graph to the most profitable order starting from the | |||
1218 | /// root node to the leaf nodes. The best order is chosen only from the nodes | |||
1219 | /// of the same size (vectorization factor). Smaller nodes are considered | |||
1220 | /// parts of subgraph with smaller VF and they are reordered independently. We | |||
1221 | /// can make it because we still need to extend smaller nodes to the wider VF | |||
1222 | /// and we can merge reordering shuffles with the widening shuffles. | |||
1223 | void reorderTopToBottom(); | |||
1224 | ||||
1225 | /// Reorders the current graph to the most profitable order starting from | |||
1226 | /// leaves to the root. It allows to rotate small subgraphs and reduce the | |||
1227 | /// number of reshuffles if the leaf nodes use the same order. In this case we | |||
1228 | /// can merge the orders and just shuffle user node instead of shuffling its | |||
1229 | /// operands. Plus, even the leaf nodes have different orders, it allows to | |||
1230 | /// sink reordering in the graph closer to the root node and merge it later | |||
1231 | /// during analysis. | |||
1232 | void reorderBottomToTop(bool IgnoreReorder = false); | |||
1233 | ||||
1234 | /// \return The vector element size in bits to use when vectorizing the | |||
1235 | /// expression tree ending at \p V. If V is a store, the size is the width of | |||
1236 | /// the stored value. Otherwise, the size is the width of the largest loaded | |||
1237 | /// value reaching V. This method is used by the vectorizer to calculate | |||
1238 | /// vectorization factors. | |||
1239 | unsigned getVectorElementSize(Value *V); | |||
1240 | ||||
1241 | /// Compute the minimum type sizes required to represent the entries in a | |||
1242 | /// vectorizable tree. | |||
1243 | void computeMinimumValueSizes(); | |||
1244 | ||||
1245 | // \returns maximum vector register size as set by TTI or overridden by cl::opt. | |||
1246 | unsigned getMaxVecRegSize() const { | |||
1247 | return MaxVecRegSize; | |||
1248 | } | |||
1249 | ||||
1250 | // \returns minimum vector register size as set by cl::opt. | |||
1251 | unsigned getMinVecRegSize() const { | |||
1252 | return MinVecRegSize; | |||
1253 | } | |||
1254 | ||||
1255 | unsigned getMinVF(unsigned Sz) const { | |||
1256 | return std::max(2U, getMinVecRegSize() / Sz); | |||
1257 | } | |||
1258 | ||||
1259 | unsigned getMaximumVF(unsigned ElemWidth, unsigned Opcode) const { | |||
1260 | unsigned MaxVF = MaxVFOption.getNumOccurrences() ? | |||
1261 | MaxVFOption : TTI->getMaximumVF(ElemWidth, Opcode); | |||
1262 | return MaxVF ? MaxVF : UINT_MAX(2147483647 *2U +1U); | |||
1263 | } | |||
1264 | ||||
1265 | /// Check if homogeneous aggregate is isomorphic to some VectorType. | |||
1266 | /// Accepts homogeneous multidimensional aggregate of scalars/vectors like | |||
1267 | /// {[4 x i16], [4 x i16]}, { <2 x float>, <2 x float> }, | |||
1268 | /// {{{i16, i16}, {i16, i16}}, {{i16, i16}, {i16, i16}}} and so on. | |||
1269 | /// | |||
1270 | /// \returns number of elements in vector if isomorphism exists, 0 otherwise. | |||
1271 | unsigned canMapToVector(Type *T, const DataLayout &DL) const; | |||
1272 | ||||
1273 | /// \returns True if the VectorizableTree is both tiny and not fully | |||
1274 | /// vectorizable. We do not vectorize such trees. | |||
1275 | bool isTreeTinyAndNotFullyVectorizable(bool ForReduction = false) const; | |||
1276 | ||||
1277 | /// Assume that a legal-sized 'or'-reduction of shifted/zexted loaded values | |||
1278 | /// can be load combined in the backend. Load combining may not be allowed in | |||
1279 | /// the IR optimizer, so we do not want to alter the pattern. For example, | |||
1280 | /// partially transforming a scalar bswap() pattern into vector code is | |||
1281 | /// effectively impossible for the backend to undo. | |||
1282 | /// TODO: If load combining is allowed in the IR optimizer, this analysis | |||
1283 | /// may not be necessary. | |||
1284 | bool isLoadCombineReductionCandidate(RecurKind RdxKind) const; | |||
1285 | ||||
1286 | /// Assume that a vector of stores of bitwise-or/shifted/zexted loaded values | |||
1287 | /// can be load combined in the backend. Load combining may not be allowed in | |||
1288 | /// the IR optimizer, so we do not want to alter the pattern. For example, | |||
1289 | /// partially transforming a scalar bswap() pattern into vector code is | |||
1290 | /// effectively impossible for the backend to undo. | |||
1291 | /// TODO: If load combining is allowed in the IR optimizer, this analysis | |||
1292 | /// may not be necessary. | |||
1293 | bool isLoadCombineCandidate() const; | |||
1294 | ||||
1295 | OptimizationRemarkEmitter *getORE() { return ORE; } | |||
1296 | ||||
1297 | /// This structure holds any data we need about the edges being traversed | |||
1298 | /// during buildTree_rec(). We keep track of: | |||
1299 | /// (i) the user TreeEntry index, and | |||
1300 | /// (ii) the index of the edge. | |||
1301 | struct EdgeInfo { | |||
1302 | EdgeInfo() = default; | |||
1303 | EdgeInfo(TreeEntry *UserTE, unsigned EdgeIdx) | |||
1304 | : UserTE(UserTE), EdgeIdx(EdgeIdx) {} | |||
1305 | /// The user TreeEntry. | |||
1306 | TreeEntry *UserTE = nullptr; | |||
1307 | /// The operand index of the use. | |||
1308 | unsigned EdgeIdx = UINT_MAX(2147483647 *2U +1U); | |||
1309 | #ifndef NDEBUG | |||
1310 | friend inline raw_ostream &operator<<(raw_ostream &OS, | |||
1311 | const BoUpSLP::EdgeInfo &EI) { | |||
1312 | EI.dump(OS); | |||
1313 | return OS; | |||
1314 | } | |||
1315 | /// Debug print. | |||
1316 | void dump(raw_ostream &OS) const { | |||
1317 | OS << "{User:" << (UserTE ? std::to_string(UserTE->Idx) : "null") | |||
1318 | << " EdgeIdx:" << EdgeIdx << "}"; | |||
1319 | } | |||
1320 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void dump() const { dump(dbgs()); } | |||
1321 | #endif | |||
1322 | }; | |||
1323 | ||||
1324 | /// A helper class used for scoring candidates for two consecutive lanes. | |||
1325 | class LookAheadHeuristics { | |||
1326 | const TargetLibraryInfo &TLI; | |||
1327 | const DataLayout &DL; | |||
1328 | ScalarEvolution &SE; | |||
1329 | const BoUpSLP &R; | |||
1330 | int NumLanes; // Total number of lanes (aka vectorization factor). | |||
1331 | int MaxLevel; // The maximum recursion depth for accumulating score. | |||
1332 | ||||
1333 | public: | |||
1334 | LookAheadHeuristics(const TargetLibraryInfo &TLI, const DataLayout &DL, | |||
1335 | ScalarEvolution &SE, const BoUpSLP &R, int NumLanes, | |||
1336 | int MaxLevel) | |||
1337 | : TLI(TLI), DL(DL), SE(SE), R(R), NumLanes(NumLanes), | |||
1338 | MaxLevel(MaxLevel) {} | |||
1339 | ||||
1340 | // The hard-coded scores listed here are not very important, though it shall | |||
1341 | // be higher for better matches to improve the resulting cost. When | |||
1342 | // computing the scores of matching one sub-tree with another, we are | |||
1343 | // basically counting the number of values that are matching. So even if all | |||
1344 | // scores are set to 1, we would still get a decent matching result. | |||
1345 | // However, sometimes we have to break ties. For example we may have to | |||
1346 | // choose between matching loads vs matching opcodes. This is what these | |||
1347 | // scores are helping us with: they provide the order of preference. Also, | |||
1348 | // this is important if the scalar is externally used or used in another | |||
1349 | // tree entry node in the different lane. | |||
1350 | ||||
1351 | /// Loads from consecutive memory addresses, e.g. load(A[i]), load(A[i+1]). | |||
1352 | static const int ScoreConsecutiveLoads = 4; | |||
1353 | /// The same load multiple times. This should have a better score than | |||
1354 | /// `ScoreSplat` because it in x86 for a 2-lane vector we can represent it | |||
1355 | /// with `movddup (%reg), xmm0` which has a throughput of 0.5 versus 0.5 for | |||
1356 | /// a vector load and 1.0 for a broadcast. | |||
1357 | static const int ScoreSplatLoads = 3; | |||
1358 | /// Loads from reversed memory addresses, e.g. load(A[i+1]), load(A[i]). | |||
1359 | static const int ScoreReversedLoads = 3; | |||
1360 | /// A load candidate for masked gather. | |||
1361 | static const int ScoreMaskedGatherCandidate = 1; | |||
1362 | /// ExtractElementInst from same vector and consecutive indexes. | |||
1363 | static const int ScoreConsecutiveExtracts = 4; | |||
1364 | /// ExtractElementInst from same vector and reversed indices. | |||
1365 | static const int ScoreReversedExtracts = 3; | |||
1366 | /// Constants. | |||
1367 | static const int ScoreConstants = 2; | |||
1368 | /// Instructions with the same opcode. | |||
1369 | static const int ScoreSameOpcode = 2; | |||
1370 | /// Instructions with alt opcodes (e.g, add + sub). | |||
1371 | static const int ScoreAltOpcodes = 1; | |||
1372 | /// Identical instructions (a.k.a. splat or broadcast). | |||
1373 | static const int ScoreSplat = 1; | |||
1374 | /// Matching with an undef is preferable to failing. | |||
1375 | static const int ScoreUndef = 1; | |||
1376 | /// Score for failing to find a decent match. | |||
1377 | static const int ScoreFail = 0; | |||
1378 | /// Score if all users are vectorized. | |||
1379 | static const int ScoreAllUserVectorized = 1; | |||
1380 | ||||
1381 | /// \returns the score of placing \p V1 and \p V2 in consecutive lanes. | |||
1382 | /// \p U1 and \p U2 are the users of \p V1 and \p V2. | |||
1383 | /// Also, checks if \p V1 and \p V2 are compatible with instructions in \p | |||
1384 | /// MainAltOps. | |||
1385 | int getShallowScore(Value *V1, Value *V2, Instruction *U1, Instruction *U2, | |||
1386 | ArrayRef<Value *> MainAltOps) const { | |||
1387 | if (!isValidElementType(V1->getType()) || | |||
1388 | !isValidElementType(V2->getType())) | |||
1389 | return LookAheadHeuristics::ScoreFail; | |||
1390 | ||||
1391 | if (V1 == V2) { | |||
1392 | if (isa<LoadInst>(V1)) { | |||
1393 | // Retruns true if the users of V1 and V2 won't need to be extracted. | |||
1394 | auto AllUsersAreInternal = [U1, U2, this](Value *V1, Value *V2) { | |||
1395 | // Bail out if we have too many uses to save compilation time. | |||
1396 | static constexpr unsigned Limit = 8; | |||
1397 | if (V1->hasNUsesOrMore(Limit) || V2->hasNUsesOrMore(Limit)) | |||
1398 | return false; | |||
1399 | ||||
1400 | auto AllUsersVectorized = [U1, U2, this](Value *V) { | |||
1401 | return llvm::all_of(V->users(), [U1, U2, this](Value *U) { | |||
1402 | return U == U1 || U == U2 || R.getTreeEntry(U) != nullptr; | |||
1403 | }); | |||
1404 | }; | |||
1405 | return AllUsersVectorized(V1) && AllUsersVectorized(V2); | |||
1406 | }; | |||
1407 | // A broadcast of a load can be cheaper on some targets. | |||
1408 | if (R.TTI->isLegalBroadcastLoad(V1->getType(), | |||
1409 | ElementCount::getFixed(NumLanes)) && | |||
1410 | ((int)V1->getNumUses() == NumLanes || | |||
1411 | AllUsersAreInternal(V1, V2))) | |||
1412 | return LookAheadHeuristics::ScoreSplatLoads; | |||
1413 | } | |||
1414 | return LookAheadHeuristics::ScoreSplat; | |||
1415 | } | |||
1416 | ||||
1417 | auto *LI1 = dyn_cast<LoadInst>(V1); | |||
1418 | auto *LI2 = dyn_cast<LoadInst>(V2); | |||
1419 | if (LI1 && LI2) { | |||
1420 | if (LI1->getParent() != LI2->getParent() || !LI1->isSimple() || | |||
1421 | !LI2->isSimple()) | |||
1422 | return LookAheadHeuristics::ScoreFail; | |||
1423 | ||||
1424 | std::optional<int> Dist = getPointersDiff( | |||
1425 | LI1->getType(), LI1->getPointerOperand(), LI2->getType(), | |||
1426 | LI2->getPointerOperand(), DL, SE, /*StrictCheck=*/true); | |||
1427 | if (!Dist || *Dist == 0) { | |||
1428 | if (getUnderlyingObject(LI1->getPointerOperand()) == | |||
1429 | getUnderlyingObject(LI2->getPointerOperand()) && | |||
1430 | R.TTI->isLegalMaskedGather( | |||
1431 | FixedVectorType::get(LI1->getType(), NumLanes), | |||
1432 | LI1->getAlign())) | |||
1433 | return LookAheadHeuristics::ScoreMaskedGatherCandidate; | |||
1434 | return LookAheadHeuristics::ScoreFail; | |||
1435 | } | |||
1436 | // The distance is too large - still may be profitable to use masked | |||
1437 | // loads/gathers. | |||
1438 | if (std::abs(*Dist) > NumLanes / 2) | |||
1439 | return LookAheadHeuristics::ScoreMaskedGatherCandidate; | |||
1440 | // This still will detect consecutive loads, but we might have "holes" | |||
1441 | // in some cases. It is ok for non-power-2 vectorization and may produce | |||
1442 | // better results. It should not affect current vectorization. | |||
1443 | return (*Dist > 0) ? LookAheadHeuristics::ScoreConsecutiveLoads | |||
1444 | : LookAheadHeuristics::ScoreReversedLoads; | |||
1445 | } | |||
1446 | ||||
1447 | auto *C1 = dyn_cast<Constant>(V1); | |||
1448 | auto *C2 = dyn_cast<Constant>(V2); | |||
1449 | if (C1 && C2) | |||
1450 | return LookAheadHeuristics::ScoreConstants; | |||
1451 | ||||
1452 | // Extracts from consecutive indexes of the same vector better score as | |||
1453 | // the extracts could be optimized away. | |||
1454 | Value *EV1; | |||
1455 | ConstantInt *Ex1Idx; | |||
1456 | if (match(V1, m_ExtractElt(m_Value(EV1), m_ConstantInt(Ex1Idx)))) { | |||
1457 | // Undefs are always profitable for extractelements. | |||
1458 | // Compiler can easily combine poison and extractelement <non-poison> or | |||
1459 | // undef and extractelement <poison>. But combining undef + | |||
1460 | // extractelement <non-poison-but-may-produce-poison> requires some | |||
1461 | // extra operations. | |||
1462 | if (isa<UndefValue>(V2)) | |||
1463 | return (isa<PoisonValue>(V2) || isUndefVector(EV1).all()) | |||
1464 | ? LookAheadHeuristics::ScoreConsecutiveExtracts | |||
1465 | : LookAheadHeuristics::ScoreSameOpcode; | |||
1466 | Value *EV2 = nullptr; | |||
1467 | ConstantInt *Ex2Idx = nullptr; | |||
1468 | if (match(V2, | |||
1469 | m_ExtractElt(m_Value(EV2), m_CombineOr(m_ConstantInt(Ex2Idx), | |||
1470 | m_Undef())))) { | |||
1471 | // Undefs are always profitable for extractelements. | |||
1472 | if (!Ex2Idx) | |||
1473 | return LookAheadHeuristics::ScoreConsecutiveExtracts; | |||
1474 | if (isUndefVector(EV2).all() && EV2->getType() == EV1->getType()) | |||
1475 | return LookAheadHeuristics::ScoreConsecutiveExtracts; | |||
1476 | if (EV2 == EV1) { | |||
1477 | int Idx1 = Ex1Idx->getZExtValue(); | |||
1478 | int Idx2 = Ex2Idx->getZExtValue(); | |||
1479 | int Dist = Idx2 - Idx1; | |||
1480 | // The distance is too large - still may be profitable to use | |||
1481 | // shuffles. | |||
1482 | if (std::abs(Dist) == 0) | |||
1483 | return LookAheadHeuristics::ScoreSplat; | |||
1484 | if (std::abs(Dist) > NumLanes / 2) | |||
1485 | return LookAheadHeuristics::ScoreSameOpcode; | |||
1486 | return (Dist > 0) ? LookAheadHeuristics::ScoreConsecutiveExtracts | |||
1487 | : LookAheadHeuristics::ScoreReversedExtracts; | |||
1488 | } | |||
1489 | return LookAheadHeuristics::ScoreAltOpcodes; | |||
1490 | } | |||
1491 | return LookAheadHeuristics::ScoreFail; | |||
1492 | } | |||
1493 | ||||
1494 | auto *I1 = dyn_cast<Instruction>(V1); | |||
1495 | auto *I2 = dyn_cast<Instruction>(V2); | |||
1496 | if (I1 && I2) { | |||
1497 | if (I1->getParent() != I2->getParent()) | |||
1498 | return LookAheadHeuristics::ScoreFail; | |||
1499 | SmallVector<Value *, 4> Ops(MainAltOps.begin(), MainAltOps.end()); | |||
1500 | Ops.push_back(I1); | |||
1501 | Ops.push_back(I2); | |||
1502 | InstructionsState S = getSameOpcode(Ops, TLI); | |||
1503 | // Note: Only consider instructions with <= 2 operands to avoid | |||
1504 | // complexity explosion. | |||
1505 | if (S.getOpcode() && | |||
1506 | (S.MainOp->getNumOperands() <= 2 || !MainAltOps.empty() || | |||
1507 | !S.isAltShuffle()) && | |||
1508 | all_of(Ops, [&S](Value *V) { | |||
1509 | return cast<Instruction>(V)->getNumOperands() == | |||
1510 | S.MainOp->getNumOperands(); | |||
1511 | })) | |||
1512 | return S.isAltShuffle() ? LookAheadHeuristics::ScoreAltOpcodes | |||
1513 | : LookAheadHeuristics::ScoreSameOpcode; | |||
1514 | } | |||
1515 | ||||
1516 | if (isa<UndefValue>(V2)) | |||
1517 | return LookAheadHeuristics::ScoreUndef; | |||
1518 | ||||
1519 | return LookAheadHeuristics::ScoreFail; | |||
1520 | } | |||
1521 | ||||
1522 | /// Go through the operands of \p LHS and \p RHS recursively until | |||
1523 | /// MaxLevel, and return the cummulative score. \p U1 and \p U2 are | |||
1524 | /// the users of \p LHS and \p RHS (that is \p LHS and \p RHS are operands | |||
1525 | /// of \p U1 and \p U2), except at the beginning of the recursion where | |||
1526 | /// these are set to nullptr. | |||
1527 | /// | |||
1528 | /// For example: | |||
1529 | /// \verbatim | |||
1530 | /// A[0] B[0] A[1] B[1] C[0] D[0] B[1] A[1] | |||
1531 | /// \ / \ / \ / \ / | |||
1532 | /// + + + + | |||
1533 | /// G1 G2 G3 G4 | |||
1534 | /// \endverbatim | |||
1535 | /// The getScoreAtLevelRec(G1, G2) function will try to match the nodes at | |||
1536 | /// each level recursively, accumulating the score. It starts from matching | |||
1537 | /// the additions at level 0, then moves on to the loads (level 1). The | |||
1538 | /// score of G1 and G2 is higher than G1 and G3, because {A[0],A[1]} and | |||
1539 | /// {B[0],B[1]} match with LookAheadHeuristics::ScoreConsecutiveLoads, while | |||
1540 | /// {A[0],C[0]} has a score of LookAheadHeuristics::ScoreFail. | |||
1541 | /// Please note that the order of the operands does not matter, as we | |||
1542 | /// evaluate the score of all profitable combinations of operands. In | |||
1543 | /// other words the score of G1 and G4 is the same as G1 and G2. This | |||
1544 | /// heuristic is based on ideas described in: | |||
1545 | /// Look-ahead SLP: Auto-vectorization in the presence of commutative | |||
1546 | /// operations, CGO 2018 by Vasileios Porpodas, Rodrigo C. O. Rocha, | |||
1547 | /// Luís F. W. Góes | |||
1548 | int getScoreAtLevelRec(Value *LHS, Value *RHS, Instruction *U1, | |||
1549 | Instruction *U2, int CurrLevel, | |||
1550 | ArrayRef<Value *> MainAltOps) const { | |||
1551 | ||||
1552 | // Get the shallow score of V1 and V2. | |||
1553 | int ShallowScoreAtThisLevel = | |||
1554 | getShallowScore(LHS, RHS, U1, U2, MainAltOps); | |||
1555 | ||||
1556 | // If reached MaxLevel, | |||
1557 | // or if V1 and V2 are not instructions, | |||
1558 | // or if they are SPLAT, | |||
1559 | // or if they are not consecutive, | |||
1560 | // or if profitable to vectorize loads or extractelements, early return | |||
1561 | // the current cost. | |||
1562 | auto *I1 = dyn_cast<Instruction>(LHS); | |||
1563 | auto *I2 = dyn_cast<Instruction>(RHS); | |||
1564 | if (CurrLevel == MaxLevel || !(I1 && I2) || I1 == I2 || | |||
1565 | ShallowScoreAtThisLevel == LookAheadHeuristics::ScoreFail || | |||
1566 | (((isa<LoadInst>(I1) && isa<LoadInst>(I2)) || | |||
1567 | (I1->getNumOperands() > 2 && I2->getNumOperands() > 2) || | |||
1568 | (isa<ExtractElementInst>(I1) && isa<ExtractElementInst>(I2))) && | |||
1569 | ShallowScoreAtThisLevel)) | |||
1570 | return ShallowScoreAtThisLevel; | |||
1571 | 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", 1571, __extension__ __PRETTY_FUNCTION__)); | |||
1572 | ||||
1573 | // Contains the I2 operand indexes that got matched with I1 operands. | |||
1574 | SmallSet<unsigned, 4> Op2Used; | |||
1575 | ||||
1576 | // Recursion towards the operands of I1 and I2. We are trying all possible | |||
1577 | // operand pairs, and keeping track of the best score. | |||
1578 | for (unsigned OpIdx1 = 0, NumOperands1 = I1->getNumOperands(); | |||
1579 | OpIdx1 != NumOperands1; ++OpIdx1) { | |||
1580 | // Try to pair op1I with the best operand of I2. | |||
1581 | int MaxTmpScore = 0; | |||
1582 | unsigned MaxOpIdx2 = 0; | |||
1583 | bool FoundBest = false; | |||
1584 | // If I2 is commutative try all combinations. | |||
1585 | unsigned FromIdx = isCommutative(I2) ? 0 : OpIdx1; | |||
1586 | unsigned ToIdx = isCommutative(I2) | |||
1587 | ? I2->getNumOperands() | |||
1588 | : std::min(I2->getNumOperands(), OpIdx1 + 1); | |||
1589 | 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", 1589, __extension__ __PRETTY_FUNCTION__)); | |||
1590 | for (unsigned OpIdx2 = FromIdx; OpIdx2 != ToIdx; ++OpIdx2) { | |||
1591 | // Skip operands already paired with OpIdx1. | |||
1592 | if (Op2Used.count(OpIdx2)) | |||
1593 | continue; | |||
1594 | // Recursively calculate the cost at each level | |||
1595 | int TmpScore = | |||
1596 | getScoreAtLevelRec(I1->getOperand(OpIdx1), I2->getOperand(OpIdx2), | |||
1597 | I1, I2, CurrLevel + 1, std::nullopt); | |||
1598 | // Look for the best score. | |||
1599 | if (TmpScore > LookAheadHeuristics::ScoreFail && | |||
1600 | TmpScore > MaxTmpScore) { | |||
1601 | MaxTmpScore = TmpScore; | |||
1602 | MaxOpIdx2 = OpIdx2; | |||
1603 | FoundBest = true; | |||
1604 | } | |||
1605 | } | |||
1606 | if (FoundBest) { | |||
1607 | // Pair {OpIdx1, MaxOpIdx2} was found to be best. Never revisit it. | |||
1608 | Op2Used.insert(MaxOpIdx2); | |||
1609 | ShallowScoreAtThisLevel += MaxTmpScore; | |||
1610 | } | |||
1611 | } | |||
1612 | return ShallowScoreAtThisLevel; | |||
1613 | } | |||
1614 | }; | |||
1615 | /// A helper data structure to hold the operands of a vector of instructions. | |||
1616 | /// This supports a fixed vector length for all operand vectors. | |||
1617 | class VLOperands { | |||
1618 | /// For each operand we need (i) the value, and (ii) the opcode that it | |||
1619 | /// would be attached to if the expression was in a left-linearized form. | |||
1620 | /// This is required to avoid illegal operand reordering. | |||
1621 | /// For example: | |||
1622 | /// \verbatim | |||
1623 | /// 0 Op1 | |||
1624 | /// |/ | |||
1625 | /// Op1 Op2 Linearized + Op2 | |||
1626 | /// \ / ----------> |/ | |||
1627 | /// - - | |||
1628 | /// | |||
1629 | /// Op1 - Op2 (0 + Op1) - Op2 | |||
1630 | /// \endverbatim | |||
1631 | /// | |||
1632 | /// Value Op1 is attached to a '+' operation, and Op2 to a '-'. | |||
1633 | /// | |||
1634 | /// Another way to think of this is to track all the operations across the | |||
1635 | /// path from the operand all the way to the root of the tree and to | |||
1636 | /// calculate the operation that corresponds to this path. For example, the | |||
1637 | /// path from Op2 to the root crosses the RHS of the '-', therefore the | |||
1638 | /// corresponding operation is a '-' (which matches the one in the | |||
1639 | /// linearized tree, as shown above). | |||
1640 | /// | |||
1641 | /// For lack of a better term, we refer to this operation as Accumulated | |||
1642 | /// Path Operation (APO). | |||
1643 | struct OperandData { | |||
1644 | OperandData() = default; | |||
1645 | OperandData(Value *V, bool APO, bool IsUsed) | |||
1646 | : V(V), APO(APO), IsUsed(IsUsed) {} | |||
1647 | /// The operand value. | |||
1648 | Value *V = nullptr; | |||
1649 | /// TreeEntries only allow a single opcode, or an alternate sequence of | |||
1650 | /// them (e.g, +, -). Therefore, we can safely use a boolean value for the | |||
1651 | /// APO. It is set to 'true' if 'V' is attached to an inverse operation | |||
1652 | /// in the left-linearized form (e.g., Sub/Div), and 'false' otherwise | |||
1653 | /// (e.g., Add/Mul) | |||
1654 | bool APO = false; | |||
1655 | /// Helper data for the reordering function. | |||
1656 | bool IsUsed = false; | |||
1657 | }; | |||
1658 | ||||
1659 | /// During operand reordering, we are trying to select the operand at lane | |||
1660 | /// that matches best with the operand at the neighboring lane. Our | |||
1661 | /// selection is based on the type of value we are looking for. For example, | |||
1662 | /// if the neighboring lane has a load, we need to look for a load that is | |||
1663 | /// accessing a consecutive address. These strategies are summarized in the | |||
1664 | /// 'ReorderingMode' enumerator. | |||
1665 | enum class ReorderingMode { | |||
1666 | Load, ///< Matching loads to consecutive memory addresses | |||
1667 | Opcode, ///< Matching instructions based on opcode (same or alternate) | |||
1668 | Constant, ///< Matching constants | |||
1669 | Splat, ///< Matching the same instruction multiple times (broadcast) | |||
1670 | Failed, ///< We failed to create a vectorizable group | |||
1671 | }; | |||
1672 | ||||
1673 | using OperandDataVec = SmallVector<OperandData, 2>; | |||
1674 | ||||
1675 | /// A vector of operand vectors. | |||
1676 | SmallVector<OperandDataVec, 4> OpsVec; | |||
1677 | ||||
1678 | const TargetLibraryInfo &TLI; | |||
1679 | const DataLayout &DL; | |||
1680 | ScalarEvolution &SE; | |||
1681 | const BoUpSLP &R; | |||
1682 | ||||
1683 | /// \returns the operand data at \p OpIdx and \p Lane. | |||
1684 | OperandData &getData(unsigned OpIdx, unsigned Lane) { | |||
1685 | return OpsVec[OpIdx][Lane]; | |||
1686 | } | |||
1687 | ||||
1688 | /// \returns the operand data at \p OpIdx and \p Lane. Const version. | |||
1689 | const OperandData &getData(unsigned OpIdx, unsigned Lane) const { | |||
1690 | return OpsVec[OpIdx][Lane]; | |||
1691 | } | |||
1692 | ||||
1693 | /// Clears the used flag for all entries. | |||
1694 | void clearUsed() { | |||
1695 | for (unsigned OpIdx = 0, NumOperands = getNumOperands(); | |||
1696 | OpIdx != NumOperands; ++OpIdx) | |||
1697 | for (unsigned Lane = 0, NumLanes = getNumLanes(); Lane != NumLanes; | |||
1698 | ++Lane) | |||
1699 | OpsVec[OpIdx][Lane].IsUsed = false; | |||
1700 | } | |||
1701 | ||||
1702 | /// Swap the operand at \p OpIdx1 with that one at \p OpIdx2. | |||
1703 | void swap(unsigned OpIdx1, unsigned OpIdx2, unsigned Lane) { | |||
1704 | std::swap(OpsVec[OpIdx1][Lane], OpsVec[OpIdx2][Lane]); | |||
1705 | } | |||
1706 | ||||
1707 | /// \param Lane lane of the operands under analysis. | |||
1708 | /// \param OpIdx operand index in \p Lane lane we're looking the best | |||
1709 | /// candidate for. | |||
1710 | /// \param Idx operand index of the current candidate value. | |||
1711 | /// \returns The additional score due to possible broadcasting of the | |||
1712 | /// elements in the lane. It is more profitable to have power-of-2 unique | |||
1713 | /// elements in the lane, it will be vectorized with higher probability | |||
1714 | /// after removing duplicates. Currently the SLP vectorizer supports only | |||
1715 | /// vectorization of the power-of-2 number of unique scalars. | |||
1716 | int getSplatScore(unsigned Lane, unsigned OpIdx, unsigned Idx) const { | |||
1717 | Value *IdxLaneV = getData(Idx, Lane).V; | |||
1718 | if (!isa<Instruction>(IdxLaneV) || IdxLaneV == getData(OpIdx, Lane).V) | |||
1719 | return 0; | |||
1720 | SmallPtrSet<Value *, 4> Uniques; | |||
1721 | for (unsigned Ln = 0, E = getNumLanes(); Ln < E; ++Ln) { | |||
1722 | if (Ln == Lane) | |||
1723 | continue; | |||
1724 | Value *OpIdxLnV = getData(OpIdx, Ln).V; | |||
1725 | if (!isa<Instruction>(OpIdxLnV)) | |||
1726 | return 0; | |||
1727 | Uniques.insert(OpIdxLnV); | |||
1728 | } | |||
1729 | int UniquesCount = Uniques.size(); | |||
1730 | int UniquesCntWithIdxLaneV = | |||
1731 | Uniques.contains(IdxLaneV) ? UniquesCount : UniquesCount + 1; | |||
1732 | Value *OpIdxLaneV = getData(OpIdx, Lane).V; | |||
1733 | int UniquesCntWithOpIdxLaneV = | |||
1734 | Uniques.contains(OpIdxLaneV) ? UniquesCount : UniquesCount + 1; | |||
1735 | if (UniquesCntWithIdxLaneV == UniquesCntWithOpIdxLaneV) | |||
1736 | return 0; | |||
1737 | return (PowerOf2Ceil(UniquesCntWithOpIdxLaneV) - | |||
1738 | UniquesCntWithOpIdxLaneV) - | |||
1739 | (PowerOf2Ceil(UniquesCntWithIdxLaneV) - UniquesCntWithIdxLaneV); | |||
1740 | } | |||
1741 | ||||
1742 | /// \param Lane lane of the operands under analysis. | |||
1743 | /// \param OpIdx operand index in \p Lane lane we're looking the best | |||
1744 | /// candidate for. | |||
1745 | /// \param Idx operand index of the current candidate value. | |||
1746 | /// \returns The additional score for the scalar which users are all | |||
1747 | /// vectorized. | |||
1748 | int getExternalUseScore(unsigned Lane, unsigned OpIdx, unsigned Idx) const { | |||
1749 | Value *IdxLaneV = getData(Idx, Lane).V; | |||
1750 | Value *OpIdxLaneV = getData(OpIdx, Lane).V; | |||
1751 | // Do not care about number of uses for vector-like instructions | |||
1752 | // (extractelement/extractvalue with constant indices), they are extracts | |||
1753 | // themselves and already externally used. Vectorization of such | |||
1754 | // instructions does not add extra extractelement instruction, just may | |||
1755 | // remove it. | |||
1756 | if (isVectorLikeInstWithConstOps(IdxLaneV) && | |||
1757 | isVectorLikeInstWithConstOps(OpIdxLaneV)) | |||
1758 | return LookAheadHeuristics::ScoreAllUserVectorized; | |||
1759 | auto *IdxLaneI = dyn_cast<Instruction>(IdxLaneV); | |||
1760 | if (!IdxLaneI || !isa<Instruction>(OpIdxLaneV)) | |||
1761 | return 0; | |||
1762 | return R.areAllUsersVectorized(IdxLaneI, std::nullopt) | |||
1763 | ? LookAheadHeuristics::ScoreAllUserVectorized | |||
1764 | : 0; | |||
1765 | } | |||
1766 | ||||
1767 | /// Score scaling factor for fully compatible instructions but with | |||
1768 | /// different number of external uses. Allows better selection of the | |||
1769 | /// instructions with less external uses. | |||
1770 | static const int ScoreScaleFactor = 10; | |||
1771 | ||||
1772 | /// \Returns the look-ahead score, which tells us how much the sub-trees | |||
1773 | /// rooted at \p LHS and \p RHS match, the more they match the higher the | |||
1774 | /// score. This helps break ties in an informed way when we cannot decide on | |||
1775 | /// the order of the operands by just considering the immediate | |||
1776 | /// predecessors. | |||
1777 | int getLookAheadScore(Value *LHS, Value *RHS, ArrayRef<Value *> MainAltOps, | |||
1778 | int Lane, unsigned OpIdx, unsigned Idx, | |||
1779 | bool &IsUsed) { | |||
1780 | LookAheadHeuristics LookAhead(TLI, DL, SE, R, getNumLanes(), | |||
1781 | LookAheadMaxDepth); | |||
1782 | // Keep track of the instruction stack as we recurse into the operands | |||
1783 | // during the look-ahead score exploration. | |||
1784 | int Score = | |||
1785 | LookAhead.getScoreAtLevelRec(LHS, RHS, /*U1=*/nullptr, /*U2=*/nullptr, | |||
1786 | /*CurrLevel=*/1, MainAltOps); | |||
1787 | if (Score) { | |||
1788 | int SplatScore = getSplatScore(Lane, OpIdx, Idx); | |||
1789 | if (Score <= -SplatScore) { | |||
1790 | // Set the minimum score for splat-like sequence to avoid setting | |||
1791 | // failed state. | |||
1792 | Score = 1; | |||
1793 | } else { | |||
1794 | Score += SplatScore; | |||
1795 | // Scale score to see the difference between different operands | |||
1796 | // and similar operands but all vectorized/not all vectorized | |||
1797 | // uses. It does not affect actual selection of the best | |||
1798 | // compatible operand in general, just allows to select the | |||
1799 | // operand with all vectorized uses. | |||
1800 | Score *= ScoreScaleFactor; | |||
1801 | Score += getExternalUseScore(Lane, OpIdx, Idx); | |||
1802 | IsUsed = true; | |||
1803 | } | |||
1804 | } | |||
1805 | return Score; | |||
1806 | } | |||
1807 | ||||
1808 | /// Best defined scores per lanes between the passes. Used to choose the | |||
1809 | /// best operand (with the highest score) between the passes. | |||
1810 | /// The key - {Operand Index, Lane}. | |||
1811 | /// The value - the best score between the passes for the lane and the | |||
1812 | /// operand. | |||
1813 | SmallDenseMap<std::pair<unsigned, unsigned>, unsigned, 8> | |||
1814 | BestScoresPerLanes; | |||
1815 | ||||
1816 | // Search all operands in Ops[*][Lane] for the one that matches best | |||
1817 | // Ops[OpIdx][LastLane] and return its opreand index. | |||
1818 | // If no good match can be found, return std::nullopt. | |||
1819 | std::optional<unsigned> | |||
1820 | getBestOperand(unsigned OpIdx, int Lane, int LastLane, | |||
1821 | ArrayRef<ReorderingMode> ReorderingModes, | |||
1822 | ArrayRef<Value *> MainAltOps) { | |||
1823 | unsigned NumOperands = getNumOperands(); | |||
1824 | ||||
1825 | // The operand of the previous lane at OpIdx. | |||
1826 | Value *OpLastLane = getData(OpIdx, LastLane).V; | |||
1827 | ||||
1828 | // Our strategy mode for OpIdx. | |||
1829 | ReorderingMode RMode = ReorderingModes[OpIdx]; | |||
1830 | if (RMode == ReorderingMode::Failed) | |||
1831 | return std::nullopt; | |||
1832 | ||||
1833 | // The linearized opcode of the operand at OpIdx, Lane. | |||
1834 | bool OpIdxAPO = getData(OpIdx, Lane).APO; | |||
1835 | ||||
1836 | // The best operand index and its score. | |||
1837 | // Sometimes we have more than one option (e.g., Opcode and Undefs), so we | |||
1838 | // are using the score to differentiate between the two. | |||
1839 | struct BestOpData { | |||
1840 | std::optional<unsigned> Idx; | |||
1841 | unsigned Score = 0; | |||
1842 | } BestOp; | |||
1843 | BestOp.Score = | |||
1844 | BestScoresPerLanes.try_emplace(std::make_pair(OpIdx, Lane), 0) | |||
1845 | .first->second; | |||
1846 | ||||
1847 | // Track if the operand must be marked as used. If the operand is set to | |||
1848 | // Score 1 explicitly (because of non power-of-2 unique scalars, we may | |||
1849 | // want to reestimate the operands again on the following iterations). | |||
1850 | bool IsUsed = | |||
1851 | RMode == ReorderingMode::Splat || RMode == ReorderingMode::Constant; | |||
1852 | // Iterate through all unused operands and look for the best. | |||
1853 | for (unsigned Idx = 0; Idx != NumOperands; ++Idx) { | |||
1854 | // Get the operand at Idx and Lane. | |||
1855 | OperandData &OpData = getData(Idx, Lane); | |||
1856 | Value *Op = OpData.V; | |||
1857 | bool OpAPO = OpData.APO; | |||
1858 | ||||
1859 | // Skip already selected operands. | |||
1860 | if (OpData.IsUsed) | |||
1861 | continue; | |||
1862 | ||||
1863 | // Skip if we are trying to move the operand to a position with a | |||
1864 | // different opcode in the linearized tree form. This would break the | |||
1865 | // semantics. | |||
1866 | if (OpAPO != OpIdxAPO) | |||
1867 | continue; | |||
1868 | ||||
1869 | // Look for an operand that matches the current mode. | |||
1870 | switch (RMode) { | |||
1871 | case ReorderingMode::Load: | |||
1872 | case ReorderingMode::Constant: | |||
1873 | case ReorderingMode::Opcode: { | |||
1874 | bool LeftToRight = Lane > LastLane; | |||
1875 | Value *OpLeft = (LeftToRight) ? OpLastLane : Op; | |||
1876 | Value *OpRight = (LeftToRight) ? Op : OpLastLane; | |||
1877 | int Score = getLookAheadScore(OpLeft, OpRight, MainAltOps, Lane, | |||
1878 | OpIdx, Idx, IsUsed); | |||
1879 | if (Score > static_cast<int>(BestOp.Score)) { | |||
1880 | BestOp.Idx = Idx; | |||
1881 | BestOp.Score = Score; | |||
1882 | BestScoresPerLanes[std::make_pair(OpIdx, Lane)] = Score; | |||
1883 | } | |||
1884 | break; | |||
1885 | } | |||
1886 | case ReorderingMode::Splat: | |||
1887 | if (Op == OpLastLane) | |||
1888 | BestOp.Idx = Idx; | |||
1889 | break; | |||
1890 | case ReorderingMode::Failed: | |||
1891 | llvm_unreachable("Not expected Failed reordering mode.")::llvm::llvm_unreachable_internal("Not expected Failed reordering mode." , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 1891); | |||
1892 | } | |||
1893 | } | |||
1894 | ||||
1895 | if (BestOp.Idx) { | |||
1896 | getData(*BestOp.Idx, Lane).IsUsed = IsUsed; | |||
1897 | return BestOp.Idx; | |||
1898 | } | |||
1899 | // If we could not find a good match return std::nullopt. | |||
1900 | return std::nullopt; | |||
1901 | } | |||
1902 | ||||
1903 | /// Helper for reorderOperandVecs. | |||
1904 | /// \returns the lane that we should start reordering from. This is the one | |||
1905 | /// which has the least number of operands that can freely move about or | |||
1906 | /// less profitable because it already has the most optimal set of operands. | |||
1907 | unsigned getBestLaneToStartReordering() const { | |||
1908 | unsigned Min = UINT_MAX(2147483647 *2U +1U); | |||
1909 | unsigned SameOpNumber = 0; | |||
1910 | // std::pair<unsigned, unsigned> is used to implement a simple voting | |||
1911 | // algorithm and choose the lane with the least number of operands that | |||
1912 | // can freely move about or less profitable because it already has the | |||
1913 | // most optimal set of operands. The first unsigned is a counter for | |||
1914 | // voting, the second unsigned is the counter of lanes with instructions | |||
1915 | // with same/alternate opcodes and same parent basic block. | |||
1916 | MapVector<unsigned, std::pair<unsigned, unsigned>> HashMap; | |||
1917 | // Try to be closer to the original results, if we have multiple lanes | |||
1918 | // with same cost. If 2 lanes have the same cost, use the one with the | |||
1919 | // lowest index. | |||
1920 | for (int I = getNumLanes(); I > 0; --I) { | |||
1921 | unsigned Lane = I - 1; | |||
1922 | OperandsOrderData NumFreeOpsHash = | |||
1923 | getMaxNumOperandsThatCanBeReordered(Lane); | |||
1924 | // Compare the number of operands that can move and choose the one with | |||
1925 | // the least number. | |||
1926 | if (NumFreeOpsHash.NumOfAPOs < Min) { | |||
1927 | Min = NumFreeOpsHash.NumOfAPOs; | |||
1928 | SameOpNumber = NumFreeOpsHash.NumOpsWithSameOpcodeParent; | |||
1929 | HashMap.clear(); | |||
1930 | HashMap[NumFreeOpsHash.Hash] = std::make_pair(1, Lane); | |||
1931 | } else if (NumFreeOpsHash.NumOfAPOs == Min && | |||
1932 | NumFreeOpsHash.NumOpsWithSameOpcodeParent < SameOpNumber) { | |||
1933 | // Select the most optimal lane in terms of number of operands that | |||
1934 | // should be moved around. | |||
1935 | SameOpNumber = NumFreeOpsHash.NumOpsWithSameOpcodeParent; | |||
1936 | HashMap[NumFreeOpsHash.Hash] = std::make_pair(1, Lane); | |||
1937 | } else if (NumFreeOpsHash.NumOfAPOs == Min && | |||
1938 | NumFreeOpsHash.NumOpsWithSameOpcodeParent == SameOpNumber) { | |||
1939 | auto It = HashMap.find(NumFreeOpsHash.Hash); | |||
1940 | if (It == HashMap.end()) | |||
1941 | HashMap[NumFreeOpsHash.Hash] = std::make_pair(1, Lane); | |||
1942 | else | |||
1943 | ++It->second.first; | |||
1944 | } | |||
1945 | } | |||
1946 | // Select the lane with the minimum counter. | |||
1947 | unsigned BestLane = 0; | |||
1948 | unsigned CntMin = UINT_MAX(2147483647 *2U +1U); | |||
1949 | for (const auto &Data : reverse(HashMap)) { | |||
1950 | if (Data.second.first < CntMin) { | |||
1951 | CntMin = Data.second.first; | |||
1952 | BestLane = Data.second.second; | |||
1953 | } | |||
1954 | } | |||
1955 | return BestLane; | |||
1956 | } | |||
1957 | ||||
1958 | /// Data structure that helps to reorder operands. | |||
1959 | struct OperandsOrderData { | |||
1960 | /// The best number of operands with the same APOs, which can be | |||
1961 | /// reordered. | |||
1962 | unsigned NumOfAPOs = UINT_MAX(2147483647 *2U +1U); | |||
1963 | /// Number of operands with the same/alternate instruction opcode and | |||
1964 | /// parent. | |||
1965 | unsigned NumOpsWithSameOpcodeParent = 0; | |||
1966 | /// Hash for the actual operands ordering. | |||
1967 | /// Used to count operands, actually their position id and opcode | |||
1968 | /// value. It is used in the voting mechanism to find the lane with the | |||
1969 | /// least number of operands that can freely move about or less profitable | |||
1970 | /// because it already has the most optimal set of operands. Can be | |||
1971 | /// replaced with SmallVector<unsigned> instead but hash code is faster | |||
1972 | /// and requires less memory. | |||
1973 | unsigned Hash = 0; | |||
1974 | }; | |||
1975 | /// \returns the maximum number of operands that are allowed to be reordered | |||
1976 | /// for \p Lane and the number of compatible instructions(with the same | |||
1977 | /// parent/opcode). This is used as a heuristic for selecting the first lane | |||
1978 | /// to start operand reordering. | |||
1979 | OperandsOrderData getMaxNumOperandsThatCanBeReordered(unsigned Lane) const { | |||
1980 | unsigned CntTrue = 0; | |||
1981 | unsigned NumOperands = getNumOperands(); | |||
1982 | // Operands with the same APO can be reordered. We therefore need to count | |||
1983 | // how many of them we have for each APO, like this: Cnt[APO] = x. | |||
1984 | // Since we only have two APOs, namely true and false, we can avoid using | |||
1985 | // a map. Instead we can simply count the number of operands that | |||
1986 | // correspond to one of them (in this case the 'true' APO), and calculate | |||
1987 | // the other by subtracting it from the total number of operands. | |||
1988 | // Operands with the same instruction opcode and parent are more | |||
1989 | // profitable since we don't need to move them in many cases, with a high | |||
1990 | // probability such lane already can be vectorized effectively. | |||
1991 | bool AllUndefs = true; | |||
1992 | unsigned NumOpsWithSameOpcodeParent = 0; | |||
1993 | Instruction *OpcodeI = nullptr; | |||
1994 | BasicBlock *Parent = nullptr; | |||
1995 | unsigned Hash = 0; | |||
1996 | for (unsigned OpIdx = 0; OpIdx != NumOperands; ++OpIdx) { | |||
1997 | const OperandData &OpData = getData(OpIdx, Lane); | |||
1998 | if (OpData.APO) | |||
1999 | ++CntTrue; | |||
2000 | // Use Boyer-Moore majority voting for finding the majority opcode and | |||
2001 | // the number of times it occurs. | |||
2002 | if (auto *I = dyn_cast<Instruction>(OpData.V)) { | |||
2003 | if (!OpcodeI || !getSameOpcode({OpcodeI, I}, TLI).getOpcode() || | |||
2004 | I->getParent() != Parent) { | |||
2005 | if (NumOpsWithSameOpcodeParent == 0) { | |||
2006 | NumOpsWithSameOpcodeParent = 1; | |||
2007 | OpcodeI = I; | |||
2008 | Parent = I->getParent(); | |||
2009 | } else { | |||
2010 | --NumOpsWithSameOpcodeParent; | |||
2011 | } | |||
2012 | } else { | |||
2013 | ++NumOpsWithSameOpcodeParent; | |||
2014 | } | |||
2015 | } | |||
2016 | Hash = hash_combine( | |||
2017 | Hash, hash_value((OpIdx + 1) * (OpData.V->getValueID() + 1))); | |||
2018 | AllUndefs = AllUndefs && isa<UndefValue>(OpData.V); | |||
2019 | } | |||
2020 | if (AllUndefs) | |||
2021 | return {}; | |||
2022 | OperandsOrderData Data; | |||
2023 | Data.NumOfAPOs = std::max(CntTrue, NumOperands - CntTrue); | |||
2024 | Data.NumOpsWithSameOpcodeParent = NumOpsWithSameOpcodeParent; | |||
2025 | Data.Hash = Hash; | |||
2026 | return Data; | |||
2027 | } | |||
2028 | ||||
2029 | /// Go through the instructions in VL and append their operands. | |||
2030 | void appendOperandsOfVL(ArrayRef<Value *> VL) { | |||
2031 | 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", 2031, __extension__ __PRETTY_FUNCTION__)); | |||
2032 | 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", 2033, __extension__ __PRETTY_FUNCTION__)) | |||
2033 | "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", 2033, __extension__ __PRETTY_FUNCTION__)); | |||
2034 | 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", 2034, __extension__ __PRETTY_FUNCTION__)); | |||
2035 | unsigned NumOperands = cast<Instruction>(VL[0])->getNumOperands(); | |||
2036 | OpsVec.resize(NumOperands); | |||
2037 | unsigned NumLanes = VL.size(); | |||
2038 | for (unsigned OpIdx = 0; OpIdx != NumOperands; ++OpIdx) { | |||
2039 | OpsVec[OpIdx].resize(NumLanes); | |||
2040 | for (unsigned Lane = 0; Lane != NumLanes; ++Lane) { | |||
2041 | 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", 2041, __extension__ __PRETTY_FUNCTION__)); | |||
2042 | // Our tree has just 3 nodes: the root and two operands. | |||
2043 | // It is therefore trivial to get the APO. We only need to check the | |||
2044 | // opcode of VL[Lane] and whether the operand at OpIdx is the LHS or | |||
2045 | // RHS operand. The LHS operand of both add and sub is never attached | |||
2046 | // to an inversese operation in the linearized form, therefore its APO | |||
2047 | // is false. The RHS is true only if VL[Lane] is an inverse operation. | |||
2048 | ||||
2049 | // Since operand reordering is performed on groups of commutative | |||
2050 | // operations or alternating sequences (e.g., +, -), we can safely | |||
2051 | // tell the inverse operations by checking commutativity. | |||
2052 | bool IsInverseOperation = !isCommutative(cast<Instruction>(VL[Lane])); | |||
2053 | bool APO = (OpIdx == 0) ? false : IsInverseOperation; | |||
2054 | OpsVec[OpIdx][Lane] = {cast<Instruction>(VL[Lane])->getOperand(OpIdx), | |||
2055 | APO, false}; | |||
2056 | } | |||
2057 | } | |||
2058 | } | |||
2059 | ||||
2060 | /// \returns the number of operands. | |||
2061 | unsigned getNumOperands() const { return OpsVec.size(); } | |||
2062 | ||||
2063 | /// \returns the number of lanes. | |||
2064 | unsigned getNumLanes() const { return OpsVec[0].size(); } | |||
2065 | ||||
2066 | /// \returns the operand value at \p OpIdx and \p Lane. | |||
2067 | Value *getValue(unsigned OpIdx, unsigned Lane) const { | |||
2068 | return getData(OpIdx, Lane).V; | |||
2069 | } | |||
2070 | ||||
2071 | /// \returns true if the data structure is empty. | |||
2072 | bool empty() const { return OpsVec.empty(); } | |||
2073 | ||||
2074 | /// Clears the data. | |||
2075 | void clear() { OpsVec.clear(); } | |||
2076 | ||||
2077 | /// \Returns true if there are enough operands identical to \p Op to fill | |||
2078 | /// the whole vector. | |||
2079 | /// Note: This modifies the 'IsUsed' flag, so a cleanUsed() must follow. | |||
2080 | bool shouldBroadcast(Value *Op, unsigned OpIdx, unsigned Lane) { | |||
2081 | bool OpAPO = getData(OpIdx, Lane).APO; | |||
2082 | for (unsigned Ln = 0, Lns = getNumLanes(); Ln != Lns; ++Ln) { | |||
2083 | if (Ln == Lane) | |||
2084 | continue; | |||
2085 | // This is set to true if we found a candidate for broadcast at Lane. | |||
2086 | bool FoundCandidate = false; | |||
2087 | for (unsigned OpI = 0, OpE = getNumOperands(); OpI != OpE; ++OpI) { | |||
2088 | OperandData &Data = getData(OpI, Ln); | |||
2089 | if (Data.APO != OpAPO || Data.IsUsed) | |||
2090 | continue; | |||
2091 | if (Data.V == Op) { | |||
2092 | FoundCandidate = true; | |||
2093 | Data.IsUsed = true; | |||
2094 | break; | |||
2095 | } | |||
2096 | } | |||
2097 | if (!FoundCandidate) | |||
2098 | return false; | |||
2099 | } | |||
2100 | return true; | |||
2101 | } | |||
2102 | ||||
2103 | public: | |||
2104 | /// Initialize with all the operands of the instruction vector \p RootVL. | |||
2105 | VLOperands(ArrayRef<Value *> RootVL, const TargetLibraryInfo &TLI, | |||
2106 | const DataLayout &DL, ScalarEvolution &SE, const BoUpSLP &R) | |||
2107 | : TLI(TLI), DL(DL), SE(SE), R(R) { | |||
2108 | // Append all the operands of RootVL. | |||
2109 | appendOperandsOfVL(RootVL); | |||
2110 | } | |||
2111 | ||||
2112 | /// \Returns a value vector with the operands across all lanes for the | |||
2113 | /// opearnd at \p OpIdx. | |||
2114 | ValueList getVL(unsigned OpIdx) const { | |||
2115 | ValueList OpVL(OpsVec[OpIdx].size()); | |||
2116 | 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", 2117, __extension__ __PRETTY_FUNCTION__)) | |||
2117 | "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", 2117, __extension__ __PRETTY_FUNCTION__)); | |||
2118 | for (unsigned Lane = 0, Lanes = getNumLanes(); Lane != Lanes; ++Lane) | |||
2119 | OpVL[Lane] = OpsVec[OpIdx][Lane].V; | |||
2120 | return OpVL; | |||
2121 | } | |||
2122 | ||||
2123 | // Performs operand reordering for 2 or more operands. | |||
2124 | // The original operands are in OrigOps[OpIdx][Lane]. | |||
2125 | // The reordered operands are returned in 'SortedOps[OpIdx][Lane]'. | |||
2126 | void reorder() { | |||
2127 | unsigned NumOperands = getNumOperands(); | |||
2128 | unsigned NumLanes = getNumLanes(); | |||
2129 | // Each operand has its own mode. We are using this mode to help us select | |||
2130 | // the instructions for each lane, so that they match best with the ones | |||
2131 | // we have selected so far. | |||
2132 | SmallVector<ReorderingMode, 2> ReorderingModes(NumOperands); | |||
2133 | ||||
2134 | // This is a greedy single-pass algorithm. We are going over each lane | |||
2135 | // once and deciding on the best order right away with no back-tracking. | |||
2136 | // However, in order to increase its effectiveness, we start with the lane | |||
2137 | // that has operands that can move the least. For example, given the | |||
2138 | // following lanes: | |||
2139 | // Lane 0 : A[0] = B[0] + C[0] // Visited 3rd | |||
2140 | // Lane 1 : A[1] = C[1] - B[1] // Visited 1st | |||
2141 | // Lane 2 : A[2] = B[2] + C[2] // Visited 2nd | |||
2142 | // Lane 3 : A[3] = C[3] - B[3] // Visited 4th | |||
2143 | // we will start at Lane 1, since the operands of the subtraction cannot | |||
2144 | // be reordered. Then we will visit the rest of the lanes in a circular | |||
2145 | // fashion. That is, Lanes 2, then Lane 0, and finally Lane 3. | |||
2146 | ||||
2147 | // Find the first lane that we will start our search from. | |||
2148 | unsigned FirstLane = getBestLaneToStartReordering(); | |||
2149 | ||||
2150 | // Initialize the modes. | |||
2151 | for (unsigned OpIdx = 0; OpIdx != NumOperands; ++OpIdx) { | |||
2152 | Value *OpLane0 = getValue(OpIdx, FirstLane); | |||
2153 | // Keep track if we have instructions with all the same opcode on one | |||
2154 | // side. | |||
2155 | if (isa<LoadInst>(OpLane0)) | |||
2156 | ReorderingModes[OpIdx] = ReorderingMode::Load; | |||
2157 | else if (isa<Instruction>(OpLane0)) { | |||
2158 | // Check if OpLane0 should be broadcast. | |||
2159 | if (shouldBroadcast(OpLane0, OpIdx, FirstLane)) | |||
2160 | ReorderingModes[OpIdx] = ReorderingMode::Splat; | |||
2161 | else | |||
2162 | ReorderingModes[OpIdx] = ReorderingMode::Opcode; | |||
2163 | } | |||
2164 | else if (isa<Constant>(OpLane0)) | |||
2165 | ReorderingModes[OpIdx] = ReorderingMode::Constant; | |||
2166 | else if (isa<Argument>(OpLane0)) | |||
2167 | // Our best hope is a Splat. It may save some cost in some cases. | |||
2168 | ReorderingModes[OpIdx] = ReorderingMode::Splat; | |||
2169 | else | |||
2170 | // NOTE: This should be unreachable. | |||
2171 | ReorderingModes[OpIdx] = ReorderingMode::Failed; | |||
2172 | } | |||
2173 | ||||
2174 | // Check that we don't have same operands. No need to reorder if operands | |||
2175 | // are just perfect diamond or shuffled diamond match. Do not do it only | |||
2176 | // for possible broadcasts or non-power of 2 number of scalars (just for | |||
2177 | // now). | |||
2178 | auto &&SkipReordering = [this]() { | |||
2179 | SmallPtrSet<Value *, 4> UniqueValues; | |||
2180 | ArrayRef<OperandData> Op0 = OpsVec.front(); | |||
2181 | for (const OperandData &Data : Op0) | |||
2182 | UniqueValues.insert(Data.V); | |||
2183 | for (ArrayRef<OperandData> Op : drop_begin(OpsVec, 1)) { | |||
2184 | if (any_of(Op, [&UniqueValues](const OperandData &Data) { | |||
2185 | return !UniqueValues.contains(Data.V); | |||
2186 | })) | |||
2187 | return false; | |||
2188 | } | |||
2189 | // TODO: Check if we can remove a check for non-power-2 number of | |||
2190 | // scalars after full support of non-power-2 vectorization. | |||
2191 | return UniqueValues.size() != 2 && isPowerOf2_32(UniqueValues.size()); | |||
2192 | }; | |||
2193 | ||||
2194 | // If the initial strategy fails for any of the operand indexes, then we | |||
2195 | // perform reordering again in a second pass. This helps avoid assigning | |||
2196 | // high priority to the failed strategy, and should improve reordering for | |||
2197 | // the non-failed operand indexes. | |||
2198 | for (int Pass = 0; Pass != 2; ++Pass) { | |||
2199 | // Check if no need to reorder operands since they're are perfect or | |||
2200 | // shuffled diamond match. | |||
2201 | // Need to to do it to avoid extra external use cost counting for | |||
2202 | // shuffled matches, which may cause regressions. | |||
2203 | if (SkipReordering()) | |||
2204 | break; | |||
2205 | // Skip the second pass if the first pass did not fail. | |||
2206 | bool StrategyFailed = false; | |||
2207 | // Mark all operand data as free to use. | |||
2208 | clearUsed(); | |||
2209 | // We keep the original operand order for the FirstLane, so reorder the | |||
2210 | // rest of the lanes. We are visiting the nodes in a circular fashion, | |||
2211 | // using FirstLane as the center point and increasing the radius | |||
2212 | // distance. | |||
2213 | SmallVector<SmallVector<Value *, 2>> MainAltOps(NumOperands); | |||
2214 | for (unsigned I = 0; I < NumOperands; ++I) | |||
2215 | MainAltOps[I].push_back(getData(I, FirstLane).V); | |||
2216 | ||||
2217 | for (unsigned Distance = 1; Distance != NumLanes; ++Distance) { | |||
2218 | // Visit the lane on the right and then the lane on the left. | |||
2219 | for (int Direction : {+1, -1}) { | |||
2220 | int Lane = FirstLane + Direction * Distance; | |||
2221 | if (Lane < 0 || Lane >= (int)NumLanes) | |||
2222 | continue; | |||
2223 | int LastLane = Lane - Direction; | |||
2224 | 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", 2225, __extension__ __PRETTY_FUNCTION__)) | |||
2225 | "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", 2225, __extension__ __PRETTY_FUNCTION__)); | |||
2226 | // Look for a good match for each operand. | |||
2227 | for (unsigned OpIdx = 0; OpIdx != NumOperands; ++OpIdx) { | |||
2228 | // Search for the operand that matches SortedOps[OpIdx][Lane-1]. | |||
2229 | std::optional<unsigned> BestIdx = getBestOperand( | |||
2230 | OpIdx, Lane, LastLane, ReorderingModes, MainAltOps[OpIdx]); | |||
2231 | // By not selecting a value, we allow the operands that follow to | |||
2232 | // select a better matching value. We will get a non-null value in | |||
2233 | // the next run of getBestOperand(). | |||
2234 | if (BestIdx) { | |||
2235 | // Swap the current operand with the one returned by | |||
2236 | // getBestOperand(). | |||
2237 | swap(OpIdx, *BestIdx, Lane); | |||
2238 | } else { | |||
2239 | // We failed to find a best operand, set mode to 'Failed'. | |||
2240 | ReorderingModes[OpIdx] = ReorderingMode::Failed; | |||
2241 | // Enable the second pass. | |||
2242 | StrategyFailed = true; | |||
2243 | } | |||
2244 | // Try to get the alternate opcode and follow it during analysis. | |||
2245 | if (MainAltOps[OpIdx].size() != 2) { | |||
2246 | OperandData &AltOp = getData(OpIdx, Lane); | |||
2247 | InstructionsState OpS = | |||
2248 | getSameOpcode({MainAltOps[OpIdx].front(), AltOp.V}, TLI); | |||
2249 | if (OpS.getOpcode() && OpS.isAltShuffle()) | |||
2250 | MainAltOps[OpIdx].push_back(AltOp.V); | |||
2251 | } | |||
2252 | } | |||
2253 | } | |||
2254 | } | |||
2255 | // Skip second pass if the strategy did not fail. | |||
2256 | if (!StrategyFailed) | |||
2257 | break; | |||
2258 | } | |||
2259 | } | |||
2260 | ||||
2261 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | |||
2262 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) static StringRef getModeStr(ReorderingMode RMode) { | |||
2263 | switch (RMode) { | |||
2264 | case ReorderingMode::Load: | |||
2265 | return "Load"; | |||
2266 | case ReorderingMode::Opcode: | |||
2267 | return "Opcode"; | |||
2268 | case ReorderingMode::Constant: | |||
2269 | return "Constant"; | |||
2270 | case ReorderingMode::Splat: | |||
2271 | return "Splat"; | |||
2272 | case ReorderingMode::Failed: | |||
2273 | return "Failed"; | |||
2274 | } | |||
2275 | llvm_unreachable("Unimplemented Reordering Type")::llvm::llvm_unreachable_internal("Unimplemented Reordering Type" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2275); | |||
2276 | } | |||
2277 | ||||
2278 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) static raw_ostream &printMode(ReorderingMode RMode, | |||
2279 | raw_ostream &OS) { | |||
2280 | return OS << getModeStr(RMode); | |||
2281 | } | |||
2282 | ||||
2283 | /// Debug print. | |||
2284 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) static void dumpMode(ReorderingMode RMode) { | |||
2285 | printMode(RMode, dbgs()); | |||
2286 | } | |||
2287 | ||||
2288 | friend raw_ostream &operator<<(raw_ostream &OS, ReorderingMode RMode) { | |||
2289 | return printMode(RMode, OS); | |||
2290 | } | |||
2291 | ||||
2292 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) raw_ostream &print(raw_ostream &OS) const { | |||
2293 | const unsigned Indent = 2; | |||
2294 | unsigned Cnt = 0; | |||
2295 | for (const OperandDataVec &OpDataVec : OpsVec) { | |||
2296 | OS << "Operand " << Cnt++ << "\n"; | |||
2297 | for (const OperandData &OpData : OpDataVec) { | |||
2298 | OS.indent(Indent) << "{"; | |||
2299 | if (Value *V = OpData.V) | |||
2300 | OS << *V; | |||
2301 | else | |||
2302 | OS << "null"; | |||
2303 | OS << ", APO:" << OpData.APO << "}\n"; | |||
2304 | } | |||
2305 | OS << "\n"; | |||
2306 | } | |||
2307 | return OS; | |||
2308 | } | |||
2309 | ||||
2310 | /// Debug print. | |||
2311 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void dump() const { print(dbgs()); } | |||
2312 | #endif | |||
2313 | }; | |||
2314 | ||||
2315 | /// Evaluate each pair in \p Candidates and return index into \p Candidates | |||
2316 | /// for a pair which have highest score deemed to have best chance to form | |||
2317 | /// root of profitable tree to vectorize. Return std::nullopt if no candidate | |||
2318 | /// scored above the LookAheadHeuristics::ScoreFail. \param Limit Lower limit | |||
2319 | /// of the cost, considered to be good enough score. | |||
2320 | std::optional<int> | |||
2321 | findBestRootPair(ArrayRef<std::pair<Value *, Value *>> Candidates, | |||
2322 | int Limit = LookAheadHeuristics::ScoreFail) { | |||
2323 | LookAheadHeuristics LookAhead(*TLI, *DL, *SE, *this, /*NumLanes=*/2, | |||
2324 | RootLookAheadMaxDepth); | |||
2325 | int BestScore = Limit; | |||
2326 | std::optional<int> Index; | |||
2327 | for (int I : seq<int>(0, Candidates.size())) { | |||
2328 | int Score = LookAhead.getScoreAtLevelRec(Candidates[I].first, | |||
2329 | Candidates[I].second, | |||
2330 | /*U1=*/nullptr, /*U2=*/nullptr, | |||
2331 | /*Level=*/1, std::nullopt); | |||
2332 | if (Score > BestScore) { | |||
2333 | BestScore = Score; | |||
2334 | Index = I; | |||
2335 | } | |||
2336 | } | |||
2337 | return Index; | |||
2338 | } | |||
2339 | ||||
2340 | /// Checks if the instruction is marked for deletion. | |||
2341 | bool isDeleted(Instruction *I) const { return DeletedInstructions.count(I); } | |||
2342 | ||||
2343 | /// Removes an instruction from its block and eventually deletes it. | |||
2344 | /// It's like Instruction::eraseFromParent() except that the actual deletion | |||
2345 | /// is delayed until BoUpSLP is destructed. | |||
2346 | void eraseInstruction(Instruction *I) { | |||
2347 | DeletedInstructions.insert(I); | |||
2348 | } | |||
2349 | ||||
2350 | /// Checks if the instruction was already analyzed for being possible | |||
2351 | /// reduction root. | |||
2352 | bool isAnalyzedReductionRoot(Instruction *I) const { | |||
2353 | return AnalyzedReductionsRoots.count(I); | |||
2354 | } | |||
2355 | /// Register given instruction as already analyzed for being possible | |||
2356 | /// reduction root. | |||
2357 | void analyzedReductionRoot(Instruction *I) { | |||
2358 | AnalyzedReductionsRoots.insert(I); | |||
2359 | } | |||
2360 | /// Checks if the provided list of reduced values was checked already for | |||
2361 | /// vectorization. | |||
2362 | bool areAnalyzedReductionVals(ArrayRef<Value *> VL) const { | |||
2363 | return AnalyzedReductionVals.contains(hash_value(VL)); | |||
2364 | } | |||
2365 | /// Adds the list of reduced values to list of already checked values for the | |||
2366 | /// vectorization. | |||
2367 | void analyzedReductionVals(ArrayRef<Value *> VL) { | |||
2368 | AnalyzedReductionVals.insert(hash_value(VL)); | |||
2369 | } | |||
2370 | /// Clear the list of the analyzed reduction root instructions. | |||
2371 | void clearReductionData() { | |||
2372 | AnalyzedReductionsRoots.clear(); | |||
2373 | AnalyzedReductionVals.clear(); | |||
2374 | } | |||
2375 | /// Checks if the given value is gathered in one of the nodes. | |||
2376 | bool isAnyGathered(const SmallDenseSet<Value *> &Vals) const { | |||
2377 | return any_of(MustGather, [&](Value *V) { return Vals.contains(V); }); | |||
2378 | } | |||
2379 | ||||
2380 | /// Check if the value is vectorized in the tree. | |||
2381 | bool isVectorized(Value *V) const { return getTreeEntry(V); } | |||
2382 | ||||
2383 | ~BoUpSLP(); | |||
2384 | ||||
2385 | private: | |||
2386 | /// Check if the operands on the edges \p Edges of the \p UserTE allows | |||
2387 | /// reordering (i.e. the operands can be reordered because they have only one | |||
2388 | /// user and reordarable). | |||
2389 | /// \param ReorderableGathers List of all gather nodes that require reordering | |||
2390 | /// (e.g., gather of extractlements or partially vectorizable loads). | |||
2391 | /// \param GatherOps List of gather operand nodes for \p UserTE that require | |||
2392 | /// reordering, subset of \p NonVectorized. | |||
2393 | bool | |||
2394 | canReorderOperands(TreeEntry *UserTE, | |||
2395 | SmallVectorImpl<std::pair<unsigned, TreeEntry *>> &Edges, | |||
2396 | ArrayRef<TreeEntry *> ReorderableGathers, | |||
2397 | SmallVectorImpl<TreeEntry *> &GatherOps); | |||
2398 | ||||
2399 | /// Checks if the given \p TE is a gather node with clustered reused scalars | |||
2400 | /// and reorders it per given \p Mask. | |||
2401 | void reorderNodeWithReuses(TreeEntry &TE, ArrayRef<int> Mask) const; | |||
2402 | ||||
2403 | /// Returns vectorized operand \p OpIdx of the node \p UserTE from the graph, | |||
2404 | /// if any. If it is not vectorized (gather node), returns nullptr. | |||
2405 | TreeEntry *getVectorizedOperand(TreeEntry *UserTE, unsigned OpIdx) { | |||
2406 | ArrayRef<Value *> VL = UserTE->getOperand(OpIdx); | |||
2407 | TreeEntry *TE = nullptr; | |||
2408 | const auto *It = find_if(VL, [this, &TE](Value *V) { | |||
2409 | TE = getTreeEntry(V); | |||
2410 | return TE; | |||
2411 | }); | |||
2412 | if (It != VL.end() && TE->isSame(VL)) | |||
2413 | return TE; | |||
2414 | return nullptr; | |||
2415 | } | |||
2416 | ||||
2417 | /// Returns vectorized operand \p OpIdx of the node \p UserTE from the graph, | |||
2418 | /// if any. If it is not vectorized (gather node), returns nullptr. | |||
2419 | const TreeEntry *getVectorizedOperand(const TreeEntry *UserTE, | |||
2420 | unsigned OpIdx) const { | |||
2421 | return const_cast<BoUpSLP *>(this)->getVectorizedOperand( | |||
2422 | const_cast<TreeEntry *>(UserTE), OpIdx); | |||
2423 | } | |||
2424 | ||||
2425 | /// Checks if all users of \p I are the part of the vectorization tree. | |||
2426 | bool areAllUsersVectorized(Instruction *I, | |||
2427 | ArrayRef<Value *> VectorizedVals) const; | |||
2428 | ||||
2429 | /// Return information about the vector formed for the specified index | |||
2430 | /// of a vector of (the same) instruction. | |||
2431 | TargetTransformInfo::OperandValueInfo getOperandInfo(ArrayRef<Value *> VL, | |||
2432 | unsigned OpIdx); | |||
2433 | ||||
2434 | /// \returns the cost of the vectorizable entry. | |||
2435 | InstructionCost getEntryCost(const TreeEntry *E, | |||
2436 | ArrayRef<Value *> VectorizedVals, | |||
2437 | SmallPtrSetImpl<Value *> &CheckedExtracts); | |||
2438 | ||||
2439 | /// This is the recursive part of buildTree. | |||
2440 | void buildTree_rec(ArrayRef<Value *> Roots, unsigned Depth, | |||
2441 | const EdgeInfo &EI); | |||
2442 | ||||
2443 | /// \returns true if the ExtractElement/ExtractValue instructions in \p VL can | |||
2444 | /// be vectorized to use the original vector (or aggregate "bitcast" to a | |||
2445 | /// vector) and sets \p CurrentOrder to the identity permutation; otherwise | |||
2446 | /// returns false, setting \p CurrentOrder to either an empty vector or a | |||
2447 | /// non-identity permutation that allows to reuse extract instructions. | |||
2448 | bool canReuseExtract(ArrayRef<Value *> VL, Value *OpValue, | |||
2449 | SmallVectorImpl<unsigned> &CurrentOrder) const; | |||
2450 | ||||
2451 | /// Vectorize a single entry in the tree. | |||
2452 | Value *vectorizeTree(TreeEntry *E); | |||
2453 | ||||
2454 | /// Vectorize a single entry in the tree, the \p Idx-th operand of the entry | |||
2455 | /// \p E. | |||
2456 | Value *vectorizeOperand(TreeEntry *E, unsigned NodeIdx); | |||
2457 | ||||
2458 | /// Create a new vector from a list of scalar values. Produces a sequence | |||
2459 | /// which exploits values reused across lanes, and arranges the inserts | |||
2460 | /// for ease of later optimization. | |||
2461 | template <typename BVTy, typename ResTy, typename... Args> | |||
2462 | ResTy processBuildVector(const TreeEntry *E, Args &...Params); | |||
2463 | ||||
2464 | /// Create a new vector from a list of scalar values. Produces a sequence | |||
2465 | /// which exploits values reused across lanes, and arranges the inserts | |||
2466 | /// for ease of later optimization. | |||
2467 | Value *createBuildVector(const TreeEntry *E); | |||
2468 | ||||
2469 | /// Returns the instruction in the bundle, which can be used as a base point | |||
2470 | /// for scheduling. Usually it is the last instruction in the bundle, except | |||
2471 | /// for the case when all operands are external (in this case, it is the first | |||
2472 | /// instruction in the list). | |||
2473 | Instruction &getLastInstructionInBundle(const TreeEntry *E); | |||
2474 | ||||
2475 | /// Checks if the gathered \p VL can be represented as shuffle(s) of previous | |||
2476 | /// tree entries. | |||
2477 | /// \param TE Tree entry checked for permutation. | |||
2478 | /// \param VL List of scalars (a subset of the TE scalar), checked for | |||
2479 | /// permutations. | |||
2480 | /// \returns ShuffleKind, if gathered values can be represented as shuffles of | |||
2481 | /// previous tree entries. \p Mask is filled with the shuffle mask. | |||
2482 | std::optional<TargetTransformInfo::ShuffleKind> | |||
2483 | isGatherShuffledEntry(const TreeEntry *TE, ArrayRef<Value *> VL, | |||
2484 | SmallVectorImpl<int> &Mask, | |||
2485 | SmallVectorImpl<const TreeEntry *> &Entries); | |||
2486 | ||||
2487 | /// \returns the scalarization cost for this list of values. Assuming that | |||
2488 | /// this subtree gets vectorized, we may need to extract the values from the | |||
2489 | /// roots. This method calculates the cost of extracting the values. | |||
2490 | /// \param ForPoisonSrc true if initial vector is poison, false otherwise. | |||
2491 | InstructionCost getGatherCost(ArrayRef<Value *> VL, bool ForPoisonSrc) const; | |||
2492 | ||||
2493 | /// Set the Builder insert point to one after the last instruction in | |||
2494 | /// the bundle | |||
2495 | void setInsertPointAfterBundle(const TreeEntry *E); | |||
2496 | ||||
2497 | /// \returns a vector from a collection of scalars in \p VL. if \p Root is not | |||
2498 | /// specified, the starting vector value is poison. | |||
2499 | Value *gather(ArrayRef<Value *> VL, Value *Root); | |||
2500 | ||||
2501 | /// \returns whether the VectorizableTree is fully vectorizable and will | |||
2502 | /// be beneficial even the tree height is tiny. | |||
2503 | bool isFullyVectorizableTinyTree(bool ForReduction) const; | |||
2504 | ||||
2505 | /// Reorder commutative or alt operands to get better probability of | |||
2506 | /// generating vectorized code. | |||
2507 | static void reorderInputsAccordingToOpcode( | |||
2508 | ArrayRef<Value *> VL, SmallVectorImpl<Value *> &Left, | |||
2509 | SmallVectorImpl<Value *> &Right, const TargetLibraryInfo &TLI, | |||
2510 | const DataLayout &DL, ScalarEvolution &SE, const BoUpSLP &R); | |||
2511 | ||||
2512 | /// Helper for `findExternalStoreUsersReorderIndices()`. It iterates over the | |||
2513 | /// users of \p TE and collects the stores. It returns the map from the store | |||
2514 | /// pointers to the collected stores. | |||
2515 | DenseMap<Value *, SmallVector<StoreInst *, 4>> | |||
2516 | collectUserStores(const BoUpSLP::TreeEntry *TE) const; | |||
2517 | ||||
2518 | /// Helper for `findExternalStoreUsersReorderIndices()`. It checks if the | |||
2519 | /// stores in \p StoresVec can form a vector instruction. If so it returns true | |||
2520 | /// and populates \p ReorderIndices with the shuffle indices of the the stores | |||
2521 | /// when compared to the sorted vector. | |||
2522 | bool canFormVector(const SmallVector<StoreInst *, 4> &StoresVec, | |||
2523 | OrdersType &ReorderIndices) const; | |||
2524 | ||||
2525 | /// Iterates through the users of \p TE, looking for scalar stores that can be | |||
2526 | /// potentially vectorized in a future SLP-tree. If found, it keeps track of | |||
2527 | /// their order and builds an order index vector for each store bundle. It | |||
2528 | /// returns all these order vectors found. | |||
2529 | /// We run this after the tree has formed, otherwise we may come across user | |||
2530 | /// instructions that are not yet in the tree. | |||
2531 | SmallVector<OrdersType, 1> | |||
2532 | findExternalStoreUsersReorderIndices(TreeEntry *TE) const; | |||
2533 | ||||
2534 | struct TreeEntry { | |||
2535 | using VecTreeTy = SmallVector<std::unique_ptr<TreeEntry>, 8>; | |||
2536 | TreeEntry(VecTreeTy &Container) : Container(Container) {} | |||
2537 | ||||
2538 | /// \returns true if the scalars in VL are equal to this entry. | |||
2539 | bool isSame(ArrayRef<Value *> VL) const { | |||
2540 | auto &&IsSame = [VL](ArrayRef<Value *> Scalars, ArrayRef<int> Mask) { | |||
2541 | if (Mask.size() != VL.size() && VL.size() == Scalars.size()) | |||
2542 | return std::equal(VL.begin(), VL.end(), Scalars.begin()); | |||
2543 | return VL.size() == Mask.size() && | |||
2544 | std::equal(VL.begin(), VL.end(), Mask.begin(), | |||
2545 | [Scalars](Value *V, int Idx) { | |||
2546 | return (isa<UndefValue>(V) && | |||
2547 | Idx == PoisonMaskElem) || | |||
2548 | (Idx != PoisonMaskElem && V == Scalars[Idx]); | |||
2549 | }); | |||
2550 | }; | |||
2551 | if (!ReorderIndices.empty()) { | |||
2552 | // TODO: implement matching if the nodes are just reordered, still can | |||
2553 | // treat the vector as the same if the list of scalars matches VL | |||
2554 | // directly, without reordering. | |||
2555 | SmallVector<int> Mask; | |||
2556 | inversePermutation(ReorderIndices, Mask); | |||
2557 | if (VL.size() == Scalars.size()) | |||
2558 | return IsSame(Scalars, Mask); | |||
2559 | if (VL.size() == ReuseShuffleIndices.size()) { | |||
2560 | ::addMask(Mask, ReuseShuffleIndices); | |||
2561 | return IsSame(Scalars, Mask); | |||
2562 | } | |||
2563 | return false; | |||
2564 | } | |||
2565 | return IsSame(Scalars, ReuseShuffleIndices); | |||
2566 | } | |||
2567 | ||||
2568 | bool isOperandGatherNode(const EdgeInfo &UserEI) const { | |||
2569 | return State == TreeEntry::NeedToGather && | |||
2570 | UserTreeIndices.front().EdgeIdx == UserEI.EdgeIdx && | |||
2571 | UserTreeIndices.front().UserTE == UserEI.UserTE; | |||
2572 | } | |||
2573 | ||||
2574 | /// \returns true if current entry has same operands as \p TE. | |||
2575 | bool hasEqualOperands(const TreeEntry &TE) const { | |||
2576 | if (TE.getNumOperands() != getNumOperands()) | |||
2577 | return false; | |||
2578 | SmallBitVector Used(getNumOperands()); | |||
2579 | for (unsigned I = 0, E = getNumOperands(); I < E; ++I) { | |||
2580 | unsigned PrevCount = Used.count(); | |||
2581 | for (unsigned K = 0; K < E; ++K) { | |||
2582 | if (Used.test(K)) | |||
2583 | continue; | |||
2584 | if (getOperand(K) == TE.getOperand(I)) { | |||
2585 | Used.set(K); | |||
2586 | break; | |||
2587 | } | |||
2588 | } | |||
2589 | // Check if we actually found the matching operand. | |||
2590 | if (PrevCount == Used.count()) | |||
2591 | return false; | |||
2592 | } | |||
2593 | return true; | |||
2594 | } | |||
2595 | ||||
2596 | /// \return Final vectorization factor for the node. Defined by the total | |||
2597 | /// number of vectorized scalars, including those, used several times in the | |||
2598 | /// entry and counted in the \a ReuseShuffleIndices, if any. | |||
2599 | unsigned getVectorFactor() const { | |||
2600 | if (!ReuseShuffleIndices.empty()) | |||
2601 | return ReuseShuffleIndices.size(); | |||
2602 | return Scalars.size(); | |||
2603 | }; | |||
2604 | ||||
2605 | /// A vector of scalars. | |||
2606 | ValueList Scalars; | |||
2607 | ||||
2608 | /// The Scalars are vectorized into this value. It is initialized to Null. | |||
2609 | WeakTrackingVH VectorizedValue = nullptr; | |||
2610 | ||||
2611 | /// Do we need to gather this sequence or vectorize it | |||
2612 | /// (either with vector instruction or with scatter/gather | |||
2613 | /// intrinsics for store/load)? | |||
2614 | enum EntryState { Vectorize, ScatterVectorize, NeedToGather }; | |||
2615 | EntryState State; | |||
2616 | ||||
2617 | /// Does this sequence require some shuffling? | |||
2618 | SmallVector<int, 4> ReuseShuffleIndices; | |||
2619 | ||||
2620 | /// Does this entry require reordering? | |||
2621 | SmallVector<unsigned, 4> ReorderIndices; | |||
2622 | ||||
2623 | /// Points back to the VectorizableTree. | |||
2624 | /// | |||
2625 | /// Only used for Graphviz right now. Unfortunately GraphTrait::NodeRef has | |||
2626 | /// to be a pointer and needs to be able to initialize the child iterator. | |||
2627 | /// Thus we need a reference back to the container to translate the indices | |||
2628 | /// to entries. | |||
2629 | VecTreeTy &Container; | |||
2630 | ||||
2631 | /// The TreeEntry index containing the user of this entry. We can actually | |||
2632 | /// have multiple users so the data structure is not truly a tree. | |||
2633 | SmallVector<EdgeInfo, 1> UserTreeIndices; | |||
2634 | ||||
2635 | /// The index of this treeEntry in VectorizableTree. | |||
2636 | int Idx = -1; | |||
2637 | ||||
2638 | private: | |||
2639 | /// The operands of each instruction in each lane Operands[op_index][lane]. | |||
2640 | /// Note: This helps avoid the replication of the code that performs the | |||
2641 | /// reordering of operands during buildTree_rec() and vectorizeTree(). | |||
2642 | SmallVector<ValueList, 2> Operands; | |||
2643 | ||||
2644 | /// The main/alternate instruction. | |||
2645 | Instruction *MainOp = nullptr; | |||
2646 | Instruction *AltOp = nullptr; | |||
2647 | ||||
2648 | public: | |||
2649 | /// Set this bundle's \p OpIdx'th operand to \p OpVL. | |||
2650 | void setOperand(unsigned OpIdx, ArrayRef<Value *> OpVL) { | |||
2651 | if (Operands.size() < OpIdx + 1) | |||
2652 | Operands.resize(OpIdx + 1); | |||
2653 | 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", 2653, __extension__ __PRETTY_FUNCTION__)); | |||
2654 | 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", 2655, __extension__ __PRETTY_FUNCTION__)) | |||
2655 | "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", 2655, __extension__ __PRETTY_FUNCTION__)); | |||
2656 | Operands[OpIdx].resize(OpVL.size()); | |||
2657 | copy(OpVL, Operands[OpIdx].begin()); | |||
2658 | } | |||
2659 | ||||
2660 | /// Set the operands of this bundle in their original order. | |||
2661 | void setOperandsInOrder() { | |||
2662 | 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", 2662, __extension__ __PRETTY_FUNCTION__)); | |||
2663 | auto *I0 = cast<Instruction>(Scalars[0]); | |||
2664 | Operands.resize(I0->getNumOperands()); | |||
2665 | unsigned NumLanes = Scalars.size(); | |||
2666 | for (unsigned OpIdx = 0, NumOperands = I0->getNumOperands(); | |||
2667 | OpIdx != NumOperands; ++OpIdx) { | |||
2668 | Operands[OpIdx].resize(NumLanes); | |||
2669 | for (unsigned Lane = 0; Lane != NumLanes; ++Lane) { | |||
2670 | auto *I = cast<Instruction>(Scalars[Lane]); | |||
2671 | 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", 2672, __extension__ __PRETTY_FUNCTION__)) | |||
2672 | "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", 2672, __extension__ __PRETTY_FUNCTION__)); | |||
2673 | Operands[OpIdx][Lane] = I->getOperand(OpIdx); | |||
2674 | } | |||
2675 | } | |||
2676 | } | |||
2677 | ||||
2678 | /// Reorders operands of the node to the given mask \p Mask. | |||
2679 | void reorderOperands(ArrayRef<int> Mask) { | |||
2680 | for (ValueList &Operand : Operands) | |||
2681 | reorderScalars(Operand, Mask); | |||
2682 | } | |||
2683 | ||||
2684 | /// \returns the \p OpIdx operand of this TreeEntry. | |||
2685 | ValueList &getOperand(unsigned OpIdx) { | |||
2686 | 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", 2686, __extension__ __PRETTY_FUNCTION__)); | |||
2687 | return Operands[OpIdx]; | |||
2688 | } | |||
2689 | ||||
2690 | /// \returns the \p OpIdx operand of this TreeEntry. | |||
2691 | ArrayRef<Value *> getOperand(unsigned OpIdx) const { | |||
2692 | 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", 2692, __extension__ __PRETTY_FUNCTION__)); | |||
2693 | return Operands[OpIdx]; | |||
2694 | } | |||
2695 | ||||
2696 | /// \returns the number of operands. | |||
2697 | unsigned getNumOperands() const { return Operands.size(); } | |||
2698 | ||||
2699 | /// \return the single \p OpIdx operand. | |||
2700 | Value *getSingleOperand(unsigned OpIdx) const { | |||
2701 | 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", 2701, __extension__ __PRETTY_FUNCTION__)); | |||
2702 | 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", 2702, __extension__ __PRETTY_FUNCTION__)); | |||
2703 | return Operands[OpIdx][0]; | |||
2704 | } | |||
2705 | ||||
2706 | /// Some of the instructions in the list have alternate opcodes. | |||
2707 | bool isAltShuffle() const { return MainOp != AltOp; } | |||
2708 | ||||
2709 | bool isOpcodeOrAlt(Instruction *I) const { | |||
2710 | unsigned CheckedOpcode = I->getOpcode(); | |||
2711 | return (getOpcode() == CheckedOpcode || | |||
2712 | getAltOpcode() == CheckedOpcode); | |||
2713 | } | |||
2714 | ||||
2715 | /// Chooses the correct key for scheduling data. If \p Op has the same (or | |||
2716 | /// alternate) opcode as \p OpValue, the key is \p Op. Otherwise the key is | |||
2717 | /// \p OpValue. | |||
2718 | Value *isOneOf(Value *Op) const { | |||
2719 | auto *I = dyn_cast<Instruction>(Op); | |||
2720 | if (I && isOpcodeOrAlt(I)) | |||
2721 | return Op; | |||
2722 | return MainOp; | |||
2723 | } | |||
2724 | ||||
2725 | void setOperations(const InstructionsState &S) { | |||
2726 | MainOp = S.MainOp; | |||
2727 | AltOp = S.AltOp; | |||
2728 | } | |||
2729 | ||||
2730 | Instruction *getMainOp() const { | |||
2731 | return MainOp; | |||
2732 | } | |||
2733 | ||||
2734 | Instruction *getAltOp() const { | |||
2735 | return AltOp; | |||
2736 | } | |||
2737 | ||||
2738 | /// The main/alternate opcodes for the list of instructions. | |||
2739 | unsigned getOpcode() const { | |||
2740 | return MainOp ? MainOp->getOpcode() : 0; | |||
2741 | } | |||
2742 | ||||
2743 | unsigned getAltOpcode() const { | |||
2744 | return AltOp ? AltOp->getOpcode() : 0; | |||
2745 | } | |||
2746 | ||||
2747 | /// When ReuseReorderShuffleIndices is empty it just returns position of \p | |||
2748 | /// V within vector of Scalars. Otherwise, try to remap on its reuse index. | |||
2749 | int findLaneForValue(Value *V) const { | |||
2750 | unsigned FoundLane = std::distance(Scalars.begin(), find(Scalars, V)); | |||
2751 | 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", 2751, __extension__ __PRETTY_FUNCTION__)); | |||
2752 | if (!ReorderIndices.empty()) | |||
2753 | FoundLane = ReorderIndices[FoundLane]; | |||
2754 | 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", 2754, __extension__ __PRETTY_FUNCTION__)); | |||
2755 | if (!ReuseShuffleIndices.empty()) { | |||
2756 | FoundLane = std::distance(ReuseShuffleIndices.begin(), | |||
2757 | find(ReuseShuffleIndices, FoundLane)); | |||
2758 | } | |||
2759 | return FoundLane; | |||
2760 | } | |||
2761 | ||||
2762 | #ifndef NDEBUG | |||
2763 | /// Debug printer. | |||
2764 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void dump() const { | |||
2765 | dbgs() << Idx << ".\n"; | |||
2766 | for (unsigned OpI = 0, OpE = Operands.size(); OpI != OpE; ++OpI) { | |||
2767 | dbgs() << "Operand " << OpI << ":\n"; | |||
2768 | for (const Value *V : Operands[OpI]) | |||
2769 | dbgs().indent(2) << *V << "\n"; | |||
2770 | } | |||
2771 | dbgs() << "Scalars: \n"; | |||
2772 | for (Value *V : Scalars) | |||
2773 | dbgs().indent(2) << *V << "\n"; | |||
2774 | dbgs() << "State: "; | |||
2775 | switch (State) { | |||
2776 | case Vectorize: | |||
2777 | dbgs() << "Vectorize\n"; | |||
2778 | break; | |||
2779 | case ScatterVectorize: | |||
2780 | dbgs() << "ScatterVectorize\n"; | |||
2781 | break; | |||
2782 | case NeedToGather: | |||
2783 | dbgs() << "NeedToGather\n"; | |||
2784 | break; | |||
2785 | } | |||
2786 | dbgs() << "MainOp: "; | |||
2787 | if (MainOp) | |||
2788 | dbgs() << *MainOp << "\n"; | |||
2789 | else | |||
2790 | dbgs() << "NULL\n"; | |||
2791 | dbgs() << "AltOp: "; | |||
2792 | if (AltOp) | |||
2793 | dbgs() << *AltOp << "\n"; | |||
2794 | else | |||
2795 | dbgs() << "NULL\n"; | |||
2796 | dbgs() << "VectorizedValue: "; | |||
2797 | if (VectorizedValue) | |||
2798 | dbgs() << *VectorizedValue << "\n"; | |||
2799 | else | |||
2800 | dbgs() << "NULL\n"; | |||
2801 | dbgs() << "ReuseShuffleIndices: "; | |||
2802 | if (ReuseShuffleIndices.empty()) | |||
2803 | dbgs() << "Empty"; | |||
2804 | else | |||
2805 | for (int ReuseIdx : ReuseShuffleIndices) | |||
2806 | dbgs() << ReuseIdx << ", "; | |||
2807 | dbgs() << "\n"; | |||
2808 | dbgs() << "ReorderIndices: "; | |||
2809 | for (unsigned ReorderIdx : ReorderIndices) | |||
2810 | dbgs() << ReorderIdx << ", "; | |||
2811 | dbgs() << "\n"; | |||
2812 | dbgs() << "UserTreeIndices: "; | |||
2813 | for (const auto &EInfo : UserTreeIndices) | |||
2814 | dbgs() << EInfo << ", "; | |||
2815 | dbgs() << "\n"; | |||
2816 | } | |||
2817 | #endif | |||
2818 | }; | |||
2819 | ||||
2820 | #ifndef NDEBUG | |||
2821 | void dumpTreeCosts(const TreeEntry *E, InstructionCost ReuseShuffleCost, | |||
2822 | InstructionCost VecCost, InstructionCost ScalarCost, | |||
2823 | StringRef Banner) const { | |||
2824 | dbgs() << "SLP: " << Banner << ":\n"; | |||
2825 | E->dump(); | |||
2826 | dbgs() << "SLP: Costs:\n"; | |||
2827 | dbgs() << "SLP: ReuseShuffleCost = " << ReuseShuffleCost << "\n"; | |||
2828 | dbgs() << "SLP: VectorCost = " << VecCost << "\n"; | |||
2829 | dbgs() << "SLP: ScalarCost = " << ScalarCost << "\n"; | |||
2830 | dbgs() << "SLP: ReuseShuffleCost + VecCost - ScalarCost = " | |||
2831 | << ReuseShuffleCost + VecCost - ScalarCost << "\n"; | |||
2832 | } | |||
2833 | #endif | |||
2834 | ||||
2835 | /// Create a new VectorizableTree entry. | |||
2836 | TreeEntry *newTreeEntry(ArrayRef<Value *> VL, | |||
2837 | std::optional<ScheduleData *> Bundle, | |||
2838 | const InstructionsState &S, | |||
2839 | const EdgeInfo &UserTreeIdx, | |||
2840 | ArrayRef<int> ReuseShuffleIndices = std::nullopt, | |||
2841 | ArrayRef<unsigned> ReorderIndices = std::nullopt) { | |||
2842 | TreeEntry::EntryState EntryState = | |||
2843 | Bundle ? TreeEntry::Vectorize : TreeEntry::NeedToGather; | |||
2844 | return newTreeEntry(VL, EntryState, Bundle, S, UserTreeIdx, | |||
2845 | ReuseShuffleIndices, ReorderIndices); | |||
2846 | } | |||
2847 | ||||
2848 | TreeEntry *newTreeEntry(ArrayRef<Value *> VL, | |||
2849 | TreeEntry::EntryState EntryState, | |||
2850 | std::optional<ScheduleData *> Bundle, | |||
2851 | const InstructionsState &S, | |||
2852 | const EdgeInfo &UserTreeIdx, | |||
2853 | ArrayRef<int> ReuseShuffleIndices = std::nullopt, | |||
2854 | ArrayRef<unsigned> ReorderIndices = std::nullopt) { | |||
2855 | 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", 2857, __extension__ __PRETTY_FUNCTION__)) | |||
2856 | (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", 2857, __extension__ __PRETTY_FUNCTION__)) | |||
2857 | "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", 2857, __extension__ __PRETTY_FUNCTION__)); | |||
2858 | VectorizableTree.push_back(std::make_unique<TreeEntry>(VectorizableTree)); | |||
2859 | TreeEntry *Last = VectorizableTree.back().get(); | |||
2860 | Last->Idx = VectorizableTree.size() - 1; | |||
2861 | Last->State = EntryState; | |||
2862 | Last->ReuseShuffleIndices.append(ReuseShuffleIndices.begin(), | |||
2863 | ReuseShuffleIndices.end()); | |||
2864 | if (ReorderIndices.empty()) { | |||
2865 | Last->Scalars.assign(VL.begin(), VL.end()); | |||
2866 | Last->setOperations(S); | |||
2867 | } else { | |||
2868 | // Reorder scalars and build final mask. | |||
2869 | Last->Scalars.assign(VL.size(), nullptr); | |||
2870 | transform(ReorderIndices, Last->Scalars.begin(), | |||
2871 | [VL](unsigned Idx) -> Value * { | |||
2872 | if (Idx >= VL.size()) | |||
2873 | return UndefValue::get(VL.front()->getType()); | |||
2874 | return VL[Idx]; | |||
2875 | }); | |||
2876 | InstructionsState S = getSameOpcode(Last->Scalars, *TLI); | |||
2877 | Last->setOperations(S); | |||
2878 | Last->ReorderIndices.append(ReorderIndices.begin(), ReorderIndices.end()); | |||
2879 | } | |||
2880 | if (Last->State != TreeEntry::NeedToGather) { | |||
2881 | for (Value *V : VL) { | |||
2882 | 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", 2882, __extension__ __PRETTY_FUNCTION__)); | |||
2883 | ScalarToTreeEntry[V] = Last; | |||
2884 | } | |||
2885 | // Update the scheduler bundle to point to this TreeEntry. | |||
2886 | ScheduleData *BundleMember = *Bundle; | |||
2887 | 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", 2890, __extension__ __PRETTY_FUNCTION__)) | |||
2888 | 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", 2890, __extension__ __PRETTY_FUNCTION__)) | |||
2889 | 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", 2890, __extension__ __PRETTY_FUNCTION__)) | |||
2890 | "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", 2890, __extension__ __PRETTY_FUNCTION__)); | |||
2891 | if (BundleMember) { | |||
2892 | for (Value *V : VL) { | |||
2893 | if (doesNotNeedToBeScheduled(V)) | |||
2894 | continue; | |||
2895 | 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", 2895, __extension__ __PRETTY_FUNCTION__)); | |||
2896 | BundleMember->TE = Last; | |||
2897 | BundleMember = BundleMember->NextInBundle; | |||
2898 | } | |||
2899 | } | |||
2900 | 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", 2900, __extension__ __PRETTY_FUNCTION__)); | |||
2901 | } else { | |||
2902 | MustGather.insert(VL.begin(), VL.end()); | |||
2903 | } | |||
2904 | ||||
2905 | if (UserTreeIdx.UserTE) | |||
2906 | Last->UserTreeIndices.push_back(UserTreeIdx); | |||
2907 | ||||
2908 | return Last; | |||
2909 | } | |||
2910 | ||||
2911 | /// -- Vectorization State -- | |||
2912 | /// Holds all of the tree entries. | |||
2913 | TreeEntry::VecTreeTy VectorizableTree; | |||
2914 | ||||
2915 | #ifndef NDEBUG | |||
2916 | /// Debug printer. | |||
2917 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void dumpVectorizableTree() const { | |||
2918 | for (unsigned Id = 0, IdE = VectorizableTree.size(); Id != IdE; ++Id) { | |||
2919 | VectorizableTree[Id]->dump(); | |||
2920 | dbgs() << "\n"; | |||
2921 | } | |||
2922 | } | |||
2923 | #endif | |||
2924 | ||||
2925 | TreeEntry *getTreeEntry(Value *V) { return ScalarToTreeEntry.lookup(V); } | |||
2926 | ||||
2927 | const TreeEntry *getTreeEntry(Value *V) const { | |||
2928 | return ScalarToTreeEntry.lookup(V); | |||
2929 | } | |||
2930 | ||||
2931 | /// Maps a specific scalar to its tree entry. | |||
2932 | SmallDenseMap<Value*, TreeEntry *> ScalarToTreeEntry; | |||
2933 | ||||
2934 | /// Maps a value to the proposed vectorizable size. | |||
2935 | SmallDenseMap<Value *, unsigned> InstrElementSize; | |||
2936 | ||||
2937 | /// A list of scalars that we found that we need to keep as scalars. | |||
2938 | ValueSet MustGather; | |||
2939 | ||||
2940 | /// A map between the vectorized entries and the last instructions in the | |||
2941 | /// bundles. The bundles are built in use order, not in the def order of the | |||
2942 | /// instructions. So, we cannot rely directly on the last instruction in the | |||
2943 | /// bundle being the last instruction in the program order during | |||
2944 | /// vectorization process since the basic blocks are affected, need to | |||
2945 | /// pre-gather them before. | |||
2946 | DenseMap<const TreeEntry *, Instruction *> EntryToLastInstruction; | |||
2947 | ||||
2948 | /// List of gather nodes, depending on other gather/vector nodes, which should | |||
2949 | /// be emitted after the vector instruction emission process to correctly | |||
2950 | /// handle order of the vector instructions and shuffles. | |||
2951 | SetVector<const TreeEntry *> PostponedGathers; | |||
2952 | ||||
2953 | using ValueToGatherNodesMap = | |||
2954 | DenseMap<Value *, SmallPtrSet<const TreeEntry *, 4>>; | |||
2955 | ValueToGatherNodesMap ValueToGatherNodes; | |||
2956 | ||||
2957 | /// This POD struct describes one external user in the vectorized tree. | |||
2958 | struct ExternalUser { | |||
2959 | ExternalUser(Value *S, llvm::User *U, int L) | |||
2960 | : Scalar(S), User(U), Lane(L) {} | |||
2961 | ||||
2962 | // Which scalar in our function. | |||
2963 | Value *Scalar; | |||
2964 | ||||
2965 | // Which user that uses the scalar. | |||
2966 | llvm::User *User; | |||
2967 | ||||
2968 | // Which lane does the scalar belong to. | |||
2969 | int Lane; | |||
2970 | }; | |||
2971 | using UserList = SmallVector<ExternalUser, 16>; | |||
2972 | ||||
2973 | /// Checks if two instructions may access the same memory. | |||
2974 | /// | |||
2975 | /// \p Loc1 is the location of \p Inst1. It is passed explicitly because it | |||
2976 | /// is invariant in the calling loop. | |||
2977 | bool isAliased(const MemoryLocation &Loc1, Instruction *Inst1, | |||
2978 | Instruction *Inst2) { | |||
2979 | // First check if the result is already in the cache. | |||
2980 | AliasCacheKey key = std::make_pair(Inst1, Inst2); | |||
2981 | std::optional<bool> &result = AliasCache[key]; | |||
2982 | if (result) { | |||
2983 | return *result; | |||
2984 | } | |||
2985 | bool aliased = true; | |||
2986 | if (Loc1.Ptr && isSimple(Inst1)) | |||
2987 | aliased = isModOrRefSet(BatchAA.getModRefInfo(Inst2, Loc1)); | |||
2988 | // Store the result in the cache. | |||
2989 | result = aliased; | |||
2990 | return aliased; | |||
2991 | } | |||
2992 | ||||
2993 | using AliasCacheKey = std::pair<Instruction *, Instruction *>; | |||
2994 | ||||
2995 | /// Cache for alias results. | |||
2996 | /// TODO: consider moving this to the AliasAnalysis itself. | |||
2997 | DenseMap<AliasCacheKey, std::optional<bool>> AliasCache; | |||
2998 | ||||
2999 | // Cache for pointerMayBeCaptured calls inside AA. This is preserved | |||
3000 | // globally through SLP because we don't perform any action which | |||
3001 | // invalidates capture results. | |||
3002 | BatchAAResults BatchAA; | |||
3003 | ||||
3004 | /// Temporary store for deleted instructions. Instructions will be deleted | |||
3005 | /// eventually when the BoUpSLP is destructed. The deferral is required to | |||
3006 | /// ensure that there are no incorrect collisions in the AliasCache, which | |||
3007 | /// can happen if a new instruction is allocated at the same address as a | |||
3008 | /// previously deleted instruction. | |||
3009 | DenseSet<Instruction *> DeletedInstructions; | |||
3010 | ||||
3011 | /// Set of the instruction, being analyzed already for reductions. | |||
3012 | SmallPtrSet<Instruction *, 16> AnalyzedReductionsRoots; | |||
3013 | ||||
3014 | /// Set of hashes for the list of reduction values already being analyzed. | |||
3015 | DenseSet<size_t> AnalyzedReductionVals; | |||
3016 | ||||
3017 | /// A list of values that need to extracted out of the tree. | |||
3018 | /// This list holds pairs of (Internal Scalar : External User). External User | |||
3019 | /// can be nullptr, it means that this Internal Scalar will be used later, | |||
3020 | /// after vectorization. | |||
3021 | UserList ExternalUses; | |||
3022 | ||||
3023 | /// Values used only by @llvm.assume calls. | |||
3024 | SmallPtrSet<const Value *, 32> EphValues; | |||
3025 | ||||
3026 | /// Holds all of the instructions that we gathered, shuffle instructions and | |||
3027 | /// extractelements. | |||
3028 | SetVector<Instruction *> GatherShuffleExtractSeq; | |||
3029 | ||||
3030 | /// A list of blocks that we are going to CSE. | |||
3031 | SetVector<BasicBlock *> CSEBlocks; | |||
3032 | ||||
3033 | /// Contains all scheduling relevant data for an instruction. | |||
3034 | /// A ScheduleData either represents a single instruction or a member of an | |||
3035 | /// instruction bundle (= a group of instructions which is combined into a | |||
3036 | /// vector instruction). | |||
3037 | struct ScheduleData { | |||
3038 | // The initial value for the dependency counters. It means that the | |||
3039 | // dependencies are not calculated yet. | |||
3040 | enum { InvalidDeps = -1 }; | |||
3041 | ||||
3042 | ScheduleData() = default; | |||
3043 | ||||
3044 | void init(int BlockSchedulingRegionID, Value *OpVal) { | |||
3045 | FirstInBundle = this; | |||
3046 | NextInBundle = nullptr; | |||
3047 | NextLoadStore = nullptr; | |||
3048 | IsScheduled = false; | |||
3049 | SchedulingRegionID = BlockSchedulingRegionID; | |||
3050 | clearDependencies(); | |||
3051 | OpValue = OpVal; | |||
3052 | TE = nullptr; | |||
3053 | } | |||
3054 | ||||
3055 | /// Verify basic self consistency properties | |||
3056 | void verify() { | |||
3057 | if (hasValidDependencies()) { | |||
3058 | assert(UnscheduledDeps <= Dependencies && "invariant")(static_cast <bool> (UnscheduledDeps <= Dependencies && "invariant") ? void (0) : __assert_fail ("UnscheduledDeps <= Dependencies && \"invariant\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 3058, __extension__ __PRETTY_FUNCTION__)); | |||
3059 | } else { | |||
3060 | assert(UnscheduledDeps == Dependencies && "invariant")(static_cast <bool> (UnscheduledDeps == Dependencies && "invariant") ? void (0) : __assert_fail ("UnscheduledDeps == Dependencies && \"invariant\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 3060, __extension__ __PRETTY_FUNCTION__)); | |||
3061 | } | |||
3062 | ||||
3063 | if (IsScheduled) { | |||
3064 | assert(isSchedulingEntity() &&(static_cast <bool> (isSchedulingEntity() && "unexpected scheduled state" ) ? void (0) : __assert_fail ("isSchedulingEntity() && \"unexpected scheduled state\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 3065, __extension__ __PRETTY_FUNCTION__)) | |||
3065 | "unexpected scheduled state")(static_cast <bool> (isSchedulingEntity() && "unexpected scheduled state" ) ? void (0) : __assert_fail ("isSchedulingEntity() && \"unexpected scheduled state\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 3065, __extension__ __PRETTY_FUNCTION__)); | |||
3066 | for (const ScheduleData *BundleMember = this; BundleMember; | |||
3067 | BundleMember = BundleMember->NextInBundle) { | |||
3068 | 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", 3070, __extension__ __PRETTY_FUNCTION__)) | |||
3069 | 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", 3070, __extension__ __PRETTY_FUNCTION__)) | |||
3070 | "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", 3070, __extension__ __PRETTY_FUNCTION__)); | |||
3071 | 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", 3072, __extension__ __PRETTY_FUNCTION__)) | |||
3072 | "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", 3072, __extension__ __PRETTY_FUNCTION__)); | |||
3073 | } | |||
3074 | } | |||
3075 | ||||
3076 | 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", 3077, __extension__ __PRETTY_FUNCTION__)) | |||
3077 | "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", 3077, __extension__ __PRETTY_FUNCTION__)); | |||
3078 | } | |||
3079 | ||||
3080 | /// Returns true if the dependency information has been calculated. | |||
3081 | /// Note that depenendency validity can vary between instructions within | |||
3082 | /// a single bundle. | |||
3083 | bool hasValidDependencies() const { return Dependencies != InvalidDeps; } | |||
3084 | ||||
3085 | /// Returns true for single instructions and for bundle representatives | |||
3086 | /// (= the head of a bundle). | |||
3087 | bool isSchedulingEntity() const { return FirstInBundle == this; } | |||
3088 | ||||
3089 | /// Returns true if it represents an instruction bundle and not only a | |||
3090 | /// single instruction. | |||
3091 | bool isPartOfBundle() const { | |||
3092 | return NextInBundle != nullptr || FirstInBundle != this || TE; | |||
3093 | } | |||
3094 | ||||
3095 | /// Returns true if it is ready for scheduling, i.e. it has no more | |||
3096 | /// unscheduled depending instructions/bundles. | |||
3097 | bool isReady() const { | |||
3098 | 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", 3099, __extension__ __PRETTY_FUNCTION__)) | |||
3099 | "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", 3099, __extension__ __PRETTY_FUNCTION__)); | |||
3100 | return unscheduledDepsInBundle() == 0 && !IsScheduled; | |||
3101 | } | |||
3102 | ||||
3103 | /// Modifies the number of unscheduled dependencies for this instruction, | |||
3104 | /// and returns the number of remaining dependencies for the containing | |||
3105 | /// bundle. | |||
3106 | int incrementUnscheduledDeps(int Incr) { | |||
3107 | 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", 3108, __extension__ __PRETTY_FUNCTION__)) | |||
3108 | "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", 3108, __extension__ __PRETTY_FUNCTION__)); | |||
3109 | UnscheduledDeps += Incr; | |||
3110 | return FirstInBundle->unscheduledDepsInBundle(); | |||
3111 | } | |||
3112 | ||||
3113 | /// Sets the number of unscheduled dependencies to the number of | |||
3114 | /// dependencies. | |||
3115 | void resetUnscheduledDeps() { | |||
3116 | UnscheduledDeps = Dependencies; | |||
3117 | } | |||
3118 | ||||
3119 | /// Clears all dependency information. | |||
3120 | void clearDependencies() { | |||
3121 | Dependencies = InvalidDeps; | |||
3122 | resetUnscheduledDeps(); | |||
3123 | MemoryDependencies.clear(); | |||
3124 | ControlDependencies.clear(); | |||
3125 | } | |||
3126 | ||||
3127 | int unscheduledDepsInBundle() const { | |||
3128 | 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", 3128, __extension__ __PRETTY_FUNCTION__)); | |||
3129 | int Sum = 0; | |||
3130 | for (const ScheduleData *BundleMember = this; BundleMember; | |||
3131 | BundleMember = BundleMember->NextInBundle) { | |||
3132 | if (BundleMember->UnscheduledDeps == InvalidDeps) | |||
3133 | return InvalidDeps; | |||
3134 | Sum += BundleMember->UnscheduledDeps; | |||
3135 | } | |||
3136 | return Sum; | |||
3137 | } | |||
3138 | ||||
3139 | void dump(raw_ostream &os) const { | |||
3140 | if (!isSchedulingEntity()) { | |||
3141 | os << "/ " << *Inst; | |||
3142 | } else if (NextInBundle) { | |||
3143 | os << '[' << *Inst; | |||
3144 | ScheduleData *SD = NextInBundle; | |||
3145 | while (SD) { | |||
3146 | os << ';' << *SD->Inst; | |||
3147 | SD = SD->NextInBundle; | |||
3148 | } | |||
3149 | os << ']'; | |||
3150 | } else { | |||
3151 | os << *Inst; | |||
3152 | } | |||
3153 | } | |||
3154 | ||||
3155 | Instruction *Inst = nullptr; | |||
3156 | ||||
3157 | /// Opcode of the current instruction in the schedule data. | |||
3158 | Value *OpValue = nullptr; | |||
3159 | ||||
3160 | /// The TreeEntry that this instruction corresponds to. | |||
3161 | TreeEntry *TE = nullptr; | |||
3162 | ||||
3163 | /// Points to the head in an instruction bundle (and always to this for | |||
3164 | /// single instructions). | |||
3165 | ScheduleData *FirstInBundle = nullptr; | |||
3166 | ||||
3167 | /// Single linked list of all instructions in a bundle. Null if it is a | |||
3168 | /// single instruction. | |||
3169 | ScheduleData *NextInBundle = nullptr; | |||
3170 | ||||
3171 | /// Single linked list of all memory instructions (e.g. load, store, call) | |||
3172 | /// in the block - until the end of the scheduling region. | |||
3173 | ScheduleData *NextLoadStore = nullptr; | |||
3174 | ||||
3175 | /// The dependent memory instructions. | |||
3176 | /// This list is derived on demand in calculateDependencies(). | |||
3177 | SmallVector<ScheduleData *, 4> MemoryDependencies; | |||
3178 | ||||
3179 | /// List of instructions which this instruction could be control dependent | |||
3180 | /// on. Allowing such nodes to be scheduled below this one could introduce | |||
3181 | /// a runtime fault which didn't exist in the original program. | |||
3182 | /// ex: this is a load or udiv following a readonly call which inf loops | |||
3183 | SmallVector<ScheduleData *, 4> ControlDependencies; | |||
3184 | ||||
3185 | /// This ScheduleData is in the current scheduling region if this matches | |||
3186 | /// the current SchedulingRegionID of BlockScheduling. | |||
3187 | int SchedulingRegionID = 0; | |||
3188 | ||||
3189 | /// Used for getting a "good" final ordering of instructions. | |||
3190 | int SchedulingPriority = 0; | |||
3191 | ||||
3192 | /// The number of dependencies. Constitutes of the number of users of the | |||
3193 | /// instruction plus the number of dependent memory instructions (if any). | |||
3194 | /// This value is calculated on demand. | |||
3195 | /// If InvalidDeps, the number of dependencies is not calculated yet. | |||
3196 | int Dependencies = InvalidDeps; | |||
3197 | ||||
3198 | /// The number of dependencies minus the number of dependencies of scheduled | |||
3199 | /// instructions. As soon as this is zero, the instruction/bundle gets ready | |||
3200 | /// for scheduling. | |||
3201 | /// Note that this is negative as long as Dependencies is not calculated. | |||
3202 | int UnscheduledDeps = InvalidDeps; | |||
3203 | ||||
3204 | /// True if this instruction is scheduled (or considered as scheduled in the | |||
3205 | /// dry-run). | |||
3206 | bool IsScheduled = false; | |||
3207 | }; | |||
3208 | ||||
3209 | #ifndef NDEBUG | |||
3210 | friend inline raw_ostream &operator<<(raw_ostream &os, | |||
3211 | const BoUpSLP::ScheduleData &SD) { | |||
3212 | SD.dump(os); | |||
3213 | return os; | |||
3214 | } | |||
3215 | #endif | |||
3216 | ||||
3217 | friend struct GraphTraits<BoUpSLP *>; | |||
3218 | friend struct DOTGraphTraits<BoUpSLP *>; | |||
3219 | ||||
3220 | /// Contains all scheduling data for a basic block. | |||
3221 | /// It does not schedules instructions, which are not memory read/write | |||
3222 | /// instructions and their operands are either constants, or arguments, or | |||
3223 | /// phis, or instructions from others blocks, or their users are phis or from | |||
3224 | /// the other blocks. The resulting vector instructions can be placed at the | |||
3225 | /// beginning of the basic block without scheduling (if operands does not need | |||
3226 | /// to be scheduled) or at the end of the block (if users are outside of the | |||
3227 | /// block). It allows to save some compile time and memory used by the | |||
3228 | /// compiler. | |||
3229 | /// ScheduleData is assigned for each instruction in between the boundaries of | |||
3230 | /// the tree entry, even for those, which are not part of the graph. It is | |||
3231 | /// required to correctly follow the dependencies between the instructions and | |||
3232 | /// their correct scheduling. The ScheduleData is not allocated for the | |||
3233 | /// instructions, which do not require scheduling, like phis, nodes with | |||
3234 | /// extractelements/insertelements only or nodes with instructions, with | |||
3235 | /// uses/operands outside of the block. | |||
3236 | struct BlockScheduling { | |||
3237 | BlockScheduling(BasicBlock *BB) | |||
3238 | : BB(BB), ChunkSize(BB->size()), ChunkPos(ChunkSize) {} | |||
3239 | ||||
3240 | void clear() { | |||
3241 | ReadyInsts.clear(); | |||
3242 | ScheduleStart = nullptr; | |||
3243 | ScheduleEnd = nullptr; | |||
3244 | FirstLoadStoreInRegion = nullptr; | |||
3245 | LastLoadStoreInRegion = nullptr; | |||
3246 | RegionHasStackSave = false; | |||
3247 | ||||
3248 | // Reduce the maximum schedule region size by the size of the | |||
3249 | // previous scheduling run. | |||
3250 | ScheduleRegionSizeLimit -= ScheduleRegionSize; | |||
3251 | if (ScheduleRegionSizeLimit < MinScheduleRegionSize) | |||
3252 | ScheduleRegionSizeLimit = MinScheduleRegionSize; | |||
3253 | ScheduleRegionSize = 0; | |||
3254 | ||||
3255 | // Make a new scheduling region, i.e. all existing ScheduleData is not | |||
3256 | // in the new region yet. | |||
3257 | ++SchedulingRegionID; | |||
3258 | } | |||
3259 | ||||
3260 | ScheduleData *getScheduleData(Instruction *I) { | |||
3261 | if (BB != I->getParent()) | |||
3262 | // Avoid lookup if can't possibly be in map. | |||
3263 | return nullptr; | |||
3264 | ScheduleData *SD = ScheduleDataMap.lookup(I); | |||
3265 | if (SD && isInSchedulingRegion(SD)) | |||
3266 | return SD; | |||
3267 | return nullptr; | |||
3268 | } | |||
3269 | ||||
3270 | ScheduleData *getScheduleData(Value *V) { | |||
3271 | if (auto *I = dyn_cast<Instruction>(V)) | |||
3272 | return getScheduleData(I); | |||
3273 | return nullptr; | |||
3274 | } | |||
3275 | ||||
3276 | ScheduleData *getScheduleData(Value *V, Value *Key) { | |||
3277 | if (V == Key) | |||
3278 | return getScheduleData(V); | |||
3279 | auto I = ExtraScheduleDataMap.find(V); | |||
3280 | if (I != ExtraScheduleDataMap.end()) { | |||
3281 | ScheduleData *SD = I->second.lookup(Key); | |||
3282 | if (SD && isInSchedulingRegion(SD)) | |||
3283 | return SD; | |||
3284 | } | |||
3285 | return nullptr; | |||
3286 | } | |||
3287 | ||||
3288 | bool isInSchedulingRegion(ScheduleData *SD) const { | |||
3289 | return SD->SchedulingRegionID == SchedulingRegionID; | |||
3290 | } | |||
3291 | ||||
3292 | /// Marks an instruction as scheduled and puts all dependent ready | |||
3293 | /// instructions into the ready-list. | |||
3294 | template <typename ReadyListType> | |||
3295 | void schedule(ScheduleData *SD, ReadyListType &ReadyList) { | |||
3296 | SD->IsScheduled = true; | |||
3297 | LLVM_DEBUG(dbgs() << "SLP: schedule " << *SD << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: schedule " << *SD << "\n"; } } while (false); | |||
3298 | ||||
3299 | for (ScheduleData *BundleMember = SD; BundleMember; | |||
3300 | BundleMember = BundleMember->NextInBundle) { | |||
3301 | if (BundleMember->Inst != BundleMember->OpValue) | |||
3302 | continue; | |||
3303 | ||||
3304 | // Handle the def-use chain dependencies. | |||
3305 | ||||
3306 | // Decrement the unscheduled counter and insert to ready list if ready. | |||
3307 | auto &&DecrUnsched = [this, &ReadyList](Instruction *I) { | |||
3308 | doForAllOpcodes(I, [&ReadyList](ScheduleData *OpDef) { | |||
3309 | if (OpDef && OpDef->hasValidDependencies() && | |||
3310 | OpDef->incrementUnscheduledDeps(-1) == 0) { | |||
3311 | // There are no more unscheduled dependencies after | |||
3312 | // decrementing, so we can put the dependent instruction | |||
3313 | // into the ready list. | |||
3314 | ScheduleData *DepBundle = OpDef->FirstInBundle; | |||
3315 | 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", 3316, __extension__ __PRETTY_FUNCTION__)) | |||
3316 | "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", 3316, __extension__ __PRETTY_FUNCTION__)); | |||
3317 | ReadyList.insert(DepBundle); | |||
3318 | LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: gets ready (def): " << *DepBundle << "\n"; } } while (false) | |||
3319 | << "SLP: gets ready (def): " << *DepBundle << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: gets ready (def): " << *DepBundle << "\n"; } } while (false); | |||
3320 | } | |||
3321 | }); | |||
3322 | }; | |||
3323 | ||||
3324 | // If BundleMember is a vector bundle, its operands may have been | |||
3325 | // reordered during buildTree(). We therefore need to get its operands | |||
3326 | // through the TreeEntry. | |||
3327 | if (TreeEntry *TE = BundleMember->TE) { | |||
3328 | // Need to search for the lane since the tree entry can be reordered. | |||
3329 | int Lane = std::distance(TE->Scalars.begin(), | |||
3330 | find(TE->Scalars, BundleMember->Inst)); | |||
3331 | 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", 3331, __extension__ __PRETTY_FUNCTION__)); | |||
3332 | ||||
3333 | // Since vectorization tree is being built recursively this assertion | |||
3334 | // ensures that the tree entry has all operands set before reaching | |||
3335 | // this code. Couple of exceptions known at the moment are extracts | |||
3336 | // where their second (immediate) operand is not added. Since | |||
3337 | // immediates do not affect scheduler behavior this is considered | |||
3338 | // okay. | |||
3339 | auto *In = BundleMember->Inst; | |||
3340 | assert(In &&(static_cast <bool> (In && (isa<ExtractValueInst , ExtractElementInst>(In) || In->getNumOperands() == TE ->getNumOperands()) && "Missed TreeEntry operands?" ) ? void (0) : __assert_fail ("In && (isa<ExtractValueInst, ExtractElementInst>(In) || In->getNumOperands() == TE->getNumOperands()) && \"Missed TreeEntry operands?\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 3343, __extension__ __PRETTY_FUNCTION__)) | |||
3341 | (isa<ExtractValueInst, ExtractElementInst>(In) ||(static_cast <bool> (In && (isa<ExtractValueInst , ExtractElementInst>(In) || In->getNumOperands() == TE ->getNumOperands()) && "Missed TreeEntry operands?" ) ? void (0) : __assert_fail ("In && (isa<ExtractValueInst, ExtractElementInst>(In) || In->getNumOperands() == TE->getNumOperands()) && \"Missed TreeEntry operands?\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 3343, __extension__ __PRETTY_FUNCTION__)) | |||
3342 | In->getNumOperands() == TE->getNumOperands()) &&(static_cast <bool> (In && (isa<ExtractValueInst , ExtractElementInst>(In) || In->getNumOperands() == TE ->getNumOperands()) && "Missed TreeEntry operands?" ) ? void (0) : __assert_fail ("In && (isa<ExtractValueInst, ExtractElementInst>(In) || In->getNumOperands() == TE->getNumOperands()) && \"Missed TreeEntry operands?\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 3343, __extension__ __PRETTY_FUNCTION__)) | |||
3343 | "Missed TreeEntry operands?")(static_cast <bool> (In && (isa<ExtractValueInst , ExtractElementInst>(In) || In->getNumOperands() == TE ->getNumOperands()) && "Missed TreeEntry operands?" ) ? void (0) : __assert_fail ("In && (isa<ExtractValueInst, ExtractElementInst>(In) || In->getNumOperands() == TE->getNumOperands()) && \"Missed TreeEntry operands?\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 3343, __extension__ __PRETTY_FUNCTION__)); | |||
3344 | (void)In; // fake use to avoid build failure when assertions disabled | |||
3345 | ||||
3346 | for (unsigned OpIdx = 0, NumOperands = TE->getNumOperands(); | |||
3347 | OpIdx != NumOperands; ++OpIdx) | |||
3348 | if (auto *I = dyn_cast<Instruction>(TE->getOperand(OpIdx)[Lane])) | |||
3349 | DecrUnsched(I); | |||
3350 | } else { | |||
3351 | // If BundleMember is a stand-alone instruction, no operand reordering | |||
3352 | // has taken place, so we directly access its operands. | |||
3353 | for (Use &U : BundleMember->Inst->operands()) | |||
3354 | if (auto *I = dyn_cast<Instruction>(U.get())) | |||
3355 | DecrUnsched(I); | |||
3356 | } | |||
3357 | // Handle the memory dependencies. | |||
3358 | for (ScheduleData *MemoryDepSD : BundleMember->MemoryDependencies) { | |||
3359 | if (MemoryDepSD->hasValidDependencies() && | |||
3360 | MemoryDepSD->incrementUnscheduledDeps(-1) == 0) { | |||
3361 | // There are no more unscheduled dependencies after decrementing, | |||
3362 | // so we can put the dependent instruction into the ready list. | |||
3363 | ScheduleData *DepBundle = MemoryDepSD->FirstInBundle; | |||
3364 | 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", 3365, __extension__ __PRETTY_FUNCTION__)) | |||
3365 | "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", 3365, __extension__ __PRETTY_FUNCTION__)); | |||
3366 | ReadyList.insert(DepBundle); | |||
3367 | LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: gets ready (mem): " << *DepBundle << "\n"; } } while (false) | |||
3368 | << "SLP: gets ready (mem): " << *DepBundle << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: gets ready (mem): " << *DepBundle << "\n"; } } while (false); | |||
3369 | } | |||
3370 | } | |||
3371 | // Handle the control dependencies. | |||
3372 | for (ScheduleData *DepSD : BundleMember->ControlDependencies) { | |||
3373 | if (DepSD->incrementUnscheduledDeps(-1) == 0) { | |||
3374 | // There are no more unscheduled dependencies after decrementing, | |||
3375 | // so we can put the dependent instruction into the ready list. | |||
3376 | ScheduleData *DepBundle = DepSD->FirstInBundle; | |||
3377 | 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", 3378, __extension__ __PRETTY_FUNCTION__)) | |||
3378 | "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", 3378, __extension__ __PRETTY_FUNCTION__)); | |||
3379 | ReadyList.insert(DepBundle); | |||
3380 | LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: gets ready (ctl): " << *DepBundle << "\n"; } } while (false) | |||
3381 | << "SLP: gets ready (ctl): " << *DepBundle << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: gets ready (ctl): " << *DepBundle << "\n"; } } while (false); | |||
3382 | } | |||
3383 | } | |||
3384 | ||||
3385 | } | |||
3386 | } | |||
3387 | ||||
3388 | /// Verify basic self consistency properties of the data structure. | |||
3389 | void verify() { | |||
3390 | if (!ScheduleStart) | |||
3391 | return; | |||
3392 | ||||
3393 | 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", 3395, __extension__ __PRETTY_FUNCTION__)) | |||
3394 | 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", 3395, __extension__ __PRETTY_FUNCTION__)) | |||
3395 | "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", 3395, __extension__ __PRETTY_FUNCTION__)); | |||
3396 | ||||
3397 | for (auto *I = ScheduleStart; I != ScheduleEnd; I = I->getNextNode()) { | |||
3398 | auto *SD = getScheduleData(I); | |||
3399 | if (!SD) | |||
3400 | continue; | |||
3401 | 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", 3402, __extension__ __PRETTY_FUNCTION__)) | |||
3402 | "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", 3402, __extension__ __PRETTY_FUNCTION__)); | |||
3403 | 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", 3404, __extension__ __PRETTY_FUNCTION__)) | |||
3404 | "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", 3404, __extension__ __PRETTY_FUNCTION__)); | |||
3405 | (void)SD; | |||
3406 | doForAllOpcodes(I, [](ScheduleData *SD) { SD->verify(); }); | |||
3407 | } | |||
3408 | ||||
3409 | for (auto *SD : ReadyInsts) { | |||
3410 | 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", 3411, __extension__ __PRETTY_FUNCTION__)) | |||
3411 | "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", 3411, __extension__ __PRETTY_FUNCTION__)); | |||
3412 | (void)SD; | |||
3413 | } | |||
3414 | } | |||
3415 | ||||
3416 | void doForAllOpcodes(Value *V, | |||
3417 | function_ref<void(ScheduleData *SD)> Action) { | |||
3418 | if (ScheduleData *SD = getScheduleData(V)) | |||
3419 | Action(SD); | |||
3420 | auto I = ExtraScheduleDataMap.find(V); | |||
3421 | if (I != ExtraScheduleDataMap.end()) | |||
3422 | for (auto &P : I->second) | |||
3423 | if (isInSchedulingRegion(P.second)) | |||
3424 | Action(P.second); | |||
3425 | } | |||
3426 | ||||
3427 | /// Put all instructions into the ReadyList which are ready for scheduling. | |||
3428 | template <typename ReadyListType> | |||
3429 | void initialFillReadyList(ReadyListType &ReadyList) { | |||
3430 | for (auto *I = ScheduleStart; I != ScheduleEnd; I = I->getNextNode()) { | |||
3431 | doForAllOpcodes(I, [&](ScheduleData *SD) { | |||
3432 | if (SD->isSchedulingEntity() && SD->hasValidDependencies() && | |||
3433 | SD->isReady()) { | |||
3434 | ReadyList.insert(SD); | |||
3435 | LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: initially in ready list: " << *SD << "\n"; } } while (false) | |||
3436 | << "SLP: initially in ready list: " << *SD << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: initially in ready list: " << *SD << "\n"; } } while (false); | |||
3437 | } | |||
3438 | }); | |||
3439 | } | |||
3440 | } | |||
3441 | ||||
3442 | /// Build a bundle from the ScheduleData nodes corresponding to the | |||
3443 | /// scalar instruction for each lane. | |||
3444 | ScheduleData *buildBundle(ArrayRef<Value *> VL); | |||
3445 | ||||
3446 | /// Checks if a bundle of instructions can be scheduled, i.e. has no | |||
3447 | /// cyclic dependencies. This is only a dry-run, no instructions are | |||
3448 | /// actually moved at this stage. | |||
3449 | /// \returns the scheduling bundle. The returned Optional value is not | |||
3450 | /// std::nullopt if \p VL is allowed to be scheduled. | |||
3451 | std::optional<ScheduleData *> | |||
3452 | tryScheduleBundle(ArrayRef<Value *> VL, BoUpSLP *SLP, | |||
3453 | const InstructionsState &S); | |||
3454 | ||||
3455 | /// Un-bundles a group of instructions. | |||
3456 | void cancelScheduling(ArrayRef<Value *> VL, Value *OpValue); | |||
3457 | ||||
3458 | /// Allocates schedule data chunk. | |||
3459 | ScheduleData *allocateScheduleDataChunks(); | |||
3460 | ||||
3461 | /// Extends the scheduling region so that V is inside the region. | |||
3462 | /// \returns true if the region size is within the limit. | |||
3463 | bool extendSchedulingRegion(Value *V, const InstructionsState &S); | |||
3464 | ||||
3465 | /// Initialize the ScheduleData structures for new instructions in the | |||
3466 | /// scheduling region. | |||
3467 | void initScheduleData(Instruction *FromI, Instruction *ToI, | |||
3468 | ScheduleData *PrevLoadStore, | |||
3469 | ScheduleData *NextLoadStore); | |||
3470 | ||||
3471 | /// Updates the dependency information of a bundle and of all instructions/ | |||
3472 | /// bundles which depend on the original bundle. | |||
3473 | void calculateDependencies(ScheduleData *SD, bool InsertInReadyList, | |||
3474 | BoUpSLP *SLP); | |||
3475 | ||||
3476 | /// Sets all instruction in the scheduling region to un-scheduled. | |||
3477 | void resetSchedule(); | |||
3478 | ||||
3479 | BasicBlock *BB; | |||
3480 | ||||
3481 | /// Simple memory allocation for ScheduleData. | |||
3482 | std::vector<std::unique_ptr<ScheduleData[]>> ScheduleDataChunks; | |||
3483 | ||||
3484 | /// The size of a ScheduleData array in ScheduleDataChunks. | |||
3485 | int ChunkSize; | |||
3486 | ||||
3487 | /// The allocator position in the current chunk, which is the last entry | |||
3488 | /// of ScheduleDataChunks. | |||
3489 | int ChunkPos; | |||
3490 | ||||
3491 | /// Attaches ScheduleData to Instruction. | |||
3492 | /// Note that the mapping survives during all vectorization iterations, i.e. | |||
3493 | /// ScheduleData structures are recycled. | |||
3494 | DenseMap<Instruction *, ScheduleData *> ScheduleDataMap; | |||
3495 | ||||
3496 | /// Attaches ScheduleData to Instruction with the leading key. | |||
3497 | DenseMap<Value *, SmallDenseMap<Value *, ScheduleData *>> | |||
3498 | ExtraScheduleDataMap; | |||
3499 | ||||
3500 | /// The ready-list for scheduling (only used for the dry-run). | |||
3501 | SetVector<ScheduleData *> ReadyInsts; | |||
3502 | ||||
3503 | /// The first instruction of the scheduling region. | |||
3504 | Instruction *ScheduleStart = nullptr; | |||
3505 | ||||
3506 | /// The first instruction _after_ the scheduling region. | |||
3507 | Instruction *ScheduleEnd = nullptr; | |||
3508 | ||||
3509 | /// The first memory accessing instruction in the scheduling region | |||
3510 | /// (can be null). | |||
3511 | ScheduleData *FirstLoadStoreInRegion = nullptr; | |||
3512 | ||||
3513 | /// The last memory accessing instruction in the scheduling region | |||
3514 | /// (can be null). | |||
3515 | ScheduleData *LastLoadStoreInRegion = nullptr; | |||
3516 | ||||
3517 | /// Is there an llvm.stacksave or llvm.stackrestore in the scheduling | |||
3518 | /// region? Used to optimize the dependence calculation for the | |||
3519 | /// common case where there isn't. | |||
3520 | bool RegionHasStackSave = false; | |||
3521 | ||||
3522 | /// The current size of the scheduling region. | |||
3523 | int ScheduleRegionSize = 0; | |||
3524 | ||||
3525 | /// The maximum size allowed for the scheduling region. | |||
3526 | int ScheduleRegionSizeLimit = ScheduleRegionSizeBudget; | |||
3527 | ||||
3528 | /// The ID of the scheduling region. For a new vectorization iteration this | |||
3529 | /// is incremented which "removes" all ScheduleData from the region. | |||
3530 | /// Make sure that the initial SchedulingRegionID is greater than the | |||
3531 | /// initial SchedulingRegionID in ScheduleData (which is 0). | |||
3532 | int SchedulingRegionID = 1; | |||
3533 | }; | |||
3534 | ||||
3535 | /// Attaches the BlockScheduling structures to basic blocks. | |||
3536 | MapVector<BasicBlock *, std::unique_ptr<BlockScheduling>> BlocksSchedules; | |||
3537 | ||||
3538 | /// Performs the "real" scheduling. Done before vectorization is actually | |||
3539 | /// performed in a basic block. | |||
3540 | void scheduleBlock(BlockScheduling *BS); | |||
3541 | ||||
3542 | /// List of users to ignore during scheduling and that don't need extracting. | |||
3543 | const SmallDenseSet<Value *> *UserIgnoreList = nullptr; | |||
3544 | ||||
3545 | /// A DenseMapInfo implementation for holding DenseMaps and DenseSets of | |||
3546 | /// sorted SmallVectors of unsigned. | |||
3547 | struct OrdersTypeDenseMapInfo { | |||
3548 | static OrdersType getEmptyKey() { | |||
3549 | OrdersType V; | |||
3550 | V.push_back(~1U); | |||
3551 | return V; | |||
3552 | } | |||
3553 | ||||
3554 | static OrdersType getTombstoneKey() { | |||
3555 | OrdersType V; | |||
3556 | V.push_back(~2U); | |||
3557 | return V; | |||
3558 | } | |||
3559 | ||||
3560 | static unsigned getHashValue(const OrdersType &V) { | |||
3561 | return static_cast<unsigned>(hash_combine_range(V.begin(), V.end())); | |||
3562 | } | |||
3563 | ||||
3564 | static bool isEqual(const OrdersType &LHS, const OrdersType &RHS) { | |||
3565 | return LHS == RHS; | |||
3566 | } | |||
3567 | }; | |||
3568 | ||||
3569 | // Analysis and block reference. | |||
3570 | Function *F; | |||
3571 | ScalarEvolution *SE; | |||
3572 | TargetTransformInfo *TTI; | |||
3573 | TargetLibraryInfo *TLI; | |||
3574 | LoopInfo *LI; | |||
3575 | DominatorTree *DT; | |||
3576 | AssumptionCache *AC; | |||
3577 | DemandedBits *DB; | |||
3578 | const DataLayout *DL; | |||
3579 | OptimizationRemarkEmitter *ORE; | |||
3580 | ||||
3581 | unsigned MaxVecRegSize; // This is set by TTI or overridden by cl::opt. | |||
3582 | unsigned MinVecRegSize; // Set by cl::opt (default: 128). | |||
3583 | ||||
3584 | /// Instruction builder to construct the vectorized tree. | |||
3585 | IRBuilder<> Builder; | |||
3586 | ||||
3587 | /// A map of scalar integer values to the smallest bit width with which they | |||
3588 | /// can legally be represented. The values map to (width, signed) pairs, | |||
3589 | /// where "width" indicates the minimum bit width and "signed" is True if the | |||
3590 | /// value must be signed-extended, rather than zero-extended, back to its | |||
3591 | /// original width. | |||
3592 | MapVector<Value *, std::pair<uint64_t, bool>> MinBWs; | |||
3593 | }; | |||
3594 | ||||
3595 | } // end namespace slpvectorizer | |||
3596 | ||||
3597 | template <> struct GraphTraits<BoUpSLP *> { | |||
3598 | using TreeEntry = BoUpSLP::TreeEntry; | |||
3599 | ||||
3600 | /// NodeRef has to be a pointer per the GraphWriter. | |||
3601 | using NodeRef = TreeEntry *; | |||
3602 | ||||
3603 | using ContainerTy = BoUpSLP::TreeEntry::VecTreeTy; | |||
3604 | ||||
3605 | /// Add the VectorizableTree to the index iterator to be able to return | |||
3606 | /// TreeEntry pointers. | |||
3607 | struct ChildIteratorType | |||
3608 | : public iterator_adaptor_base< | |||
3609 | ChildIteratorType, SmallVector<BoUpSLP::EdgeInfo, 1>::iterator> { | |||
3610 | ContainerTy &VectorizableTree; | |||
3611 | ||||
3612 | ChildIteratorType(SmallVector<BoUpSLP::EdgeInfo, 1>::iterator W, | |||
3613 | ContainerTy &VT) | |||
3614 | : ChildIteratorType::iterator_adaptor_base(W), VectorizableTree(VT) {} | |||
3615 | ||||
3616 | NodeRef operator*() { return I->UserTE; } | |||
3617 | }; | |||
3618 | ||||
3619 | static NodeRef getEntryNode(BoUpSLP &R) { | |||
3620 | return R.VectorizableTree[0].get(); | |||
3621 | } | |||
3622 | ||||
3623 | static ChildIteratorType child_begin(NodeRef N) { | |||
3624 | return {N->UserTreeIndices.begin(), N->Container}; | |||
3625 | } | |||
3626 | ||||
3627 | static ChildIteratorType child_end(NodeRef N) { | |||
3628 | return {N->UserTreeIndices.end(), N->Container}; | |||
3629 | } | |||
3630 | ||||
3631 | /// For the node iterator we just need to turn the TreeEntry iterator into a | |||
3632 | /// TreeEntry* iterator so that it dereferences to NodeRef. | |||
3633 | class nodes_iterator { | |||
3634 | using ItTy = ContainerTy::iterator; | |||
3635 | ItTy It; | |||
3636 | ||||
3637 | public: | |||
3638 | nodes_iterator(const ItTy &It2) : It(It2) {} | |||
3639 | NodeRef operator*() { return It->get(); } | |||
3640 | nodes_iterator operator++() { | |||
3641 | ++It; | |||
3642 | return *this; | |||
3643 | } | |||
3644 | bool operator!=(const nodes_iterator &N2) const { return N2.It != It; } | |||
3645 | }; | |||
3646 | ||||
3647 | static nodes_iterator nodes_begin(BoUpSLP *R) { | |||
3648 | return nodes_iterator(R->VectorizableTree.begin()); | |||
3649 | } | |||
3650 | ||||
3651 | static nodes_iterator nodes_end(BoUpSLP *R) { | |||
3652 | return nodes_iterator(R->VectorizableTree.end()); | |||
3653 | } | |||
3654 | ||||
3655 | static unsigned size(BoUpSLP *R) { return R->VectorizableTree.size(); } | |||
3656 | }; | |||
3657 | ||||
3658 | template <> struct DOTGraphTraits<BoUpSLP *> : public DefaultDOTGraphTraits { | |||
3659 | using TreeEntry = BoUpSLP::TreeEntry; | |||
3660 | ||||
3661 | DOTGraphTraits(bool isSimple = false) : DefaultDOTGraphTraits(isSimple) {} | |||
3662 | ||||
3663 | std::string getNodeLabel(const TreeEntry *Entry, const BoUpSLP *R) { | |||
3664 | std::string Str; | |||
3665 | raw_string_ostream OS(Str); | |||
3666 | OS << Entry->Idx << ".\n"; | |||
3667 | if (isSplat(Entry->Scalars)) | |||
3668 | OS << "<splat> "; | |||
3669 | for (auto *V : Entry->Scalars) { | |||
3670 | OS << *V; | |||
3671 | if (llvm::any_of(R->ExternalUses, [&](const BoUpSLP::ExternalUser &EU) { | |||
3672 | return EU.Scalar == V; | |||
3673 | })) | |||
3674 | OS << " <extract>"; | |||
3675 | OS << "\n"; | |||
3676 | } | |||
3677 | return Str; | |||
3678 | } | |||
3679 | ||||
3680 | static std::string getNodeAttributes(const TreeEntry *Entry, | |||
3681 | const BoUpSLP *) { | |||
3682 | if (Entry->State == TreeEntry::NeedToGather) | |||
3683 | return "color=red"; | |||
3684 | if (Entry->State == TreeEntry::ScatterVectorize) | |||
3685 | return "color=blue"; | |||
3686 | return ""; | |||
3687 | } | |||
3688 | }; | |||
3689 | ||||
3690 | } // end namespace llvm | |||
3691 | ||||
3692 | BoUpSLP::~BoUpSLP() { | |||
3693 | SmallVector<WeakTrackingVH> DeadInsts; | |||
3694 | for (auto *I : DeletedInstructions) { | |||
3695 | for (Use &U : I->operands()) { | |||
3696 | auto *Op = dyn_cast<Instruction>(U.get()); | |||
3697 | if (Op && !DeletedInstructions.count(Op) && Op->hasOneUser() && | |||
3698 | wouldInstructionBeTriviallyDead(Op, TLI)) | |||
3699 | DeadInsts.emplace_back(Op); | |||
3700 | } | |||
3701 | I->dropAllReferences(); | |||
3702 | } | |||
3703 | for (auto *I : DeletedInstructions) { | |||
3704 | 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", 3705, __extension__ __PRETTY_FUNCTION__)) | |||
3705 | "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", 3705, __extension__ __PRETTY_FUNCTION__)); | |||
3706 | I->eraseFromParent(); | |||
3707 | } | |||
3708 | ||||
3709 | // Cleanup any dead scalar code feeding the vectorized instructions | |||
3710 | RecursivelyDeleteTriviallyDeadInstructions(DeadInsts, TLI); | |||
3711 | ||||
3712 | #ifdef EXPENSIVE_CHECKS | |||
3713 | // If we could guarantee that this call is not extremely slow, we could | |||
3714 | // remove the ifdef limitation (see PR47712). | |||
3715 | assert(!verifyFunction(*F, &dbgs()))(static_cast <bool> (!verifyFunction(*F, &dbgs())) ? void (0) : __assert_fail ("!verifyFunction(*F, &dbgs())" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 3715, __extension__ __PRETTY_FUNCTION__)); | |||
3716 | #endif | |||
3717 | } | |||
3718 | ||||
3719 | /// Reorders the given \p Reuses mask according to the given \p Mask. \p Reuses | |||
3720 | /// contains original mask for the scalars reused in the node. Procedure | |||
3721 | /// transform this mask in accordance with the given \p Mask. | |||
3722 | static void reorderReuses(SmallVectorImpl<int> &Reuses, ArrayRef<int> Mask) { | |||
3723 | 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", 3724, __extension__ __PRETTY_FUNCTION__)) | |||
3724 | "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", 3724, __extension__ __PRETTY_FUNCTION__)); | |||
3725 | SmallVector<int> Prev(Reuses.begin(), Reuses.end()); | |||
3726 | Prev.swap(Reuses); | |||
3727 | for (unsigned I = 0, E = Prev.size(); I < E; ++I) | |||
3728 | if (Mask[I] != PoisonMaskElem) | |||
3729 | Reuses[Mask[I]] = Prev[I]; | |||
3730 | } | |||
3731 | ||||
3732 | /// Reorders the given \p Order according to the given \p Mask. \p Order - is | |||
3733 | /// the original order of the scalars. Procedure transforms the provided order | |||
3734 | /// in accordance with the given \p Mask. If the resulting \p Order is just an | |||
3735 | /// identity order, \p Order is cleared. | |||
3736 | static void reorderOrder(SmallVectorImpl<unsigned> &Order, ArrayRef<int> Mask) { | |||
3737 | 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", 3737, __extension__ __PRETTY_FUNCTION__)); | |||
3738 | SmallVector<int> MaskOrder; | |||
3739 | if (Order.empty()) { | |||
3740 | MaskOrder.resize(Mask.size()); | |||
3741 | std::iota(MaskOrder.begin(), MaskOrder.end(), 0); | |||
3742 | } else { | |||
3743 | inversePermutation(Order, MaskOrder); | |||
3744 | } | |||
3745 | reorderReuses(MaskOrder, Mask); | |||
3746 | if (ShuffleVectorInst::isIdentityMask(MaskOrder)) { | |||
3747 | Order.clear(); | |||
3748 | return; | |||
3749 | } | |||
3750 | Order.assign(Mask.size(), Mask.size()); | |||
3751 | for (unsigned I = 0, E = Mask.size(); I < E; ++I) | |||
3752 | if (MaskOrder[I] != PoisonMaskElem) | |||
3753 | Order[MaskOrder[I]] = I; | |||
3754 | fixupOrderingIndices(Order); | |||
3755 | } | |||
3756 | ||||
3757 | std::optional<BoUpSLP::OrdersType> | |||
3758 | BoUpSLP::findReusedOrderedScalars(const BoUpSLP::TreeEntry &TE) { | |||
3759 | 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", 3759, __extension__ __PRETTY_FUNCTION__)); | |||
3760 | unsigned NumScalars = TE.Scalars.size(); | |||
3761 | OrdersType CurrentOrder(NumScalars, NumScalars); | |||
3762 | SmallVector<int> Positions; | |||
3763 | SmallBitVector UsedPositions(NumScalars); | |||
3764 | const TreeEntry *STE = nullptr; | |||
3765 | // Try to find all gathered scalars that are gets vectorized in other | |||
3766 | // vectorize node. Here we can have only one single tree vector node to | |||
3767 | // correctly identify order of the gathered scalars. | |||
3768 | for (unsigned I = 0; I < NumScalars; ++I) { | |||
3769 | Value *V = TE.Scalars[I]; | |||
3770 | if (!isa<LoadInst, ExtractElementInst, ExtractValueInst>(V)) | |||
3771 | continue; | |||
3772 | if (const auto *LocalSTE = getTreeEntry(V)) { | |||
3773 | if (!STE) | |||
3774 | STE = LocalSTE; | |||
3775 | else if (STE != LocalSTE) | |||
3776 | // Take the order only from the single vector node. | |||
3777 | return std::nullopt; | |||
3778 | unsigned Lane = | |||
3779 | std::distance(STE->Scalars.begin(), find(STE->Scalars, V)); | |||
3780 | if (Lane >= NumScalars) | |||
3781 | return std::nullopt; | |||
3782 | if (CurrentOrder[Lane] != NumScalars) { | |||
3783 | if (Lane != I) | |||
3784 | continue; | |||
3785 | UsedPositions.reset(CurrentOrder[Lane]); | |||
3786 | } | |||
3787 | // The partial identity (where only some elements of the gather node are | |||
3788 | // in the identity order) is good. | |||
3789 | CurrentOrder[Lane] = I; | |||
3790 | UsedPositions.set(I); | |||
3791 | } | |||
3792 | } | |||
3793 | // Need to keep the order if we have a vector entry and at least 2 scalars or | |||
3794 | // the vectorized entry has just 2 scalars. | |||
3795 | if (STE && (UsedPositions.count() > 1 || STE->Scalars.size() == 2)) { | |||
3796 | auto &&IsIdentityOrder = [NumScalars](ArrayRef<unsigned> CurrentOrder) { | |||
3797 | for (unsigned I = 0; I < NumScalars; ++I) | |||
3798 | if (CurrentOrder[I] != I && CurrentOrder[I] != NumScalars) | |||
3799 | return false; | |||
3800 | return true; | |||
3801 | }; | |||
3802 | if (IsIdentityOrder(CurrentOrder)) | |||
3803 | return OrdersType(); | |||
3804 | auto *It = CurrentOrder.begin(); | |||
3805 | for (unsigned I = 0; I < NumScalars;) { | |||
3806 | if (UsedPositions.test(I)) { | |||
3807 | ++I; | |||
3808 | continue; | |||
3809 | } | |||
3810 | if (*It == NumScalars) { | |||
3811 | *It = I; | |||
3812 | ++I; | |||
3813 | } | |||
3814 | ++It; | |||
3815 | } | |||
3816 | return std::move(CurrentOrder); | |||
3817 | } | |||
3818 | return std::nullopt; | |||
3819 | } | |||
3820 | ||||
3821 | namespace { | |||
3822 | /// Tracks the state we can represent the loads in the given sequence. | |||
3823 | enum class LoadsState { Gather, Vectorize, ScatterVectorize }; | |||
3824 | } // anonymous namespace | |||
3825 | ||||
3826 | static bool arePointersCompatible(Value *Ptr1, Value *Ptr2, | |||
3827 | const TargetLibraryInfo &TLI, | |||
3828 | bool CompareOpcodes = true) { | |||
3829 | if (getUnderlyingObject(Ptr1) != getUnderlyingObject(Ptr2)) | |||
3830 | return false; | |||
3831 | auto *GEP1 = dyn_cast<GetElementPtrInst>(Ptr1); | |||
3832 | if (!GEP1) | |||
3833 | return false; | |||
3834 | auto *GEP2 = dyn_cast<GetElementPtrInst>(Ptr2); | |||
3835 | if (!GEP2) | |||
3836 | return false; | |||
3837 | return GEP1->getNumOperands() == 2 && GEP2->getNumOperands() == 2 && | |||
3838 | ((isConstant(GEP1->getOperand(1)) && | |||
3839 | isConstant(GEP2->getOperand(1))) || | |||
3840 | !CompareOpcodes || | |||
3841 | getSameOpcode({GEP1->getOperand(1), GEP2->getOperand(1)}, TLI) | |||
3842 | .getOpcode()); | |||
3843 | } | |||
3844 | ||||
3845 | /// Checks if the given array of loads can be represented as a vectorized, | |||
3846 | /// scatter or just simple gather. | |||
3847 | static LoadsState canVectorizeLoads(ArrayRef<Value *> VL, const Value *VL0, | |||
3848 | const TargetTransformInfo &TTI, | |||
3849 | const DataLayout &DL, ScalarEvolution &SE, | |||
3850 | LoopInfo &LI, const TargetLibraryInfo &TLI, | |||
3851 | SmallVectorImpl<unsigned> &Order, | |||
3852 | SmallVectorImpl<Value *> &PointerOps) { | |||
3853 | // Check that a vectorized load would load the same memory as a scalar | |||
3854 | // load. For example, we don't want to vectorize loads that are smaller | |||
3855 | // than 8-bit. Even though we have a packed struct {<i2, i2, i2, i2>} LLVM | |||
3856 | // treats loading/storing it as an i8 struct. If we vectorize loads/stores | |||
3857 | // from such a struct, we read/write packed bits disagreeing with the | |||
3858 | // unvectorized version. | |||
3859 | Type *ScalarTy = VL0->getType(); | |||
3860 | ||||
3861 | if (DL.getTypeSizeInBits(ScalarTy) != DL.getTypeAllocSizeInBits(ScalarTy)) | |||
3862 | return LoadsState::Gather; | |||
3863 | ||||
3864 | // Make sure all loads in the bundle are simple - we can't vectorize | |||
3865 | // atomic or volatile loads. | |||
3866 | PointerOps.clear(); | |||
3867 | PointerOps.resize(VL.size()); | |||
3868 | auto *POIter = PointerOps.begin(); | |||
3869 | for (Value *V : VL) { | |||
3870 | auto *L = cast<LoadInst>(V); | |||
3871 | if (!L->isSimple()) | |||
3872 | return LoadsState::Gather; | |||
3873 | *POIter = L->getPointerOperand(); | |||
3874 | ++POIter; | |||
3875 | } | |||
3876 | ||||
3877 | Order.clear(); | |||
3878 | // Check the order of pointer operands or that all pointers are the same. | |||
3879 | bool IsSorted = sortPtrAccesses(PointerOps, ScalarTy, DL, SE, Order); | |||
3880 | if (IsSorted || all_of(PointerOps, [&](Value *P) { | |||
3881 | return arePointersCompatible(P, PointerOps.front(), TLI); | |||
3882 | })) { | |||
3883 | if (IsSorted) { | |||
3884 | Value *Ptr0; | |||
3885 | Value *PtrN; | |||
3886 | if (Order.empty()) { | |||
3887 | Ptr0 = PointerOps.front(); | |||
3888 | PtrN = PointerOps.back(); | |||
3889 | } else { | |||
3890 | Ptr0 = PointerOps[Order.front()]; | |||
3891 | PtrN = PointerOps[Order.back()]; | |||
3892 | } | |||
3893 | std::optional<int> Diff = | |||
3894 | getPointersDiff(ScalarTy, Ptr0, ScalarTy, PtrN, DL, SE); | |||
3895 | // Check that the sorted loads are consecutive. | |||
3896 | if (static_cast<unsigned>(*Diff) == VL.size() - 1) | |||
3897 | return LoadsState::Vectorize; | |||
3898 | } | |||
3899 | // TODO: need to improve analysis of the pointers, if not all of them are | |||
3900 | // GEPs or have > 2 operands, we end up with a gather node, which just | |||
3901 | // increases the cost. | |||
3902 | Loop *L = LI.getLoopFor(cast<LoadInst>(VL0)->getParent()); | |||
3903 | bool ProfitableGatherPointers = | |||
3904 | static_cast<unsigned>(count_if(PointerOps, [L](Value *V) { | |||
3905 | return L && L->isLoopInvariant(V); | |||
3906 | })) <= VL.size() / 2 && VL.size() > 2; | |||
3907 | if (ProfitableGatherPointers || all_of(PointerOps, [IsSorted](Value *P) { | |||
3908 | auto *GEP = dyn_cast<GetElementPtrInst>(P); | |||
3909 | return (IsSorted && !GEP && doesNotNeedToBeScheduled(P)) || | |||
3910 | (GEP && GEP->getNumOperands() == 2); | |||
3911 | })) { | |||
3912 | Align CommonAlignment = cast<LoadInst>(VL0)->getAlign(); | |||
3913 | for (Value *V : VL) | |||
3914 | CommonAlignment = | |||
3915 | std::min(CommonAlignment, cast<LoadInst>(V)->getAlign()); | |||
3916 | auto *VecTy = FixedVectorType::get(ScalarTy, VL.size()); | |||
3917 | if (TTI.isLegalMaskedGather(VecTy, CommonAlignment) && | |||
3918 | !TTI.forceScalarizeMaskedGather(VecTy, CommonAlignment)) | |||
3919 | return LoadsState::ScatterVectorize; | |||
3920 | } | |||
3921 | } | |||
3922 | ||||
3923 | return LoadsState::Gather; | |||
3924 | } | |||
3925 | ||||
3926 | static bool clusterSortPtrAccesses(ArrayRef<Value *> VL, Type *ElemTy, | |||
3927 | const DataLayout &DL, ScalarEvolution &SE, | |||
3928 | SmallVectorImpl<unsigned> &SortedIndices) { | |||
3929 | assert(llvm::all_of((static_cast <bool> (llvm::all_of( VL, [](const Value * V) { return V->getType()->isPointerTy(); }) && "Expected list of pointer operands." ) ? void (0) : __assert_fail ("llvm::all_of( VL, [](const Value *V) { return V->getType()->isPointerTy(); }) && \"Expected list of pointer operands.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 3931, __extension__ __PRETTY_FUNCTION__)) | |||
3930 | VL, [](const Value *V) { return V->getType()->isPointerTy(); }) &&(static_cast <bool> (llvm::all_of( VL, [](const Value * V) { return V->getType()->isPointerTy(); }) && "Expected list of pointer operands." ) ? void (0) : __assert_fail ("llvm::all_of( VL, [](const Value *V) { return V->getType()->isPointerTy(); }) && \"Expected list of pointer operands.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 3931, __extension__ __PRETTY_FUNCTION__)) | |||
3931 | "Expected list of pointer operands.")(static_cast <bool> (llvm::all_of( VL, [](const Value * V) { return V->getType()->isPointerTy(); }) && "Expected list of pointer operands." ) ? void (0) : __assert_fail ("llvm::all_of( VL, [](const Value *V) { return V->getType()->isPointerTy(); }) && \"Expected list of pointer operands.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 3931, __extension__ __PRETTY_FUNCTION__)); | |||
3932 | // Map from bases to a vector of (Ptr, Offset, OrigIdx), which we insert each | |||
3933 | // Ptr into, sort and return the sorted indices with values next to one | |||
3934 | // another. | |||
3935 | MapVector<Value *, SmallVector<std::tuple<Value *, int, unsigned>>> Bases; | |||
3936 | Bases[VL[0]].push_back(std::make_tuple(VL[0], 0U, 0U)); | |||
3937 | ||||
3938 | unsigned Cnt = 1; | |||
3939 | for (Value *Ptr : VL.drop_front()) { | |||
3940 | bool Found = any_of(Bases, [&](auto &Base) { | |||
3941 | std::optional<int> Diff = | |||
3942 | getPointersDiff(ElemTy, Base.first, ElemTy, Ptr, DL, SE, | |||
3943 | /*StrictCheck=*/true); | |||
3944 | if (!Diff) | |||
3945 | return false; | |||
3946 | ||||
3947 | Base.second.emplace_back(Ptr, *Diff, Cnt++); | |||
3948 | return true; | |||
3949 | }); | |||
3950 | ||||
3951 | if (!Found) { | |||
3952 | // If we haven't found enough to usefully cluster, return early. | |||
3953 | if (Bases.size() > VL.size() / 2 - 1) | |||
3954 | return false; | |||
3955 | ||||
3956 | // Not found already - add a new Base | |||
3957 | Bases[Ptr].emplace_back(Ptr, 0, Cnt++); | |||
3958 | } | |||
3959 | } | |||
3960 | ||||
3961 | // For each of the bases sort the pointers by Offset and check if any of the | |||
3962 | // base become consecutively allocated. | |||
3963 | bool AnyConsecutive = false; | |||
3964 | for (auto &Base : Bases) { | |||
3965 | auto &Vec = Base.second; | |||
3966 | if (Vec.size() > 1) { | |||
3967 | llvm::stable_sort(Vec, [](const std::tuple<Value *, int, unsigned> &X, | |||
3968 | const std::tuple<Value *, int, unsigned> &Y) { | |||
3969 | return std::get<1>(X) < std::get<1>(Y); | |||
3970 | }); | |||
3971 | int InitialOffset = std::get<1>(Vec[0]); | |||
3972 | AnyConsecutive |= all_of(enumerate(Vec), [InitialOffset](const auto &P) { | |||
3973 | return std::get<1>(P.value()) == int(P.index()) + InitialOffset; | |||
3974 | }); | |||
3975 | } | |||
3976 | } | |||
3977 | ||||
3978 | // Fill SortedIndices array only if it looks worth-while to sort the ptrs. | |||
3979 | SortedIndices.clear(); | |||
3980 | if (!AnyConsecutive) | |||
3981 | return false; | |||
3982 | ||||
3983 | for (auto &Base : Bases) { | |||
3984 | for (auto &T : Base.second) | |||
3985 | SortedIndices.push_back(std::get<2>(T)); | |||
3986 | } | |||
3987 | ||||
3988 | assert(SortedIndices.size() == VL.size() &&(static_cast <bool> (SortedIndices.size() == VL.size() && "Expected SortedIndices to be the size of VL") ? void (0) : __assert_fail ("SortedIndices.size() == VL.size() && \"Expected SortedIndices to be the size of VL\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 3989, __extension__ __PRETTY_FUNCTION__)) | |||
3989 | "Expected SortedIndices to be the size of VL")(static_cast <bool> (SortedIndices.size() == VL.size() && "Expected SortedIndices to be the size of VL") ? void (0) : __assert_fail ("SortedIndices.size() == VL.size() && \"Expected SortedIndices to be the size of VL\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 3989, __extension__ __PRETTY_FUNCTION__)); | |||
3990 | return true; | |||
3991 | } | |||
3992 | ||||
3993 | std::optional<BoUpSLP::OrdersType> | |||
3994 | BoUpSLP::findPartiallyOrderedLoads(const BoUpSLP::TreeEntry &TE) { | |||
3995 | 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", 3995, __extension__ __PRETTY_FUNCTION__)); | |||
3996 | Type *ScalarTy = TE.Scalars[0]->getType(); | |||
3997 | ||||
3998 | SmallVector<Value *> Ptrs; | |||
3999 | Ptrs.reserve(TE.Scalars.size()); | |||
4000 | for (Value *V : TE.Scalars) { | |||
4001 | auto *L = dyn_cast<LoadInst>(V); | |||
4002 | if (!L || !L->isSimple()) | |||
4003 | return std::nullopt; | |||
4004 | Ptrs.push_back(L->getPointerOperand()); | |||
4005 | } | |||
4006 | ||||
4007 | BoUpSLP::OrdersType Order; | |||
4008 | if (clusterSortPtrAccesses(Ptrs, ScalarTy, *DL, *SE, Order)) | |||
4009 | return std::move(Order); | |||
4010 | return std::nullopt; | |||
4011 | } | |||
4012 | ||||
4013 | /// Check if two insertelement instructions are from the same buildvector. | |||
4014 | static bool areTwoInsertFromSameBuildVector( | |||
4015 | InsertElementInst *VU, InsertElementInst *V, | |||
4016 | function_ref<Value *(InsertElementInst *)> GetBaseOperand) { | |||
4017 | // Instructions must be from the same basic blocks. | |||
4018 | if (VU->getParent() != V->getParent()) | |||
4019 | return false; | |||
4020 | // Checks if 2 insertelements are from the same buildvector. | |||
4021 | if (VU->getType() != V->getType()) | |||
4022 | return false; | |||
4023 | // Multiple used inserts are separate nodes. | |||
4024 | if (!VU->hasOneUse() && !V->hasOneUse()) | |||
4025 | return false; | |||
4026 | auto *IE1 = VU; | |||
4027 | auto *IE2 = V; | |||
4028 | std::optional<unsigned> Idx1 = getInsertIndex(IE1); | |||
4029 | std::optional<unsigned> Idx2 = getInsertIndex(IE2); | |||
4030 | if (Idx1 == std::nullopt || Idx2 == std::nullopt) | |||
4031 | return false; | |||
4032 | // Go through the vector operand of insertelement instructions trying to find | |||
4033 | // either VU as the original vector for IE2 or V as the original vector for | |||
4034 | // IE1. | |||
4035 | SmallSet<int, 8> ReusedIdx; | |||
4036 | bool IsReusedIdx = false; | |||
4037 | do { | |||
4038 | if (IE2 == VU && !IE1) | |||
4039 | return VU->hasOneUse(); | |||
4040 | if (IE1 == V && !IE2) | |||
4041 | return V->hasOneUse(); | |||
4042 | if (IE1 && IE1 != V) { | |||
4043 | IsReusedIdx |= | |||
4044 | !ReusedIdx.insert(getInsertIndex(IE1).value_or(*Idx2)).second; | |||
4045 | if ((IE1 != VU && !IE1->hasOneUse()) || IsReusedIdx) | |||
4046 | IE1 = nullptr; | |||
4047 | else | |||
4048 | IE1 = dyn_cast_or_null<InsertElementInst>(GetBaseOperand(IE1)); | |||
4049 | } | |||
4050 | if (IE2 && IE2 != VU) { | |||
4051 | IsReusedIdx |= | |||
4052 | !ReusedIdx.insert(getInsertIndex(IE2).value_or(*Idx1)).second; | |||
4053 | if ((IE2 != V && !IE2->hasOneUse()) || IsReusedIdx) | |||
4054 | IE2 = nullptr; | |||
4055 | else | |||
4056 | IE2 = dyn_cast_or_null<InsertElementInst>(GetBaseOperand(IE2)); | |||
4057 | } | |||
4058 | } while (!IsReusedIdx && (IE1 || IE2)); | |||
4059 | return false; | |||
4060 | } | |||
4061 | ||||
4062 | std::optional<BoUpSLP::OrdersType> | |||
4063 | BoUpSLP::getReorderingData(const TreeEntry &TE, bool TopToBottom) { | |||
4064 | // No need to reorder if need to shuffle reuses, still need to shuffle the | |||
4065 | // node. | |||
4066 | if (!TE.ReuseShuffleIndices.empty()) { | |||
4067 | // Check if reuse shuffle indices can be improved by reordering. | |||
4068 | // For this, check that reuse mask is "clustered", i.e. each scalar values | |||
4069 | // is used once in each submask of size <number_of_scalars>. | |||
4070 | // Example: 4 scalar values. | |||
4071 | // ReuseShuffleIndices mask: 0, 1, 2, 3, 3, 2, 0, 1 - clustered. | |||
4072 | // 0, 1, 2, 3, 3, 3, 1, 0 - not clustered, because | |||
4073 | // element 3 is used twice in the second submask. | |||
4074 | unsigned Sz = TE.Scalars.size(); | |||
4075 | if (!ShuffleVectorInst::isOneUseSingleSourceMask(TE.ReuseShuffleIndices, | |||
4076 | Sz)) | |||
4077 | return std::nullopt; | |||
4078 | unsigned VF = TE.getVectorFactor(); | |||
4079 | // Try build correct order for extractelement instructions. | |||
4080 | SmallVector<int> ReusedMask(TE.ReuseShuffleIndices.begin(), | |||
4081 | TE.ReuseShuffleIndices.end()); | |||
4082 | if (TE.getOpcode() == Instruction::ExtractElement && !TE.isAltShuffle() && | |||
4083 | all_of(TE.Scalars, [Sz](Value *V) { | |||
4084 | std::optional<unsigned> Idx = getExtractIndex(cast<Instruction>(V)); | |||
4085 | return Idx && *Idx < Sz; | |||
4086 | })) { | |||
4087 | SmallVector<int> ReorderMask(Sz, PoisonMaskElem); | |||
4088 | if (TE.ReorderIndices.empty()) | |||
4089 | std::iota(ReorderMask.begin(), ReorderMask.end(), 0); | |||
4090 | else | |||
4091 | inversePermutation(TE.ReorderIndices, ReorderMask); | |||
4092 | for (unsigned I = 0; I < VF; ++I) { | |||
4093 | int &Idx = ReusedMask[I]; | |||
4094 | if (Idx == PoisonMaskElem) | |||
4095 | continue; | |||
4096 | Value *V = TE.Scalars[ReorderMask[Idx]]; | |||
4097 | std::optional<unsigned> EI = getExtractIndex(cast<Instruction>(V)); | |||
4098 | Idx = std::distance(ReorderMask.begin(), find(ReorderMask, *EI)); | |||
4099 | } | |||
4100 | } | |||
4101 | // Build the order of the VF size, need to reorder reuses shuffles, they are | |||
4102 | // always of VF size. | |||
4103 | OrdersType ResOrder(VF); | |||
4104 | std::iota(ResOrder.begin(), ResOrder.end(), 0); | |||
4105 | auto *It = ResOrder.begin(); | |||
4106 | for (unsigned K = 0; K < VF; K += Sz) { | |||
4107 | OrdersType CurrentOrder(TE.ReorderIndices); | |||
4108 | SmallVector<int> SubMask{ArrayRef(ReusedMask).slice(K, Sz)}; | |||
4109 | if (SubMask.front() == PoisonMaskElem) | |||
4110 | std::iota(SubMask.begin(), SubMask.end(), 0); | |||
4111 | reorderOrder(CurrentOrder, SubMask); | |||
4112 | transform(CurrentOrder, It, [K](unsigned Pos) { return Pos + K; }); | |||
4113 | std::advance(It, Sz); | |||
4114 | } | |||
4115 | if (all_of(enumerate(ResOrder), | |||
4116 | [](const auto &Data) { return Data.index() == Data.value(); })) | |||
4117 | return std::nullopt; // No need to reorder. | |||
4118 | return std::move(ResOrder); | |||
4119 | } | |||
4120 | if (TE.State == TreeEntry::Vectorize && | |||
4121 | (isa<LoadInst, ExtractElementInst, ExtractValueInst>(TE.getMainOp()) || | |||
4122 | (TopToBottom && isa<StoreInst, InsertElementInst>(TE.getMainOp()))) && | |||
4123 | !TE.isAltShuffle()) | |||
4124 | return TE.ReorderIndices; | |||
4125 | if (TE.State == TreeEntry::Vectorize && TE.getOpcode() == Instruction::PHI) { | |||
4126 | auto PHICompare = [](llvm::Value *V1, llvm::Value *V2) { | |||
4127 | if (!V1->hasOneUse() || !V2->hasOneUse()) | |||
4128 | return false; | |||
4129 | auto *FirstUserOfPhi1 = cast<Instruction>(*V1->user_begin()); | |||
4130 | auto *FirstUserOfPhi2 = cast<Instruction>(*V2->user_begin()); | |||
4131 | if (auto *IE1 = dyn_cast<InsertElementInst>(FirstUserOfPhi1)) | |||
4132 | if (auto *IE2 = dyn_cast<InsertElementInst>(FirstUserOfPhi2)) { | |||
4133 | if (!areTwoInsertFromSameBuildVector( | |||
4134 | IE1, IE2, | |||
4135 | [](InsertElementInst *II) { return II->getOperand(0); })) | |||
4136 | return false; | |||
4137 | std::optional<unsigned> Idx1 = getInsertIndex(IE1); | |||
4138 | std::optional<unsigned> Idx2 = getInsertIndex(IE2); | |||
4139 | if (Idx1 == std::nullopt || Idx2 == std::nullopt) | |||
4140 | return false; | |||
4141 | return *Idx1 < *Idx2; | |||
4142 | } | |||
4143 | if (auto *EE1 = dyn_cast<ExtractElementInst>(FirstUserOfPhi1)) | |||
4144 | if (auto *EE2 = dyn_cast<ExtractElementInst>(FirstUserOfPhi2)) { | |||
4145 | if (EE1->getOperand(0) != EE2->getOperand(0)) | |||
4146 | return false; | |||
4147 | std::optional<unsigned> Idx1 = getExtractIndex(EE1); | |||
4148 | std::optional<unsigned> Idx2 = getExtractIndex(EE2); | |||
4149 | if (Idx1 == std::nullopt || Idx2 == std::nullopt) | |||
4150 | return false; | |||
4151 | return *Idx1 < *Idx2; | |||
4152 | } | |||
4153 | return false; | |||
4154 | }; | |||
4155 | auto IsIdentityOrder = [](const OrdersType &Order) { | |||
4156 | for (unsigned Idx : seq<unsigned>(0, Order.size())) | |||
4157 | if (Idx != Order[Idx]) | |||
4158 | return false; | |||
4159 | return true; | |||
4160 | }; | |||
4161 | if (!TE.ReorderIndices.empty()) | |||
4162 | return TE.ReorderIndices; | |||
4163 | DenseMap<Value *, unsigned> PhiToId; | |||
4164 | SmallVector<Value *, 4> Phis; | |||
4165 | OrdersType ResOrder(TE.Scalars.size()); | |||
4166 | for (unsigned Id = 0, Sz = TE.Scalars.size(); Id < Sz; ++Id) { | |||
4167 | PhiToId[TE.Scalars[Id]] = Id; | |||
4168 | Phis.push_back(TE.Scalars[Id]); | |||
4169 | } | |||
4170 | llvm::stable_sort(Phis, PHICompare); | |||
4171 | for (unsigned Id = 0, Sz = Phis.size(); Id < Sz; ++Id) | |||
4172 | ResOrder[Id] = PhiToId[Phis[Id]]; | |||
4173 | if (IsIdentityOrder(ResOrder)) | |||
4174 | return std::nullopt; // No need to reorder. | |||
4175 | return std::move(ResOrder); | |||
4176 | } | |||
4177 | if (TE.State == TreeEntry::NeedToGather) { | |||
4178 | // TODO: add analysis of other gather nodes with extractelement | |||
4179 | // instructions and other values/instructions, not only undefs. | |||
4180 | if (((TE.getOpcode() == Instruction::ExtractElement && | |||
4181 | !TE.isAltShuffle()) || | |||
4182 | (all_of(TE.Scalars, | |||
4183 | [](Value *V) { | |||
4184 | return isa<UndefValue, ExtractElementInst>(V); | |||
4185 | }) && | |||
4186 | any_of(TE.Scalars, | |||
4187 | [](Value *V) { return isa<ExtractElementInst>(V); }))) && | |||
4188 | all_of(TE.Scalars, | |||
4189 | [](Value *V) { | |||
4190 | auto *EE = dyn_cast<ExtractElementInst>(V); | |||
4191 | return !EE || isa<FixedVectorType>(EE->getVectorOperandType()); | |||
4192 | }) && | |||
4193 | allSameType(TE.Scalars)) { | |||
4194 | // Check that gather of extractelements can be represented as | |||
4195 | // just a shuffle of a single vector. | |||
4196 | OrdersType CurrentOrder; | |||
4197 | bool Reuse = canReuseExtract(TE.Scalars, TE.getMainOp(), CurrentOrder); | |||
4198 | if (Reuse || !CurrentOrder.empty()) { | |||
4199 | if (!CurrentOrder.empty()) | |||
4200 | fixupOrderingIndices(CurrentOrder); | |||
4201 | return std::move(CurrentOrder); | |||
4202 | } | |||
4203 | } | |||
4204 | // If the gather node is <undef, v, .., poison> and | |||
4205 | // insertelement poison, v, 0 [+ permute] | |||
4206 | // is cheaper than | |||
4207 | // insertelement poison, v, n - try to reorder. | |||
4208 | // If rotating the whole graph, exclude the permute cost, the whole graph | |||
4209 | // might be transformed. | |||
4210 | int Sz = TE.Scalars.size(); | |||
4211 | if (isSplat(TE.Scalars) && !allConstant(TE.Scalars) && | |||
4212 | count_if(TE.Scalars, UndefValue::classof) == Sz - 1) { | |||
4213 | const auto *It = | |||
4214 | find_if(TE.Scalars, [](Value *V) { return !isConstant(V); }); | |||
4215 | if (It == TE.Scalars.begin()) | |||
4216 | return OrdersType(); | |||
4217 | auto *Ty = FixedVectorType::get(TE.Scalars.front()->getType(), Sz); | |||
4218 | if (It != TE.Scalars.end()) { | |||
4219 | OrdersType Order(Sz, Sz); | |||
4220 | unsigned Idx = std::distance(TE.Scalars.begin(), It); | |||
4221 | Order[Idx] = 0; | |||
4222 | fixupOrderingIndices(Order); | |||
4223 | SmallVector<int> Mask; | |||
4224 | inversePermutation(Order, Mask); | |||
4225 | InstructionCost PermuteCost = | |||
4226 | TopToBottom | |||
4227 | ? 0 | |||
4228 | : TTI->getShuffleCost(TTI::SK_PermuteSingleSrc, Ty, Mask); | |||
4229 | InstructionCost InsertFirstCost = TTI->getVectorInstrCost( | |||
4230 | Instruction::InsertElement, Ty, TTI::TCK_RecipThroughput, 0, | |||
4231 | PoisonValue::get(Ty), *It); | |||
4232 | InstructionCost InsertIdxCost = TTI->getVectorInstrCost( | |||
4233 | Instruction::InsertElement, Ty, TTI::TCK_RecipThroughput, Idx, | |||
4234 | PoisonValue::get(Ty), *It); | |||
4235 | if (InsertFirstCost + PermuteCost < InsertIdxCost) | |||
4236 | return std::move(Order); | |||
4237 | } | |||
4238 | } | |||
4239 | if (std::optional<OrdersType> CurrentOrder = findReusedOrderedScalars(TE)) | |||
4240 | return CurrentOrder; | |||
4241 | if (TE.Scalars.size() >= 4) | |||
4242 | if (std::optional<OrdersType> Order = findPartiallyOrderedLoads(TE)) | |||
4243 | return Order; | |||
4244 | } | |||
4245 | return std::nullopt; | |||
4246 | } | |||
4247 | ||||
4248 | /// Checks if the given mask is a "clustered" mask with the same clusters of | |||
4249 | /// size \p Sz, which are not identity submasks. | |||
4250 | static bool isRepeatedNonIdentityClusteredMask(ArrayRef<int> Mask, | |||
4251 | unsigned Sz) { | |||
4252 | ArrayRef<int> FirstCluster = Mask.slice(0, Sz); | |||
4253 | if (ShuffleVectorInst::isIdentityMask(FirstCluster)) | |||
4254 | return false; | |||
4255 | for (unsigned I = Sz, E = Mask.size(); I < E; I += Sz) { | |||
4256 | ArrayRef<int> Cluster = Mask.slice(I, Sz); | |||
4257 | if (Cluster != FirstCluster) | |||
4258 | return false; | |||
4259 | } | |||
4260 | return true; | |||
4261 | } | |||
4262 | ||||
4263 | void BoUpSLP::reorderNodeWithReuses(TreeEntry &TE, ArrayRef<int> Mask) const { | |||
4264 | // Reorder reuses mask. | |||
4265 | reorderReuses(TE.ReuseShuffleIndices, Mask); | |||
4266 | const unsigned Sz = TE.Scalars.size(); | |||
4267 | // For vectorized and non-clustered reused no need to do anything else. | |||
4268 | if (TE.State != TreeEntry::NeedToGather || | |||
4269 | !ShuffleVectorInst::isOneUseSingleSourceMask(TE.ReuseShuffleIndices, | |||
4270 | Sz) || | |||
4271 | !isRepeatedNonIdentityClusteredMask(TE.ReuseShuffleIndices, Sz)) | |||
4272 | return; | |||
4273 | SmallVector<int> NewMask; | |||
4274 | inversePermutation(TE.ReorderIndices, NewMask); | |||
4275 | addMask(NewMask, TE.ReuseShuffleIndices); | |||
4276 | // Clear reorder since it is going to be applied to the new mask. | |||
4277 | TE.ReorderIndices.clear(); | |||
4278 | // Try to improve gathered nodes with clustered reuses, if possible. | |||
4279 | ArrayRef<int> Slice = ArrayRef(NewMask).slice(0, Sz); | |||
4280 | SmallVector<unsigned> NewOrder(Slice.begin(), Slice.end()); | |||
4281 | inversePermutation(NewOrder, NewMask); | |||
4282 | reorderScalars(TE.Scalars, NewMask); | |||
4283 | // Fill the reuses mask with the identity submasks. | |||
4284 | for (auto *It = TE.ReuseShuffleIndices.begin(), | |||
4285 | *End = TE.ReuseShuffleIndices.end(); | |||
4286 | It != End; std::advance(It, Sz)) | |||
4287 | std::iota(It, std::next(It, Sz), 0); | |||
4288 | } | |||
4289 | ||||
4290 | void BoUpSLP::reorderTopToBottom() { | |||
4291 | // Maps VF to the graph nodes. | |||
4292 | DenseMap<unsigned, SetVector<TreeEntry *>> VFToOrderedEntries; | |||
4293 | // ExtractElement gather nodes which can be vectorized and need to handle | |||
4294 | // their ordering. | |||
4295 | DenseMap<const TreeEntry *, OrdersType> GathersToOrders; | |||
4296 | ||||
4297 | // Phi nodes can have preferred ordering based on their result users | |||
4298 | DenseMap<const TreeEntry *, OrdersType> PhisToOrders; | |||
4299 | ||||
4300 | // AltShuffles can also have a preferred ordering that leads to fewer | |||
4301 | // instructions, e.g., the addsub instruction in x86. | |||
4302 | DenseMap<const TreeEntry *, OrdersType> AltShufflesToOrders; | |||
4303 | ||||
4304 | // Maps a TreeEntry to the reorder indices of external users. | |||
4305 | DenseMap<const TreeEntry *, SmallVector<OrdersType, 1>> | |||
4306 | ExternalUserReorderMap; | |||
4307 | // FIXME: Workaround for syntax error reported by MSVC buildbots. | |||
4308 | TargetTransformInfo &TTIRef = *TTI; | |||
4309 | // Find all reorderable nodes with the given VF. | |||
4310 | // Currently the are vectorized stores,loads,extracts + some gathering of | |||
4311 | // extracts. | |||
4312 | for_each(VectorizableTree, [this, &TTIRef, &VFToOrderedEntries, | |||
4313 | &GathersToOrders, &ExternalUserReorderMap, | |||
4314 | &AltShufflesToOrders, &PhisToOrders]( | |||
4315 | const std::unique_ptr<TreeEntry> &TE) { | |||
4316 | // Look for external users that will probably be vectorized. | |||
4317 | SmallVector<OrdersType, 1> ExternalUserReorderIndices = | |||
4318 | findExternalStoreUsersReorderIndices(TE.get()); | |||
4319 | if (!ExternalUserReorderIndices.empty()) { | |||
4320 | VFToOrderedEntries[TE->getVectorFactor()].insert(TE.get()); | |||
4321 | ExternalUserReorderMap.try_emplace(TE.get(), | |||
4322 | std::move(ExternalUserReorderIndices)); | |||
4323 | } | |||
4324 | ||||
4325 | // Patterns like [fadd,fsub] can be combined into a single instruction in | |||
4326 | // x86. Reordering them into [fsub,fadd] blocks this pattern. So we need | |||
4327 | // to take into account their order when looking for the most used order. | |||
4328 | if (TE->isAltShuffle()) { | |||
4329 | VectorType *VecTy = | |||
4330 | FixedVectorType::get(TE->Scalars[0]->getType(), TE->Scalars.size()); | |||
4331 | unsigned Opcode0 = TE->getOpcode(); | |||
4332 | unsigned Opcode1 = TE->getAltOpcode(); | |||
4333 | // The opcode mask selects between the two opcodes. | |||
4334 | SmallBitVector OpcodeMask(TE->Scalars.size(), false); | |||
4335 | for (unsigned Lane : seq<unsigned>(0, TE->Scalars.size())) | |||
4336 | if (cast<Instruction>(TE->Scalars[Lane])->getOpcode() == Opcode1) | |||
4337 | OpcodeMask.set(Lane); | |||
4338 | // If this pattern is supported by the target then we consider the order. | |||
4339 | if (TTIRef.isLegalAltInstr(VecTy, Opcode0, Opcode1, OpcodeMask)) { | |||
4340 | VFToOrderedEntries[TE->getVectorFactor()].insert(TE.get()); | |||
4341 | AltShufflesToOrders.try_emplace(TE.get(), OrdersType()); | |||
4342 | } | |||
4343 | // TODO: Check the reverse order too. | |||
4344 | } | |||
4345 | ||||
4346 | if (std::optional<OrdersType> CurrentOrder = | |||
4347 | getReorderingData(*TE, /*TopToBottom=*/true)) { | |||
4348 | // Do not include ordering for nodes used in the alt opcode vectorization, | |||
4349 | // better to reorder them during bottom-to-top stage. If follow the order | |||
4350 | // here, it causes reordering of the whole graph though actually it is | |||
4351 | // profitable just to reorder the subgraph that starts from the alternate | |||
4352 | // opcode vectorization node. Such nodes already end-up with the shuffle | |||
4353 | // instruction and it is just enough to change this shuffle rather than | |||
4354 | // rotate the scalars for the whole graph. | |||
4355 | unsigned Cnt = 0; | |||
4356 | const TreeEntry *UserTE = TE.get(); | |||
4357 | while (UserTE && Cnt < RecursionMaxDepth) { | |||
4358 | if (UserTE->UserTreeIndices.size() != 1) | |||
4359 | break; | |||
4360 | if (all_of(UserTE->UserTreeIndices, [](const EdgeInfo &EI) { | |||
4361 | return EI.UserTE->State == TreeEntry::Vectorize && | |||
4362 | EI.UserTE->isAltShuffle() && EI.UserTE->Idx != 0; | |||
4363 | })) | |||
4364 | return; | |||
4365 | UserTE = UserTE->UserTreeIndices.back().UserTE; | |||
4366 | ++Cnt; | |||
4367 | } | |||
4368 | VFToOrderedEntries[TE->getVectorFactor()].insert(TE.get()); | |||
4369 | if (TE->State != TreeEntry::Vectorize || !TE->ReuseShuffleIndices.empty()) | |||
4370 | GathersToOrders.try_emplace(TE.get(), *CurrentOrder); | |||
4371 | if (TE->State == TreeEntry::Vectorize && | |||
4372 | TE->getOpcode() == Instruction::PHI) | |||
4373 | PhisToOrders.try_emplace(TE.get(), *CurrentOrder); | |||
4374 | } | |||
4375 | }); | |||
4376 | ||||
4377 | // Reorder the graph nodes according to their vectorization factor. | |||
4378 | for (unsigned VF = VectorizableTree.front()->getVectorFactor(); VF > 1; | |||
4379 | VF /= 2) { | |||
4380 | auto It = VFToOrderedEntries.find(VF); | |||
4381 | if (It == VFToOrderedEntries.end()) | |||
4382 | continue; | |||
4383 | // Try to find the most profitable order. We just are looking for the most | |||
4384 | // used order and reorder scalar elements in the nodes according to this | |||
4385 | // mostly used order. | |||
4386 | ArrayRef<TreeEntry *> OrderedEntries = It->second.getArrayRef(); | |||
4387 | // All operands are reordered and used only in this node - propagate the | |||
4388 | // most used order to the user node. | |||
4389 | MapVector<OrdersType, unsigned, | |||
4390 | DenseMap<OrdersType, unsigned, OrdersTypeDenseMapInfo>> | |||
4391 | OrdersUses; | |||
4392 | SmallPtrSet<const TreeEntry *, 4> VisitedOps; | |||
4393 | for (const TreeEntry *OpTE : OrderedEntries) { | |||
4394 | // No need to reorder this nodes, still need to extend and to use shuffle, | |||
4395 | // just need to merge reordering shuffle and the reuse shuffle. | |||
4396 | if (!OpTE->ReuseShuffleIndices.empty() && !GathersToOrders.count(OpTE)) | |||
4397 | continue; | |||
4398 | // Count number of orders uses. | |||
4399 | const auto &Order = [OpTE, &GathersToOrders, &AltShufflesToOrders, | |||
4400 | &PhisToOrders]() -> const OrdersType & { | |||
4401 | if (OpTE->State == TreeEntry::NeedToGather || | |||
4402 | !OpTE->ReuseShuffleIndices.empty()) { | |||
4403 | auto It = GathersToOrders.find(OpTE); | |||
4404 | if (It != GathersToOrders.end()) | |||
4405 | return It->second; | |||
4406 | } | |||
4407 | if (OpTE->isAltShuffle()) { | |||
4408 | auto It = AltShufflesToOrders.find(OpTE); | |||
4409 | if (It != AltShufflesToOrders.end()) | |||
4410 | return It->second; | |||
4411 | } | |||
4412 | if (OpTE->State == TreeEntry::Vectorize && | |||
4413 | OpTE->getOpcode() == Instruction::PHI) { | |||
4414 | auto It = PhisToOrders.find(OpTE); | |||
4415 | if (It != PhisToOrders.end()) | |||
4416 | return It->second; | |||
4417 | } | |||
4418 | return OpTE->ReorderIndices; | |||
4419 | }(); | |||
4420 | // First consider the order of the external scalar users. | |||
4421 | auto It = ExternalUserReorderMap.find(OpTE); | |||
4422 | if (It != ExternalUserReorderMap.end()) { | |||
4423 | const auto &ExternalUserReorderIndices = It->second; | |||
4424 | // If the OpTE vector factor != number of scalars - use natural order, | |||
4425 | // it is an attempt to reorder node with reused scalars but with | |||
4426 | // external uses. | |||
4427 | if (OpTE->getVectorFactor() != OpTE->Scalars.size()) { | |||
4428 | OrdersUses.insert(std::make_pair(OrdersType(), 0)).first->second += | |||
4429 | ExternalUserReorderIndices.size(); | |||
4430 | } else { | |||
4431 | for (const OrdersType &ExtOrder : ExternalUserReorderIndices) | |||
4432 | ++OrdersUses.insert(std::make_pair(ExtOrder, 0)).first->second; | |||
4433 | } | |||
4434 | // No other useful reorder data in this entry. | |||
4435 | if (Order.empty()) | |||
4436 | continue; | |||
4437 | } | |||
4438 | // Stores actually store the mask, not the order, need to invert. | |||
4439 | if (OpTE->State == TreeEntry::Vectorize && !OpTE->isAltShuffle() && | |||
4440 | OpTE->getOpcode() == Instruction::Store && !Order.empty()) { | |||
4441 | SmallVector<int> Mask; | |||
4442 | inversePermutation(Order, Mask); | |||
4443 | unsigned E = Order.size(); | |||
4444 | OrdersType CurrentOrder(E, E); | |||
4445 | transform(Mask, CurrentOrder.begin(), [E](int Idx) { | |||
4446 | return Idx == PoisonMaskElem ? E : static_cast<unsigned>(Idx); | |||
4447 | }); | |||
4448 | fixupOrderingIndices(CurrentOrder); | |||
4449 | ++OrdersUses.insert(std::make_pair(CurrentOrder, 0)).first->second; | |||
4450 | } else { | |||
4451 | ++OrdersUses.insert(std::make_pair(Order, 0)).first->second; | |||
4452 | } | |||
4453 | } | |||
4454 | // Set order of the user node. | |||
4455 | if (OrdersUses.empty()) | |||
4456 | continue; | |||
4457 | // Choose the most used order. | |||
4458 | ArrayRef<unsigned> BestOrder = OrdersUses.front().first; | |||
4459 | unsigned Cnt = OrdersUses.front().second; | |||
4460 | for (const auto &Pair : drop_begin(OrdersUses)) { | |||
4461 | if (Cnt < Pair.second || (Cnt == Pair.second && Pair.first.empty())) { | |||
4462 | BestOrder = Pair.first; | |||
4463 | Cnt = Pair.second; | |||
4464 | } | |||
4465 | } | |||
4466 | // Set order of the user node. | |||
4467 | if (BestOrder.empty()) | |||
4468 | continue; | |||
4469 | SmallVector<int> Mask; | |||
4470 | inversePermutation(BestOrder, Mask); | |||
4471 | SmallVector<int> MaskOrder(BestOrder.size(), PoisonMaskElem); | |||
4472 | unsigned E = BestOrder.size(); | |||
4473 | transform(BestOrder, MaskOrder.begin(), [E](unsigned I) { | |||
4474 | return I < E ? static_cast<int>(I) : PoisonMaskElem; | |||
4475 | }); | |||
4476 | // Do an actual reordering, if profitable. | |||
4477 | for (std::unique_ptr<TreeEntry> &TE : VectorizableTree) { | |||
4478 | // Just do the reordering for the nodes with the given VF. | |||
4479 | if (TE->Scalars.size() != VF) { | |||
4480 | if (TE->ReuseShuffleIndices.size() == VF) { | |||
4481 | // Need to reorder the reuses masks of the operands with smaller VF to | |||
4482 | // be able to find the match between the graph nodes and scalar | |||
4483 | // operands of the given node during vectorization/cost estimation. | |||
4484 | 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", 4490, __extension__ __PRETTY_FUNCTION__)) | |||
4485 | [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", 4490, __extension__ __PRETTY_FUNCTION__)) | |||
4486 | 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", 4490, __extension__ __PRETTY_FUNCTION__)) | |||
4487 | 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", 4490, __extension__ __PRETTY_FUNCTION__)) | |||
4488 | 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", 4490, __extension__ __PRETTY_FUNCTION__)) | |||
4489 | }) &&(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", 4490, __extension__ __PRETTY_FUNCTION__)) | |||
4490 | "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", 4490, __extension__ __PRETTY_FUNCTION__)); | |||
4491 | // Update ordering of the operands with the smaller VF than the given | |||
4492 | // one. | |||
4493 | reorderNodeWithReuses(*TE, Mask); | |||
4494 | } | |||
4495 | continue; | |||
4496 | } | |||
4497 | if (TE->State == TreeEntry::Vectorize && | |||
4498 | isa<ExtractElementInst, ExtractValueInst, LoadInst, StoreInst, | |||
4499 | InsertElementInst>(TE->getMainOp()) && | |||
4500 | !TE->isAltShuffle()) { | |||
4501 | // Build correct orders for extract{element,value}, loads and | |||
4502 | // stores. | |||
4503 | reorderOrder(TE->ReorderIndices, Mask); | |||
4504 | if (isa<InsertElementInst, StoreInst>(TE->getMainOp())) | |||
4505 | TE->reorderOperands(Mask); | |||
4506 | } else { | |||
4507 | // Reorder the node and its operands. | |||
4508 | TE->reorderOperands(Mask); | |||
4509 | 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", 4510, __extension__ __PRETTY_FUNCTION__)) | |||
4510 | "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", 4510, __extension__ __PRETTY_FUNCTION__)); | |||
4511 | reorderScalars(TE->Scalars, Mask); | |||
4512 | } | |||
4513 | if (!TE->ReuseShuffleIndices.empty()) { | |||
4514 | // Apply reversed order to keep the original ordering of the reused | |||
4515 | // elements to avoid extra reorder indices shuffling. | |||
4516 | OrdersType CurrentOrder; | |||
4517 | reorderOrder(CurrentOrder, MaskOrder); | |||
4518 | SmallVector<int> NewReuses; | |||
4519 | inversePermutation(CurrentOrder, NewReuses); | |||
4520 | addMask(NewReuses, TE->ReuseShuffleIndices); | |||
4521 | TE->ReuseShuffleIndices.swap(NewReuses); | |||
4522 | } | |||
4523 | } | |||
4524 | } | |||
4525 | } | |||
4526 | ||||
4527 | bool BoUpSLP::canReorderOperands( | |||
4528 | TreeEntry *UserTE, SmallVectorImpl<std::pair<unsigned, TreeEntry *>> &Edges, | |||
4529 | ArrayRef<TreeEntry *> ReorderableGathers, | |||
4530 | SmallVectorImpl<TreeEntry *> &GatherOps) { | |||
4531 | for (unsigned I = 0, E = UserTE->getNumOperands(); I < E; ++I) { | |||
4532 | if (any_of(Edges, [I](const std::pair<unsigned, TreeEntry *> &OpData) { | |||
4533 | return OpData.first == I && | |||
4534 | OpData.second->State == TreeEntry::Vectorize; | |||
4535 | })) | |||
4536 | continue; | |||
4537 | if (TreeEntry *TE = getVectorizedOperand(UserTE, I)) { | |||
4538 | // Do not reorder if operand node is used by many user nodes. | |||
4539 | if (any_of(TE->UserTreeIndices, | |||
4540 | [UserTE](const EdgeInfo &EI) { return EI.UserTE != UserTE; })) | |||
4541 | return false; | |||
4542 | // Add the node to the list of the ordered nodes with the identity | |||
4543 | // order. | |||
4544 | Edges.emplace_back(I, TE); | |||
4545 | // Add ScatterVectorize nodes to the list of operands, where just | |||
4546 | // reordering of the scalars is required. Similar to the gathers, so | |||
4547 | // simply add to the list of gathered ops. | |||
4548 | // If there are reused scalars, process this node as a regular vectorize | |||
4549 | // node, just reorder reuses mask. | |||
4550 | if (TE->State != TreeEntry::Vectorize && TE->ReuseShuffleIndices.empty()) | |||
4551 | GatherOps.push_back(TE); | |||
4552 | continue; | |||
4553 | } | |||
4554 | TreeEntry *Gather = nullptr; | |||
4555 | if (count_if(ReorderableGathers, | |||
4556 | [&Gather, UserTE, I](TreeEntry *TE) { | |||
4557 | 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", 4558, __extension__ __PRETTY_FUNCTION__)) | |||
4558 | "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", 4558, __extension__ __PRETTY_FUNCTION__)); | |||
4559 | if (any_of(TE->UserTreeIndices, | |||
4560 | [UserTE, I](const EdgeInfo &EI) { | |||
4561 | return EI.UserTE == UserTE && EI.EdgeIdx == I; | |||
4562 | })) { | |||
4563 | assert(TE->isSame(UserTE->getOperand(I)) &&(static_cast <bool> (TE->isSame(UserTE->getOperand (I)) && "Operand entry does not match operands.") ? void (0) : __assert_fail ("TE->isSame(UserTE->getOperand(I)) && \"Operand entry does not match operands.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 4564, __extension__ __PRETTY_FUNCTION__)) | |||
4564 | "Operand entry does not match operands.")(static_cast <bool> (TE->isSame(UserTE->getOperand (I)) && "Operand entry does not match operands.") ? void (0) : __assert_fail ("TE->isSame(UserTE->getOperand(I)) && \"Operand entry does not match operands.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 4564, __extension__ __PRETTY_FUNCTION__)); | |||
4565 | Gather = TE; | |||
4566 | return true; | |||
4567 | } | |||
4568 | return false; | |||
4569 | }) > 1 && | |||
4570 | !allConstant(UserTE->getOperand(I))) | |||
4571 | return false; | |||
4572 | if (Gather) | |||
4573 | GatherOps.push_back(Gather); | |||
4574 | } | |||
4575 | return true; | |||
4576 | } | |||
4577 | ||||
4578 | void BoUpSLP::reorderBottomToTop(bool IgnoreReorder) { | |||
4579 | SetVector<TreeEntry *> OrderedEntries; | |||
4580 | DenseMap<const TreeEntry *, OrdersType> GathersToOrders; | |||
4581 | // Find all reorderable leaf nodes with the given VF. | |||
4582 | // Currently the are vectorized loads,extracts without alternate operands + | |||
4583 | // some gathering of extracts. | |||
4584 | SmallVector<TreeEntry *> NonVectorized; | |||
4585 | for_each(VectorizableTree, [this, &OrderedEntries, &GathersToOrders, | |||
4586 | &NonVectorized]( | |||
4587 | const std::unique_ptr<TreeEntry> &TE) { | |||
4588 | if (TE->State != TreeEntry::Vectorize) | |||
4589 | NonVectorized.push_back(TE.get()); | |||
4590 | if (std::optional<OrdersType> CurrentOrder = | |||
4591 | getReorderingData(*TE, /*TopToBottom=*/false)) { | |||
4592 | OrderedEntries.insert(TE.get()); | |||
4593 | if (TE->State != TreeEntry::Vectorize || !TE->ReuseShuffleIndices.empty()) | |||
4594 | GathersToOrders.try_emplace(TE.get(), *CurrentOrder); | |||
4595 | } | |||
4596 | }); | |||
4597 | ||||
4598 | // 1. Propagate order to the graph nodes, which use only reordered nodes. | |||
4599 | // I.e., if the node has operands, that are reordered, try to make at least | |||
4600 | // one operand order in the natural order and reorder others + reorder the | |||
4601 | // user node itself. | |||
4602 | SmallPtrSet<const TreeEntry *, 4> Visited; | |||
4603 | while (!OrderedEntries.empty()) { | |||
4604 | // 1. Filter out only reordered nodes. | |||
4605 | // 2. If the entry has multiple uses - skip it and jump to the next node. | |||
4606 | DenseMap<TreeEntry *, SmallVector<std::pair<unsigned, TreeEntry *>>> Users; | |||
4607 | SmallVector<TreeEntry *> Filtered; | |||
4608 | for (TreeEntry *TE : OrderedEntries) { | |||
4609 | if (!(TE->State == TreeEntry::Vectorize || | |||
4610 | (TE->State == TreeEntry::NeedToGather && | |||
4611 | GathersToOrders.count(TE))) || | |||
4612 | TE->UserTreeIndices.empty() || !TE->ReuseShuffleIndices.empty() || | |||
4613 | !all_of(drop_begin(TE->UserTreeIndices), | |||
4614 | [TE](const EdgeInfo &EI) { | |||
4615 | return EI.UserTE == TE->UserTreeIndices.front().UserTE; | |||
4616 | }) || | |||
4617 | !Visited.insert(TE).second) { | |||
4618 | Filtered.push_back(TE); | |||
4619 | continue; | |||
4620 | } | |||
4621 | // Build a map between user nodes and their operands order to speedup | |||
4622 | // search. The graph currently does not provide this dependency directly. | |||
4623 | for (EdgeInfo &EI : TE->UserTreeIndices) { | |||
4624 | TreeEntry *UserTE = EI.UserTE; | |||
4625 | auto It = Users.find(UserTE); | |||
4626 | if (It == Users.end()) | |||
4627 | It = Users.insert({UserTE, {}}).first; | |||
4628 | It->second.emplace_back(EI.EdgeIdx, TE); | |||
4629 | } | |||
4630 | } | |||
4631 | // Erase filtered entries. | |||
4632 | for_each(Filtered, | |||
4633 | [&OrderedEntries](TreeEntry *TE) { OrderedEntries.remove(TE); }); | |||
4634 | SmallVector< | |||
4635 | std::pair<TreeEntry *, SmallVector<std::pair<unsigned, TreeEntry *>>>> | |||
4636 | UsersVec(Users.begin(), Users.end()); | |||
4637 | sort(UsersVec, [](const auto &Data1, const auto &Data2) { | |||
4638 | return Data1.first->Idx > Data2.first->Idx; | |||
4639 | }); | |||
4640 | for (auto &Data : UsersVec) { | |||
4641 | // Check that operands are used only in the User node. | |||
4642 | SmallVector<TreeEntry *> GatherOps; | |||
4643 | if (!canReorderOperands(Data.first, Data.second, NonVectorized, | |||
4644 | GatherOps)) { | |||
4645 | for_each(Data.second, | |||
4646 | [&OrderedEntries](const std::pair<unsigned, TreeEntry *> &Op) { | |||
4647 | OrderedEntries.remove(Op.second); | |||
4648 | }); | |||
4649 | continue; | |||
4650 | } | |||
4651 | // All operands are reordered and used only in this node - propagate the | |||
4652 | // most used order to the user node. | |||
4653 | MapVector<OrdersType, unsigned, | |||
4654 | DenseMap<OrdersType, unsigned, OrdersTypeDenseMapInfo>> | |||
4655 | OrdersUses; | |||
4656 | // Do the analysis for each tree entry only once, otherwise the order of | |||
4657 | // the same node my be considered several times, though might be not | |||
4658 | // profitable. | |||
4659 | SmallPtrSet<const TreeEntry *, 4> VisitedOps; | |||
4660 | SmallPtrSet<const TreeEntry *, 4> VisitedUsers; | |||
4661 | for (const auto &Op : Data.second) { | |||
4662 | TreeEntry *OpTE = Op.second; | |||
4663 | if (!VisitedOps.insert(OpTE).second) | |||
4664 | continue; | |||
4665 | if (!OpTE->ReuseShuffleIndices.empty() && !GathersToOrders.count(OpTE)) | |||
4666 | continue; | |||
4667 | const auto &Order = [OpTE, &GathersToOrders]() -> const OrdersType & { | |||
4668 | if (OpTE->State == TreeEntry::NeedToGather || | |||
4669 | !OpTE->ReuseShuffleIndices.empty()) | |||
4670 | return GathersToOrders.find(OpTE)->second; | |||
4671 | return OpTE->ReorderIndices; | |||
4672 | }(); | |||
4673 | unsigned NumOps = count_if( | |||
4674 | Data.second, [OpTE](const std::pair<unsigned, TreeEntry *> &P) { | |||
4675 | return P.second == OpTE; | |||
4676 | }); | |||
4677 | // Stores actually store the mask, not the order, need to invert. | |||
4678 | if (OpTE->State == TreeEntry::Vectorize && !OpTE->isAltShuffle() && | |||
4679 | OpTE->getOpcode() == Instruction::Store && !Order.empty()) { | |||
4680 | SmallVector<int> Mask; | |||
4681 | inversePermutation(Order, Mask); | |||
4682 | unsigned E = Order.size(); | |||
4683 | OrdersType CurrentOrder(E, E); | |||
4684 | transform(Mask, CurrentOrder.begin(), [E](int Idx) { | |||
4685 | return Idx == PoisonMaskElem ? E : static_cast<unsigned>(Idx); | |||
4686 | }); | |||
4687 | fixupOrderingIndices(CurrentOrder); | |||
4688 | OrdersUses.insert(std::make_pair(CurrentOrder, 0)).first->second += | |||
4689 | NumOps; | |||
4690 | } else { | |||
4691 | OrdersUses.insert(std::make_pair(Order, 0)).first->second += NumOps; | |||
4692 | } | |||
4693 | auto Res = OrdersUses.insert(std::make_pair(OrdersType(), 0)); | |||
4694 | const auto &&AllowsReordering = [IgnoreReorder, &GathersToOrders]( | |||
4695 | const TreeEntry *TE) { | |||
4696 | if (!TE->ReorderIndices.empty() || !TE->ReuseShuffleIndices.empty() || | |||
4697 | (TE->State == TreeEntry::Vectorize && TE->isAltShuffle()) || | |||
4698 | (IgnoreReorder && TE->Idx == 0)) | |||
4699 | return true; | |||
4700 | if (TE->State == TreeEntry::NeedToGather) { | |||
4701 | auto It = GathersToOrders.find(TE); | |||
4702 | if (It != GathersToOrders.end()) | |||
4703 | return !It->second.empty(); | |||
4704 | return true; | |||
4705 | } | |||
4706 | return false; | |||
4707 | }; | |||
4708 | for (const EdgeInfo &EI : OpTE->UserTreeIndices) { | |||
4709 | TreeEntry *UserTE = EI.UserTE; | |||
4710 | if (!VisitedUsers.insert(UserTE).second) | |||
4711 | continue; | |||
4712 | // May reorder user node if it requires reordering, has reused | |||
4713 | // scalars, is an alternate op vectorize node or its op nodes require | |||
4714 | // reordering. | |||
4715 | if (AllowsReordering(UserTE)) | |||
4716 | continue; | |||
4717 | // Check if users allow reordering. | |||
4718 | // Currently look up just 1 level of operands to avoid increase of | |||
4719 | // the compile time. | |||
4720 | // Profitable to reorder if definitely more operands allow | |||
4721 | // reordering rather than those with natural order. | |||
4722 | ArrayRef<std::pair<unsigned, TreeEntry *>> Ops = Users[UserTE]; | |||
4723 | if (static_cast<unsigned>(count_if( | |||
4724 | Ops, [UserTE, &AllowsReordering]( | |||
4725 | const std::pair<unsigned, TreeEntry *> &Op) { | |||
4726 | return AllowsReordering(Op.second) && | |||
4727 | all_of(Op.second->UserTreeIndices, | |||
4728 | [UserTE](const EdgeInfo &EI) { | |||
4729 | return EI.UserTE == UserTE; | |||
4730 | }); | |||
4731 | })) <= Ops.size() / 2) | |||
4732 | ++Res.first->second; | |||
4733 | } | |||
4734 | } | |||
4735 | // If no orders - skip current nodes and jump to the next one, if any. | |||
4736 | if (OrdersUses.empty()) { | |||
4737 | for_each(Data.second, | |||
4738 | [&OrderedEntries](const std::pair<unsigned, TreeEntry *> &Op) { | |||
4739 | OrderedEntries.remove(Op.second); | |||
4740 | }); | |||
4741 | continue; | |||
4742 | } | |||
4743 | // Choose the best order. | |||
4744 | ArrayRef<unsigned> BestOrder = OrdersUses.front().first; | |||
4745 | unsigned Cnt = OrdersUses.front().second; | |||
4746 | for (const auto &Pair : drop_begin(OrdersUses)) { | |||
4747 | if (Cnt < Pair.second || (Cnt == Pair.second && Pair.first.empty())) { | |||
4748 | BestOrder = Pair.first; | |||
4749 | Cnt = Pair.second; | |||
4750 | } | |||
4751 | } | |||
4752 | // Set order of the user node (reordering of operands and user nodes). | |||
4753 | if (BestOrder.empty()) { | |||
4754 | for_each(Data.second, | |||
4755 | [&OrderedEntries](const std::pair<unsigned, TreeEntry *> &Op) { | |||
4756 | OrderedEntries.remove(Op.second); | |||
4757 | }); | |||
4758 | continue; | |||
4759 | } | |||
4760 | // Erase operands from OrderedEntries list and adjust their orders. | |||
4761 | VisitedOps.clear(); | |||
4762 | SmallVector<int> Mask; | |||
4763 | inversePermutation(BestOrder, Mask); | |||
4764 | SmallVector<int> MaskOrder(BestOrder.size(), PoisonMaskElem); | |||
4765 | unsigned E = BestOrder.size(); | |||
4766 | transform(BestOrder, MaskOrder.begin(), [E](unsigned I) { | |||
4767 | return I < E ? static_cast<int>(I) : PoisonMaskElem; | |||
4768 | }); | |||
4769 | for (const std::pair<unsigned, TreeEntry *> &Op : Data.second) { | |||
4770 | TreeEntry *TE = Op.second; | |||
4771 | OrderedEntries.remove(TE); | |||
4772 | if (!VisitedOps.insert(TE).second) | |||
4773 | continue; | |||
4774 | if (TE->ReuseShuffleIndices.size() == BestOrder.size()) { | |||
4775 | reorderNodeWithReuses(*TE, Mask); | |||
4776 | continue; | |||
4777 | } | |||
4778 | // Gathers are processed separately. | |||
4779 | if (TE->State != TreeEntry::Vectorize) | |||
4780 | continue; | |||
4781 | 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", 4783, __extension__ __PRETTY_FUNCTION__)) | |||
4782 | 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", 4783, __extension__ __PRETTY_FUNCTION__)) | |||
4783 | "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", 4783, __extension__ __PRETTY_FUNCTION__)); | |||
4784 | reorderOrder(TE->ReorderIndices, Mask); | |||
4785 | if (IgnoreReorder && TE == VectorizableTree.front().get()) | |||
4786 | IgnoreReorder = false; | |||
4787 | } | |||
4788 | // For gathers just need to reorder its scalars. | |||
4789 | for (TreeEntry *Gather : GatherOps) { | |||
4790 | 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", 4791, __extension__ __PRETTY_FUNCTION__)) | |||
4791 | "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", 4791, __extension__ __PRETTY_FUNCTION__)); | |||
4792 | if (!Gather->ReuseShuffleIndices.empty()) { | |||
4793 | // Just reorder reuses indices. | |||
4794 | reorderReuses(Gather->ReuseShuffleIndices, Mask); | |||
4795 | continue; | |||
4796 | } | |||
4797 | reorderScalars(Gather->Scalars, Mask); | |||
4798 | OrderedEntries.remove(Gather); | |||
4799 | } | |||
4800 | // Reorder operands of the user node and set the ordering for the user | |||
4801 | // node itself. | |||
4802 | if (Data.first->State != TreeEntry::Vectorize || | |||
4803 | !isa<ExtractElementInst, ExtractValueInst, LoadInst>( | |||
4804 | Data.first->getMainOp()) || | |||
4805 | Data.first->isAltShuffle()) | |||
4806 | Data.first->reorderOperands(Mask); | |||
4807 | if (!isa<InsertElementInst, StoreInst>(Data.first->getMainOp()) || | |||
4808 | Data.first->isAltShuffle()) { | |||
4809 | reorderScalars(Data.first->Scalars, Mask); | |||
4810 | reorderOrder(Data.first->ReorderIndices, MaskOrder); | |||
4811 | if (Data.first->ReuseShuffleIndices.empty() && | |||
4812 | !Data.first->ReorderIndices.empty() && | |||
4813 | !Data.first->isAltShuffle()) { | |||
4814 | // Insert user node to the list to try to sink reordering deeper in | |||
4815 | // the graph. | |||
4816 | OrderedEntries.insert(Data.first); | |||
4817 | } | |||
4818 | } else { | |||
4819 | reorderOrder(Data.first->ReorderIndices, Mask); | |||
4820 | } | |||
4821 | } | |||
4822 | } | |||
4823 | // If the reordering is unnecessary, just remove the reorder. | |||
4824 | if (IgnoreReorder && !VectorizableTree.front()->ReorderIndices.empty() && | |||
4825 | VectorizableTree.front()->ReuseShuffleIndices.empty()) | |||
4826 | VectorizableTree.front()->ReorderIndices.clear(); | |||
4827 | } | |||
4828 | ||||
4829 | void BoUpSLP::buildExternalUses( | |||
4830 | const ExtraValueToDebugLocsMap &ExternallyUsedValues) { | |||
4831 | // Collect the values that we need to extract from the tree. | |||
4832 | for (auto &TEPtr : VectorizableTree) { | |||
4833 | TreeEntry *Entry = TEPtr.get(); | |||
4834 | ||||
4835 | // No need to handle users of gathered values. | |||
4836 | if (Entry->State == TreeEntry::NeedToGather) | |||
4837 | continue; | |||
4838 | ||||
4839 | // For each lane: | |||
4840 | for (int Lane = 0, LE = Entry->Scalars.size(); Lane != LE; ++Lane) { | |||
4841 | Value *Scalar = Entry->Scalars[Lane]; | |||
4842 | int FoundLane = Entry->findLaneForValue(Scalar); | |||
4843 | ||||
4844 | // Check if the scalar is externally used as an extra arg. | |||
4845 | auto ExtI = ExternallyUsedValues.find(Scalar); | |||
4846 | if (ExtI != ExternallyUsedValues.end()) { | |||
4847 | 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) | |||
4848 | << 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); | |||
4849 | ExternalUses.emplace_back(Scalar, nullptr, FoundLane); | |||
4850 | } | |||
4851 | for (User *U : Scalar->users()) { | |||
4852 | LLVM_DEBUG(dbgs() << "SLP: Checking user:" << *U << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Checking user:" << *U << ".\n"; } } while (false); | |||
4853 | ||||
4854 | Instruction *UserInst = dyn_cast<Instruction>(U); | |||
4855 | if (!UserInst) | |||
4856 | continue; | |||
4857 | ||||
4858 | if (isDeleted(UserInst)) | |||
4859 | continue; | |||
4860 | ||||
4861 | // Skip in-tree scalars that become vectors | |||
4862 | if (TreeEntry *UseEntry = getTreeEntry(U)) { | |||
4863 | Value *UseScalar = UseEntry->Scalars[0]; | |||
4864 | // Some in-tree scalars will remain as scalar in vectorized | |||
4865 | // instructions. If that is the case, the one in Lane 0 will | |||
4866 | // be used. | |||
4867 | if (UseScalar != U || | |||
4868 | UseEntry->State == TreeEntry::ScatterVectorize || | |||
4869 | !InTreeUserNeedToExtract(Scalar, UserInst, TLI)) { | |||
4870 | 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) | |||
4871 | << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: \tInternal user will be removed:" << *U << ".\n"; } } while (false); | |||
4872 | 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", 4872, __extension__ __PRETTY_FUNCTION__)); | |||
4873 | continue; | |||
4874 | } | |||
4875 | } | |||
4876 | ||||
4877 | // Ignore users in the user ignore list. | |||
4878 | if (UserIgnoreList && UserIgnoreList->contains(UserInst)) | |||
4879 | continue; | |||
4880 | ||||
4881 | 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) | |||
4882 | << 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); | |||
4883 | ExternalUses.push_back(ExternalUser(Scalar, U, FoundLane)); | |||
4884 | } | |||
4885 | } | |||
4886 | } | |||
4887 | } | |||
4888 | ||||
4889 | DenseMap<Value *, SmallVector<StoreInst *, 4>> | |||
4890 | BoUpSLP::collectUserStores(const BoUpSLP::TreeEntry *TE) const { | |||
4891 | DenseMap<Value *, SmallVector<StoreInst *, 4>> PtrToStoresMap; | |||
4892 | for (unsigned Lane : seq<unsigned>(0, TE->Scalars.size())) { | |||
4893 | Value *V = TE->Scalars[Lane]; | |||
4894 | // To save compilation time we don't visit if we have too many users. | |||
4895 | static constexpr unsigned UsersLimit = 4; | |||
4896 | if (V->hasNUsesOrMore(UsersLimit)) | |||
4897 | break; | |||
4898 | ||||
4899 | // Collect stores per pointer object. | |||
4900 | for (User *U : V->users()) { | |||
4901 | auto *SI = dyn_cast<StoreInst>(U); | |||
4902 | if (SI == nullptr || !SI->isSimple() || | |||
4903 | !isValidElementType(SI->getValueOperand()->getType())) | |||
4904 | continue; | |||
4905 | // Skip entry if already | |||
4906 | if (getTreeEntry(U)) | |||
4907 | continue; | |||
4908 | ||||
4909 | Value *Ptr = getUnderlyingObject(SI->getPointerOperand()); | |||
4910 | auto &StoresVec = PtrToStoresMap[Ptr]; | |||
4911 | // For now just keep one store per pointer object per lane. | |||
4912 | // TODO: Extend this to support multiple stores per pointer per lane | |||
4913 | if (StoresVec.size() > Lane) | |||
4914 | continue; | |||
4915 | // Skip if in different BBs. | |||
4916 | if (!StoresVec.empty() && | |||
4917 | SI->getParent() != StoresVec.back()->getParent()) | |||
4918 | continue; | |||
4919 | // Make sure that the stores are of the same type. | |||
4920 | if (!StoresVec.empty() && | |||
4921 | SI->getValueOperand()->getType() != | |||
4922 | StoresVec.back()->getValueOperand()->getType()) | |||
4923 | continue; | |||
4924 | StoresVec.push_back(SI); | |||
4925 | } | |||
4926 | } | |||
4927 | return PtrToStoresMap; | |||
4928 | } | |||
4929 | ||||
4930 | bool BoUpSLP::canFormVector(const SmallVector<StoreInst *, 4> &StoresVec, | |||
4931 | OrdersType &ReorderIndices) const { | |||
4932 | // We check whether the stores in StoreVec can form a vector by sorting them | |||
4933 | // and checking whether they are consecutive. | |||
4934 | ||||
4935 | // To avoid calling getPointersDiff() while sorting we create a vector of | |||
4936 | // pairs {store, offset from first} and sort this instead. | |||
4937 | SmallVector<std::pair<StoreInst *, int>, 4> StoreOffsetVec(StoresVec.size()); | |||
4938 | StoreInst *S0 = StoresVec[0]; | |||
4939 | StoreOffsetVec[0] = {S0, 0}; | |||
4940 | Type *S0Ty = S0->getValueOperand()->getType(); | |||
4941 | Value *S0Ptr = S0->getPointerOperand(); | |||
4942 | for (unsigned Idx : seq<unsigned>(1, StoresVec.size())) { | |||
4943 | StoreInst *SI = StoresVec[Idx]; | |||
4944 | std::optional<int> Diff = | |||
4945 | getPointersDiff(S0Ty, S0Ptr, SI->getValueOperand()->getType(), | |||
4946 | SI->getPointerOperand(), *DL, *SE, | |||
4947 | /*StrictCheck=*/true); | |||
4948 | // We failed to compare the pointers so just abandon this StoresVec. | |||
4949 | if (!Diff) | |||
4950 | return false; | |||
4951 | StoreOffsetVec[Idx] = {StoresVec[Idx], *Diff}; | |||
4952 | } | |||
4953 | ||||
4954 | // Sort the vector based on the pointers. We create a copy because we may | |||
4955 | // need the original later for calculating the reorder (shuffle) indices. | |||
4956 | stable_sort(StoreOffsetVec, [](const std::pair<StoreInst *, int> &Pair1, | |||
4957 | const std::pair<StoreInst *, int> &Pair2) { | |||
4958 | int Offset1 = Pair1.second; | |||
4959 | int Offset2 = Pair2.second; | |||
4960 | return Offset1 < Offset2; | |||
4961 | }); | |||
4962 | ||||
4963 | // Check if the stores are consecutive by checking if their difference is 1. | |||
4964 | for (unsigned Idx : seq<unsigned>(1, StoreOffsetVec.size())) | |||
4965 | if (StoreOffsetVec[Idx].second != StoreOffsetVec[Idx-1].second + 1) | |||
4966 | return false; | |||
4967 | ||||
4968 | // Calculate the shuffle indices according to their offset against the sorted | |||
4969 | // StoreOffsetVec. | |||
4970 | ReorderIndices.reserve(StoresVec.size()); | |||
4971 | for (StoreInst *SI : StoresVec) { | |||
4972 | unsigned Idx = find_if(StoreOffsetVec, | |||
4973 | [SI](const std::pair<StoreInst *, int> &Pair) { | |||
4974 | return Pair.first == SI; | |||
4975 | }) - | |||
4976 | StoreOffsetVec.begin(); | |||
4977 | ReorderIndices.push_back(Idx); | |||
4978 | } | |||
4979 | // Identity order (e.g., {0,1,2,3}) is modeled as an empty OrdersType in | |||
4980 | // reorderTopToBottom() and reorderBottomToTop(), so we are following the | |||
4981 | // same convention here. | |||
4982 | auto IsIdentityOrder = [](const OrdersType &Order) { | |||
4983 | for (unsigned Idx : seq<unsigned>(0, Order.size())) | |||
4984 | if (Idx != Order[Idx]) | |||
4985 | return false; | |||
4986 | return true; | |||
4987 | }; | |||
4988 | if (IsIdentityOrder(ReorderIndices)) | |||
4989 | ReorderIndices.clear(); | |||
4990 | ||||
4991 | return true; | |||
4992 | } | |||
4993 | ||||
4994 | #ifndef NDEBUG | |||
4995 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) static void dumpOrder(const BoUpSLP::OrdersType &Order) { | |||
4996 | for (unsigned Idx : Order) | |||
4997 | dbgs() << Idx << ", "; | |||
4998 | dbgs() << "\n"; | |||
4999 | } | |||
5000 | #endif | |||
5001 | ||||
5002 | SmallVector<BoUpSLP::OrdersType, 1> | |||
5003 | BoUpSLP::findExternalStoreUsersReorderIndices(TreeEntry *TE) const { | |||
5004 | unsigned NumLanes = TE->Scalars.size(); | |||
5005 | ||||
5006 | DenseMap<Value *, SmallVector<StoreInst *, 4>> PtrToStoresMap = | |||
5007 | collectUserStores(TE); | |||
5008 | ||||
5009 | // Holds the reorder indices for each candidate store vector that is a user of | |||
5010 | // the current TreeEntry. | |||
5011 | SmallVector<OrdersType, 1> ExternalReorderIndices; | |||
5012 | ||||
5013 | // Now inspect the stores collected per pointer and look for vectorization | |||
5014 | // candidates. For each candidate calculate the reorder index vector and push | |||
5015 | // it into `ExternalReorderIndices` | |||
5016 | for (const auto &Pair : PtrToStoresMap) { | |||
5017 | auto &StoresVec = Pair.second; | |||
5018 | // If we have fewer than NumLanes stores, then we can't form a vector. | |||
5019 | if (StoresVec.size() != NumLanes) | |||
5020 | continue; | |||
5021 | ||||
5022 | // If the stores are not consecutive then abandon this StoresVec. | |||
5023 | OrdersType ReorderIndices; | |||
5024 | if (!canFormVector(StoresVec, ReorderIndices)) | |||
5025 | continue; | |||
5026 | ||||
5027 | // We now know that the scalars in StoresVec can form a vector instruction, | |||
5028 | // so set the reorder indices. | |||
5029 | ExternalReorderIndices.push_back(ReorderIndices); | |||
5030 | } | |||
5031 | return ExternalReorderIndices; | |||
5032 | } | |||
5033 | ||||
5034 | void BoUpSLP::buildTree(ArrayRef<Value *> Roots, | |||
5035 | const SmallDenseSet<Value *> &UserIgnoreLst) { | |||
5036 | deleteTree(); | |||
5037 | UserIgnoreList = &UserIgnoreLst; | |||
5038 | if (!allSameType(Roots)) | |||
5039 | return; | |||
5040 | buildTree_rec(Roots, 0, EdgeInfo()); | |||
5041 | } | |||
5042 | ||||
5043 | void BoUpSLP::buildTree(ArrayRef<Value *> Roots) { | |||
5044 | deleteTree(); | |||
5045 | if (!allSameType(Roots)) | |||
5046 | return; | |||
5047 | buildTree_rec(Roots, 0, EdgeInfo()); | |||
5048 | } | |||
5049 | ||||
5050 | /// \return true if the specified list of values has only one instruction that | |||
5051 | /// requires scheduling, false otherwise. | |||
5052 | #ifndef NDEBUG | |||
5053 | static bool needToScheduleSingleInstruction(ArrayRef<Value *> VL) { | |||
5054 | Value *NeedsScheduling = nullptr; | |||
5055 | for (Value *V : VL) { | |||
5056 | if (doesNotNeedToBeScheduled(V)) | |||
5057 | continue; | |||
5058 | if (!NeedsScheduling) { | |||
5059 | NeedsScheduling = V; | |||
5060 | continue; | |||
5061 | } | |||
5062 | return false; | |||
5063 | } | |||
5064 | return NeedsScheduling; | |||
5065 | } | |||
5066 | #endif | |||
5067 | ||||
5068 | /// Generates key/subkey pair for the given value to provide effective sorting | |||
5069 | /// of the values and better detection of the vectorizable values sequences. The | |||
5070 | /// keys/subkeys can be used for better sorting of the values themselves (keys) | |||
5071 | /// and in values subgroups (subkeys). | |||
5072 | static std::pair<size_t, size_t> generateKeySubkey( | |||
5073 | Value *V, const TargetLibraryInfo *TLI, | |||
5074 | function_ref<hash_code(size_t, LoadInst *)> LoadsSubkeyGenerator, | |||
5075 | bool AllowAlternate) { | |||
5076 | hash_code Key = hash_value(V->getValueID() + 2); | |||
5077 | hash_code SubKey = hash_value(0); | |||
5078 | // Sort the loads by the distance between the pointers. | |||
5079 | if (auto *LI = dyn_cast<LoadInst>(V)) { | |||
5080 | Key = hash_combine(LI->getType(), hash_value(Instruction::Load), Key); | |||
5081 | if (LI->isSimple()) | |||
5082 | SubKey = hash_value(LoadsSubkeyGenerator(Key, LI)); | |||
5083 | else | |||
5084 | Key = SubKey = hash_value(LI); | |||
5085 | } else if (isVectorLikeInstWithConstOps(V)) { | |||
5086 | // Sort extracts by the vector operands. | |||
5087 | if (isa<ExtractElementInst, UndefValue>(V)) | |||
5088 | Key = hash_value(Value::UndefValueVal + 1); | |||
5089 | if (auto *EI = dyn_cast<ExtractElementInst>(V)) { | |||
5090 | if (!isUndefVector(EI->getVectorOperand()).all() && | |||
5091 | !isa<UndefValue>(EI->getIndexOperand())) | |||
5092 | SubKey = hash_value(EI->getVectorOperand()); | |||
5093 | } | |||
5094 | } else if (auto *I = dyn_cast<Instruction>(V)) { | |||
5095 | // Sort other instructions just by the opcodes except for CMPInst. | |||
5096 | // For CMP also sort by the predicate kind. | |||
5097 | if ((isa<BinaryOperator, CastInst>(I)) && | |||
5098 | isValidForAlternation(I->getOpcode())) { | |||
5099 | if (AllowAlternate) | |||
5100 | Key = hash_value(isa<BinaryOperator>(I) ? 1 : 0); | |||
5101 | else | |||
5102 | Key = hash_combine(hash_value(I->getOpcode()), Key); | |||
5103 | SubKey = hash_combine( | |||
5104 | hash_value(I->getOpcode()), hash_value(I->getType()), | |||
5105 | hash_value(isa<BinaryOperator>(I) | |||
5106 | ? I->getType() | |||
5107 | : cast<CastInst>(I)->getOperand(0)->getType())); | |||
5108 | // For casts, look through the only operand to improve compile time. | |||
5109 | if (isa<CastInst>(I)) { | |||
5110 | std::pair<size_t, size_t> OpVals = | |||
5111 | generateKeySubkey(I->getOperand(0), TLI, LoadsSubkeyGenerator, | |||
5112 | /*AllowAlternate=*/true); | |||
5113 | Key = hash_combine(OpVals.first, Key); | |||
5114 | SubKey = hash_combine(OpVals.first, SubKey); | |||
5115 | } | |||
5116 | } else if (auto *CI = dyn_cast<CmpInst>(I)) { | |||
5117 | CmpInst::Predicate Pred = CI->getPredicate(); | |||
5118 | if (CI->isCommutative()) | |||
5119 | Pred = std::min(Pred, CmpInst::getInversePredicate(Pred)); | |||
5120 | CmpInst::Predicate SwapPred = CmpInst::getSwappedPredicate(Pred); | |||
5121 | SubKey = hash_combine(hash_value(I->getOpcode()), hash_value(Pred), | |||
5122 | hash_value(SwapPred), | |||
5123 | hash_value(CI->getOperand(0)->getType())); | |||
5124 | } else if (auto *Call = dyn_cast<CallInst>(I)) { | |||
5125 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(Call, TLI); | |||
5126 | if (isTriviallyVectorizable(ID)) { | |||
5127 | SubKey = hash_combine(hash_value(I->getOpcode()), hash_value(ID)); | |||
5128 | } else if (!VFDatabase(*Call).getMappings(*Call).empty()) { | |||
5129 | SubKey = hash_combine(hash_value(I->getOpcode()), | |||
5130 | hash_value(Call->getCalledFunction())); | |||
5131 | } else { | |||
5132 | Key = hash_combine(hash_value(Call), Key); | |||
5133 | SubKey = hash_combine(hash_value(I->getOpcode()), hash_value(Call)); | |||
5134 | } | |||
5135 | for (const CallBase::BundleOpInfo &Op : Call->bundle_op_infos()) | |||
5136 | SubKey = hash_combine(hash_value(Op.Begin), hash_value(Op.End), | |||
5137 | hash_value(Op.Tag), SubKey); | |||
5138 | } else if (auto *Gep = dyn_cast<GetElementPtrInst>(I)) { | |||
5139 | if (Gep->getNumOperands() == 2 && isa<ConstantInt>(Gep->getOperand(1))) | |||
5140 | SubKey = hash_value(Gep->getPointerOperand()); | |||
5141 | else | |||
5142 | SubKey = hash_value(Gep); | |||
5143 | } else if (BinaryOperator::isIntDivRem(I->getOpcode()) && | |||
5144 | !isa<ConstantInt>(I->getOperand(1))) { | |||
5145 | // Do not try to vectorize instructions with potentially high cost. | |||
5146 | SubKey = hash_value(I); | |||
5147 | } else { | |||
5148 | SubKey = hash_value(I->getOpcode()); | |||
5149 | } | |||
5150 | Key = hash_combine(hash_value(I->getParent()), Key); | |||
5151 | } | |||
5152 | return std::make_pair(Key, SubKey); | |||
5153 | } | |||
5154 | ||||
5155 | /// Checks if the specified instruction \p I is an alternate operation for | |||
5156 | /// the given \p MainOp and \p AltOp instructions. | |||
5157 | static bool isAlternateInstruction(const Instruction *I, | |||
5158 | const Instruction *MainOp, | |||
5159 | const Instruction *AltOp, | |||
5160 | const TargetLibraryInfo &TLI); | |||
5161 | ||||
5162 | void BoUpSLP::buildTree_rec(ArrayRef<Value *> VL, unsigned Depth, | |||
5163 | const EdgeInfo &UserTreeIdx) { | |||
5164 | 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", 5164, __extension__ __PRETTY_FUNCTION__)); | |||
5165 | ||||
5166 | SmallVector<int> ReuseShuffleIndicies; | |||
5167 | SmallVector<Value *> UniqueValues; | |||
5168 | auto &&TryToFindDuplicates = [&VL, &ReuseShuffleIndicies, &UniqueValues, | |||
5169 | &UserTreeIdx, | |||
5170 | this](const InstructionsState &S) { | |||
5171 | // Check that every instruction appears once in this bundle. | |||
5172 | DenseMap<Value *, unsigned> UniquePositions(VL.size()); | |||
5173 | for (Value *V : VL) { | |||
5174 | if (isConstant(V)) { | |||
5175 | ReuseShuffleIndicies.emplace_back( | |||
5176 | isa<UndefValue>(V) ? PoisonMaskElem : UniqueValues.size()); | |||
5177 | UniqueValues.emplace_back(V); | |||
5178 | continue; | |||
5179 | } | |||
5180 | auto Res = UniquePositions.try_emplace(V, UniqueValues.size()); | |||
5181 | ReuseShuffleIndicies.emplace_back(Res.first->second); | |||
5182 | if (Res.second) | |||
5183 | UniqueValues.emplace_back(V); | |||
5184 | } | |||
5185 | size_t NumUniqueScalarValues = UniqueValues.size(); | |||
5186 | if (NumUniqueScalarValues == VL.size()) { | |||
5187 | ReuseShuffleIndicies.clear(); | |||
5188 | } else { | |||
5189 | 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); | |||
5190 | if (NumUniqueScalarValues <= 1 || | |||
5191 | (UniquePositions.size() == 1 && all_of(UniqueValues, | |||
5192 | [](Value *V) { | |||
5193 | return isa<UndefValue>(V) || | |||
5194 | !isConstant(V); | |||
5195 | })) || | |||
5196 | !llvm::has_single_bit<uint32_t>(NumUniqueScalarValues)) { | |||
5197 | 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); | |||
5198 | newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx); | |||
5199 | return false; | |||
5200 | } | |||
5201 | VL = UniqueValues; | |||
5202 | } | |||
5203 | return true; | |||
5204 | }; | |||
5205 | ||||
5206 | InstructionsState S = getSameOpcode(VL, *TLI); | |||
5207 | ||||
5208 | // Gather if we hit the RecursionMaxDepth, unless this is a load (or z/sext of | |||
5209 | // a load), in which case peek through to include it in the tree, without | |||
5210 | // ballooning over-budget. | |||
5211 | if (Depth >= RecursionMaxDepth && | |||
5212 | !(S.MainOp && isa<Instruction>(S.MainOp) && S.MainOp == S.AltOp && | |||
5213 | VL.size() >= 4 && | |||
5214 | (match(S.MainOp, m_Load(m_Value())) || all_of(VL, [&S](const Value *I) { | |||
5215 | return match(I, | |||
5216 | m_OneUse(m_ZExtOrSExt(m_OneUse(m_Load(m_Value()))))) && | |||
5217 | cast<Instruction>(I)->getOpcode() == | |||
5218 | cast<Instruction>(S.MainOp)->getOpcode(); | |||
5219 | })))) { | |||
5220 | 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); | |||
5221 | if (TryToFindDuplicates(S)) | |||
5222 | newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx, | |||
5223 | ReuseShuffleIndicies); | |||
5224 | return; | |||
5225 | } | |||
5226 | ||||
5227 | // Don't handle scalable vectors | |||
5228 | if (S.getOpcode() == Instruction::ExtractElement && | |||
5229 | isa<ScalableVectorType>( | |||
5230 | cast<ExtractElementInst>(S.OpValue)->getVectorOperandType())) { | |||
5231 | 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); | |||
5232 | if (TryToFindDuplicates(S)) | |||
5233 | newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx, | |||
5234 | ReuseShuffleIndicies); | |||
5235 | return; | |||
5236 | } | |||
5237 | ||||
5238 | // Don't handle vectors. | |||
5239 | if (S.OpValue->getType()->isVectorTy() && | |||
5240 | !isa<InsertElementInst>(S.OpValue)) { | |||
5241 | 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); | |||
5242 | newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx); | |||
5243 | return; | |||
5244 | } | |||
5245 | ||||
5246 | if (StoreInst *SI = dyn_cast<StoreInst>(S.OpValue)) | |||
5247 | if (SI->getValueOperand()->getType()->isVectorTy()) { | |||
5248 | 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); | |||
5249 | newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx); | |||
5250 | return; | |||
5251 | } | |||
5252 | ||||
5253 | // If all of the operands are identical or constant we have a simple solution. | |||
5254 | // If we deal with insert/extract instructions, they all must have constant | |||
5255 | // indices, otherwise we should gather them, not try to vectorize. | |||
5256 | // If alternate op node with 2 elements with gathered operands - do not | |||
5257 | // vectorize. | |||
5258 | auto &&NotProfitableForVectorization = [&S, this, | |||
5259 | Depth](ArrayRef<Value *> VL) { | |||
5260 | if (!S.getOpcode() || !S.isAltShuffle() || VL.size() > 2) | |||
5261 | return false; | |||
5262 | if (VectorizableTree.size() < MinTreeSize) | |||
5263 | return false; | |||
5264 | if (Depth >= RecursionMaxDepth - 1) | |||
5265 | return true; | |||
5266 | // Check if all operands are extracts, part of vector node or can build a | |||
5267 | // regular vectorize node. | |||
5268 | SmallVector<unsigned, 2> InstsCount(VL.size(), 0); | |||
5269 | for (Value *V : VL) { | |||
5270 | auto *I = cast<Instruction>(V); | |||
5271 | InstsCount.push_back(count_if(I->operand_values(), [](Value *Op) { | |||
5272 | return isa<Instruction>(Op) || isVectorLikeInstWithConstOps(Op); | |||
5273 | })); | |||
5274 | } | |||
5275 | bool IsCommutative = isCommutative(S.MainOp) || isCommutative(S.AltOp); | |||
5276 | if ((IsCommutative && | |||
5277 | std::accumulate(InstsCount.begin(), InstsCount.end(), 0) < 2) || | |||
5278 | (!IsCommutative && | |||
5279 | all_of(InstsCount, [](unsigned ICnt) { return ICnt < 2; }))) | |||
5280 | return true; | |||
5281 | assert(VL.size() == 2 && "Expected only 2 alternate op instructions.")(static_cast <bool> (VL.size() == 2 && "Expected only 2 alternate op instructions." ) ? void (0) : __assert_fail ("VL.size() == 2 && \"Expected only 2 alternate op instructions.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 5281, __extension__ __PRETTY_FUNCTION__)); | |||
5282 | SmallVector<SmallVector<std::pair<Value *, Value *>>> Candidates; | |||
5283 | auto *I1 = cast<Instruction>(VL.front()); | |||
5284 | auto *I2 = cast<Instruction>(VL.back()); | |||
5285 | for (int Op = 0, E = S.MainOp->getNumOperands(); Op < E; ++Op) | |||
5286 | Candidates.emplace_back().emplace_back(I1->getOperand(Op), | |||
5287 | I2->getOperand(Op)); | |||
5288 | if (static_cast<unsigned>(count_if( | |||
5289 | Candidates, [this](ArrayRef<std::pair<Value *, Value *>> Cand) { | |||
5290 | return findBestRootPair(Cand, LookAheadHeuristics::ScoreSplat); | |||
5291 | })) >= S.MainOp->getNumOperands() / 2) | |||
5292 | return false; | |||
5293 | if (S.MainOp->getNumOperands() > 2) | |||
5294 | return true; | |||
5295 | if (IsCommutative) { | |||
5296 | // Check permuted operands. | |||
5297 | Candidates.clear(); | |||
5298 | for (int Op = 0, E = S.MainOp->getNumOperands(); Op < E; ++Op) | |||
5299 | Candidates.emplace_back().emplace_back(I1->getOperand(Op), | |||
5300 | I2->getOperand((Op + 1) % E)); | |||
5301 | if (any_of( | |||
5302 | Candidates, [this](ArrayRef<std::pair<Value *, Value *>> Cand) { | |||
5303 | return findBestRootPair(Cand, LookAheadHeuristics::ScoreSplat); | |||
5304 | })) | |||
5305 | return false; | |||
5306 | } | |||
5307 | return true; | |||
5308 | }; | |||
5309 | SmallVector<unsigned> SortedIndices; | |||
5310 | BasicBlock *BB = nullptr; | |||
5311 | bool IsScatterVectorizeUserTE = | |||
5312 | UserTreeIdx.UserTE && | |||
5313 | UserTreeIdx.UserTE->State == TreeEntry::ScatterVectorize; | |||
5314 | bool AreAllSameInsts = | |||
5315 | (S.getOpcode() && allSameBlock(VL)) || | |||
5316 | (S.OpValue->getType()->isPointerTy() && IsScatterVectorizeUserTE && | |||
5317 | VL.size() > 2 && | |||
5318 | all_of(VL, | |||
5319 | [&BB](Value *V) { | |||
5320 | auto *I = dyn_cast<GetElementPtrInst>(V); | |||
5321 | if (!I) | |||
5322 | return doesNotNeedToBeScheduled(V); | |||
5323 | if (!BB) | |||
5324 | BB = I->getParent(); | |||
5325 | return BB == I->getParent() && I->getNumOperands() == 2; | |||
5326 | }) && | |||
5327 | BB && | |||
5328 | sortPtrAccesses(VL, UserTreeIdx.UserTE->getMainOp()->getType(), *DL, *SE, | |||
5329 | SortedIndices)); | |||
5330 | if (!AreAllSameInsts || allConstant(VL) || isSplat(VL) || | |||
5331 | (isa<InsertElementInst, ExtractValueInst, ExtractElementInst>( | |||
5332 | S.OpValue) && | |||
5333 | !all_of(VL, isVectorLikeInstWithConstOps)) || | |||
5334 | NotProfitableForVectorization(VL)) { | |||
5335 | LLVM_DEBUG(dbgs() << "SLP: Gathering due to C,S,B,O, small shuffle. \n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering due to C,S,B,O, small shuffle. \n" ; } } while (false); | |||
5336 | if (TryToFindDuplicates(S)) | |||
5337 | newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx, | |||
5338 | ReuseShuffleIndicies); | |||
5339 | return; | |||
5340 | } | |||
5341 | ||||
5342 | // We now know that this is a vector of instructions of the same type from | |||
5343 | // the same block. | |||
5344 | ||||
5345 | // Don't vectorize ephemeral values. | |||
5346 | if (!EphValues.empty()) { | |||
5347 | for (Value *V : VL) { | |||
5348 | if (EphValues.count(V)) { | |||
5349 | LLVM_DEBUG(dbgs() << "SLP: The instruction (" << *Vdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: The instruction (" << * V << ") is ephemeral.\n"; } } while (false) | |||
5350 | << ") is ephemeral.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: The instruction (" << * V << ") is ephemeral.\n"; } } while (false); | |||
5351 | newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx); | |||
5352 | return; | |||
5353 | } | |||
5354 | } | |||
5355 | } | |||
5356 | ||||
5357 | // Check if this is a duplicate of another entry. | |||
5358 | if (TreeEntry *E = getTreeEntry(S.OpValue)) { | |||
5359 | 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); | |||
5360 | if (!E->isSame(VL)) { | |||
5361 | 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); | |||
5362 | if (TryToFindDuplicates(S)) | |||
5363 | newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx, | |||
5364 | ReuseShuffleIndicies); | |||
5365 | return; | |||
5366 | } | |||
5367 | // Record the reuse of the tree node. FIXME, currently this is only used to | |||
5368 | // properly draw the graph rather than for the actual vectorization. | |||
5369 | E->UserTreeIndices.push_back(UserTreeIdx); | |||
5370 | 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) | |||
5371 | << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Perfect diamond merge at " << *S.OpValue << ".\n"; } } while (false); | |||
5372 | return; | |||
5373 | } | |||
5374 | ||||
5375 | // Check that none of the instructions in the bundle are already in the tree. | |||
5376 | for (Value *V : VL) { | |||
5377 | if (!IsScatterVectorizeUserTE && !isa<Instruction>(V)) | |||
5378 | continue; | |||
5379 | if (getTreeEntry(V)) { | |||
5380 | 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) | |||
5381 | << ") is already in tree.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: The instruction (" << * V << ") is already in tree.\n"; } } while (false); | |||
5382 | if (TryToFindDuplicates(S)) | |||
5383 | newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx, | |||
5384 | ReuseShuffleIndicies); | |||
5385 | return; | |||
5386 | } | |||
5387 | } | |||
5388 | ||||
5389 | // The reduction nodes (stored in UserIgnoreList) also should stay scalar. | |||
5390 | if (UserIgnoreList && !UserIgnoreList->empty()) { | |||
5391 | for (Value *V : VL) { | |||
5392 | if (UserIgnoreList && UserIgnoreList->contains(V)) { | |||
5393 | 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); | |||
5394 | if (TryToFindDuplicates(S)) | |||
5395 | newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx, | |||
5396 | ReuseShuffleIndicies); | |||
5397 | return; | |||
5398 | } | |||
5399 | } | |||
5400 | } | |||
5401 | ||||
5402 | // Special processing for sorted pointers for ScatterVectorize node with | |||
5403 | // constant indeces only. | |||
5404 | if (AreAllSameInsts && UserTreeIdx.UserTE && | |||
5405 | UserTreeIdx.UserTE->State == TreeEntry::ScatterVectorize && | |||
5406 | !(S.getOpcode() && allSameBlock(VL))) { | |||
5407 | assert(S.OpValue->getType()->isPointerTy() &&(static_cast <bool> (S.OpValue->getType()->isPointerTy () && count_if(VL, [](Value *V) { return isa<GetElementPtrInst >(V); }) >= 2 && "Expected pointers only.") ? void (0) : __assert_fail ("S.OpValue->getType()->isPointerTy() && count_if(VL, [](Value *V) { return isa<GetElementPtrInst>(V); }) >= 2 && \"Expected pointers only.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 5410, __extension__ __PRETTY_FUNCTION__)) | |||
5408 | count_if(VL, [](Value *V) { return isa<GetElementPtrInst>(V); }) >=(static_cast <bool> (S.OpValue->getType()->isPointerTy () && count_if(VL, [](Value *V) { return isa<GetElementPtrInst >(V); }) >= 2 && "Expected pointers only.") ? void (0) : __assert_fail ("S.OpValue->getType()->isPointerTy() && count_if(VL, [](Value *V) { return isa<GetElementPtrInst>(V); }) >= 2 && \"Expected pointers only.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 5410, __extension__ __PRETTY_FUNCTION__)) | |||
5409 | 2 &&(static_cast <bool> (S.OpValue->getType()->isPointerTy () && count_if(VL, [](Value *V) { return isa<GetElementPtrInst >(V); }) >= 2 && "Expected pointers only.") ? void (0) : __assert_fail ("S.OpValue->getType()->isPointerTy() && count_if(VL, [](Value *V) { return isa<GetElementPtrInst>(V); }) >= 2 && \"Expected pointers only.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 5410, __extension__ __PRETTY_FUNCTION__)) | |||
5410 | "Expected pointers only.")(static_cast <bool> (S.OpValue->getType()->isPointerTy () && count_if(VL, [](Value *V) { return isa<GetElementPtrInst >(V); }) >= 2 && "Expected pointers only.") ? void (0) : __assert_fail ("S.OpValue->getType()->isPointerTy() && count_if(VL, [](Value *V) { return isa<GetElementPtrInst>(V); }) >= 2 && \"Expected pointers only.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 5410, __extension__ __PRETTY_FUNCTION__)); | |||
5411 | // Reset S to make it GetElementPtr kind of node. | |||
5412 | const auto *It = find_if(VL, [](Value *V) { return isa<GetElementPtrInst>(V); }); | |||
5413 | assert(It != VL.end() && "Expected at least one GEP.")(static_cast <bool> (It != VL.end() && "Expected at least one GEP." ) ? void (0) : __assert_fail ("It != VL.end() && \"Expected at least one GEP.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 5413, __extension__ __PRETTY_FUNCTION__)); | |||
5414 | S = getSameOpcode(*It, *TLI); | |||
5415 | } | |||
5416 | ||||
5417 | // Check that all of the users of the scalars that we want to vectorize are | |||
5418 | // schedulable. | |||
5419 | auto *VL0 = cast<Instruction>(S.OpValue); | |||
5420 | BB = VL0->getParent(); | |||
5421 | ||||
5422 | if (!DT->isReachableFromEntry(BB)) { | |||
5423 | // Don't go into unreachable blocks. They may contain instructions with | |||
5424 | // dependency cycles which confuse the final scheduling. | |||
5425 | 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); | |||
5426 | newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx); | |||
5427 | return; | |||
5428 | } | |||
5429 | ||||
5430 | // Don't go into catchswitch blocks, which can happen with PHIs. | |||
5431 | // Such blocks can only have PHIs and the catchswitch. There is no | |||
5432 | // place to insert a shuffle if we need to, so just avoid that issue. | |||
5433 | if (isa<CatchSwitchInst>(BB->getTerminator())) { | |||
5434 | LLVM_DEBUG(dbgs() << "SLP: bundle in catchswitch block.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: bundle in catchswitch block.\n" ; } } while (false); | |||
5435 | newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx); | |||
5436 | return; | |||
5437 | } | |||
5438 | ||||
5439 | // Check that every instruction appears once in this bundle. | |||
5440 | if (!TryToFindDuplicates(S)) | |||
5441 | return; | |||
5442 | ||||
5443 | auto &BSRef = BlocksSchedules[BB]; | |||
5444 | if (!BSRef) | |||
5445 | BSRef = std::make_unique<BlockScheduling>(BB); | |||
5446 | ||||
5447 | BlockScheduling &BS = *BSRef; | |||
5448 | ||||
5449 | std::optional<ScheduleData *> Bundle = BS.tryScheduleBundle(VL, this, S); | |||
5450 | #ifdef EXPENSIVE_CHECKS | |||
5451 | // Make sure we didn't break any internal invariants | |||
5452 | BS.verify(); | |||
5453 | #endif | |||
5454 | if (!Bundle) { | |||
5455 | 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); | |||
5456 | 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", 5458, __extension__ __PRETTY_FUNCTION__)) | |||
5457 | !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", 5458, __extension__ __PRETTY_FUNCTION__)) | |||
5458 | "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", 5458, __extension__ __PRETTY_FUNCTION__)); | |||
5459 | newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx, | |||
5460 | ReuseShuffleIndicies); | |||
5461 | return; | |||
5462 | } | |||
5463 | 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); | |||
5464 | ||||
5465 | unsigned ShuffleOrOp = S.isAltShuffle() ? | |||
5466 | (unsigned) Instruction::ShuffleVector : S.getOpcode(); | |||
5467 | switch (ShuffleOrOp) { | |||
5468 | case Instruction::PHI: { | |||
5469 | auto *PH = cast<PHINode>(VL0); | |||
5470 | ||||
5471 | // Check for terminator values (e.g. invoke). | |||
5472 | for (Value *V : VL) | |||
5473 | for (Value *Incoming : cast<PHINode>(V)->incoming_values()) { | |||
5474 | Instruction *Term = dyn_cast<Instruction>(Incoming); | |||
5475 | if (Term && Term->isTerminator()) { | |||
5476 | LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Need to swizzle PHINodes (terminator use).\n" ; } } while (false) | |||
5477 | << "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); | |||
5478 | BS.cancelScheduling(VL, VL0); | |||
5479 | newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx, | |||
5480 | ReuseShuffleIndicies); | |||
5481 | return; | |||
5482 | } | |||
5483 | } | |||
5484 | ||||
5485 | TreeEntry *TE = | |||
5486 | newTreeEntry(VL, Bundle, S, UserTreeIdx, ReuseShuffleIndicies); | |||
5487 | 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); | |||
5488 | ||||
5489 | // Keeps the reordered operands to avoid code duplication. | |||
5490 | SmallVector<ValueList, 2> OperandsVec; | |||
5491 | for (unsigned I = 0, E = PH->getNumIncomingValues(); I < E; ++I) { | |||
5492 | if (!DT->isReachableFromEntry(PH->getIncomingBlock(I))) { | |||
5493 | ValueList Operands(VL.size(), PoisonValue::get(PH->getType())); | |||
5494 | TE->setOperand(I, Operands); | |||
5495 | OperandsVec.push_back(Operands); | |||
5496 | continue; | |||
5497 | } | |||
5498 | ValueList Operands; | |||
5499 | // Prepare the operand vector. | |||
5500 | for (Value *V : VL) | |||
5501 | Operands.push_back(cast<PHINode>(V)->getIncomingValueForBlock( | |||
5502 | PH->getIncomingBlock(I))); | |||
5503 | TE->setOperand(I, Operands); | |||
5504 | OperandsVec.push_back(Operands); | |||
5505 | } | |||
5506 | for (unsigned OpIdx = 0, OpE = OperandsVec.size(); OpIdx != OpE; ++OpIdx) | |||
5507 | buildTree_rec(OperandsVec[OpIdx], Depth + 1, {TE, OpIdx}); | |||
5508 | return; | |||
5509 | } | |||
5510 | case Instruction::ExtractValue: | |||
5511 | case Instruction::ExtractElement: { | |||
5512 | OrdersType CurrentOrder; | |||
5513 | bool Reuse = canReuseExtract(VL, VL0, CurrentOrder); | |||
5514 | if (Reuse) { | |||
5515 | 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); | |||
5516 | newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, | |||
5517 | ReuseShuffleIndicies); | |||
5518 | // This is a special case, as it does not gather, but at the same time | |||
5519 | // we are not extending buildTree_rec() towards the operands. | |||
5520 | ValueList Op0; | |||
5521 | Op0.assign(VL.size(), VL0->getOperand(0)); | |||
5522 | VectorizableTree.back()->setOperand(0, Op0); | |||
5523 | return; | |||
5524 | } | |||
5525 | if (!CurrentOrder.empty()) { | |||
5526 | 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) | |||
5527 | 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) | |||
5528 | "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) | |||
5529 | 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) | |||
5530 | 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) | |||
5531 | 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) | |||
5532 | })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); | |||
5533 | fixupOrderingIndices(CurrentOrder); | |||
5534 | // Insert new order with initial value 0, if it does not exist, | |||
5535 | // otherwise return the iterator to the existing one. | |||
5536 | newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, | |||
5537 | ReuseShuffleIndicies, CurrentOrder); | |||
5538 | // This is a special case, as it does not gather, but at the same time | |||
5539 | // we are not extending buildTree_rec() towards the operands. | |||
5540 | ValueList Op0; | |||
5541 | Op0.assign(VL.size(), VL0->getOperand(0)); | |||
5542 | VectorizableTree.back()->setOperand(0, Op0); | |||
5543 | return; | |||
5544 | } | |||
5545 | LLVM_DEBUG(dbgs() << "SLP: Gather extract sequence.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gather extract sequence.\n"; } } while (false); | |||
5546 | newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx, | |||
5547 | ReuseShuffleIndicies); | |||
5548 | BS.cancelScheduling(VL, VL0); | |||
5549 | return; | |||
5550 | } | |||
5551 | case Instruction::InsertElement: { | |||
5552 | 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", 5552, __extension__ __PRETTY_FUNCTION__)); | |||
5553 | ||||
5554 | // Check that we have a buildvector and not a shuffle of 2 or more | |||
5555 | // different vectors. | |||
5556 | ValueSet SourceVectors; | |||
5557 | for (Value *V : VL) { | |||
5558 | SourceVectors.insert(cast<Instruction>(V)->getOperand(0)); | |||
5559 | assert(getInsertIndex(V) != std::nullopt &&(static_cast <bool> (getInsertIndex(V) != std::nullopt && "Non-constant or undef index?") ? void (0) : __assert_fail ( "getInsertIndex(V) != std::nullopt && \"Non-constant or undef index?\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 5560, __extension__ __PRETTY_FUNCTION__)) | |||
5560 | "Non-constant or undef index?")(static_cast <bool> (getInsertIndex(V) != std::nullopt && "Non-constant or undef index?") ? void (0) : __assert_fail ( "getInsertIndex(V) != std::nullopt && \"Non-constant or undef index?\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 5560, __extension__ __PRETTY_FUNCTION__)); | |||
5561 | } | |||
5562 | ||||
5563 | if (count_if(VL, [&SourceVectors](Value *V) { | |||
5564 | return !SourceVectors.contains(V); | |||
5565 | }) >= 2) { | |||
5566 | // Found 2nd source vector - cancel. | |||
5567 | 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) | |||
5568 | "different source vectors.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gather of insertelement vectors with " "different source vectors.\n"; } } while (false); | |||
5569 | newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx); | |||
5570 | BS.cancelScheduling(VL, VL0); | |||
5571 | return; | |||
5572 | } | |||
5573 | ||||
5574 | auto OrdCompare = [](const std::pair<int, int> &P1, | |||
5575 | const std::pair<int, int> &P2) { | |||
5576 | return P1.first > P2.first; | |||
5577 | }; | |||
5578 | PriorityQueue<std::pair<int, int>, SmallVector<std::pair<int, int>>, | |||
5579 | decltype(OrdCompare)> | |||
5580 | Indices(OrdCompare); | |||
5581 | for (int I = 0, E = VL.size(); I < E; ++I) { | |||
5582 | unsigned Idx = *getInsertIndex(VL[I]); | |||
5583 | Indices.emplace(Idx, I); | |||
5584 | } | |||
5585 | OrdersType CurrentOrder(VL.size(), VL.size()); | |||
5586 | bool IsIdentity = true; | |||
5587 | for (int I = 0, E = VL.size(); I < E; ++I) { | |||
5588 | CurrentOrder[Indices.top().second] = I; | |||
5589 | IsIdentity &= Indices.top().second == I; | |||
5590 | Indices.pop(); | |||
5591 | } | |||
5592 | if (IsIdentity) | |||
5593 | CurrentOrder.clear(); | |||
5594 | TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, | |||
5595 | std::nullopt, CurrentOrder); | |||
5596 | LLVM_DEBUG(dbgs() << "SLP: added inserts bundle.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: added inserts bundle.\n"; } } while (false); | |||
5597 | ||||
5598 | constexpr int NumOps = 2; | |||
5599 | ValueList VectorOperands[NumOps]; | |||
5600 | for (int I = 0; I < NumOps; ++I) { | |||
5601 | for (Value *V : VL) | |||
5602 | VectorOperands[I].push_back(cast<Instruction>(V)->getOperand(I)); | |||
5603 | ||||
5604 | TE->setOperand(I, VectorOperands[I]); | |||
5605 | } | |||
5606 | buildTree_rec(VectorOperands[NumOps - 1], Depth + 1, {TE, NumOps - 1}); | |||
5607 | return; | |||
5608 | } | |||
5609 | case Instruction::Load: { | |||
5610 | // Check that a vectorized load would load the same memory as a scalar | |||
5611 | // load. For example, we don't want to vectorize loads that are smaller | |||
5612 | // than 8-bit. Even though we have a packed struct {<i2, i2, i2, i2>} LLVM | |||
5613 | // treats loading/storing it as an i8 struct. If we vectorize loads/stores | |||
5614 | // from such a struct, we read/write packed bits disagreeing with the | |||
5615 | // unvectorized version. | |||
5616 | SmallVector<Value *> PointerOps; | |||
5617 | OrdersType CurrentOrder; | |||
5618 | TreeEntry *TE = nullptr; | |||
5619 | switch (canVectorizeLoads(VL, VL0, *TTI, *DL, *SE, *LI, *TLI, | |||
5620 | CurrentOrder, PointerOps)) { | |||
5621 | case LoadsState::Vectorize: | |||
5622 | if (CurrentOrder.empty()) { | |||
5623 | // Original loads are consecutive and does not require reordering. | |||
5624 | TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, | |||
5625 | ReuseShuffleIndicies); | |||
5626 | 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); | |||
5627 | } else { | |||
5628 | fixupOrderingIndices(CurrentOrder); | |||
5629 | // Need to reorder. | |||
5630 | TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, | |||
5631 | ReuseShuffleIndicies, CurrentOrder); | |||
5632 | 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); | |||
5633 | } | |||
5634 | TE->setOperandsInOrder(); | |||
5635 | break; | |||
5636 | case LoadsState::ScatterVectorize: | |||
5637 | // Vectorizing non-consecutive loads with `llvm.masked.gather`. | |||
5638 | TE = newTreeEntry(VL, TreeEntry::ScatterVectorize, Bundle, S, | |||
5639 | UserTreeIdx, ReuseShuffleIndicies); | |||
5640 | TE->setOperandsInOrder(); | |||
5641 | buildTree_rec(PointerOps, Depth + 1, {TE, 0}); | |||
5642 | 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); | |||
5643 | break; | |||
5644 | case LoadsState::Gather: | |||
5645 | BS.cancelScheduling(VL, VL0); | |||
5646 | newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx, | |||
5647 | ReuseShuffleIndicies); | |||
5648 | #ifndef NDEBUG | |||
5649 | Type *ScalarTy = VL0->getType(); | |||
5650 | if (DL->getTypeSizeInBits(ScalarTy) != | |||
5651 | DL->getTypeAllocSizeInBits(ScalarTy)) | |||
5652 | 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); | |||
5653 | else if (any_of(VL, [](Value *V) { | |||
5654 | return !cast<LoadInst>(V)->isSimple(); | |||
5655 | })) | |||
5656 | 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); | |||
5657 | else | |||
5658 | 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); | |||
5659 | #endif // NDEBUG | |||
5660 | break; | |||
5661 | } | |||
5662 | return; | |||
5663 | } | |||
5664 | case Instruction::ZExt: | |||
5665 | case Instruction::SExt: | |||
5666 | case Instruction::FPToUI: | |||
5667 | case Instruction::FPToSI: | |||
5668 | case Instruction::FPExt: | |||
5669 | case Instruction::PtrToInt: | |||
5670 | case Instruction::IntToPtr: | |||
5671 | case Instruction::SIToFP: | |||
5672 | case Instruction::UIToFP: | |||
5673 | case Instruction::Trunc: | |||
5674 | case Instruction::FPTrunc: | |||
5675 | case Instruction::BitCast: { | |||
5676 | Type *SrcTy = VL0->getOperand(0)->getType(); | |||
5677 | for (Value *V : VL) { | |||
5678 | Type *Ty = cast<Instruction>(V)->getOperand(0)->getType(); | |||
5679 | if (Ty != SrcTy || !isValidElementType(Ty)) { | |||
5680 | BS.cancelScheduling(VL, VL0); | |||
5681 | newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx, | |||
5682 | ReuseShuffleIndicies); | |||
5683 | LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering casts with different src types.\n" ; } } while (false) | |||
5684 | << "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); | |||
5685 | return; | |||
5686 | } | |||
5687 | } | |||
5688 | TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, | |||
5689 | ReuseShuffleIndicies); | |||
5690 | 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); | |||
5691 | ||||
5692 | TE->setOperandsInOrder(); | |||
5693 | for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) { | |||
5694 | ValueList Operands; | |||
5695 | // Prepare the operand vector. | |||
5696 | for (Value *V : VL) | |||
5697 | Operands.push_back(cast<Instruction>(V)->getOperand(i)); | |||
5698 | ||||
5699 | buildTree_rec(Operands, Depth + 1, {TE, i}); | |||
5700 | } | |||
5701 | return; | |||
5702 | } | |||
5703 | case Instruction::ICmp: | |||
5704 | case Instruction::FCmp: { | |||
5705 | // Check that all of the compares have the same predicate. | |||
5706 | CmpInst::Predicate P0 = cast<CmpInst>(VL0)->getPredicate(); | |||
5707 | CmpInst::Predicate SwapP0 = CmpInst::getSwappedPredicate(P0); | |||
5708 | Type *ComparedTy = VL0->getOperand(0)->getType(); | |||
5709 | for (Value *V : VL) { | |||
5710 | CmpInst *Cmp = cast<CmpInst>(V); | |||
5711 | if ((Cmp->getPredicate() != P0 && Cmp->getPredicate() != SwapP0) || | |||
5712 | Cmp->getOperand(0)->getType() != ComparedTy) { | |||
5713 | BS.cancelScheduling(VL, VL0); | |||
5714 | newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx, | |||
5715 | ReuseShuffleIndicies); | |||
5716 | LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering cmp with different predicate.\n" ; } } while (false) | |||
5717 | << "SLP: Gathering cmp with different predicate.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering cmp with different predicate.\n" ; } } while (false); | |||
5718 | return; | |||
5719 | } | |||
5720 | } | |||
5721 | ||||
5722 | TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, | |||
5723 | ReuseShuffleIndicies); | |||
5724 | 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); | |||
5725 | ||||
5726 | ValueList Left, Right; | |||
5727 | if (cast<CmpInst>(VL0)->isCommutative()) { | |||
5728 | // Commutative predicate - collect + sort operands of the instructions | |||
5729 | // so that each side is more likely to have the same opcode. | |||
5730 | 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", 5730, __extension__ __PRETTY_FUNCTION__)); | |||
5731 | reorderInputsAccordingToOpcode(VL, Left, Right, *TLI, *DL, *SE, *this); | |||
5732 | } else { | |||
5733 | // Collect operands - commute if it uses the swapped predicate. | |||
5734 | for (Value *V : VL) { | |||
5735 | auto *Cmp = cast<CmpInst>(V); | |||
5736 | Value *LHS = Cmp->getOperand(0); | |||
5737 | Value *RHS = Cmp->getOperand(1); | |||
5738 | if (Cmp->getPredicate() != P0) | |||
5739 | std::swap(LHS, RHS); | |||
5740 | Left.push_back(LHS); | |||
5741 | Right.push_back(RHS); | |||
5742 | } | |||
5743 | } | |||
5744 | TE->setOperand(0, Left); | |||
5745 | TE->setOperand(1, Right); | |||
5746 | buildTree_rec(Left, Depth + 1, {TE, 0}); | |||
5747 | buildTree_rec(Right, Depth + 1, {TE, 1}); | |||
5748 | return; | |||
5749 | } | |||
5750 | case Instruction::Select: | |||
5751 | case Instruction::FNeg: | |||
5752 | case Instruction::Add: | |||
5753 | case Instruction::FAdd: | |||
5754 | case Instruction::Sub: | |||
5755 | case Instruction::FSub: | |||
5756 | case Instruction::Mul: | |||
5757 | case Instruction::FMul: | |||
5758 | case Instruction::UDiv: | |||
5759 | case Instruction::SDiv: | |||
5760 | case Instruction::FDiv: | |||
5761 | case Instruction::URem: | |||
5762 | case Instruction::SRem: | |||
5763 | case Instruction::FRem: | |||
5764 | case Instruction::Shl: | |||
5765 | case Instruction::LShr: | |||
5766 | case Instruction::AShr: | |||
5767 | case Instruction::And: | |||
5768 | case Instruction::Or: | |||
5769 | case Instruction::Xor: { | |||
5770 | TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, | |||
5771 | ReuseShuffleIndicies); | |||
5772 | 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); | |||
5773 | ||||
5774 | // Sort operands of the instructions so that each side is more likely to | |||
5775 | // have the same opcode. | |||
5776 | if (isa<BinaryOperator>(VL0) && VL0->isCommutative()) { | |||
5777 | ValueList Left, Right; | |||
5778 | reorderInputsAccordingToOpcode(VL, Left, Right, *TLI, *DL, *SE, *this); | |||
5779 | TE->setOperand(0, Left); | |||
5780 | TE->setOperand(1, Right); | |||
5781 | buildTree_rec(Left, Depth + 1, {TE, 0}); | |||
5782 | buildTree_rec(Right, Depth + 1, {TE, 1}); | |||
5783 | return; | |||
5784 | } | |||
5785 | ||||
5786 | TE->setOperandsInOrder(); | |||
5787 | for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) { | |||
5788 | ValueList Operands; | |||
5789 | // Prepare the operand vector. | |||
5790 | for (Value *V : VL) | |||
5791 | Operands.push_back(cast<Instruction>(V)->getOperand(i)); | |||
5792 | ||||
5793 | buildTree_rec(Operands, Depth + 1, {TE, i}); | |||
5794 | } | |||
5795 | return; | |||
5796 | } | |||
5797 | case Instruction::GetElementPtr: { | |||
5798 | // We don't combine GEPs with complicated (nested) indexing. | |||
5799 | for (Value *V : VL) { | |||
5800 | auto *I = dyn_cast<GetElementPtrInst>(V); | |||
5801 | if (!I) | |||
5802 | continue; | |||
5803 | if (I->getNumOperands() != 2) { | |||
5804 | 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); | |||
5805 | BS.cancelScheduling(VL, VL0); | |||
5806 | newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx, | |||
5807 | ReuseShuffleIndicies); | |||
5808 | return; | |||
5809 | } | |||
5810 | } | |||
5811 | ||||
5812 | // We can't combine several GEPs into one vector if they operate on | |||
5813 | // different types. | |||
5814 | Type *Ty0 = cast<GEPOperator>(VL0)->getSourceElementType(); | |||
5815 | for (Value *V : VL) { | |||
5816 | auto *GEP = dyn_cast<GEPOperator>(V); | |||
5817 | if (!GEP) | |||
5818 | continue; | |||
5819 | Type *CurTy = GEP->getSourceElementType(); | |||
5820 | if (Ty0 != CurTy) { | |||
5821 | LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: not-vectorizable GEP (different types).\n" ; } } while (false) | |||
5822 | << "SLP: not-vectorizable GEP (different types).\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: not-vectorizable GEP (different types).\n" ; } } while (false); | |||
5823 | BS.cancelScheduling(VL, VL0); | |||
5824 | newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx, | |||
5825 | ReuseShuffleIndicies); | |||
5826 | return; | |||
5827 | } | |||
5828 | } | |||
5829 | ||||
5830 | // We don't combine GEPs with non-constant indexes. | |||
5831 | Type *Ty1 = VL0->getOperand(1)->getType(); | |||
5832 | for (Value *V : VL) { | |||
5833 | auto *I = dyn_cast<GetElementPtrInst>(V); | |||
5834 | if (!I) | |||
5835 | continue; | |||
5836 | auto *Op = I->getOperand(1); | |||
5837 | if ((!IsScatterVectorizeUserTE && !isa<ConstantInt>(Op)) || | |||
5838 | (Op->getType() != Ty1 && | |||
5839 | ((IsScatterVectorizeUserTE && !isa<ConstantInt>(Op)) || | |||
5840 | Op->getType()->getScalarSizeInBits() > | |||
5841 | DL->getIndexSizeInBits( | |||
5842 | V->getType()->getPointerAddressSpace())))) { | |||
5843 | LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: not-vectorizable GEP (non-constant indexes).\n" ; } } while (false) | |||
5844 | << "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); | |||
5845 | BS.cancelScheduling(VL, VL0); | |||
5846 | newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx, | |||
5847 | ReuseShuffleIndicies); | |||
5848 | return; | |||
5849 | } | |||
5850 | } | |||
5851 | ||||
5852 | TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, | |||
5853 | ReuseShuffleIndicies); | |||
5854 | 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); | |||
5855 | SmallVector<ValueList, 2> Operands(2); | |||
5856 | // Prepare the operand vector for pointer operands. | |||
5857 | for (Value *V : VL) { | |||
5858 | auto *GEP = dyn_cast<GetElementPtrInst>(V); | |||
5859 | if (!GEP) { | |||
5860 | Operands.front().push_back(V); | |||
5861 | continue; | |||
5862 | } | |||
5863 | Operands.front().push_back(GEP->getPointerOperand()); | |||
5864 | } | |||
5865 | TE->setOperand(0, Operands.front()); | |||
5866 | // Need to cast all indices to the same type before vectorization to | |||
5867 | // avoid crash. | |||
5868 | // Required to be able to find correct matches between different gather | |||
5869 | // nodes and reuse the vectorized values rather than trying to gather them | |||
5870 | // again. | |||
5871 | int IndexIdx = 1; | |||
5872 | Type *VL0Ty = VL0->getOperand(IndexIdx)->getType(); | |||
5873 | Type *Ty = all_of(VL, | |||
5874 | [VL0Ty, IndexIdx](Value *V) { | |||
5875 | auto *GEP = dyn_cast<GetElementPtrInst>(V); | |||
5876 | if (!GEP) | |||
5877 | return true; | |||
5878 | return VL0Ty == GEP->getOperand(IndexIdx)->getType(); | |||
5879 | }) | |||
5880 | ? VL0Ty | |||
5881 | : DL->getIndexType(cast<GetElementPtrInst>(VL0) | |||
5882 | ->getPointerOperandType() | |||
5883 | ->getScalarType()); | |||
5884 | // Prepare the operand vector. | |||
5885 | for (Value *V : VL) { | |||
5886 | auto *I = dyn_cast<GetElementPtrInst>(V); | |||
5887 | if (!I) { | |||
5888 | Operands.back().push_back( | |||
5889 | ConstantInt::get(Ty, 0, /*isSigned=*/false)); | |||
5890 | continue; | |||
5891 | } | |||
5892 | auto *Op = I->getOperand(IndexIdx); | |||
5893 | auto *CI = dyn_cast<ConstantInt>(Op); | |||
5894 | if (!CI) | |||
5895 | Operands.back().push_back(Op); | |||
5896 | else | |||
5897 | Operands.back().push_back(ConstantExpr::getIntegerCast( | |||
5898 | CI, Ty, CI->getValue().isSignBitSet())); | |||
5899 | } | |||
5900 | TE->setOperand(IndexIdx, Operands.back()); | |||
5901 | ||||
5902 | for (unsigned I = 0, Ops = Operands.size(); I < Ops; ++I) | |||
5903 | buildTree_rec(Operands[I], Depth + 1, {TE, I}); | |||
5904 | return; | |||
5905 | } | |||
5906 | case Instruction::Store: { | |||
5907 | // Check if the stores are consecutive or if we need to swizzle them. | |||
5908 | llvm::Type *ScalarTy = cast<StoreInst>(VL0)->getValueOperand()->getType(); | |||
5909 | // Avoid types that are padded when being allocated as scalars, while | |||
5910 | // being packed together in a vector (such as i1). | |||
5911 | if (DL->getTypeSizeInBits(ScalarTy) != | |||
5912 | DL->getTypeAllocSizeInBits(ScalarTy)) { | |||
5913 | BS.cancelScheduling(VL, VL0); | |||
5914 | newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx, | |||
5915 | ReuseShuffleIndicies); | |||
5916 | 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); | |||
5917 | return; | |||
5918 | } | |||
5919 | // Make sure all stores in the bundle are simple - we can't vectorize | |||
5920 | // atomic or volatile stores. | |||
5921 | SmallVector<Value *, 4> PointerOps(VL.size()); | |||
5922 | ValueList Operands(VL.size()); | |||
5923 | auto POIter = PointerOps.begin(); | |||
5924 | auto OIter = Operands.begin(); | |||
5925 | for (Value *V : VL) { | |||
5926 | auto *SI = cast<StoreInst>(V); | |||
5927 | if (!SI->isSimple()) { | |||
5928 | BS.cancelScheduling(VL, VL0); | |||
5929 | newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx, | |||
5930 | ReuseShuffleIndicies); | |||
5931 | 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); | |||
5932 | return; | |||
5933 | } | |||
5934 | *POIter = SI->getPointerOperand(); | |||
5935 | *OIter = SI->getValueOperand(); | |||
5936 | ++POIter; | |||
5937 | ++OIter; | |||
5938 | } | |||
5939 | ||||
5940 | OrdersType CurrentOrder; | |||
5941 | // Check the order of pointer operands. | |||
5942 | if (llvm::sortPtrAccesses(PointerOps, ScalarTy, *DL, *SE, CurrentOrder)) { | |||
5943 | Value *Ptr0; | |||
5944 | Value *PtrN; | |||
5945 | if (CurrentOrder.empty()) { | |||
5946 | Ptr0 = PointerOps.front(); | |||
5947 | PtrN = PointerOps.back(); | |||
5948 | } else { | |||
5949 | Ptr0 = PointerOps[CurrentOrder.front()]; | |||
5950 | PtrN = PointerOps[CurrentOrder.back()]; | |||
5951 | } | |||
5952 | std::optional<int> Dist = | |||
5953 | getPointersDiff(ScalarTy, Ptr0, ScalarTy, PtrN, *DL, *SE); | |||
5954 | // Check that the sorted pointer operands are consecutive. | |||
5955 | if (static_cast<unsigned>(*Dist) == VL.size() - 1) { | |||
5956 | if (CurrentOrder.empty()) { | |||
5957 | // Original stores are consecutive and does not require reordering. | |||
5958 | TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, | |||
5959 | UserTreeIdx, ReuseShuffleIndicies); | |||
5960 | TE->setOperandsInOrder(); | |||
5961 | buildTree_rec(Operands, Depth + 1, {TE, 0}); | |||
5962 | 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); | |||
5963 | } else { | |||
5964 | fixupOrderingIndices(CurrentOrder); | |||
5965 | TreeEntry *TE = | |||
5966 | newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, | |||
5967 | ReuseShuffleIndicies, CurrentOrder); | |||
5968 | TE->setOperandsInOrder(); | |||
5969 | buildTree_rec(Operands, Depth + 1, {TE, 0}); | |||
5970 | 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); | |||
5971 | } | |||
5972 | return; | |||
5973 | } | |||
5974 | } | |||
5975 | ||||
5976 | BS.cancelScheduling(VL, VL0); | |||
5977 | newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx, | |||
5978 | ReuseShuffleIndicies); | |||
5979 | LLVM_DEBUG(dbgs() << "SLP: Non-consecutive store.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Non-consecutive store.\n"; } } while (false); | |||
5980 | return; | |||
5981 | } | |||
5982 | case Instruction::Call: { | |||
5983 | // Check if the calls are all to the same vectorizable intrinsic or | |||
5984 | // library function. | |||
5985 | CallInst *CI = cast<CallInst>(VL0); | |||
5986 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); | |||
5987 | ||||
5988 | VFShape Shape = VFShape::get( | |||
5989 | *CI, ElementCount::getFixed(static_cast<unsigned int>(VL.size())), | |||
5990 | false /*HasGlobalPred*/); | |||
5991 | Function *VecFunc = VFDatabase(*CI).getVectorizedFunction(Shape); | |||
5992 | ||||
5993 | if (!VecFunc && !isTriviallyVectorizable(ID)) { | |||
5994 | BS.cancelScheduling(VL, VL0); | |||
5995 | newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx, | |||
5996 | ReuseShuffleIndicies); | |||
5997 | LLVM_DEBUG(dbgs() << "SLP: Non-vectorizable call.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Non-vectorizable call.\n"; } } while (false); | |||
5998 | return; | |||
5999 | } | |||
6000 | Function *F = CI->getCalledFunction(); | |||
6001 | unsigned NumArgs = CI->arg_size(); | |||
6002 | SmallVector<Value*, 4> ScalarArgs(NumArgs, nullptr); | |||
6003 | for (unsigned j = 0; j != NumArgs; ++j) | |||
6004 | if (isVectorIntrinsicWithScalarOpAtArg(ID, j)) | |||
6005 | ScalarArgs[j] = CI->getArgOperand(j); | |||
6006 | for (Value *V : VL) { | |||
6007 | CallInst *CI2 = dyn_cast<CallInst>(V); | |||
6008 | if (!CI2 || CI2->getCalledFunction() != F || | |||
6009 | getVectorIntrinsicIDForCall(CI2, TLI) != ID || | |||
6010 | (VecFunc && | |||
6011 | VecFunc != VFDatabase(*CI2).getVectorizedFunction(Shape)) || | |||
6012 | !CI->hasIdenticalOperandBundleSchema(*CI2)) { | |||
6013 | BS.cancelScheduling(VL, VL0); | |||
6014 | newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx, | |||
6015 | ReuseShuffleIndicies); | |||
6016 | LLVM_DEBUG(dbgs() << "SLP: mismatched calls:" << *CI << "!=" << *Vdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: mismatched calls:" << * CI << "!=" << *V << "\n"; } } while (false) | |||
6017 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: mismatched calls:" << * CI << "!=" << *V << "\n"; } } while (false); | |||
6018 | return; | |||
6019 | } | |||
6020 | // Some intrinsics have scalar arguments and should be same in order for | |||
6021 | // them to be vectorized. | |||
6022 | for (unsigned j = 0; j != NumArgs; ++j) { | |||
6023 | if (isVectorIntrinsicWithScalarOpAtArg(ID, j)) { | |||
6024 | Value *A1J = CI2->getArgOperand(j); | |||
6025 | if (ScalarArgs[j] != A1J) { | |||
6026 | BS.cancelScheduling(VL, VL0); | |||
6027 | newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx, | |||
6028 | ReuseShuffleIndicies); | |||
6029 | 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) | |||
6030 | << " argument " << ScalarArgs[j] << "!=" << A1Jdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: mismatched arguments in call:" << *CI << " argument " << ScalarArgs[j] << "!=" << A1J << "\n"; } } while (false) | |||
6031 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: mismatched arguments in call:" << *CI << " argument " << ScalarArgs[j] << "!=" << A1J << "\n"; } } while (false); | |||
6032 | return; | |||
6033 | } | |||
6034 | } | |||
6035 | } | |||
6036 | // Verify that the bundle operands are identical between the two calls. | |||
6037 | if (CI->hasOperandBundles() && | |||
6038 | !std::equal(CI->op_begin() + CI->getBundleOperandsStartIndex(), | |||
6039 | CI->op_begin() + CI->getBundleOperandsEndIndex(), | |||
6040 | CI2->op_begin() + CI2->getBundleOperandsStartIndex())) { | |||
6041 | BS.cancelScheduling(VL, VL0); | |||
6042 | newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx, | |||
6043 | ReuseShuffleIndicies); | |||
6044 | 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) | |||
6045 | << *CI << "!=" << *V << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: mismatched bundle operands in calls:" << *CI << "!=" << *V << '\n'; } } while (false); | |||
6046 | return; | |||
6047 | } | |||
6048 | } | |||
6049 | ||||
6050 | TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, | |||
6051 | ReuseShuffleIndicies); | |||
6052 | TE->setOperandsInOrder(); | |||
6053 | for (unsigned i = 0, e = CI->arg_size(); i != e; ++i) { | |||
6054 | // For scalar operands no need to to create an entry since no need to | |||
6055 | // vectorize it. | |||
6056 | if (isVectorIntrinsicWithScalarOpAtArg(ID, i)) | |||
6057 | continue; | |||
6058 | ValueList Operands; | |||
6059 | // Prepare the operand vector. | |||
6060 | for (Value *V : VL) { | |||
6061 | auto *CI2 = cast<CallInst>(V); | |||
6062 | Operands.push_back(CI2->getArgOperand(i)); | |||
6063 | } | |||
6064 | buildTree_rec(Operands, Depth + 1, {TE, i}); | |||
6065 | } | |||
6066 | return; | |||
6067 | } | |||
6068 | case Instruction::ShuffleVector: { | |||
6069 | // If this is not an alternate sequence of opcode like add-sub | |||
6070 | // then do not vectorize this instruction. | |||
6071 | if (!S.isAltShuffle()) { | |||
6072 | BS.cancelScheduling(VL, VL0); | |||
6073 | newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx, | |||
6074 | ReuseShuffleIndicies); | |||
6075 | 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); | |||
6076 | return; | |||
6077 | } | |||
6078 | TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, | |||
6079 | ReuseShuffleIndicies); | |||
6080 | 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); | |||
6081 | ||||
6082 | // Reorder operands if reordering would enable vectorization. | |||
6083 | auto *CI = dyn_cast<CmpInst>(VL0); | |||
6084 | if (isa<BinaryOperator>(VL0) || CI) { | |||
6085 | ValueList Left, Right; | |||
6086 | if (!CI || all_of(VL, [](Value *V) { | |||
6087 | return cast<CmpInst>(V)->isCommutative(); | |||
6088 | })) { | |||
6089 | reorderInputsAccordingToOpcode(VL, Left, Right, *TLI, *DL, *SE, | |||
6090 | *this); | |||
6091 | } else { | |||
6092 | auto *MainCI = cast<CmpInst>(S.MainOp); | |||
6093 | auto *AltCI = cast<CmpInst>(S.AltOp); | |||
6094 | CmpInst::Predicate MainP = MainCI->getPredicate(); | |||
6095 | CmpInst::Predicate AltP = AltCI->getPredicate(); | |||
6096 | assert(MainP != AltP &&(static_cast <bool> (MainP != AltP && "Expected different main/alternate predicates." ) ? void (0) : __assert_fail ("MainP != AltP && \"Expected different main/alternate predicates.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 6097, __extension__ __PRETTY_FUNCTION__)) | |||
6097 | "Expected different main/alternate predicates.")(static_cast <bool> (MainP != AltP && "Expected different main/alternate predicates." ) ? void (0) : __assert_fail ("MainP != AltP && \"Expected different main/alternate predicates.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 6097, __extension__ __PRETTY_FUNCTION__)); | |||
6098 | // Collect operands - commute if it uses the swapped predicate or | |||
6099 | // alternate operation. | |||
6100 | for (Value *V : VL) { | |||
6101 | auto *Cmp = cast<CmpInst>(V); | |||
6102 | Value *LHS = Cmp->getOperand(0); | |||
6103 | Value *RHS = Cmp->getOperand(1); | |||
6104 | ||||
6105 | if (isAlternateInstruction(Cmp, MainCI, AltCI, *TLI)) { | |||
6106 | if (AltP == CmpInst::getSwappedPredicate(Cmp->getPredicate())) | |||
6107 | std::swap(LHS, RHS); | |||
6108 | } else { | |||
6109 | if (MainP == CmpInst::getSwappedPredicate(Cmp->getPredicate())) | |||
6110 | std::swap(LHS, RHS); | |||
6111 | } | |||
6112 | Left.push_back(LHS); | |||
6113 | Right.push_back(RHS); | |||
6114 | } | |||
6115 | } | |||
6116 | TE->setOperand(0, Left); | |||
6117 | TE->setOperand(1, Right); | |||
6118 | buildTree_rec(Left, Depth + 1, {TE, 0}); | |||
6119 | buildTree_rec(Right, Depth + 1, {TE, 1}); | |||
6120 | return; | |||
6121 | } | |||
6122 | ||||
6123 | TE->setOperandsInOrder(); | |||
6124 | for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) { | |||
6125 | ValueList Operands; | |||
6126 | // Prepare the operand vector. | |||
6127 | for (Value *V : VL) | |||
6128 | Operands.push_back(cast<Instruction>(V)->getOperand(i)); | |||
6129 | ||||
6130 | buildTree_rec(Operands, Depth + 1, {TE, i}); | |||
6131 | } | |||
6132 | return; | |||
6133 | } | |||
6134 | default: | |||
6135 | BS.cancelScheduling(VL, VL0); | |||
6136 | newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx, | |||
6137 | ReuseShuffleIndicies); | |||
6138 | LLVM_DEBUG(dbgs() << "SLP: Gathering unknown instruction.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering unknown instruction.\n" ; } } while (false); | |||
6139 | return; | |||
6140 | } | |||
6141 | } | |||
6142 | ||||
6143 | unsigned BoUpSLP::canMapToVector(Type *T, const DataLayout &DL) const { | |||
6144 | unsigned N = 1; | |||
6145 | Type *EltTy = T; | |||
6146 | ||||
6147 | while (isa<StructType, ArrayType, VectorType>(EltTy)) { | |||
6148 | if (auto *ST = dyn_cast<StructType>(EltTy)) { | |||
6149 | // Check that struct is homogeneous. | |||
6150 | for (const auto *Ty : ST->elements()) | |||
6151 | if (Ty != *ST->element_begin()) | |||
6152 | return 0; | |||
6153 | N *= ST->getNumElements(); | |||
6154 | EltTy = *ST->element_begin(); | |||
6155 | } else if (auto *AT = dyn_cast<ArrayType>(EltTy)) { | |||
6156 | N *= AT->getNumElements(); | |||
6157 | EltTy = AT->getElementType(); | |||
6158 | } else { | |||
6159 | auto *VT = cast<FixedVectorType>(EltTy); | |||
6160 | N *= VT->getNumElements(); | |||
6161 | EltTy = VT->getElementType(); | |||
6162 | } | |||
6163 | } | |||
6164 | ||||
6165 | if (!isValidElementType(EltTy)) | |||
6166 | return 0; | |||
6167 | uint64_t VTSize = DL.getTypeStoreSizeInBits(FixedVectorType::get(EltTy, N)); | |||
6168 | if (VTSize < MinVecRegSize || VTSize > MaxVecRegSize || VTSize != DL.getTypeStoreSizeInBits(T)) | |||
6169 | return 0; | |||
6170 | return N; | |||
6171 | } | |||
6172 | ||||
6173 | bool BoUpSLP::canReuseExtract(ArrayRef<Value *> VL, Value *OpValue, | |||
6174 | SmallVectorImpl<unsigned> &CurrentOrder) const { | |||
6175 | const auto *It = find_if(VL, [](Value *V) { | |||
6176 | return isa<ExtractElementInst, ExtractValueInst>(V); | |||
6177 | }); | |||
6178 | 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", 6178, __extension__ __PRETTY_FUNCTION__)); | |||
6179 | auto *E0 = cast<Instruction>(*It); | |||
6180 | 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", 6185, __extension__ __PRETTY_FUNCTION__)) | |||
6181 | [](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", 6185, __extension__ __PRETTY_FUNCTION__)) | |||
6182 | 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", 6185, __extension__ __PRETTY_FUNCTION__)) | |||
6183 | 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", 6185, __extension__ __PRETTY_FUNCTION__)) | |||
6184 | }) &&(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", 6185, __extension__ __PRETTY_FUNCTION__)) | |||
6185 | "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", 6185, __extension__ __PRETTY_FUNCTION__)); | |||
6186 | // Check if all of the extracts come from the same vector and from the | |||
6187 | // correct offset. | |||
6188 | Value *Vec = E0->getOperand(0); | |||
6189 | ||||
6190 | CurrentOrder.clear(); | |||
6191 | ||||
6192 | // We have to extract from a vector/aggregate with the same number of elements. | |||
6193 | unsigned NElts; | |||
6194 | if (E0->getOpcode() == Instruction::ExtractValue) { | |||
6195 | const DataLayout &DL = E0->getModule()->getDataLayout(); | |||
6196 | NElts = canMapToVector(Vec->getType(), DL); | |||
6197 | if (!NElts) | |||
6198 | return false; | |||
6199 | // Check if load can be rewritten as load of vector. | |||
6200 | LoadInst *LI = dyn_cast<LoadInst>(Vec); | |||
6201 | if (!LI || !LI->isSimple() || !LI->hasNUses(VL.size())) | |||
6202 | return false; | |||
6203 | } else { | |||
6204 | NElts = cast<FixedVectorType>(Vec->getType())->getNumElements(); | |||
6205 | } | |||
6206 | ||||
6207 | if (NElts != VL.size()) | |||
6208 | return false; | |||
6209 | ||||
6210 | // Check that all of the indices extract from the correct offset. | |||
6211 | bool ShouldKeepOrder = true; | |||
6212 | unsigned E = VL.size(); | |||
6213 | // Assign to all items the initial value E + 1 so we can check if the extract | |||
6214 | // instruction index was used already. | |||
6215 | // Also, later we can check that all the indices are used and we have a | |||
6216 | // consecutive access in the extract instructions, by checking that no | |||
6217 | // element of CurrentOrder still has value E + 1. | |||
6218 | CurrentOrder.assign(E, E); | |||
6219 | unsigned I = 0; | |||
6220 | for (; I < E; ++I) { | |||
6221 | auto *Inst = dyn_cast<Instruction>(VL[I]); | |||
6222 | if (!Inst) | |||
6223 | continue; | |||
6224 | if (Inst->getOperand(0) != Vec) | |||
6225 | break; | |||
6226 | if (auto *EE = dyn_cast<ExtractElementInst>(Inst)) | |||
6227 | if (isa<UndefValue>(EE->getIndexOperand())) | |||
6228 | continue; | |||
6229 | std::optional<unsigned> Idx = getExtractIndex(Inst); | |||
6230 | if (!Idx) | |||
6231 | break; | |||
6232 | const unsigned ExtIdx = *Idx; | |||
6233 | if (ExtIdx != I) { | |||
6234 | if (ExtIdx >= E || CurrentOrder[ExtIdx] != E) | |||
6235 | break; | |||
6236 | ShouldKeepOrder = false; | |||
6237 | CurrentOrder[ExtIdx] = I; | |||
6238 | } else { | |||
6239 | if (CurrentOrder[I] != E) | |||
6240 | break; | |||
6241 | CurrentOrder[I] = I; | |||
6242 | } | |||
6243 | } | |||
6244 | if (I < E) { | |||
6245 | CurrentOrder.clear(); | |||
6246 | return false; | |||
6247 | } | |||
6248 | if (ShouldKeepOrder) | |||
6249 | CurrentOrder.clear(); | |||
6250 | ||||
6251 | return ShouldKeepOrder; | |||
6252 | } | |||
6253 | ||||
6254 | bool BoUpSLP::areAllUsersVectorized(Instruction *I, | |||
6255 | ArrayRef<Value *> VectorizedVals) const { | |||
6256 | return (I->hasOneUse() && is_contained(VectorizedVals, I)) || | |||
6257 | all_of(I->users(), [this](User *U) { | |||
6258 | return ScalarToTreeEntry.count(U) > 0 || | |||
6259 | isVectorLikeInstWithConstOps(U) || | |||
6260 | (isa<ExtractElementInst>(U) && MustGather.contains(U)); | |||
6261 | }); | |||
6262 | } | |||
6263 | ||||
6264 | static std::pair<InstructionCost, InstructionCost> | |||
6265 | getVectorCallCosts(CallInst *CI, FixedVectorType *VecTy, | |||
6266 | TargetTransformInfo *TTI, TargetLibraryInfo *TLI) { | |||
6267 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); | |||
6268 | ||||
6269 | // Calculate the cost of the scalar and vector calls. | |||
6270 | SmallVector<Type *, 4> VecTys; | |||
6271 | for (Use &Arg : CI->args()) | |||
6272 | VecTys.push_back( | |||
6273 | FixedVectorType::get(Arg->getType(), VecTy->getNumElements())); | |||
6274 | FastMathFlags FMF; | |||
6275 | if (auto *FPCI = dyn_cast<FPMathOperator>(CI)) | |||
6276 | FMF = FPCI->getFastMathFlags(); | |||
6277 | SmallVector<const Value *> Arguments(CI->args()); | |||
6278 | IntrinsicCostAttributes CostAttrs(ID, VecTy, Arguments, VecTys, FMF, | |||
6279 | dyn_cast<IntrinsicInst>(CI)); | |||
6280 | auto IntrinsicCost = | |||
6281 | TTI->getIntrinsicInstrCost(CostAttrs, TTI::TCK_RecipThroughput); | |||
6282 | ||||
6283 | auto Shape = VFShape::get(*CI, ElementCount::getFixed(static_cast<unsigned>( | |||
6284 | VecTy->getNumElements())), | |||
6285 | false /*HasGlobalPred*/); | |||
6286 | Function *VecFunc = VFDatabase(*CI).getVectorizedFunction(Shape); | |||
6287 | auto LibCost = IntrinsicCost; | |||
6288 | if (!CI->isNoBuiltin() && VecFunc) { | |||
6289 | // Calculate the cost of the vector library call. | |||
6290 | // If the corresponding vector call is cheaper, return its cost. | |||
6291 | LibCost = TTI->getCallInstrCost(nullptr, VecTy, VecTys, | |||
6292 | TTI::TCK_RecipThroughput); | |||
6293 | } | |||
6294 | return {IntrinsicCost, LibCost}; | |||
6295 | } | |||
6296 | ||||
6297 | /// Build shuffle mask for shuffle graph entries and lists of main and alternate | |||
6298 | /// operations operands. | |||
6299 | static void | |||
6300 | buildShuffleEntryMask(ArrayRef<Value *> VL, ArrayRef<unsigned> ReorderIndices, | |||
6301 | ArrayRef<int> ReusesIndices, | |||
6302 | const function_ref<bool(Instruction *)> IsAltOp, | |||
6303 | SmallVectorImpl<int> &Mask, | |||
6304 | SmallVectorImpl<Value *> *OpScalars = nullptr, | |||
6305 | SmallVectorImpl<Value *> *AltScalars = nullptr) { | |||
6306 | unsigned Sz = VL.size(); | |||
6307 | Mask.assign(Sz, PoisonMaskElem); | |||
6308 | SmallVector<int> OrderMask; | |||
6309 | if (!ReorderIndices.empty()) | |||
6310 | inversePermutation(ReorderIndices, OrderMask); | |||
6311 | for (unsigned I = 0; I < Sz; ++I) { | |||
6312 | unsigned Idx = I; | |||
6313 | if (!ReorderIndices.empty()) | |||
6314 | Idx = OrderMask[I]; | |||
6315 | auto *OpInst = cast<Instruction>(VL[Idx]); | |||
6316 | if (IsAltOp(OpInst)) { | |||
6317 | Mask[I] = Sz + Idx; | |||
6318 | if (AltScalars) | |||
6319 | AltScalars->push_back(OpInst); | |||
6320 | } else { | |||
6321 | Mask[I] = Idx; | |||
6322 | if (OpScalars) | |||
6323 | OpScalars->push_back(OpInst); | |||
6324 | } | |||
6325 | } | |||
6326 | if (!ReusesIndices.empty()) { | |||
6327 | SmallVector<int> NewMask(ReusesIndices.size(), PoisonMaskElem); | |||
6328 | transform(ReusesIndices, NewMask.begin(), [&Mask](int Idx) { | |||
6329 | return Idx != PoisonMaskElem ? Mask[Idx] : PoisonMaskElem; | |||
6330 | }); | |||
6331 | Mask.swap(NewMask); | |||
6332 | } | |||
6333 | } | |||
6334 | ||||
6335 | static bool isAlternateInstruction(const Instruction *I, | |||
6336 | const Instruction *MainOp, | |||
6337 | const Instruction *AltOp, | |||
6338 | const TargetLibraryInfo &TLI) { | |||
6339 | if (auto *MainCI = dyn_cast<CmpInst>(MainOp)) { | |||
6340 | auto *AltCI = cast<CmpInst>(AltOp); | |||
6341 | CmpInst::Predicate MainP = MainCI->getPredicate(); | |||
6342 | CmpInst::Predicate AltP = AltCI->getPredicate(); | |||
6343 | assert(MainP != AltP && "Expected different main/alternate predicates.")(static_cast <bool> (MainP != AltP && "Expected different main/alternate predicates." ) ? void (0) : __assert_fail ("MainP != AltP && \"Expected different main/alternate predicates.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 6343, __extension__ __PRETTY_FUNCTION__)); | |||
6344 | auto *CI = cast<CmpInst>(I); | |||
6345 | if (isCmpSameOrSwapped(MainCI, CI, TLI)) | |||
6346 | return false; | |||
6347 | if (isCmpSameOrSwapped(AltCI, CI, TLI)) | |||
6348 | return true; | |||
6349 | CmpInst::Predicate P = CI->getPredicate(); | |||
6350 | CmpInst::Predicate SwappedP = CmpInst::getSwappedPredicate(P); | |||
6351 | ||||
6352 | assert((MainP == P || AltP == P || MainP == SwappedP || AltP == SwappedP) &&(static_cast <bool> ((MainP == P || AltP == P || MainP == SwappedP || AltP == SwappedP) && "CmpInst expected to match either main or alternate predicate or " "their swap.") ? void (0) : __assert_fail ("(MainP == P || AltP == P || MainP == SwappedP || AltP == SwappedP) && \"CmpInst expected to match either main or alternate predicate or \" \"their swap.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 6354, __extension__ __PRETTY_FUNCTION__)) | |||
6353 | "CmpInst expected to match either main or alternate predicate or "(static_cast <bool> ((MainP == P || AltP == P || MainP == SwappedP || AltP == SwappedP) && "CmpInst expected to match either main or alternate predicate or " "their swap.") ? void (0) : __assert_fail ("(MainP == P || AltP == P || MainP == SwappedP || AltP == SwappedP) && \"CmpInst expected to match either main or alternate predicate or \" \"their swap.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 6354, __extension__ __PRETTY_FUNCTION__)) | |||
6354 | "their swap.")(static_cast <bool> ((MainP == P || AltP == P || MainP == SwappedP || AltP == SwappedP) && "CmpInst expected to match either main or alternate predicate or " "their swap.") ? void (0) : __assert_fail ("(MainP == P || AltP == P || MainP == SwappedP || AltP == SwappedP) && \"CmpInst expected to match either main or alternate predicate or \" \"their swap.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 6354, __extension__ __PRETTY_FUNCTION__)); | |||
6355 | (void)AltP; | |||
6356 | return MainP != P && MainP != SwappedP; | |||
6357 | } | |||
6358 | return I->getOpcode() == AltOp->getOpcode(); | |||
6359 | } | |||
6360 | ||||
6361 | TTI::OperandValueInfo BoUpSLP::getOperandInfo(ArrayRef<Value *> VL, | |||
6362 | unsigned OpIdx) { | |||
6363 | assert(!VL.empty())(static_cast <bool> (!VL.empty()) ? void (0) : __assert_fail ("!VL.empty()", "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 6363, __extension__ __PRETTY_FUNCTION__)); | |||
6364 | const auto *I0 = cast<Instruction>(*find_if(VL, Instruction::classof)); | |||
6365 | const auto *Op0 = I0->getOperand(OpIdx); | |||
6366 | ||||
6367 | const bool IsConstant = all_of(VL, [&](Value *V) { | |||
6368 | // TODO: We should allow undef elements here | |||
6369 | const auto *I = dyn_cast<Instruction>(V); | |||
6370 | if (!I) | |||
6371 | return true; | |||
6372 | auto *Op = I->getOperand(OpIdx); | |||
6373 | return isConstant(Op) && !isa<UndefValue>(Op); | |||
6374 | }); | |||
6375 | const bool IsUniform = all_of(VL, [&](Value *V) { | |||
6376 | // TODO: We should allow undef elements here | |||
6377 | const auto *I = dyn_cast<Instruction>(V); | |||
6378 | if (!I) | |||
6379 | return false; | |||
6380 | return I->getOperand(OpIdx) == Op0; | |||
6381 | }); | |||
6382 | const bool IsPowerOfTwo = all_of(VL, [&](Value *V) { | |||
6383 | // TODO: We should allow undef elements here | |||
6384 | const auto *I = dyn_cast<Instruction>(V); | |||
6385 | if (!I) { | |||
6386 | assert((isa<UndefValue>(V) ||(static_cast <bool> ((isa<UndefValue>(V) || I0-> getOpcode() == Instruction::GetElementPtr) && "Expected undef or GEP." ) ? void (0) : __assert_fail ("(isa<UndefValue>(V) || I0->getOpcode() == Instruction::GetElementPtr) && \"Expected undef or GEP.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 6388, __extension__ __PRETTY_FUNCTION__)) | |||
6387 | I0->getOpcode() == Instruction::GetElementPtr) &&(static_cast <bool> ((isa<UndefValue>(V) || I0-> getOpcode() == Instruction::GetElementPtr) && "Expected undef or GEP." ) ? void (0) : __assert_fail ("(isa<UndefValue>(V) || I0->getOpcode() == Instruction::GetElementPtr) && \"Expected undef or GEP.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 6388, __extension__ __PRETTY_FUNCTION__)) | |||
6388 | "Expected undef or GEP.")(static_cast <bool> ((isa<UndefValue>(V) || I0-> getOpcode() == Instruction::GetElementPtr) && "Expected undef or GEP." ) ? void (0) : __assert_fail ("(isa<UndefValue>(V) || I0->getOpcode() == Instruction::GetElementPtr) && \"Expected undef or GEP.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 6388, __extension__ __PRETTY_FUNCTION__)); | |||
6389 | return true; | |||
6390 | } | |||
6391 | auto *Op = I->getOperand(OpIdx); | |||
6392 | if (auto *CI = dyn_cast<ConstantInt>(Op)) | |||
6393 | return CI->getValue().isPowerOf2(); | |||
6394 | return false; | |||
6395 | }); | |||
6396 | const bool IsNegatedPowerOfTwo = all_of(VL, [&](Value *V) { | |||
6397 | // TODO: We should allow undef elements here | |||
6398 | const auto *I = dyn_cast<Instruction>(V); | |||
6399 | if (!I) { | |||
6400 | assert((isa<UndefValue>(V) ||(static_cast <bool> ((isa<UndefValue>(V) || I0-> getOpcode() == Instruction::GetElementPtr) && "Expected undef or GEP." ) ? void (0) : __assert_fail ("(isa<UndefValue>(V) || I0->getOpcode() == Instruction::GetElementPtr) && \"Expected undef or GEP.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 6402, __extension__ __PRETTY_FUNCTION__)) | |||
6401 | I0->getOpcode() == Instruction::GetElementPtr) &&(static_cast <bool> ((isa<UndefValue>(V) || I0-> getOpcode() == Instruction::GetElementPtr) && "Expected undef or GEP." ) ? void (0) : __assert_fail ("(isa<UndefValue>(V) || I0->getOpcode() == Instruction::GetElementPtr) && \"Expected undef or GEP.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 6402, __extension__ __PRETTY_FUNCTION__)) | |||
6402 | "Expected undef or GEP.")(static_cast <bool> ((isa<UndefValue>(V) || I0-> getOpcode() == Instruction::GetElementPtr) && "Expected undef or GEP." ) ? void (0) : __assert_fail ("(isa<UndefValue>(V) || I0->getOpcode() == Instruction::GetElementPtr) && \"Expected undef or GEP.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 6402, __extension__ __PRETTY_FUNCTION__)); | |||
6403 | return true; | |||
6404 | } | |||
6405 | const auto *Op = I->getOperand(OpIdx); | |||
6406 | if (auto *CI = dyn_cast<ConstantInt>(Op)) | |||
6407 | return CI->getValue().isNegatedPowerOf2(); | |||
6408 | return false; | |||
6409 | }); | |||
6410 | ||||
6411 | TTI::OperandValueKind VK = TTI::OK_AnyValue; | |||
6412 | if (IsConstant && IsUniform) | |||
6413 | VK = TTI::OK_UniformConstantValue; | |||
6414 | else if (IsConstant) | |||
6415 | VK = TTI::OK_NonUniformConstantValue; | |||
6416 | else if (IsUniform) | |||
6417 | VK = TTI::OK_UniformValue; | |||
6418 | ||||
6419 | TTI::OperandValueProperties VP = TTI::OP_None; | |||
6420 | VP = IsPowerOfTwo ? TTI::OP_PowerOf2 : VP; | |||
6421 | VP = IsNegatedPowerOfTwo ? TTI::OP_NegatedPowerOf2 : VP; | |||
6422 | ||||
6423 | return {VK, VP}; | |||
6424 | } | |||
6425 | ||||
6426 | namespace { | |||
6427 | /// The base class for shuffle instruction emission and shuffle cost estimation. | |||
6428 | class BaseShuffleAnalysis { | |||
6429 | protected: | |||
6430 | /// Checks if the mask is an identity mask. | |||
6431 | /// \param IsStrict if is true the function returns false if mask size does | |||
6432 | /// not match vector size. | |||
6433 | static bool isIdentityMask(ArrayRef<int> Mask, const FixedVectorType *VecTy, | |||
6434 | bool IsStrict) { | |||
6435 | int Limit = Mask.size(); | |||
6436 | int VF = VecTy->getNumElements(); | |||
6437 | return (VF == Limit || !IsStrict) && | |||
6438 | all_of(Mask, [Limit](int Idx) { return Idx < Limit; }) && | |||
6439 | ShuffleVectorInst::isIdentityMask(Mask); | |||
6440 | } | |||
6441 | ||||
6442 | /// Tries to combine 2 different masks into single one. | |||
6443 | /// \param LocalVF Vector length of the permuted input vector. \p Mask may | |||
6444 | /// change the size of the vector, \p LocalVF is the original size of the | |||
6445 | /// shuffled vector. | |||
6446 | static void combineMasks(unsigned LocalVF, SmallVectorImpl<int> &Mask, | |||
6447 | ArrayRef<int> ExtMask) { | |||
6448 | unsigned VF = Mask.size(); | |||
6449 | SmallVector<int> NewMask(ExtMask.size(), PoisonMaskElem); | |||
6450 | for (int I = 0, Sz = ExtMask.size(); I < Sz; ++I) { | |||
6451 | if (ExtMask[I] == PoisonMaskElem) | |||
6452 | continue; | |||
6453 | int MaskedIdx = Mask[ExtMask[I] % VF]; | |||
6454 | NewMask[I] = | |||
6455 | MaskedIdx == PoisonMaskElem ? PoisonMaskElem : MaskedIdx % LocalVF; | |||
6456 | } | |||
6457 | Mask.swap(NewMask); | |||
6458 | } | |||
6459 | ||||
6460 | /// Looks through shuffles trying to reduce final number of shuffles in the | |||
6461 | /// code. The function looks through the previously emitted shuffle | |||
6462 | /// instructions and properly mark indices in mask as undef. | |||
6463 | /// For example, given the code | |||
6464 | /// \code | |||
6465 | /// %s1 = shufflevector <2 x ty> %0, poison, <1, 0> | |||
6466 | /// %s2 = shufflevector <2 x ty> %1, poison, <1, 0> | |||
6467 | /// \endcode | |||
6468 | /// and if need to emit shuffle of %s1 and %s2 with mask <1, 0, 3, 2>, it will | |||
6469 | /// look through %s1 and %s2 and select vectors %0 and %1 with mask | |||
6470 | /// <0, 1, 2, 3> for the shuffle. | |||
6471 | /// If 2 operands are of different size, the smallest one will be resized and | |||
6472 | /// the mask recalculated properly. | |||
6473 | /// For example, given the code | |||
6474 | /// \code | |||
6475 | /// %s1 = shufflevector <2 x ty> %0, poison, <1, 0, 1, 0> | |||
6476 | /// %s2 = shufflevector <2 x ty> %1, poison, <1, 0, 1, 0> | |||
6477 | /// \endcode | |||
6478 | /// and if need to emit shuffle of %s1 and %s2 with mask <1, 0, 5, 4>, it will | |||
6479 | /// look through %s1 and %s2 and select vectors %0 and %1 with mask | |||
6480 | /// <0, 1, 2, 3> for the shuffle. | |||
6481 | /// So, it tries to transform permutations to simple vector merge, if | |||
6482 | /// possible. | |||
6483 | /// \param V The input vector which must be shuffled using the given \p Mask. | |||
6484 | /// If the better candidate is found, \p V is set to this best candidate | |||
6485 | /// vector. | |||
6486 | /// \param Mask The input mask for the shuffle. If the best candidate is found | |||
6487 | /// during looking-through-shuffles attempt, it is updated accordingly. | |||
6488 | /// \param SinglePermute true if the shuffle operation is originally a | |||
6489 | /// single-value-permutation. In this case the look-through-shuffles procedure | |||
6490 | /// may look for resizing shuffles as the best candidates. | |||
6491 | /// \return true if the shuffle results in the non-resizing identity shuffle | |||
6492 | /// (and thus can be ignored), false - otherwise. | |||
6493 | static bool peekThroughShuffles(Value *&V, SmallVectorImpl<int> &Mask, | |||
6494 | bool SinglePermute) { | |||
6495 | Value *Op = V; | |||
6496 | ShuffleVectorInst *IdentityOp = nullptr; | |||
6497 | SmallVector<int> IdentityMask; | |||
6498 | while (auto *SV = dyn_cast<ShuffleVectorInst>(Op)) { | |||
6499 | // Exit if not a fixed vector type or changing size shuffle. | |||
6500 | auto *SVTy = dyn_cast<FixedVectorType>(SV->getType()); | |||
6501 | if (!SVTy) | |||
6502 | break; | |||
6503 | // Remember the identity or broadcast mask, if it is not a resizing | |||
6504 | // shuffle. If no better candidates are found, this Op and Mask will be | |||
6505 | // used in the final shuffle. | |||
6506 | if (isIdentityMask(Mask, SVTy, /*IsStrict=*/false)) { | |||
6507 | if (!IdentityOp || !SinglePermute || | |||
6508 | (isIdentityMask(Mask, SVTy, /*IsStrict=*/true) && | |||
6509 | !ShuffleVectorInst::isZeroEltSplatMask(IdentityMask))) { | |||
6510 | IdentityOp = SV; | |||
6511 | // Store current mask in the IdentityMask so later we did not lost | |||
6512 | // this info if IdentityOp is selected as the best candidate for the | |||
6513 | // permutation. | |||
6514 | IdentityMask.assign(Mask); | |||
6515 | } | |||
6516 | } | |||
6517 | // Remember the broadcast mask. If no better candidates are found, this Op | |||
6518 | // and Mask will be used in the final shuffle. | |||
6519 | // Zero splat can be used as identity too, since it might be used with | |||
6520 | // mask <0, 1, 2, ...>, i.e. identity mask without extra reshuffling. | |||
6521 | // E.g. if need to shuffle the vector with the mask <3, 1, 2, 0>, which is | |||
6522 | // expensive, the analysis founds out, that the source vector is just a | |||
6523 | // broadcast, this original mask can be transformed to identity mask <0, | |||
6524 | // 1, 2, 3>. | |||
6525 | // \code | |||
6526 | // %0 = shuffle %v, poison, zeroinitalizer | |||
6527 | // %res = shuffle %0, poison, <3, 1, 2, 0> | |||
6528 | // \endcode | |||
6529 | // may be transformed to | |||
6530 | // \code | |||
6531 | // %0 = shuffle %v, poison, zeroinitalizer | |||
6532 | // %res = shuffle %0, poison, <0, 1, 2, 3> | |||
6533 | // \endcode | |||
6534 | if (SV->isZeroEltSplat()) { | |||
6535 | IdentityOp = SV; | |||
6536 | IdentityMask.assign(Mask); | |||
6537 | } | |||
6538 | int LocalVF = Mask.size(); | |||
6539 | if (auto *SVOpTy = | |||
6540 | dyn_cast<FixedVectorType>(SV->getOperand(0)->getType())) | |||
6541 | LocalVF = SVOpTy->getNumElements(); | |||
6542 | SmallVector<int> ExtMask(Mask.size(), PoisonMaskElem); | |||
6543 | for (auto [Idx, I] : enumerate(Mask)) { | |||
6544 | if (I == PoisonMaskElem || | |||
6545 | static_cast<unsigned>(I) >= SV->getShuffleMask().size()) | |||
6546 | continue; | |||
6547 | ExtMask[Idx] = SV->getMaskValue(I); | |||
6548 | } | |||
6549 | bool IsOp1Undef = | |||
6550 | isUndefVector(SV->getOperand(0), | |||
6551 | buildUseMask(LocalVF, ExtMask, UseMask::FirstArg)) | |||
6552 | .all(); | |||
6553 | bool IsOp2Undef = | |||
6554 | isUndefVector(SV->getOperand(1), | |||
6555 | buildUseMask(LocalVF, ExtMask, UseMask::SecondArg)) | |||
6556 | .all(); | |||
6557 | if (!IsOp1Undef && !IsOp2Undef) { | |||
6558 | // Update mask and mark undef elems. | |||
6559 | for (int &I : Mask) { | |||
6560 | if (I == PoisonMaskElem) | |||
6561 | continue; | |||
6562 | if (SV->getMaskValue(I % SV->getShuffleMask().size()) == | |||
6563 | PoisonMaskElem) | |||
6564 | I = PoisonMaskElem; | |||
6565 | } | |||
6566 | break; | |||
6567 | } | |||
6568 | SmallVector<int> ShuffleMask(SV->getShuffleMask().begin(), | |||
6569 | SV->getShuffleMask().end()); | |||
6570 | combineMasks(LocalVF, ShuffleMask, Mask); | |||
6571 | Mask.swap(ShuffleMask); | |||
6572 | if (IsOp2Undef) | |||
6573 | Op = SV->getOperand(0); | |||
6574 | else | |||
6575 | Op = SV->getOperand(1); | |||
6576 | } | |||
6577 | if (auto *OpTy = dyn_cast<FixedVectorType>(Op->getType()); | |||
6578 | !OpTy || !isIdentityMask(Mask, OpTy, SinglePermute) || | |||
6579 | ShuffleVectorInst::isZeroEltSplatMask(Mask)) { | |||
6580 | if (IdentityOp) { | |||
6581 | V = IdentityOp; | |||
6582 | assert(Mask.size() == IdentityMask.size() &&(static_cast <bool> (Mask.size() == IdentityMask.size() && "Expected masks of same sizes.") ? void (0) : __assert_fail ("Mask.size() == IdentityMask.size() && \"Expected masks of same sizes.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 6583, __extension__ __PRETTY_FUNCTION__)) | |||
6583 | "Expected masks of same sizes.")(static_cast <bool> (Mask.size() == IdentityMask.size() && "Expected masks of same sizes.") ? void (0) : __assert_fail ("Mask.size() == IdentityMask.size() && \"Expected masks of same sizes.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 6583, __extension__ __PRETTY_FUNCTION__)); | |||
6584 | // Clear known poison elements. | |||
6585 | for (auto [I, Idx] : enumerate(Mask)) | |||
6586 | if (Idx == PoisonMaskElem) | |||
6587 | IdentityMask[I] = PoisonMaskElem; | |||
6588 | Mask.swap(IdentityMask); | |||
6589 | auto *Shuffle = dyn_cast<ShuffleVectorInst>(V); | |||
6590 | return SinglePermute && | |||
6591 | (isIdentityMask(Mask, cast<FixedVectorType>(V->getType()), | |||
6592 | /*IsStrict=*/true) || | |||
6593 | (Shuffle && Mask.size() == Shuffle->getShuffleMask().size() && | |||
6594 | Shuffle->isZeroEltSplat() && | |||
6595 | ShuffleVectorInst::isZeroEltSplatMask(Mask))); | |||
6596 | } | |||
6597 | V = Op; | |||
6598 | return false; | |||
6599 | } | |||
6600 | V = Op; | |||
6601 | return true; | |||
6602 | } | |||
6603 | ||||
6604 | /// Smart shuffle instruction emission, walks through shuffles trees and | |||
6605 | /// tries to find the best matching vector for the actual shuffle | |||
6606 | /// instruction. | |||
6607 | template <typename T, typename ShuffleBuilderTy> | |||
6608 | static T createShuffle(Value *V1, Value *V2, ArrayRef<int> Mask, | |||
6609 | ShuffleBuilderTy &Builder) { | |||
6610 | assert(V1 && "Expected at least one vector value.")(static_cast <bool> (V1 && "Expected at least one vector value." ) ? void (0) : __assert_fail ("V1 && \"Expected at least one vector value.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 6610, __extension__ __PRETTY_FUNCTION__)); | |||
6611 | if (V2) | |||
6612 | Builder.resizeToMatch(V1, V2); | |||
6613 | int VF = Mask.size(); | |||
6614 | if (auto *FTy = dyn_cast<FixedVectorType>(V1->getType())) | |||
6615 | VF = FTy->getNumElements(); | |||
6616 | if (V2 && | |||
6617 | !isUndefVector(V2, buildUseMask(VF, Mask, UseMask::SecondArg)).all()) { | |||
6618 | // Peek through shuffles. | |||
6619 | Value *Op1 = V1; | |||
6620 | Value *Op2 = V2; | |||
6621 | int VF = | |||
6622 | cast<VectorType>(V1->getType())->getElementCount().getKnownMinValue(); | |||
6623 | SmallVector<int> CombinedMask1(Mask.size(), PoisonMaskElem); | |||
6624 | SmallVector<int> CombinedMask2(Mask.size(), PoisonMaskElem); | |||
6625 | for (int I = 0, E = Mask.size(); I < E; ++I) { | |||
6626 | if (Mask[I] < VF) | |||
6627 | CombinedMask1[I] = Mask[I]; | |||
6628 | else | |||
6629 | CombinedMask2[I] = Mask[I] - VF; | |||
6630 | } | |||
6631 | Value *PrevOp1; | |||
6632 | Value *PrevOp2; | |||
6633 | do { | |||
6634 | PrevOp1 = Op1; | |||
6635 | PrevOp2 = Op2; | |||
6636 | (void)peekThroughShuffles(Op1, CombinedMask1, /*SinglePermute=*/false); | |||
6637 | (void)peekThroughShuffles(Op2, CombinedMask2, /*SinglePermute=*/false); | |||
6638 | // Check if we have 2 resizing shuffles - need to peek through operands | |||
6639 | // again. | |||
6640 | if (auto *SV1 = dyn_cast<ShuffleVectorInst>(Op1)) | |||
6641 | if (auto *SV2 = dyn_cast<ShuffleVectorInst>(Op2)) { | |||
6642 | SmallVector<int> ExtMask1(Mask.size(), PoisonMaskElem); | |||
6643 | for (auto [Idx, I] : enumerate(CombinedMask1)) { | |||
6644 | if (I == PoisonMaskElem) | |||
6645 | continue; | |||
6646 | ExtMask1[Idx] = SV1->getMaskValue(I); | |||
6647 | } | |||
6648 | SmallBitVector UseMask1 = buildUseMask( | |||
6649 | cast<FixedVectorType>(SV1->getOperand(1)->getType()) | |||
6650 | ->getNumElements(), | |||
6651 | ExtMask1, UseMask::SecondArg); | |||
6652 | SmallVector<int> ExtMask2(CombinedMask2.size(), PoisonMaskElem); | |||
6653 | for (auto [Idx, I] : enumerate(CombinedMask2)) { | |||
6654 | if (I == PoisonMaskElem) | |||
6655 | continue; | |||
6656 | ExtMask2[Idx] = SV2->getMaskValue(I); | |||
6657 | } | |||
6658 | SmallBitVector UseMask2 = buildUseMask( | |||
6659 | cast<FixedVectorType>(SV2->getOperand(1)->getType()) | |||
6660 | ->getNumElements(), | |||
6661 | ExtMask2, UseMask::SecondArg); | |||
6662 | if (SV1->getOperand(0)->getType() == | |||
6663 | SV2->getOperand(0)->getType() && | |||
6664 | SV1->getOperand(0)->getType() != SV1->getType() && | |||
6665 | isUndefVector(SV1->getOperand(1), UseMask1).all() && | |||
6666 | isUndefVector(SV2->getOperand(1), UseMask2).all()) { | |||
6667 | Op1 = SV1->getOperand(0); | |||
6668 | Op2 = SV2->getOperand(0); | |||
6669 | SmallVector<int> ShuffleMask1(SV1->getShuffleMask().begin(), | |||
6670 | SV1->getShuffleMask().end()); | |||
6671 | int LocalVF = ShuffleMask1.size(); | |||
6672 | if (auto *FTy = dyn_cast<FixedVectorType>(Op1->getType())) | |||
6673 | LocalVF = FTy->getNumElements(); | |||
6674 | combineMasks(LocalVF, ShuffleMask1, CombinedMask1); | |||
6675 | CombinedMask1.swap(ShuffleMask1); | |||
6676 | SmallVector<int> ShuffleMask2(SV2->getShuffleMask().begin(), | |||
6677 | SV2->getShuffleMask().end()); | |||
6678 | LocalVF = ShuffleMask2.size(); | |||
6679 | if (auto *FTy = dyn_cast<FixedVectorType>(Op2->getType())) | |||
6680 | LocalVF = FTy->getNumElements(); | |||
6681 | combineMasks(LocalVF, ShuffleMask2, CombinedMask2); | |||
6682 | CombinedMask2.swap(ShuffleMask2); | |||
6683 | } | |||
6684 | } | |||
6685 | } while (PrevOp1 != Op1 || PrevOp2 != Op2); | |||
6686 | Builder.resizeToMatch(Op1, Op2); | |||
6687 | VF = std::max(cast<VectorType>(Op1->getType()) | |||
6688 | ->getElementCount() | |||
6689 | .getKnownMinValue(), | |||
6690 | cast<VectorType>(Op2->getType()) | |||
6691 | ->getElementCount() | |||
6692 | .getKnownMinValue()); | |||
6693 | for (int I = 0, E = Mask.size(); I < E; ++I) { | |||
6694 | if (CombinedMask2[I] != PoisonMaskElem) { | |||
6695 | assert(CombinedMask1[I] == PoisonMaskElem &&(static_cast <bool> (CombinedMask1[I] == PoisonMaskElem && "Expected undefined mask element") ? void (0) : __assert_fail ("CombinedMask1[I] == PoisonMaskElem && \"Expected undefined mask element\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 6696, __extension__ __PRETTY_FUNCTION__)) | |||
6696 | "Expected undefined mask element")(static_cast <bool> (CombinedMask1[I] == PoisonMaskElem && "Expected undefined mask element") ? void (0) : __assert_fail ("CombinedMask1[I] == PoisonMaskElem && \"Expected undefined mask element\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 6696, __extension__ __PRETTY_FUNCTION__)); | |||
6697 | CombinedMask1[I] = CombinedMask2[I] + (Op1 == Op2 ? 0 : VF); | |||
6698 | } | |||
6699 | } | |||
6700 | const int Limit = CombinedMask1.size() * 2; | |||
6701 | if (Op1 == Op2 && Limit == 2 * VF && | |||
6702 | all_of(CombinedMask1, [=](int Idx) { return Idx < Limit; }) && | |||
6703 | (ShuffleVectorInst::isIdentityMask(CombinedMask1) || | |||
6704 | (ShuffleVectorInst::isZeroEltSplatMask(CombinedMask1) && | |||
6705 | isa<ShuffleVectorInst>(Op1) && | |||
6706 | cast<ShuffleVectorInst>(Op1)->getShuffleMask() == | |||
6707 | ArrayRef(CombinedMask1)))) | |||
6708 | return Builder.createIdentity(Op1); | |||
6709 | return Builder.createShuffleVector( | |||
6710 | Op1, Op1 == Op2 ? PoisonValue::get(Op1->getType()) : Op2, | |||
6711 | CombinedMask1); | |||
6712 | } | |||
6713 | if (isa<PoisonValue>(V1)) | |||
6714 | return Builder.createPoison( | |||
6715 | cast<VectorType>(V1->getType())->getElementType(), Mask.size()); | |||
6716 | SmallVector<int> NewMask(Mask.begin(), Mask.end()); | |||
6717 | bool IsIdentity = peekThroughShuffles(V1, NewMask, /*SinglePermute=*/true); | |||
6718 | assert(V1 && "Expected non-null value after looking through shuffles.")(static_cast <bool> (V1 && "Expected non-null value after looking through shuffles." ) ? void (0) : __assert_fail ("V1 && \"Expected non-null value after looking through shuffles.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 6718, __extension__ __PRETTY_FUNCTION__)); | |||
6719 | ||||
6720 | if (!IsIdentity) | |||
6721 | return Builder.createShuffleVector(V1, NewMask); | |||
6722 | return Builder.createIdentity(V1); | |||
6723 | } | |||
6724 | }; | |||
6725 | } // namespace | |||
6726 | ||||
6727 | /// Merges shuffle masks and emits final shuffle instruction, if required. It | |||
6728 | /// supports shuffling of 2 input vectors. It implements lazy shuffles emission, | |||
6729 | /// when the actual shuffle instruction is generated only if this is actually | |||
6730 | /// required. Otherwise, the shuffle instruction emission is delayed till the | |||
6731 | /// end of the process, to reduce the number of emitted instructions and further | |||
6732 | /// analysis/transformations. | |||
6733 | class BoUpSLP::ShuffleCostEstimator : public BaseShuffleAnalysis { | |||
6734 | bool IsFinalized = false; | |||
6735 | SmallVector<int> CommonMask; | |||
6736 | SmallVector<PointerUnion<Value *, const TreeEntry *> , 2> InVectors; | |||
6737 | const TargetTransformInfo &TTI; | |||
6738 | InstructionCost Cost = 0; | |||
6739 | ArrayRef<Value *> VectorizedVals; | |||
6740 | BoUpSLP &R; | |||
6741 | SmallPtrSetImpl<Value *> &CheckedExtracts; | |||
6742 | constexpr static TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput; | |||
6743 | ||||
6744 | InstructionCost getBuildVectorCost(ArrayRef<Value *> VL, Value *Root) { | |||
6745 | if ((!Root && allConstant(VL)) || all_of(VL, UndefValue::classof)) | |||
6746 | return TTI::TCC_Free; | |||
6747 | auto *VecTy = FixedVectorType::get(VL.front()->getType(), VL.size()); | |||
6748 | InstructionCost GatherCost = 0; | |||
6749 | SmallVector<Value *> Gathers(VL.begin(), VL.end()); | |||
6750 | // Improve gather cost for gather of loads, if we can group some of the | |||
6751 | // loads into vector loads. | |||
6752 | InstructionsState S = getSameOpcode(VL, *R.TLI); | |||
6753 | if (VL.size() > 2 && S.getOpcode() == Instruction::Load && | |||
6754 | !S.isAltShuffle() && | |||
6755 | !all_of(Gathers, [&](Value *V) { return R.getTreeEntry(V); }) && | |||
6756 | !isSplat(Gathers)) { | |||
6757 | BoUpSLP::ValueSet VectorizedLoads; | |||
6758 | unsigned StartIdx = 0; | |||
6759 | unsigned VF = VL.size() / 2; | |||
6760 | unsigned VectorizedCnt = 0; | |||
6761 | unsigned ScatterVectorizeCnt = 0; | |||
6762 | const unsigned Sz = R.DL->getTypeSizeInBits(S.MainOp->getType()); | |||
6763 | for (unsigned MinVF = R.getMinVF(2 * Sz); VF >= MinVF; VF /= 2) { | |||
6764 | for (unsigned Cnt = StartIdx, End = VL.size(); Cnt + VF <= End; | |||
6765 | Cnt += VF) { | |||
6766 | ArrayRef<Value *> Slice = VL.slice(Cnt, VF); | |||
6767 | if (!VectorizedLoads.count(Slice.front()) && | |||
6768 | !VectorizedLoads.count(Slice.back()) && allSameBlock(Slice)) { | |||
6769 | SmallVector<Value *> PointerOps; | |||
6770 | OrdersType CurrentOrder; | |||
6771 | LoadsState LS = | |||
6772 | canVectorizeLoads(Slice, Slice.front(), TTI, *R.DL, *R.SE, | |||
6773 | *R.LI, *R.TLI, CurrentOrder, PointerOps); | |||
6774 | switch (LS) { | |||
6775 | case LoadsState::Vectorize: | |||
6776 | case LoadsState::ScatterVectorize: | |||
6777 | // Mark the vectorized loads so that we don't vectorize them | |||
6778 | // again. | |||
6779 | if (LS == LoadsState::Vectorize) | |||
6780 | ++VectorizedCnt; | |||
6781 | else | |||
6782 | ++ScatterVectorizeCnt; | |||
6783 | VectorizedLoads.insert(Slice.begin(), Slice.end()); | |||
6784 | // If we vectorized initial block, no need to try to vectorize | |||
6785 | // it again. | |||
6786 | if (Cnt == StartIdx) | |||
6787 | StartIdx += VF; | |||
6788 | break; | |||
6789 | case LoadsState::Gather: | |||
6790 | break; | |||
6791 | } | |||
6792 | } | |||
6793 | } | |||
6794 | // Check if the whole array was vectorized already - exit. | |||
6795 | if (StartIdx >= VL.size()) | |||
6796 | break; | |||
6797 | // Found vectorizable parts - exit. | |||
6798 | if (!VectorizedLoads.empty()) | |||
6799 | break; | |||
6800 | } | |||
6801 | if (!VectorizedLoads.empty()) { | |||
6802 | unsigned NumParts = TTI.getNumberOfParts(VecTy); | |||
6803 | bool NeedInsertSubvectorAnalysis = | |||
6804 | !NumParts || (VL.size() / VF) > NumParts; | |||
6805 | // Get the cost for gathered loads. | |||
6806 | for (unsigned I = 0, End = VL.size(); I < End; I += VF) { | |||
6807 | if (VectorizedLoads.contains(VL[I])) | |||
6808 | continue; | |||
6809 | GatherCost += getBuildVectorCost(VL.slice(I, VF), Root); | |||
6810 | } | |||
6811 | // Exclude potentially vectorized loads from list of gathered | |||
6812 | // scalars. | |||
6813 | auto *LI = cast<LoadInst>(S.MainOp); | |||
6814 | Gathers.assign(Gathers.size(), PoisonValue::get(LI->getType())); | |||
6815 | // The cost for vectorized loads. | |||
6816 | InstructionCost ScalarsCost = 0; | |||
6817 | for (Value *V : VectorizedLoads) { | |||
6818 | auto *LI = cast<LoadInst>(V); | |||
6819 | ScalarsCost += | |||
6820 | TTI.getMemoryOpCost(Instruction::Load, LI->getType(), | |||
6821 | LI->getAlign(), LI->getPointerAddressSpace(), | |||
6822 | CostKind, TTI::OperandValueInfo(), LI); | |||
6823 | } | |||
6824 | auto *LoadTy = FixedVectorType::get(LI->getType(), VF); | |||
6825 | Align Alignment = LI->getAlign(); | |||
6826 | GatherCost += | |||
6827 | VectorizedCnt * | |||
6828 | TTI.getMemoryOpCost(Instruction::Load, LoadTy, Alignment, | |||
6829 | LI->getPointerAddressSpace(), CostKind, | |||
6830 | TTI::OperandValueInfo(), LI); | |||
6831 | GatherCost += ScatterVectorizeCnt * | |||
6832 | TTI.getGatherScatterOpCost( | |||
6833 | Instruction::Load, LoadTy, LI->getPointerOperand(), | |||
6834 | /*VariableMask=*/false, Alignment, CostKind, LI); | |||
6835 | if (NeedInsertSubvectorAnalysis) { | |||
6836 | // Add the cost for the subvectors insert. | |||
6837 | for (int I = VF, E = VL.size(); I < E; I += VF) | |||
6838 | GatherCost += TTI.getShuffleCost(TTI::SK_InsertSubvector, VecTy, | |||
6839 | std::nullopt, CostKind, I, LoadTy); | |||
6840 | } | |||
6841 | GatherCost -= ScalarsCost; | |||
6842 | } | |||
6843 | } else if (!Root && isSplat(VL)) { | |||
6844 | // Found the broadcasting of the single scalar, calculate the cost as | |||
6845 | // the broadcast. | |||
6846 | const auto *It = | |||
6847 | find_if(VL, [](Value *V) { return !isa<UndefValue>(V); }); | |||
6848 | assert(It != VL.end() && "Expected at least one non-undef value.")(static_cast <bool> (It != VL.end() && "Expected at least one non-undef value." ) ? void (0) : __assert_fail ("It != VL.end() && \"Expected at least one non-undef value.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 6848, __extension__ __PRETTY_FUNCTION__)); | |||
6849 | // Add broadcast for non-identity shuffle only. | |||
6850 | bool NeedShuffle = | |||
6851 | count(VL, *It) > 1 && | |||
6852 | (VL.front() != *It || !all_of(VL.drop_front(), UndefValue::classof)); | |||
6853 | InstructionCost InsertCost = TTI.getVectorInstrCost( | |||
6854 | Instruction::InsertElement, VecTy, CostKind, | |||
6855 | NeedShuffle ? 0 : std::distance(VL.begin(), It), | |||
6856 | PoisonValue::get(VecTy), *It); | |||
6857 | return InsertCost + | |||
6858 | (NeedShuffle ? TTI.getShuffleCost( | |||
6859 | TargetTransformInfo::SK_Broadcast, VecTy, | |||
6860 | /*Mask=*/std::nullopt, CostKind, /*Index=*/0, | |||
6861 | /*SubTp=*/nullptr, /*Args=*/*It) | |||
6862 | : TTI::TCC_Free); | |||
6863 | } | |||
6864 | return GatherCost + | |||
6865 | (all_of(Gathers, UndefValue::classof) | |||
6866 | ? TTI::TCC_Free | |||
6867 | : R.getGatherCost(Gathers, !Root && VL.equals(Gathers))); | |||
6868 | }; | |||
6869 | ||||
6870 | /// Compute the cost of creating a vector of type \p VecTy containing the | |||
6871 | /// extracted values from \p VL. | |||
6872 | InstructionCost computeExtractCost(ArrayRef<Value *> VL, ArrayRef<int> Mask, | |||
6873 | TTI::ShuffleKind ShuffleKind) { | |||
6874 | auto *VecTy = FixedVectorType::get(VL.front()->getType(), VL.size()); | |||
6875 | unsigned NumOfParts = TTI.getNumberOfParts(VecTy); | |||
6876 | ||||
6877 | if (ShuffleKind != TargetTransformInfo::SK_PermuteSingleSrc || | |||
6878 | !NumOfParts || VecTy->getNumElements() < NumOfParts) | |||
6879 | return TTI.getShuffleCost(ShuffleKind, VecTy, Mask); | |||
6880 | ||||
6881 | bool AllConsecutive = true; | |||
6882 | unsigned EltsPerVector = VecTy->getNumElements() / NumOfParts; | |||
6883 | unsigned Idx = -1; | |||
6884 | InstructionCost Cost = 0; | |||
6885 | ||||
6886 | // Process extracts in blocks of EltsPerVector to check if the source vector | |||
6887 | // operand can be re-used directly. If not, add the cost of creating a | |||
6888 | // shuffle to extract the values into a vector register. | |||
6889 | SmallVector<int> RegMask(EltsPerVector, PoisonMaskElem); | |||
6890 | for (auto *V : VL) { | |||
6891 | ++Idx; | |||
6892 | ||||
6893 | // Reached the start of a new vector registers. | |||
6894 | if (Idx % EltsPerVector == 0) { | |||
6895 | RegMask.assign(EltsPerVector, PoisonMaskElem); | |||
6896 | AllConsecutive = true; | |||
6897 | continue; | |||
6898 | } | |||
6899 | ||||
6900 | // Need to exclude undefs from analysis. | |||
6901 | if (isa<UndefValue>(V) || Mask[Idx] == PoisonMaskElem) | |||
6902 | continue; | |||
6903 | ||||
6904 | // Check all extracts for a vector register on the target directly | |||
6905 | // extract values in order. | |||
6906 | unsigned CurrentIdx = *getExtractIndex(cast<Instruction>(V)); | |||
6907 | if (!isa<UndefValue>(VL[Idx - 1]) && Mask[Idx - 1] != PoisonMaskElem) { | |||
6908 | unsigned PrevIdx = *getExtractIndex(cast<Instruction>(VL[Idx - 1])); | |||
6909 | AllConsecutive &= PrevIdx + 1 == CurrentIdx && | |||
6910 | CurrentIdx % EltsPerVector == Idx % EltsPerVector; | |||
6911 | RegMask[Idx % EltsPerVector] = CurrentIdx % EltsPerVector; | |||
6912 | } | |||
6913 | ||||
6914 | if (AllConsecutive) | |||
6915 | continue; | |||
6916 | ||||
6917 | // Skip all indices, except for the last index per vector block. | |||
6918 | if ((Idx + 1) % EltsPerVector != 0 && Idx + 1 != VL.size()) | |||
6919 | continue; | |||
6920 | ||||
6921 | // If we have a series of extracts which are not consecutive and hence | |||
6922 | // cannot re-use the source vector register directly, compute the shuffle | |||
6923 | // cost to extract the vector with EltsPerVector elements. | |||
6924 | Cost += TTI.getShuffleCost( | |||
6925 | TargetTransformInfo::SK_PermuteSingleSrc, | |||
6926 | FixedVectorType::get(VecTy->getElementType(), EltsPerVector), | |||
6927 | RegMask); | |||
6928 | } | |||
6929 | return Cost; | |||
6930 | } | |||
6931 | ||||
6932 | class ShuffleCostBuilder { | |||
6933 | const TargetTransformInfo &TTI; | |||
6934 | ||||
6935 | static bool isEmptyOrIdentity(ArrayRef<int> Mask, unsigned VF) { | |||
6936 | int Limit = 2 * VF; | |||
6937 | return Mask.empty() || | |||
6938 | (VF == Mask.size() && | |||
6939 | all_of(Mask, [Limit](int Idx) { return Idx < Limit; }) && | |||
6940 | ShuffleVectorInst::isIdentityMask(Mask)); | |||
6941 | } | |||
6942 | ||||
6943 | public: | |||
6944 | ShuffleCostBuilder(const TargetTransformInfo &TTI) : TTI(TTI) {} | |||
6945 | ~ShuffleCostBuilder() = default; | |||
6946 | InstructionCost createShuffleVector(Value *V1, Value *, | |||
6947 | ArrayRef<int> Mask) const { | |||
6948 | // Empty mask or identity mask are free. | |||
6949 | unsigned VF = | |||
6950 | cast<VectorType>(V1->getType())->getElementCount().getKnownMinValue(); | |||
6951 | if (isEmptyOrIdentity(Mask, VF)) | |||
6952 | return TTI::TCC_Free; | |||
6953 | return TTI.getShuffleCost( | |||
6954 | TTI::SK_PermuteTwoSrc, | |||
6955 | FixedVectorType::get( | |||
6956 | cast<VectorType>(V1->getType())->getElementType(), Mask.size()), | |||
6957 | Mask); | |||
6958 | } | |||
6959 | InstructionCost createShuffleVector(Value *V1, ArrayRef<int> Mask) const { | |||
6960 | // Empty mask or identity mask are free. | |||
6961 | if (isEmptyOrIdentity(Mask, Mask.size())) | |||
6962 | return TTI::TCC_Free; | |||
6963 | return TTI.getShuffleCost( | |||
6964 | TTI::SK_PermuteSingleSrc, | |||
6965 | FixedVectorType::get( | |||
6966 | cast<VectorType>(V1->getType())->getElementType(), Mask.size()), | |||
6967 | Mask); | |||
6968 | } | |||
6969 | InstructionCost createIdentity(Value *) const { return TTI::TCC_Free; } | |||
6970 | InstructionCost createPoison(Type *Ty, unsigned VF) const { | |||
6971 | return TTI::TCC_Free; | |||
6972 | } | |||
6973 | void resizeToMatch(Value *&, Value *&) const {} | |||
6974 | }; | |||
6975 | ||||
6976 | /// Smart shuffle instruction emission, walks through shuffles trees and | |||
6977 | /// tries to find the best matching vector for the actual shuffle | |||
6978 | /// instruction. | |||
6979 | InstructionCost | |||
6980 | createShuffle(const PointerUnion<Value *, const TreeEntry *> &P1, | |||
6981 | const PointerUnion<Value *, const TreeEntry *> &P2, | |||
6982 | ArrayRef<int> Mask) { | |||
6983 | ShuffleCostBuilder Builder(TTI); | |||
6984 | Value *V1 = P1.dyn_cast<Value *>(), *V2 = P2.dyn_cast<Value *>(); | |||
6985 | unsigned CommonVF = 0; | |||
6986 | if (!V1) { | |||
6987 | const TreeEntry *E = P1.get<const TreeEntry *>(); | |||
6988 | unsigned VF = E->getVectorFactor(); | |||
6989 | if (V2) { | |||
6990 | unsigned V2VF = cast<FixedVectorType>(V2->getType())->getNumElements(); | |||
6991 | if (V2VF != VF && V2VF == E->Scalars.size()) | |||
6992 | VF = E->Scalars.size(); | |||
6993 | } else if (!P2.isNull()) { | |||
6994 | const TreeEntry *E2 = P2.get<const TreeEntry *>(); | |||
6995 | if (E->Scalars.size() == E2->Scalars.size()) | |||
6996 | CommonVF = VF = E->Scalars.size(); | |||
6997 | } | |||
6998 | V1 = Constant::getNullValue( | |||
6999 | FixedVectorType::get(E->Scalars.front()->getType(), VF)); | |||
7000 | } | |||
7001 | if (!V2 && !P2.isNull()) { | |||
7002 | const TreeEntry *E = P2.get<const TreeEntry *>(); | |||
7003 | unsigned VF = E->getVectorFactor(); | |||
7004 | unsigned V1VF = cast<FixedVectorType>(V1->getType())->getNumElements(); | |||
7005 | if (!CommonVF && V1VF == E->Scalars.size()) | |||
7006 | CommonVF = E->Scalars.size(); | |||
7007 | if (CommonVF) | |||
7008 | VF = CommonVF; | |||
7009 | V2 = Constant::getNullValue( | |||
7010 | FixedVectorType::get(E->Scalars.front()->getType(), VF)); | |||
7011 | } | |||
7012 | return BaseShuffleAnalysis::createShuffle<InstructionCost>(V1, V2, Mask, | |||
7013 | Builder); | |||
7014 | } | |||
7015 | ||||
7016 | public: | |||
7017 | ShuffleCostEstimator(TargetTransformInfo &TTI, | |||
7018 | ArrayRef<Value *> VectorizedVals, BoUpSLP &R, | |||
7019 | SmallPtrSetImpl<Value *> &CheckedExtracts) | |||
7020 | : TTI(TTI), VectorizedVals(VectorizedVals), R(R), | |||
7021 | CheckedExtracts(CheckedExtracts) {} | |||
7022 | Value *adjustExtracts(const TreeEntry *E, ArrayRef<int> Mask, | |||
7023 | TTI::ShuffleKind ShuffleKind) { | |||
7024 | if (Mask.empty()) | |||
7025 | return nullptr; | |||
7026 | Value *VecBase = nullptr; | |||
7027 | ArrayRef<Value *> VL = E->Scalars; | |||
7028 | auto *VecTy = FixedVectorType::get(VL.front()->getType(), VL.size()); | |||
7029 | // If the resulting type is scalarized, do not adjust the cost. | |||
7030 | unsigned VecNumParts = TTI.getNumberOfParts(VecTy); | |||
7031 | if (VecNumParts == VecTy->getNumElements()) { | |||
7032 | InVectors.assign(1, E); | |||
7033 | return nullptr; | |||
7034 | } | |||
7035 | DenseMap<Value *, int> ExtractVectorsTys; | |||
7036 | for (auto [I, V] : enumerate(VL)) { | |||
7037 | // Ignore non-extractelement scalars. | |||
7038 | if (isa<UndefValue>(V) || (!Mask.empty() && Mask[I] == PoisonMaskElem)) | |||
7039 | continue; | |||
7040 | // If all users of instruction are going to be vectorized and this | |||
7041 | // instruction itself is not going to be vectorized, consider this | |||
7042 | // instruction as dead and remove its cost from the final cost of the | |||
7043 | // vectorized tree. | |||
7044 | // Also, avoid adjusting the cost for extractelements with multiple uses | |||
7045 | // in different graph entries. | |||
7046 | const TreeEntry *VE = R.getTreeEntry(V); | |||
7047 | if (!CheckedExtracts.insert(V).second || | |||
7048 | !R.areAllUsersVectorized(cast<Instruction>(V), VectorizedVals) || | |||
7049 | (VE && VE != E)) | |||
7050 | continue; | |||
7051 | auto *EE = cast<ExtractElementInst>(V); | |||
7052 | VecBase = EE->getVectorOperand(); | |||
7053 | std::optional<unsigned> EEIdx = getExtractIndex(EE); | |||
7054 | if (!EEIdx) | |||
7055 | continue; | |||
7056 | unsigned Idx = *EEIdx; | |||
7057 | if (VecNumParts != TTI.getNumberOfParts(EE->getVectorOperandType())) { | |||
7058 | auto It = | |||
7059 | ExtractVectorsTys.try_emplace(EE->getVectorOperand(), Idx).first; | |||
7060 | It->getSecond() = std::min<int>(It->second, Idx); | |||
7061 | } | |||
7062 | // Take credit for instruction that will become dead. | |||
7063 | if (EE->hasOneUse()) { | |||
7064 | Instruction *Ext = EE->user_back(); | |||
7065 | if (isa<SExtInst, ZExtInst>(Ext) && all_of(Ext->users(), [](User *U) { | |||
7066 | return isa<GetElementPtrInst>(U); | |||
7067 | })) { | |||
7068 | // Use getExtractWithExtendCost() to calculate the cost of | |||
7069 | // extractelement/ext pair. | |||
7070 | Cost -= TTI.getExtractWithExtendCost(Ext->getOpcode(), Ext->getType(), | |||
7071 | EE->getVectorOperandType(), Idx); | |||
7072 | // Add back the cost of s|zext which is subtracted separately. | |||
7073 | Cost += TTI.getCastInstrCost( | |||
7074 | Ext->getOpcode(), Ext->getType(), EE->getType(), | |||
7075 | TTI::getCastContextHint(Ext), CostKind, Ext); | |||
7076 | continue; | |||
7077 | } | |||
7078 | } | |||
7079 | Cost -= TTI.getVectorInstrCost(*EE, EE->getVectorOperandType(), CostKind, | |||
7080 | Idx); | |||
7081 | } | |||
7082 | // Add a cost for subvector extracts/inserts if required. | |||
7083 | for (const auto &Data : ExtractVectorsTys) { | |||
7084 | auto *EEVTy = cast<FixedVectorType>(Data.first->getType()); | |||
7085 | unsigned NumElts = VecTy->getNumElements(); | |||
7086 | if (Data.second % NumElts == 0) | |||
7087 | continue; | |||
7088 | if (TTI.getNumberOfParts(EEVTy) > VecNumParts) { | |||
7089 | unsigned Idx = (Data.second / NumElts) * NumElts; | |||
7090 | unsigned EENumElts = EEVTy->getNumElements(); | |||
7091 | if (Idx % NumElts == 0) | |||
7092 | continue; | |||
7093 | if (Idx + NumElts <= EENumElts) { | |||
7094 | Cost += TTI.getShuffleCost(TargetTransformInfo::SK_ExtractSubvector, | |||
7095 | EEVTy, std::nullopt, CostKind, Idx, VecTy); | |||
7096 | } else { | |||
7097 | // Need to round up the subvector type vectorization factor to avoid a | |||
7098 | // crash in cost model functions. Make SubVT so that Idx + VF of SubVT | |||
7099 | // <= EENumElts. | |||
7100 | auto *SubVT = | |||
7101 | FixedVectorType::get(VecTy->getElementType(), EENumElts - Idx); | |||
7102 | Cost += TTI.getShuffleCost(TargetTransformInfo::SK_ExtractSubvector, | |||
7103 | EEVTy, std::nullopt, CostKind, Idx, SubVT); | |||
7104 | } | |||
7105 | } else { | |||
7106 | Cost += TTI.getShuffleCost(TargetTransformInfo::SK_InsertSubvector, | |||
7107 | VecTy, std::nullopt, CostKind, 0, EEVTy); | |||
7108 | } | |||
7109 | } | |||
7110 | // Check that gather of extractelements can be represented as just a | |||
7111 | // shuffle of a single/two vectors the scalars are extracted from. | |||
7112 | // Found the bunch of extractelement instructions that must be gathered | |||
7113 | // into a vector and can be represented as a permutation elements in a | |||
7114 | // single input vector or of 2 input vectors. | |||
7115 | Cost += computeExtractCost(VL, Mask, ShuffleKind); | |||
7116 | InVectors.assign(1, E); | |||
7117 | return VecBase; | |||
7118 | } | |||
7119 | void add(const TreeEntry *E1, const TreeEntry *E2, ArrayRef<int> Mask) { | |||
7120 | CommonMask.assign(Mask.begin(), Mask.end()); | |||
7121 | InVectors.assign({E1, E2}); | |||
7122 | } | |||
7123 | void add(const TreeEntry *E1, ArrayRef<int> Mask) { | |||
7124 | CommonMask.assign(Mask.begin(), Mask.end()); | |||
7125 | InVectors.assign(1, E1); | |||
7126 | } | |||
7127 | void gather(ArrayRef<Value *> VL, Value *Root = nullptr) { | |||
7128 | Cost += getBuildVectorCost(VL, Root); | |||
7129 | if (!Root) { | |||
7130 | assert(InVectors.empty() && "Unexpected input vectors for buildvector.")(static_cast <bool> (InVectors.empty() && "Unexpected input vectors for buildvector." ) ? void (0) : __assert_fail ("InVectors.empty() && \"Unexpected input vectors for buildvector.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7130, __extension__ __PRETTY_FUNCTION__)); | |||
7131 | // FIXME: Need to find a way to avoid use of getNullValue here. | |||
7132 | InVectors.assign(1, Constant::getNullValue(FixedVectorType::get( | |||
7133 | VL.front()->getType(), VL.size()))); | |||
7134 | } | |||
7135 | } | |||
7136 | /// Finalize emission of the shuffles. | |||
7137 | InstructionCost finalize(ArrayRef<int> ExtMask) { | |||
7138 | IsFinalized = true; | |||
7139 | ::addMask(CommonMask, ExtMask, /*ExtendingManyInputs=*/true); | |||
7140 | if (CommonMask.empty()) | |||
7141 | return Cost; | |||
7142 | int Limit = CommonMask.size() * 2; | |||
7143 | if (all_of(CommonMask, [=](int Idx) { return Idx < Limit; }) && | |||
7144 | ShuffleVectorInst::isIdentityMask(CommonMask)) | |||
7145 | return Cost; | |||
7146 | return Cost + | |||
7147 | createShuffle(InVectors.front(), | |||
7148 | InVectors.size() == 2 ? InVectors.back() : nullptr, | |||
7149 | CommonMask); | |||
7150 | } | |||
7151 | ||||
7152 | ~ShuffleCostEstimator() { | |||
7153 | assert((IsFinalized || CommonMask.empty()) &&(static_cast <bool> ((IsFinalized || CommonMask.empty() ) && "Shuffle construction must be finalized.") ? void (0) : __assert_fail ("(IsFinalized || CommonMask.empty()) && \"Shuffle construction must be finalized.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7154, __extension__ __PRETTY_FUNCTION__)) | |||
7154 | "Shuffle construction must be finalized.")(static_cast <bool> ((IsFinalized || CommonMask.empty() ) && "Shuffle construction must be finalized.") ? void (0) : __assert_fail ("(IsFinalized || CommonMask.empty()) && \"Shuffle construction must be finalized.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7154, __extension__ __PRETTY_FUNCTION__)); | |||
7155 | } | |||
7156 | }; | |||
7157 | ||||
7158 | InstructionCost | |||
7159 | BoUpSLP::getEntryCost(const TreeEntry *E, ArrayRef<Value *> VectorizedVals, | |||
7160 | SmallPtrSetImpl<Value *> &CheckedExtracts) { | |||
7161 | ArrayRef<Value *> VL = E->Scalars; | |||
7162 | ||||
7163 | Type *ScalarTy = VL[0]->getType(); | |||
7164 | if (auto *SI = dyn_cast<StoreInst>(VL[0])) | |||
7165 | ScalarTy = SI->getValueOperand()->getType(); | |||
7166 | else if (auto *CI = dyn_cast<CmpInst>(VL[0])) | |||
7167 | ScalarTy = CI->getOperand(0)->getType(); | |||
7168 | else if (auto *IE = dyn_cast<InsertElementInst>(VL[0])) | |||
7169 | ScalarTy = IE->getOperand(1)->getType(); | |||
7170 | auto *VecTy = FixedVectorType::get(ScalarTy, VL.size()); | |||
7171 | TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput; | |||
7172 | ||||
7173 | // If we have computed a smaller type for the expression, update VecTy so | |||
7174 | // that the costs will be accurate. | |||
7175 | if (MinBWs.count(VL[0])) | |||
7176 | VecTy = FixedVectorType::get( | |||
7177 | IntegerType::get(F->getContext(), MinBWs[VL[0]].first), VL.size()); | |||
7178 | unsigned EntryVF = E->getVectorFactor(); | |||
7179 | auto *FinalVecTy = FixedVectorType::get(VecTy->getElementType(), EntryVF); | |||
7180 | ||||
7181 | bool NeedToShuffleReuses = !E->ReuseShuffleIndices.empty(); | |||
7182 | if (E->State == TreeEntry::NeedToGather) { | |||
7183 | if (allConstant(VL)) | |||
7184 | return 0; | |||
7185 | if (isa<InsertElementInst>(VL[0])) | |||
7186 | return InstructionCost::getInvalid(); | |||
7187 | ShuffleCostEstimator Estimator(*TTI, VectorizedVals, *this, | |||
7188 | CheckedExtracts); | |||
7189 | unsigned VF = E->getVectorFactor(); | |||
7190 | SmallVector<int> ReuseShuffleIndicies(E->ReuseShuffleIndices.begin(), | |||
7191 | E->ReuseShuffleIndices.end()); | |||
7192 | SmallVector<Value *> GatheredScalars(E->Scalars.begin(), E->Scalars.end()); | |||
7193 | // Build a mask out of the reorder indices and reorder scalars per this | |||
7194 | // mask. | |||
7195 | SmallVector<int> ReorderMask; | |||
7196 | inversePermutation(E->ReorderIndices, ReorderMask); | |||
7197 | if (!ReorderMask.empty()) | |||
7198 | reorderScalars(GatheredScalars, ReorderMask); | |||
7199 | SmallVector<int> Mask; | |||
7200 | SmallVector<int> ExtractMask; | |||
7201 | std::optional<TargetTransformInfo::ShuffleKind> ExtractShuffle; | |||
7202 | std::optional<TargetTransformInfo::ShuffleKind> GatherShuffle; | |||
7203 | SmallVector<const TreeEntry *> Entries; | |||
7204 | Type *ScalarTy = GatheredScalars.front()->getType(); | |||
7205 | // Check for gathered extracts. | |||
7206 | ExtractShuffle = tryToGatherExtractElements(GatheredScalars, ExtractMask); | |||
7207 | SmallVector<Value *> IgnoredVals; | |||
7208 | if (UserIgnoreList) | |||
7209 | IgnoredVals.assign(UserIgnoreList->begin(), UserIgnoreList->end()); | |||
7210 | ||||
7211 | bool Resized = false; | |||
7212 | if (Value *VecBase = Estimator.adjustExtracts( | |||
7213 | E, ExtractMask, ExtractShuffle.value_or(TTI::SK_PermuteTwoSrc))) | |||
7214 | if (auto *VecBaseTy = dyn_cast<FixedVectorType>(VecBase->getType())) | |||
7215 | if (VF == VecBaseTy->getNumElements() && GatheredScalars.size() != VF) { | |||
7216 | Resized = true; | |||
7217 | GatheredScalars.append(VF - GatheredScalars.size(), | |||
7218 | PoisonValue::get(ScalarTy)); | |||
7219 | } | |||
7220 | ||||
7221 | // Do not try to look for reshuffled loads for gathered loads (they will be | |||
7222 | // handled later), for vectorized scalars, and cases, which are definitely | |||
7223 | // not profitable (splats and small gather nodes.) | |||
7224 | if (ExtractShuffle || E->getOpcode() != Instruction::Load || | |||
7225 | E->isAltShuffle() || | |||
7226 | all_of(E->Scalars, [this](Value *V) { return getTreeEntry(V); }) || | |||
7227 | isSplat(E->Scalars) || | |||
7228 | (E->Scalars != GatheredScalars && GatheredScalars.size() <= 2)) | |||
7229 | GatherShuffle = isGatherShuffledEntry(E, GatheredScalars, Mask, Entries); | |||
7230 | if (GatherShuffle) { | |||
7231 | 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", 7232, __extension__ __PRETTY_FUNCTION__)) | |||
7232 | "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", 7232, __extension__ __PRETTY_FUNCTION__)); | |||
7233 | if (*GatherShuffle == TTI::SK_PermuteSingleSrc && | |||
7234 | Entries.front()->isSame(E->Scalars)) { | |||
7235 | // Perfect match in the graph, will reuse the previously vectorized | |||
7236 | // node. Cost is 0. | |||
7237 | 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) | |||
7238 | dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: perfect diamond match for gather bundle that starts with " << *VL.front() << ".\n"; } } while (false) | |||
7239 | << "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) | |||
7240 | << *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); | |||
7241 | return 0; | |||
7242 | } | |||
7243 | if (!Resized) { | |||
7244 | unsigned VF1 = Entries.front()->getVectorFactor(); | |||
7245 | unsigned VF2 = Entries.back()->getVectorFactor(); | |||
7246 | if ((VF == VF1 || VF == VF2) && GatheredScalars.size() != VF) | |||
7247 | GatheredScalars.append(VF - GatheredScalars.size(), | |||
7248 | PoisonValue::get(ScalarTy)); | |||
7249 | } | |||
7250 | // Remove shuffled elements from list of gathers. | |||
7251 | for (int I = 0, Sz = Mask.size(); I < Sz; ++I) { | |||
7252 | if (Mask[I] != PoisonMaskElem) | |||
7253 | GatheredScalars[I] = PoisonValue::get(ScalarTy); | |||
7254 | } | |||
7255 | 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) | |||
7256 | << " 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) | |||
7257 | << *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); | |||
7258 | if (Entries.size() == 1) | |||
7259 | Estimator.add(Entries.front(), Mask); | |||
7260 | else | |||
7261 | Estimator.add(Entries.front(), Entries.back(), Mask); | |||
7262 | Estimator.gather( | |||
7263 | GatheredScalars, | |||
7264 | Constant::getNullValue(FixedVectorType::get( | |||
7265 | GatheredScalars.front()->getType(), GatheredScalars.size()))); | |||
7266 | return Estimator.finalize(E->ReuseShuffleIndices); | |||
7267 | } | |||
7268 | Estimator.gather( | |||
7269 | GatheredScalars, | |||
7270 | VL.equals(GatheredScalars) | |||
7271 | ? nullptr | |||
7272 | : Constant::getNullValue(FixedVectorType::get( | |||
7273 | GatheredScalars.front()->getType(), GatheredScalars.size()))); | |||
7274 | return Estimator.finalize(E->ReuseShuffleIndices); | |||
7275 | } | |||
7276 | InstructionCost CommonCost = 0; | |||
7277 | SmallVector<int> Mask; | |||
7278 | if (!E->ReorderIndices.empty()) { | |||
7279 | SmallVector<int> NewMask; | |||
7280 | if (E->getOpcode() == Instruction::Store) { | |||
7281 | // For stores the order is actually a mask. | |||
7282 | NewMask.resize(E->ReorderIndices.size()); | |||
7283 | copy(E->ReorderIndices, NewMask.begin()); | |||
7284 | } else { | |||
7285 | inversePermutation(E->ReorderIndices, NewMask); | |||
7286 | } | |||
7287 | ::addMask(Mask, NewMask); | |||
7288 | } | |||
7289 | if (NeedToShuffleReuses) | |||
7290 | ::addMask(Mask, E->ReuseShuffleIndices); | |||
7291 | if (!Mask.empty() && !ShuffleVectorInst::isIdentityMask(Mask)) | |||
7292 | CommonCost = | |||
7293 | TTI->getShuffleCost(TTI::SK_PermuteSingleSrc, FinalVecTy, Mask); | |||
7294 | 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", 7296, __extension__ __PRETTY_FUNCTION__)) | |||
7295 | 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", 7296, __extension__ __PRETTY_FUNCTION__)) | |||
7296 | "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", 7296, __extension__ __PRETTY_FUNCTION__)); | |||
7297 | assert(E->getOpcode() &&(static_cast <bool> (E->getOpcode() && ((allSameType (VL) && allSameBlock(VL)) || (E->getOpcode() == Instruction ::GetElementPtr && E->getMainOp()->getType()-> isPointerTy())) && "Invalid VL") ? void (0) : __assert_fail ("E->getOpcode() && ((allSameType(VL) && allSameBlock(VL)) || (E->getOpcode() == Instruction::GetElementPtr && E->getMainOp()->getType()->isPointerTy())) && \"Invalid VL\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7301, __extension__ __PRETTY_FUNCTION__)) | |||
7298 | ((allSameType(VL) && allSameBlock(VL)) ||(static_cast <bool> (E->getOpcode() && ((allSameType (VL) && allSameBlock(VL)) || (E->getOpcode() == Instruction ::GetElementPtr && E->getMainOp()->getType()-> isPointerTy())) && "Invalid VL") ? void (0) : __assert_fail ("E->getOpcode() && ((allSameType(VL) && allSameBlock(VL)) || (E->getOpcode() == Instruction::GetElementPtr && E->getMainOp()->getType()->isPointerTy())) && \"Invalid VL\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7301, __extension__ __PRETTY_FUNCTION__)) | |||
7299 | (E->getOpcode() == Instruction::GetElementPtr &&(static_cast <bool> (E->getOpcode() && ((allSameType (VL) && allSameBlock(VL)) || (E->getOpcode() == Instruction ::GetElementPtr && E->getMainOp()->getType()-> isPointerTy())) && "Invalid VL") ? void (0) : __assert_fail ("E->getOpcode() && ((allSameType(VL) && allSameBlock(VL)) || (E->getOpcode() == Instruction::GetElementPtr && E->getMainOp()->getType()->isPointerTy())) && \"Invalid VL\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7301, __extension__ __PRETTY_FUNCTION__)) | |||
7300 | E->getMainOp()->getType()->isPointerTy())) &&(static_cast <bool> (E->getOpcode() && ((allSameType (VL) && allSameBlock(VL)) || (E->getOpcode() == Instruction ::GetElementPtr && E->getMainOp()->getType()-> isPointerTy())) && "Invalid VL") ? void (0) : __assert_fail ("E->getOpcode() && ((allSameType(VL) && allSameBlock(VL)) || (E->getOpcode() == Instruction::GetElementPtr && E->getMainOp()->getType()->isPointerTy())) && \"Invalid VL\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7301, __extension__ __PRETTY_FUNCTION__)) | |||
7301 | "Invalid VL")(static_cast <bool> (E->getOpcode() && ((allSameType (VL) && allSameBlock(VL)) || (E->getOpcode() == Instruction ::GetElementPtr && E->getMainOp()->getType()-> isPointerTy())) && "Invalid VL") ? void (0) : __assert_fail ("E->getOpcode() && ((allSameType(VL) && allSameBlock(VL)) || (E->getOpcode() == Instruction::GetElementPtr && E->getMainOp()->getType()->isPointerTy())) && \"Invalid VL\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7301, __extension__ __PRETTY_FUNCTION__)); | |||
7302 | Instruction *VL0 = E->getMainOp(); | |||
7303 | unsigned ShuffleOrOp = | |||
7304 | E->isAltShuffle() ? (unsigned)Instruction::ShuffleVector : E->getOpcode(); | |||
7305 | const unsigned Sz = VL.size(); | |||
7306 | auto GetCostDiff = | |||
7307 | [=](function_ref<InstructionCost(unsigned)> ScalarEltCost, | |||
7308 | function_ref<InstructionCost(InstructionCost)> VectorCost) { | |||
7309 | // Calculate the cost of this instruction. | |||
7310 | InstructionCost ScalarCost = 0; | |||
7311 | if (isa<CastInst, CmpInst, SelectInst, CallInst>(VL0)) { | |||
7312 | // For some of the instructions no need to calculate cost for each | |||
7313 | // particular instruction, we can use the cost of the single | |||
7314 | // instruction x total number of scalar instructions. | |||
7315 | ScalarCost = Sz * ScalarEltCost(0); | |||
7316 | } else { | |||
7317 | for (unsigned I = 0; I < Sz; ++I) | |||
7318 | ScalarCost += ScalarEltCost(I); | |||
7319 | } | |||
7320 | ||||
7321 | InstructionCost VecCost = VectorCost(CommonCost); | |||
7322 | LLVM_DEBUG(dumpTreeCosts(E, CommonCost, VecCost - CommonCost,do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dumpTreeCosts(E, CommonCost, VecCost - CommonCost, ScalarCost, "Calculated costs for Tree"); } } while (false) | |||
7323 | ScalarCost, "Calculated costs for Tree"))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dumpTreeCosts(E, CommonCost, VecCost - CommonCost, ScalarCost, "Calculated costs for Tree"); } } while (false); | |||
7324 | return VecCost - ScalarCost; | |||
7325 | }; | |||
7326 | // Calculate cost difference from vectorizing set of GEPs. | |||
7327 | // Negative value means vectorizing is profitable. | |||
7328 | auto GetGEPCostDiff = [=](ArrayRef<Value *> Ptrs, Value *BasePtr) { | |||
7329 | InstructionCost ScalarCost = 0; | |||
7330 | InstructionCost VecCost = 0; | |||
7331 | // Here we differentiate two cases: (1) when Ptrs represent a regular | |||
7332 | // vectorization tree node (as they are pointer arguments of scattered | |||
7333 | // loads) or (2) when Ptrs are the arguments of loads or stores being | |||
7334 | // vectorized as plane wide unit-stride load/store since all the | |||
7335 | // loads/stores are known to be from/to adjacent locations. | |||
7336 | assert(E->State == TreeEntry::Vectorize &&(static_cast <bool> (E->State == TreeEntry::Vectorize && "Entry state expected to be Vectorize here.") ? void (0) : __assert_fail ("E->State == TreeEntry::Vectorize && \"Entry state expected to be Vectorize here.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7337, __extension__ __PRETTY_FUNCTION__)) | |||
7337 | "Entry state expected to be Vectorize here.")(static_cast <bool> (E->State == TreeEntry::Vectorize && "Entry state expected to be Vectorize here.") ? void (0) : __assert_fail ("E->State == TreeEntry::Vectorize && \"Entry state expected to be Vectorize here.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7337, __extension__ __PRETTY_FUNCTION__)); | |||
7338 | if (isa<LoadInst, StoreInst>(VL0)) { | |||
7339 | // Case 2: estimate costs for pointer related costs when vectorizing to | |||
7340 | // a wide load/store. | |||
7341 | // Scalar cost is estimated as a set of pointers with known relationship | |||
7342 | // between them. | |||
7343 | // For vector code we will use BasePtr as argument for the wide load/store | |||
7344 | // but we also need to account all the instructions which are going to | |||
7345 | // stay in vectorized code due to uses outside of these scalar | |||
7346 | // loads/stores. | |||
7347 | ScalarCost = TTI->getPointersChainCost( | |||
7348 | Ptrs, BasePtr, TTI::PointersChainInfo::getKnownUniformStrided(), | |||
7349 | CostKind); | |||
7350 | ||||
7351 | SmallVector<const Value *> PtrsRetainedInVecCode; | |||
7352 | for (Value *V : Ptrs) { | |||
7353 | if (V == BasePtr) { | |||
7354 | PtrsRetainedInVecCode.push_back(V); | |||
7355 | continue; | |||
7356 | } | |||
7357 | auto *Ptr = dyn_cast<GetElementPtrInst>(V); | |||
7358 | // For simplicity assume Ptr to stay in vectorized code if it's not a | |||
7359 | // GEP instruction. We don't care since it's cost considered free. | |||
7360 | // TODO: We should check for any uses outside of vectorizable tree | |||
7361 | // rather than just single use. | |||
7362 | if (!Ptr || !Ptr->hasOneUse()) | |||
7363 | PtrsRetainedInVecCode.push_back(V); | |||
7364 | } | |||
7365 | ||||
7366 | if (PtrsRetainedInVecCode.size() == Ptrs.size()) { | |||
7367 | // If all pointers stay in vectorized code then we don't have | |||
7368 | // any savings on that. | |||
7369 | LLVM_DEBUG(dumpTreeCosts(E, 0, ScalarCost, ScalarCost,do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dumpTreeCosts(E, 0, ScalarCost, ScalarCost, "Calculated GEPs cost for Tree" ); } } while (false) | |||
7370 | "Calculated GEPs cost for Tree"))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dumpTreeCosts(E, 0, ScalarCost, ScalarCost, "Calculated GEPs cost for Tree" ); } } while (false); | |||
7371 | return InstructionCost{TTI::TCC_Free}; | |||
7372 | } | |||
7373 | VecCost = TTI->getPointersChainCost( | |||
7374 | PtrsRetainedInVecCode, BasePtr, | |||
7375 | TTI::PointersChainInfo::getKnownNonUniformStrided(), CostKind); | |||
7376 | } else { | |||
7377 | // Case 1: Ptrs are the arguments of loads that we are going to transform | |||
7378 | // into masked gather load intrinsic. | |||
7379 | // All the scalar GEPs will be removed as a result of vectorization. | |||
7380 | // For any external uses of some lanes extract element instructions will | |||
7381 | // be generated (which cost is estimated separately). | |||
7382 | TTI::PointersChainInfo PtrsInfo = | |||
7383 | all_of(Ptrs, | |||
7384 | [](const Value *V) { | |||
7385 | auto *Ptr = dyn_cast<GetElementPtrInst>(V); | |||
7386 | return Ptr && !Ptr->hasAllConstantIndices(); | |||
7387 | }) | |||
7388 | ? TTI::PointersChainInfo::getNonUniformStrided() | |||
7389 | : TTI::PointersChainInfo::getKnownNonUniformStrided(); | |||
7390 | ||||
7391 | ScalarCost = TTI->getPointersChainCost(Ptrs, BasePtr, PtrsInfo, CostKind); | |||
7392 | ||||
7393 | // Remark: it not quite correct to use scalar GEP cost for a vector GEP, | |||
7394 | // but it's not clear how to do that without having vector GEP arguments | |||
7395 | // ready. | |||
7396 | // Perhaps using just TTI::TCC_Free/TTI::TCC_Basic would be better option. | |||
7397 | if (const auto *Base = dyn_cast<GetElementPtrInst>(BasePtr)) { | |||
7398 | SmallVector<const Value *> Indices(Base->indices()); | |||
7399 | VecCost = TTI->getGEPCost(Base->getSourceElementType(), | |||
7400 | Base->getPointerOperand(), Indices, CostKind); | |||
7401 | } | |||
7402 | } | |||
7403 | ||||
7404 | LLVM_DEBUG(dumpTreeCosts(E, 0, VecCost, ScalarCost,do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dumpTreeCosts(E, 0, VecCost, ScalarCost, "Calculated GEPs cost for Tree" ); } } while (false) | |||
7405 | "Calculated GEPs cost for Tree"))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dumpTreeCosts(E, 0, VecCost, ScalarCost, "Calculated GEPs cost for Tree" ); } } while (false); | |||
7406 | ||||
7407 | return VecCost - ScalarCost; | |||
7408 | }; | |||
7409 | ||||
7410 | switch (ShuffleOrOp) { | |||
7411 | case Instruction::PHI: { | |||
7412 | // Count reused scalars. | |||
7413 | InstructionCost ScalarCost = 0; | |||
7414 | SmallPtrSet<const TreeEntry *, 4> CountedOps; | |||
7415 | for (Value *V : VL) { | |||
7416 | auto *PHI = dyn_cast<PHINode>(V); | |||
7417 | if (!PHI) | |||
7418 | continue; | |||
7419 | ||||
7420 | ValueList Operands(PHI->getNumIncomingValues(), nullptr); | |||
7421 | for (unsigned I = 0, N = PHI->getNumIncomingValues(); I < N; ++I) { | |||
7422 | Value *Op = PHI->getIncomingValue(I); | |||
7423 | Operands[I] = Op; | |||
7424 | } | |||
7425 | if (const TreeEntry *OpTE = getTreeEntry(Operands.front())) | |||
7426 | if (OpTE->isSame(Operands) && CountedOps.insert(OpTE).second) | |||
7427 | if (!OpTE->ReuseShuffleIndices.empty()) | |||
7428 | ScalarCost += TTI::TCC_Basic * (OpTE->ReuseShuffleIndices.size() - | |||
7429 | OpTE->Scalars.size()); | |||
7430 | } | |||
7431 | ||||
7432 | return CommonCost - ScalarCost; | |||
7433 | } | |||
7434 | case Instruction::ExtractValue: | |||
7435 | case Instruction::ExtractElement: { | |||
7436 | auto GetScalarCost = [=](unsigned Idx) { | |||
7437 | auto *I = cast<Instruction>(VL[Idx]); | |||
7438 | VectorType *SrcVecTy; | |||
7439 | if (ShuffleOrOp == Instruction::ExtractElement) { | |||
7440 | auto *EE = cast<ExtractElementInst>(I); | |||
7441 | SrcVecTy = EE->getVectorOperandType(); | |||
7442 | } else { | |||
7443 | auto *EV = cast<ExtractValueInst>(I); | |||
7444 | Type *AggregateTy = EV->getAggregateOperand()->getType(); | |||
7445 | unsigned NumElts; | |||
7446 | if (auto *ATy = dyn_cast<ArrayType>(AggregateTy)) | |||
7447 | NumElts = ATy->getNumElements(); | |||
7448 | else | |||
7449 | NumElts = AggregateTy->getStructNumElements(); | |||
7450 | SrcVecTy = FixedVectorType::get(ScalarTy, NumElts); | |||
7451 | } | |||
7452 | if (I->hasOneUse()) { | |||
7453 | Instruction *Ext = I->user_back(); | |||
7454 | if ((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) && | |||
7455 | all_of(Ext->users(), | |||
7456 | [](User *U) { return isa<GetElementPtrInst>(U); })) { | |||
7457 | // Use getExtractWithExtendCost() to calculate the cost of | |||
7458 | // extractelement/ext pair. | |||
7459 | InstructionCost Cost = TTI->getExtractWithExtendCost( | |||
7460 | Ext->getOpcode(), Ext->getType(), SrcVecTy, *getExtractIndex(I)); | |||
7461 | // Subtract the cost of s|zext which is subtracted separately. | |||
7462 | Cost -= TTI->getCastInstrCost( | |||
7463 | Ext->getOpcode(), Ext->getType(), I->getType(), | |||
7464 | TTI::getCastContextHint(Ext), CostKind, Ext); | |||
7465 | return Cost; | |||
7466 | } | |||
7467 | } | |||
7468 | return TTI->getVectorInstrCost(Instruction::ExtractElement, SrcVecTy, | |||
7469 | CostKind, *getExtractIndex(I)); | |||
7470 | }; | |||
7471 | auto GetVectorCost = [](InstructionCost CommonCost) { return CommonCost; }; | |||
7472 | return GetCostDiff(GetScalarCost, GetVectorCost); | |||
7473 | } | |||
7474 | case Instruction::InsertElement: { | |||
7475 | 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", 7476, __extension__ __PRETTY_FUNCTION__)) | |||
7476 | "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", 7476, __extension__ __PRETTY_FUNCTION__)); | |||
7477 | auto *SrcVecTy = cast<FixedVectorType>(VL0->getType()); | |||
7478 | unsigned const NumElts = SrcVecTy->getNumElements(); | |||
7479 | unsigned const NumScalars = VL.size(); | |||
7480 | ||||
7481 | unsigned NumOfParts = TTI->getNumberOfParts(SrcVecTy); | |||
7482 | ||||
7483 | SmallVector<int> InsertMask(NumElts, PoisonMaskElem); | |||
7484 | unsigned OffsetBeg = *getInsertIndex(VL.front()); | |||
7485 | unsigned OffsetEnd = OffsetBeg; | |||
7486 | InsertMask[OffsetBeg] = 0; | |||
7487 | for (auto [I, V] : enumerate(VL.drop_front())) { | |||
7488 | unsigned Idx = *getInsertIndex(V); | |||
7489 | if (OffsetBeg > Idx) | |||
7490 | OffsetBeg = Idx; | |||
7491 | else if (OffsetEnd < Idx) | |||
7492 | OffsetEnd = Idx; | |||
7493 | InsertMask[Idx] = I + 1; | |||
7494 | } | |||
7495 | unsigned VecScalarsSz = PowerOf2Ceil(NumElts); | |||
7496 | if (NumOfParts > 0) | |||
7497 | VecScalarsSz = PowerOf2Ceil((NumElts + NumOfParts - 1) / NumOfParts); | |||
7498 | unsigned VecSz = (1 + OffsetEnd / VecScalarsSz - OffsetBeg / VecScalarsSz) * | |||
7499 | VecScalarsSz; | |||
7500 | unsigned Offset = VecScalarsSz * (OffsetBeg / VecScalarsSz); | |||
7501 | unsigned InsertVecSz = std::min<unsigned>( | |||
7502 | PowerOf2Ceil(OffsetEnd - OffsetBeg + 1), | |||
7503 | ((OffsetEnd - OffsetBeg + VecScalarsSz) / VecScalarsSz) * VecScalarsSz); | |||
7504 | bool IsWholeSubvector = | |||
7505 | OffsetBeg == Offset && ((OffsetEnd + 1) % VecScalarsSz == 0); | |||
7506 | // Check if we can safely insert a subvector. If it is not possible, just | |||
7507 | // generate a whole-sized vector and shuffle the source vector and the new | |||
7508 | // subvector. | |||
7509 | if (OffsetBeg + InsertVecSz > VecSz) { | |||
7510 | // Align OffsetBeg to generate correct mask. | |||
7511 | OffsetBeg = alignDown(OffsetBeg, VecSz, Offset); | |||
7512 | InsertVecSz = VecSz; | |||
7513 | } | |||
7514 | ||||
7515 | APInt DemandedElts = APInt::getZero(NumElts); | |||
7516 | // TODO: Add support for Instruction::InsertValue. | |||
7517 | SmallVector<int> Mask; | |||
7518 | if (!E->ReorderIndices.empty()) { | |||
7519 | inversePermutation(E->ReorderIndices, Mask); | |||
7520 | Mask.append(InsertVecSz - Mask.size(), PoisonMaskElem); | |||
7521 | } else { | |||
7522 | Mask.assign(VecSz, PoisonMaskElem); | |||
7523 | std::iota(Mask.begin(), std::next(Mask.begin(), InsertVecSz), 0); | |||
7524 | } | |||
7525 | bool IsIdentity = true; | |||
7526 | SmallVector<int> PrevMask(InsertVecSz, PoisonMaskElem); | |||
7527 | Mask.swap(PrevMask); | |||
7528 | for (unsigned I = 0; I < NumScalars; ++I) { | |||
7529 | unsigned InsertIdx = *getInsertIndex(VL[PrevMask[I]]); | |||
7530 | DemandedElts.setBit(InsertIdx); | |||
7531 | IsIdentity &= InsertIdx - OffsetBeg == I; | |||
7532 | Mask[InsertIdx - OffsetBeg] = I; | |||
7533 | } | |||
7534 | 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", 7534, __extension__ __PRETTY_FUNCTION__)); | |||
7535 | ||||
7536 | InstructionCost Cost = 0; | |||
7537 | Cost -= TTI->getScalarizationOverhead(SrcVecTy, DemandedElts, | |||
7538 | /*Insert*/ true, /*Extract*/ false, | |||
7539 | CostKind); | |||
7540 | ||||
7541 | // First cost - resize to actual vector size if not identity shuffle or | |||
7542 | // need to shift the vector. | |||
7543 | // Do not calculate the cost if the actual size is the register size and | |||
7544 | // we can merge this shuffle with the following SK_Select. | |||
7545 | auto *InsertVecTy = | |||
7546 | FixedVectorType::get(SrcVecTy->getElementType(), InsertVecSz); | |||
7547 | if (!IsIdentity) | |||
7548 | Cost += TTI->getShuffleCost(TargetTransformInfo::SK_PermuteSingleSrc, | |||
7549 | InsertVecTy, Mask); | |||
7550 | auto *FirstInsert = cast<Instruction>(*find_if(E->Scalars, [E](Value *V) { | |||
7551 | return !is_contained(E->Scalars, cast<Instruction>(V)->getOperand(0)); | |||
7552 | })); | |||
7553 | // Second cost - permutation with subvector, if some elements are from the | |||
7554 | // initial vector or inserting a subvector. | |||
7555 | // TODO: Implement the analysis of the FirstInsert->getOperand(0) | |||
7556 | // subvector of ActualVecTy. | |||
7557 | SmallBitVector InMask = | |||
7558 | isUndefVector(FirstInsert->getOperand(0), | |||
7559 | buildUseMask(NumElts, InsertMask, UseMask::UndefsAsMask)); | |||
7560 | if (!InMask.all() && NumScalars != NumElts && !IsWholeSubvector) { | |||
7561 | if (InsertVecSz != VecSz) { | |||
7562 | auto *ActualVecTy = | |||
7563 | FixedVectorType::get(SrcVecTy->getElementType(), VecSz); | |||
7564 | Cost += TTI->getShuffleCost(TTI::SK_InsertSubvector, ActualVecTy, | |||
7565 | std::nullopt, CostKind, OffsetBeg - Offset, | |||
7566 | InsertVecTy); | |||
7567 | } else { | |||
7568 | for (unsigned I = 0, End = OffsetBeg - Offset; I < End; ++I) | |||
7569 | Mask[I] = InMask.test(I) ? PoisonMaskElem : I; | |||
7570 | for (unsigned I = OffsetBeg - Offset, End = OffsetEnd - Offset; | |||
7571 | I <= End; ++I) | |||
7572 | if (Mask[I] != PoisonMaskElem) | |||
7573 | Mask[I] = I + VecSz; | |||
7574 | for (unsigned I = OffsetEnd + 1 - Offset; I < VecSz; ++I) | |||
7575 | Mask[I] = | |||
7576 | ((I >= InMask.size()) || InMask.test(I)) ? PoisonMaskElem : I; | |||
7577 | Cost += TTI->getShuffleCost(TTI::SK_PermuteTwoSrc, InsertVecTy, Mask); | |||
7578 | } | |||
7579 | } | |||
7580 | return Cost; | |||
7581 | } | |||
7582 | case Instruction::ZExt: | |||
7583 | case Instruction::SExt: | |||
7584 | case Instruction::FPToUI: | |||
7585 | case Instruction::FPToSI: | |||
7586 | case Instruction::FPExt: | |||
7587 | case Instruction::PtrToInt: | |||
7588 | case Instruction::IntToPtr: | |||
7589 | case Instruction::SIToFP: | |||
7590 | case Instruction::UIToFP: | |||
7591 | case Instruction::Trunc: | |||
7592 | case Instruction::FPTrunc: | |||
7593 | case Instruction::BitCast: { | |||
7594 | auto GetScalarCost = [=](unsigned Idx) { | |||
7595 | auto *VI = cast<Instruction>(VL[Idx]); | |||
7596 | return TTI->getCastInstrCost(E->getOpcode(), ScalarTy, | |||
7597 | VI->getOperand(0)->getType(), | |||
7598 | TTI::getCastContextHint(VI), CostKind, VI); | |||
7599 | }; | |||
7600 | auto GetVectorCost = [=](InstructionCost CommonCost) { | |||
7601 | Type *SrcTy = VL0->getOperand(0)->getType(); | |||
7602 | auto *SrcVecTy = FixedVectorType::get(SrcTy, VL.size()); | |||
7603 | InstructionCost VecCost = CommonCost; | |||
7604 | // Check if the values are candidates to demote. | |||
7605 | if (!MinBWs.count(VL0) || VecTy != SrcVecTy) | |||
7606 | VecCost += | |||
7607 | TTI->getCastInstrCost(E->getOpcode(), VecTy, SrcVecTy, | |||
7608 | TTI::getCastContextHint(VL0), CostKind, VL0); | |||
7609 | return VecCost; | |||
7610 | }; | |||
7611 | return GetCostDiff(GetScalarCost, GetVectorCost); | |||
7612 | } | |||
7613 | case Instruction::FCmp: | |||
7614 | case Instruction::ICmp: | |||
7615 | case Instruction::Select: { | |||
7616 | CmpInst::Predicate VecPred, SwappedVecPred; | |||
7617 | auto MatchCmp = m_Cmp(VecPred, m_Value(), m_Value()); | |||
7618 | if (match(VL0, m_Select(MatchCmp, m_Value(), m_Value())) || | |||
7619 | match(VL0, MatchCmp)) | |||
7620 | SwappedVecPred = CmpInst::getSwappedPredicate(VecPred); | |||
7621 | else | |||
7622 | SwappedVecPred = VecPred = ScalarTy->isFloatingPointTy() | |||
7623 | ? CmpInst::BAD_FCMP_PREDICATE | |||
7624 | : CmpInst::BAD_ICMP_PREDICATE; | |||
7625 | auto GetScalarCost = [&](unsigned Idx) { | |||
7626 | auto *VI = cast<Instruction>(VL[Idx]); | |||
7627 | CmpInst::Predicate CurrentPred = ScalarTy->isFloatingPointTy() | |||
7628 | ? CmpInst::BAD_FCMP_PREDICATE | |||
7629 | : CmpInst::BAD_ICMP_PREDICATE; | |||
7630 | auto MatchCmp = m_Cmp(CurrentPred, m_Value(), m_Value()); | |||
7631 | if ((!match(VI, m_Select(MatchCmp, m_Value(), m_Value())) && | |||
7632 | !match(VI, MatchCmp)) || | |||
7633 | (CurrentPred != VecPred && CurrentPred != SwappedVecPred)) | |||
7634 | VecPred = SwappedVecPred = ScalarTy->isFloatingPointTy() | |||
7635 | ? CmpInst::BAD_FCMP_PREDICATE | |||
7636 | : CmpInst::BAD_ICMP_PREDICATE; | |||
7637 | ||||
7638 | return TTI->getCmpSelInstrCost(E->getOpcode(), ScalarTy, | |||
7639 | Builder.getInt1Ty(), CurrentPred, CostKind, | |||
7640 | VI); | |||
7641 | }; | |||
7642 | auto GetVectorCost = [&](InstructionCost CommonCost) { | |||
7643 | auto *MaskTy = FixedVectorType::get(Builder.getInt1Ty(), VL.size()); | |||
7644 | ||||
7645 | InstructionCost VecCost = TTI->getCmpSelInstrCost( | |||
7646 | E->getOpcode(), VecTy, MaskTy, VecPred, CostKind, VL0); | |||
7647 | // Check if it is possible and profitable to use min/max for selects | |||
7648 | // in VL. | |||
7649 | // | |||
7650 | auto IntrinsicAndUse = canConvertToMinOrMaxIntrinsic(VL); | |||
7651 | if (IntrinsicAndUse.first != Intrinsic::not_intrinsic) { | |||
7652 | IntrinsicCostAttributes CostAttrs(IntrinsicAndUse.first, VecTy, | |||
7653 | {VecTy, VecTy}); | |||
7654 | InstructionCost IntrinsicCost = | |||
7655 | TTI->getIntrinsicInstrCost(CostAttrs, CostKind); | |||
7656 | // If the selects are the only uses of the compares, they will be | |||
7657 | // dead and we can adjust the cost by removing their cost. | |||
7658 | if (IntrinsicAndUse.second) | |||
7659 | IntrinsicCost -= TTI->getCmpSelInstrCost(Instruction::ICmp, VecTy, | |||
7660 | MaskTy, VecPred, CostKind); | |||
7661 | VecCost = std::min(VecCost, IntrinsicCost); | |||
7662 | } | |||
7663 | return VecCost + CommonCost; | |||
7664 | }; | |||
7665 | return GetCostDiff(GetScalarCost, GetVectorCost); | |||
7666 | } | |||
7667 | case Instruction::FNeg: | |||
7668 | case Instruction::Add: | |||
7669 | case Instruction::FAdd: | |||
7670 | case Instruction::Sub: | |||
7671 | case Instruction::FSub: | |||
7672 | case Instruction::Mul: | |||
7673 | case Instruction::FMul: | |||
7674 | case Instruction::UDiv: | |||
7675 | case Instruction::SDiv: | |||
7676 | case Instruction::FDiv: | |||
7677 | case Instruction::URem: | |||
7678 | case Instruction::SRem: | |||
7679 | case Instruction::FRem: | |||
7680 | case Instruction::Shl: | |||
7681 | case Instruction::LShr: | |||
7682 | case Instruction::AShr: | |||
7683 | case Instruction::And: | |||
7684 | case Instruction::Or: | |||
7685 | case Instruction::Xor: { | |||
7686 | auto GetScalarCost = [=](unsigned Idx) { | |||
7687 | auto *VI = cast<Instruction>(VL[Idx]); | |||
7688 | unsigned OpIdx = isa<UnaryOperator>(VI) ? 0 : 1; | |||
7689 | TTI::OperandValueInfo Op1Info = TTI::getOperandInfo(VI->getOperand(0)); | |||
7690 | TTI::OperandValueInfo Op2Info = | |||
7691 | TTI::getOperandInfo(VI->getOperand(OpIdx)); | |||
7692 | SmallVector<const Value *> Operands(VI->operand_values()); | |||
7693 | return TTI->getArithmeticInstrCost(ShuffleOrOp, ScalarTy, CostKind, | |||
7694 | Op1Info, Op2Info, Operands, VI); | |||
7695 | }; | |||
7696 | auto GetVectorCost = [=](InstructionCost CommonCost) { | |||
7697 | unsigned OpIdx = isa<UnaryOperator>(VL0) ? 0 : 1; | |||
7698 | TTI::OperandValueInfo Op1Info = getOperandInfo(VL, 0); | |||
7699 | TTI::OperandValueInfo Op2Info = getOperandInfo(VL, OpIdx); | |||
7700 | return TTI->getArithmeticInstrCost(ShuffleOrOp, VecTy, CostKind, Op1Info, | |||
7701 | Op2Info) + | |||
7702 | CommonCost; | |||
7703 | }; | |||
7704 | return GetCostDiff(GetScalarCost, GetVectorCost); | |||
7705 | } | |||
7706 | case Instruction::GetElementPtr: { | |||
7707 | return CommonCost + GetGEPCostDiff(VL, VL0); | |||
7708 | } | |||
7709 | case Instruction::Load: { | |||
7710 | auto GetScalarCost = [=](unsigned Idx) { | |||
7711 | auto *VI = cast<LoadInst>(VL[Idx]); | |||
7712 | return TTI->getMemoryOpCost(Instruction::Load, ScalarTy, VI->getAlign(), | |||
7713 | VI->getPointerAddressSpace(), CostKind, | |||
7714 | TTI::OperandValueInfo(), VI); | |||
7715 | }; | |||
7716 | auto *LI0 = cast<LoadInst>(VL0); | |||
7717 | auto GetVectorCost = [=](InstructionCost CommonCost) { | |||
7718 | InstructionCost VecLdCost; | |||
7719 | if (E->State == TreeEntry::Vectorize) { | |||
7720 | VecLdCost = TTI->getMemoryOpCost( | |||
7721 | Instruction::Load, VecTy, LI0->getAlign(), | |||
7722 | LI0->getPointerAddressSpace(), CostKind, TTI::OperandValueInfo()); | |||
7723 | } else { | |||
7724 | 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", 7724, __extension__ __PRETTY_FUNCTION__)); | |||
7725 | Align CommonAlignment = LI0->getAlign(); | |||
7726 | for (Value *V : VL) | |||
7727 | CommonAlignment = | |||
7728 | std::min(CommonAlignment, cast<LoadInst>(V)->getAlign()); | |||
7729 | VecLdCost = TTI->getGatherScatterOpCost( | |||
7730 | Instruction::Load, VecTy, LI0->getPointerOperand(), | |||
7731 | /*VariableMask=*/false, CommonAlignment, CostKind); | |||
7732 | } | |||
7733 | return VecLdCost + CommonCost; | |||
7734 | }; | |||
7735 | ||||
7736 | InstructionCost Cost = GetCostDiff(GetScalarCost, GetVectorCost); | |||
7737 | // If this node generates masked gather load then it is not a terminal node. | |||
7738 | // Hence address operand cost is estimated separately. | |||
7739 | if (E->State == TreeEntry::ScatterVectorize) | |||
7740 | return Cost; | |||
7741 | ||||
7742 | // Estimate cost of GEPs since this tree node is a terminator. | |||
7743 | SmallVector<Value *> PointerOps(VL.size()); | |||
7744 | for (auto [I, V] : enumerate(VL)) | |||
7745 | PointerOps[I] = cast<LoadInst>(V)->getPointerOperand(); | |||
7746 | return Cost + GetGEPCostDiff(PointerOps, LI0->getPointerOperand()); | |||
7747 | } | |||
7748 | case Instruction::Store: { | |||
7749 | bool IsReorder = !E->ReorderIndices.empty(); | |||
7750 | auto GetScalarCost = [=](unsigned Idx) { | |||
7751 | auto *VI = cast<StoreInst>(VL[Idx]); | |||
7752 | TTI::OperandValueInfo OpInfo = getOperandInfo(VI, 0); | |||
7753 | return TTI->getMemoryOpCost(Instruction::Store, ScalarTy, VI->getAlign(), | |||
7754 | VI->getPointerAddressSpace(), CostKind, | |||
7755 | OpInfo, VI); | |||
7756 | }; | |||
7757 | auto *BaseSI = | |||
7758 | cast<StoreInst>(IsReorder ? VL[E->ReorderIndices.front()] : VL0); | |||
7759 | auto GetVectorCost = [=](InstructionCost CommonCost) { | |||
7760 | // We know that we can merge the stores. Calculate the cost. | |||
7761 | TTI::OperandValueInfo OpInfo = getOperandInfo(VL, 0); | |||
7762 | return TTI->getMemoryOpCost(Instruction::Store, VecTy, BaseSI->getAlign(), | |||
7763 | BaseSI->getPointerAddressSpace(), CostKind, | |||
7764 | OpInfo) + | |||
7765 | CommonCost; | |||
7766 | }; | |||
7767 | SmallVector<Value *> PointerOps(VL.size()); | |||
7768 | for (auto [I, V] : enumerate(VL)) { | |||
7769 | unsigned Idx = IsReorder ? E->ReorderIndices[I] : I; | |||
7770 | PointerOps[Idx] = cast<StoreInst>(V)->getPointerOperand(); | |||
7771 | } | |||
7772 | ||||
7773 | return GetCostDiff(GetScalarCost, GetVectorCost) + | |||
7774 | GetGEPCostDiff(PointerOps, BaseSI->getPointerOperand()); | |||
7775 | } | |||
7776 | case Instruction::Call: { | |||
7777 | auto GetScalarCost = [=](unsigned Idx) { | |||
7778 | auto *CI = cast<CallInst>(VL[Idx]); | |||
7779 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); | |||
7780 | if (ID != Intrinsic::not_intrinsic) { | |||
7781 | IntrinsicCostAttributes CostAttrs(ID, *CI, 1); | |||
7782 | return TTI->getIntrinsicInstrCost(CostAttrs, CostKind); | |||
7783 | } | |||
7784 | return TTI->getCallInstrCost(CI->getCalledFunction(), | |||
7785 | CI->getFunctionType()->getReturnType(), | |||
7786 | CI->getFunctionType()->params(), CostKind); | |||
7787 | }; | |||
7788 | auto GetVectorCost = [=](InstructionCost CommonCost) { | |||
7789 | auto *CI = cast<CallInst>(VL0); | |||
7790 | auto VecCallCosts = getVectorCallCosts(CI, VecTy, TTI, TLI); | |||
7791 | return std::min(VecCallCosts.first, VecCallCosts.second) + CommonCost; | |||
7792 | }; | |||
7793 | return GetCostDiff(GetScalarCost, GetVectorCost); | |||
7794 | } | |||
7795 | case Instruction::ShuffleVector: { | |||
7796 | 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", 7802, __extension__ __PRETTY_FUNCTION__)) | |||
7797 | ((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", 7802, __extension__ __PRETTY_FUNCTION__)) | |||
7798 | 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", 7802, __extension__ __PRETTY_FUNCTION__)) | |||
7799 | (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", 7802, __extension__ __PRETTY_FUNCTION__)) | |||
7800 | 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", 7802, __extension__ __PRETTY_FUNCTION__)) | |||
7801 | (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", 7802, __extension__ __PRETTY_FUNCTION__)) | |||
7802 | "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", 7802, __extension__ __PRETTY_FUNCTION__)); | |||
7803 | // Try to find the previous shuffle node with the same operands and same | |||
7804 | // main/alternate ops. | |||
7805 | auto TryFindNodeWithEqualOperands = [=]() { | |||
7806 | for (const std::unique_ptr<TreeEntry> &TE : VectorizableTree) { | |||
7807 | if (TE.get() == E) | |||
7808 | break; | |||
7809 | if (TE->isAltShuffle() && | |||
7810 | ((TE->getOpcode() == E->getOpcode() && | |||
7811 | TE->getAltOpcode() == E->getAltOpcode()) || | |||
7812 | (TE->getOpcode() == E->getAltOpcode() && | |||
7813 | TE->getAltOpcode() == E->getOpcode())) && | |||
7814 | TE->hasEqualOperands(*E)) | |||
7815 | return true; | |||
7816 | } | |||
7817 | return false; | |||
7818 | }; | |||
7819 | auto GetScalarCost = [=](unsigned Idx) { | |||
7820 | auto *VI = cast<Instruction>(VL[Idx]); | |||
7821 | assert(E->isOpcodeOrAlt(VI) && "Unexpected main/alternate opcode")(static_cast <bool> (E->isOpcodeOrAlt(VI) && "Unexpected main/alternate opcode") ? void (0) : __assert_fail ("E->isOpcodeOrAlt(VI) && \"Unexpected main/alternate opcode\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7821, __extension__ __PRETTY_FUNCTION__)); | |||
7822 | (void)E; | |||
7823 | return TTI->getInstructionCost(VI, CostKind); | |||
7824 | }; | |||
7825 | // Need to clear CommonCost since the final shuffle cost is included into | |||
7826 | // vector cost. | |||
7827 | auto GetVectorCost = [&](InstructionCost) { | |||
7828 | // VecCost is equal to sum of the cost of creating 2 vectors | |||
7829 | // and the cost of creating shuffle. | |||
7830 | InstructionCost VecCost = 0; | |||
7831 | if (TryFindNodeWithEqualOperands()) { | |||
7832 | LLVM_DEBUG({do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { { dbgs() << "SLP: diamond match for alternate node found.\n" ; E->dump(); }; } } while (false) | |||
7833 | 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) | |||
7834 | E->dump();do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { { dbgs() << "SLP: diamond match for alternate node found.\n" ; E->dump(); }; } } while (false) | |||
7835 | })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { { dbgs() << "SLP: diamond match for alternate node found.\n" ; E->dump(); }; } } while (false); | |||
7836 | // No need to add new vector costs here since we're going to reuse | |||
7837 | // same main/alternate vector ops, just do different shuffling. | |||
7838 | } else if (Instruction::isBinaryOp(E->getOpcode())) { | |||
7839 | VecCost = TTI->getArithmeticInstrCost(E->getOpcode(), VecTy, CostKind); | |||
7840 | VecCost += | |||
7841 | TTI->getArithmeticInstrCost(E->getAltOpcode(), VecTy, CostKind); | |||
7842 | } else if (auto *CI0 = dyn_cast<CmpInst>(VL0)) { | |||
7843 | VecCost = TTI->getCmpSelInstrCost(E->getOpcode(), ScalarTy, | |||
7844 | Builder.getInt1Ty(), | |||
7845 | CI0->getPredicate(), CostKind, VL0); | |||
7846 | VecCost += TTI->getCmpSelInstrCost( | |||
7847 | E->getOpcode(), ScalarTy, Builder.getInt1Ty(), | |||
7848 | cast<CmpInst>(E->getAltOp())->getPredicate(), CostKind, | |||
7849 | E->getAltOp()); | |||
7850 | } else { | |||
7851 | Type *Src0SclTy = E->getMainOp()->getOperand(0)->getType(); | |||
7852 | Type *Src1SclTy = E->getAltOp()->getOperand(0)->getType(); | |||
7853 | auto *Src0Ty = FixedVectorType::get(Src0SclTy, VL.size()); | |||
7854 | auto *Src1Ty = FixedVectorType::get(Src1SclTy, VL.size()); | |||
7855 | VecCost = TTI->getCastInstrCost(E->getOpcode(), VecTy, Src0Ty, | |||
7856 | TTI::CastContextHint::None, CostKind); | |||
7857 | VecCost += TTI->getCastInstrCost(E->getAltOpcode(), VecTy, Src1Ty, | |||
7858 | TTI::CastContextHint::None, CostKind); | |||
7859 | } | |||
7860 | if (E->ReuseShuffleIndices.empty()) { | |||
7861 | VecCost += | |||
7862 | TTI->getShuffleCost(TargetTransformInfo::SK_Select, FinalVecTy); | |||
7863 | } else { | |||
7864 | SmallVector<int> Mask; | |||
7865 | buildShuffleEntryMask( | |||
7866 | E->Scalars, E->ReorderIndices, E->ReuseShuffleIndices, | |||
7867 | [E](Instruction *I) { | |||
7868 | 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", 7868, __extension__ __PRETTY_FUNCTION__)); | |||
7869 | return I->getOpcode() == E->getAltOpcode(); | |||
7870 | }, | |||
7871 | Mask); | |||
7872 | VecCost += TTI->getShuffleCost(TargetTransformInfo::SK_PermuteTwoSrc, | |||
7873 | FinalVecTy, Mask); | |||
7874 | } | |||
7875 | return VecCost; | |||
7876 | }; | |||
7877 | return GetCostDiff(GetScalarCost, GetVectorCost); | |||
7878 | } | |||
7879 | default: | |||
7880 | llvm_unreachable("Unknown instruction")::llvm::llvm_unreachable_internal("Unknown instruction", "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 7880); | |||
7881 | } | |||
7882 | } | |||
7883 | ||||
7884 | bool BoUpSLP::isFullyVectorizableTinyTree(bool ForReduction) const { | |||
7885 | 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) | |||
7886 | << 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); | |||
7887 | ||||
7888 | auto &&AreVectorizableGathers = [this](const TreeEntry *TE, unsigned Limit) { | |||
7889 | SmallVector<int> Mask; | |||
7890 | return TE->State == TreeEntry::NeedToGather && | |||
7891 | !any_of(TE->Scalars, | |||
7892 | [this](Value *V) { return EphValues.contains(V); }) && | |||
7893 | (allConstant(TE->Scalars) || isSplat(TE->Scalars) || | |||
7894 | TE->Scalars.size() < Limit || | |||
7895 | ((TE->getOpcode() == Instruction::ExtractElement || | |||
7896 | all_of(TE->Scalars, | |||
7897 | [](Value *V) { | |||
7898 | return isa<ExtractElementInst, UndefValue>(V); | |||
7899 | })) && | |||
7900 | isFixedVectorShuffle(TE->Scalars, Mask)) || | |||
7901 | (TE->State == TreeEntry::NeedToGather && | |||
7902 | TE->getOpcode() == Instruction::Load && !TE->isAltShuffle())); | |||
7903 | }; | |||
7904 | ||||
7905 | // We only handle trees of heights 1 and 2. | |||
7906 | if (VectorizableTree.size() == 1 && | |||
7907 | (VectorizableTree[0]->State == TreeEntry::Vectorize || | |||
7908 | (ForReduction && | |||
7909 | AreVectorizableGathers(VectorizableTree[0].get(), | |||
7910 | VectorizableTree[0]->Scalars.size()) && | |||
7911 | VectorizableTree[0]->getVectorFactor() > 2))) | |||
7912 | return true; | |||
7913 | ||||
7914 | if (VectorizableTree.size() != 2) | |||
7915 | return false; | |||
7916 | ||||
7917 | // Handle splat and all-constants stores. Also try to vectorize tiny trees | |||
7918 | // with the second gather nodes if they have less scalar operands rather than | |||
7919 | // the initial tree element (may be profitable to shuffle the second gather) | |||
7920 | // or they are extractelements, which form shuffle. | |||
7921 | SmallVector<int> Mask; | |||
7922 | if (VectorizableTree[0]->State == TreeEntry::Vectorize && | |||
7923 | AreVectorizableGathers(VectorizableTree[1].get(), | |||
7924 | VectorizableTree[0]->Scalars.size())) | |||
7925 | return true; | |||
7926 | ||||
7927 | // Gathering cost would be too much for tiny trees. | |||
7928 | if (VectorizableTree[0]->State == TreeEntry::NeedToGather || | |||
7929 | (VectorizableTree[1]->State == TreeEntry::NeedToGather && | |||
7930 | VectorizableTree[0]->State != TreeEntry::ScatterVectorize)) | |||
7931 | return false; | |||
7932 | ||||
7933 | return true; | |||
7934 | } | |||
7935 | ||||
7936 | static bool isLoadCombineCandidateImpl(Value *Root, unsigned NumElts, | |||
7937 | TargetTransformInfo *TTI, | |||
7938 | bool MustMatchOrInst) { | |||
7939 | // Look past the root to find a source value. Arbitrarily follow the | |||
7940 | // path through operand 0 of any 'or'. Also, peek through optional | |||
7941 | // shift-left-by-multiple-of-8-bits. | |||
7942 | Value *ZextLoad = Root; | |||
7943 | const APInt *ShAmtC; | |||
7944 | bool FoundOr = false; | |||
7945 | while (!isa<ConstantExpr>(ZextLoad) && | |||
7946 | (match(ZextLoad, m_Or(m_Value(), m_Value())) || | |||
7947 | (match(ZextLoad, m_Shl(m_Value(), m_APInt(ShAmtC))) && | |||
7948 | ShAmtC->urem(8) == 0))) { | |||
7949 | auto *BinOp = cast<BinaryOperator>(ZextLoad); | |||
7950 | ZextLoad = BinOp->getOperand(0); | |||
7951 | if (BinOp->getOpcode() == Instruction::Or) | |||
7952 | FoundOr = true; | |||
7953 | } | |||
7954 | // Check if the input is an extended load of the required or/shift expression. | |||
7955 | Value *Load; | |||
7956 | if ((MustMatchOrInst && !FoundOr) || ZextLoad == Root || | |||
7957 | !match(ZextLoad, m_ZExt(m_Value(Load))) || !isa<LoadInst>(Load)) | |||
7958 | return false; | |||
7959 | ||||
7960 | // Require that the total load bit width is a legal integer type. | |||
7961 | // For example, <8 x i8> --> i64 is a legal integer on a 64-bit target. | |||
7962 | // But <16 x i8> --> i128 is not, so the backend probably can't reduce it. | |||
7963 | Type *SrcTy = Load->getType(); | |||
7964 | unsigned LoadBitWidth = SrcTy->getIntegerBitWidth() * NumElts; | |||
7965 | if (!TTI->isTypeLegal(IntegerType::get(Root->getContext(), LoadBitWidth))) | |||
7966 | return false; | |||
7967 | ||||
7968 | // Everything matched - assume that we can fold the whole sequence using | |||
7969 | // load combining. | |||
7970 | 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) | |||
7971 | << *(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); | |||
7972 | ||||
7973 | return true; | |||
7974 | } | |||
7975 | ||||
7976 | bool BoUpSLP::isLoadCombineReductionCandidate(RecurKind RdxKind) const { | |||
7977 | if (RdxKind != RecurKind::Or) | |||
7978 | return false; | |||
7979 | ||||
7980 | unsigned NumElts = VectorizableTree[0]->Scalars.size(); | |||
7981 | Value *FirstReduced = VectorizableTree[0]->Scalars[0]; | |||
7982 | return isLoadCombineCandidateImpl(FirstReduced, NumElts, TTI, | |||
7983 | /* MatchOr */ false); | |||
7984 | } | |||
7985 | ||||
7986 | bool BoUpSLP::isLoadCombineCandidate() const { | |||
7987 | // Peek through a final sequence of stores and check if all operations are | |||
7988 | // likely to be load-combined. | |||
7989 | unsigned NumElts = VectorizableTree[0]->Scalars.size(); | |||
7990 | for (Value *Scalar : VectorizableTree[0]->Scalars) { | |||
7991 | Value *X; | |||
7992 | if (!match(Scalar, m_Store(m_Value(X), m_Value())) || | |||
7993 | !isLoadCombineCandidateImpl(X, NumElts, TTI, /* MatchOr */ true)) | |||
7994 | return false; | |||
7995 | } | |||
7996 | return true; | |||
7997 | } | |||
7998 | ||||
7999 | bool BoUpSLP::isTreeTinyAndNotFullyVectorizable(bool ForReduction) const { | |||
8000 | // No need to vectorize inserts of gathered values. | |||
8001 | if (VectorizableTree.size() == 2 && | |||
8002 | isa<InsertElementInst>(VectorizableTree[0]->Scalars[0]) && | |||
8003 | VectorizableTree[1]->State == TreeEntry::NeedToGather && | |||
8004 | (VectorizableTree[1]->getVectorFactor() <= 2 || | |||
8005 | !(isSplat(VectorizableTree[1]->Scalars) || | |||
8006 | allConstant(VectorizableTree[1]->Scalars)))) | |||
8007 | return true; | |||
8008 | ||||
8009 | // We can vectorize the tree if its size is greater than or equal to the | |||
8010 | // minimum size specified by the MinTreeSize command line option. | |||
8011 | if (VectorizableTree.size() >= MinTreeSize) | |||
8012 | return false; | |||
8013 | ||||
8014 | // If we have a tiny tree (a tree whose size is less than MinTreeSize), we | |||
8015 | // can vectorize it if we can prove it fully vectorizable. | |||
8016 | if (isFullyVectorizableTinyTree(ForReduction)) | |||
8017 | return false; | |||
8018 | ||||
8019 | 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", 8021, __extension__ __PRETTY_FUNCTION__)) | |||
8020 | ? 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", 8021, __extension__ __PRETTY_FUNCTION__)) | |||
8021 | : 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", 8021, __extension__ __PRETTY_FUNCTION__)); | |||
8022 | ||||
8023 | // Otherwise, we can't vectorize the tree. It is both tiny and not fully | |||
8024 | // vectorizable. | |||
8025 | return true; | |||
8026 | } | |||
8027 | ||||
8028 | InstructionCost BoUpSLP::getSpillCost() const { | |||
8029 | // Walk from the bottom of the tree to the top, tracking which values are | |||
8030 | // live. When we see a call instruction that is not part of our tree, | |||
8031 | // query TTI to see if there is a cost to keeping values live over it | |||
8032 | // (for example, if spills and fills are required). | |||
8033 | unsigned BundleWidth = VectorizableTree.front()->Scalars.size(); | |||
8034 | InstructionCost Cost = 0; | |||
8035 | ||||
8036 | SmallPtrSet<Instruction*, 4> LiveValues; | |||
8037 | Instruction *PrevInst = nullptr; | |||
8038 | ||||
8039 | // The entries in VectorizableTree are not necessarily ordered by their | |||
8040 | // position in basic blocks. Collect them and order them by dominance so later | |||
8041 | // instructions are guaranteed to be visited first. For instructions in | |||
8042 | // different basic blocks, we only scan to the beginning of the block, so | |||
8043 | // their order does not matter, as long as all instructions in a basic block | |||
8044 | // are grouped together. Using dominance ensures a deterministic order. | |||
8045 | SmallVector<Instruction *, 16> OrderedScalars; | |||
8046 | for (const auto &TEPtr : VectorizableTree) { | |||
8047 | Instruction *Inst = dyn_cast<Instruction>(TEPtr->Scalars[0]); | |||
8048 | if (!Inst) | |||
8049 | continue; | |||
8050 | OrderedScalars.push_back(Inst); | |||
8051 | } | |||
8052 | llvm::sort(OrderedScalars, [&](Instruction *A, Instruction *B) { | |||
8053 | auto *NodeA = DT->getNode(A->getParent()); | |||
8054 | auto *NodeB = DT->getNode(B->getParent()); | |||
8055 | 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", 8055, __extension__ __PRETTY_FUNCTION__)); | |||
8056 | 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", 8056, __extension__ __PRETTY_FUNCTION__)); | |||
8057 | 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", 8058, __extension__ __PRETTY_FUNCTION__)) | |||
8058 | "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", 8058, __extension__ __PRETTY_FUNCTION__)); | |||
8059 | if (NodeA != NodeB) | |||
8060 | return NodeA->getDFSNumIn() < NodeB->getDFSNumIn(); | |||
8061 | return B->comesBefore(A); | |||
8062 | }); | |||
8063 | ||||
8064 | for (Instruction *Inst : OrderedScalars) { | |||
8065 | if (!PrevInst) { | |||
8066 | PrevInst = Inst; | |||
8067 | continue; | |||
8068 | } | |||
8069 | ||||
8070 | // Update LiveValues. | |||
8071 | LiveValues.erase(PrevInst); | |||
8072 | for (auto &J : PrevInst->operands()) { | |||
8073 | if (isa<Instruction>(&*J) && getTreeEntry(&*J)) | |||
8074 | LiveValues.insert(cast<Instruction>(&*J)); | |||
8075 | } | |||
8076 | ||||
8077 | 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) | |||
8078 | 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) | |||
8079 | 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) | |||
8080 | 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) | |||
8081 | 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) | |||
8082 | 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) | |||
8083 | })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); | |||
8084 | ||||
8085 | // Now find the sequence of instructions between PrevInst and Inst. | |||
8086 | unsigned NumCalls = 0; | |||
8087 | BasicBlock::reverse_iterator InstIt = ++Inst->getIterator().getReverse(), | |||
8088 | PrevInstIt = | |||
8089 | PrevInst->getIterator().getReverse(); | |||
8090 | while (InstIt != PrevInstIt) { | |||
8091 | if (PrevInstIt == PrevInst->getParent()->rend()) { | |||
8092 | PrevInstIt = Inst->getParent()->rbegin(); | |||
8093 | continue; | |||
8094 | } | |||
8095 | ||||
8096 | auto NoCallIntrinsic = [this](Instruction *I) { | |||
8097 | if (auto *II = dyn_cast<IntrinsicInst>(I)) { | |||
8098 | if (II->isAssumeLikeIntrinsic()) | |||
8099 | return true; | |||
8100 | FastMathFlags FMF; | |||
8101 | SmallVector<Type *, 4> Tys; | |||
8102 | for (auto &ArgOp : II->args()) | |||
8103 | Tys.push_back(ArgOp->getType()); | |||
8104 | if (auto *FPMO = dyn_cast<FPMathOperator>(II)) | |||
8105 | FMF = FPMO->getFastMathFlags(); | |||
8106 | IntrinsicCostAttributes ICA(II->getIntrinsicID(), II->getType(), Tys, | |||
8107 | FMF); | |||
8108 | InstructionCost IntrCost = | |||
8109 | TTI->getIntrinsicInstrCost(ICA, TTI::TCK_RecipThroughput); | |||
8110 | InstructionCost CallCost = TTI->getCallInstrCost( | |||
8111 | nullptr, II->getType(), Tys, TTI::TCK_RecipThroughput); | |||
8112 | if (IntrCost < CallCost) | |||
8113 | return true; | |||
8114 | } | |||
8115 | return false; | |||
8116 | }; | |||
8117 | ||||
8118 | // Debug information does not impact spill cost. | |||
8119 | if (isa<CallInst>(&*PrevInstIt) && !NoCallIntrinsic(&*PrevInstIt) && | |||
8120 | &*PrevInstIt != PrevInst) | |||
8121 | NumCalls++; | |||
8122 | ||||
8123 | ++PrevInstIt; | |||
8124 | } | |||
8125 | ||||
8126 | if (NumCalls) { | |||
8127 | SmallVector<Type*, 4> V; | |||
8128 | for (auto *II : LiveValues) { | |||
8129 | auto *ScalarTy = II->getType(); | |||
8130 | if (auto *VectorTy = dyn_cast<FixedVectorType>(ScalarTy)) | |||
8131 | ScalarTy = VectorTy->getElementType(); | |||
8132 | V.push_back(FixedVectorType::get(ScalarTy, BundleWidth)); | |||
8133 | } | |||
8134 | Cost += NumCalls * TTI->getCostOfKeepingLiveOverCall(V); | |||
8135 | } | |||
8136 | ||||
8137 | PrevInst = Inst; | |||
8138 | } | |||
8139 | ||||
8140 | return Cost; | |||
8141 | } | |||
8142 | ||||
8143 | /// Checks if the \p IE1 instructions is followed by \p IE2 instruction in the | |||
8144 | /// buildvector sequence. | |||
8145 | static bool isFirstInsertElement(const InsertElementInst *IE1, | |||
8146 | const InsertElementInst *IE2) { | |||
8147 | if (IE1 == IE2) | |||
8148 | return false; | |||
8149 | const auto *I1 = IE1; | |||
8150 | const auto *I2 = IE2; | |||
8151 | const InsertElementInst *PrevI1; | |||
8152 | const InsertElementInst *PrevI2; | |||
8153 | unsigned Idx1 = *getInsertIndex(IE1); | |||
8154 | unsigned Idx2 = *getInsertIndex(IE2); | |||
8155 | do { | |||
8156 | if (I2 == IE1) | |||
8157 | return true; | |||
8158 | if (I1 == IE2) | |||
8159 | return false; | |||
8160 | PrevI1 = I1; | |||
8161 | PrevI2 = I2; | |||
8162 | if (I1 && (I1 == IE1 || I1->hasOneUse()) && | |||
8163 | getInsertIndex(I1).value_or(Idx2) != Idx2) | |||
8164 | I1 = dyn_cast<InsertElementInst>(I1->getOperand(0)); | |||
8165 | if (I2 && ((I2 == IE2 || I2->hasOneUse())) && | |||
8166 | getInsertIndex(I2).value_or(Idx1) != Idx1) | |||
8167 | I2 = dyn_cast<InsertElementInst>(I2->getOperand(0)); | |||
8168 | } while ((I1 && PrevI1 != I1) || (I2 && PrevI2 != I2)); | |||
8169 | llvm_unreachable("Two different buildvectors not expected.")::llvm::llvm_unreachable_internal("Two different buildvectors not expected." , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8169); | |||
8170 | } | |||
8171 | ||||
8172 | namespace { | |||
8173 | /// Returns incoming Value *, if the requested type is Value * too, or a default | |||
8174 | /// value, otherwise. | |||
8175 | struct ValueSelect { | |||
8176 | template <typename U> | |||
8177 | static std::enable_if_t<std::is_same_v<Value *, U>, Value *> get(Value *V) { | |||
8178 | return V; | |||
8179 | } | |||
8180 | template <typename U> | |||
8181 | static std::enable_if_t<!std::is_same_v<Value *, U>, U> get(Value *) { | |||
8182 | return U(); | |||
8183 | } | |||
8184 | }; | |||
8185 | } // namespace | |||
8186 | ||||
8187 | /// Does the analysis of the provided shuffle masks and performs the requested | |||
8188 | /// actions on the vectors with the given shuffle masks. It tries to do it in | |||
8189 | /// several steps. | |||
8190 | /// 1. If the Base vector is not undef vector, resizing the very first mask to | |||
8191 | /// have common VF and perform action for 2 input vectors (including non-undef | |||
8192 | /// Base). Other shuffle masks are combined with the resulting after the 1 stage | |||
8193 | /// and processed as a shuffle of 2 elements. | |||
8194 | /// 2. If the Base is undef vector and have only 1 shuffle mask, perform the | |||
8195 | /// action only for 1 vector with the given mask, if it is not the identity | |||
8196 | /// mask. | |||
8197 | /// 3. If > 2 masks are used, perform the remaining shuffle actions for 2 | |||
8198 | /// vectors, combing the masks properly between the steps. | |||
8199 | template <typename T> | |||
8200 | static T *performExtractsShuffleAction( | |||
8201 | MutableArrayRef<std::pair<T *, SmallVector<int>>> ShuffleMask, Value *Base, | |||
8202 | function_ref<unsigned(T *)> GetVF, | |||
8203 | function_ref<std::pair<T *, bool>(T *, ArrayRef<int>, bool)> ResizeAction, | |||
8204 | function_ref<T *(ArrayRef<int>, ArrayRef<T *>)> Action) { | |||
8205 | assert(!ShuffleMask.empty() && "Empty list of shuffles for inserts.")(static_cast <bool> (!ShuffleMask.empty() && "Empty list of shuffles for inserts." ) ? void (0) : __assert_fail ("!ShuffleMask.empty() && \"Empty list of shuffles for inserts.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8205, __extension__ __PRETTY_FUNCTION__)); | |||
8206 | SmallVector<int> Mask(ShuffleMask.begin()->second); | |||
8207 | auto VMIt = std::next(ShuffleMask.begin()); | |||
8208 | T *Prev = nullptr; | |||
8209 | SmallBitVector UseMask = | |||
8210 | buildUseMask(Mask.size(), Mask, UseMask::UndefsAsMask); | |||
8211 | SmallBitVector IsBaseUndef = isUndefVector(Base, UseMask); | |||
8212 | if (!IsBaseUndef.all()) { | |||
8213 | // Base is not undef, need to combine it with the next subvectors. | |||
8214 | std::pair<T *, bool> Res = | |||
8215 | ResizeAction(ShuffleMask.begin()->first, Mask, /*ForSingleMask=*/false); | |||
8216 | SmallBitVector IsBasePoison = isUndefVector<true>(Base, UseMask); | |||
8217 | for (unsigned Idx = 0, VF = Mask.size(); Idx < VF; ++Idx) { | |||
8218 | if (Mask[Idx] == PoisonMaskElem) | |||
8219 | Mask[Idx] = IsBasePoison.test(Idx) ? PoisonMaskElem : Idx; | |||
8220 | else | |||
8221 | Mask[Idx] = (Res.second ? Idx : Mask[Idx]) + VF; | |||
8222 | } | |||
8223 | auto *V = ValueSelect::get<T *>(Base); | |||
8224 | (void)V; | |||
8225 | assert((!V || GetVF(V) == Mask.size()) &&(static_cast <bool> ((!V || GetVF(V) == Mask.size()) && "Expected base vector of VF number of elements.") ? void (0) : __assert_fail ("(!V || GetVF(V) == Mask.size()) && \"Expected base vector of VF number of elements.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8226, __extension__ __PRETTY_FUNCTION__)) | |||
8226 | "Expected base vector of VF number of elements.")(static_cast <bool> ((!V || GetVF(V) == Mask.size()) && "Expected base vector of VF number of elements.") ? void (0) : __assert_fail ("(!V || GetVF(V) == Mask.size()) && \"Expected base vector of VF number of elements.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8226, __extension__ __PRETTY_FUNCTION__)); | |||
8227 | Prev = Action(Mask, {nullptr, Res.first}); | |||
8228 | } else if (ShuffleMask.size() == 1) { | |||
8229 | // Base is undef and only 1 vector is shuffled - perform the action only for | |||
8230 | // single vector, if the mask is not the identity mask. | |||
8231 | std::pair<T *, bool> Res = ResizeAction(ShuffleMask.begin()->first, Mask, | |||
8232 | /*ForSingleMask=*/true); | |||
8233 | if (Res.second) | |||
8234 | // Identity mask is found. | |||
8235 | Prev = Res.first; | |||
8236 | else | |||
8237 | Prev = Action(Mask, {ShuffleMask.begin()->first}); | |||
8238 | } else { | |||
8239 | // Base is undef and at least 2 input vectors shuffled - perform 2 vectors | |||
8240 | // shuffles step by step, combining shuffle between the steps. | |||
8241 | unsigned Vec1VF = GetVF(ShuffleMask.begin()->first); | |||
8242 | unsigned Vec2VF = GetVF(VMIt->first); | |||
8243 | if (Vec1VF == Vec2VF) { | |||
8244 | // No need to resize the input vectors since they are of the same size, we | |||
8245 | // can shuffle them directly. | |||
8246 | ArrayRef<int> SecMask = VMIt->second; | |||
8247 | for (unsigned I = 0, VF = Mask.size(); I < VF; ++I) { | |||
8248 | if (SecMask[I] != PoisonMaskElem) { | |||
8249 | assert(Mask[I] == PoisonMaskElem && "Multiple uses of scalars.")(static_cast <bool> (Mask[I] == PoisonMaskElem && "Multiple uses of scalars.") ? void (0) : __assert_fail ("Mask[I] == PoisonMaskElem && \"Multiple uses of scalars.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8249, __extension__ __PRETTY_FUNCTION__)); | |||
8250 | Mask[I] = SecMask[I] + Vec1VF; | |||
8251 | } | |||
8252 | } | |||
8253 | Prev = Action(Mask, {ShuffleMask.begin()->first, VMIt->first}); | |||
8254 | } else { | |||
8255 | // Vectors of different sizes - resize and reshuffle. | |||
8256 | std::pair<T *, bool> Res1 = ResizeAction(ShuffleMask.begin()->first, Mask, | |||
8257 | /*ForSingleMask=*/false); | |||
8258 | std::pair<T *, bool> Res2 = | |||
8259 | ResizeAction(VMIt->first, VMIt->second, /*ForSingleMask=*/false); | |||
8260 | ArrayRef<int> SecMask = VMIt->second; | |||
8261 | for (unsigned I = 0, VF = Mask.size(); I < VF; ++I) { | |||
8262 | if (Mask[I] != PoisonMaskElem) { | |||
8263 | assert(SecMask[I] == PoisonMaskElem && "Multiple uses of scalars.")(static_cast <bool> (SecMask[I] == PoisonMaskElem && "Multiple uses of scalars.") ? void (0) : __assert_fail ("SecMask[I] == PoisonMaskElem && \"Multiple uses of scalars.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8263, __extension__ __PRETTY_FUNCTION__)); | |||
8264 | if (Res1.second) | |||
8265 | Mask[I] = I; | |||
8266 | } else if (SecMask[I] != PoisonMaskElem) { | |||
8267 | assert(Mask[I] == PoisonMaskElem && "Multiple uses of scalars.")(static_cast <bool> (Mask[I] == PoisonMaskElem && "Multiple uses of scalars.") ? void (0) : __assert_fail ("Mask[I] == PoisonMaskElem && \"Multiple uses of scalars.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8267, __extension__ __PRETTY_FUNCTION__)); | |||
8268 | Mask[I] = (Res2.second ? I : SecMask[I]) + VF; | |||
8269 | } | |||
8270 | } | |||
8271 | Prev = Action(Mask, {Res1.first, Res2.first}); | |||
8272 | } | |||
8273 | VMIt = std::next(VMIt); | |||
8274 | } | |||
8275 | bool IsBaseNotUndef = !IsBaseUndef.all(); | |||
8276 | (void)IsBaseNotUndef; | |||
8277 | // Perform requested actions for the remaining masks/vectors. | |||
8278 | for (auto E = ShuffleMask.end(); VMIt != E; ++VMIt) { | |||
8279 | // Shuffle other input vectors, if any. | |||
8280 | std::pair<T *, bool> Res = | |||
8281 | ResizeAction(VMIt->first, VMIt->second, /*ForSingleMask=*/false); | |||
8282 | ArrayRef<int> SecMask = VMIt->second; | |||
8283 | for (unsigned I = 0, VF = Mask.size(); I < VF; ++I) { | |||
8284 | if (SecMask[I] != PoisonMaskElem) { | |||
8285 | assert((Mask[I] == PoisonMaskElem || IsBaseNotUndef) &&(static_cast <bool> ((Mask[I] == PoisonMaskElem || IsBaseNotUndef ) && "Multiple uses of scalars.") ? void (0) : __assert_fail ("(Mask[I] == PoisonMaskElem || IsBaseNotUndef) && \"Multiple uses of scalars.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8286, __extension__ __PRETTY_FUNCTION__)) | |||
8286 | "Multiple uses of scalars.")(static_cast <bool> ((Mask[I] == PoisonMaskElem || IsBaseNotUndef ) && "Multiple uses of scalars.") ? void (0) : __assert_fail ("(Mask[I] == PoisonMaskElem || IsBaseNotUndef) && \"Multiple uses of scalars.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8286, __extension__ __PRETTY_FUNCTION__)); | |||
8287 | Mask[I] = (Res.second ? I : SecMask[I]) + VF; | |||
8288 | } else if (Mask[I] != PoisonMaskElem) { | |||
8289 | Mask[I] = I; | |||
8290 | } | |||
8291 | } | |||
8292 | Prev = Action(Mask, {Prev, Res.first}); | |||
8293 | } | |||
8294 | return Prev; | |||
8295 | } | |||
8296 | ||||
8297 | InstructionCost BoUpSLP::getTreeCost(ArrayRef<Value *> VectorizedVals) { | |||
8298 | // Build a map for gathered scalars to the nodes where they are used. | |||
8299 | ValueToGatherNodes.clear(); | |||
8300 | for (const std::unique_ptr<TreeEntry> &EntryPtr : VectorizableTree) { | |||
8301 | if (EntryPtr->State != TreeEntry::NeedToGather) | |||
8302 | continue; | |||
8303 | for (Value *V : EntryPtr->Scalars) | |||
8304 | if (!isConstant(V)) | |||
8305 | ValueToGatherNodes.try_emplace(V).first->getSecond().insert( | |||
8306 | EntryPtr.get()); | |||
8307 | } | |||
8308 | InstructionCost Cost = 0; | |||
8309 | 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) | |||
8310 | << VectorizableTree.size() << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Calculating cost for tree of size " << VectorizableTree.size() << ".\n"; } } while ( false); | |||
8311 | ||||
8312 | unsigned BundleWidth = VectorizableTree[0]->Scalars.size(); | |||
8313 | ||||
8314 | SmallPtrSet<Value *, 4> CheckedExtracts; | |||
8315 | for (unsigned I = 0, E = VectorizableTree.size(); I < E; ++I) { | |||
8316 | TreeEntry &TE = *VectorizableTree[I]; | |||
8317 | if (TE.State == TreeEntry::NeedToGather) { | |||
8318 | if (const TreeEntry *E = getTreeEntry(TE.getMainOp()); | |||
8319 | E && E->getVectorFactor() == TE.getVectorFactor() && | |||
8320 | E->isSame(TE.Scalars)) { | |||
8321 | // Some gather nodes might be absolutely the same as some vectorizable | |||
8322 | // nodes after reordering, need to handle it. | |||
8323 | LLVM_DEBUG(dbgs() << "SLP: Adding cost 0 for bundle that starts with "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Adding cost 0 for bundle that starts with " << *TE.Scalars[0] << ".\n" << "SLP: Current total cost = " << Cost << "\n"; } } while (false) | |||
8324 | << *TE.Scalars[0] << ".\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Adding cost 0 for bundle that starts with " << *TE.Scalars[0] << ".\n" << "SLP: Current total cost = " << Cost << "\n"; } } while (false) | |||
8325 | << "SLP: Current total cost = " << Cost << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Adding cost 0 for bundle that starts with " << *TE.Scalars[0] << ".\n" << "SLP: Current total cost = " << Cost << "\n"; } } while (false); | |||
8326 | continue; | |||
8327 | } | |||
8328 | } | |||
8329 | ||||
8330 | InstructionCost C = getEntryCost(&TE, VectorizedVals, CheckedExtracts); | |||
8331 | Cost += C; | |||
8332 | 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) | |||
8333 | << " 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) | |||
8334 | << ".\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) | |||
8335 | << "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); | |||
8336 | } | |||
8337 | ||||
8338 | SmallPtrSet<Value *, 16> ExtractCostCalculated; | |||
8339 | InstructionCost ExtractCost = 0; | |||
8340 | SmallVector<MapVector<const TreeEntry *, SmallVector<int>>> ShuffleMasks; | |||
8341 | SmallVector<std::pair<Value *, const TreeEntry *>> FirstUsers; | |||
8342 | SmallVector<APInt> DemandedElts; | |||
8343 | for (ExternalUser &EU : ExternalUses) { | |||
8344 | // We only add extract cost once for the same scalar. | |||
8345 | if (!isa_and_nonnull<InsertElementInst>(EU.User) && | |||
8346 | !ExtractCostCalculated.insert(EU.Scalar).second) | |||
8347 | continue; | |||
8348 | ||||
8349 | // Uses by ephemeral values are free (because the ephemeral value will be | |||
8350 | // removed prior to code generation, and so the extraction will be | |||
8351 | // removed as well). | |||
8352 | if (EphValues.count(EU.User)) | |||
8353 | continue; | |||
8354 | ||||
8355 | // No extract cost for vector "scalar" | |||
8356 | if (isa<FixedVectorType>(EU.Scalar->getType())) | |||
8357 | continue; | |||
8358 | ||||
8359 | // If found user is an insertelement, do not calculate extract cost but try | |||
8360 | // to detect it as a final shuffled/identity match. | |||
8361 | if (auto *VU = dyn_cast_or_null<InsertElementInst>(EU.User)) { | |||
8362 | if (auto *FTy = dyn_cast<FixedVectorType>(VU->getType())) { | |||
8363 | std::optional<unsigned> InsertIdx = getInsertIndex(VU); | |||
8364 | if (InsertIdx) { | |||
8365 | const TreeEntry *ScalarTE = getTreeEntry(EU.Scalar); | |||
8366 | auto *It = find_if( | |||
8367 | FirstUsers, | |||
8368 | [this, VU](const std::pair<Value *, const TreeEntry *> &Pair) { | |||
8369 | return areTwoInsertFromSameBuildVector( | |||
8370 | VU, cast<InsertElementInst>(Pair.first), | |||
8371 | [this](InsertElementInst *II) -> Value * { | |||
8372 | Value *Op0 = II->getOperand(0); | |||
8373 | if (getTreeEntry(II) && !getTreeEntry(Op0)) | |||
8374 | return nullptr; | |||
8375 | return Op0; | |||
8376 | }); | |||
8377 | }); | |||
8378 | int VecId = -1; | |||
8379 | if (It == FirstUsers.end()) { | |||
8380 | (void)ShuffleMasks.emplace_back(); | |||
8381 | SmallVectorImpl<int> &Mask = ShuffleMasks.back()[ScalarTE]; | |||
8382 | if (Mask.empty()) | |||
8383 | Mask.assign(FTy->getNumElements(), PoisonMaskElem); | |||
8384 | // Find the insertvector, vectorized in tree, if any. | |||
8385 | Value *Base = VU; | |||
8386 | while (auto *IEBase = dyn_cast<InsertElementInst>(Base)) { | |||
8387 | if (IEBase != EU.User && | |||
8388 | (!IEBase->hasOneUse() || | |||
8389 | getInsertIndex(IEBase).value_or(*InsertIdx) == *InsertIdx)) | |||
8390 | break; | |||
8391 | // Build the mask for the vectorized insertelement instructions. | |||
8392 | if (const TreeEntry *E = getTreeEntry(IEBase)) { | |||
8393 | VU = IEBase; | |||
8394 | do { | |||
8395 | IEBase = cast<InsertElementInst>(Base); | |||
8396 | int Idx = *getInsertIndex(IEBase); | |||
8397 | assert(Mask[Idx] == PoisonMaskElem &&(static_cast <bool> (Mask[Idx] == PoisonMaskElem && "InsertElementInstruction used already.") ? void (0) : __assert_fail ("Mask[Idx] == PoisonMaskElem && \"InsertElementInstruction used already.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8398, __extension__ __PRETTY_FUNCTION__)) | |||
8398 | "InsertElementInstruction used already.")(static_cast <bool> (Mask[Idx] == PoisonMaskElem && "InsertElementInstruction used already.") ? void (0) : __assert_fail ("Mask[Idx] == PoisonMaskElem && \"InsertElementInstruction used already.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8398, __extension__ __PRETTY_FUNCTION__)); | |||
8399 | Mask[Idx] = Idx; | |||
8400 | Base = IEBase->getOperand(0); | |||
8401 | } while (E == getTreeEntry(Base)); | |||
8402 | break; | |||
8403 | } | |||
8404 | Base = cast<InsertElementInst>(Base)->getOperand(0); | |||
8405 | } | |||
8406 | FirstUsers.emplace_back(VU, ScalarTE); | |||
8407 | DemandedElts.push_back(APInt::getZero(FTy->getNumElements())); | |||
8408 | VecId = FirstUsers.size() - 1; | |||
8409 | } else { | |||
8410 | if (isFirstInsertElement(VU, cast<InsertElementInst>(It->first))) | |||
8411 | It->first = VU; | |||
8412 | VecId = std::distance(FirstUsers.begin(), It); | |||
8413 | } | |||
8414 | int InIdx = *InsertIdx; | |||
8415 | SmallVectorImpl<int> &Mask = ShuffleMasks[VecId][ScalarTE]; | |||
8416 | if (Mask.empty()) | |||
8417 | Mask.assign(FTy->getNumElements(), PoisonMaskElem); | |||
8418 | Mask[InIdx] = EU.Lane; | |||
8419 | DemandedElts[VecId].setBit(InIdx); | |||
8420 | continue; | |||
8421 | } | |||
8422 | } | |||
8423 | } | |||
8424 | ||||
8425 | // If we plan to rewrite the tree in a smaller type, we will need to sign | |||
8426 | // extend the extracted value back to the original type. Here, we account | |||
8427 | // for the extract and the added cost of the sign extend if needed. | |||
8428 | auto *VecTy = FixedVectorType::get(EU.Scalar->getType(), BundleWidth); | |||
8429 | TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput; | |||
8430 | auto *ScalarRoot = VectorizableTree[0]->Scalars[0]; | |||
8431 | if (MinBWs.count(ScalarRoot)) { | |||
8432 | auto *MinTy = IntegerType::get(F->getContext(), MinBWs[ScalarRoot].first); | |||
8433 | auto Extend = | |||
8434 | MinBWs[ScalarRoot].second ? Instruction::SExt : Instruction::ZExt; | |||
8435 | VecTy = FixedVectorType::get(MinTy, BundleWidth); | |||
8436 | ExtractCost += TTI->getExtractWithExtendCost(Extend, EU.Scalar->getType(), | |||
8437 | VecTy, EU.Lane); | |||
8438 | } else { | |||
8439 | ExtractCost += TTI->getVectorInstrCost(Instruction::ExtractElement, VecTy, | |||
8440 | CostKind, EU.Lane); | |||
8441 | } | |||
8442 | } | |||
8443 | ||||
8444 | InstructionCost SpillCost = getSpillCost(); | |||
8445 | Cost += SpillCost + ExtractCost; | |||
8446 | auto &&ResizeToVF = [this, &Cost](const TreeEntry *TE, ArrayRef<int> Mask, | |||
8447 | bool) { | |||
8448 | InstructionCost C = 0; | |||
8449 | unsigned VF = Mask.size(); | |||
8450 | unsigned VecVF = TE->getVectorFactor(); | |||
8451 | if (VF != VecVF && | |||
8452 | (any_of(Mask, [VF](int Idx) { return Idx >= static_cast<int>(VF); }) || | |||
8453 | (all_of(Mask, | |||
8454 | [VF](int Idx) { return Idx < 2 * static_cast<int>(VF); }) && | |||
8455 | !ShuffleVectorInst::isIdentityMask(Mask)))) { | |||
8456 | SmallVector<int> OrigMask(VecVF, PoisonMaskElem); | |||
8457 | std::copy(Mask.begin(), std::next(Mask.begin(), std::min(VF, VecVF)), | |||
8458 | OrigMask.begin()); | |||
8459 | C = TTI->getShuffleCost( | |||
8460 | TTI::SK_PermuteSingleSrc, | |||
8461 | FixedVectorType::get(TE->getMainOp()->getType(), VecVF), OrigMask); | |||
8462 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Adding cost " << C << " for final shuffle of insertelement external users.\n"; TE-> dump(); dbgs() << "SLP: Current total cost = " << Cost << "\n"; } } while (false) | |||
8463 | dbgs() << "SLP: Adding cost " << Cdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Adding cost " << C << " for final shuffle of insertelement external users.\n"; TE-> dump(); dbgs() << "SLP: Current total cost = " << Cost << "\n"; } } while (false) | |||
8464 | << " for final shuffle of insertelement external users.\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Adding cost " << C << " for final shuffle of insertelement external users.\n"; TE-> dump(); dbgs() << "SLP: Current total cost = " << Cost << "\n"; } } while (false) | |||
8465 | TE->dump(); dbgs() << "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.\n"; TE-> dump(); dbgs() << "SLP: Current total cost = " << Cost << "\n"; } } while (false); | |||
8466 | Cost += C; | |||
8467 | return std::make_pair(TE, true); | |||
8468 | } | |||
8469 | return std::make_pair(TE, false); | |||
8470 | }; | |||
8471 | // Calculate the cost of the reshuffled vectors, if any. | |||
8472 | for (int I = 0, E = FirstUsers.size(); I < E; ++I) { | |||
8473 | Value *Base = cast<Instruction>(FirstUsers[I].first)->getOperand(0); | |||
8474 | unsigned VF = ShuffleMasks[I].begin()->second.size(); | |||
8475 | auto *FTy = FixedVectorType::get( | |||
8476 | cast<VectorType>(FirstUsers[I].first->getType())->getElementType(), VF); | |||
8477 | auto Vector = ShuffleMasks[I].takeVector(); | |||
8478 | auto &&EstimateShufflesCost = [this, FTy, | |||
8479 | &Cost](ArrayRef<int> Mask, | |||
8480 | ArrayRef<const TreeEntry *> TEs) { | |||
8481 | assert((TEs.size() == 1 || TEs.size() == 2) &&(static_cast <bool> ((TEs.size() == 1 || TEs.size() == 2 ) && "Expected exactly 1 or 2 tree entries.") ? void ( 0) : __assert_fail ("(TEs.size() == 1 || TEs.size() == 2) && \"Expected exactly 1 or 2 tree entries.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8482, __extension__ __PRETTY_FUNCTION__)) | |||
8482 | "Expected exactly 1 or 2 tree entries.")(static_cast <bool> ((TEs.size() == 1 || TEs.size() == 2 ) && "Expected exactly 1 or 2 tree entries.") ? void ( 0) : __assert_fail ("(TEs.size() == 1 || TEs.size() == 2) && \"Expected exactly 1 or 2 tree entries.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8482, __extension__ __PRETTY_FUNCTION__)); | |||
8483 | if (TEs.size() == 1) { | |||
8484 | int Limit = 2 * Mask.size(); | |||
8485 | if (!all_of(Mask, [Limit](int Idx) { return Idx < Limit; }) || | |||
8486 | !ShuffleVectorInst::isIdentityMask(Mask)) { | |||
8487 | InstructionCost C = | |||
8488 | TTI->getShuffleCost(TTI::SK_PermuteSingleSrc, FTy, Mask); | |||
8489 | 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.\n"; TEs .front()->dump(); dbgs() << "SLP: Current total cost = " << Cost << "\n"; } } while (false) | |||
8490 | << " for final shuffle of insertelement "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Adding cost " << C << " for final shuffle of insertelement " "external users.\n"; TEs .front()->dump(); dbgs() << "SLP: Current total cost = " << Cost << "\n"; } } while (false) | |||
8491 | "external users.\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Adding cost " << C << " for final shuffle of insertelement " "external users.\n"; TEs .front()->dump(); dbgs() << "SLP: Current total cost = " << Cost << "\n"; } } while (false) | |||
8492 | TEs.front()->dump();do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Adding cost " << C << " for final shuffle of insertelement " "external users.\n"; TEs .front()->dump(); dbgs() << "SLP: Current total cost = " << Cost << "\n"; } } while (false) | |||
8493 | dbgs() << "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.\n"; TEs .front()->dump(); dbgs() << "SLP: Current total cost = " << Cost << "\n"; } } while (false); | |||
8494 | Cost += C; | |||
8495 | } | |||
8496 | } else { | |||
8497 | InstructionCost C = | |||
8498 | TTI->getShuffleCost(TTI::SK_PermuteTwoSrc, FTy, Mask); | |||
8499 | 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.\n" ; if (TEs.front()) { TEs.front()->dump(); } TEs.back()-> dump(); dbgs() << "SLP: Current total cost = " << Cost << "\n"; } } while (false) | |||
8500 | << " 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.\n" ; if (TEs.front()) { TEs.front()->dump(); } TEs.back()-> dump(); dbgs() << "SLP: Current total cost = " << Cost << "\n"; } } while (false) | |||
8501 | "insertelement users.\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Adding cost " << C << " for final shuffle of vector node and external " "insertelement users.\n" ; if (TEs.front()) { TEs.front()->dump(); } TEs.back()-> dump(); dbgs() << "SLP: Current total cost = " << Cost << "\n"; } } while (false) | |||
8502 | if (TEs.front()) { TEs.front()->dump(); } TEs.back()->dump();do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Adding cost " << C << " for final shuffle of vector node and external " "insertelement users.\n" ; if (TEs.front()) { TEs.front()->dump(); } TEs.back()-> dump(); dbgs() << "SLP: Current total cost = " << Cost << "\n"; } } while (false) | |||
8503 | dbgs() << "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.\n" ; if (TEs.front()) { TEs.front()->dump(); } TEs.back()-> dump(); dbgs() << "SLP: Current total cost = " << Cost << "\n"; } } while (false); | |||
8504 | Cost += C; | |||
8505 | } | |||
8506 | return TEs.back(); | |||
8507 | }; | |||
8508 | (void)performExtractsShuffleAction<const TreeEntry>( | |||
8509 | MutableArrayRef(Vector.data(), Vector.size()), Base, | |||
8510 | [](const TreeEntry *E) { return E->getVectorFactor(); }, ResizeToVF, | |||
8511 | EstimateShufflesCost); | |||
8512 | InstructionCost InsertCost = TTI->getScalarizationOverhead( | |||
8513 | cast<FixedVectorType>(FirstUsers[I].first->getType()), DemandedElts[I], | |||
8514 | /*Insert*/ true, /*Extract*/ false, TTI::TCK_RecipThroughput); | |||
8515 | Cost -= InsertCost; | |||
8516 | } | |||
8517 | ||||
8518 | #ifndef NDEBUG | |||
8519 | SmallString<256> Str; | |||
8520 | { | |||
8521 | raw_svector_ostream OS(Str); | |||
8522 | OS << "SLP: Spill Cost = " << SpillCost << ".\n" | |||
8523 | << "SLP: Extract Cost = " << ExtractCost << ".\n" | |||
8524 | << "SLP: Total Cost = " << Cost << ".\n"; | |||
8525 | } | |||
8526 | LLVM_DEBUG(dbgs() << Str)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << Str; } } while (false); | |||
8527 | if (ViewSLPTree) | |||
8528 | ViewGraph(this, "SLP" + F->getName(), false, Str); | |||
8529 | #endif | |||
8530 | ||||
8531 | return Cost; | |||
8532 | } | |||
8533 | ||||
8534 | std::optional<TargetTransformInfo::ShuffleKind> | |||
8535 | BoUpSLP::isGatherShuffledEntry(const TreeEntry *TE, ArrayRef<Value *> VL, | |||
8536 | SmallVectorImpl<int> &Mask, | |||
8537 | SmallVectorImpl<const TreeEntry *> &Entries) { | |||
8538 | Entries.clear(); | |||
8539 | // No need to check for the topmost gather node. | |||
8540 | if (TE == VectorizableTree.front().get()) | |||
8541 | return std::nullopt; | |||
8542 | Mask.assign(VL.size(), PoisonMaskElem); | |||
8543 | assert(TE->UserTreeIndices.size() == 1 &&(static_cast <bool> (TE->UserTreeIndices.size() == 1 && "Expected only single user of the gather node.") ? void (0) : __assert_fail ("TE->UserTreeIndices.size() == 1 && \"Expected only single user of the gather node.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8544, __extension__ __PRETTY_FUNCTION__)) | |||
8544 | "Expected only single user of the gather node.")(static_cast <bool> (TE->UserTreeIndices.size() == 1 && "Expected only single user of the gather node.") ? void (0) : __assert_fail ("TE->UserTreeIndices.size() == 1 && \"Expected only single user of the gather node.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8544, __extension__ __PRETTY_FUNCTION__)); | |||
8545 | // TODO: currently checking only for Scalars in the tree entry, need to count | |||
8546 | // reused elements too for better cost estimation. | |||
8547 | Instruction &UserInst = | |||
8548 | getLastInstructionInBundle(TE->UserTreeIndices.front().UserTE); | |||
8549 | BasicBlock *ParentBB = nullptr; | |||
8550 | // Main node of PHI entries keeps the correct order of operands/incoming | |||
8551 | // blocks. | |||
8552 | if (auto *PHI = | |||
8553 | dyn_cast<PHINode>(TE->UserTreeIndices.front().UserTE->getMainOp())) { | |||
8554 | ParentBB = PHI->getIncomingBlock(TE->UserTreeIndices.front().EdgeIdx); | |||
8555 | } else { | |||
8556 | ParentBB = UserInst.getParent(); | |||
8557 | } | |||
8558 | auto *NodeUI = DT->getNode(ParentBB); | |||
8559 | assert(NodeUI && "Should only process reachable instructions")(static_cast <bool> (NodeUI && "Should only process reachable instructions" ) ? void (0) : __assert_fail ("NodeUI && \"Should only process reachable instructions\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8559, __extension__ __PRETTY_FUNCTION__)); | |||
8560 | SmallPtrSet<Value *, 4> GatheredScalars(VL.begin(), VL.end()); | |||
8561 | auto CheckOrdering = [&](Instruction *LastEI) { | |||
8562 | // Check if the user node of the TE comes after user node of EntryPtr, | |||
8563 | // otherwise EntryPtr depends on TE. | |||
8564 | // Gather nodes usually are not scheduled and inserted before their first | |||
8565 | // user node. So, instead of checking dependency between the gather nodes | |||
8566 | // themselves, we check the dependency between their user nodes. | |||
8567 | // If one user node comes before the second one, we cannot use the second | |||
8568 | // gather node as the source vector for the first gather node, because in | |||
8569 | // the list of instructions it will be emitted later. | |||
8570 | auto *EntryParent = LastEI->getParent(); | |||
8571 | auto *NodeEUI = DT->getNode(EntryParent); | |||
8572 | if (!NodeEUI) | |||
8573 | return false; | |||
8574 | assert((NodeUI == NodeEUI) ==(static_cast <bool> ((NodeUI == NodeEUI) == (NodeUI-> getDFSNumIn() == NodeEUI->getDFSNumIn()) && "Different nodes should have different DFS numbers" ) ? void (0) : __assert_fail ("(NodeUI == NodeEUI) == (NodeUI->getDFSNumIn() == NodeEUI->getDFSNumIn()) && \"Different nodes should have different DFS numbers\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8576, __extension__ __PRETTY_FUNCTION__)) | |||
8575 | (NodeUI->getDFSNumIn() == NodeEUI->getDFSNumIn()) &&(static_cast <bool> ((NodeUI == NodeEUI) == (NodeUI-> getDFSNumIn() == NodeEUI->getDFSNumIn()) && "Different nodes should have different DFS numbers" ) ? void (0) : __assert_fail ("(NodeUI == NodeEUI) == (NodeUI->getDFSNumIn() == NodeEUI->getDFSNumIn()) && \"Different nodes should have different DFS numbers\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8576, __extension__ __PRETTY_FUNCTION__)) | |||
8576 | "Different nodes should have different DFS numbers")(static_cast <bool> ((NodeUI == NodeEUI) == (NodeUI-> getDFSNumIn() == NodeEUI->getDFSNumIn()) && "Different nodes should have different DFS numbers" ) ? void (0) : __assert_fail ("(NodeUI == NodeEUI) == (NodeUI->getDFSNumIn() == NodeEUI->getDFSNumIn()) && \"Different nodes should have different DFS numbers\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8576, __extension__ __PRETTY_FUNCTION__)); | |||
8577 | // Check the order of the gather nodes users. | |||
8578 | if (UserInst.getParent() != EntryParent && | |||
8579 | (DT->dominates(NodeUI, NodeEUI) || !DT->dominates(NodeEUI, NodeUI))) | |||
8580 | return false; | |||
8581 | if (UserInst.getParent() == EntryParent && UserInst.comesBefore(LastEI)) | |||
8582 | return false; | |||
8583 | return true; | |||
8584 | }; | |||
8585 | // Find all tree entries used by the gathered values. If no common entries | |||
8586 | // found - not a shuffle. | |||
8587 | // Here we build a set of tree nodes for each gathered value and trying to | |||
8588 | // find the intersection between these sets. If we have at least one common | |||
8589 | // tree node for each gathered value - we have just a permutation of the | |||
8590 | // single vector. If we have 2 different sets, we're in situation where we | |||
8591 | // have a permutation of 2 input vectors. | |||
8592 | SmallVector<SmallPtrSet<const TreeEntry *, 4>> UsedTEs; | |||
8593 | DenseMap<Value *, int> UsedValuesEntry; | |||
8594 | for (Value *V : VL) { | |||
8595 | if (isConstant(V)) | |||
8596 | continue; | |||
8597 | // Build a list of tree entries where V is used. | |||
8598 | SmallPtrSet<const TreeEntry *, 4> VToTEs; | |||
8599 | for (const TreeEntry *TEPtr : ValueToGatherNodes.find(V)->second) { | |||
8600 | if (TEPtr == TE) | |||
8601 | continue; | |||
8602 | if (!any_of(TEPtr->Scalars, [&GatheredScalars](Value *V) { | |||
8603 | return GatheredScalars.contains(V); | |||
8604 | })) | |||
8605 | continue; | |||
8606 | assert(TEPtr->UserTreeIndices.size() == 1 &&(static_cast <bool> (TEPtr->UserTreeIndices.size() == 1 && "Expected only single user of the gather node." ) ? void (0) : __assert_fail ("TEPtr->UserTreeIndices.size() == 1 && \"Expected only single user of the gather node.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8607, __extension__ __PRETTY_FUNCTION__)) | |||
8607 | "Expected only single user of the gather node.")(static_cast <bool> (TEPtr->UserTreeIndices.size() == 1 && "Expected only single user of the gather node." ) ? void (0) : __assert_fail ("TEPtr->UserTreeIndices.size() == 1 && \"Expected only single user of the gather node.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8607, __extension__ __PRETTY_FUNCTION__)); | |||
8608 | Instruction &EntryUserInst = | |||
8609 | getLastInstructionInBundle(TEPtr->UserTreeIndices.front().UserTE); | |||
8610 | PHINode *EntryPHI = | |||
8611 | dyn_cast<PHINode>(TEPtr->UserTreeIndices.front().UserTE->getMainOp()); | |||
8612 | if (&UserInst == &EntryUserInst && !EntryPHI) { | |||
8613 | // If 2 gathers are operands of the same entry, compare operands | |||
8614 | // indices, use the earlier one as the base. | |||
8615 | if (TE->UserTreeIndices.front().UserTE == | |||
8616 | TEPtr->UserTreeIndices.front().UserTE && | |||
8617 | TE->UserTreeIndices.front().EdgeIdx < | |||
8618 | TEPtr->UserTreeIndices.front().EdgeIdx) | |||
8619 | continue; | |||
8620 | } | |||
8621 | // Check if the user node of the TE comes after user node of EntryPtr, | |||
8622 | // otherwise EntryPtr depends on TE. | |||
8623 | auto *EntryI = EntryPHI | |||
8624 | ? EntryPHI | |||
8625 | ->getIncomingBlock( | |||
8626 | TEPtr->UserTreeIndices.front().EdgeIdx) | |||
8627 | ->getTerminator() | |||
8628 | : &EntryUserInst; | |||
8629 | if (!CheckOrdering(EntryI) && | |||
8630 | (ParentBB != EntryI->getParent() || | |||
8631 | TE->UserTreeIndices.front().UserTE != | |||
8632 | TEPtr->UserTreeIndices.front().UserTE || | |||
8633 | TE->UserTreeIndices.front().EdgeIdx < | |||
8634 | TEPtr->UserTreeIndices.front().EdgeIdx)) | |||
8635 | continue; | |||
8636 | VToTEs.insert(TEPtr); | |||
8637 | } | |||
8638 | if (const TreeEntry *VTE = getTreeEntry(V)) { | |||
8639 | Instruction &EntryUserInst = getLastInstructionInBundle(VTE); | |||
8640 | if (&EntryUserInst == &UserInst || !CheckOrdering(&EntryUserInst)) | |||
8641 | continue; | |||
8642 | VToTEs.insert(VTE); | |||
8643 | } | |||
8644 | if (VToTEs.empty()) | |||
8645 | continue; | |||
8646 | if (UsedTEs.empty()) { | |||
8647 | // The first iteration, just insert the list of nodes to vector. | |||
8648 | UsedTEs.push_back(VToTEs); | |||
8649 | UsedValuesEntry.try_emplace(V, 0); | |||
8650 | } else { | |||
8651 | // Need to check if there are any previously used tree nodes which use V. | |||
8652 | // If there are no such nodes, consider that we have another one input | |||
8653 | // vector. | |||
8654 | SmallPtrSet<const TreeEntry *, 4> SavedVToTEs(VToTEs); | |||
8655 | unsigned Idx = 0; | |||
8656 | for (SmallPtrSet<const TreeEntry *, 4> &Set : UsedTEs) { | |||
8657 | // Do we have a non-empty intersection of previously listed tree entries | |||
8658 | // and tree entries using current V? | |||
8659 | set_intersect(VToTEs, Set); | |||
8660 | if (!VToTEs.empty()) { | |||
8661 | // Yes, write the new subset and continue analysis for the next | |||
8662 | // scalar. | |||
8663 | Set.swap(VToTEs); | |||
8664 | break; | |||
8665 | } | |||
8666 | VToTEs = SavedVToTEs; | |||
8667 | ++Idx; | |||
8668 | } | |||
8669 | // No non-empty intersection found - need to add a second set of possible | |||
8670 | // source vectors. | |||
8671 | if (Idx == UsedTEs.size()) { | |||
8672 | // If the number of input vectors is greater than 2 - not a permutation, | |||
8673 | // fallback to the regular gather. | |||
8674 | // TODO: support multiple reshuffled nodes. | |||
8675 | if (UsedTEs.size() == 2) | |||
8676 | continue; | |||
8677 | UsedTEs.push_back(SavedVToTEs); | |||
8678 | Idx = UsedTEs.size() - 1; | |||
8679 | } | |||
8680 | UsedValuesEntry.try_emplace(V, Idx); | |||
8681 | } | |||
8682 | } | |||
8683 | ||||
8684 | if (UsedTEs.empty()) | |||
8685 | return std::nullopt; | |||
8686 | ||||
8687 | unsigned VF = 0; | |||
8688 | if (UsedTEs.size() == 1) { | |||
8689 | // Keep the order to avoid non-determinism. | |||
8690 | SmallVector<const TreeEntry *> FirstEntries(UsedTEs.front().begin(), | |||
8691 | UsedTEs.front().end()); | |||
8692 | sort(FirstEntries, [](const TreeEntry *TE1, const TreeEntry *TE2) { | |||
8693 | return TE1->Idx < TE2->Idx; | |||
8694 | }); | |||
8695 | // Try to find the perfect match in another gather node at first. | |||
8696 | auto *It = find_if(FirstEntries, [=](const TreeEntry *EntryPtr) { | |||
8697 | return EntryPtr->isSame(VL) || EntryPtr->isSame(TE->Scalars); | |||
8698 | }); | |||
8699 | if (It != FirstEntries.end() && (*It)->getVectorFactor() == VL.size()) { | |||
8700 | Entries.push_back(*It); | |||
8701 | std::iota(Mask.begin(), Mask.end(), 0); | |||
8702 | // Clear undef scalars. | |||
8703 | for (int I = 0, Sz = VL.size(); I < Sz; ++I) | |||
8704 | if (isa<PoisonValue>(VL[I])) | |||
8705 | Mask[I] = PoisonMaskElem; | |||
8706 | return TargetTransformInfo::SK_PermuteSingleSrc; | |||
8707 | } | |||
8708 | // No perfect match, just shuffle, so choose the first tree node from the | |||
8709 | // tree. | |||
8710 | Entries.push_back(FirstEntries.front()); | |||
8711 | } else { | |||
8712 | // Try to find nodes with the same vector factor. | |||
8713 | 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", 8713, __extension__ __PRETTY_FUNCTION__)); | |||
8714 | // Keep the order of tree nodes to avoid non-determinism. | |||
8715 | DenseMap<int, const TreeEntry *> VFToTE; | |||
8716 | for (const TreeEntry *TE : UsedTEs.front()) { | |||
8717 | unsigned VF = TE->getVectorFactor(); | |||
8718 | auto It = VFToTE.find(VF); | |||
8719 | if (It != VFToTE.end()) { | |||
8720 | if (It->second->Idx > TE->Idx) | |||
8721 | It->getSecond() = TE; | |||
8722 | continue; | |||
8723 | } | |||
8724 | VFToTE.try_emplace(VF, TE); | |||
8725 | } | |||
8726 | // Same, keep the order to avoid non-determinism. | |||
8727 | SmallVector<const TreeEntry *> SecondEntries(UsedTEs.back().begin(), | |||
8728 | UsedTEs.back().end()); | |||
8729 | sort(SecondEntries, [](const TreeEntry *TE1, const TreeEntry *TE2) { | |||
8730 | return TE1->Idx < TE2->Idx; | |||
8731 | }); | |||
8732 | for (const TreeEntry *TE : SecondEntries) { | |||
8733 | auto It = VFToTE.find(TE->getVectorFactor()); | |||
8734 | if (It != VFToTE.end()) { | |||
8735 | VF = It->first; | |||
8736 | Entries.push_back(It->second); | |||
8737 | Entries.push_back(TE); | |||
8738 | break; | |||
8739 | } | |||
8740 | } | |||
8741 | // No 2 source vectors with the same vector factor - just choose 2 with max | |||
8742 | // index. | |||
8743 | if (Entries.empty()) { | |||
8744 | Entries.push_back( | |||
8745 | *std::max_element(UsedTEs.front().begin(), UsedTEs.front().end(), | |||
8746 | [](const TreeEntry *TE1, const TreeEntry *TE2) { | |||
8747 | return TE1->Idx < TE2->Idx; | |||
8748 | })); | |||
8749 | Entries.push_back(SecondEntries.front()); | |||
8750 | VF = std::max(Entries.front()->getVectorFactor(), | |||
8751 | Entries.back()->getVectorFactor()); | |||
8752 | } | |||
8753 | } | |||
8754 | ||||
8755 | bool IsSplatOrUndefs = isSplat(VL) || all_of(VL, UndefValue::classof); | |||
8756 | // Checks if the 2 PHIs are compatible in terms of high possibility to be | |||
8757 | // vectorized. | |||
8758 | auto AreCompatiblePHIs = [&](Value *V, Value *V1) { | |||
8759 | auto *PHI = cast<PHINode>(V); | |||
8760 | auto *PHI1 = cast<PHINode>(V1); | |||
8761 | // Check that all incoming values are compatible/from same parent (if they | |||
8762 | // are instructions). | |||
8763 | // The incoming values are compatible if they all are constants, or | |||
8764 | // instruction with the same/alternate opcodes from the same basic block. | |||
8765 | for (int I = 0, E = PHI->getNumIncomingValues(); I < E; ++I) { | |||
8766 | Value *In = PHI->getIncomingValue(I); | |||
8767 | Value *In1 = PHI1->getIncomingValue(I); | |||
8768 | if (isConstant(In) && isConstant(In1)) | |||
8769 | continue; | |||
8770 | if (!getSameOpcode({In, In1}, *TLI).getOpcode()) | |||
8771 | return false; | |||
8772 | if (cast<Instruction>(In)->getParent() != | |||
8773 | cast<Instruction>(In1)->getParent()) | |||
8774 | return false; | |||
8775 | } | |||
8776 | return true; | |||
8777 | }; | |||
8778 | // Check if the value can be ignored during analysis for shuffled gathers. | |||
8779 | // We suppose it is better to ignore instruction, which do not form splats, | |||
8780 | // are not vectorized/not extractelements (these instructions will be handled | |||
8781 | // by extractelements processing) or may form vector node in future. | |||
8782 | auto MightBeIgnored = [=](Value *V) { | |||
8783 | auto *I = dyn_cast<Instruction>(V); | |||
8784 | SmallVector<Value *> IgnoredVals; | |||
8785 | if (UserIgnoreList) | |||
8786 | IgnoredVals.assign(UserIgnoreList->begin(), UserIgnoreList->end()); | |||
8787 | return I && !IsSplatOrUndefs && !ScalarToTreeEntry.count(I) && | |||
8788 | !isVectorLikeInstWithConstOps(I) && | |||
8789 | !areAllUsersVectorized(I, IgnoredVals) && isSimple(I); | |||
8790 | }; | |||
8791 | // Check that the neighbor instruction may form a full vector node with the | |||
8792 | // current instruction V. It is possible, if they have same/alternate opcode | |||
8793 | // and same parent basic block. | |||
8794 | auto NeighborMightBeIgnored = [&](Value *V, int Idx) { | |||
8795 | Value *V1 = VL[Idx]; | |||
8796 | bool UsedInSameVTE = false; | |||
8797 | auto It = UsedValuesEntry.find(V1); | |||
8798 | if (It != UsedValuesEntry.end()) | |||
8799 | UsedInSameVTE = It->second == UsedValuesEntry.find(V)->second; | |||
8800 | return V != V1 && MightBeIgnored(V1) && !UsedInSameVTE && | |||
8801 | getSameOpcode({V, V1}, *TLI).getOpcode() && | |||
8802 | cast<Instruction>(V)->getParent() == | |||
8803 | cast<Instruction>(V1)->getParent() && | |||
8804 | (!isa<PHINode>(V1) || AreCompatiblePHIs(V, V1)); | |||
8805 | }; | |||
8806 | // Build a shuffle mask for better cost estimation and vector emission. | |||
8807 | SmallBitVector UsedIdxs(Entries.size()); | |||
8808 | SmallVector<std::pair<unsigned, int>> EntryLanes; | |||
8809 | for (int I = 0, E = VL.size(); I < E; ++I) { | |||
8810 | Value *V = VL[I]; | |||
8811 | auto It = UsedValuesEntry.find(V); | |||
8812 | if (It == UsedValuesEntry.end()) | |||
8813 | continue; | |||
8814 | // Do not try to shuffle scalars, if they are constants, or instructions | |||
8815 | // that can be vectorized as a result of the following vector build | |||
8816 | // vectorization. | |||
8817 | if (isConstant(V) || (MightBeIgnored(V) && | |||
8818 | ((I > 0 && NeighborMightBeIgnored(V, I - 1)) || | |||
8819 | (I != E - 1 && NeighborMightBeIgnored(V, I + 1))))) | |||
8820 | continue; | |||
8821 | unsigned Idx = It->second; | |||
8822 | EntryLanes.emplace_back(Idx, I); | |||
8823 | UsedIdxs.set(Idx); | |||
8824 | } | |||
8825 | // Iterate through all shuffled scalars and select entries, which can be used | |||
8826 | // for final shuffle. | |||
8827 | SmallVector<const TreeEntry *> TempEntries; | |||
8828 | for (unsigned I = 0, Sz = Entries.size(); I < Sz; ++I) { | |||
8829 | if (!UsedIdxs.test(I)) | |||
8830 | continue; | |||
8831 | // Fix the entry number for the given scalar. If it is the first entry, set | |||
8832 | // Pair.first to 0, otherwise to 1 (currently select at max 2 nodes). | |||
8833 | // These indices are used when calculating final shuffle mask as the vector | |||
8834 | // offset. | |||
8835 | for (std::pair<unsigned, int> &Pair : EntryLanes) | |||
8836 | if (Pair.first == I) | |||
8837 | Pair.first = TempEntries.size(); | |||
8838 | TempEntries.push_back(Entries[I]); | |||
8839 | } | |||
8840 | Entries.swap(TempEntries); | |||
8841 | if (EntryLanes.size() == Entries.size() && !VL.equals(TE->Scalars)) { | |||
8842 | // We may have here 1 or 2 entries only. If the number of scalars is equal | |||
8843 | // to the number of entries, no need to do the analysis, it is not very | |||
8844 | // profitable. Since VL is not the same as TE->Scalars, it means we already | |||
8845 | // have some shuffles before. Cut off not profitable case. | |||
8846 | Entries.clear(); | |||
8847 | return std::nullopt; | |||
8848 | } | |||
8849 | // Build the final mask, check for the identity shuffle, if possible. | |||
8850 | bool IsIdentity = Entries.size() == 1; | |||
8851 | // Pair.first is the offset to the vector, while Pair.second is the index of | |||
8852 | // scalar in the list. | |||
8853 | for (const std::pair<unsigned, int> &Pair : EntryLanes) { | |||
8854 | Mask[Pair.second] = Pair.first * VF + | |||
8855 | Entries[Pair.first]->findLaneForValue(VL[Pair.second]); | |||
8856 | IsIdentity &= Mask[Pair.second] == Pair.second; | |||
8857 | } | |||
8858 | switch (Entries.size()) { | |||
8859 | case 1: | |||
8860 | if (IsIdentity || EntryLanes.size() > 1 || VL.size() <= 2) | |||
8861 | return TargetTransformInfo::SK_PermuteSingleSrc; | |||
8862 | break; | |||
8863 | case 2: | |||
8864 | if (EntryLanes.size() > 2 || VL.size() <= 2) | |||
8865 | return TargetTransformInfo::SK_PermuteTwoSrc; | |||
8866 | break; | |||
8867 | default: | |||
8868 | break; | |||
8869 | } | |||
8870 | Entries.clear(); | |||
8871 | return std::nullopt; | |||
8872 | } | |||
8873 | ||||
8874 | InstructionCost BoUpSLP::getGatherCost(ArrayRef<Value *> VL, | |||
8875 | bool ForPoisonSrc) const { | |||
8876 | // Find the type of the operands in VL. | |||
8877 | Type *ScalarTy = VL[0]->getType(); | |||
8878 | if (StoreInst *SI = dyn_cast<StoreInst>(VL[0])) | |||
8879 | ScalarTy = SI->getValueOperand()->getType(); | |||
8880 | auto *VecTy = FixedVectorType::get(ScalarTy, VL.size()); | |||
8881 | bool DuplicateNonConst = false; | |||
8882 | // Find the cost of inserting/extracting values from the vector. | |||
8883 | // Check if the same elements are inserted several times and count them as | |||
8884 | // shuffle candidates. | |||
8885 | APInt ShuffledElements = APInt::getZero(VL.size()); | |||
8886 | DenseSet<Value *> UniqueElements; | |||
8887 | constexpr TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput; | |||
8888 | InstructionCost Cost; | |||
8889 | auto EstimateInsertCost = [&](unsigned I, Value *V) { | |||
8890 | if (!ForPoisonSrc) | |||
8891 | Cost += | |||
8892 | TTI->getVectorInstrCost(Instruction::InsertElement, VecTy, CostKind, | |||
8893 | I, Constant::getNullValue(VecTy), V); | |||
8894 | }; | |||
8895 | for (unsigned I = 0, E = VL.size(); I < E; ++I) { | |||
8896 | Value *V = VL[I]; | |||
8897 | // No need to shuffle duplicates for constants. | |||
8898 | if ((ForPoisonSrc && isConstant(V)) || isa<UndefValue>(V)) { | |||
8899 | ShuffledElements.setBit(I); | |||
8900 | continue; | |||
8901 | } | |||
8902 | if (!UniqueElements.insert(V).second) { | |||
8903 | DuplicateNonConst = true; | |||
8904 | ShuffledElements.setBit(I); | |||
8905 | continue; | |||
8906 | } | |||
8907 | EstimateInsertCost(I, V); | |||
8908 | } | |||
8909 | if (ForPoisonSrc) | |||
8910 | Cost = | |||
8911 | TTI->getScalarizationOverhead(VecTy, ~ShuffledElements, /*Insert*/ true, | |||
8912 | /*Extract*/ false, CostKind); | |||
8913 | if (DuplicateNonConst) | |||
8914 | Cost += | |||
8915 | TTI->getShuffleCost(TargetTransformInfo::SK_PermuteSingleSrc, VecTy); | |||
8916 | return Cost; | |||
8917 | } | |||
8918 | ||||
8919 | // Perform operand reordering on the instructions in VL and return the reordered | |||
8920 | // operands in Left and Right. | |||
8921 | void BoUpSLP::reorderInputsAccordingToOpcode( | |||
8922 | ArrayRef<Value *> VL, SmallVectorImpl<Value *> &Left, | |||
8923 | SmallVectorImpl<Value *> &Right, const TargetLibraryInfo &TLI, | |||
8924 | const DataLayout &DL, ScalarEvolution &SE, const BoUpSLP &R) { | |||
8925 | if (VL.empty()) | |||
8926 | return; | |||
8927 | VLOperands Ops(VL, TLI, DL, SE, R); | |||
8928 | // Reorder the operands in place. | |||
8929 | Ops.reorder(); | |||
8930 | Left = Ops.getVL(0); | |||
8931 | Right = Ops.getVL(1); | |||
8932 | } | |||
8933 | ||||
8934 | Instruction &BoUpSLP::getLastInstructionInBundle(const TreeEntry *E) { | |||
8935 | // Get the basic block this bundle is in. All instructions in the bundle | |||
8936 | // should be in this block (except for extractelement-like instructions with | |||
8937 | // constant indeces). | |||
8938 | auto *Front = E->getMainOp(); | |||
8939 | auto *BB = Front->getParent(); | |||
8940 | assert(llvm::all_of(E->Scalars, [=](Value *V) -> bool {(static_cast <bool> (llvm::all_of(E->Scalars, [=](Value *V) -> bool { if (E->getOpcode() == Instruction::GetElementPtr && !isa<GetElementPtrInst>(V)) return true; auto *I = cast<Instruction>(V); return !E->isOpcodeOrAlt (I) || I->getParent() == BB || isVectorLikeInstWithConstOps (I); })) ? void (0) : __assert_fail ("llvm::all_of(E->Scalars, [=](Value *V) -> bool { if (E->getOpcode() == Instruction::GetElementPtr && !isa<GetElementPtrInst>(V)) return true; auto *I = cast<Instruction>(V); return !E->isOpcodeOrAlt(I) || I->getParent() == BB || isVectorLikeInstWithConstOps(I); })" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8947, __extension__ __PRETTY_FUNCTION__)) | |||
8941 | if (E->getOpcode() == Instruction::GetElementPtr &&(static_cast <bool> (llvm::all_of(E->Scalars, [=](Value *V) -> bool { if (E->getOpcode() == Instruction::GetElementPtr && !isa<GetElementPtrInst>(V)) return true; auto *I = cast<Instruction>(V); return !E->isOpcodeOrAlt (I) || I->getParent() == BB || isVectorLikeInstWithConstOps (I); })) ? void (0) : __assert_fail ("llvm::all_of(E->Scalars, [=](Value *V) -> bool { if (E->getOpcode() == Instruction::GetElementPtr && !isa<GetElementPtrInst>(V)) return true; auto *I = cast<Instruction>(V); return !E->isOpcodeOrAlt(I) || I->getParent() == BB || isVectorLikeInstWithConstOps(I); })" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8947, __extension__ __PRETTY_FUNCTION__)) | |||
8942 | !isa<GetElementPtrInst>(V))(static_cast <bool> (llvm::all_of(E->Scalars, [=](Value *V) -> bool { if (E->getOpcode() == Instruction::GetElementPtr && !isa<GetElementPtrInst>(V)) return true; auto *I = cast<Instruction>(V); return !E->isOpcodeOrAlt (I) || I->getParent() == BB || isVectorLikeInstWithConstOps (I); })) ? void (0) : __assert_fail ("llvm::all_of(E->Scalars, [=](Value *V) -> bool { if (E->getOpcode() == Instruction::GetElementPtr && !isa<GetElementPtrInst>(V)) return true; auto *I = cast<Instruction>(V); return !E->isOpcodeOrAlt(I) || I->getParent() == BB || isVectorLikeInstWithConstOps(I); })" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8947, __extension__ __PRETTY_FUNCTION__)) | |||
8943 | return true;(static_cast <bool> (llvm::all_of(E->Scalars, [=](Value *V) -> bool { if (E->getOpcode() == Instruction::GetElementPtr && !isa<GetElementPtrInst>(V)) return true; auto *I = cast<Instruction>(V); return !E->isOpcodeOrAlt (I) || I->getParent() == BB || isVectorLikeInstWithConstOps (I); })) ? void (0) : __assert_fail ("llvm::all_of(E->Scalars, [=](Value *V) -> bool { if (E->getOpcode() == Instruction::GetElementPtr && !isa<GetElementPtrInst>(V)) return true; auto *I = cast<Instruction>(V); return !E->isOpcodeOrAlt(I) || I->getParent() == BB || isVectorLikeInstWithConstOps(I); })" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8947, __extension__ __PRETTY_FUNCTION__)) | |||
8944 | auto *I = cast<Instruction>(V);(static_cast <bool> (llvm::all_of(E->Scalars, [=](Value *V) -> bool { if (E->getOpcode() == Instruction::GetElementPtr && !isa<GetElementPtrInst>(V)) return true; auto *I = cast<Instruction>(V); return !E->isOpcodeOrAlt (I) || I->getParent() == BB || isVectorLikeInstWithConstOps (I); })) ? void (0) : __assert_fail ("llvm::all_of(E->Scalars, [=](Value *V) -> bool { if (E->getOpcode() == Instruction::GetElementPtr && !isa<GetElementPtrInst>(V)) return true; auto *I = cast<Instruction>(V); return !E->isOpcodeOrAlt(I) || I->getParent() == BB || isVectorLikeInstWithConstOps(I); })" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8947, __extension__ __PRETTY_FUNCTION__)) | |||
8945 | return !E->isOpcodeOrAlt(I) || I->getParent() == BB ||(static_cast <bool> (llvm::all_of(E->Scalars, [=](Value *V) -> bool { if (E->getOpcode() == Instruction::GetElementPtr && !isa<GetElementPtrInst>(V)) return true; auto *I = cast<Instruction>(V); return !E->isOpcodeOrAlt (I) || I->getParent() == BB || isVectorLikeInstWithConstOps (I); })) ? void (0) : __assert_fail ("llvm::all_of(E->Scalars, [=](Value *V) -> bool { if (E->getOpcode() == Instruction::GetElementPtr && !isa<GetElementPtrInst>(V)) return true; auto *I = cast<Instruction>(V); return !E->isOpcodeOrAlt(I) || I->getParent() == BB || isVectorLikeInstWithConstOps(I); })" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8947, __extension__ __PRETTY_FUNCTION__)) | |||
8946 | isVectorLikeInstWithConstOps(I);(static_cast <bool> (llvm::all_of(E->Scalars, [=](Value *V) -> bool { if (E->getOpcode() == Instruction::GetElementPtr && !isa<GetElementPtrInst>(V)) return true; auto *I = cast<Instruction>(V); return !E->isOpcodeOrAlt (I) || I->getParent() == BB || isVectorLikeInstWithConstOps (I); })) ? void (0) : __assert_fail ("llvm::all_of(E->Scalars, [=](Value *V) -> bool { if (E->getOpcode() == Instruction::GetElementPtr && !isa<GetElementPtrInst>(V)) return true; auto *I = cast<Instruction>(V); return !E->isOpcodeOrAlt(I) || I->getParent() == BB || isVectorLikeInstWithConstOps(I); })" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8947, __extension__ __PRETTY_FUNCTION__)) | |||
8947 | }))(static_cast <bool> (llvm::all_of(E->Scalars, [=](Value *V) -> bool { if (E->getOpcode() == Instruction::GetElementPtr && !isa<GetElementPtrInst>(V)) return true; auto *I = cast<Instruction>(V); return !E->isOpcodeOrAlt (I) || I->getParent() == BB || isVectorLikeInstWithConstOps (I); })) ? void (0) : __assert_fail ("llvm::all_of(E->Scalars, [=](Value *V) -> bool { if (E->getOpcode() == Instruction::GetElementPtr && !isa<GetElementPtrInst>(V)) return true; auto *I = cast<Instruction>(V); return !E->isOpcodeOrAlt(I) || I->getParent() == BB || isVectorLikeInstWithConstOps(I); })" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8947, __extension__ __PRETTY_FUNCTION__)); | |||
8948 | ||||
8949 | auto &&FindLastInst = [E, Front, this, &BB]() { | |||
8950 | Instruction *LastInst = Front; | |||
8951 | for (Value *V : E->Scalars) { | |||
8952 | auto *I = dyn_cast<Instruction>(V); | |||
8953 | if (!I) | |||
8954 | continue; | |||
8955 | if (LastInst->getParent() == I->getParent()) { | |||
8956 | if (LastInst->comesBefore(I)) | |||
8957 | LastInst = I; | |||
8958 | continue; | |||
8959 | } | |||
8960 | assert(((E->getOpcode() == Instruction::GetElementPtr &&(static_cast <bool> (((E->getOpcode() == Instruction ::GetElementPtr && !isa<GetElementPtrInst>(I)) || (isVectorLikeInstWithConstOps(LastInst) && isVectorLikeInstWithConstOps (I))) && "Expected vector-like or non-GEP in GEP node insts only." ) ? void (0) : __assert_fail ("((E->getOpcode() == Instruction::GetElementPtr && !isa<GetElementPtrInst>(I)) || (isVectorLikeInstWithConstOps(LastInst) && isVectorLikeInstWithConstOps(I))) && \"Expected vector-like or non-GEP in GEP node insts only.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8964, __extension__ __PRETTY_FUNCTION__)) | |||
8961 | !isa<GetElementPtrInst>(I)) ||(static_cast <bool> (((E->getOpcode() == Instruction ::GetElementPtr && !isa<GetElementPtrInst>(I)) || (isVectorLikeInstWithConstOps(LastInst) && isVectorLikeInstWithConstOps (I))) && "Expected vector-like or non-GEP in GEP node insts only." ) ? void (0) : __assert_fail ("((E->getOpcode() == Instruction::GetElementPtr && !isa<GetElementPtrInst>(I)) || (isVectorLikeInstWithConstOps(LastInst) && isVectorLikeInstWithConstOps(I))) && \"Expected vector-like or non-GEP in GEP node insts only.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8964, __extension__ __PRETTY_FUNCTION__)) | |||
8962 | (isVectorLikeInstWithConstOps(LastInst) &&(static_cast <bool> (((E->getOpcode() == Instruction ::GetElementPtr && !isa<GetElementPtrInst>(I)) || (isVectorLikeInstWithConstOps(LastInst) && isVectorLikeInstWithConstOps (I))) && "Expected vector-like or non-GEP in GEP node insts only." ) ? void (0) : __assert_fail ("((E->getOpcode() == Instruction::GetElementPtr && !isa<GetElementPtrInst>(I)) || (isVectorLikeInstWithConstOps(LastInst) && isVectorLikeInstWithConstOps(I))) && \"Expected vector-like or non-GEP in GEP node insts only.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8964, __extension__ __PRETTY_FUNCTION__)) | |||
8963 | isVectorLikeInstWithConstOps(I))) &&(static_cast <bool> (((E->getOpcode() == Instruction ::GetElementPtr && !isa<GetElementPtrInst>(I)) || (isVectorLikeInstWithConstOps(LastInst) && isVectorLikeInstWithConstOps (I))) && "Expected vector-like or non-GEP in GEP node insts only." ) ? void (0) : __assert_fail ("((E->getOpcode() == Instruction::GetElementPtr && !isa<GetElementPtrInst>(I)) || (isVectorLikeInstWithConstOps(LastInst) && isVectorLikeInstWithConstOps(I))) && \"Expected vector-like or non-GEP in GEP node insts only.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8964, __extension__ __PRETTY_FUNCTION__)) | |||
8964 | "Expected vector-like or non-GEP in GEP node insts only.")(static_cast <bool> (((E->getOpcode() == Instruction ::GetElementPtr && !isa<GetElementPtrInst>(I)) || (isVectorLikeInstWithConstOps(LastInst) && isVectorLikeInstWithConstOps (I))) && "Expected vector-like or non-GEP in GEP node insts only." ) ? void (0) : __assert_fail ("((E->getOpcode() == Instruction::GetElementPtr && !isa<GetElementPtrInst>(I)) || (isVectorLikeInstWithConstOps(LastInst) && isVectorLikeInstWithConstOps(I))) && \"Expected vector-like or non-GEP in GEP node insts only.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8964, __extension__ __PRETTY_FUNCTION__)); | |||
8965 | if (!DT->isReachableFromEntry(LastInst->getParent())) { | |||
8966 | LastInst = I; | |||
8967 | continue; | |||
8968 | } | |||
8969 | if (!DT->isReachableFromEntry(I->getParent())) | |||
8970 | continue; | |||
8971 | auto *NodeA = DT->getNode(LastInst->getParent()); | |||
8972 | auto *NodeB = DT->getNode(I->getParent()); | |||
8973 | 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", 8973, __extension__ __PRETTY_FUNCTION__)); | |||
8974 | 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", 8974, __extension__ __PRETTY_FUNCTION__)); | |||
8975 | assert((NodeA == NodeB) ==(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", 8977, __extension__ __PRETTY_FUNCTION__)) | |||
8976 | (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", 8977, __extension__ __PRETTY_FUNCTION__)) | |||
8977 | "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", 8977, __extension__ __PRETTY_FUNCTION__)); | |||
8978 | if (NodeA->getDFSNumIn() < NodeB->getDFSNumIn()) | |||
8979 | LastInst = I; | |||
8980 | } | |||
8981 | BB = LastInst->getParent(); | |||
8982 | return LastInst; | |||
8983 | }; | |||
8984 | ||||
8985 | auto &&FindFirstInst = [E, Front, this]() { | |||
8986 | Instruction *FirstInst = Front; | |||
8987 | for (Value *V : E->Scalars) { | |||
8988 | auto *I = dyn_cast<Instruction>(V); | |||
8989 | if (!I) | |||
8990 | continue; | |||
8991 | if (FirstInst->getParent() == I->getParent()) { | |||
8992 | if (I->comesBefore(FirstInst)) | |||
8993 | FirstInst = I; | |||
8994 | continue; | |||
8995 | } | |||
8996 | assert(((E->getOpcode() == Instruction::GetElementPtr &&(static_cast <bool> (((E->getOpcode() == Instruction ::GetElementPtr && !isa<GetElementPtrInst>(I)) || (isVectorLikeInstWithConstOps(FirstInst) && isVectorLikeInstWithConstOps (I))) && "Expected vector-like or non-GEP in GEP node insts only." ) ? void (0) : __assert_fail ("((E->getOpcode() == Instruction::GetElementPtr && !isa<GetElementPtrInst>(I)) || (isVectorLikeInstWithConstOps(FirstInst) && isVectorLikeInstWithConstOps(I))) && \"Expected vector-like or non-GEP in GEP node insts only.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 9000, __extension__ __PRETTY_FUNCTION__)) | |||
8997 | !isa<GetElementPtrInst>(I)) ||(static_cast <bool> (((E->getOpcode() == Instruction ::GetElementPtr && !isa<GetElementPtrInst>(I)) || (isVectorLikeInstWithConstOps(FirstInst) && isVectorLikeInstWithConstOps (I))) && "Expected vector-like or non-GEP in GEP node insts only." ) ? void (0) : __assert_fail ("((E->getOpcode() == Instruction::GetElementPtr && !isa<GetElementPtrInst>(I)) || (isVectorLikeInstWithConstOps(FirstInst) && isVectorLikeInstWithConstOps(I))) && \"Expected vector-like or non-GEP in GEP node insts only.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 9000, __extension__ __PRETTY_FUNCTION__)) | |||
8998 | (isVectorLikeInstWithConstOps(FirstInst) &&(static_cast <bool> (((E->getOpcode() == Instruction ::GetElementPtr && !isa<GetElementPtrInst>(I)) || (isVectorLikeInstWithConstOps(FirstInst) && isVectorLikeInstWithConstOps (I))) && "Expected vector-like or non-GEP in GEP node insts only." ) ? void (0) : __assert_fail ("((E->getOpcode() == Instruction::GetElementPtr && !isa<GetElementPtrInst>(I)) || (isVectorLikeInstWithConstOps(FirstInst) && isVectorLikeInstWithConstOps(I))) && \"Expected vector-like or non-GEP in GEP node insts only.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 9000, __extension__ __PRETTY_FUNCTION__)) | |||
8999 | isVectorLikeInstWithConstOps(I))) &&(static_cast <bool> (((E->getOpcode() == Instruction ::GetElementPtr && !isa<GetElementPtrInst>(I)) || (isVectorLikeInstWithConstOps(FirstInst) && isVectorLikeInstWithConstOps (I))) && "Expected vector-like or non-GEP in GEP node insts only." ) ? void (0) : __assert_fail ("((E->getOpcode() == Instruction::GetElementPtr && !isa<GetElementPtrInst>(I)) || (isVectorLikeInstWithConstOps(FirstInst) && isVectorLikeInstWithConstOps(I))) && \"Expected vector-like or non-GEP in GEP node insts only.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 9000, __extension__ __PRETTY_FUNCTION__)) | |||
9000 | "Expected vector-like or non-GEP in GEP node insts only.")(static_cast <bool> (((E->getOpcode() == Instruction ::GetElementPtr && !isa<GetElementPtrInst>(I)) || (isVectorLikeInstWithConstOps(FirstInst) && isVectorLikeInstWithConstOps (I))) && "Expected vector-like or non-GEP in GEP node insts only." ) ? void (0) : __assert_fail ("((E->getOpcode() == Instruction::GetElementPtr && !isa<GetElementPtrInst>(I)) || (isVectorLikeInstWithConstOps(FirstInst) && isVectorLikeInstWithConstOps(I))) && \"Expected vector-like or non-GEP in GEP node insts only.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 9000, __extension__ __PRETTY_FUNCTION__)); | |||
9001 | if (!DT->isReachableFromEntry(FirstInst->getParent())) { | |||
9002 | FirstInst = I; | |||
9003 | continue; | |||
9004 | } | |||
9005 | if (!DT->isReachableFromEntry(I->getParent())) | |||
9006 | continue; | |||
9007 | auto *NodeA = DT->getNode(FirstInst->getParent()); | |||
9008 | auto *NodeB = DT->getNode(I->getParent()); | |||
9009 | 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", 9009, __extension__ __PRETTY_FUNCTION__)); | |||
9010 | 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", 9010, __extension__ __PRETTY_FUNCTION__)); | |||
9011 | assert((NodeA == NodeB) ==(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", 9013, __extension__ __PRETTY_FUNCTION__)) | |||
9012 | (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", 9013, __extension__ __PRETTY_FUNCTION__)) | |||
9013 | "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", 9013, __extension__ __PRETTY_FUNCTION__)); | |||
9014 | if (NodeA->getDFSNumIn() > NodeB->getDFSNumIn()) | |||
9015 | FirstInst = I; | |||
9016 | } | |||
9017 | return FirstInst; | |||
9018 | }; | |||
9019 | ||||
9020 | // Set the insert point to the beginning of the basic block if the entry | |||
9021 | // should not be scheduled. | |||
9022 | if (E->State != TreeEntry::NeedToGather && | |||
9023 | (doesNotNeedToSchedule(E->Scalars) || | |||
9024 | all_of(E->Scalars, isVectorLikeInstWithConstOps))) { | |||
9025 | Instruction *InsertInst; | |||
9026 | if ((E->getOpcode() == Instruction::GetElementPtr && | |||
9027 | any_of(E->Scalars, | |||
9028 | [](Value *V) { | |||
9029 | return !isa<GetElementPtrInst>(V) && isa<Instruction>(V); | |||
9030 | })) || | |||
9031 | all_of(E->Scalars, [](Value *V) { | |||
9032 | return !isVectorLikeInstWithConstOps(V) && isUsedOutsideBlock(V); | |||
9033 | })) | |||
9034 | InsertInst = FindLastInst(); | |||
9035 | else | |||
9036 | InsertInst = FindFirstInst(); | |||
9037 | return *InsertInst; | |||
9038 | } | |||
9039 | ||||
9040 | // The last instruction in the bundle in program order. | |||
9041 | Instruction *LastInst = nullptr; | |||
9042 | ||||
9043 | // Find the last instruction. The common case should be that BB has been | |||
9044 | // scheduled, and the last instruction is VL.back(). So we start with | |||
9045 | // VL.back() and iterate over schedule data until we reach the end of the | |||
9046 | // bundle. The end of the bundle is marked by null ScheduleData. | |||
9047 | if (BlocksSchedules.count(BB)) { | |||
9048 | Value *V = E->isOneOf(E->Scalars.back()); | |||
9049 | if (doesNotNeedToBeScheduled(V)) | |||
9050 | V = *find_if_not(E->Scalars, doesNotNeedToBeScheduled); | |||
9051 | auto *Bundle = BlocksSchedules[BB]->getScheduleData(V); | |||
9052 | if (Bundle && Bundle->isPartOfBundle()) | |||
9053 | for (; Bundle; Bundle = Bundle->NextInBundle) | |||
9054 | if (Bundle->OpValue == Bundle->Inst) | |||
9055 | LastInst = Bundle->Inst; | |||
9056 | } | |||
9057 | ||||
9058 | // LastInst can still be null at this point if there's either not an entry | |||
9059 | // for BB in BlocksSchedules or there's no ScheduleData available for | |||
9060 | // VL.back(). This can be the case if buildTree_rec aborts for various | |||
9061 | // reasons (e.g., the maximum recursion depth is reached, the maximum region | |||
9062 | // size is reached, etc.). ScheduleData is initialized in the scheduling | |||
9063 | // "dry-run". | |||
9064 | // | |||
9065 | // If this happens, we can still find the last instruction by brute force. We | |||
9066 | // iterate forwards from Front (inclusive) until we either see all | |||
9067 | // instructions in the bundle or reach the end of the block. If Front is the | |||
9068 | // last instruction in program order, LastInst will be set to Front, and we | |||
9069 | // will visit all the remaining instructions in the block. | |||
9070 | // | |||
9071 | // One of the reasons we exit early from buildTree_rec is to place an upper | |||
9072 | // bound on compile-time. Thus, taking an additional compile-time hit here is | |||
9073 | // not ideal. However, this should be exceedingly rare since it requires that | |||
9074 | // we both exit early from buildTree_rec and that the bundle be out-of-order | |||
9075 | // (causing us to iterate all the way to the end of the block). | |||
9076 | if (!LastInst) | |||
9077 | LastInst = FindLastInst(); | |||
9078 | 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", 9078, __extension__ __PRETTY_FUNCTION__)); | |||
9079 | return *LastInst; | |||
9080 | } | |||
9081 | ||||
9082 | void BoUpSLP::setInsertPointAfterBundle(const TreeEntry *E) { | |||
9083 | auto *Front = E->getMainOp(); | |||
9084 | Instruction *LastInst = EntryToLastInstruction.lookup(E); | |||
9085 | 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", 9085, __extension__ __PRETTY_FUNCTION__)); | |||
9086 | // If the instruction is PHI, set the insert point after all the PHIs. | |||
9087 | bool IsPHI = isa<PHINode>(LastInst); | |||
9088 | if (IsPHI) | |||
9089 | LastInst = LastInst->getParent()->getFirstNonPHI(); | |||
9090 | if (IsPHI || (E->State != TreeEntry::NeedToGather && | |||
9091 | doesNotNeedToSchedule(E->Scalars))) { | |||
9092 | Builder.SetInsertPoint(LastInst); | |||
9093 | } else { | |||
9094 | // Set the insertion point after the last instruction in the bundle. Set the | |||
9095 | // debug location to Front. | |||
9096 | Builder.SetInsertPoint(LastInst->getParent(), | |||
9097 | std::next(LastInst->getIterator())); | |||
9098 | } | |||
9099 | Builder.SetCurrentDebugLocation(Front->getDebugLoc()); | |||
9100 | } | |||
9101 | ||||
9102 | Value *BoUpSLP::gather(ArrayRef<Value *> VL, Value *Root) { | |||
9103 | // List of instructions/lanes from current block and/or the blocks which are | |||
9104 | // part of the current loop. These instructions will be inserted at the end to | |||
9105 | // make it possible to optimize loops and hoist invariant instructions out of | |||
9106 | // the loops body with better chances for success. | |||
9107 | SmallVector<std::pair<Value *, unsigned>, 4> PostponedInsts; | |||
9108 | SmallSet<int, 4> PostponedIndices; | |||
9109 | Loop *L = LI->getLoopFor(Builder.GetInsertBlock()); | |||
9110 | auto &&CheckPredecessor = [](BasicBlock *InstBB, BasicBlock *InsertBB) { | |||
9111 | SmallPtrSet<BasicBlock *, 4> Visited; | |||
9112 | while (InsertBB && InsertBB != InstBB && Visited.insert(InsertBB).second) | |||
9113 | InsertBB = InsertBB->getSinglePredecessor(); | |||
9114 | return InsertBB && InsertBB == InstBB; | |||
9115 | }; | |||
9116 | for (int I = 0, E = VL.size(); I < E; ++I) { | |||
9117 | if (auto *Inst = dyn_cast<Instruction>(VL[I])) | |||
9118 | if ((CheckPredecessor(Inst->getParent(), Builder.GetInsertBlock()) || | |||
9119 | getTreeEntry(Inst) || | |||
9120 | (L && (!Root || L->isLoopInvariant(Root)) && L->contains(Inst))) && | |||
9121 | PostponedIndices.insert(I).second) | |||
9122 | PostponedInsts.emplace_back(Inst, I); | |||
9123 | } | |||
9124 | ||||
9125 | auto &&CreateInsertElement = [this](Value *Vec, Value *V, unsigned Pos) { | |||
9126 | Vec = Builder.CreateInsertElement(Vec, V, Builder.getInt32(Pos)); | |||
9127 | auto *InsElt = dyn_cast<InsertElementInst>(Vec); | |||
9128 | if (!InsElt) | |||
9129 | return Vec; | |||
9130 | GatherShuffleExtractSeq.insert(InsElt); | |||
9131 | CSEBlocks.insert(InsElt->getParent()); | |||
9132 | // Add to our 'need-to-extract' list. | |||
9133 | if (TreeEntry *Entry = getTreeEntry(V)) { | |||
9134 | // Find which lane we need to extract. | |||
9135 | unsigned FoundLane = Entry->findLaneForValue(V); | |||
9136 | ExternalUses.emplace_back(V, InsElt, FoundLane); | |||
9137 | } | |||
9138 | return Vec; | |||
9139 | }; | |||
9140 | Value *Val0 = | |||
9141 | isa<StoreInst>(VL[0]) ? cast<StoreInst>(VL[0])->getValueOperand() : VL[0]; | |||
9142 | FixedVectorType *VecTy = FixedVectorType::get(Val0->getType(), VL.size()); | |||
9143 | Value *Vec = Root ? Root : PoisonValue::get(VecTy); | |||
9144 | SmallVector<int> NonConsts; | |||
9145 | // Insert constant values at first. | |||
9146 | for (int I = 0, E = VL.size(); I < E; ++I) { | |||
9147 | if (PostponedIndices.contains(I)) | |||
9148 | continue; | |||
9149 | if (!isConstant(VL[I])) { | |||
9150 | NonConsts.push_back(I); | |||
9151 | continue; | |||
9152 | } | |||
9153 | if (Root) { | |||
9154 | if (!isa<UndefValue>(VL[I])) { | |||
9155 | NonConsts.push_back(I); | |||
9156 | continue; | |||
9157 | } | |||
9158 | if (isa<PoisonValue>(VL[I])) | |||
9159 | continue; | |||
9160 | if (auto *SV = dyn_cast<ShuffleVectorInst>(Root)) { | |||
9161 | if (SV->getMaskValue(I) == PoisonMaskElem) | |||
9162 | continue; | |||
9163 | } | |||
9164 | } | |||
9165 | Vec = CreateInsertElement(Vec, VL[I], I); | |||
9166 | } | |||
9167 | // Insert non-constant values. | |||
9168 | for (int I : NonConsts) | |||
9169 | Vec = CreateInsertElement(Vec, VL[I], I); | |||
9170 | // Append instructions, which are/may be part of the loop, in the end to make | |||
9171 | // it possible to hoist non-loop-based instructions. | |||
9172 | for (const std::pair<Value *, unsigned> &Pair : PostponedInsts) | |||
9173 | Vec = CreateInsertElement(Vec, Pair.first, Pair.second); | |||
9174 | ||||
9175 | return Vec; | |||
9176 | } | |||
9177 | ||||
9178 | /// Merges shuffle masks and emits final shuffle instruction, if required. It | |||
9179 | /// supports shuffling of 2 input vectors. It implements lazy shuffles emission, | |||
9180 | /// when the actual shuffle instruction is generated only if this is actually | |||
9181 | /// required. Otherwise, the shuffle instruction emission is delayed till the | |||
9182 | /// end of the process, to reduce the number of emitted instructions and further | |||
9183 | /// analysis/transformations. | |||
9184 | /// The class also will look through the previously emitted shuffle instructions | |||
9185 | /// and properly mark indices in mask as undef. | |||
9186 | /// For example, given the code | |||
9187 | /// \code | |||
9188 | /// %s1 = shufflevector <2 x ty> %0, poison, <1, 0> | |||
9189 | /// %s2 = shufflevector <2 x ty> %1, poison, <1, 0> | |||
9190 | /// \endcode | |||
9191 | /// and if need to emit shuffle of %s1 and %s2 with mask <1, 0, 3, 2>, it will | |||
9192 | /// look through %s1 and %s2 and emit | |||
9193 | /// \code | |||
9194 | /// %res = shufflevector <2 x ty> %0, %1, <0, 1, 2, 3> | |||
9195 | /// \endcode | |||
9196 | /// instead. | |||
9197 | /// If 2 operands are of different size, the smallest one will be resized and | |||
9198 | /// the mask recalculated properly. | |||
9199 | /// For example, given the code | |||
9200 | /// \code | |||
9201 | /// %s1 = shufflevector <2 x ty> %0, poison, <1, 0, 1, 0> | |||
9202 | /// %s2 = shufflevector <2 x ty> %1, poison, <1, 0, 1, 0> | |||
9203 | /// \endcode | |||
9204 | /// and if need to emit shuffle of %s1 and %s2 with mask <1, 0, 5, 4>, it will | |||
9205 | /// look through %s1 and %s2 and emit | |||
9206 | /// \code | |||
9207 | /// %res = shufflevector <2 x ty> %0, %1, <0, 1, 2, 3> | |||
9208 | /// \endcode | |||
9209 | /// instead. | |||
9210 | class BoUpSLP::ShuffleInstructionBuilder final : public BaseShuffleAnalysis { | |||
9211 | bool IsFinalized = false; | |||
9212 | /// Combined mask for all applied operands and masks. It is built during | |||
9213 | /// analysis and actual emission of shuffle vector instructions. | |||
9214 | SmallVector<int> CommonMask; | |||
9215 | /// List of operands for the shuffle vector instruction. It hold at max 2 | |||
9216 | /// operands, if the 3rd is going to be added, the first 2 are combined into | |||
9217 | /// shuffle with \p CommonMask mask, the first operand sets to be the | |||
9218 | /// resulting shuffle and the second operand sets to be the newly added | |||
9219 | /// operand. The \p CommonMask is transformed in the proper way after that. | |||
9220 | SmallVector<Value *, 2> InVectors; | |||
9221 | IRBuilderBase &Builder; | |||
9222 | BoUpSLP &R; | |||
9223 | ||||
9224 | class ShuffleIRBuilder { | |||
9225 | IRBuilderBase &Builder; | |||
9226 | /// Holds all of the instructions that we gathered. | |||
9227 | SetVector<Instruction *> &GatherShuffleExtractSeq; | |||
9228 | /// A list of blocks that we are going to CSE. | |||
9229 | SetVector<BasicBlock *> &CSEBlocks; | |||
9230 | ||||
9231 | public: | |||
9232 | ShuffleIRBuilder(IRBuilderBase &Builder, | |||
9233 | SetVector<Instruction *> &GatherShuffleExtractSeq, | |||
9234 | SetVector<BasicBlock *> &CSEBlocks) | |||
9235 | : Builder(Builder), GatherShuffleExtractSeq(GatherShuffleExtractSeq), | |||
9236 | CSEBlocks(CSEBlocks) {} | |||
9237 | ~ShuffleIRBuilder() = default; | |||
9238 | /// Creates shufflevector for the 2 operands with the given mask. | |||
9239 | Value *createShuffleVector(Value *V1, Value *V2, ArrayRef<int> Mask) { | |||
9240 | Value *Vec = Builder.CreateShuffleVector(V1, V2, Mask); | |||
9241 | if (auto *I = dyn_cast<Instruction>(Vec)) { | |||
9242 | GatherShuffleExtractSeq.insert(I); | |||
9243 | CSEBlocks.insert(I->getParent()); | |||
9244 | } | |||
9245 | return Vec; | |||
9246 | } | |||
9247 | /// Creates permutation of the single vector operand with the given mask, if | |||
9248 | /// it is not identity mask. | |||
9249 | Value *createShuffleVector(Value *V1, ArrayRef<int> Mask) { | |||
9250 | if (Mask.empty()) | |||
9251 | return V1; | |||
9252 | unsigned VF = Mask.size(); | |||
9253 | unsigned LocalVF = cast<FixedVectorType>(V1->getType())->getNumElements(); | |||
9254 | if (VF == LocalVF && ShuffleVectorInst::isIdentityMask(Mask)) | |||
9255 | return V1; | |||
9256 | Value *Vec = Builder.CreateShuffleVector(V1, Mask); | |||
9257 | if (auto *I = dyn_cast<Instruction>(Vec)) { | |||
9258 | GatherShuffleExtractSeq.insert(I); | |||
9259 | CSEBlocks.insert(I->getParent()); | |||
9260 | } | |||
9261 | return Vec; | |||
9262 | } | |||
9263 | Value *createIdentity(Value *V) { return V; } | |||
9264 | Value *createPoison(Type *Ty, unsigned VF) { | |||
9265 | return PoisonValue::get(FixedVectorType::get(Ty, VF)); | |||
9266 | } | |||
9267 | /// Resizes 2 input vector to match the sizes, if the they are not equal | |||
9268 | /// yet. The smallest vector is resized to the size of the larger vector. | |||
9269 | void resizeToMatch(Value *&V1, Value *&V2) { | |||
9270 | if (V1->getType() == V2->getType()) | |||
9271 | return; | |||
9272 | int V1VF = cast<FixedVectorType>(V1->getType())->getNumElements(); | |||
9273 | int V2VF = cast<FixedVectorType>(V2->getType())->getNumElements(); | |||
9274 | int VF = std::max(V1VF, V2VF); | |||
9275 | int MinVF = std::min(V1VF, V2VF); | |||
9276 | SmallVector<int> IdentityMask(VF, PoisonMaskElem); | |||
9277 | std::iota(IdentityMask.begin(), std::next(IdentityMask.begin(), MinVF), | |||
9278 | 0); | |||
9279 | Value *&Op = MinVF == V1VF ? V1 : V2; | |||
9280 | Op = Builder.CreateShuffleVector(Op, IdentityMask); | |||
9281 | if (auto *I = dyn_cast<Instruction>(Op)) { | |||
9282 | GatherShuffleExtractSeq.insert(I); | |||
9283 | CSEBlocks.insert(I->getParent()); | |||
9284 | } | |||
9285 | if (MinVF == V1VF) | |||
9286 | V1 = Op; | |||
9287 | else | |||
9288 | V2 = Op; | |||
9289 | } | |||
9290 | }; | |||
9291 | ||||
9292 | /// Smart shuffle instruction emission, walks through shuffles trees and | |||
9293 | /// tries to find the best matching vector for the actual shuffle | |||
9294 | /// instruction. | |||
9295 | Value *createShuffle(Value *V1, Value *V2, ArrayRef<int> Mask) { | |||
9296 | assert(V1 && "Expected at least one vector value.")(static_cast <bool> (V1 && "Expected at least one vector value." ) ? void (0) : __assert_fail ("V1 && \"Expected at least one vector value.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 9296, __extension__ __PRETTY_FUNCTION__)); | |||
9297 | ShuffleIRBuilder ShuffleBuilder(Builder, R.GatherShuffleExtractSeq, | |||
9298 | R.CSEBlocks); | |||
9299 | return BaseShuffleAnalysis::createShuffle<Value *>(V1, V2, Mask, | |||
9300 | ShuffleBuilder); | |||
9301 | } | |||
9302 | ||||
9303 | /// Transforms mask \p CommonMask per given \p Mask to make proper set after | |||
9304 | /// shuffle emission. | |||
9305 | static void transformMaskAfterShuffle(MutableArrayRef<int> CommonMask, | |||
9306 | ArrayRef<int> Mask) { | |||
9307 | for (unsigned Idx = 0, Sz = CommonMask.size(); Idx < Sz; ++Idx) | |||
9308 | if (Mask[Idx] != PoisonMaskElem) | |||
9309 | CommonMask[Idx] = Idx; | |||
9310 | } | |||
9311 | ||||
9312 | public: | |||
9313 | ShuffleInstructionBuilder(IRBuilderBase &Builder, BoUpSLP &R) | |||
9314 | : Builder(Builder), R(R) {} | |||
9315 | ||||
9316 | /// Adjusts extractelements after reusing them. | |||
9317 | Value *adjustExtracts(const TreeEntry *E, ArrayRef<int> Mask) { | |||
9318 | Value *VecBase = nullptr; | |||
9319 | for (int I = 0, Sz = Mask.size(); I < Sz; ++I) { | |||
9320 | int Idx = Mask[I]; | |||
9321 | if (Idx == PoisonMaskElem) | |||
9322 | continue; | |||
9323 | auto *EI = cast<ExtractElementInst>(E->Scalars[I]); | |||
9324 | VecBase = EI->getVectorOperand(); | |||
9325 | // If the only one use is vectorized - can delete the extractelement | |||
9326 | // itself. | |||
9327 | if (!EI->hasOneUse() || any_of(EI->users(), [&](User *U) { | |||
9328 | return !R.ScalarToTreeEntry.count(U); | |||
9329 | })) | |||
9330 | continue; | |||
9331 | R.eraseInstruction(EI); | |||
9332 | } | |||
9333 | return VecBase; | |||
9334 | } | |||
9335 | /// Checks if the specified entry \p E needs to be delayed because of its | |||
9336 | /// dependency nodes. | |||
9337 | Value *needToDelay(const TreeEntry *E, ArrayRef<const TreeEntry *> Deps) { | |||
9338 | // No need to delay emission if all deps are ready. | |||
9339 | if (all_of(Deps, [](const TreeEntry *TE) { return TE->VectorizedValue; })) | |||
9340 | return nullptr; | |||
9341 | // Postpone gather emission, will be emitted after the end of the | |||
9342 | // process to keep correct order. | |||
9343 | auto *VecTy = FixedVectorType::get(E->Scalars.front()->getType(), | |||
9344 | E->getVectorFactor()); | |||
9345 | Value *Vec = Builder.CreateAlignedLoad( | |||
9346 | VecTy, PoisonValue::get(VecTy->getPointerTo()), MaybeAlign()); | |||
9347 | return Vec; | |||
9348 | } | |||
9349 | /// Adds 2 input vectors and the mask for their shuffling. | |||
9350 | void add(Value *V1, Value *V2, ArrayRef<int> Mask) { | |||
9351 | assert(V1 && V2 && !Mask.empty() && "Expected non-empty input vectors.")(static_cast <bool> (V1 && V2 && !Mask. empty() && "Expected non-empty input vectors.") ? void (0) : __assert_fail ("V1 && V2 && !Mask.empty() && \"Expected non-empty input vectors.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 9351, __extension__ __PRETTY_FUNCTION__)); | |||
9352 | if (InVectors.empty()) { | |||
9353 | InVectors.push_back(V1); | |||
9354 | InVectors.push_back(V2); | |||
9355 | CommonMask.assign(Mask.begin(), Mask.end()); | |||
9356 | return; | |||
9357 | } | |||
9358 | Value *Vec = InVectors.front(); | |||
9359 | if (InVectors.size() == 2) { | |||
9360 | Vec = createShuffle(Vec, InVectors.back(), CommonMask); | |||
9361 | transformMaskAfterShuffle(CommonMask, CommonMask); | |||
9362 | } else if (cast<FixedVectorType>(Vec->getType())->getNumElements() != | |||
9363 | Mask.size()) { | |||
9364 | Vec = createShuffle(Vec, nullptr, CommonMask); | |||
9365 | transformMaskAfterShuffle(CommonMask, CommonMask); | |||
9366 | } | |||
9367 | V1 = createShuffle(V1, V2, Mask); | |||
9368 | for (unsigned Idx = 0, Sz = CommonMask.size(); Idx < Sz; ++Idx) | |||
9369 | if (Mask[Idx] != PoisonMaskElem) | |||
9370 | CommonMask[Idx] = Idx + Sz; | |||
9371 | InVectors.front() = Vec; | |||
9372 | if (InVectors.size() == 2) | |||
9373 | InVectors.back() = V1; | |||
9374 | else | |||
9375 | InVectors.push_back(V1); | |||
9376 | } | |||
9377 | /// Adds another one input vector and the mask for the shuffling. | |||
9378 | void add(Value *V1, ArrayRef<int> Mask) { | |||
9379 | if (InVectors.empty()) { | |||
9380 | if (!isa<FixedVectorType>(V1->getType())) { | |||
9381 | V1 = createShuffle(V1, nullptr, CommonMask); | |||
9382 | CommonMask.assign(Mask.size(), PoisonMaskElem); | |||
9383 | transformMaskAfterShuffle(CommonMask, Mask); | |||
9384 | } | |||
9385 | InVectors.push_back(V1); | |||
9386 | CommonMask.assign(Mask.begin(), Mask.end()); | |||
9387 | return; | |||
9388 | } | |||
9389 | const auto *It = find(InVectors, V1); | |||
9390 | if (It == InVectors.end()) { | |||
9391 | if (InVectors.size() == 2 || | |||
9392 | InVectors.front()->getType() != V1->getType() || | |||
9393 | !isa<FixedVectorType>(V1->getType())) { | |||
9394 | Value *V = InVectors.front(); | |||
9395 | if (InVectors.size() == 2) { | |||
9396 | V = createShuffle(InVectors.front(), InVectors.back(), CommonMask); | |||
9397 | transformMaskAfterShuffle(CommonMask, CommonMask); | |||
9398 | } else if (cast<FixedVectorType>(V->getType())->getNumElements() != | |||
9399 | CommonMask.size()) { | |||
9400 | V = createShuffle(InVectors.front(), nullptr, CommonMask); | |||
9401 | transformMaskAfterShuffle(CommonMask, CommonMask); | |||
9402 | } | |||
9403 | for (unsigned Idx = 0, Sz = CommonMask.size(); Idx < Sz; ++Idx) | |||
9404 | if (CommonMask[Idx] == PoisonMaskElem && Mask[Idx] != PoisonMaskElem) | |||
9405 | CommonMask[Idx] = | |||
9406 | V->getType() != V1->getType() | |||
9407 | ? Idx + Sz | |||
9408 | : Mask[Idx] + cast<FixedVectorType>(V1->getType()) | |||
9409 | ->getNumElements(); | |||
9410 | if (V->getType() != V1->getType()) | |||
9411 | V1 = createShuffle(V1, nullptr, Mask); | |||
9412 | InVectors.front() = V; | |||
9413 | if (InVectors.size() == 2) | |||
9414 | InVectors.back() = V1; | |||
9415 | else | |||
9416 | InVectors.push_back(V1); | |||
9417 | return; | |||
9418 | } | |||
9419 | // Check if second vector is required if the used elements are already | |||
9420 | // used from the first one. | |||
9421 | for (unsigned Idx = 0, Sz = CommonMask.size(); Idx < Sz; ++Idx) | |||
9422 | if (Mask[Idx] != PoisonMaskElem && CommonMask[Idx] == PoisonMaskElem) { | |||
9423 | InVectors.push_back(V1); | |||
9424 | break; | |||
9425 | } | |||
9426 | } | |||
9427 | int VF = CommonMask.size(); | |||
9428 | if (auto *FTy = dyn_cast<FixedVectorType>(V1->getType())) | |||
9429 | VF = FTy->getNumElements(); | |||
9430 | for (unsigned Idx = 0, Sz = CommonMask.size(); Idx < Sz; ++Idx) | |||
9431 | if (Mask[Idx] != PoisonMaskElem && CommonMask[Idx] == PoisonMaskElem) | |||
9432 | CommonMask[Idx] = Mask[Idx] + (It == InVectors.begin() ? 0 : VF); | |||
9433 | } | |||
9434 | /// Adds another one input vector and the mask for the shuffling. | |||
9435 | void addOrdered(Value *V1, ArrayRef<unsigned> Order) { | |||
9436 | SmallVector<int> NewMask; | |||
9437 | inversePermutation(Order, NewMask); | |||
9438 | add(V1, NewMask); | |||
9439 | } | |||
9440 | Value *gather(ArrayRef<Value *> VL, Value *Root = nullptr) { | |||
9441 | return R.gather(VL, Root); | |||
9442 | } | |||
9443 | Value *createFreeze(Value *V) { return Builder.CreateFreeze(V); } | |||
9444 | /// Finalize emission of the shuffles. | |||
9445 | /// \param Action the action (if any) to be performed before final applying of | |||
9446 | /// the \p ExtMask mask. | |||
9447 | Value * | |||
9448 | finalize(ArrayRef<int> ExtMask, unsigned VF = 0, | |||
9449 | function_ref<void(Value *&, SmallVectorImpl<int> &)> Action = {}) { | |||
9450 | IsFinalized = true; | |||
9451 | if (Action) { | |||
9452 | Value *Vec = InVectors.front(); | |||
9453 | if (InVectors.size() == 2) { | |||
9454 | Vec = createShuffle(Vec, InVectors.back(), CommonMask); | |||
9455 | InVectors.pop_back(); | |||
9456 | } else { | |||
9457 | Vec = createShuffle(Vec, nullptr, CommonMask); | |||
9458 | } | |||
9459 | for (unsigned Idx = 0, Sz = CommonMask.size(); Idx < Sz; ++Idx) | |||
9460 | if (CommonMask[Idx] != PoisonMaskElem) | |||
9461 | CommonMask[Idx] = Idx; | |||
9462 | assert(VF > 0 &&(static_cast <bool> (VF > 0 && "Expected vector length for the final value before action." ) ? void (0) : __assert_fail ("VF > 0 && \"Expected vector length for the final value before action.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 9463, __extension__ __PRETTY_FUNCTION__)) | |||
9463 | "Expected vector length for the final value before action.")(static_cast <bool> (VF > 0 && "Expected vector length for the final value before action." ) ? void (0) : __assert_fail ("VF > 0 && \"Expected vector length for the final value before action.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 9463, __extension__ __PRETTY_FUNCTION__)); | |||
9464 | unsigned VecVF = cast<FixedVectorType>(Vec->getType())->getNumElements(); | |||
9465 | if (VecVF < VF) { | |||
9466 | SmallVector<int> ResizeMask(VF, PoisonMaskElem); | |||
9467 | std::iota(ResizeMask.begin(), std::next(ResizeMask.begin(), VecVF), 0); | |||
9468 | Vec = createShuffle(Vec, nullptr, ResizeMask); | |||
9469 | } | |||
9470 | Action(Vec, CommonMask); | |||
9471 | InVectors.front() = Vec; | |||
9472 | } | |||
9473 | if (!ExtMask.empty()) { | |||
9474 | if (CommonMask.empty()) { | |||
9475 | CommonMask.assign(ExtMask.begin(), ExtMask.end()); | |||
9476 | } else { | |||
9477 | SmallVector<int> NewMask(ExtMask.size(), PoisonMaskElem); | |||
9478 | for (int I = 0, Sz = ExtMask.size(); I < Sz; ++I) { | |||
9479 | if (ExtMask[I] == PoisonMaskElem) | |||
9480 | continue; | |||
9481 | NewMask[I] = CommonMask[ExtMask[I]]; | |||
9482 | } | |||
9483 | CommonMask.swap(NewMask); | |||
9484 | } | |||
9485 | } | |||
9486 | if (CommonMask.empty()) { | |||
9487 | assert(InVectors.size() == 1 && "Expected only one vector with no mask")(static_cast <bool> (InVectors.size() == 1 && "Expected only one vector with no mask" ) ? void (0) : __assert_fail ("InVectors.size() == 1 && \"Expected only one vector with no mask\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 9487, __extension__ __PRETTY_FUNCTION__)); | |||
9488 | return InVectors.front(); | |||
9489 | } | |||
9490 | if (InVectors.size() == 2) | |||
9491 | return createShuffle(InVectors.front(), InVectors.back(), CommonMask); | |||
9492 | return createShuffle(InVectors.front(), nullptr, CommonMask); | |||
9493 | } | |||
9494 | ||||
9495 | ~ShuffleInstructionBuilder() { | |||
9496 | assert((IsFinalized || CommonMask.empty()) &&(static_cast <bool> ((IsFinalized || CommonMask.empty() ) && "Shuffle construction must be finalized.") ? void (0) : __assert_fail ("(IsFinalized || CommonMask.empty()) && \"Shuffle construction must be finalized.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 9497, __extension__ __PRETTY_FUNCTION__)) | |||
9497 | "Shuffle construction must be finalized.")(static_cast <bool> ((IsFinalized || CommonMask.empty() ) && "Shuffle construction must be finalized.") ? void (0) : __assert_fail ("(IsFinalized || CommonMask.empty()) && \"Shuffle construction must be finalized.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 9497, __extension__ __PRETTY_FUNCTION__)); | |||
9498 | } | |||
9499 | }; | |||
9500 | ||||
9501 | Value *BoUpSLP::vectorizeOperand(TreeEntry *E, unsigned NodeIdx) { | |||
9502 | ArrayRef<Value *> VL = E->getOperand(NodeIdx); | |||
9503 | const unsigned VF = VL.size(); | |||
9504 | InstructionsState S = getSameOpcode(VL, *TLI); | |||
9505 | // Special processing for GEPs bundle, which may include non-gep values. | |||
9506 | if (!S.getOpcode() && VL.front()->getType()->isPointerTy()) { | |||
9507 | const auto *It = | |||
9508 | find_if(VL, [](Value *V) { return isa<GetElementPtrInst>(V); }); | |||
9509 | if (It != VL.end()) | |||
9510 | S = getSameOpcode(*It, *TLI); | |||
9511 | } | |||
9512 | if (S.getOpcode()) { | |||
9513 | if (TreeEntry *VE = getTreeEntry(S.OpValue); | |||
9514 | VE && VE->isSame(VL) && | |||
9515 | (any_of(VE->UserTreeIndices, | |||
9516 | [E, NodeIdx](const EdgeInfo &EI) { | |||
9517 | return EI.UserTE == E && EI.EdgeIdx == NodeIdx; | |||
9518 | }) || | |||
9519 | any_of(VectorizableTree, | |||
9520 | [E, NodeIdx, VE](const std::unique_ptr<TreeEntry> &TE) { | |||
9521 | return TE->isOperandGatherNode({E, NodeIdx}) && | |||
9522 | VE->isSame(TE->Scalars); | |||
9523 | }))) { | |||
9524 | auto FinalShuffle = [&](Value *V, ArrayRef<int> Mask) { | |||
9525 | ShuffleInstructionBuilder ShuffleBuilder(Builder, *this); | |||
9526 | ShuffleBuilder.add(V, Mask); | |||
9527 | return ShuffleBuilder.finalize(std::nullopt); | |||
9528 | }; | |||
9529 | Value *V = vectorizeTree(VE); | |||
9530 | if (VF != cast<FixedVectorType>(V->getType())->getNumElements()) { | |||
9531 | if (!VE->ReuseShuffleIndices.empty()) { | |||
9532 | // Reshuffle to get only unique values. | |||
9533 | // If some of the scalars are duplicated in the vectorization | |||
9534 | // tree entry, we do not vectorize them but instead generate a | |||
9535 | // mask for the reuses. But if there are several users of the | |||
9536 | // same entry, they may have different vectorization factors. | |||
9537 | // This is especially important for PHI nodes. In this case, we | |||
9538 | // need to adapt the resulting instruction for the user | |||
9539 | // vectorization factor and have to reshuffle it again to take | |||
9540 | // only unique elements of the vector. Without this code the | |||
9541 | // function incorrectly returns reduced vector instruction with | |||
9542 | // the same elements, not with the unique ones. | |||
9543 | ||||
9544 | // block: | |||
9545 | // %phi = phi <2 x > { .., %entry} {%shuffle, %block} | |||
9546 | // %2 = shuffle <2 x > %phi, poison, <4 x > <1, 1, 0, 0> | |||
9547 | // ... (use %2) | |||
9548 | // %shuffle = shuffle <2 x> %2, poison, <2 x> {2, 0} | |||
9549 | // br %block | |||
9550 | SmallVector<int> UniqueIdxs(VF, PoisonMaskElem); | |||
9551 | SmallSet<int, 4> UsedIdxs; | |||
9552 | int Pos = 0; | |||
9553 | for (int Idx : VE->ReuseShuffleIndices) { | |||
9554 | if (Idx != static_cast<int>(VF) && Idx != PoisonMaskElem && | |||
9555 | UsedIdxs.insert(Idx).second) | |||
9556 | UniqueIdxs[Idx] = Pos; | |||
9557 | ++Pos; | |||
9558 | } | |||
9559 | 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", 9560, __extension__ __PRETTY_FUNCTION__)) | |||
9560 | "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", 9560, __extension__ __PRETTY_FUNCTION__)); | |||
9561 | UniqueIdxs.append(VF - UsedIdxs.size(), PoisonMaskElem); | |||
9562 | V = FinalShuffle(V, UniqueIdxs); | |||
9563 | } else { | |||
9564 | 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", 9566, __extension__ __PRETTY_FUNCTION__)) | |||
9565 | "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", 9566, __extension__ __PRETTY_FUNCTION__)) | |||
9566 | "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", 9566, __extension__ __PRETTY_FUNCTION__)); | |||
9567 | SmallVector<int> UniformMask(VF, 0); | |||
9568 | std::iota(UniformMask.begin(), UniformMask.end(), 0); | |||
9569 | V = FinalShuffle(V, UniformMask); | |||
9570 | } | |||
9571 | } | |||
9572 | // Need to update the operand gather node, if actually the operand is not a | |||
9573 | // vectorized node, but the buildvector/gather node, which matches one of | |||
9574 | // the vectorized nodes. | |||
9575 | if (find_if(VE->UserTreeIndices, [&](const EdgeInfo &EI) { | |||
9576 | return EI.UserTE == E && EI.EdgeIdx == NodeIdx; | |||
9577 | }) == VE->UserTreeIndices.end()) { | |||
9578 | auto *It = find_if( | |||
9579 | VectorizableTree, [&](const std::unique_ptr<TreeEntry> &TE) { | |||
9580 | return TE->State == TreeEntry::NeedToGather && | |||
9581 | TE->UserTreeIndices.front().UserTE == E && | |||
9582 | TE->UserTreeIndices.front().EdgeIdx == NodeIdx; | |||
9583 | }); | |||
9584 | assert(It != VectorizableTree.end() && "Expected gather node operand.")(static_cast <bool> (It != VectorizableTree.end() && "Expected gather node operand.") ? void (0) : __assert_fail ( "It != VectorizableTree.end() && \"Expected gather node operand.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 9584, __extension__ __PRETTY_FUNCTION__)); | |||
9585 | (*It)->VectorizedValue = V; | |||
9586 | } | |||
9587 | return V; | |||
9588 | } | |||
9589 | } | |||
9590 | ||||
9591 | // Find the corresponding gather entry and vectorize it. | |||
9592 | // Allows to be more accurate with tree/graph transformations, checks for the | |||
9593 | // correctness of the transformations in many cases. | |||
9594 | auto *I = find_if(VectorizableTree, | |||
9595 | [E, NodeIdx](const std::unique_ptr<TreeEntry> &TE) { | |||
9596 | return TE->isOperandGatherNode({E, NodeIdx}); | |||
9597 | }); | |||
9598 | assert(I != VectorizableTree.end() && "Gather node is not in the graph.")(static_cast <bool> (I != VectorizableTree.end() && "Gather node is not in the graph.") ? void (0) : __assert_fail ("I != VectorizableTree.end() && \"Gather node is not in the graph.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 9598, __extension__ __PRETTY_FUNCTION__)); | |||
9599 | assert(I->get()->UserTreeIndices.size() == 1 &&(static_cast <bool> (I->get()->UserTreeIndices.size () == 1 && "Expected only single user for the gather node." ) ? void (0) : __assert_fail ("I->get()->UserTreeIndices.size() == 1 && \"Expected only single user for the gather node.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 9600, __extension__ __PRETTY_FUNCTION__)) | |||
9600 | "Expected only single user for the gather node.")(static_cast <bool> (I->get()->UserTreeIndices.size () == 1 && "Expected only single user for the gather node." ) ? void (0) : __assert_fail ("I->get()->UserTreeIndices.size() == 1 && \"Expected only single user for the gather node.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 9600, __extension__ __PRETTY_FUNCTION__)); | |||
9601 | assert(I->get()->isSame(VL) && "Expected same list of scalars.")(static_cast <bool> (I->get()->isSame(VL) && "Expected same list of scalars.") ? void (0) : __assert_fail ("I->get()->isSame(VL) && \"Expected same list of scalars.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 9601, __extension__ __PRETTY_FUNCTION__)); | |||
9602 | IRBuilder<>::InsertPointGuard Guard(Builder); | |||
9603 | if (E->getOpcode() != Instruction::InsertElement && | |||
9604 | E->getOpcode() != Instruction::PHI) { | |||
9605 | Instruction *LastInst = EntryToLastInstruction.lookup(E); | |||
9606 | 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", 9606, __extension__ __PRETTY_FUNCTION__)); | |||
9607 | Builder.SetInsertPoint(LastInst); | |||
9608 | } | |||
9609 | return vectorizeTree(I->get()); | |||
9610 | } | |||
9611 | ||||
9612 | template <typename BVTy, typename ResTy, typename... Args> | |||
9613 | ResTy BoUpSLP::processBuildVector(const TreeEntry *E, Args &...Params) { | |||
9614 | assert(E->State == TreeEntry::NeedToGather && "Expected gather node.")(static_cast <bool> (E->State == TreeEntry::NeedToGather && "Expected gather node.") ? void (0) : __assert_fail ("E->State == TreeEntry::NeedToGather && \"Expected gather node.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 9614, __extension__ __PRETTY_FUNCTION__)); | |||
9615 | unsigned VF = E->getVectorFactor(); | |||
9616 | ||||
9617 | bool NeedFreeze = false; | |||
9618 | SmallVector<int> ReuseShuffleIndicies(E->ReuseShuffleIndices.begin(), | |||
9619 | E->ReuseShuffleIndices.end()); | |||
9620 | SmallVector<Value *> GatheredScalars(E->Scalars.begin(), E->Scalars.end()); | |||
9621 | // Build a mask out of the reorder indices and reorder scalars per this | |||
9622 | // mask. | |||
9623 | SmallVector<int> ReorderMask; | |||
9624 | inversePermutation(E->ReorderIndices, ReorderMask); | |||
9625 | if (!ReorderMask.empty()) | |||
9626 | reorderScalars(GatheredScalars, ReorderMask); | |||
9627 | auto FindReusedSplat = [&](SmallVectorImpl<int> &Mask) { | |||
9628 | if (!isSplat(E->Scalars) || none_of(E->Scalars, [](Value *V) { | |||
9629 | return isa<UndefValue>(V) && !isa<PoisonValue>(V); | |||
9630 | })) | |||
9631 | return false; | |||
9632 | TreeEntry *UserTE = E->UserTreeIndices.back().UserTE; | |||
9633 | unsigned EdgeIdx = E->UserTreeIndices.back().EdgeIdx; | |||
9634 | if (UserTE->getNumOperands() != 2) | |||
9635 | return false; | |||
9636 | auto *It = | |||
9637 | find_if(VectorizableTree, [=](const std::unique_ptr<TreeEntry> &TE) { | |||
9638 | return find_if(TE->UserTreeIndices, [=](const EdgeInfo &EI) { | |||
9639 | return EI.UserTE == UserTE && EI.EdgeIdx != EdgeIdx; | |||
9640 | }) != TE->UserTreeIndices.end(); | |||
9641 | }); | |||
9642 | if (It == VectorizableTree.end()) | |||
9643 | return false; | |||
9644 | unsigned I = | |||
9645 | *find_if_not(Mask, [](int Idx) { return Idx == PoisonMaskElem; }); | |||
9646 | int Sz = Mask.size(); | |||
9647 | if (all_of(Mask, [Sz](int Idx) { return Idx < 2 * Sz; }) && | |||
9648 | ShuffleVectorInst::isIdentityMask(Mask)) | |||
9649 | std::iota(Mask.begin(), Mask.end(), 0); | |||
9650 | else | |||
9651 | std::fill(Mask.begin(), Mask.end(), I); | |||
9652 | return true; | |||
9653 | }; | |||
9654 | BVTy ShuffleBuilder(Params...); | |||
9655 | ResTy Res = ResTy(); | |||
9656 | SmallVector<int> Mask; | |||
9657 | SmallVector<int> ExtractMask; | |||
9658 | std::optional<TargetTransformInfo::ShuffleKind> ExtractShuffle; | |||
9659 | std::optional<TargetTransformInfo::ShuffleKind> GatherShuffle; | |||
9660 | SmallVector<const TreeEntry *> Entries; | |||
9661 | Type *ScalarTy = GatheredScalars.front()->getType(); | |||
9662 | if (!all_of(GatheredScalars, UndefValue::classof)) { | |||
9663 | // Check for gathered extracts. | |||
9664 | ExtractShuffle = tryToGatherExtractElements(GatheredScalars, ExtractMask); | |||
9665 | SmallVector<Value *> IgnoredVals; | |||
9666 | if (UserIgnoreList) | |||
9667 | IgnoredVals.assign(UserIgnoreList->begin(), UserIgnoreList->end()); | |||
9668 | bool Resized = false; | |||
9669 | if (Value *VecBase = ShuffleBuilder.adjustExtracts(E, ExtractMask)) | |||
9670 | if (auto *VecBaseTy = dyn_cast<FixedVectorType>(VecBase->getType())) | |||
9671 | if (VF == VecBaseTy->getNumElements() && GatheredScalars.size() != VF) { | |||
9672 | Resized = true; | |||
9673 | GatheredScalars.append(VF - GatheredScalars.size(), | |||
9674 | PoisonValue::get(ScalarTy)); | |||
9675 | } | |||
9676 | // Gather extracts after we check for full matched gathers only. | |||
9677 | if (ExtractShuffle || E->getOpcode() != Instruction::Load || | |||
9678 | E->isAltShuffle() || | |||
9679 | all_of(E->Scalars, [this](Value *V) { return getTreeEntry(V); }) || | |||
9680 | isSplat(E->Scalars) || | |||
9681 | (E->Scalars != GatheredScalars && GatheredScalars.size() <= 2)) { | |||
9682 | GatherShuffle = isGatherShuffledEntry(E, GatheredScalars, Mask, Entries); | |||
9683 | } | |||
9684 | if (GatherShuffle) { | |||
9685 | if (Value *Delayed = ShuffleBuilder.needToDelay(E, Entries)) { | |||
9686 | // Delay emission of gathers which are not ready yet. | |||
9687 | PostponedGathers.insert(E); | |||
9688 | // Postpone gather emission, will be emitted after the end of the | |||
9689 | // process to keep correct order. | |||
9690 | return Delayed; | |||
9691 | } | |||
9692 | 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", 9693, __extension__ __PRETTY_FUNCTION__)) | |||
9693 | "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", 9693, __extension__ __PRETTY_FUNCTION__)); | |||
9694 | if (*GatherShuffle == TTI::SK_PermuteSingleSrc && | |||
9695 | Entries.front()->isSame(E->Scalars)) { | |||
9696 | // Perfect match in the graph, will reuse the previously vectorized | |||
9697 | // node. Cost is 0. | |||
9698 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: perfect diamond match for gather bundle that starts with " << *E->Scalars.front() << ".\n"; } } while (false ) | |||
9699 | dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: perfect diamond match for gather bundle that starts with " << *E->Scalars.front() << ".\n"; } } while (false ) | |||
9700 | << "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 " << *E->Scalars.front() << ".\n"; } } while (false ) | |||
9701 | << *E->Scalars.front() << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: perfect diamond match for gather bundle that starts with " << *E->Scalars.front() << ".\n"; } } while (false ); | |||
9702 | // Restore the mask for previous partially matched values. | |||
9703 | for (auto [I, V] : enumerate(E->Scalars)) { | |||
9704 | if (isa<PoisonValue>(V)) { | |||
9705 | Mask[I] = PoisonMaskElem; | |||
9706 | continue; | |||
9707 | } | |||
9708 | if (Mask[I] == PoisonMaskElem) | |||
9709 | Mask[I] = Entries.front()->findLaneForValue(V); | |||
9710 | } | |||
9711 | ShuffleBuilder.add(Entries.front()->VectorizedValue, Mask); | |||
9712 | Res = ShuffleBuilder.finalize(E->ReuseShuffleIndices); | |||
9713 | return Res; | |||
9714 | } | |||
9715 | if (!Resized) { | |||
9716 | unsigned VF1 = Entries.front()->getVectorFactor(); | |||
9717 | unsigned VF2 = Entries.back()->getVectorFactor(); | |||
9718 | if ((VF == VF1 || VF == VF2) && GatheredScalars.size() != VF) | |||
9719 | GatheredScalars.append(VF - GatheredScalars.size(), | |||
9720 | PoisonValue::get(ScalarTy)); | |||
9721 | } | |||
9722 | // Remove shuffled elements from list of gathers. | |||
9723 | for (int I = 0, Sz = Mask.size(); I < Sz; ++I) { | |||
9724 | if (Mask[I] != PoisonMaskElem) | |||
9725 | GatheredScalars[I] = PoisonValue::get(ScalarTy); | |||
9726 | } | |||
9727 | } | |||
9728 | } | |||
9729 | auto TryPackScalars = [&](SmallVectorImpl<Value *> &Scalars, | |||
9730 | SmallVectorImpl<int> &ReuseMask, | |||
9731 | bool IsRootPoison) { | |||
9732 | // For splats with can emit broadcasts instead of gathers, so try to find | |||
9733 | // such sequences. | |||
9734 | bool IsSplat = IsRootPoison && isSplat(Scalars) && | |||
9735 | (Scalars.size() > 2 || Scalars.front() == Scalars.back()); | |||
9736 | Scalars.append(VF - Scalars.size(), PoisonValue::get(ScalarTy)); | |||
9737 | SmallVector<int> UndefPos; | |||
9738 | DenseMap<Value *, unsigned> UniquePositions; | |||
9739 | // Gather unique non-const values and all constant values. | |||
9740 | // For repeated values, just shuffle them. | |||
9741 | int NumNonConsts = 0; | |||
9742 | int SinglePos = 0; | |||
9743 | for (auto [I, V] : enumerate(Scalars)) { | |||
9744 | if (isa<UndefValue>(V)) { | |||
9745 | if (!isa<PoisonValue>(V)) { | |||
9746 | ReuseMask[I] = I; | |||
9747 | UndefPos.push_back(I); | |||
9748 | } | |||
9749 | continue; | |||
9750 | } | |||
9751 | if (isConstant(V)) { | |||
9752 | ReuseMask[I] = I; | |||
9753 | continue; | |||
9754 | } | |||
9755 | ++NumNonConsts; | |||
9756 | SinglePos = I; | |||
9757 | Value *OrigV = V; | |||
9758 | Scalars[I] = PoisonValue::get(ScalarTy); | |||
9759 | if (IsSplat) { | |||
9760 | Scalars.front() = OrigV; | |||
9761 | ReuseMask[I] = 0; | |||
9762 | } else { | |||
9763 | const auto Res = UniquePositions.try_emplace(OrigV, I); | |||
9764 | Scalars[Res.first->second] = OrigV; | |||
9765 | ReuseMask[I] = Res.first->second; | |||
9766 | } | |||
9767 | } | |||
9768 | if (NumNonConsts == 1) { | |||
9769 | // Restore single insert element. | |||
9770 | if (IsSplat) { | |||
9771 | ReuseMask.assign(VF, PoisonMaskElem); | |||
9772 | std::swap(Scalars.front(), Scalars[SinglePos]); | |||
9773 | if (!UndefPos.empty() && UndefPos.front() == 0) | |||
9774 | Scalars.front() = UndefValue::get(ScalarTy); | |||
9775 | } | |||
9776 | ReuseMask[SinglePos] = SinglePos; | |||
9777 | } else if (!UndefPos.empty() && IsSplat) { | |||
9778 | // For undef values, try to replace them with the simple broadcast. | |||
9779 | // We can do it if the broadcasted value is guaranteed to be | |||
9780 | // non-poisonous, or by freezing the incoming scalar value first. | |||
9781 | auto *It = find_if(Scalars, [this, E](Value *V) { | |||
9782 | return !isa<UndefValue>(V) && | |||
9783 | (getTreeEntry(V) || isGuaranteedNotToBePoison(V) || | |||
9784 | (E->UserTreeIndices.size() == 1 && | |||
9785 | any_of(V->uses(), [E](const Use &U) { | |||
9786 | // Check if the value already used in the same operation in | |||
9787 | // one of the nodes already. | |||
9788 | return E->UserTreeIndices.front().EdgeIdx != | |||
9789 | U.getOperandNo() && | |||
9790 | is_contained( | |||
9791 | E->UserTreeIndices.front().UserTE->Scalars, | |||
9792 | U.getUser()); | |||
9793 | }))); | |||
9794 | }); | |||
9795 | if (It != Scalars.end()) { | |||
9796 | // Replace undefs by the non-poisoned scalars and emit broadcast. | |||
9797 | int Pos = std::distance(Scalars.begin(), It); | |||
9798 | for_each(UndefPos, [&](int I) { | |||
9799 | // Set the undef position to the non-poisoned scalar. | |||
9800 | ReuseMask[I] = Pos; | |||
9801 | // Replace the undef by the poison, in the mask it is replaced by | |||
9802 | // non-poisoned scalar already. | |||
9803 | if (I != Pos) | |||
9804 | Scalars[I] = PoisonValue::get(ScalarTy); | |||
9805 | }); | |||
9806 | } else { | |||
9807 | // Replace undefs by the poisons, emit broadcast and then emit | |||
9808 | // freeze. | |||
9809 | for_each(UndefPos, [&](int I) { | |||
9810 | ReuseMask[I] = PoisonMaskElem; | |||
9811 | if (isa<UndefValue>(Scalars[I])) | |||
9812 | Scalars[I] = PoisonValue::get(ScalarTy); | |||
9813 | }); | |||
9814 | NeedFreeze = true; | |||
9815 | } | |||
9816 | } | |||
9817 | }; | |||
9818 | if (ExtractShuffle || GatherShuffle) { | |||
9819 | bool IsNonPoisoned = true; | |||
9820 | bool IsUsedInExpr = false; | |||
9821 | Value *Vec1 = nullptr; | |||
9822 | if (ExtractShuffle) { | |||
9823 | // Gather of extractelements can be represented as just a shuffle of | |||
9824 | // a single/two vectors the scalars are extracted from. | |||
9825 | // Find input vectors. | |||
9826 | Value *Vec2 = nullptr; | |||
9827 | for (unsigned I = 0, Sz = ExtractMask.size(); I < Sz; ++I) { | |||
9828 | if (ExtractMask[I] == PoisonMaskElem || | |||
9829 | (!Mask.empty() && Mask[I] != PoisonMaskElem)) { | |||
9830 | ExtractMask[I] = PoisonMaskElem; | |||
9831 | continue; | |||
9832 | } | |||
9833 | if (isa<UndefValue>(E->Scalars[I])) | |||
9834 | continue; | |||
9835 | auto *EI = cast<ExtractElementInst>(E->Scalars[I]); | |||
9836 | if (!Vec1) { | |||
9837 | Vec1 = EI->getVectorOperand(); | |||
9838 | } else if (Vec1 != EI->getVectorOperand()) { | |||
9839 | assert((!Vec2 || Vec2 == EI->getVectorOperand()) &&(static_cast <bool> ((!Vec2 || Vec2 == EI->getVectorOperand ()) && "Expected only 1 or 2 vectors shuffle.") ? void (0) : __assert_fail ("(!Vec2 || Vec2 == EI->getVectorOperand()) && \"Expected only 1 or 2 vectors shuffle.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 9840, __extension__ __PRETTY_FUNCTION__)) | |||
9840 | "Expected only 1 or 2 vectors shuffle.")(static_cast <bool> ((!Vec2 || Vec2 == EI->getVectorOperand ()) && "Expected only 1 or 2 vectors shuffle.") ? void (0) : __assert_fail ("(!Vec2 || Vec2 == EI->getVectorOperand()) && \"Expected only 1 or 2 vectors shuffle.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 9840, __extension__ __PRETTY_FUNCTION__)); | |||
9841 | Vec2 = EI->getVectorOperand(); | |||
9842 | } | |||
9843 | } | |||
9844 | if (Vec2) { | |||
9845 | IsNonPoisoned &= | |||
9846 | isGuaranteedNotToBePoison(Vec1) && isGuaranteedNotToBePoison(Vec2); | |||
9847 | ShuffleBuilder.add(Vec1, Vec2, ExtractMask); | |||
9848 | } else if (Vec1) { | |||
9849 | IsUsedInExpr = FindReusedSplat(ExtractMask); | |||
9850 | ShuffleBuilder.add(Vec1, ExtractMask); | |||
9851 | IsNonPoisoned &= isGuaranteedNotToBePoison(Vec1); | |||
9852 | } else { | |||
9853 | ShuffleBuilder.add(PoisonValue::get(FixedVectorType::get( | |||
9854 | ScalarTy, GatheredScalars.size())), | |||
9855 | ExtractMask); | |||
9856 | } | |||
9857 | } | |||
9858 | if (GatherShuffle) { | |||
9859 | if (Entries.size() == 1) { | |||
9860 | IsUsedInExpr = FindReusedSplat(Mask); | |||
9861 | ShuffleBuilder.add(Entries.front()->VectorizedValue, Mask); | |||
9862 | IsNonPoisoned &= | |||
9863 | isGuaranteedNotToBePoison(Entries.front()->VectorizedValue); | |||
9864 | } else { | |||
9865 | ShuffleBuilder.add(Entries.front()->VectorizedValue, | |||
9866 | Entries.back()->VectorizedValue, Mask); | |||
9867 | IsNonPoisoned &= | |||
9868 | isGuaranteedNotToBePoison(Entries.front()->VectorizedValue) && | |||
9869 | isGuaranteedNotToBePoison(Entries.back()->VectorizedValue); | |||
9870 | } | |||
9871 | } | |||
9872 | // Try to figure out best way to combine values: build a shuffle and insert | |||
9873 | // elements or just build several shuffles. | |||
9874 | // Insert non-constant scalars. | |||
9875 | SmallVector<Value *> NonConstants(GatheredScalars); | |||
9876 | int EMSz = ExtractMask.size(); | |||
9877 | int MSz = Mask.size(); | |||
9878 | // Try to build constant vector and shuffle with it only if currently we | |||
9879 | // have a single permutation and more than 1 scalar constants. | |||
9880 | bool IsSingleShuffle = !ExtractShuffle || !GatherShuffle; | |||
9881 | bool IsIdentityShuffle = | |||
9882 | (ExtractShuffle.value_or(TTI::SK_PermuteTwoSrc) == | |||
9883 | TTI::SK_PermuteSingleSrc && | |||
9884 | none_of(ExtractMask, [&](int I) { return I >= EMSz; }) && | |||
9885 | ShuffleVectorInst::isIdentityMask(ExtractMask)) || | |||
9886 | (GatherShuffle.value_or(TTI::SK_PermuteTwoSrc) == | |||
9887 | TTI::SK_PermuteSingleSrc && | |||
9888 | none_of(Mask, [&](int I) { return I >= MSz; }) && | |||
9889 | ShuffleVectorInst::isIdentityMask(Mask)); | |||
9890 | bool EnoughConstsForShuffle = | |||
9891 | IsSingleShuffle && | |||
9892 | (none_of(GatheredScalars, | |||
9893 | [](Value *V) { | |||
9894 | return isa<UndefValue>(V) && !isa<PoisonValue>(V); | |||
9895 | }) || | |||
9896 | any_of(GatheredScalars, | |||
9897 | [](Value *V) { | |||
9898 | return isa<Constant>(V) && !isa<UndefValue>(V); | |||
9899 | })) && | |||
9900 | (!IsIdentityShuffle || | |||
9901 | (GatheredScalars.size() == 2 && | |||
9902 | any_of(GatheredScalars, | |||
9903 | [](Value *V) { return !isa<UndefValue>(V); })) || | |||
9904 | count_if(GatheredScalars, [](Value *V) { | |||
9905 | return isa<Constant>(V) && !isa<PoisonValue>(V); | |||
9906 | }) > 1); | |||
9907 | // NonConstants array contains just non-constant values, GatheredScalars | |||
9908 | // contains only constant to build final vector and then shuffle. | |||
9909 | for (int I = 0, Sz = GatheredScalars.size(); I < Sz; ++I) { | |||
9910 | if (EnoughConstsForShuffle && isa<Constant>(GatheredScalars[I])) | |||
9911 | NonConstants[I] = PoisonValue::get(ScalarTy); | |||
9912 | else | |||
9913 | GatheredScalars[I] = PoisonValue::get(ScalarTy); | |||
9914 | } | |||
9915 | // Generate constants for final shuffle and build a mask for them. | |||
9916 | if (!all_of(GatheredScalars, PoisonValue::classof)) { | |||
9917 | SmallVector<int> BVMask(GatheredScalars.size(), PoisonMaskElem); | |||
9918 | TryPackScalars(GatheredScalars, BVMask, /*IsRootPoison=*/true); | |||
9919 | Value *BV = ShuffleBuilder.gather(GatheredScalars); | |||
9920 | ShuffleBuilder.add(BV, BVMask); | |||
9921 | } | |||
9922 | if (all_of(NonConstants, [=](Value *V) { | |||
9923 | return isa<PoisonValue>(V) || | |||
9924 | (IsSingleShuffle && ((IsIdentityShuffle && | |||
9925 | IsNonPoisoned) || IsUsedInExpr) && isa<UndefValue>(V)); | |||
9926 | })) | |||
9927 | Res = ShuffleBuilder.finalize(E->ReuseShuffleIndices); | |||
9928 | else | |||
9929 | Res = ShuffleBuilder.finalize( | |||
9930 | E->ReuseShuffleIndices, E->Scalars.size(), | |||
9931 | [&](Value *&Vec, SmallVectorImpl<int> &Mask) { | |||
9932 | TryPackScalars(NonConstants, Mask, /*IsRootPoison=*/false); | |||
9933 | Vec = ShuffleBuilder.gather(NonConstants, Vec); | |||
9934 | }); | |||
9935 | } else if (!allConstant(GatheredScalars)) { | |||
9936 | // Gather unique scalars and all constants. | |||
9937 | SmallVector<int> ReuseMask(GatheredScalars.size(), PoisonMaskElem); | |||
9938 | TryPackScalars(GatheredScalars, ReuseMask, /*IsRootPoison=*/true); | |||
9939 | Value *BV = ShuffleBuilder.gather(GatheredScalars); | |||
9940 | ShuffleBuilder.add(BV, ReuseMask); | |||
9941 | Res = ShuffleBuilder.finalize(E->ReuseShuffleIndices); | |||
9942 | } else { | |||
9943 | // Gather all constants. | |||
9944 | SmallVector<int> Mask(E->Scalars.size(), PoisonMaskElem); | |||
9945 | for (auto [I, V] : enumerate(E->Scalars)) { | |||
9946 | if (!isa<PoisonValue>(V)) | |||
9947 | Mask[I] = I; | |||
9948 | } | |||
9949 | Value *BV = ShuffleBuilder.gather(E->Scalars); | |||
9950 | ShuffleBuilder.add(BV, Mask); | |||
9951 | Res = ShuffleBuilder.finalize(E->ReuseShuffleIndices); | |||
9952 | } | |||
9953 | ||||
9954 | if (NeedFreeze) | |||
9955 | Res = ShuffleBuilder.createFreeze(Res); | |||
9956 | return Res; | |||
9957 | } | |||
9958 | ||||
9959 | Value *BoUpSLP::createBuildVector(const TreeEntry *E) { | |||
9960 | return processBuildVector<ShuffleInstructionBuilder, Value *>(E, Builder, | |||
9961 | *this); | |||
9962 | } | |||
9963 | ||||
9964 | Value *BoUpSLP::vectorizeTree(TreeEntry *E) { | |||
9965 | IRBuilder<>::InsertPointGuard Guard(Builder); | |||
9966 | ||||
9967 | if (E->VectorizedValue) { | |||
9968 | 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); | |||
9969 | return E->VectorizedValue; | |||
9970 | } | |||
9971 | ||||
9972 | if (E->State == TreeEntry::NeedToGather) { | |||
9973 | if (E->getMainOp() && E->Idx == 0) | |||
9974 | setInsertPointAfterBundle(E); | |||
9975 | Value *Vec = createBuildVector(E); | |||
9976 | E->VectorizedValue = Vec; | |||
9977 | return Vec; | |||
9978 | } | |||
9979 | ||||
9980 | auto FinalShuffle = [&](Value *V, const TreeEntry *E) { | |||
9981 | ShuffleInstructionBuilder ShuffleBuilder(Builder, *this); | |||
9982 | if (E->getOpcode() == Instruction::Store) { | |||
9983 | ArrayRef<int> Mask = | |||
9984 | ArrayRef(reinterpret_cast<const int *>(E->ReorderIndices.begin()), | |||
9985 | E->ReorderIndices.size()); | |||
9986 | ShuffleBuilder.add(V, Mask); | |||
9987 | } else { | |||
9988 | ShuffleBuilder.addOrdered(V, E->ReorderIndices); | |||
9989 | } | |||
9990 | return ShuffleBuilder.finalize(E->ReuseShuffleIndices); | |||
9991 | }; | |||
9992 | ||||
9993 | 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", 9995, __extension__ __PRETTY_FUNCTION__)) | |||
9994 | 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", 9995, __extension__ __PRETTY_FUNCTION__)) | |||
9995 | "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", 9995, __extension__ __PRETTY_FUNCTION__)); | |||
9996 | unsigned ShuffleOrOp = | |||
9997 | E->isAltShuffle() ? (unsigned)Instruction::ShuffleVector : E->getOpcode(); | |||
9998 | Instruction *VL0 = E->getMainOp(); | |||
9999 | Type *ScalarTy = VL0->getType(); | |||
10000 | if (auto *Store = dyn_cast<StoreInst>(VL0)) | |||
10001 | ScalarTy = Store->getValueOperand()->getType(); | |||
10002 | else if (auto *IE = dyn_cast<InsertElementInst>(VL0)) | |||
10003 | ScalarTy = IE->getOperand(1)->getType(); | |||
10004 | auto *VecTy = FixedVectorType::get(ScalarTy, E->Scalars.size()); | |||
10005 | switch (ShuffleOrOp) { | |||
10006 | case Instruction::PHI: { | |||
10007 | assert((E->ReorderIndices.empty() ||(static_cast <bool> ((E->ReorderIndices.empty() || E != VectorizableTree.front().get() || !E->UserTreeIndices. empty()) && "PHI reordering is free.") ? void (0) : __assert_fail ("(E->ReorderIndices.empty() || E != VectorizableTree.front().get() || !E->UserTreeIndices.empty()) && \"PHI reordering is free.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 10010, __extension__ __PRETTY_FUNCTION__)) | |||
10008 | E != VectorizableTree.front().get() ||(static_cast <bool> ((E->ReorderIndices.empty() || E != VectorizableTree.front().get() || !E->UserTreeIndices. empty()) && "PHI reordering is free.") ? void (0) : __assert_fail ("(E->ReorderIndices.empty() || E != VectorizableTree.front().get() || !E->UserTreeIndices.empty()) && \"PHI reordering is free.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 10010, __extension__ __PRETTY_FUNCTION__)) | |||
10009 | !E->UserTreeIndices.empty()) &&(static_cast <bool> ((E->ReorderIndices.empty() || E != VectorizableTree.front().get() || !E->UserTreeIndices. empty()) && "PHI reordering is free.") ? void (0) : __assert_fail ("(E->ReorderIndices.empty() || E != VectorizableTree.front().get() || !E->UserTreeIndices.empty()) && \"PHI reordering is free.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 10010, __extension__ __PRETTY_FUNCTION__)) | |||
10010 | "PHI reordering is free.")(static_cast <bool> ((E->ReorderIndices.empty() || E != VectorizableTree.front().get() || !E->UserTreeIndices. empty()) && "PHI reordering is free.") ? void (0) : __assert_fail ("(E->ReorderIndices.empty() || E != VectorizableTree.front().get() || !E->UserTreeIndices.empty()) && \"PHI reordering is free.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 10010, __extension__ __PRETTY_FUNCTION__)); | |||
10011 | auto *PH = cast<PHINode>(VL0); | |||
10012 | Builder.SetInsertPoint(PH->getParent()->getFirstNonPHI()); | |||
10013 | Builder.SetCurrentDebugLocation(PH->getDebugLoc()); | |||
10014 | PHINode *NewPhi = Builder.CreatePHI(VecTy, PH->getNumIncomingValues()); | |||
10015 | Value *V = NewPhi; | |||
10016 | ||||
10017 | // Adjust insertion point once all PHI's have been generated. | |||
10018 | Builder.SetInsertPoint(&*PH->getParent()->getFirstInsertionPt()); | |||
10019 | Builder.SetCurrentDebugLocation(PH->getDebugLoc()); | |||
10020 | ||||
10021 | V = FinalShuffle(V, E); | |||
10022 | ||||
10023 | E->VectorizedValue = V; | |||
10024 | ||||
10025 | // PHINodes may have multiple entries from the same block. We want to | |||
10026 | // visit every block once. | |||
10027 | SmallPtrSet<BasicBlock*, 4> VisitedBBs; | |||
10028 | ||||
10029 | for (unsigned i = 0, e = PH->getNumIncomingValues(); i < e; ++i) { | |||
10030 | ValueList Operands; | |||
10031 | BasicBlock *IBB = PH->getIncomingBlock(i); | |||
10032 | ||||
10033 | // Stop emission if all incoming values are generated. | |||
10034 | if (NewPhi->getNumIncomingValues() == PH->getNumIncomingValues()) { | |||
10035 | 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); | |||
10036 | return V; | |||
10037 | } | |||
10038 | ||||
10039 | if (!VisitedBBs.insert(IBB).second) { | |||
10040 | NewPhi->addIncoming(NewPhi->getIncomingValueForBlock(IBB), IBB); | |||
10041 | continue; | |||
10042 | } | |||
10043 | ||||
10044 | Builder.SetInsertPoint(IBB->getTerminator()); | |||
10045 | Builder.SetCurrentDebugLocation(PH->getDebugLoc()); | |||
10046 | Value *Vec = vectorizeOperand(E, i); | |||
10047 | NewPhi->addIncoming(Vec, IBB); | |||
10048 | } | |||
10049 | ||||
10050 | 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", 10051, __extension__ __PRETTY_FUNCTION__)) | |||
10051 | "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", 10051, __extension__ __PRETTY_FUNCTION__)); | |||
10052 | return V; | |||
10053 | } | |||
10054 | ||||
10055 | case Instruction::ExtractElement: { | |||
10056 | Value *V = E->getSingleOperand(0); | |||
10057 | setInsertPointAfterBundle(E); | |||
10058 | V = FinalShuffle(V, E); | |||
10059 | E->VectorizedValue = V; | |||
10060 | return V; | |||
10061 | } | |||
10062 | case Instruction::ExtractValue: { | |||
10063 | auto *LI = cast<LoadInst>(E->getSingleOperand(0)); | |||
10064 | Builder.SetInsertPoint(LI); | |||
10065 | auto *PtrTy = PointerType::get(VecTy, LI->getPointerAddressSpace()); | |||
10066 | Value *Ptr = Builder.CreateBitCast(LI->getOperand(0), PtrTy); | |||
10067 | LoadInst *V = Builder.CreateAlignedLoad(VecTy, Ptr, LI->getAlign()); | |||
10068 | Value *NewV = propagateMetadata(V, E->Scalars); | |||
10069 | NewV = FinalShuffle(NewV, E); | |||
10070 | E->VectorizedValue = NewV; | |||
10071 | return NewV; | |||
10072 | } | |||
10073 | case Instruction::InsertElement: { | |||
10074 | 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", 10074, __extension__ __PRETTY_FUNCTION__)); | |||
10075 | Builder.SetInsertPoint(cast<Instruction>(E->Scalars.back())); | |||
10076 | Value *V = vectorizeOperand(E, 1); | |||
10077 | ||||
10078 | // Create InsertVector shuffle if necessary | |||
10079 | auto *FirstInsert = cast<Instruction>(*find_if(E->Scalars, [E](Value *V) { | |||
10080 | return !is_contained(E->Scalars, cast<Instruction>(V)->getOperand(0)); | |||
10081 | })); | |||
10082 | const unsigned NumElts = | |||
10083 | cast<FixedVectorType>(FirstInsert->getType())->getNumElements(); | |||
10084 | const unsigned NumScalars = E->Scalars.size(); | |||
10085 | ||||
10086 | unsigned Offset = *getInsertIndex(VL0); | |||
10087 | 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", 10087, __extension__ __PRETTY_FUNCTION__)); | |||
10088 | ||||
10089 | // Create shuffle to resize vector | |||
10090 | SmallVector<int> Mask; | |||
10091 | if (!E->ReorderIndices.empty()) { | |||
10092 | inversePermutation(E->ReorderIndices, Mask); | |||
10093 | Mask.append(NumElts - NumScalars, PoisonMaskElem); | |||
10094 | } else { | |||
10095 | Mask.assign(NumElts, PoisonMaskElem); | |||
10096 | std::iota(Mask.begin(), std::next(Mask.begin(), NumScalars), 0); | |||
10097 | } | |||
10098 | // Create InsertVector shuffle if necessary | |||
10099 | bool IsIdentity = true; | |||
10100 | SmallVector<int> PrevMask(NumElts, PoisonMaskElem); | |||
10101 | Mask.swap(PrevMask); | |||
10102 | for (unsigned I = 0; I < NumScalars; ++I) { | |||
10103 | Value *Scalar = E->Scalars[PrevMask[I]]; | |||
10104 | unsigned InsertIdx = *getInsertIndex(Scalar); | |||
10105 | IsIdentity &= InsertIdx - Offset == I; | |||
10106 | Mask[InsertIdx - Offset] = I; | |||
10107 | } | |||
10108 | if (!IsIdentity || NumElts != NumScalars) { | |||
10109 | V = Builder.CreateShuffleVector(V, Mask); | |||
10110 | if (auto *I = dyn_cast<Instruction>(V)) { | |||
10111 | GatherShuffleExtractSeq.insert(I); | |||
10112 | CSEBlocks.insert(I->getParent()); | |||
10113 | } | |||
10114 | } | |||
10115 | ||||
10116 | SmallVector<int> InsertMask(NumElts, PoisonMaskElem); | |||
10117 | for (unsigned I = 0; I < NumElts; I++) { | |||
10118 | if (Mask[I] != PoisonMaskElem) | |||
10119 | InsertMask[Offset + I] = I; | |||
10120 | } | |||
10121 | SmallBitVector UseMask = | |||
10122 | buildUseMask(NumElts, InsertMask, UseMask::UndefsAsMask); | |||
10123 | SmallBitVector IsFirstUndef = | |||
10124 | isUndefVector(FirstInsert->getOperand(0), UseMask); | |||
10125 | if ((!IsIdentity || Offset != 0 || !IsFirstUndef.all()) && | |||
10126 | NumElts != NumScalars) { | |||
10127 | if (IsFirstUndef.all()) { | |||
10128 | if (!ShuffleVectorInst::isIdentityMask(InsertMask)) { | |||
10129 | SmallBitVector IsFirstPoison = | |||
10130 | isUndefVector<true>(FirstInsert->getOperand(0), UseMask); | |||
10131 | if (!IsFirstPoison.all()) { | |||
10132 | for (unsigned I = 0; I < NumElts; I++) { | |||
10133 | if (InsertMask[I] == PoisonMaskElem && !IsFirstPoison.test(I)) | |||
10134 | InsertMask[I] = I + NumElts; | |||
10135 | } | |||
10136 | } | |||
10137 | V = Builder.CreateShuffleVector( | |||
10138 | V, | |||
10139 | IsFirstPoison.all() ? PoisonValue::get(V->getType()) | |||
10140 | : FirstInsert->getOperand(0), | |||
10141 | InsertMask, cast<Instruction>(E->Scalars.back())->getName()); | |||
10142 | if (auto *I = dyn_cast<Instruction>(V)) { | |||
10143 | GatherShuffleExtractSeq.insert(I); | |||
10144 | CSEBlocks.insert(I->getParent()); | |||
10145 | } | |||
10146 | } | |||
10147 | } else { | |||
10148 | SmallBitVector IsFirstPoison = | |||
10149 | isUndefVector<true>(FirstInsert->getOperand(0), UseMask); | |||
10150 | for (unsigned I = 0; I < NumElts; I++) { | |||
10151 | if (InsertMask[I] == PoisonMaskElem) | |||
10152 | InsertMask[I] = IsFirstPoison.test(I) ? PoisonMaskElem : I; | |||
10153 | else | |||
10154 | InsertMask[I] += NumElts; | |||
10155 | } | |||
10156 | V = Builder.CreateShuffleVector( | |||
10157 | FirstInsert->getOperand(0), V, InsertMask, | |||
10158 | cast<Instruction>(E->Scalars.back())->getName()); | |||
10159 | if (auto *I = dyn_cast<Instruction>(V)) { | |||
10160 | GatherShuffleExtractSeq.insert(I); | |||
10161 | CSEBlocks.insert(I->getParent()); | |||
10162 | } | |||
10163 | } | |||
10164 | } | |||
10165 | ||||
10166 | ++NumVectorInstructions; | |||
10167 | E->VectorizedValue = V; | |||
10168 | return V; | |||
10169 | } | |||
10170 | case Instruction::ZExt: | |||
10171 | case Instruction::SExt: | |||
10172 | case Instruction::FPToUI: | |||
10173 | case Instruction::FPToSI: | |||
10174 | case Instruction::FPExt: | |||
10175 | case Instruction::PtrToInt: | |||
10176 | case Instruction::IntToPtr: | |||
10177 | case Instruction::SIToFP: | |||
10178 | case Instruction::UIToFP: | |||
10179 | case Instruction::Trunc: | |||
10180 | case Instruction::FPTrunc: | |||
10181 | case Instruction::BitCast: { | |||
10182 | setInsertPointAfterBundle(E); | |||
10183 | ||||
10184 | Value *InVec = vectorizeOperand(E, 0); | |||
10185 | if (E->VectorizedValue) { | |||
10186 | 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); | |||
10187 | return E->VectorizedValue; | |||
10188 | } | |||
10189 | ||||
10190 | auto *CI = cast<CastInst>(VL0); | |||
10191 | Value *V = Builder.CreateCast(CI->getOpcode(), InVec, VecTy); | |||
10192 | V = FinalShuffle(V, E); | |||
10193 | ||||
10194 | E->VectorizedValue = V; | |||
10195 | ++NumVectorInstructions; | |||
10196 | return V; | |||
10197 | } | |||
10198 | case Instruction::FCmp: | |||
10199 | case Instruction::ICmp: { | |||
10200 | setInsertPointAfterBundle(E); | |||
10201 | ||||
10202 | Value *L = vectorizeOperand(E, 0); | |||
10203 | if (E->VectorizedValue) { | |||
10204 | 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); | |||
10205 | return E->VectorizedValue; | |||
10206 | } | |||
10207 | Value *R = vectorizeOperand(E, 1); | |||
10208 | if (E->VectorizedValue) { | |||
10209 | 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); | |||
10210 | return E->VectorizedValue; | |||
10211 | } | |||
10212 | ||||
10213 | CmpInst::Predicate P0 = cast<CmpInst>(VL0)->getPredicate(); | |||
10214 | Value *V = Builder.CreateCmp(P0, L, R); | |||
10215 | propagateIRFlags(V, E->Scalars, VL0); | |||
10216 | V = FinalShuffle(V, E); | |||
10217 | ||||
10218 | E->VectorizedValue = V; | |||
10219 | ++NumVectorInstructions; | |||
10220 | return V; | |||
10221 | } | |||
10222 | case Instruction::Select: { | |||
10223 | setInsertPointAfterBundle(E); | |||
10224 | ||||
10225 | Value *Cond = vectorizeOperand(E, 0); | |||
10226 | if (E->VectorizedValue) { | |||
10227 | 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); | |||
10228 | return E->VectorizedValue; | |||
10229 | } | |||
10230 | Value *True = vectorizeOperand(E, 1); | |||
10231 | if (E->VectorizedValue) { | |||
10232 | 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); | |||
10233 | return E->VectorizedValue; | |||
10234 | } | |||
10235 | Value *False = vectorizeOperand(E, 2); | |||
10236 | if (E->VectorizedValue) { | |||
10237 | 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); | |||
10238 | return E->VectorizedValue; | |||
10239 | } | |||
10240 | ||||
10241 | Value *V = Builder.CreateSelect(Cond, True, False); | |||
10242 | V = FinalShuffle(V, E); | |||
10243 | ||||
10244 | E->VectorizedValue = V; | |||
10245 | ++NumVectorInstructions; | |||
10246 | return V; | |||
10247 | } | |||
10248 | case Instruction::FNeg: { | |||
10249 | setInsertPointAfterBundle(E); | |||
10250 | ||||
10251 | Value *Op = vectorizeOperand(E, 0); | |||
10252 | ||||
10253 | if (E->VectorizedValue) { | |||
10254 | 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); | |||
10255 | return E->VectorizedValue; | |||
10256 | } | |||
10257 | ||||
10258 | Value *V = Builder.CreateUnOp( | |||
10259 | static_cast<Instruction::UnaryOps>(E->getOpcode()), Op); | |||
10260 | propagateIRFlags(V, E->Scalars, VL0); | |||
10261 | if (auto *I = dyn_cast<Instruction>(V)) | |||
10262 | V = propagateMetadata(I, E->Scalars); | |||
10263 | ||||
10264 | V = FinalShuffle(V, E); | |||
10265 | ||||
10266 | E->VectorizedValue = V; | |||
10267 | ++NumVectorInstructions; | |||
10268 | ||||
10269 | return V; | |||
10270 | } | |||
10271 | case Instruction::Add: | |||
10272 | case Instruction::FAdd: | |||
10273 | case Instruction::Sub: | |||
10274 | case Instruction::FSub: | |||
10275 | case Instruction::Mul: | |||
10276 | case Instruction::FMul: | |||
10277 | case Instruction::UDiv: | |||
10278 | case Instruction::SDiv: | |||
10279 | case Instruction::FDiv: | |||
10280 | case Instruction::URem: | |||
10281 | case Instruction::SRem: | |||
10282 | case Instruction::FRem: | |||
10283 | case Instruction::Shl: | |||
10284 | case Instruction::LShr: | |||
10285 | case Instruction::AShr: | |||
10286 | case Instruction::And: | |||
10287 | case Instruction::Or: | |||
10288 | case Instruction::Xor: { | |||
10289 | setInsertPointAfterBundle(E); | |||
10290 | ||||
10291 | Value *LHS = vectorizeOperand(E, 0); | |||
10292 | if (E->VectorizedValue) { | |||
10293 | 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); | |||
10294 | return E->VectorizedValue; | |||
10295 | } | |||
10296 | Value *RHS = vectorizeOperand(E, 1); | |||
10297 | if (E->VectorizedValue) { | |||
10298 | 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); | |||
10299 | return E->VectorizedValue; | |||
10300 | } | |||
10301 | ||||
10302 | Value *V = Builder.CreateBinOp( | |||
10303 | static_cast<Instruction::BinaryOps>(E->getOpcode()), LHS, | |||
10304 | RHS); | |||
10305 | propagateIRFlags(V, E->Scalars, VL0); | |||
10306 | if (auto *I = dyn_cast<Instruction>(V)) | |||
10307 | V = propagateMetadata(I, E->Scalars); | |||
10308 | ||||
10309 | V = FinalShuffle(V, E); | |||
10310 | ||||
10311 | E->VectorizedValue = V; | |||
10312 | ++NumVectorInstructions; | |||
10313 | ||||
10314 | return V; | |||
10315 | } | |||
10316 | case Instruction::Load: { | |||
10317 | // Loads are inserted at the head of the tree because we don't want to | |||
10318 | // sink them all the way down past store instructions. | |||
10319 | setInsertPointAfterBundle(E); | |||
10320 | ||||
10321 | LoadInst *LI = cast<LoadInst>(VL0); | |||
10322 | Instruction *NewLI; | |||
10323 | unsigned AS = LI->getPointerAddressSpace(); | |||
10324 | Value *PO = LI->getPointerOperand(); | |||
10325 | if (E->State == TreeEntry::Vectorize) { | |||
10326 | Value *VecPtr = Builder.CreateBitCast(PO, VecTy->getPointerTo(AS)); | |||
10327 | NewLI = Builder.CreateAlignedLoad(VecTy, VecPtr, LI->getAlign()); | |||
10328 | ||||
10329 | // The pointer operand uses an in-tree scalar so we add the new BitCast | |||
10330 | // or LoadInst to ExternalUses list to make sure that an extract will | |||
10331 | // be generated in the future. | |||
10332 | if (TreeEntry *Entry = getTreeEntry(PO)) { | |||
10333 | // Find which lane we need to extract. | |||
10334 | unsigned FoundLane = Entry->findLaneForValue(PO); | |||
10335 | ExternalUses.emplace_back( | |||
10336 | PO, PO != VecPtr ? cast<User>(VecPtr) : NewLI, FoundLane); | |||
10337 | } | |||
10338 | } else { | |||
10339 | 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", 10339, __extension__ __PRETTY_FUNCTION__)); | |||
10340 | Value *VecPtr = vectorizeOperand(E, 0); | |||
10341 | if (E->VectorizedValue) { | |||
10342 | 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); | |||
10343 | return E->VectorizedValue; | |||
10344 | } | |||
10345 | // Use the minimum alignment of the gathered loads. | |||
10346 | Align CommonAlignment = LI->getAlign(); | |||
10347 | for (Value *V : E->Scalars) | |||
10348 | CommonAlignment = | |||
10349 | std::min(CommonAlignment, cast<LoadInst>(V)->getAlign()); | |||
10350 | NewLI = Builder.CreateMaskedGather(VecTy, VecPtr, CommonAlignment); | |||
10351 | } | |||
10352 | Value *V = propagateMetadata(NewLI, E->Scalars); | |||
10353 | ||||
10354 | V = FinalShuffle(V, E); | |||
10355 | E->VectorizedValue = V; | |||
10356 | ++NumVectorInstructions; | |||
10357 | return V; | |||
10358 | } | |||
10359 | case Instruction::Store: { | |||
10360 | auto *SI = cast<StoreInst>(VL0); | |||
10361 | unsigned AS = SI->getPointerAddressSpace(); | |||
10362 | ||||
10363 | setInsertPointAfterBundle(E); | |||
10364 | ||||
10365 | Value *VecValue = vectorizeOperand(E, 0); | |||
10366 | VecValue = FinalShuffle(VecValue, E); | |||
10367 | ||||
10368 | Value *ScalarPtr = SI->getPointerOperand(); | |||
10369 | Value *VecPtr = Builder.CreateBitCast( | |||
10370 | ScalarPtr, VecValue->getType()->getPointerTo(AS)); | |||
10371 | StoreInst *ST = | |||
10372 | Builder.CreateAlignedStore(VecValue, VecPtr, SI->getAlign()); | |||
10373 | ||||
10374 | // The pointer operand uses an in-tree scalar, so add the new BitCast or | |||
10375 | // StoreInst to ExternalUses to make sure that an extract will be | |||
10376 | // generated in the future. | |||
10377 | if (TreeEntry *Entry = getTreeEntry(ScalarPtr)) { | |||
10378 | // Find which lane we need to extract. | |||
10379 | unsigned FoundLane = Entry->findLaneForValue(ScalarPtr); | |||
10380 | ExternalUses.push_back(ExternalUser( | |||
10381 | ScalarPtr, ScalarPtr != VecPtr ? cast<User>(VecPtr) : ST, | |||
10382 | FoundLane)); | |||
10383 | } | |||
10384 | ||||
10385 | Value *V = propagateMetadata(ST, E->Scalars); | |||
10386 | ||||
10387 | E->VectorizedValue = V; | |||
10388 | ++NumVectorInstructions; | |||
10389 | return V; | |||
10390 | } | |||
10391 | case Instruction::GetElementPtr: { | |||
10392 | auto *GEP0 = cast<GetElementPtrInst>(VL0); | |||
10393 | setInsertPointAfterBundle(E); | |||
10394 | ||||
10395 | Value *Op0 = vectorizeOperand(E, 0); | |||
10396 | if (E->VectorizedValue) { | |||
10397 | 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); | |||
10398 | return E->VectorizedValue; | |||
10399 | } | |||
10400 | ||||
10401 | SmallVector<Value *> OpVecs; | |||
10402 | for (int J = 1, N = GEP0->getNumOperands(); J < N; ++J) { | |||
10403 | Value *OpVec = vectorizeOperand(E, J); | |||
10404 | if (E->VectorizedValue) { | |||
10405 | 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); | |||
10406 | return E->VectorizedValue; | |||
10407 | } | |||
10408 | OpVecs.push_back(OpVec); | |||
10409 | } | |||
10410 | ||||
10411 | Value *V = Builder.CreateGEP(GEP0->getSourceElementType(), Op0, OpVecs); | |||
10412 | if (Instruction *I = dyn_cast<GetElementPtrInst>(V)) { | |||
10413 | SmallVector<Value *> GEPs; | |||
10414 | for (Value *V : E->Scalars) { | |||
10415 | if (isa<GetElementPtrInst>(V)) | |||
10416 | GEPs.push_back(V); | |||
10417 | } | |||
10418 | V = propagateMetadata(I, GEPs); | |||
10419 | } | |||
10420 | ||||
10421 | V = FinalShuffle(V, E); | |||
10422 | ||||
10423 | E->VectorizedValue = V; | |||
10424 | ++NumVectorInstructions; | |||
10425 | ||||
10426 | return V; | |||
10427 | } | |||
10428 | case Instruction::Call: { | |||
10429 | CallInst *CI = cast<CallInst>(VL0); | |||
10430 | setInsertPointAfterBundle(E); | |||
10431 | ||||
10432 | Intrinsic::ID IID = Intrinsic::not_intrinsic; | |||
10433 | if (Function *FI = CI->getCalledFunction()) | |||
10434 | IID = FI->getIntrinsicID(); | |||
10435 | ||||
10436 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); | |||
10437 | ||||
10438 | auto VecCallCosts = getVectorCallCosts(CI, VecTy, TTI, TLI); | |||
10439 | bool UseIntrinsic = ID != Intrinsic::not_intrinsic && | |||
10440 | VecCallCosts.first <= VecCallCosts.second; | |||
10441 | ||||
10442 | Value *ScalarArg = nullptr; | |||
10443 | std::vector<Value *> OpVecs; | |||
10444 | SmallVector<Type *, 2> TysForDecl; | |||
10445 | // Add return type if intrinsic is overloaded on it. | |||
10446 | if (isVectorIntrinsicWithOverloadTypeAtArg(IID, -1)) | |||
10447 | TysForDecl.push_back( | |||
10448 | FixedVectorType::get(CI->getType(), E->Scalars.size())); | |||
10449 | for (int j = 0, e = CI->arg_size(); j < e; ++j) { | |||
10450 | ValueList OpVL; | |||
10451 | // Some intrinsics have scalar arguments. This argument should not be | |||
10452 | // vectorized. | |||
10453 | if (UseIntrinsic && isVectorIntrinsicWithScalarOpAtArg(IID, j)) { | |||
10454 | CallInst *CEI = cast<CallInst>(VL0); | |||
10455 | ScalarArg = CEI->getArgOperand(j); | |||
10456 | OpVecs.push_back(CEI->getArgOperand(j)); | |||
10457 | if (isVectorIntrinsicWithOverloadTypeAtArg(IID, j)) | |||
10458 | TysForDecl.push_back(ScalarArg->getType()); | |||
10459 | continue; | |||
10460 | } | |||
10461 | ||||
10462 | Value *OpVec = vectorizeOperand(E, j); | |||
10463 | if (E->VectorizedValue) { | |||
10464 | 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); | |||
10465 | return E->VectorizedValue; | |||
10466 | } | |||
10467 | LLVM_DEBUG(dbgs() << "SLP: OpVec[" << j << "]: " << *OpVec << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: OpVec[" << j << "]: " << *OpVec << "\n"; } } while (false); | |||
10468 | OpVecs.push_back(OpVec); | |||
10469 | if (isVectorIntrinsicWithOverloadTypeAtArg(IID, j)) | |||
10470 | TysForDecl.push_back(OpVec->getType()); | |||
10471 | } | |||
10472 | ||||
10473 | Function *CF; | |||
10474 | if (!UseIntrinsic) { | |||
10475 | VFShape Shape = | |||
10476 | VFShape::get(*CI, ElementCount::getFixed(static_cast<unsigned>( | |||
10477 | VecTy->getNumElements())), | |||
10478 | false /*HasGlobalPred*/); | |||
10479 | CF = VFDatabase(*CI).getVectorizedFunction(Shape); | |||
10480 | } else { | |||
10481 | CF = Intrinsic::getDeclaration(F->getParent(), ID, TysForDecl); | |||
10482 | } | |||
10483 | ||||
10484 | SmallVector<OperandBundleDef, 1> OpBundles; | |||
10485 | CI->getOperandBundlesAsDefs(OpBundles); | |||
10486 | Value *V = Builder.CreateCall(CF, OpVecs, OpBundles); | |||
10487 | ||||
10488 | // The scalar argument uses an in-tree scalar so we add the new vectorized | |||
10489 | // call to ExternalUses list to make sure that an extract will be | |||
10490 | // generated in the future. | |||
10491 | if (ScalarArg) { | |||
10492 | if (TreeEntry *Entry = getTreeEntry(ScalarArg)) { | |||
10493 | // Find which lane we need to extract. | |||
10494 | unsigned FoundLane = Entry->findLaneForValue(ScalarArg); | |||
10495 | ExternalUses.push_back( | |||
10496 | ExternalUser(ScalarArg, cast<User>(V), FoundLane)); | |||
10497 | } | |||
10498 | } | |||
10499 | ||||
10500 | propagateIRFlags(V, E->Scalars, VL0); | |||
10501 | V = FinalShuffle(V, E); | |||
10502 | ||||
10503 | E->VectorizedValue = V; | |||
10504 | ++NumVectorInstructions; | |||
10505 | return V; | |||
10506 | } | |||
10507 | case Instruction::ShuffleVector: { | |||
10508 | 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", 10514, __extension__ __PRETTY_FUNCTION__)) | |||
10509 | ((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", 10514, __extension__ __PRETTY_FUNCTION__)) | |||
10510 | 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", 10514, __extension__ __PRETTY_FUNCTION__)) | |||
10511 | (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", 10514, __extension__ __PRETTY_FUNCTION__)) | |||
10512 | 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", 10514, __extension__ __PRETTY_FUNCTION__)) | |||
10513 | (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", 10514, __extension__ __PRETTY_FUNCTION__)) | |||
10514 | "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", 10514, __extension__ __PRETTY_FUNCTION__)); | |||
10515 | ||||
10516 | Value *LHS = nullptr, *RHS = nullptr; | |||
10517 | if (Instruction::isBinaryOp(E->getOpcode()) || isa<CmpInst>(VL0)) { | |||
10518 | setInsertPointAfterBundle(E); | |||
10519 | LHS = vectorizeOperand(E, 0); | |||
10520 | if (E->VectorizedValue) { | |||
10521 | 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); | |||
10522 | return E->VectorizedValue; | |||
10523 | } | |||
10524 | RHS = vectorizeOperand(E, 1); | |||
10525 | } else { | |||
10526 | setInsertPointAfterBundle(E); | |||
10527 | LHS = vectorizeOperand(E, 0); | |||
10528 | } | |||
10529 | if (E->VectorizedValue) { | |||
10530 | 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); | |||
10531 | return E->VectorizedValue; | |||
10532 | } | |||
10533 | ||||
10534 | Value *V0, *V1; | |||
10535 | if (Instruction::isBinaryOp(E->getOpcode())) { | |||
10536 | V0 = Builder.CreateBinOp( | |||
10537 | static_cast<Instruction::BinaryOps>(E->getOpcode()), LHS, RHS); | |||
10538 | V1 = Builder.CreateBinOp( | |||
10539 | static_cast<Instruction::BinaryOps>(E->getAltOpcode()), LHS, RHS); | |||
10540 | } else if (auto *CI0 = dyn_cast<CmpInst>(VL0)) { | |||
10541 | V0 = Builder.CreateCmp(CI0->getPredicate(), LHS, RHS); | |||
10542 | auto *AltCI = cast<CmpInst>(E->getAltOp()); | |||
10543 | CmpInst::Predicate AltPred = AltCI->getPredicate(); | |||
10544 | V1 = Builder.CreateCmp(AltPred, LHS, RHS); | |||
10545 | } else { | |||
10546 | V0 = Builder.CreateCast( | |||
10547 | static_cast<Instruction::CastOps>(E->getOpcode()), LHS, VecTy); | |||
10548 | V1 = Builder.CreateCast( | |||
10549 | static_cast<Instruction::CastOps>(E->getAltOpcode()), LHS, VecTy); | |||
10550 | } | |||
10551 | // Add V0 and V1 to later analysis to try to find and remove matching | |||
10552 | // instruction, if any. | |||
10553 | for (Value *V : {V0, V1}) { | |||
10554 | if (auto *I = dyn_cast<Instruction>(V)) { | |||
10555 | GatherShuffleExtractSeq.insert(I); | |||
10556 | CSEBlocks.insert(I->getParent()); | |||
10557 | } | |||
10558 | } | |||
10559 | ||||
10560 | // Create shuffle to take alternate operations from the vector. | |||
10561 | // Also, gather up main and alt scalar ops to propagate IR flags to | |||
10562 | // each vector operation. | |||
10563 | ValueList OpScalars, AltScalars; | |||
10564 | SmallVector<int> Mask; | |||
10565 | buildShuffleEntryMask( | |||
10566 | E->Scalars, E->ReorderIndices, E->ReuseShuffleIndices, | |||
10567 | [E, this](Instruction *I) { | |||
10568 | 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", 10568, __extension__ __PRETTY_FUNCTION__)); | |||
10569 | return isAlternateInstruction(I, E->getMainOp(), E->getAltOp(), | |||
10570 | *TLI); | |||
10571 | }, | |||
10572 | Mask, &OpScalars, &AltScalars); | |||
10573 | ||||
10574 | propagateIRFlags(V0, OpScalars); | |||
10575 | propagateIRFlags(V1, AltScalars); | |||
10576 | ||||
10577 | Value *V = Builder.CreateShuffleVector(V0, V1, Mask); | |||
10578 | if (auto *I = dyn_cast<Instruction>(V)) { | |||
10579 | V = propagateMetadata(I, E->Scalars); | |||
10580 | GatherShuffleExtractSeq.insert(I); | |||
10581 | CSEBlocks.insert(I->getParent()); | |||
10582 | } | |||
10583 | ||||
10584 | E->VectorizedValue = V; | |||
10585 | ++NumVectorInstructions; | |||
10586 | ||||
10587 | return V; | |||
10588 | } | |||
10589 | default: | |||
10590 | llvm_unreachable("unknown inst")::llvm::llvm_unreachable_internal("unknown inst", "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 10590); | |||
10591 | } | |||
10592 | return nullptr; | |||
10593 | } | |||
10594 | ||||
10595 | Value *BoUpSLP::vectorizeTree() { | |||
10596 | ExtraValueToDebugLocsMap ExternallyUsedValues; | |||
10597 | SmallVector<std::pair<Value *, Value *>> ReplacedExternals; | |||
10598 | return vectorizeTree(ExternallyUsedValues, ReplacedExternals); | |||
10599 | } | |||
10600 | ||||
10601 | namespace { | |||
10602 | /// Data type for handling buildvector sequences with the reused scalars from | |||
10603 | /// other tree entries. | |||
10604 | struct ShuffledInsertData { | |||
10605 | /// List of insertelements to be replaced by shuffles. | |||
10606 | SmallVector<InsertElementInst *> InsertElements; | |||
10607 | /// The parent vectors and shuffle mask for the given list of inserts. | |||
10608 | MapVector<Value *, SmallVector<int>> ValueMasks; | |||
10609 | }; | |||
10610 | } // namespace | |||
10611 | ||||
10612 | Value *BoUpSLP::vectorizeTree( | |||
10613 | const ExtraValueToDebugLocsMap &ExternallyUsedValues, | |||
10614 | SmallVectorImpl<std::pair<Value *, Value *>> &ReplacedExternals, | |||
10615 | Instruction *ReductionRoot) { | |||
10616 | // All blocks must be scheduled before any instructions are inserted. | |||
10617 | for (auto &BSIter : BlocksSchedules) { | |||
10618 | scheduleBlock(BSIter.second.get()); | |||
10619 | } | |||
10620 | ||||
10621 | // Pre-gather last instructions. | |||
10622 | for (const std::unique_ptr<TreeEntry> &E : VectorizableTree) { | |||
10623 | if ((E->State == TreeEntry::NeedToGather && | |||
10624 | (!E->getMainOp() || E->Idx > 0)) || | |||
10625 | (E->State != TreeEntry::NeedToGather && | |||
10626 | E->getOpcode() == Instruction::ExtractValue) || | |||
10627 | E->getOpcode() == Instruction::InsertElement) | |||
10628 | continue; | |||
10629 | Instruction *LastInst = &getLastInstructionInBundle(E.get()); | |||
10630 | EntryToLastInstruction.try_emplace(E.get(), LastInst); | |||
10631 | } | |||
10632 | ||||
10633 | Builder.SetInsertPoint(ReductionRoot ? ReductionRoot | |||
10634 | : &F->getEntryBlock().front()); | |||
10635 | auto *VectorRoot = vectorizeTree(VectorizableTree[0].get()); | |||
10636 | // Run through the list of postponed gathers and emit them, replacing the temp | |||
10637 | // emitted allocas with actual vector instructions. | |||
10638 | ArrayRef<const TreeEntry *> PostponedNodes = PostponedGathers.getArrayRef(); | |||
10639 | DenseMap<Value *, SmallVector<TreeEntry *>> PostponedValues; | |||
10640 | for (const TreeEntry *E : PostponedNodes) { | |||
10641 | auto *TE = const_cast<TreeEntry *>(E); | |||
10642 | if (auto *VecTE = getTreeEntry(TE->Scalars.front())) | |||
10643 | if (VecTE->isSame(TE->UserTreeIndices.front().UserTE->getOperand( | |||
10644 | TE->UserTreeIndices.front().EdgeIdx))) | |||
10645 | // Found gather node which is absolutely the same as one of the | |||
10646 | // vectorized nodes. It may happen after reordering. | |||
10647 | continue; | |||
10648 | auto *PrevVec = cast<Instruction>(TE->VectorizedValue); | |||
10649 | TE->VectorizedValue = nullptr; | |||
10650 | auto *UserI = | |||
10651 | cast<Instruction>(TE->UserTreeIndices.front().UserTE->VectorizedValue); | |||
10652 | Builder.SetInsertPoint(PrevVec); | |||
10653 | Builder.SetCurrentDebugLocation(UserI->getDebugLoc()); | |||
10654 | Value *Vec = vectorizeTree(TE); | |||
10655 | PrevVec->replaceAllUsesWith(Vec); | |||
10656 | PostponedValues.try_emplace(Vec).first->second.push_back(TE); | |||
10657 | // Replace the stub vector node, if it was used before for one of the | |||
10658 | // buildvector nodes already. | |||
10659 | auto It = PostponedValues.find(PrevVec); | |||
10660 | if (It != PostponedValues.end()) { | |||
10661 | for (TreeEntry *VTE : It->getSecond()) | |||
10662 | VTE->VectorizedValue = Vec; | |||
10663 | } | |||
10664 | eraseInstruction(PrevVec); | |||
10665 | } | |||
10666 | ||||
10667 | // If the vectorized tree can be rewritten in a smaller type, we truncate the | |||
10668 | // vectorized root. InstCombine will then rewrite the entire expression. We | |||
10669 | // sign extend the extracted values below. | |||
10670 | auto *ScalarRoot = VectorizableTree[0]->Scalars[0]; | |||
10671 | if (MinBWs.count(ScalarRoot)) { | |||
10672 | if (auto *I = dyn_cast<Instruction>(VectorRoot)) { | |||
10673 | // If current instr is a phi and not the last phi, insert it after the | |||
10674 | // last phi node. | |||
10675 | if (isa<PHINode>(I)) | |||
10676 | Builder.SetInsertPoint(&*I->getParent()->getFirstInsertionPt()); | |||
10677 | else | |||
10678 | Builder.SetInsertPoint(&*++BasicBlock::iterator(I)); | |||
10679 | } | |||
10680 | auto BundleWidth = VectorizableTree[0]->Scalars.size(); | |||
10681 | auto *MinTy = IntegerType::get(F->getContext(), MinBWs[ScalarRoot].first); | |||
10682 | auto *VecTy = FixedVectorType::get(MinTy, BundleWidth); | |||
10683 | auto *Trunc = Builder.CreateTrunc(VectorRoot, VecTy); | |||
10684 | VectorizableTree[0]->VectorizedValue = Trunc; | |||
10685 | } | |||
10686 | ||||
10687 | LLVM_DEBUG(dbgs() << "SLP: Extracting " << ExternalUses.size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Extracting " << ExternalUses .size() << " values .\n"; } } while (false) | |||
10688 | << " values .\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Extracting " << ExternalUses .size() << " values .\n"; } } while (false); | |||
10689 | ||||
10690 | SmallVector<ShuffledInsertData> ShuffledInserts; | |||
10691 | // Maps vector instruction to original insertelement instruction | |||
10692 | DenseMap<Value *, InsertElementInst *> VectorToInsertElement; | |||
10693 | // Maps extract Scalar to the corresponding extractelement instruction in the | |||
10694 | // basic block. Only one extractelement per block should be emitted. | |||
10695 | DenseMap<Value *, DenseMap<BasicBlock *, Instruction *>> ScalarToEEs; | |||
10696 | // Extract all of the elements with the external uses. | |||
10697 | for (const auto &ExternalUse : ExternalUses) { | |||
10698 | Value *Scalar = ExternalUse.Scalar; | |||
10699 | llvm::User *User = ExternalUse.User; | |||
10700 | ||||
10701 | // Skip users that we already RAUW. This happens when one instruction | |||
10702 | // has multiple uses of the same value. | |||
10703 | if (User && !is_contained(Scalar->users(), User)) | |||
10704 | continue; | |||
10705 | TreeEntry *E = getTreeEntry(Scalar); | |||
10706 | assert(E && "Invalid scalar")(static_cast <bool> (E && "Invalid scalar") ? void (0) : __assert_fail ("E && \"Invalid scalar\"", "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 10706, __extension__ __PRETTY_FUNCTION__)); | |||
10707 | 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", 10708, __extension__ __PRETTY_FUNCTION__)) | |||
10708 | "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", 10708, __extension__ __PRETTY_FUNCTION__)); | |||
10709 | // Non-instruction pointers are not deleted, just skip them. | |||
10710 | if (E->getOpcode() == Instruction::GetElementPtr && | |||
10711 | !isa<GetElementPtrInst>(Scalar)) | |||
10712 | continue; | |||
10713 | ||||
10714 | Value *Vec = E->VectorizedValue; | |||
10715 | 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", 10715, __extension__ __PRETTY_FUNCTION__)); | |||
10716 | ||||
10717 | Value *Lane = Builder.getInt32(ExternalUse.Lane); | |||
10718 | auto ExtractAndExtendIfNeeded = [&](Value *Vec) { | |||
10719 | if (Scalar->getType() != Vec->getType()) { | |||
10720 | Value *Ex = nullptr; | |||
10721 | auto It = ScalarToEEs.find(Scalar); | |||
10722 | if (It != ScalarToEEs.end()) { | |||
10723 | // No need to emit many extracts, just move the only one in the | |||
10724 | // current block. | |||
10725 | auto EEIt = It->second.find(Builder.GetInsertBlock()); | |||
10726 | if (EEIt != It->second.end()) { | |||
10727 | Instruction *I = EEIt->second; | |||
10728 | if (Builder.GetInsertPoint() != Builder.GetInsertBlock()->end() && | |||
10729 | Builder.GetInsertPoint()->comesBefore(I)) | |||
10730 | I->moveBefore(&*Builder.GetInsertPoint()); | |||
10731 | Ex = I; | |||
10732 | } | |||
10733 | } | |||
10734 | if (!Ex) { | |||
10735 | // "Reuse" the existing extract to improve final codegen. | |||
10736 | if (auto *ES = dyn_cast<ExtractElementInst>(Scalar)) { | |||
10737 | Ex = Builder.CreateExtractElement(ES->getOperand(0), | |||
10738 | ES->getOperand(1)); | |||
10739 | } else { | |||
10740 | Ex = Builder.CreateExtractElement(Vec, Lane); | |||
10741 | } | |||
10742 | if (auto *I = dyn_cast<Instruction>(Ex)) | |||
10743 | ScalarToEEs[Scalar].try_emplace(Builder.GetInsertBlock(), I); | |||
10744 | } | |||
10745 | // The then branch of the previous if may produce constants, since 0 | |||
10746 | // operand might be a constant. | |||
10747 | if (auto *ExI = dyn_cast<Instruction>(Ex)) { | |||
10748 | GatherShuffleExtractSeq.insert(ExI); | |||
10749 | CSEBlocks.insert(ExI->getParent()); | |||
10750 | } | |||
10751 | // If necessary, sign-extend or zero-extend ScalarRoot | |||
10752 | // to the larger type. | |||
10753 | if (!MinBWs.count(ScalarRoot)) | |||
10754 | return Ex; | |||
10755 | if (MinBWs[ScalarRoot].second) | |||
10756 | return Builder.CreateSExt(Ex, Scalar->getType()); | |||
10757 | return Builder.CreateZExt(Ex, Scalar->getType()); | |||
10758 | } | |||
10759 | 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", 10761, __extension__ __PRETTY_FUNCTION__)) | |||
10760 | 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", 10761, __extension__ __PRETTY_FUNCTION__)) | |||
10761 | "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", 10761, __extension__ __PRETTY_FUNCTION__)); | |||
10762 | auto *IE = cast<InsertElementInst>(Scalar); | |||
10763 | VectorToInsertElement.try_emplace(Vec, IE); | |||
10764 | return Vec; | |||
10765 | }; | |||
10766 | // If User == nullptr, the Scalar is used as extra arg. Generate | |||
10767 | // ExtractElement instruction and update the record for this scalar in | |||
10768 | // ExternallyUsedValues. | |||
10769 | if (!User) { | |||
10770 | 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", 10772, __extension__ __PRETTY_FUNCTION__)) | |||
10771 | "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", 10772, __extension__ __PRETTY_FUNCTION__)) | |||
10772 | "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", 10772, __extension__ __PRETTY_FUNCTION__)); | |||
10773 | if (auto *VecI = dyn_cast<Instruction>(Vec)) { | |||
10774 | if (auto *PHI = dyn_cast<PHINode>(VecI)) | |||
10775 | Builder.SetInsertPoint(PHI->getParent()->getFirstNonPHI()); | |||
10776 | else | |||
10777 | Builder.SetInsertPoint(VecI->getParent(), | |||
10778 | std::next(VecI->getIterator())); | |||
10779 | } else { | |||
10780 | Builder.SetInsertPoint(&F->getEntryBlock().front()); | |||
10781 | } | |||
10782 | Value *NewInst = ExtractAndExtendIfNeeded(Vec); | |||
10783 | // Required to update internally referenced instructions. | |||
10784 | Scalar->replaceAllUsesWith(NewInst); | |||
10785 | ReplacedExternals.emplace_back(Scalar, NewInst); | |||
10786 | continue; | |||
10787 | } | |||
10788 | ||||
10789 | if (auto *VU = dyn_cast<InsertElementInst>(User)) { | |||
10790 | // Skip if the scalar is another vector op or Vec is not an instruction. | |||
10791 | if (!Scalar->getType()->isVectorTy() && isa<Instruction>(Vec)) { | |||
10792 | if (auto *FTy = dyn_cast<FixedVectorType>(User->getType())) { | |||
10793 | std::optional<unsigned> InsertIdx = getInsertIndex(VU); | |||
10794 | if (InsertIdx) { | |||
10795 | // Need to use original vector, if the root is truncated. | |||
10796 | if (MinBWs.count(Scalar) && | |||
10797 | VectorizableTree[0]->VectorizedValue == Vec) | |||
10798 | Vec = VectorRoot; | |||
10799 | auto *It = | |||
10800 | find_if(ShuffledInserts, [VU](const ShuffledInsertData &Data) { | |||
10801 | // Checks if 2 insertelements are from the same buildvector. | |||
10802 | InsertElementInst *VecInsert = Data.InsertElements.front(); | |||
10803 | return areTwoInsertFromSameBuildVector( | |||
10804 | VU, VecInsert, | |||
10805 | [](InsertElementInst *II) { return II->getOperand(0); }); | |||
10806 | }); | |||
10807 | unsigned Idx = *InsertIdx; | |||
10808 | if (It == ShuffledInserts.end()) { | |||
10809 | (void)ShuffledInserts.emplace_back(); | |||
10810 | It = std::next(ShuffledInserts.begin(), | |||
10811 | ShuffledInserts.size() - 1); | |||
10812 | SmallVectorImpl<int> &Mask = It->ValueMasks[Vec]; | |||
10813 | if (Mask.empty()) | |||
10814 | Mask.assign(FTy->getNumElements(), PoisonMaskElem); | |||
10815 | // Find the insertvector, vectorized in tree, if any. | |||
10816 | Value *Base = VU; | |||
10817 | while (auto *IEBase = dyn_cast<InsertElementInst>(Base)) { | |||
10818 | if (IEBase != User && | |||
10819 | (!IEBase->hasOneUse() || | |||
10820 | getInsertIndex(IEBase).value_or(Idx) == Idx)) | |||
10821 | break; | |||
10822 | // Build the mask for the vectorized insertelement instructions. | |||
10823 | if (const TreeEntry *E = getTreeEntry(IEBase)) { | |||
10824 | do { | |||
10825 | IEBase = cast<InsertElementInst>(Base); | |||
10826 | int IEIdx = *getInsertIndex(IEBase); | |||
10827 | assert(Mask[Idx] == PoisonMaskElem &&(static_cast <bool> (Mask[Idx] == PoisonMaskElem && "InsertElementInstruction used already.") ? void (0) : __assert_fail ("Mask[Idx] == PoisonMaskElem && \"InsertElementInstruction used already.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 10828, __extension__ __PRETTY_FUNCTION__)) | |||
10828 | "InsertElementInstruction used already.")(static_cast <bool> (Mask[Idx] == PoisonMaskElem && "InsertElementInstruction used already.") ? void (0) : __assert_fail ("Mask[Idx] == PoisonMaskElem && \"InsertElementInstruction used already.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 10828, __extension__ __PRETTY_FUNCTION__)); | |||
10829 | Mask[IEIdx] = IEIdx; | |||
10830 | Base = IEBase->getOperand(0); | |||
10831 | } while (E == getTreeEntry(Base)); | |||
10832 | break; | |||
10833 | } | |||
10834 | Base = cast<InsertElementInst>(Base)->getOperand(0); | |||
10835 | // After the vectorization the def-use chain has changed, need | |||
10836 | // to look through original insertelement instructions, if they | |||
10837 | // get replaced by vector instructions. | |||
10838 | auto It = VectorToInsertElement.find(Base); | |||
10839 | if (It != VectorToInsertElement.end()) | |||
10840 | Base = It->second; | |||
10841 | } | |||
10842 | } | |||
10843 | SmallVectorImpl<int> &Mask = It->ValueMasks[Vec]; | |||
10844 | if (Mask.empty()) | |||
10845 | Mask.assign(FTy->getNumElements(), PoisonMaskElem); | |||
10846 | Mask[Idx] = ExternalUse.Lane; | |||
10847 | It->InsertElements.push_back(cast<InsertElementInst>(User)); | |||
10848 | continue; | |||
10849 | } | |||
10850 | } | |||
10851 | } | |||
10852 | } | |||
10853 | ||||
10854 | // Generate extracts for out-of-tree users. | |||
10855 | // Find the insertion point for the extractelement lane. | |||
10856 | if (auto *VecI = dyn_cast<Instruction>(Vec)) { | |||
10857 | if (PHINode *PH = dyn_cast<PHINode>(User)) { | |||
10858 | for (int i = 0, e = PH->getNumIncomingValues(); i != e; ++i) { | |||
10859 | if (PH->getIncomingValue(i) == Scalar) { | |||
10860 | Instruction *IncomingTerminator = | |||
10861 | PH->getIncomingBlock(i)->getTerminator(); | |||
10862 | if (isa<CatchSwitchInst>(IncomingTerminator)) { | |||
10863 | Builder.SetInsertPoint(VecI->getParent(), | |||
10864 | std::next(VecI->getIterator())); | |||
10865 | } else { | |||
10866 | Builder.SetInsertPoint(PH->getIncomingBlock(i)->getTerminator()); | |||
10867 | } | |||
10868 | Value *NewInst = ExtractAndExtendIfNeeded(Vec); | |||
10869 | PH->setOperand(i, NewInst); | |||
10870 | } | |||
10871 | } | |||
10872 | } else { | |||
10873 | Builder.SetInsertPoint(cast<Instruction>(User)); | |||
10874 | Value *NewInst = ExtractAndExtendIfNeeded(Vec); | |||
10875 | User->replaceUsesOfWith(Scalar, NewInst); | |||
10876 | } | |||
10877 | } else { | |||
10878 | Builder.SetInsertPoint(&F->getEntryBlock().front()); | |||
10879 | Value *NewInst = ExtractAndExtendIfNeeded(Vec); | |||
10880 | User->replaceUsesOfWith(Scalar, NewInst); | |||
10881 | } | |||
10882 | ||||
10883 | LLVM_DEBUG(dbgs() << "SLP: Replaced:" << *User << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Replaced:" << *User << ".\n"; } } while (false); | |||
10884 | } | |||
10885 | ||||
10886 | auto CreateShuffle = [&](Value *V1, Value *V2, ArrayRef<int> Mask) { | |||
10887 | SmallVector<int> CombinedMask1(Mask.size(), PoisonMaskElem); | |||
10888 | SmallVector<int> CombinedMask2(Mask.size(), PoisonMaskElem); | |||
10889 | int VF = cast<FixedVectorType>(V1->getType())->getNumElements(); | |||
10890 | for (int I = 0, E = Mask.size(); I < E; ++I) { | |||
10891 | if (Mask[I] < VF) | |||
10892 | CombinedMask1[I] = Mask[I]; | |||
10893 | else | |||
10894 | CombinedMask2[I] = Mask[I] - VF; | |||
10895 | } | |||
10896 | ShuffleInstructionBuilder ShuffleBuilder(Builder, *this); | |||
10897 | ShuffleBuilder.add(V1, CombinedMask1); | |||
10898 | if (V2) | |||
10899 | ShuffleBuilder.add(V2, CombinedMask2); | |||
10900 | return ShuffleBuilder.finalize(std::nullopt); | |||
10901 | }; | |||
10902 | ||||
10903 | auto &&ResizeToVF = [&CreateShuffle](Value *Vec, ArrayRef<int> Mask, | |||
10904 | bool ForSingleMask) { | |||
10905 | unsigned VF = Mask.size(); | |||
10906 | unsigned VecVF = cast<FixedVectorType>(Vec->getType())->getNumElements(); | |||
10907 | if (VF != VecVF) { | |||
10908 | if (any_of(Mask, [VF](int Idx) { return Idx >= static_cast<int>(VF); })) { | |||
10909 | Vec = CreateShuffle(Vec, nullptr, Mask); | |||
10910 | return std::make_pair(Vec, true); | |||
10911 | } | |||
10912 | if (!ForSingleMask) { | |||
10913 | SmallVector<int> ResizeMask(VF, PoisonMaskElem); | |||
10914 | for (unsigned I = 0; I < VF; ++I) { | |||
10915 | if (Mask[I] != PoisonMaskElem) | |||
10916 | ResizeMask[Mask[I]] = Mask[I]; | |||
10917 | } | |||
10918 | Vec = CreateShuffle(Vec, nullptr, ResizeMask); | |||
10919 | } | |||
10920 | } | |||
10921 | ||||
10922 | return std::make_pair(Vec, false); | |||
10923 | }; | |||
10924 | // Perform shuffling of the vectorize tree entries for better handling of | |||
10925 | // external extracts. | |||
10926 | for (int I = 0, E = ShuffledInserts.size(); I < E; ++I) { | |||
10927 | // Find the first and the last instruction in the list of insertelements. | |||
10928 | sort(ShuffledInserts[I].InsertElements, isFirstInsertElement); | |||
10929 | InsertElementInst *FirstInsert = ShuffledInserts[I].InsertElements.front(); | |||
10930 | InsertElementInst *LastInsert = ShuffledInserts[I].InsertElements.back(); | |||
10931 | Builder.SetInsertPoint(LastInsert); | |||
10932 | auto Vector = ShuffledInserts[I].ValueMasks.takeVector(); | |||
10933 | Value *NewInst = performExtractsShuffleAction<Value>( | |||
10934 | MutableArrayRef(Vector.data(), Vector.size()), | |||
10935 | FirstInsert->getOperand(0), | |||
10936 | [](Value *Vec) { | |||
10937 | return cast<VectorType>(Vec->getType()) | |||
10938 | ->getElementCount() | |||
10939 | .getKnownMinValue(); | |||
10940 | }, | |||
10941 | ResizeToVF, | |||
10942 | [FirstInsert, &CreateShuffle](ArrayRef<int> Mask, | |||
10943 | ArrayRef<Value *> Vals) { | |||
10944 | assert((Vals.size() == 1 || Vals.size() == 2) &&(static_cast <bool> ((Vals.size() == 1 || Vals.size() == 2) && "Expected exactly 1 or 2 input values.") ? void (0) : __assert_fail ("(Vals.size() == 1 || Vals.size() == 2) && \"Expected exactly 1 or 2 input values.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 10945, __extension__ __PRETTY_FUNCTION__)) | |||
10945 | "Expected exactly 1 or 2 input values.")(static_cast <bool> ((Vals.size() == 1 || Vals.size() == 2) && "Expected exactly 1 or 2 input values.") ? void (0) : __assert_fail ("(Vals.size() == 1 || Vals.size() == 2) && \"Expected exactly 1 or 2 input values.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 10945, __extension__ __PRETTY_FUNCTION__)); | |||
10946 | if (Vals.size() == 1) { | |||
10947 | // Do not create shuffle if the mask is a simple identity | |||
10948 | // non-resizing mask. | |||
10949 | if (Mask.size() != cast<FixedVectorType>(Vals.front()->getType()) | |||
10950 | ->getNumElements() || | |||
10951 | !ShuffleVectorInst::isIdentityMask(Mask)) | |||
10952 | return CreateShuffle(Vals.front(), nullptr, Mask); | |||
10953 | return Vals.front(); | |||
10954 | } | |||
10955 | return CreateShuffle(Vals.front() ? Vals.front() | |||
10956 | : FirstInsert->getOperand(0), | |||
10957 | Vals.back(), Mask); | |||
10958 | }); | |||
10959 | auto It = ShuffledInserts[I].InsertElements.rbegin(); | |||
10960 | // Rebuild buildvector chain. | |||
10961 | InsertElementInst *II = nullptr; | |||
10962 | if (It != ShuffledInserts[I].InsertElements.rend()) | |||
10963 | II = *It; | |||
10964 | SmallVector<Instruction *> Inserts; | |||
10965 | while (It != ShuffledInserts[I].InsertElements.rend()) { | |||
10966 | assert(II && "Must be an insertelement instruction.")(static_cast <bool> (II && "Must be an insertelement instruction." ) ? void (0) : __assert_fail ("II && \"Must be an insertelement instruction.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 10966, __extension__ __PRETTY_FUNCTION__)); | |||
10967 | if (*It == II) | |||
10968 | ++It; | |||
10969 | else | |||
10970 | Inserts.push_back(cast<Instruction>(II)); | |||
10971 | II = dyn_cast<InsertElementInst>(II->getOperand(0)); | |||
10972 | } | |||
10973 | for (Instruction *II : reverse(Inserts)) { | |||
10974 | II->replaceUsesOfWith(II->getOperand(0), NewInst); | |||
10975 | if (auto *NewI = dyn_cast<Instruction>(NewInst)) | |||
10976 | if (II->getParent() == NewI->getParent() && II->comesBefore(NewI)) | |||
10977 | II->moveAfter(NewI); | |||
10978 | NewInst = II; | |||
10979 | } | |||
10980 | LastInsert->replaceAllUsesWith(NewInst); | |||
10981 | for (InsertElementInst *IE : reverse(ShuffledInserts[I].InsertElements)) { | |||
10982 | IE->replaceUsesOfWith(IE->getOperand(0), | |||
10983 | PoisonValue::get(IE->getOperand(0)->getType())); | |||
10984 | IE->replaceUsesOfWith(IE->getOperand(1), | |||
10985 | PoisonValue::get(IE->getOperand(1)->getType())); | |||
10986 | eraseInstruction(IE); | |||
10987 | } | |||
10988 | CSEBlocks.insert(LastInsert->getParent()); | |||
10989 | } | |||
10990 | ||||
10991 | SmallVector<Instruction *> RemovedInsts; | |||
10992 | // For each vectorized value: | |||
10993 | for (auto &TEPtr : VectorizableTree) { | |||
10994 | TreeEntry *Entry = TEPtr.get(); | |||
10995 | ||||
10996 | // No need to handle users of gathered values. | |||
10997 | if (Entry->State == TreeEntry::NeedToGather) | |||
10998 | continue; | |||
10999 | ||||
11000 | 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", 11000, __extension__ __PRETTY_FUNCTION__)); | |||
11001 | ||||
11002 | // For each lane: | |||
11003 | for (int Lane = 0, LE = Entry->Scalars.size(); Lane != LE; ++Lane) { | |||
11004 | Value *Scalar = Entry->Scalars[Lane]; | |||
11005 | ||||
11006 | if (Entry->getOpcode() == Instruction::GetElementPtr && | |||
11007 | !isa<GetElementPtrInst>(Scalar)) | |||
11008 | continue; | |||
11009 | #ifndef NDEBUG | |||
11010 | Type *Ty = Scalar->getType(); | |||
11011 | if (!Ty->isVoidTy()) { | |||
11012 | for (User *U : Scalar->users()) { | |||
11013 | LLVM_DEBUG(dbgs() << "SLP: \tvalidating user:" << *U << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: \tvalidating user:" << *U << ".\n"; } } while (false); | |||
11014 | ||||
11015 | // It is legal to delete users in the ignorelist. | |||
11016 | assert((getTreeEntry(U) ||(static_cast <bool> ((getTreeEntry(U) || (UserIgnoreList && UserIgnoreList->contains(U)) || (isa_and_nonnull <Instruction>(U) && isDeleted(cast<Instruction >(U)))) && "Deleting out-of-tree value") ? void (0 ) : __assert_fail ("(getTreeEntry(U) || (UserIgnoreList && UserIgnoreList->contains(U)) || (isa_and_nonnull<Instruction>(U) && isDeleted(cast<Instruction>(U)))) && \"Deleting out-of-tree value\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 11020, __extension__ __PRETTY_FUNCTION__)) | |||
11017 | (UserIgnoreList && UserIgnoreList->contains(U)) ||(static_cast <bool> ((getTreeEntry(U) || (UserIgnoreList && UserIgnoreList->contains(U)) || (isa_and_nonnull <Instruction>(U) && isDeleted(cast<Instruction >(U)))) && "Deleting out-of-tree value") ? void (0 ) : __assert_fail ("(getTreeEntry(U) || (UserIgnoreList && UserIgnoreList->contains(U)) || (isa_and_nonnull<Instruction>(U) && isDeleted(cast<Instruction>(U)))) && \"Deleting out-of-tree value\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 11020, __extension__ __PRETTY_FUNCTION__)) | |||
11018 | (isa_and_nonnull<Instruction>(U) &&(static_cast <bool> ((getTreeEntry(U) || (UserIgnoreList && UserIgnoreList->contains(U)) || (isa_and_nonnull <Instruction>(U) && isDeleted(cast<Instruction >(U)))) && "Deleting out-of-tree value") ? void (0 ) : __assert_fail ("(getTreeEntry(U) || (UserIgnoreList && UserIgnoreList->contains(U)) || (isa_and_nonnull<Instruction>(U) && isDeleted(cast<Instruction>(U)))) && \"Deleting out-of-tree value\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 11020, __extension__ __PRETTY_FUNCTION__)) | |||
11019 | isDeleted(cast<Instruction>(U)))) &&(static_cast <bool> ((getTreeEntry(U) || (UserIgnoreList && UserIgnoreList->contains(U)) || (isa_and_nonnull <Instruction>(U) && isDeleted(cast<Instruction >(U)))) && "Deleting out-of-tree value") ? void (0 ) : __assert_fail ("(getTreeEntry(U) || (UserIgnoreList && UserIgnoreList->contains(U)) || (isa_and_nonnull<Instruction>(U) && isDeleted(cast<Instruction>(U)))) && \"Deleting out-of-tree value\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 11020, __extension__ __PRETTY_FUNCTION__)) | |||
11020 | "Deleting out-of-tree value")(static_cast <bool> ((getTreeEntry(U) || (UserIgnoreList && UserIgnoreList->contains(U)) || (isa_and_nonnull <Instruction>(U) && isDeleted(cast<Instruction >(U)))) && "Deleting out-of-tree value") ? void (0 ) : __assert_fail ("(getTreeEntry(U) || (UserIgnoreList && UserIgnoreList->contains(U)) || (isa_and_nonnull<Instruction>(U) && isDeleted(cast<Instruction>(U)))) && \"Deleting out-of-tree value\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 11020, __extension__ __PRETTY_FUNCTION__)); | |||
11021 | } | |||
11022 | } | |||
11023 | #endif | |||
11024 | LLVM_DEBUG(dbgs() << "SLP: \tErasing scalar:" << *Scalar << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: \tErasing scalar:" << * Scalar << ".\n"; } } while (false); | |||
11025 | eraseInstruction(cast<Instruction>(Scalar)); | |||
11026 | // Retain to-be-deleted instructions for some debug-info | |||
11027 | // bookkeeping. NOTE: eraseInstruction only marks the instruction for | |||
11028 | // deletion - instructions are not deleted until later. | |||
11029 | RemovedInsts.push_back(cast<Instruction>(Scalar)); | |||
11030 | } | |||
11031 | } | |||
11032 | ||||
11033 | // Merge the DIAssignIDs from the about-to-be-deleted instructions into the | |||
11034 | // new vector instruction. | |||
11035 | if (auto *V = dyn_cast<Instruction>(VectorizableTree[0]->VectorizedValue)) | |||
11036 | V->mergeDIAssignID(RemovedInsts); | |||
11037 | ||||
11038 | Builder.ClearInsertionPoint(); | |||
11039 | InstrElementSize.clear(); | |||
11040 | ||||
11041 | return VectorizableTree[0]->VectorizedValue; | |||
11042 | } | |||
11043 | ||||
11044 | void BoUpSLP::optimizeGatherSequence() { | |||
11045 | LLVM_DEBUG(dbgs() << "SLP: Optimizing " << GatherShuffleExtractSeq.size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Optimizing " << GatherShuffleExtractSeq .size() << " gather sequences instructions.\n"; } } while (false) | |||
11046 | << " gather sequences instructions.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Optimizing " << GatherShuffleExtractSeq .size() << " gather sequences instructions.\n"; } } while (false); | |||
11047 | // LICM InsertElementInst sequences. | |||
11048 | for (Instruction *I : GatherShuffleExtractSeq) { | |||
11049 | if (isDeleted(I)) | |||
11050 | continue; | |||
11051 | ||||
11052 | // Check if this block is inside a loop. | |||
11053 | Loop *L = LI->getLoopFor(I->getParent()); | |||
11054 | if (!L) | |||
11055 | continue; | |||
11056 | ||||
11057 | // Check if it has a preheader. | |||
11058 | BasicBlock *PreHeader = L->getLoopPreheader(); | |||
11059 | if (!PreHeader) | |||
11060 | continue; | |||
11061 | ||||
11062 | // If the vector or the element that we insert into it are | |||
11063 | // instructions that are defined in this basic block then we can't | |||
11064 | // hoist this instruction. | |||
11065 | if (any_of(I->operands(), [L](Value *V) { | |||
11066 | auto *OpI = dyn_cast<Instruction>(V); | |||
11067 | return OpI && L->contains(OpI); | |||
11068 | })) | |||
11069 | continue; | |||
11070 | ||||
11071 | // We can hoist this instruction. Move it to the pre-header. | |||
11072 | I->moveBefore(PreHeader->getTerminator()); | |||
11073 | CSEBlocks.insert(PreHeader); | |||
11074 | } | |||
11075 | ||||
11076 | // Make a list of all reachable blocks in our CSE queue. | |||
11077 | SmallVector<const DomTreeNode *, 8> CSEWorkList; | |||
11078 | CSEWorkList.reserve(CSEBlocks.size()); | |||
11079 | for (BasicBlock *BB : CSEBlocks) | |||
11080 | if (DomTreeNode *N = DT->getNode(BB)) { | |||
11081 | assert(DT->isReachableFromEntry(N))(static_cast <bool> (DT->isReachableFromEntry(N)) ? void (0) : __assert_fail ("DT->isReachableFromEntry(N)", "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 11081, __extension__ __PRETTY_FUNCTION__)); | |||
11082 | CSEWorkList.push_back(N); | |||
11083 | } | |||
11084 | ||||
11085 | // Sort blocks by domination. This ensures we visit a block after all blocks | |||
11086 | // dominating it are visited. | |||
11087 | llvm::sort(CSEWorkList, [](const DomTreeNode *A, const DomTreeNode *B) { | |||
11088 | 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", 11089, __extension__ __PRETTY_FUNCTION__)) | |||
11089 | "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", 11089, __extension__ __PRETTY_FUNCTION__)); | |||
11090 | return A->getDFSNumIn() < B->getDFSNumIn(); | |||
11091 | }); | |||
11092 | ||||
11093 | // Less defined shuffles can be replaced by the more defined copies. | |||
11094 | // Between two shuffles one is less defined if it has the same vector operands | |||
11095 | // and its mask indeces are the same as in the first one or undefs. E.g. | |||
11096 | // shuffle %0, poison, <0, 0, 0, undef> is less defined than shuffle %0, | |||
11097 | // poison, <0, 0, 0, 0>. | |||
11098 | auto &&IsIdenticalOrLessDefined = [this](Instruction *I1, Instruction *I2, | |||
11099 | SmallVectorImpl<int> &NewMask) { | |||
11100 | if (I1->getType() != I2->getType()) | |||
11101 | return false; | |||
11102 | auto *SI1 = dyn_cast<ShuffleVectorInst>(I1); | |||
11103 | auto *SI2 = dyn_cast<ShuffleVectorInst>(I2); | |||
11104 | if (!SI1 || !SI2) | |||
11105 | return I1->isIdenticalTo(I2); | |||
11106 | if (SI1->isIdenticalTo(SI2)) | |||
11107 | return true; | |||
11108 | for (int I = 0, E = SI1->getNumOperands(); I < E; ++I) | |||
11109 | if (SI1->getOperand(I) != SI2->getOperand(I)) | |||
11110 | return false; | |||
11111 | // Check if the second instruction is more defined than the first one. | |||
11112 | NewMask.assign(SI2->getShuffleMask().begin(), SI2->getShuffleMask().end()); | |||
11113 | ArrayRef<int> SM1 = SI1->getShuffleMask(); | |||
11114 | // Count trailing undefs in the mask to check the final number of used | |||
11115 | // registers. | |||
11116 | unsigned LastUndefsCnt = 0; | |||
11117 | for (int I = 0, E = NewMask.size(); I < E; ++I) { | |||
11118 | if (SM1[I] == PoisonMaskElem) | |||
11119 | ++LastUndefsCnt; | |||
11120 | else | |||
11121 | LastUndefsCnt = 0; | |||
11122 | if (NewMask[I] != PoisonMaskElem && SM1[I] != PoisonMaskElem && | |||
11123 | NewMask[I] != SM1[I]) | |||
11124 | return false; | |||
11125 | if (NewMask[I] == PoisonMaskElem) | |||
11126 | NewMask[I] = SM1[I]; | |||
11127 | } | |||
11128 | // Check if the last undefs actually change the final number of used vector | |||
11129 | // registers. | |||
11130 | return SM1.size() - LastUndefsCnt > 1 && | |||
11131 | TTI->getNumberOfParts(SI1->getType()) == | |||
11132 | TTI->getNumberOfParts( | |||
11133 | FixedVectorType::get(SI1->getType()->getElementType(), | |||
11134 | SM1.size() - LastUndefsCnt)); | |||
11135 | }; | |||
11136 | // Perform O(N^2) search over the gather/shuffle sequences and merge identical | |||
11137 | // instructions. TODO: We can further optimize this scan if we split the | |||
11138 | // instructions into different buckets based on the insert lane. | |||
11139 | SmallVector<Instruction *, 16> Visited; | |||
11140 | for (auto I = CSEWorkList.begin(), E = CSEWorkList.end(); I != E; ++I) { | |||
11141 | 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", 11143, __extension__ __PRETTY_FUNCTION__)) | |||
11142 | (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", 11143, __extension__ __PRETTY_FUNCTION__)) | |||
11143 | "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", 11143, __extension__ __PRETTY_FUNCTION__)); | |||
11144 | BasicBlock *BB = (*I)->getBlock(); | |||
11145 | // For all instructions in blocks containing gather sequences: | |||
11146 | for (Instruction &In : llvm::make_early_inc_range(*BB)) { | |||
11147 | if (isDeleted(&In)) | |||
11148 | continue; | |||
11149 | if (!isa<InsertElementInst, ExtractElementInst, ShuffleVectorInst>(&In) && | |||
11150 | !GatherShuffleExtractSeq.contains(&In)) | |||
11151 | continue; | |||
11152 | ||||
11153 | // Check if we can replace this instruction with any of the | |||
11154 | // visited instructions. | |||
11155 | bool Replaced = false; | |||
11156 | for (Instruction *&V : Visited) { | |||
11157 | SmallVector<int> NewMask; | |||
11158 | if (IsIdenticalOrLessDefined(&In, V, NewMask) && | |||
11159 | DT->dominates(V->getParent(), In.getParent())) { | |||
11160 | In.replaceAllUsesWith(V); | |||
11161 | eraseInstruction(&In); | |||
11162 | if (auto *SI = dyn_cast<ShuffleVectorInst>(V)) | |||
11163 | if (!NewMask.empty()) | |||
11164 | SI->setShuffleMask(NewMask); | |||
11165 | Replaced = true; | |||
11166 | break; | |||
11167 | } | |||
11168 | if (isa<ShuffleVectorInst>(In) && isa<ShuffleVectorInst>(V) && | |||
11169 | GatherShuffleExtractSeq.contains(V) && | |||
11170 | IsIdenticalOrLessDefined(V, &In, NewMask) && | |||
11171 | DT->dominates(In.getParent(), V->getParent())) { | |||
11172 | In.moveAfter(V); | |||
11173 | V->replaceAllUsesWith(&In); | |||
11174 | eraseInstruction(V); | |||
11175 | if (auto *SI = dyn_cast<ShuffleVectorInst>(&In)) | |||
11176 | if (!NewMask.empty()) | |||
11177 | SI->setShuffleMask(NewMask); | |||
11178 | V = &In; | |||
11179 | Replaced = true; | |||
11180 | break; | |||
11181 | } | |||
11182 | } | |||
11183 | if (!Replaced) { | |||
11184 | 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", 11184, __extension__ __PRETTY_FUNCTION__)); | |||
11185 | Visited.push_back(&In); | |||
11186 | } | |||
11187 | } | |||
11188 | } | |||
11189 | CSEBlocks.clear(); | |||
11190 | GatherShuffleExtractSeq.clear(); | |||
11191 | } | |||
11192 | ||||
11193 | BoUpSLP::ScheduleData * | |||
11194 | BoUpSLP::BlockScheduling::buildBundle(ArrayRef<Value *> VL) { | |||
11195 | ScheduleData *Bundle = nullptr; | |||
11196 | ScheduleData *PrevInBundle = nullptr; | |||
11197 | for (Value *V : VL) { | |||
11198 | if (doesNotNeedToBeScheduled(V)) | |||
11199 | continue; | |||
11200 | ScheduleData *BundleMember = getScheduleData(V); | |||
11201 | 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", 11203, __extension__ __PRETTY_FUNCTION__)) | |||
11202 | "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", 11203, __extension__ __PRETTY_FUNCTION__)) | |||
11203 | "(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", 11203, __extension__ __PRETTY_FUNCTION__)); | |||
11204 | 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", 11205, __extension__ __PRETTY_FUNCTION__)) | |||
11205 | "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", 11205, __extension__ __PRETTY_FUNCTION__)); | |||
11206 | if (PrevInBundle) { | |||
11207 | PrevInBundle->NextInBundle = BundleMember; | |||
11208 | } else { | |||
11209 | Bundle = BundleMember; | |||
11210 | } | |||
11211 | ||||
11212 | // Group the instructions to a bundle. | |||
11213 | BundleMember->FirstInBundle = Bundle; | |||
11214 | PrevInBundle = BundleMember; | |||
11215 | } | |||
11216 | 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", 11216, __extension__ __PRETTY_FUNCTION__)); | |||
11217 | return Bundle; | |||
11218 | } | |||
11219 | ||||
11220 | // Groups the instructions to a bundle (which is then a single scheduling entity) | |||
11221 | // and schedules instructions until the bundle gets ready. | |||
11222 | std::optional<BoUpSLP::ScheduleData *> | |||
11223 | BoUpSLP::BlockScheduling::tryScheduleBundle(ArrayRef<Value *> VL, BoUpSLP *SLP, | |||
11224 | const InstructionsState &S) { | |||
11225 | // No need to schedule PHIs, insertelement, extractelement and extractvalue | |||
11226 | // instructions. | |||
11227 | if (isa<PHINode>(S.OpValue) || isVectorLikeInstWithConstOps(S.OpValue) || | |||
11228 | doesNotNeedToSchedule(VL)) | |||
11229 | return nullptr; | |||
11230 | ||||
11231 | // Initialize the instruction bundle. | |||
11232 | Instruction *OldScheduleEnd = ScheduleEnd; | |||
11233 | LLVM_DEBUG(dbgs() << "SLP: bundle: " << *S.OpValue << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: bundle: " << *S.OpValue << "\n"; } } while (false); | |||
11234 | ||||
11235 | auto TryScheduleBundleImpl = [this, OldScheduleEnd, SLP](bool ReSchedule, | |||
11236 | ScheduleData *Bundle) { | |||
11237 | // The scheduling region got new instructions at the lower end (or it is a | |||
11238 | // new region for the first bundle). This makes it necessary to | |||
11239 | // recalculate all dependencies. | |||
11240 | // It is seldom that this needs to be done a second time after adding the | |||
11241 | // initial bundle to the region. | |||
11242 | if (ScheduleEnd != OldScheduleEnd) { | |||
11243 | for (auto *I = ScheduleStart; I != ScheduleEnd; I = I->getNextNode()) | |||
11244 | doForAllOpcodes(I, [](ScheduleData *SD) { SD->clearDependencies(); }); | |||
11245 | ReSchedule = true; | |||
11246 | } | |||
11247 | if (Bundle) { | |||
11248 | 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) | |||
11249 | << " 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); | |||
11250 | calculateDependencies(Bundle, /*InsertInReadyList=*/true, SLP); | |||
11251 | } | |||
11252 | ||||
11253 | if (ReSchedule) { | |||
11254 | resetSchedule(); | |||
11255 | initialFillReadyList(ReadyInsts); | |||
11256 | } | |||
11257 | ||||
11258 | // Now try to schedule the new bundle or (if no bundle) just calculate | |||
11259 | // dependencies. As soon as the bundle is "ready" it means that there are no | |||
11260 | // cyclic dependencies and we can schedule it. Note that's important that we | |||
11261 | // don't "schedule" the bundle yet (see cancelScheduling). | |||
11262 | while (((!Bundle && ReSchedule) || (Bundle && !Bundle->isReady())) && | |||
11263 | !ReadyInsts.empty()) { | |||
11264 | ScheduleData *Picked = ReadyInsts.pop_back_val(); | |||
11265 | 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", 11266, __extension__ __PRETTY_FUNCTION__)) | |||
11266 | "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", 11266, __extension__ __PRETTY_FUNCTION__)); | |||
11267 | schedule(Picked, ReadyInsts); | |||
11268 | } | |||
11269 | }; | |||
11270 | ||||
11271 | // Make sure that the scheduling region contains all | |||
11272 | // instructions of the bundle. | |||
11273 | for (Value *V : VL) { | |||
11274 | if (doesNotNeedToBeScheduled(V)) | |||
11275 | continue; | |||
11276 | if (!extendSchedulingRegion(V, S)) { | |||
11277 | // If the scheduling region got new instructions at the lower end (or it | |||
11278 | // is a new region for the first bundle). This makes it necessary to | |||
11279 | // recalculate all dependencies. | |||
11280 | // Otherwise the compiler may crash trying to incorrectly calculate | |||
11281 | // dependencies and emit instruction in the wrong order at the actual | |||
11282 | // scheduling. | |||
11283 | TryScheduleBundleImpl(/*ReSchedule=*/false, nullptr); | |||
11284 | return std::nullopt; | |||
11285 | } | |||
11286 | } | |||
11287 | ||||
11288 | bool ReSchedule = false; | |||
11289 | for (Value *V : VL) { | |||
11290 | if (doesNotNeedToBeScheduled(V)) | |||
11291 | continue; | |||
11292 | ScheduleData *BundleMember = getScheduleData(V); | |||
11293 | 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", 11294, __extension__ __PRETTY_FUNCTION__)) | |||
11294 | "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", 11294, __extension__ __PRETTY_FUNCTION__)); | |||
11295 | ||||
11296 | // Make sure we don't leave the pieces of the bundle in the ready list when | |||
11297 | // whole bundle might not be ready. | |||
11298 | ReadyInsts.remove(BundleMember); | |||
11299 | ||||
11300 | if (!BundleMember->IsScheduled) | |||
11301 | continue; | |||
11302 | // A bundle member was scheduled as single instruction before and now | |||
11303 | // needs to be scheduled as part of the bundle. We just get rid of the | |||
11304 | // existing schedule. | |||
11305 | 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) | |||
11306 | << " was already scheduled\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: reset schedule because " << *BundleMember << " was already scheduled\n"; } } while (false); | |||
11307 | ReSchedule = true; | |||
11308 | } | |||
11309 | ||||
11310 | auto *Bundle = buildBundle(VL); | |||
11311 | TryScheduleBundleImpl(ReSchedule, Bundle); | |||
11312 | if (!Bundle->isReady()) { | |||
11313 | cancelScheduling(VL, S.OpValue); | |||
11314 | return std::nullopt; | |||
11315 | } | |||
11316 | return Bundle; | |||
11317 | } | |||
11318 | ||||
11319 | void BoUpSLP::BlockScheduling::cancelScheduling(ArrayRef<Value *> VL, | |||
11320 | Value *OpValue) { | |||
11321 | if (isa<PHINode>(OpValue) || isVectorLikeInstWithConstOps(OpValue) || | |||
11322 | doesNotNeedToSchedule(VL)) | |||
11323 | return; | |||
11324 | ||||
11325 | if (doesNotNeedToBeScheduled(OpValue)) | |||
11326 | OpValue = *find_if_not(VL, doesNotNeedToBeScheduled); | |||
11327 | ScheduleData *Bundle = getScheduleData(OpValue); | |||
11328 | 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); | |||
11329 | 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", 11330, __extension__ __PRETTY_FUNCTION__)) | |||
11330 | "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", 11330, __extension__ __PRETTY_FUNCTION__)); | |||
11331 | 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", 11333, __extension__ __PRETTY_FUNCTION__)) | |||
11332 | (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", 11333, __extension__ __PRETTY_FUNCTION__)) | |||
11333 | "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", 11333, __extension__ __PRETTY_FUNCTION__)); | |||
11334 | ||||
11335 | // Remove the bundle from the ready list. | |||
11336 | if (Bundle->isReady()) | |||
11337 | ReadyInsts.remove(Bundle); | |||
11338 | ||||
11339 | // Un-bundle: make single instructions out of the bundle. | |||
11340 | ScheduleData *BundleMember = Bundle; | |||
11341 | while (BundleMember) { | |||
11342 | 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", 11342, __extension__ __PRETTY_FUNCTION__)); | |||
11343 | BundleMember->FirstInBundle = BundleMember; | |||
11344 | ScheduleData *Next = BundleMember->NextInBundle; | |||
11345 | BundleMember->NextInBundle = nullptr; | |||
11346 | BundleMember->TE = nullptr; | |||
11347 | if (BundleMember->unscheduledDepsInBundle() == 0) { | |||
11348 | ReadyInsts.insert(BundleMember); | |||
11349 | } | |||
11350 | BundleMember = Next; | |||
11351 | } | |||
11352 | } | |||
11353 | ||||
11354 | BoUpSLP::ScheduleData *BoUpSLP::BlockScheduling::allocateScheduleDataChunks() { | |||
11355 | // Allocate a new ScheduleData for the instruction. | |||
11356 | if (ChunkPos >= ChunkSize) { | |||
11357 | ScheduleDataChunks.push_back(std::make_unique<ScheduleData[]>(ChunkSize)); | |||
11358 | ChunkPos = 0; | |||
11359 | } | |||
11360 | return &(ScheduleDataChunks.back()[ChunkPos++]); | |||
11361 | } | |||
11362 | ||||
11363 | bool BoUpSLP::BlockScheduling::extendSchedulingRegion(Value *V, | |||
11364 | const InstructionsState &S) { | |||
11365 | if (getScheduleData(V, isOneOf(S, V))) | |||
11366 | return true; | |||
11367 | Instruction *I = dyn_cast<Instruction>(V); | |||
11368 | 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", 11368, __extension__ __PRETTY_FUNCTION__)); | |||
11369 | 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", 11372, __extension__ __PRETTY_FUNCTION__)) | |||
11370 | !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", 11372, __extension__ __PRETTY_FUNCTION__)) | |||
11371 | "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", 11372, __extension__ __PRETTY_FUNCTION__)) | |||
11372 | "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", 11372, __extension__ __PRETTY_FUNCTION__)); | |||
11373 | auto &&CheckScheduleForI = [this, &S](Instruction *I) -> bool { | |||
11374 | ScheduleData *ISD = getScheduleData(I); | |||
11375 | if (!ISD) | |||
11376 | return false; | |||
11377 | 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", 11378, __extension__ __PRETTY_FUNCTION__)) | |||
11378 | "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", 11378, __extension__ __PRETTY_FUNCTION__)); | |||
11379 | ScheduleData *SD = allocateScheduleDataChunks(); | |||
11380 | SD->Inst = I; | |||
11381 | SD->init(SchedulingRegionID, S.OpValue); | |||
11382 | ExtraScheduleDataMap[I][S.OpValue] = SD; | |||
11383 | return true; | |||
11384 | }; | |||
11385 | if (CheckScheduleForI(I)) | |||
11386 | return true; | |||
11387 | if (!ScheduleStart) { | |||
11388 | // It's the first instruction in the new region. | |||
11389 | initScheduleData(I, I->getNextNode(), nullptr, nullptr); | |||
11390 | ScheduleStart = I; | |||
11391 | ScheduleEnd = I->getNextNode(); | |||
11392 | if (isOneOf(S, I) != I) | |||
11393 | CheckScheduleForI(I); | |||
11394 | 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", 11394, __extension__ __PRETTY_FUNCTION__)); | |||
11395 | 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); | |||
11396 | return true; | |||
11397 | } | |||
11398 | // Search up and down at the same time, because we don't know if the new | |||
11399 | // instruction is above or below the existing scheduling region. | |||
11400 | BasicBlock::reverse_iterator UpIter = | |||
11401 | ++ScheduleStart->getIterator().getReverse(); | |||
11402 | BasicBlock::reverse_iterator UpperEnd = BB->rend(); | |||
11403 | BasicBlock::iterator DownIter = ScheduleEnd->getIterator(); | |||
11404 | BasicBlock::iterator LowerEnd = BB->end(); | |||
11405 | while (UpIter != UpperEnd && DownIter != LowerEnd && &*UpIter != I && | |||
11406 | &*DownIter != I) { | |||
11407 | if (++ScheduleRegionSize > ScheduleRegionSizeLimit) { | |||
11408 | 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); | |||
11409 | return false; | |||
11410 | } | |||
11411 | ||||
11412 | ++UpIter; | |||
11413 | ++DownIter; | |||
11414 | } | |||
11415 | if (DownIter == LowerEnd || (UpIter != UpperEnd && &*UpIter == I)) { | |||
11416 | 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", 11417, __extension__ __PRETTY_FUNCTION__)) | |||
11417 | "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", 11417, __extension__ __PRETTY_FUNCTION__)); | |||
11418 | initScheduleData(I, ScheduleStart, nullptr, FirstLoadStoreInRegion); | |||
11419 | ScheduleStart = I; | |||
11420 | if (isOneOf(S, I) != I) | |||
11421 | CheckScheduleForI(I); | |||
11422 | 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) | |||
11423 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: extend schedule region start to " << *I << "\n"; } } while (false); | |||
11424 | return true; | |||
11425 | } | |||
11426 | 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", 11428, __extension__ __PRETTY_FUNCTION__)) | |||
11427 | "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", 11428, __extension__ __PRETTY_FUNCTION__)) | |||
11428 | "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", 11428, __extension__ __PRETTY_FUNCTION__)); | |||
11429 | 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", 11430, __extension__ __PRETTY_FUNCTION__)) | |||
11430 | "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", 11430, __extension__ __PRETTY_FUNCTION__)); | |||
11431 | initScheduleData(ScheduleEnd, I->getNextNode(), LastLoadStoreInRegion, | |||
11432 | nullptr); | |||
11433 | ScheduleEnd = I->getNextNode(); | |||
11434 | if (isOneOf(S, I) != I) | |||
11435 | CheckScheduleForI(I); | |||
11436 | 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", 11436, __extension__ __PRETTY_FUNCTION__)); | |||
11437 | 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); | |||
11438 | return true; | |||
11439 | } | |||
11440 | ||||
11441 | void BoUpSLP::BlockScheduling::initScheduleData(Instruction *FromI, | |||
11442 | Instruction *ToI, | |||
11443 | ScheduleData *PrevLoadStore, | |||
11444 | ScheduleData *NextLoadStore) { | |||
11445 | ScheduleData *CurrentLoadStore = PrevLoadStore; | |||
11446 | for (Instruction *I = FromI; I != ToI; I = I->getNextNode()) { | |||
11447 | // No need to allocate data for non-schedulable instructions. | |||
11448 | if (doesNotNeedToBeScheduled(I)) | |||
11449 | continue; | |||
11450 | ScheduleData *SD = ScheduleDataMap.lookup(I); | |||
11451 | if (!SD) { | |||
11452 | SD = allocateScheduleDataChunks(); | |||
11453 | ScheduleDataMap[I] = SD; | |||
11454 | SD->Inst = I; | |||
11455 | } | |||
11456 | 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", 11457, __extension__ __PRETTY_FUNCTION__)) | |||
11457 | "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", 11457, __extension__ __PRETTY_FUNCTION__)); | |||
11458 | SD->init(SchedulingRegionID, I); | |||
11459 | ||||
11460 | if (I->mayReadOrWriteMemory() && | |||
11461 | (!isa<IntrinsicInst>(I) || | |||
11462 | (cast<IntrinsicInst>(I)->getIntrinsicID() != Intrinsic::sideeffect && | |||
11463 | cast<IntrinsicInst>(I)->getIntrinsicID() != | |||
11464 | Intrinsic::pseudoprobe))) { | |||
11465 | // Update the linked list of memory accessing instructions. | |||
11466 | if (CurrentLoadStore) { | |||
11467 | CurrentLoadStore->NextLoadStore = SD; | |||
11468 | } else { | |||
11469 | FirstLoadStoreInRegion = SD; | |||
11470 | } | |||
11471 | CurrentLoadStore = SD; | |||
11472 | } | |||
11473 | ||||
11474 | if (match(I, m_Intrinsic<Intrinsic::stacksave>()) || | |||
11475 | match(I, m_Intrinsic<Intrinsic::stackrestore>())) | |||
11476 | RegionHasStackSave = true; | |||
11477 | } | |||
11478 | if (NextLoadStore) { | |||
11479 | if (CurrentLoadStore) | |||
11480 | CurrentLoadStore->NextLoadStore = NextLoadStore; | |||
11481 | } else { | |||
11482 | LastLoadStoreInRegion = CurrentLoadStore; | |||
11483 | } | |||
11484 | } | |||
11485 | ||||
11486 | void BoUpSLP::BlockScheduling::calculateDependencies(ScheduleData *SD, | |||
11487 | bool InsertInReadyList, | |||
11488 | BoUpSLP *SLP) { | |||
11489 | assert(SD->isSchedulingEntity())(static_cast <bool> (SD->isSchedulingEntity()) ? void (0) : __assert_fail ("SD->isSchedulingEntity()", "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 11489, __extension__ __PRETTY_FUNCTION__)); | |||
| ||||
11490 | ||||
11491 | SmallVector<ScheduleData *, 10> WorkList; | |||
11492 | WorkList.push_back(SD); | |||
11493 | ||||
11494 | while (!WorkList.empty()) { | |||
11495 | ScheduleData *SD = WorkList.pop_back_val(); | |||
11496 | for (ScheduleData *BundleMember = SD; BundleMember; | |||
11497 | BundleMember = BundleMember->NextInBundle) { | |||
11498 | assert(isInSchedulingRegion(BundleMember))(static_cast <bool> (isInSchedulingRegion(BundleMember) ) ? void (0) : __assert_fail ("isInSchedulingRegion(BundleMember)" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 11498, __extension__ __PRETTY_FUNCTION__)); | |||
11499 | if (BundleMember->hasValidDependencies()) | |||
11500 | continue; | |||
11501 | ||||
11502 | LLVM_DEBUG(dbgs() << "SLP: update deps of " << *BundleMemberdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: update deps of " << *BundleMember << "\n"; } } while (false) | |||
11503 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: update deps of " << *BundleMember << "\n"; } } while (false); | |||
11504 | BundleMember->Dependencies = 0; | |||
11505 | BundleMember->resetUnscheduledDeps(); | |||
11506 | ||||
11507 | // Handle def-use chain dependencies. | |||
11508 | if (BundleMember->OpValue != BundleMember->Inst) { | |||
11509 | if (ScheduleData *UseSD = getScheduleData(BundleMember->Inst)) { | |||
11510 | BundleMember->Dependencies++; | |||
11511 | ScheduleData *DestBundle = UseSD->FirstInBundle; | |||
11512 | if (!DestBundle->IsScheduled) | |||
11513 | BundleMember->incrementUnscheduledDeps(1); | |||
11514 | if (!DestBundle->hasValidDependencies()) | |||
11515 | WorkList.push_back(DestBundle); | |||
11516 | } | |||
11517 | } else { | |||
11518 | for (User *U : BundleMember->Inst->users()) { | |||
11519 | if (ScheduleData *UseSD = getScheduleData(cast<Instruction>(U))) { | |||
11520 | BundleMember->Dependencies++; | |||
11521 | ScheduleData *DestBundle = UseSD->FirstInBundle; | |||
11522 | if (!DestBundle->IsScheduled) | |||
11523 | BundleMember->incrementUnscheduledDeps(1); | |||
11524 | if (!DestBundle->hasValidDependencies()) | |||
11525 | WorkList.push_back(DestBundle); | |||
11526 | } | |||
11527 | } | |||
11528 | } | |||
11529 | ||||
11530 | auto makeControlDependent = [&](Instruction *I) { | |||
11531 | auto *DepDest = getScheduleData(I); | |||
11532 | 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", 11532, __extension__ __PRETTY_FUNCTION__)); | |||
11533 | DepDest->ControlDependencies.push_back(BundleMember); | |||
11534 | BundleMember->Dependencies++; | |||
11535 | ScheduleData *DestBundle = DepDest->FirstInBundle; | |||
11536 | if (!DestBundle->IsScheduled) | |||
11537 | BundleMember->incrementUnscheduledDeps(1); | |||
11538 | if (!DestBundle->hasValidDependencies()) | |||
11539 | WorkList.push_back(DestBundle); | |||
11540 | }; | |||
11541 | ||||
11542 | // Any instruction which isn't safe to speculate at the beginning of the | |||
11543 | // block is control dependend on any early exit or non-willreturn call | |||
11544 | // which proceeds it. | |||
11545 | if (!isGuaranteedToTransferExecutionToSuccessor(BundleMember->Inst)) { | |||
11546 | for (Instruction *I = BundleMember->Inst->getNextNode(); | |||
11547 | I != ScheduleEnd; I = I->getNextNode()) { | |||
11548 | if (isSafeToSpeculativelyExecute(I, &*BB->begin(), SLP->AC)) | |||
11549 | continue; | |||
11550 | ||||
11551 | // Add the dependency | |||
11552 | makeControlDependent(I); | |||
11553 | ||||
11554 | if (!isGuaranteedToTransferExecutionToSuccessor(I)) | |||
11555 | // Everything past here must be control dependent on I. | |||
11556 | break; | |||
11557 | } | |||
11558 | } | |||
11559 | ||||
11560 | if (RegionHasStackSave) { | |||
11561 | // If we have an inalloc alloca instruction, it needs to be scheduled | |||
11562 | // after any preceeding stacksave. We also need to prevent any alloca | |||
11563 | // from reordering above a preceeding stackrestore. | |||
11564 | if (match(BundleMember->Inst, m_Intrinsic<Intrinsic::stacksave>()) || | |||
11565 | match(BundleMember->Inst, m_Intrinsic<Intrinsic::stackrestore>())) { | |||
11566 | for (Instruction *I = BundleMember->Inst->getNextNode(); | |||
11567 | I != ScheduleEnd; I = I->getNextNode()) { | |||
11568 | if (match(I, m_Intrinsic<Intrinsic::stacksave>()) || | |||
11569 | match(I, m_Intrinsic<Intrinsic::stackrestore>())) | |||
11570 | // Any allocas past here must be control dependent on I, and I | |||
11571 | // must be memory dependend on BundleMember->Inst. | |||
11572 | break; | |||
11573 | ||||
11574 | if (!isa<AllocaInst>(I)) | |||
11575 | continue; | |||
11576 | ||||
11577 | // Add the dependency | |||
11578 | makeControlDependent(I); | |||
11579 | } | |||
11580 | } | |||
11581 | ||||
11582 | // In addition to the cases handle just above, we need to prevent | |||
11583 | // allocas and loads/stores from moving below a stacksave or a | |||
11584 | // stackrestore. Avoiding moving allocas below stackrestore is currently | |||
11585 | // thought to be conservatism. Moving loads/stores below a stackrestore | |||
11586 | // can lead to incorrect code. | |||
11587 | if (isa<AllocaInst>(BundleMember->Inst) || | |||
11588 | BundleMember->Inst->mayReadOrWriteMemory()) { | |||
11589 | for (Instruction *I = BundleMember->Inst->getNextNode(); | |||
11590 | I != ScheduleEnd; I = I->getNextNode()) { | |||
11591 | if (!match(I, m_Intrinsic<Intrinsic::stacksave>()) && | |||
11592 | !match(I, m_Intrinsic<Intrinsic::stackrestore>())) | |||
11593 | continue; | |||
11594 | ||||
11595 | // Add the dependency | |||
11596 | makeControlDependent(I); | |||
11597 | break; | |||
11598 | } | |||
11599 | } | |||
11600 | } | |||
11601 | ||||
11602 | // Handle the memory dependencies (if any). | |||
11603 | ScheduleData *DepDest = BundleMember->NextLoadStore; | |||
11604 | if (!DepDest) | |||
11605 | continue; | |||
11606 | Instruction *SrcInst = BundleMember->Inst; | |||
11607 | 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", 11608, __extension__ __PRETTY_FUNCTION__)) | |||
11608 | "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", 11608, __extension__ __PRETTY_FUNCTION__)); | |||
11609 | MemoryLocation SrcLoc = getLocation(SrcInst); | |||
11610 | bool SrcMayWrite = BundleMember->Inst->mayWriteToMemory(); | |||
11611 | unsigned numAliased = 0; | |||
11612 | unsigned DistToSrc = 1; | |||
11613 | ||||
11614 | for ( ; DepDest; DepDest = DepDest->NextLoadStore) { | |||
11615 | assert(isInSchedulingRegion(DepDest))(static_cast <bool> (isInSchedulingRegion(DepDest)) ? void (0) : __assert_fail ("isInSchedulingRegion(DepDest)", "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 11615, __extension__ __PRETTY_FUNCTION__)); | |||
11616 | ||||
11617 | // We have two limits to reduce the complexity: | |||
11618 | // 1) AliasedCheckLimit: It's a small limit to reduce calls to | |||
11619 | // SLP->isAliased (which is the expensive part in this loop). | |||
11620 | // 2) MaxMemDepDistance: It's for very large blocks and it aborts | |||
11621 | // the whole loop (even if the loop is fast, it's quadratic). | |||
11622 | // It's important for the loop break condition (see below) to | |||
11623 | // check this limit even between two read-only instructions. | |||
11624 | if (DistToSrc >= MaxMemDepDistance || | |||
11625 | ((SrcMayWrite || DepDest->Inst->mayWriteToMemory()) && | |||
11626 | (numAliased >= AliasedCheckLimit || | |||
11627 | SLP->isAliased(SrcLoc, SrcInst, DepDest->Inst)))) { | |||
11628 | ||||
11629 | // We increment the counter only if the locations are aliased | |||
11630 | // (instead of counting all alias checks). This gives a better | |||
11631 | // balance between reduced runtime and accurate dependencies. | |||
11632 | numAliased++; | |||
11633 | ||||
11634 | DepDest->MemoryDependencies.push_back(BundleMember); | |||
11635 | BundleMember->Dependencies++; | |||
11636 | ScheduleData *DestBundle = DepDest->FirstInBundle; | |||
11637 | if (!DestBundle->IsScheduled) { | |||
11638 | BundleMember->incrementUnscheduledDeps(1); | |||
11639 | } | |||
11640 | if (!DestBundle->hasValidDependencies()) { | |||
11641 | WorkList.push_back(DestBundle); | |||
11642 | } | |||
11643 | } | |||
11644 | ||||
11645 | // Example, explaining the loop break condition: Let's assume our | |||
11646 | // starting instruction is i0 and MaxMemDepDistance = 3. | |||
11647 | // | |||
11648 | // +--------v--v--v | |||
11649 | // i0,i1,i2,i3,i4,i5,i6,i7,i8 | |||
11650 | // +--------^--^--^ | |||
11651 | // | |||
11652 | // MaxMemDepDistance let us stop alias-checking at i3 and we add | |||
11653 | // dependencies from i0 to i3,i4,.. (even if they are not aliased). | |||
11654 | // Previously we already added dependencies from i3 to i6,i7,i8 | |||
11655 | // (because of MaxMemDepDistance). As we added a dependency from | |||
11656 | // i0 to i3, we have transitive dependencies from i0 to i6,i7,i8 | |||
11657 | // and we can abort this loop at i6. | |||
11658 | if (DistToSrc >= 2 * MaxMemDepDistance) | |||
11659 | break; | |||
11660 | DistToSrc++; | |||
11661 | } | |||
11662 | } | |||
11663 | if (InsertInReadyList && SD->isReady()) { | |||
| ||||
11664 | ReadyInsts.insert(SD); | |||
11665 | 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) | |||
11666 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: gets ready on update: " << *SD->Inst << "\n"; } } while (false); | |||
11667 | } | |||
11668 | } | |||
11669 | } | |||
11670 | ||||
11671 | void BoUpSLP::BlockScheduling::resetSchedule() { | |||
11672 | 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", 11673, __extension__ __PRETTY_FUNCTION__)) | |||
11673 | "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", 11673, __extension__ __PRETTY_FUNCTION__)); | |||
11674 | for (Instruction *I = ScheduleStart; I != ScheduleEnd; I = I->getNextNode()) { | |||
11675 | doForAllOpcodes(I, [&](ScheduleData *SD) { | |||
11676 | 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", 11677, __extension__ __PRETTY_FUNCTION__)) | |||
11677 | "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", 11677, __extension__ __PRETTY_FUNCTION__)); | |||
11678 | SD->IsScheduled = false; | |||
11679 | SD->resetUnscheduledDeps(); | |||
11680 | }); | |||
11681 | } | |||
11682 | ReadyInsts.clear(); | |||
11683 | } | |||
11684 | ||||
11685 | void BoUpSLP::scheduleBlock(BlockScheduling *BS) { | |||
11686 | if (!BS->ScheduleStart) | |||
11687 | return; | |||
11688 | ||||
11689 | 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); | |||
11690 | ||||
11691 | // A key point - if we got here, pre-scheduling was able to find a valid | |||
11692 | // scheduling of the sub-graph of the scheduling window which consists | |||
11693 | // of all vector bundles and their transitive users. As such, we do not | |||
11694 | // need to reschedule anything *outside of* that subgraph. | |||
11695 | ||||
11696 | BS->resetSchedule(); | |||
11697 | ||||
11698 | // For the real scheduling we use a more sophisticated ready-list: it is | |||
11699 | // sorted by the original instruction location. This lets the final schedule | |||
11700 | // be as close as possible to the original instruction order. | |||
11701 | // WARNING: If changing this order causes a correctness issue, that means | |||
11702 | // there is some missing dependence edge in the schedule data graph. | |||
11703 | struct ScheduleDataCompare { | |||
11704 | bool operator()(ScheduleData *SD1, ScheduleData *SD2) const { | |||
11705 | return SD2->SchedulingPriority < SD1->SchedulingPriority; | |||
11706 | } | |||
11707 | }; | |||
11708 | std::set<ScheduleData *, ScheduleDataCompare> ReadyInsts; | |||
11709 | ||||
11710 | // Ensure that all dependency data is updated (for nodes in the sub-graph) | |||
11711 | // and fill the ready-list with initial instructions. | |||
11712 | int Idx = 0; | |||
11713 | for (auto *I = BS->ScheduleStart; I != BS->ScheduleEnd; | |||
11714 | I = I->getNextNode()) { | |||
11715 | BS->doForAllOpcodes(I, [this, &Idx, BS](ScheduleData *SD) { | |||
11716 | TreeEntry *SDTE = getTreeEntry(SD->Inst); | |||
11717 | (void)SDTE; | |||
11718 | 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", 11721, __extension__ __PRETTY_FUNCTION__)) | |||
11719 | 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", 11721, __extension__ __PRETTY_FUNCTION__)) | |||
11720 | (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", 11721, __extension__ __PRETTY_FUNCTION__)) | |||
11721 | "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", 11721, __extension__ __PRETTY_FUNCTION__)); | |||
11722 | SD->FirstInBundle->SchedulingPriority = Idx++; | |||
11723 | ||||
11724 | if (SD->isSchedulingEntity() && SD->isPartOfBundle()) | |||
11725 | BS->calculateDependencies(SD, false, this); | |||
11726 | }); | |||
11727 | } | |||
11728 | BS->initialFillReadyList(ReadyInsts); | |||
11729 | ||||
11730 | Instruction *LastScheduledInst = BS->ScheduleEnd; | |||
11731 | ||||
11732 | // Do the "real" scheduling. | |||
11733 | while (!ReadyInsts.empty()) { | |||
11734 | ScheduleData *picked = *ReadyInsts.begin(); | |||
11735 | ReadyInsts.erase(ReadyInsts.begin()); | |||
11736 | ||||
11737 | // Move the scheduled instruction(s) to their dedicated places, if not | |||
11738 | // there yet. | |||
11739 | for (ScheduleData *BundleMember = picked; BundleMember; | |||
11740 | BundleMember = BundleMember->NextInBundle) { | |||
11741 | Instruction *pickedInst = BundleMember->Inst; | |||
11742 | if (pickedInst->getNextNode() != LastScheduledInst) | |||
11743 | pickedInst->moveBefore(LastScheduledInst); | |||
11744 | LastScheduledInst = pickedInst; | |||
11745 | } | |||
11746 | ||||
11747 | BS->schedule(picked, ReadyInsts); | |||
11748 | } | |||
11749 | ||||
11750 | // Check that we didn't break any of our invariants. | |||
11751 | #ifdef EXPENSIVE_CHECKS | |||
11752 | BS->verify(); | |||
11753 | #endif | |||
11754 | ||||
11755 | #if !defined(NDEBUG) || defined(EXPENSIVE_CHECKS) | |||
11756 | // Check that all schedulable entities got scheduled | |||
11757 | for (auto *I = BS->ScheduleStart; I != BS->ScheduleEnd; I = I->getNextNode()) { | |||
11758 | BS->doForAllOpcodes(I, [&](ScheduleData *SD) { | |||
11759 | if (SD->isSchedulingEntity() && SD->hasValidDependencies()) { | |||
11760 | 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", 11760, __extension__ __PRETTY_FUNCTION__)); | |||
11761 | } | |||
11762 | }); | |||
11763 | } | |||
11764 | #endif | |||
11765 | ||||
11766 | // Avoid duplicate scheduling of the block. | |||
11767 | BS->ScheduleStart = nullptr; | |||
11768 | } | |||
11769 | ||||
11770 | unsigned BoUpSLP::getVectorElementSize(Value *V) { | |||
11771 | // If V is a store, just return the width of the stored value (or value | |||
11772 | // truncated just before storing) without traversing the expression tree. | |||
11773 | // This is the common case. | |||
11774 | if (auto *Store = dyn_cast<StoreInst>(V)) | |||
11775 | return DL->getTypeSizeInBits(Store->getValueOperand()->getType()); | |||
11776 | ||||
11777 | if (auto *IEI = dyn_cast<InsertElementInst>(V)) | |||
11778 | return getVectorElementSize(IEI->getOperand(1)); | |||
11779 | ||||
11780 | auto E = InstrElementSize.find(V); | |||
11781 | if (E != InstrElementSize.end()) | |||
11782 | return E->second; | |||
11783 | ||||
11784 | // If V is not a store, we can traverse the expression tree to find loads | |||
11785 | // that feed it. The type of the loaded value may indicate a more suitable | |||
11786 | // width than V's type. We want to base the vector element size on the width | |||
11787 | // of memory operations where possible. | |||
11788 | SmallVector<std::pair<Instruction *, BasicBlock *>, 16> Worklist; | |||
11789 | SmallPtrSet<Instruction *, 16> Visited; | |||
11790 | if (auto *I = dyn_cast<Instruction>(V)) { | |||
11791 | Worklist.emplace_back(I, I->getParent()); | |||
11792 | Visited.insert(I); | |||
11793 | } | |||
11794 | ||||
11795 | // Traverse the expression tree in bottom-up order looking for loads. If we | |||
11796 | // encounter an instruction we don't yet handle, we give up. | |||
11797 | auto Width = 0u; | |||
11798 | while (!Worklist.empty()) { | |||
11799 | Instruction *I; | |||
11800 | BasicBlock *Parent; | |||
11801 | std::tie(I, Parent) = Worklist.pop_back_val(); | |||
11802 | ||||
11803 | // We should only be looking at scalar instructions here. If the current | |||
11804 | // instruction has a vector type, skip. | |||
11805 | auto *Ty = I->getType(); | |||
11806 | if (isa<VectorType>(Ty)) | |||
11807 | continue; | |||
11808 | ||||
11809 | // If the current instruction is a load, update MaxWidth to reflect the | |||
11810 | // width of the loaded value. | |||
11811 | if (isa<LoadInst, ExtractElementInst, ExtractValueInst>(I)) | |||
11812 | Width = std::max<unsigned>(Width, DL->getTypeSizeInBits(Ty)); | |||
11813 | ||||
11814 | // Otherwise, we need to visit the operands of the instruction. We only | |||
11815 | // handle the interesting cases from buildTree here. If an operand is an | |||
11816 | // instruction we haven't yet visited and from the same basic block as the | |||
11817 | // user or the use is a PHI node, we add it to the worklist. | |||
11818 | else if (isa<PHINode, CastInst, GetElementPtrInst, CmpInst, SelectInst, | |||
11819 | BinaryOperator, UnaryOperator>(I)) { | |||
11820 | for (Use &U : I->operands()) | |||
11821 | if (auto *J = dyn_cast<Instruction>(U.get())) | |||
11822 | if (Visited.insert(J).second && | |||
11823 | (isa<PHINode>(I) || J->getParent() == Parent)) | |||
11824 | Worklist.emplace_back(J, J->getParent()); | |||
11825 | } else { | |||
11826 | break; | |||
11827 | } | |||
11828 | } | |||
11829 | ||||
11830 | // If we didn't encounter a memory access in the expression tree, or if we | |||
11831 | // gave up for some reason, just return the width of V. Otherwise, return the | |||
11832 | // maximum width we found. | |||
11833 | if (!Width) { | |||
11834 | if (auto *CI = dyn_cast<CmpInst>(V)) | |||
11835 | V = CI->getOperand(0); | |||
11836 | Width = DL->getTypeSizeInBits(V->getType()); | |||
11837 | } | |||
11838 | ||||
11839 | for (Instruction *I : Visited) | |||
11840 | InstrElementSize[I] = Width; | |||
11841 | ||||
11842 | return Width; | |||
11843 | } | |||
11844 | ||||
11845 | // Determine if a value V in a vectorizable expression Expr can be demoted to a | |||
11846 | // smaller type with a truncation. We collect the values that will be demoted | |||
11847 | // in ToDemote and additional roots that require investigating in Roots. | |||
11848 | static bool collectValuesToDemote(Value *V, SmallPtrSetImpl<Value *> &Expr, | |||
11849 | SmallVectorImpl<Value *> &ToDemote, | |||
11850 | SmallVectorImpl<Value *> &Roots) { | |||
11851 | // We can always demote constants. | |||
11852 | if (isa<Constant>(V)) { | |||
11853 | ToDemote.push_back(V); | |||
11854 | return true; | |||
11855 | } | |||
11856 | ||||
11857 | // If the value is not an instruction in the expression with only one use, it | |||
11858 | // cannot be demoted. | |||
11859 | auto *I = dyn_cast<Instruction>(V); | |||
11860 | if (!I || !I->hasOneUse() || !Expr.count(I)) | |||
11861 | return false; | |||
11862 | ||||
11863 | switch (I->getOpcode()) { | |||
11864 | ||||
11865 | // We can always demote truncations and extensions. Since truncations can | |||
11866 | // seed additional demotion, we save the truncated value. | |||
11867 | case Instruction::Trunc: | |||
11868 | Roots.push_back(I->getOperand(0)); | |||
11869 | break; | |||
11870 | case Instruction::ZExt: | |||
11871 | case Instruction::SExt: | |||
11872 | if (isa<ExtractElementInst, InsertElementInst>(I->getOperand(0))) | |||
11873 | return false; | |||
11874 | break; | |||
11875 | ||||
11876 | // We can demote certain binary operations if we can demote both of their | |||
11877 | // operands. | |||
11878 | case Instruction::Add: | |||
11879 | case Instruction::Sub: | |||
11880 | case Instruction::Mul: | |||
11881 | case Instruction::And: | |||
11882 | case Instruction::Or: | |||
11883 | case Instruction::Xor: | |||
11884 | if (!collectValuesToDemote(I->getOperand(0), Expr, ToDemote, Roots) || | |||
11885 | !collectValuesToDemote(I->getOperand(1), Expr, ToDemote, Roots)) | |||
11886 | return false; | |||
11887 | break; | |||
11888 | ||||
11889 | // We can demote selects if we can demote their true and false values. | |||
11890 | case Instruction::Select: { | |||
11891 | SelectInst *SI = cast<SelectInst>(I); | |||
11892 | if (!collectValuesToDemote(SI->getTrueValue(), Expr, ToDemote, Roots) || | |||
11893 | !collectValuesToDemote(SI->getFalseValue(), Expr, ToDemote, Roots)) | |||
11894 | return false; | |||
11895 | break; | |||
11896 | } | |||
11897 | ||||
11898 | // We can demote phis if we can demote all their incoming operands. Note that | |||
11899 | // we don't need to worry about cycles since we ensure single use above. | |||
11900 | case Instruction::PHI: { | |||
11901 | PHINode *PN = cast<PHINode>(I); | |||
11902 | for (Value *IncValue : PN->incoming_values()) | |||
11903 | if (!collectValuesToDemote(IncValue, Expr, ToDemote, Roots)) | |||
11904 | return false; | |||
11905 | break; | |||
11906 | } | |||
11907 | ||||
11908 | // Otherwise, conservatively give up. | |||
11909 | default: | |||
11910 | return false; | |||
11911 | } | |||
11912 | ||||
11913 | // Record the value that we can demote. | |||
11914 | ToDemote.push_back(V); | |||
11915 | return true; | |||
11916 | } | |||
11917 | ||||
11918 | void BoUpSLP::computeMinimumValueSizes() { | |||
11919 | // If there are no external uses, the expression tree must be rooted by a | |||
11920 | // store. We can't demote in-memory values, so there is nothing to do here. | |||
11921 | if (ExternalUses.empty()) | |||
11922 | return; | |||
11923 | ||||
11924 | // We only attempt to truncate integer expressions. | |||
11925 | auto &TreeRoot = VectorizableTree[0]->Scalars; | |||
11926 | auto *TreeRootIT = dyn_cast<IntegerType>(TreeRoot[0]->getType()); | |||
11927 | if (!TreeRootIT) | |||
11928 | return; | |||
11929 | ||||
11930 | // If the expression is not rooted by a store, these roots should have | |||
11931 | // external uses. We will rely on InstCombine to rewrite the expression in | |||
11932 | // the narrower type. However, InstCombine only rewrites single-use values. | |||
11933 | // This means that if a tree entry other than a root is used externally, it | |||
11934 | // must have multiple uses and InstCombine will not rewrite it. The code | |||
11935 | // below ensures that only the roots are used externally. | |||
11936 | SmallPtrSet<Value *, 32> Expr(TreeRoot.begin(), TreeRoot.end()); | |||
11937 | for (auto &EU : ExternalUses) | |||
11938 | if (!Expr.erase(EU.Scalar)) | |||
11939 | return; | |||
11940 | if (!Expr.empty()) | |||
11941 | return; | |||
11942 | ||||
11943 | // Collect the scalar values of the vectorizable expression. We will use this | |||
11944 | // context to determine which values can be demoted. If we see a truncation, | |||
11945 | // we mark it as seeding another demotion. | |||
11946 | for (auto &EntryPtr : VectorizableTree) | |||
11947 | Expr.insert(EntryPtr->Scalars.begin(), EntryPtr->Scalars.end()); | |||
11948 | ||||
11949 | // Ensure the roots of the vectorizable tree don't form a cycle. They must | |||
11950 | // have a single external user that is not in the vectorizable tree. | |||
11951 | for (auto *Root : TreeRoot) | |||
11952 | if (!Root->hasOneUse() || Expr.count(*Root->user_begin())) | |||
11953 | return; | |||
11954 | ||||
11955 | // Conservatively determine if we can actually truncate the roots of the | |||
11956 | // expression. Collect the values that can be demoted in ToDemote and | |||
11957 | // additional roots that require investigating in Roots. | |||
11958 | SmallVector<Value *, 32> ToDemote; | |||
11959 | SmallVector<Value *, 4> Roots; | |||
11960 | for (auto *Root : TreeRoot) | |||
11961 | if (!collectValuesToDemote(Root, Expr, ToDemote, Roots)) | |||
11962 | return; | |||
11963 | ||||
11964 | // The maximum bit width required to represent all the values that can be | |||
11965 | // demoted without loss of precision. It would be safe to truncate the roots | |||
11966 | // of the expression to this width. | |||
11967 | auto MaxBitWidth = 8u; | |||
11968 | ||||
11969 | // We first check if all the bits of the roots are demanded. If they're not, | |||
11970 | // we can truncate the roots to this narrower type. | |||
11971 | for (auto *Root : TreeRoot) { | |||
11972 | auto Mask = DB->getDemandedBits(cast<Instruction>(Root)); | |||
11973 | MaxBitWidth = std::max<unsigned>(Mask.getBitWidth() - Mask.countl_zero(), | |||
11974 | MaxBitWidth); | |||
11975 | } | |||
11976 | ||||
11977 | // True if the roots can be zero-extended back to their original type, rather | |||
11978 | // than sign-extended. We know that if the leading bits are not demanded, we | |||
11979 | // can safely zero-extend. So we initialize IsKnownPositive to True. | |||
11980 | bool IsKnownPositive = true; | |||
11981 | ||||
11982 | // If all the bits of the roots are demanded, we can try a little harder to | |||
11983 | // compute a narrower type. This can happen, for example, if the roots are | |||
11984 | // getelementptr indices. InstCombine promotes these indices to the pointer | |||
11985 | // width. Thus, all their bits are technically demanded even though the | |||
11986 | // address computation might be vectorized in a smaller type. | |||
11987 | // | |||
11988 | // We start by looking at each entry that can be demoted. We compute the | |||
11989 | // maximum bit width required to store the scalar by using ValueTracking to | |||
11990 | // compute the number of high-order bits we can truncate. | |||
11991 | if (MaxBitWidth == DL->getTypeSizeInBits(TreeRoot[0]->getType()) && | |||
11992 | llvm::all_of(TreeRoot, [](Value *R) { | |||
11993 | 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", 11993, __extension__ __PRETTY_FUNCTION__)); | |||
11994 | return isa<GetElementPtrInst>(R->user_back()); | |||
11995 | })) { | |||
11996 | MaxBitWidth = 8u; | |||
11997 | ||||
11998 | // Determine if the sign bit of all the roots is known to be zero. If not, | |||
11999 | // IsKnownPositive is set to False. | |||
12000 | IsKnownPositive = llvm::all_of(TreeRoot, [&](Value *R) { | |||
12001 | KnownBits Known = computeKnownBits(R, *DL); | |||
12002 | return Known.isNonNegative(); | |||
12003 | }); | |||
12004 | ||||
12005 | // Determine the maximum number of bits required to store the scalar | |||
12006 | // values. | |||
12007 | for (auto *Scalar : ToDemote) { | |||
12008 | auto NumSignBits = ComputeNumSignBits(Scalar, *DL, 0, AC, nullptr, DT); | |||
12009 | auto NumTypeBits = DL->getTypeSizeInBits(Scalar->getType()); | |||
12010 | MaxBitWidth = std::max<unsigned>(NumTypeBits - NumSignBits, MaxBitWidth); | |||
12011 | } | |||
12012 | ||||
12013 | // If we can't prove that the sign bit is zero, we must add one to the | |||
12014 | // maximum bit width to account for the unknown sign bit. This preserves | |||
12015 | // the existing sign bit so we can safely sign-extend the root back to the | |||
12016 | // original type. Otherwise, if we know the sign bit is zero, we will | |||
12017 | // zero-extend the root instead. | |||
12018 | // | |||
12019 | // FIXME: This is somewhat suboptimal, as there will be cases where adding | |||
12020 | // one to the maximum bit width will yield a larger-than-necessary | |||
12021 | // type. In general, we need to add an extra bit only if we can't | |||
12022 | // prove that the upper bit of the original type is equal to the | |||
12023 | // upper bit of the proposed smaller type. If these two bits are the | |||
12024 | // same (either zero or one) we know that sign-extending from the | |||
12025 | // smaller type will result in the same value. Here, since we can't | |||
12026 | // yet prove this, we are just making the proposed smaller type | |||
12027 | // larger to ensure correctness. | |||
12028 | if (!IsKnownPositive) | |||
12029 | ++MaxBitWidth; | |||
12030 | } | |||
12031 | ||||
12032 | // Round MaxBitWidth up to the next power-of-two. | |||
12033 | MaxBitWidth = llvm::bit_ceil(MaxBitWidth); | |||
12034 | ||||
12035 | // If the maximum bit width we compute is less than the with of the roots' | |||
12036 | // type, we can proceed with the narrowing. Otherwise, do nothing. | |||
12037 | if (MaxBitWidth >= TreeRootIT->getBitWidth()) | |||
12038 | return; | |||
12039 | ||||
12040 | // If we can truncate the root, we must collect additional values that might | |||
12041 | // be demoted as a result. That is, those seeded by truncations we will | |||
12042 | // modify. | |||
12043 | while (!Roots.empty()) | |||
12044 | collectValuesToDemote(Roots.pop_back_val(), Expr, ToDemote, Roots); | |||
12045 | ||||
12046 | // Finally, map the values we can demote to the maximum bit with we computed. | |||
12047 | for (auto *Scalar : ToDemote) | |||
12048 | MinBWs[Scalar] = std::make_pair(MaxBitWidth, !IsKnownPositive); | |||
12049 | } | |||
12050 | ||||
12051 | PreservedAnalyses SLPVectorizerPass::run(Function &F, FunctionAnalysisManager &AM) { | |||
12052 | auto *SE = &AM.getResult<ScalarEvolutionAnalysis>(F); | |||
12053 | auto *TTI = &AM.getResult<TargetIRAnalysis>(F); | |||
12054 | auto *TLI = AM.getCachedResult<TargetLibraryAnalysis>(F); | |||
12055 | auto *AA = &AM.getResult<AAManager>(F); | |||
12056 | auto *LI = &AM.getResult<LoopAnalysis>(F); | |||
12057 | auto *DT = &AM.getResult<DominatorTreeAnalysis>(F); | |||
12058 | auto *AC = &AM.getResult<AssumptionAnalysis>(F); | |||
12059 | auto *DB = &AM.getResult<DemandedBitsAnalysis>(F); | |||
12060 | auto *ORE = &AM.getResult<OptimizationRemarkEmitterAnalysis>(F); | |||
12061 | ||||
12062 | bool Changed = runImpl(F, SE, TTI, TLI, AA, LI, DT, AC, DB, ORE); | |||
12063 | if (!Changed) | |||
12064 | return PreservedAnalyses::all(); | |||
12065 | ||||
12066 | PreservedAnalyses PA; | |||
12067 | PA.preserveSet<CFGAnalyses>(); | |||
12068 | return PA; | |||
12069 | } | |||
12070 | ||||
12071 | bool SLPVectorizerPass::runImpl(Function &F, ScalarEvolution *SE_, | |||
12072 | TargetTransformInfo *TTI_, | |||
12073 | TargetLibraryInfo *TLI_, AAResults *AA_, | |||
12074 | LoopInfo *LI_, DominatorTree *DT_, | |||
12075 | AssumptionCache *AC_, DemandedBits *DB_, | |||
12076 | OptimizationRemarkEmitter *ORE_) { | |||
12077 | if (!RunSLPVectorization) | |||
12078 | return false; | |||
12079 | SE = SE_; | |||
12080 | TTI = TTI_; | |||
12081 | TLI = TLI_; | |||
12082 | AA = AA_; | |||
12083 | LI = LI_; | |||
12084 | DT = DT_; | |||
12085 | AC = AC_; | |||
12086 | DB = DB_; | |||
12087 | DL = &F.getParent()->getDataLayout(); | |||
12088 | ||||
12089 | Stores.clear(); | |||
12090 | GEPs.clear(); | |||
12091 | bool Changed = false; | |||
12092 | ||||
12093 | // If the target claims to have no vector registers don't attempt | |||
12094 | // vectorization. | |||
12095 | if (!TTI->getNumberOfRegisters(TTI->getRegisterClassForType(true))) { | |||
12096 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Didn't find any vector registers for target, abort.\n" ; } } while (false) | |||
12097 | 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); | |||
12098 | return false; | |||
12099 | } | |||
12100 | ||||
12101 | // Don't vectorize when the attribute NoImplicitFloat is used. | |||
12102 | if (F.hasFnAttribute(Attribute::NoImplicitFloat)) | |||
12103 | return false; | |||
12104 | ||||
12105 | 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); | |||
12106 | ||||
12107 | // Use the bottom up slp vectorizer to construct chains that start with | |||
12108 | // store instructions. | |||
12109 | BoUpSLP R(&F, SE, TTI, TLI, AA, LI, DT, AC, DB, DL, ORE_); | |||
12110 | ||||
12111 | // A general note: the vectorizer must use BoUpSLP::eraseInstruction() to | |||
12112 | // delete instructions. | |||
12113 | ||||
12114 | // Update DFS numbers now so that we can use them for ordering. | |||
12115 | DT->updateDFSNumbers(); | |||
12116 | ||||
12117 | // Scan the blocks in the function in post order. | |||
12118 | for (auto *BB : post_order(&F.getEntryBlock())) { | |||
12119 | // Start new block - clear the list of reduction roots. | |||
12120 | R.clearReductionData(); | |||
12121 | collectSeedInstructions(BB); | |||
12122 | ||||
12123 | // Vectorize trees that end at stores. | |||
12124 | if (!Stores.empty()) { | |||
12125 | 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) | |||
12126 | << " underlying objects.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Found stores for " << Stores .size() << " underlying objects.\n"; } } while (false); | |||
12127 | Changed |= vectorizeStoreChains(R); | |||
12128 | } | |||
12129 | ||||
12130 | // Vectorize trees that end at reductions. | |||
12131 | Changed |= vectorizeChainsInBlock(BB, R); | |||
12132 | ||||
12133 | // Vectorize the index computations of getelementptr instructions. This | |||
12134 | // is primarily intended to catch gather-like idioms ending at | |||
12135 | // non-consecutive loads. | |||
12136 | if (!GEPs.empty()) { | |||
12137 | 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) | |||
12138 | << " underlying objects.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Found GEPs for " << GEPs .size() << " underlying objects.\n"; } } while (false); | |||
12139 | Changed |= vectorizeGEPIndices(BB, R); | |||
12140 | } | |||
12141 | } | |||
12142 | ||||
12143 | if (Changed) { | |||
12144 | R.optimizeGatherSequence(); | |||
12145 | LLVM_DEBUG(dbgs() << "SLP: vectorized \"" << F.getName() << "\"\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: vectorized \"" << F.getName () << "\"\n"; } } while (false); | |||
12146 | } | |||
12147 | return Changed; | |||
12148 | } | |||
12149 | ||||
12150 | bool SLPVectorizerPass::vectorizeStoreChain(ArrayRef<Value *> Chain, BoUpSLP &R, | |||
12151 | unsigned Idx, unsigned MinVF) { | |||
12152 | 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) | |||
12153 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing a store chain of length " << Chain.size() << "\n"; } } while (false); | |||
12154 | const unsigned Sz = R.getVectorElementSize(Chain[0]); | |||
12155 | unsigned VF = Chain.size(); | |||
12156 | ||||
12157 | if (!isPowerOf2_32(Sz) || !isPowerOf2_32(VF) || VF < 2 || VF < MinVF) | |||
12158 | return false; | |||
12159 | ||||
12160 | 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) | |||
12161 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing " << VF << " stores at offset " << Idx << "\n"; } } while ( false); | |||
12162 | ||||
12163 | R.buildTree(Chain); | |||
12164 | if (R.isTreeTinyAndNotFullyVectorizable()) | |||
12165 | return false; | |||
12166 | if (R.isLoadCombineCandidate()) | |||
12167 | return false; | |||
12168 | R.reorderTopToBottom(); | |||
12169 | R.reorderBottomToTop(); | |||
12170 | R.buildExternalUses(); | |||
12171 | ||||
12172 | R.computeMinimumValueSizes(); | |||
12173 | ||||
12174 | InstructionCost Cost = R.getTreeCost(); | |||
12175 | ||||
12176 | 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 ); | |||
12177 | if (Cost < -SLPCostThreshold) { | |||
12178 | 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); | |||
12179 | ||||
12180 | using namespace ore; | |||
12181 | ||||
12182 | R.getORE()->emit(OptimizationRemark(SV_NAME"slp-vectorizer", "StoresVectorized", | |||
12183 | cast<StoreInst>(Chain[0])) | |||
12184 | << "Stores SLP vectorized with cost " << NV("Cost", Cost) | |||
12185 | << " and with tree size " | |||
12186 | << NV("TreeSize", R.getTreeSize())); | |||
12187 | ||||
12188 | R.vectorizeTree(); | |||
12189 | return true; | |||
12190 | } | |||
12191 | ||||
12192 | return false; | |||
12193 | } | |||
12194 | ||||
12195 | bool SLPVectorizerPass::vectorizeStores(ArrayRef<StoreInst *> Stores, | |||
12196 | BoUpSLP &R) { | |||
12197 | // We may run into multiple chains that merge into a single chain. We mark the | |||
12198 | // stores that we vectorized so that we don't visit the same store twice. | |||
12199 | BoUpSLP::ValueSet VectorizedStores; | |||
12200 | bool Changed = false; | |||
12201 | ||||
12202 | int E = Stores.size(); | |||
12203 | SmallBitVector Tails(E, false); | |||
12204 | int MaxIter = MaxStoreLookup.getValue(); | |||
12205 | SmallVector<std::pair<int, int>, 16> ConsecutiveChain( | |||
12206 | E, std::make_pair(E, INT_MAX2147483647)); | |||
12207 | SmallVector<SmallBitVector, 4> CheckedPairs(E, SmallBitVector(E, false)); | |||
12208 | int IterCnt; | |||
12209 | auto &&FindConsecutiveAccess = [this, &Stores, &Tails, &IterCnt, MaxIter, | |||
12210 | &CheckedPairs, | |||
12211 | &ConsecutiveChain](int K, int Idx) { | |||
12212 | if (IterCnt >= MaxIter) | |||
12213 | return true; | |||
12214 | if (CheckedPairs[Idx].test(K)) | |||
12215 | return ConsecutiveChain[K].second == 1 && | |||
12216 | ConsecutiveChain[K].first == Idx; | |||
12217 | ++IterCnt; | |||
12218 | CheckedPairs[Idx].set(K); | |||
12219 | CheckedPairs[K].set(Idx); | |||
12220 | std::optional<int> Diff = getPointersDiff( | |||
12221 | Stores[K]->getValueOperand()->getType(), Stores[K]->getPointerOperand(), | |||
12222 | Stores[Idx]->getValueOperand()->getType(), | |||
12223 | Stores[Idx]->getPointerOperand(), *DL, *SE, /*StrictCheck=*/true); | |||
12224 | if (!Diff || *Diff == 0) | |||
12225 | return false; | |||
12226 | int Val = *Diff; | |||
12227 | if (Val < 0) { | |||
12228 | if (ConsecutiveChain[Idx].second > -Val) { | |||
12229 | Tails.set(K); | |||
12230 | ConsecutiveChain[Idx] = std::make_pair(K, -Val); | |||
12231 | } | |||
12232 | return false; | |||
12233 | } | |||
12234 | if (ConsecutiveChain[K].second <= Val) | |||
12235 | return false; | |||
12236 | ||||
12237 | Tails.set(Idx); | |||
12238 | ConsecutiveChain[K] = std::make_pair(Idx, Val); | |||
12239 | return Val == 1; | |||
12240 | }; | |||
12241 | // Do a quadratic search on all of the given stores in reverse order and find | |||
12242 | // all of the pairs of stores that follow each other. | |||
12243 | for (int Idx = E - 1; Idx >= 0; --Idx) { | |||
12244 | // If a store has multiple consecutive store candidates, search according | |||
12245 | // to the sequence: Idx-1, Idx+1, Idx-2, Idx+2, ... | |||
12246 | // This is because usually pairing with immediate succeeding or preceding | |||
12247 | // candidate create the best chance to find slp vectorization opportunity. | |||
12248 | const int MaxLookDepth = std::max(E - Idx, Idx + 1); | |||
12249 | IterCnt = 0; | |||
12250 | for (int Offset = 1, F = MaxLookDepth; Offset < F; ++Offset) | |||
12251 | if ((Idx >= Offset && FindConsecutiveAccess(Idx - Offset, Idx)) || | |||
12252 | (Idx + Offset < E && FindConsecutiveAccess(Idx + Offset, Idx))) | |||
12253 | break; | |||
12254 | } | |||
12255 | ||||
12256 | // Tracks if we tried to vectorize stores starting from the given tail | |||
12257 | // already. | |||
12258 | SmallBitVector TriedTails(E, false); | |||
12259 | // For stores that start but don't end a link in the chain: | |||
12260 | for (int Cnt = E; Cnt > 0; --Cnt) { | |||
12261 | int I = Cnt - 1; | |||
12262 | if (ConsecutiveChain[I].first == E || Tails.test(I)) | |||
12263 | continue; | |||
12264 | // We found a store instr that starts a chain. Now follow the chain and try | |||
12265 | // to vectorize it. | |||
12266 | BoUpSLP::ValueList Operands; | |||
12267 | // Collect the chain into a list. | |||
12268 | while (I != E && !VectorizedStores.count(Stores[I])) { | |||
12269 | Operands.push_back(Stores[I]); | |||
12270 | Tails.set(I); | |||
12271 | if (ConsecutiveChain[I].second != 1) { | |||
12272 | // Mark the new end in the chain and go back, if required. It might be | |||
12273 | // required if the original stores come in reversed order, for example. | |||
12274 | if (ConsecutiveChain[I].first != E && | |||
12275 | Tails.test(ConsecutiveChain[I].first) && !TriedTails.test(I) && | |||
12276 | !VectorizedStores.count(Stores[ConsecutiveChain[I].first])) { | |||
12277 | TriedTails.set(I); | |||
12278 | Tails.reset(ConsecutiveChain[I].first); | |||
12279 | if (Cnt < ConsecutiveChain[I].first + 2) | |||
12280 | Cnt = ConsecutiveChain[I].first + 2; | |||
12281 | } | |||
12282 | break; | |||
12283 | } | |||
12284 | // Move to the next value in the chain. | |||
12285 | I = ConsecutiveChain[I].first; | |||
12286 | } | |||
12287 | 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", 12287, __extension__ __PRETTY_FUNCTION__)); | |||
12288 | ||||
12289 | unsigned MaxVecRegSize = R.getMaxVecRegSize(); | |||
12290 | unsigned EltSize = R.getVectorElementSize(Operands[0]); | |||
12291 | unsigned MaxElts = llvm::bit_floor(MaxVecRegSize / EltSize); | |||
12292 | ||||
12293 | unsigned MaxVF = std::min(R.getMaximumVF(EltSize, Instruction::Store), | |||
12294 | MaxElts); | |||
12295 | auto *Store = cast<StoreInst>(Operands[0]); | |||
12296 | Type *StoreTy = Store->getValueOperand()->getType(); | |||
12297 | Type *ValueTy = StoreTy; | |||
12298 | if (auto *Trunc = dyn_cast<TruncInst>(Store->getValueOperand())) | |||
12299 | ValueTy = Trunc->getSrcTy(); | |||
12300 | unsigned MinVF = TTI->getStoreMinimumVF( | |||
12301 | R.getMinVF(DL->getTypeSizeInBits(ValueTy)), StoreTy, ValueTy); | |||
12302 | ||||
12303 | if (MaxVF <= MinVF) { | |||
12304 | LLVM_DEBUG(dbgs() << "SLP: Vectorization infeasible as MaxVF (" << MaxVF << ") <= "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Vectorization infeasible as MaxVF (" << MaxVF << ") <= " << "MinVF (" << MinVF << ")\n"; } } while (false) | |||
12305 | << "MinVF (" << MinVF << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Vectorization infeasible as MaxVF (" << MaxVF << ") <= " << "MinVF (" << MinVF << ")\n"; } } while (false); | |||
12306 | } | |||
12307 | ||||
12308 | // FIXME: Is division-by-2 the correct step? Should we assert that the | |||
12309 | // register size is a power-of-2? | |||
12310 | unsigned StartIdx = 0; | |||
12311 | for (unsigned Size = MaxVF; Size >= MinVF; Size /= 2) { | |||
12312 | for (unsigned Cnt = StartIdx, E = Operands.size(); Cnt + Size <= E;) { | |||
12313 | ArrayRef<Value *> Slice = ArrayRef(Operands).slice(Cnt, Size); | |||
12314 | if (!VectorizedStores.count(Slice.front()) && | |||
12315 | !VectorizedStores.count(Slice.back()) && | |||
12316 | vectorizeStoreChain(Slice, R, Cnt, MinVF)) { | |||
12317 | // Mark the vectorized stores so that we don't vectorize them again. | |||
12318 | VectorizedStores.insert(Slice.begin(), Slice.end()); | |||
12319 | Changed = true; | |||
12320 | // If we vectorized initial block, no need to try to vectorize it | |||
12321 | // again. | |||
12322 | if (Cnt == StartIdx) | |||
12323 | StartIdx += Size; | |||
12324 | Cnt += Size; | |||
12325 | continue; | |||
12326 | } | |||
12327 | ++Cnt; | |||
12328 | } | |||
12329 | // Check if the whole array was vectorized already - exit. | |||
12330 | if (StartIdx >= Operands.size()) | |||
12331 | break; | |||
12332 | } | |||
12333 | } | |||
12334 | ||||
12335 | return Changed; | |||
12336 | } | |||
12337 | ||||
12338 | void SLPVectorizerPass::collectSeedInstructions(BasicBlock *BB) { | |||
12339 | // Initialize the collections. We will make a single pass over the block. | |||
12340 | Stores.clear(); | |||
12341 | GEPs.clear(); | |||
12342 | ||||
12343 | // Visit the store and getelementptr instructions in BB and organize them in | |||
12344 | // Stores and GEPs according to the underlying objects of their pointer | |||
12345 | // operands. | |||
12346 | for (Instruction &I : *BB) { | |||
12347 | // Ignore store instructions that are volatile or have a pointer operand | |||
12348 | // that doesn't point to a scalar type. | |||
12349 | if (auto *SI = dyn_cast<StoreInst>(&I)) { | |||
12350 | if (!SI->isSimple()) | |||
12351 | continue; | |||
12352 | if (!isValidElementType(SI->getValueOperand()->getType())) | |||
12353 | continue; | |||
12354 | Stores[getUnderlyingObject(SI->getPointerOperand())].push_back(SI); | |||
12355 | } | |||
12356 | ||||
12357 | // Ignore getelementptr instructions that have more than one index, a | |||
12358 | // constant index, or a pointer operand that doesn't point to a scalar | |||
12359 | // type. | |||
12360 | else if (auto *GEP = dyn_cast<GetElementPtrInst>(&I)) { | |||
12361 | auto Idx = GEP->idx_begin()->get(); | |||
12362 | if (GEP->getNumIndices() > 1 || isa<Constant>(Idx)) | |||
12363 | continue; | |||
12364 | if (!isValidElementType(Idx->getType())) | |||
12365 | continue; | |||
12366 | if (GEP->getType()->isVectorTy()) | |||
12367 | continue; | |||
12368 | GEPs[GEP->getPointerOperand()].push_back(GEP); | |||
12369 | } | |||
12370 | } | |||
12371 | } | |||
12372 | ||||
12373 | bool SLPVectorizerPass::tryToVectorizePair(Value *A, Value *B, BoUpSLP &R) { | |||
12374 | if (!A || !B) | |||
12375 | return false; | |||
12376 | if (isa<InsertElementInst>(A) || isa<InsertElementInst>(B)) | |||
12377 | return false; | |||
12378 | Value *VL[] = {A, B}; | |||
12379 | return tryToVectorizeList(VL, R); | |||
12380 | } | |||
12381 | ||||
12382 | bool SLPVectorizerPass::tryToVectorizeList(ArrayRef<Value *> VL, BoUpSLP &R, | |||
12383 | bool LimitForRegisterSize) { | |||
12384 | if (VL.size() < 2) | |||
12385 | return false; | |||
12386 | ||||
12387 | 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) | |||
12388 | << VL.size() << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Trying to vectorize a list of length = " << VL.size() << ".\n"; } } while (false); | |||
12389 | ||||
12390 | // Check that all of the parts are instructions of the same type, | |||
12391 | // we permit an alternate opcode via InstructionsState. | |||
12392 | InstructionsState S = getSameOpcode(VL, *TLI); | |||
12393 | if (!S.getOpcode()) | |||
12394 | return false; | |||
12395 | ||||
12396 | Instruction *I0 = cast<Instruction>(S.OpValue); | |||
12397 | // Make sure invalid types (including vector type) are rejected before | |||
12398 | // determining vectorization factor for scalar instructions. | |||
12399 | for (Value *V : VL) { | |||
12400 | Type *Ty = V->getType(); | |||
12401 | if (!isa<InsertElementInst>(V) && !isValidElementType(Ty)) { | |||
12402 | // NOTE: the following will give user internal llvm type name, which may | |||
12403 | // not be useful. | |||
12404 | R.getORE()->emit([&]() { | |||
12405 | std::string type_str; | |||
12406 | llvm::raw_string_ostream rso(type_str); | |||
12407 | Ty->print(rso); | |||
12408 | return OptimizationRemarkMissed(SV_NAME"slp-vectorizer", "UnsupportedType", I0) | |||
12409 | << "Cannot SLP vectorize list: type " | |||
12410 | << rso.str() + " is unsupported by vectorizer"; | |||
12411 | }); | |||
12412 | return false; | |||
12413 | } | |||
12414 | } | |||
12415 | ||||
12416 | unsigned Sz = R.getVectorElementSize(I0); | |||
12417 | unsigned MinVF = R.getMinVF(Sz); | |||
12418 | unsigned MaxVF = std::max<unsigned>(llvm::bit_floor(VL.size()), MinVF); | |||
12419 | MaxVF = std::min(R.getMaximumVF(Sz, S.getOpcode()), MaxVF); | |||
12420 | if (MaxVF < 2) { | |||
12421 | R.getORE()->emit([&]() { | |||
12422 | return OptimizationRemarkMissed(SV_NAME"slp-vectorizer", "SmallVF", I0) | |||
12423 | << "Cannot SLP vectorize list: vectorization factor " | |||
12424 | << "less than 2 is not supported"; | |||
12425 | }); | |||
12426 | return false; | |||
12427 | } | |||
12428 | ||||
12429 | bool Changed = false; | |||
12430 | bool CandidateFound = false; | |||
12431 | InstructionCost MinCost = SLPCostThreshold.getValue(); | |||
12432 | Type *ScalarTy = VL[0]->getType(); | |||
12433 | if (auto *IE = dyn_cast<InsertElementInst>(VL[0])) | |||
12434 | ScalarTy = IE->getOperand(1)->getType(); | |||
12435 | ||||
12436 | unsigned NextInst = 0, MaxInst = VL.size(); | |||
12437 | for (unsigned VF = MaxVF; NextInst + 1 < MaxInst && VF >= MinVF; VF /= 2) { | |||
12438 | // No actual vectorization should happen, if number of parts is the same as | |||
12439 | // provided vectorization factor (i.e. the scalar type is used for vector | |||
12440 | // code during codegen). | |||
12441 | auto *VecTy = FixedVectorType::get(ScalarTy, VF); | |||
12442 | if (TTI->getNumberOfParts(VecTy) == VF) | |||
12443 | continue; | |||
12444 | for (unsigned I = NextInst; I < MaxInst; ++I) { | |||
12445 | unsigned OpsWidth = 0; | |||
12446 | ||||
12447 | if (I + VF > MaxInst) | |||
12448 | OpsWidth = MaxInst - I; | |||
12449 | else | |||
12450 | OpsWidth = VF; | |||
12451 | ||||
12452 | if (!isPowerOf2_32(OpsWidth)) | |||
12453 | continue; | |||
12454 | ||||
12455 | if ((LimitForRegisterSize && OpsWidth < MaxVF) || | |||
12456 | (VF > MinVF && OpsWidth <= VF / 2) || (VF == MinVF && OpsWidth < 2)) | |||
12457 | break; | |||
12458 | ||||
12459 | ArrayRef<Value *> Ops = VL.slice(I, OpsWidth); | |||
12460 | // Check that a previous iteration of this loop did not delete the Value. | |||
12461 | if (llvm::any_of(Ops, [&R](Value *V) { | |||
12462 | auto *I = dyn_cast<Instruction>(V); | |||
12463 | return I && R.isDeleted(I); | |||
12464 | })) | |||
12465 | continue; | |||
12466 | ||||
12467 | LLVM_DEBUG(dbgs() << "SLP: Analyzing " << OpsWidth << " operations "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing " << OpsWidth << " operations " << "\n"; } } while (false) | |||
12468 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing " << OpsWidth << " operations " << "\n"; } } while (false); | |||
12469 | ||||
12470 | R.buildTree(Ops); | |||
12471 | if (R.isTreeTinyAndNotFullyVectorizable()) | |||
12472 | continue; | |||
12473 | R.reorderTopToBottom(); | |||
12474 | R.reorderBottomToTop( | |||
12475 | /*IgnoreReorder=*/!isa<InsertElementInst>(Ops.front()) && | |||
12476 | !R.doesRootHaveInTreeUses()); | |||
12477 | R.buildExternalUses(); | |||
12478 | ||||
12479 | R.computeMinimumValueSizes(); | |||
12480 | InstructionCost Cost = R.getTreeCost(); | |||
12481 | CandidateFound = true; | |||
12482 | MinCost = std::min(MinCost, Cost); | |||
12483 | ||||
12484 | LLVM_DEBUG(dbgs() << "SLP: Found cost = " << Costdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Found cost = " << Cost << " for VF=" << OpsWidth << "\n"; } } while (false) | |||
12485 | << " for VF=" << OpsWidth << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Found cost = " << Cost << " for VF=" << OpsWidth << "\n"; } } while (false); | |||
12486 | if (Cost < -SLPCostThreshold) { | |||
12487 | 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); | |||
12488 | R.getORE()->emit(OptimizationRemark(SV_NAME"slp-vectorizer", "VectorizedList", | |||
12489 | cast<Instruction>(Ops[0])) | |||
12490 | << "SLP vectorized with cost " << ore::NV("Cost", Cost) | |||
12491 | << " and with tree size " | |||
12492 | << ore::NV("TreeSize", R.getTreeSize())); | |||
12493 | ||||
12494 | R.vectorizeTree(); | |||
12495 | // Move to the next bundle. | |||
12496 | I += VF - 1; | |||
12497 | NextInst = I + 1; | |||
12498 | Changed = true; | |||
12499 | } | |||
12500 | } | |||
12501 | } | |||
12502 | ||||
12503 | if (!Changed && CandidateFound) { | |||
12504 | R.getORE()->emit([&]() { | |||
12505 | return OptimizationRemarkMissed(SV_NAME"slp-vectorizer", "NotBeneficial", I0) | |||
12506 | << "List vectorization was possible but not beneficial with cost " | |||
12507 | << ore::NV("Cost", MinCost) << " >= " | |||
12508 | << ore::NV("Treshold", -SLPCostThreshold); | |||
12509 | }); | |||
12510 | } else if (!Changed) { | |||
12511 | R.getORE()->emit([&]() { | |||
12512 | return OptimizationRemarkMissed(SV_NAME"slp-vectorizer", "NotPossible", I0) | |||
12513 | << "Cannot SLP vectorize list: vectorization was impossible" | |||
12514 | << " with available vectorization factors"; | |||
12515 | }); | |||
12516 | } | |||
12517 | return Changed; | |||
12518 | } | |||
12519 | ||||
12520 | bool SLPVectorizerPass::tryToVectorize(Instruction *I, BoUpSLP &R) { | |||
12521 | if (!I) | |||
12522 | return false; | |||
12523 | ||||
12524 | if (!isa<BinaryOperator, CmpInst>(I) || isa<VectorType>(I->getType())) | |||
12525 | return false; | |||
12526 | ||||
12527 | Value *P = I->getParent(); | |||
12528 | ||||
12529 | // Vectorize in current basic block only. | |||
12530 | auto *Op0 = dyn_cast<Instruction>(I->getOperand(0)); | |||
12531 | auto *Op1 = dyn_cast<Instruction>(I->getOperand(1)); | |||
12532 | if (!Op0 || !Op1 || Op0->getParent() != P || Op1->getParent() != P) | |||
12533 | return false; | |||
12534 | ||||
12535 | // First collect all possible candidates | |||
12536 | SmallVector<std::pair<Value *, Value *>, 4> Candidates; | |||
12537 | Candidates.emplace_back(Op0, Op1); | |||
12538 | ||||
12539 | auto *A = dyn_cast<BinaryOperator>(Op0); | |||
12540 | auto *B = dyn_cast<BinaryOperator>(Op1); | |||
12541 | // Try to skip B. | |||
12542 | if (A && B && B->hasOneUse()) { | |||
12543 | auto *B0 = dyn_cast<BinaryOperator>(B->getOperand(0)); | |||
12544 | auto *B1 = dyn_cast<BinaryOperator>(B->getOperand(1)); | |||
12545 | if (B0 && B0->getParent() == P) | |||
12546 | Candidates.emplace_back(A, B0); | |||
12547 | if (B1 && B1->getParent() == P) | |||
12548 | Candidates.emplace_back(A, B1); | |||
12549 | } | |||
12550 | // Try to skip A. | |||
12551 | if (B && A && A->hasOneUse()) { | |||
12552 | auto *A0 = dyn_cast<BinaryOperator>(A->getOperand(0)); | |||
12553 | auto *A1 = dyn_cast<BinaryOperator>(A->getOperand(1)); | |||
12554 | if (A0 && A0->getParent() == P) | |||
12555 | Candidates.emplace_back(A0, B); | |||
12556 | if (A1 && A1->getParent() == P) | |||
12557 | Candidates.emplace_back(A1, B); | |||
12558 | } | |||
12559 | ||||
12560 | if (Candidates.size() == 1) | |||
12561 | return tryToVectorizePair(Op0, Op1, R); | |||
12562 | ||||
12563 | // We have multiple options. Try to pick the single best. | |||
12564 | std::optional<int> BestCandidate = R.findBestRootPair(Candidates); | |||
12565 | if (!BestCandidate) | |||
12566 | return false; | |||
12567 | return tryToVectorizePair(Candidates[*BestCandidate].first, | |||
12568 | Candidates[*BestCandidate].second, R); | |||
12569 | } | |||
12570 | ||||
12571 | namespace { | |||
12572 | ||||
12573 | /// Model horizontal reductions. | |||
12574 | /// | |||
12575 | /// A horizontal reduction is a tree of reduction instructions that has values | |||
12576 | /// that can be put into a vector as its leaves. For example: | |||
12577 | /// | |||
12578 | /// mul mul mul mul | |||
12579 | /// \ / \ / | |||
12580 | /// + + | |||
12581 | /// \ / | |||
12582 | /// + | |||
12583 | /// This tree has "mul" as its leaf values and "+" as its reduction | |||
12584 | /// instructions. A reduction can feed into a store or a binary operation | |||
12585 | /// feeding a phi. | |||
12586 | /// ... | |||
12587 | /// \ / | |||
12588 | /// + | |||
12589 | /// | | |||
12590 | /// phi += | |||
12591 | /// | |||
12592 | /// Or: | |||
12593 | /// ... | |||
12594 | /// \ / | |||
12595 | /// + | |||
12596 | /// | | |||
12597 | /// *p = | |||
12598 | /// | |||
12599 | class HorizontalReduction { | |||
12600 | using ReductionOpsType = SmallVector<Value *, 16>; | |||
12601 | using ReductionOpsListType = SmallVector<ReductionOpsType, 2>; | |||
12602 | ReductionOpsListType ReductionOps; | |||
12603 | /// List of possibly reduced values. | |||
12604 | SmallVector<SmallVector<Value *>> ReducedVals; | |||
12605 | /// Maps reduced value to the corresponding reduction operation. | |||
12606 | DenseMap<Value *, SmallVector<Instruction *>> ReducedValsToOps; | |||
12607 | // Use map vector to make stable output. | |||
12608 | MapVector<Instruction *, Value *> ExtraArgs; | |||
12609 | WeakTrackingVH ReductionRoot; | |||
12610 | /// The type of reduction operation. | |||
12611 | RecurKind RdxKind; | |||
12612 | /// Checks if the optimization of original scalar identity operations on | |||
12613 | /// matched horizontal reductions is enabled and allowed. | |||
12614 | bool IsSupportedHorRdxIdentityOp = false; | |||
12615 | ||||
12616 | static bool isCmpSelMinMax(Instruction *I) { | |||
12617 | return match(I, m_Select(m_Cmp(), m_Value(), m_Value())) && | |||
12618 | RecurrenceDescriptor::isMinMaxRecurrenceKind(getRdxKind(I)); | |||
12619 | } | |||
12620 | ||||
12621 | // And/or are potentially poison-safe logical patterns like: | |||
12622 | // select x, y, false | |||
12623 | // select x, true, y | |||
12624 | static bool isBoolLogicOp(Instruction *I) { | |||
12625 | return isa<SelectInst>(I) && | |||
12626 | (match(I, m_LogicalAnd()) || match(I, m_LogicalOr())); | |||
12627 | } | |||
12628 | ||||
12629 | /// Checks if instruction is associative and can be vectorized. | |||
12630 | static bool isVectorizable(RecurKind Kind, Instruction *I) { | |||
12631 | if (Kind == RecurKind::None) | |||
12632 | return false; | |||
12633 | ||||
12634 | // Integer ops that map to select instructions or intrinsics are fine. | |||
12635 | if (RecurrenceDescriptor::isIntMinMaxRecurrenceKind(Kind) || | |||
12636 | isBoolLogicOp(I)) | |||
12637 | return true; | |||
12638 | ||||
12639 | if (Kind == RecurKind::FMax || Kind == RecurKind::FMin) { | |||
12640 | // FP min/max are associative except for NaN and -0.0. We do not | |||
12641 | // have to rule out -0.0 here because the intrinsic semantics do not | |||
12642 | // specify a fixed result for it. | |||
12643 | return I->getFastMathFlags().noNaNs(); | |||
12644 | } | |||
12645 | ||||
12646 | return I->isAssociative(); | |||
12647 | } | |||
12648 | ||||
12649 | static Value *getRdxOperand(Instruction *I, unsigned Index) { | |||
12650 | // Poison-safe 'or' takes the form: select X, true, Y | |||
12651 | // To make that work with the normal operand processing, we skip the | |||
12652 | // true value operand. | |||
12653 | // TODO: Change the code and data structures to handle this without a hack. | |||
12654 | if (getRdxKind(I) == RecurKind::Or && isa<SelectInst>(I) && Index == 1) | |||
12655 | return I->getOperand(2); | |||
12656 | return I->getOperand(Index); | |||
12657 | } | |||
12658 | ||||
12659 | /// Creates reduction operation with the current opcode. | |||
12660 | static Value *createOp(IRBuilder<> &Builder, RecurKind Kind, Value *LHS, | |||
12661 | Value *RHS, const Twine &Name, bool UseSelect) { | |||
12662 | unsigned RdxOpcode = RecurrenceDescriptor::getOpcode(Kind); | |||
12663 | bool IsConstant = isConstant(LHS) && isConstant(RHS); | |||
12664 | switch (Kind) { | |||
12665 | case RecurKind::Or: | |||
12666 | if (UseSelect && | |||
12667 | LHS->getType() == CmpInst::makeCmpResultType(LHS->getType())) | |||
12668 | return Builder.CreateSelect(LHS, Builder.getTrue(), RHS, Name); | |||
12669 | return Builder.CreateBinOp((Instruction::BinaryOps)RdxOpcode, LHS, RHS, | |||
12670 | Name); | |||
12671 | case RecurKind::And: | |||
12672 | if (UseSelect && | |||
12673 | LHS->getType() == CmpInst::makeCmpResultType(LHS->getType())) | |||
12674 | return Builder.CreateSelect(LHS, RHS, Builder.getFalse(), Name); | |||
12675 | return Builder.CreateBinOp((Instruction::BinaryOps)RdxOpcode, LHS, RHS, | |||
12676 | Name); | |||
12677 | case RecurKind::Add: | |||
12678 | case RecurKind::Mul: | |||
12679 | case RecurKind::Xor: | |||
12680 | case RecurKind::FAdd: | |||
12681 | case RecurKind::FMul: | |||
12682 | return Builder.CreateBinOp((Instruction::BinaryOps)RdxOpcode, LHS, RHS, | |||
12683 | Name); | |||
12684 | case RecurKind::FMax: | |||
12685 | if (IsConstant) | |||
12686 | return ConstantFP::get(LHS->getType(), | |||
12687 | maxnum(cast<ConstantFP>(LHS)->getValueAPF(), | |||
12688 | cast<ConstantFP>(RHS)->getValueAPF())); | |||
12689 | return Builder.CreateBinaryIntrinsic(Intrinsic::maxnum, LHS, RHS); | |||
12690 | case RecurKind::FMin: | |||
12691 | if (IsConstant) | |||
12692 | return ConstantFP::get(LHS->getType(), | |||
12693 | minnum(cast<ConstantFP>(LHS)->getValueAPF(), | |||
12694 | cast<ConstantFP>(RHS)->getValueAPF())); | |||
12695 | return Builder.CreateBinaryIntrinsic(Intrinsic::minnum, LHS, RHS); | |||
12696 | case RecurKind::SMax: | |||
12697 | if (IsConstant || UseSelect) { | |||
12698 | Value *Cmp = Builder.CreateICmpSGT(LHS, RHS, Name); | |||
12699 | return Builder.CreateSelect(Cmp, LHS, RHS, Name); | |||
12700 | } | |||
12701 | return Builder.CreateBinaryIntrinsic(Intrinsic::smax, LHS, RHS); | |||
12702 | case RecurKind::SMin: | |||
12703 | if (IsConstant || UseSelect) { | |||
12704 | Value *Cmp = Builder.CreateICmpSLT(LHS, RHS, Name); | |||
12705 | return Builder.CreateSelect(Cmp, LHS, RHS, Name); | |||
12706 | } | |||
12707 | return Builder.CreateBinaryIntrinsic(Intrinsic::smin, LHS, RHS); | |||
12708 | case RecurKind::UMax: | |||
12709 | if (IsConstant || UseSelect) { | |||
12710 | Value *Cmp = Builder.CreateICmpUGT(LHS, RHS, Name); | |||
12711 | return Builder.CreateSelect(Cmp, LHS, RHS, Name); | |||
12712 | } | |||
12713 | return Builder.CreateBinaryIntrinsic(Intrinsic::umax, LHS, RHS); | |||
12714 | case RecurKind::UMin: | |||
12715 | if (IsConstant || UseSelect) { | |||
12716 | Value *Cmp = Builder.CreateICmpULT(LHS, RHS, Name); | |||
12717 | return Builder.CreateSelect(Cmp, LHS, RHS, Name); | |||
12718 | } | |||
12719 | return Builder.CreateBinaryIntrinsic(Intrinsic::umin, LHS, RHS); | |||
12720 | default: | |||
12721 | llvm_unreachable("Unknown reduction operation.")::llvm::llvm_unreachable_internal("Unknown reduction operation." , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 12721); | |||
12722 | } | |||
12723 | } | |||
12724 | ||||
12725 | /// Creates reduction operation with the current opcode with the IR flags | |||
12726 | /// from \p ReductionOps, dropping nuw/nsw flags. | |||
12727 | static Value *createOp(IRBuilder<> &Builder, RecurKind RdxKind, Value *LHS, | |||
12728 | Value *RHS, const Twine &Name, | |||
12729 | const ReductionOpsListType &ReductionOps) { | |||
12730 | bool UseSelect = ReductionOps.size() == 2 || | |||
12731 | // Logical or/and. | |||
12732 | (ReductionOps.size() == 1 && | |||
12733 | isa<SelectInst>(ReductionOps.front().front())); | |||
12734 | 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", 12736, __extension__ __PRETTY_FUNCTION__)) | |||
12735 | 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", 12736, __extension__ __PRETTY_FUNCTION__)) | |||
12736 | "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", 12736, __extension__ __PRETTY_FUNCTION__)); | |||
12737 | Value *Op = createOp(Builder, RdxKind, LHS, RHS, Name, UseSelect); | |||
12738 | if (RecurrenceDescriptor::isIntMinMaxRecurrenceKind(RdxKind)) { | |||
12739 | if (auto *Sel = dyn_cast<SelectInst>(Op)) { | |||
12740 | propagateIRFlags(Sel->getCondition(), ReductionOps[0], nullptr, | |||
12741 | /*IncludeWrapFlags=*/false); | |||
12742 | propagateIRFlags(Op, ReductionOps[1], nullptr, | |||
12743 | /*IncludeWrapFlags=*/false); | |||
12744 | return Op; | |||
12745 | } | |||
12746 | } | |||
12747 | propagateIRFlags(Op, ReductionOps[0], nullptr, /*IncludeWrapFlags=*/false); | |||
12748 | return Op; | |||
12749 | } | |||
12750 | ||||
12751 | public: | |||
12752 | static RecurKind getRdxKind(Value *V) { | |||
12753 | auto *I = dyn_cast<Instruction>(V); | |||
12754 | if (!I) | |||
12755 | return RecurKind::None; | |||
12756 | if (match(I, m_Add(m_Value(), m_Value()))) | |||
12757 | return RecurKind::Add; | |||
12758 | if (match(I, m_Mul(m_Value(), m_Value()))) | |||
12759 | return RecurKind::Mul; | |||
12760 | if (match(I, m_And(m_Value(), m_Value())) || | |||
12761 | match(I, m_LogicalAnd(m_Value(), m_Value()))) | |||
12762 | return RecurKind::And; | |||
12763 | if (match(I, m_Or(m_Value(), m_Value())) || | |||
12764 | match(I, m_LogicalOr(m_Value(), m_Value()))) | |||
12765 | return RecurKind::Or; | |||
12766 | if (match(I, m_Xor(m_Value(), m_Value()))) | |||
12767 | return RecurKind::Xor; | |||
12768 | if (match(I, m_FAdd(m_Value(), m_Value()))) | |||
12769 | return RecurKind::FAdd; | |||
12770 | if (match(I, m_FMul(m_Value(), m_Value()))) | |||
12771 | return RecurKind::FMul; | |||
12772 | ||||
12773 | if (match(I, m_Intrinsic<Intrinsic::maxnum>(m_Value(), m_Value()))) | |||
12774 | return RecurKind::FMax; | |||
12775 | if (match(I, m_Intrinsic<Intrinsic::minnum>(m_Value(), m_Value()))) | |||
12776 | return RecurKind::FMin; | |||
12777 | ||||
12778 | // This matches either cmp+select or intrinsics. SLP is expected to handle | |||
12779 | // either form. | |||
12780 | // TODO: If we are canonicalizing to intrinsics, we can remove several | |||
12781 | // special-case paths that deal with selects. | |||
12782 | if (match(I, m_SMax(m_Value(), m_Value()))) | |||
12783 | return RecurKind::SMax; | |||
12784 | if (match(I, m_SMin(m_Value(), m_Value()))) | |||
12785 | return RecurKind::SMin; | |||
12786 | if (match(I, m_UMax(m_Value(), m_Value()))) | |||
12787 | return RecurKind::UMax; | |||
12788 | if (match(I, m_UMin(m_Value(), m_Value()))) | |||
12789 | return RecurKind::UMin; | |||
12790 | ||||
12791 | if (auto *Select = dyn_cast<SelectInst>(I)) { | |||
12792 | // Try harder: look for min/max pattern based on instructions producing | |||
12793 | // same values such as: select ((cmp Inst1, Inst2), Inst1, Inst2). | |||
12794 | // During the intermediate stages of SLP, it's very common to have | |||
12795 | // pattern like this (since optimizeGatherSequence is run only once | |||
12796 | // at the end): | |||
12797 | // %1 = extractelement <2 x i32> %a, i32 0 | |||
12798 | // %2 = extractelement <2 x i32> %a, i32 1 | |||
12799 | // %cond = icmp sgt i32 %1, %2 | |||
12800 | // %3 = extractelement <2 x i32> %a, i32 0 | |||
12801 | // %4 = extractelement <2 x i32> %a, i32 1 | |||
12802 | // %select = select i1 %cond, i32 %3, i32 %4 | |||
12803 | CmpInst::Predicate Pred; | |||
12804 | Instruction *L1; | |||
12805 | Instruction *L2; | |||
12806 | ||||
12807 | Value *LHS = Select->getTrueValue(); | |||
12808 | Value *RHS = Select->getFalseValue(); | |||
12809 | Value *Cond = Select->getCondition(); | |||
12810 | ||||
12811 | // TODO: Support inverse predicates. | |||
12812 | if (match(Cond, m_Cmp(Pred, m_Specific(LHS), m_Instruction(L2)))) { | |||
12813 | if (!isa<ExtractElementInst>(RHS) || | |||
12814 | !L2->isIdenticalTo(cast<Instruction>(RHS))) | |||
12815 | return RecurKind::None; | |||
12816 | } else if (match(Cond, m_Cmp(Pred, m_Instruction(L1), m_Specific(RHS)))) { | |||
12817 | if (!isa<ExtractElementInst>(LHS) || | |||
12818 | !L1->isIdenticalTo(cast<Instruction>(LHS))) | |||
12819 | return RecurKind::None; | |||
12820 | } else { | |||
12821 | if (!isa<ExtractElementInst>(LHS) || !isa<ExtractElementInst>(RHS)) | |||
12822 | return RecurKind::None; | |||
12823 | if (!match(Cond, m_Cmp(Pred, m_Instruction(L1), m_Instruction(L2))) || | |||
12824 | !L1->isIdenticalTo(cast<Instruction>(LHS)) || | |||
12825 | !L2->isIdenticalTo(cast<Instruction>(RHS))) | |||
12826 | return RecurKind::None; | |||
12827 | } | |||
12828 | ||||
12829 | switch (Pred) { | |||
12830 | default: | |||
12831 | return RecurKind::None; | |||
12832 | case CmpInst::ICMP_SGT: | |||
12833 | case CmpInst::ICMP_SGE: | |||
12834 | return RecurKind::SMax; | |||
12835 | case CmpInst::ICMP_SLT: | |||
12836 | case CmpInst::ICMP_SLE: | |||
12837 | return RecurKind::SMin; | |||
12838 | case CmpInst::ICMP_UGT: | |||
12839 | case CmpInst::ICMP_UGE: | |||
12840 | return RecurKind::UMax; | |||
12841 | case CmpInst::ICMP_ULT: | |||
12842 | case CmpInst::ICMP_ULE: | |||
12843 | return RecurKind::UMin; | |||
12844 | } | |||
12845 | } | |||
12846 | return RecurKind::None; | |||
12847 | } | |||
12848 | ||||
12849 | /// Get the index of the first operand. | |||
12850 | static unsigned getFirstOperandIndex(Instruction *I) { | |||
12851 | return isCmpSelMinMax(I) ? 1 : 0; | |||
12852 | } | |||
12853 | ||||
12854 | private: | |||
12855 | /// Total number of operands in the reduction operation. | |||
12856 | static unsigned getNumberOfOperands(Instruction *I) { | |||
12857 | return isCmpSelMinMax(I) ? 3 : 2; | |||
12858 | } | |||
12859 | ||||
12860 | /// Checks if the instruction is in basic block \p BB. | |||
12861 | /// For a cmp+sel min/max reduction check that both ops are in \p BB. | |||
12862 | static bool hasSameParent(Instruction *I, BasicBlock *BB) { | |||
12863 | if (isCmpSelMinMax(I) || isBoolLogicOp(I)) { | |||
12864 | auto *Sel = cast<SelectInst>(I); | |||
12865 | auto *Cmp = dyn_cast<Instruction>(Sel->getCondition()); | |||
12866 | return Sel->getParent() == BB && Cmp && Cmp->getParent() == BB; | |||
12867 | } | |||
12868 | return I->getParent() == BB; | |||
12869 | } | |||
12870 | ||||
12871 | /// Expected number of uses for reduction operations/reduced values. | |||
12872 | static bool hasRequiredNumberOfUses(bool IsCmpSelMinMax, Instruction *I) { | |||
12873 | if (IsCmpSelMinMax) { | |||
12874 | // SelectInst must be used twice while the condition op must have single | |||
12875 | // use only. | |||
12876 | if (auto *Sel = dyn_cast<SelectInst>(I)) | |||
12877 | return Sel->hasNUses(2) && Sel->getCondition()->hasOneUse(); | |||
12878 | return I->hasNUses(2); | |||
12879 | } | |||
12880 | ||||
12881 | // Arithmetic reduction operation must be used once only. | |||
12882 | return I->hasOneUse(); | |||
12883 | } | |||
12884 | ||||
12885 | /// Initializes the list of reduction operations. | |||
12886 | void initReductionOps(Instruction *I) { | |||
12887 | if (isCmpSelMinMax(I)) | |||
12888 | ReductionOps.assign(2, ReductionOpsType()); | |||
12889 | else | |||
12890 | ReductionOps.assign(1, ReductionOpsType()); | |||
12891 | } | |||
12892 | ||||
12893 | /// Add all reduction operations for the reduction instruction \p I. | |||
12894 | void addReductionOps(Instruction *I) { | |||
12895 | if (isCmpSelMinMax(I)) { | |||
12896 | ReductionOps[0].emplace_back(cast<SelectInst>(I)->getCondition()); | |||
12897 | ReductionOps[1].emplace_back(I); | |||
12898 | } else { | |||
12899 | ReductionOps[0].emplace_back(I); | |||
12900 | } | |||
12901 | } | |||
12902 | ||||
12903 | static bool isGoodForReduction(ArrayRef<Value *> Data) { | |||
12904 | int Sz = Data.size(); | |||
12905 | auto *I = dyn_cast<Instruction>(Data.front()); | |||
12906 | return Sz > 1 || isConstant(Data.front()) || | |||
12907 | (I && !isa<LoadInst>(I) && isValidForAlternation(I->getOpcode())); | |||
12908 | } | |||
12909 | ||||
12910 | public: | |||
12911 | HorizontalReduction() = default; | |||
12912 | ||||
12913 | /// Try to find a reduction tree. | |||
12914 | bool matchAssociativeReduction(BoUpSLP &R, Instruction *Root, | |||
12915 | ScalarEvolution &SE, const DataLayout &DL, | |||
12916 | const TargetLibraryInfo &TLI) { | |||
12917 | RdxKind = HorizontalReduction::getRdxKind(Root); | |||
12918 | if (!isVectorizable(RdxKind, Root)) | |||
12919 | return false; | |||
12920 | ||||
12921 | // Analyze "regular" integer/FP types for reductions - no target-specific | |||
12922 | // types or pointers. | |||
12923 | Type *Ty = Root->getType(); | |||
12924 | if (!isValidElementType(Ty) || Ty->isPointerTy()) | |||
12925 | return false; | |||
12926 | ||||
12927 | // Though the ultimate reduction may have multiple uses, its condition must | |||
12928 | // have only single use. | |||
12929 | if (auto *Sel = dyn_cast<SelectInst>(Root)) | |||
12930 | if (!Sel->getCondition()->hasOneUse()) | |||
12931 | return false; | |||
12932 | ||||
12933 | ReductionRoot = Root; | |||
12934 | ||||
12935 | // Iterate through all the operands of the possible reduction tree and | |||
12936 | // gather all the reduced values, sorting them by their value id. | |||
12937 | BasicBlock *BB = Root->getParent(); | |||
12938 | bool IsCmpSelMinMax = isCmpSelMinMax(Root); | |||
12939 | SmallVector<Instruction *> Worklist(1, Root); | |||
12940 | // Checks if the operands of the \p TreeN instruction are also reduction | |||
12941 | // operations or should be treated as reduced values or an extra argument, | |||
12942 | // which is not part of the reduction. | |||
12943 | auto CheckOperands = [&](Instruction *TreeN, | |||
12944 | SmallVectorImpl<Value *> &ExtraArgs, | |||
12945 | SmallVectorImpl<Value *> &PossibleReducedVals, | |||
12946 | SmallVectorImpl<Instruction *> &ReductionOps) { | |||
12947 | for (int I = getFirstOperandIndex(TreeN), | |||
12948 | End = getNumberOfOperands(TreeN); | |||
12949 | I < End; ++I) { | |||
12950 | Value *EdgeVal = getRdxOperand(TreeN, I); | |||
12951 | ReducedValsToOps[EdgeVal].push_back(TreeN); | |||
12952 | auto *EdgeInst = dyn_cast<Instruction>(EdgeVal); | |||
12953 | // Edge has wrong parent - mark as an extra argument. | |||
12954 | if (EdgeInst && !isVectorLikeInstWithConstOps(EdgeInst) && | |||
12955 | !hasSameParent(EdgeInst, BB)) { | |||
12956 | ExtraArgs.push_back(EdgeVal); | |||
12957 | continue; | |||
12958 | } | |||
12959 | // If the edge is not an instruction, or it is different from the main | |||
12960 | // reduction opcode or has too many uses - possible reduced value. | |||
12961 | // Also, do not try to reduce const values, if the operation is not | |||
12962 | // foldable. | |||
12963 | if (!EdgeInst || getRdxKind(EdgeInst) != RdxKind || | |||
12964 | IsCmpSelMinMax != isCmpSelMinMax(EdgeInst) || | |||
12965 | !hasRequiredNumberOfUses(IsCmpSelMinMax, EdgeInst) || | |||
12966 | !isVectorizable(RdxKind, EdgeInst) || | |||
12967 | (R.isAnalyzedReductionRoot(EdgeInst) && | |||
12968 | all_of(EdgeInst->operands(), Constant::classof))) { | |||
12969 | PossibleReducedVals.push_back(EdgeVal); | |||
12970 | continue; | |||
12971 | } | |||
12972 | ReductionOps.push_back(EdgeInst); | |||
12973 | } | |||
12974 | }; | |||
12975 | // Try to regroup reduced values so that it gets more profitable to try to | |||
12976 | // reduce them. Values are grouped by their value ids, instructions - by | |||
12977 | // instruction op id and/or alternate op id, plus do extra analysis for | |||
12978 | // loads (grouping them by the distabce between pointers) and cmp | |||
12979 | // instructions (grouping them by the predicate). | |||
12980 | MapVector<size_t, MapVector<size_t, MapVector<Value *, unsigned>>> | |||
12981 | PossibleReducedVals; | |||
12982 | initReductionOps(Root); | |||
12983 | DenseMap<Value *, SmallVector<LoadInst *>> LoadsMap; | |||
12984 | SmallSet<size_t, 2> LoadKeyUsed; | |||
12985 | SmallPtrSet<Value *, 4> DoNotReverseVals; | |||
12986 | ||||
12987 | auto GenerateLoadsSubkey = [&](size_t Key, LoadInst *LI) { | |||
12988 | Value *Ptr = getUnderlyingObject(LI->getPointerOperand()); | |||
12989 | if (LoadKeyUsed.contains(Key)) { | |||
12990 | auto LIt = LoadsMap.find(Ptr); | |||
12991 | if (LIt != LoadsMap.end()) { | |||
12992 | for (LoadInst *RLI : LIt->second) { | |||
12993 | if (getPointersDiff(RLI->getType(), RLI->getPointerOperand(), | |||
12994 | LI->getType(), LI->getPointerOperand(), DL, SE, | |||
12995 | /*StrictCheck=*/true)) | |||
12996 | return hash_value(RLI->getPointerOperand()); | |||
12997 | } | |||
12998 | for (LoadInst *RLI : LIt->second) { | |||
12999 | if (arePointersCompatible(RLI->getPointerOperand(), | |||
13000 | LI->getPointerOperand(), TLI)) { | |||
13001 | hash_code SubKey = hash_value(RLI->getPointerOperand()); | |||
13002 | DoNotReverseVals.insert(RLI); | |||
13003 | return SubKey; | |||
13004 | } | |||
13005 | } | |||
13006 | if (LIt->second.size() > 2) { | |||
13007 | hash_code SubKey = | |||
13008 | hash_value(LIt->second.back()->getPointerOperand()); | |||
13009 | DoNotReverseVals.insert(LIt->second.back()); | |||
13010 | return SubKey; | |||
13011 | } | |||
13012 | } | |||
13013 | } | |||
13014 | LoadKeyUsed.insert(Key); | |||
13015 | LoadsMap.try_emplace(Ptr).first->second.push_back(LI); | |||
13016 | return hash_value(LI->getPointerOperand()); | |||
13017 | }; | |||
13018 | ||||
13019 | while (!Worklist.empty()) { | |||
13020 | Instruction *TreeN = Worklist.pop_back_val(); | |||
13021 | SmallVector<Value *> Args; | |||
13022 | SmallVector<Value *> PossibleRedVals; | |||
13023 | SmallVector<Instruction *> PossibleReductionOps; | |||
13024 | CheckOperands(TreeN, Args, PossibleRedVals, PossibleReductionOps); | |||
13025 | // If too many extra args - mark the instruction itself as a reduction | |||
13026 | // value, not a reduction operation. | |||
13027 | if (Args.size() < 2) { | |||
13028 | addReductionOps(TreeN); | |||
13029 | // Add extra args. | |||
13030 | if (!Args.empty()) { | |||
13031 | assert(Args.size() == 1 && "Expected only single argument.")(static_cast <bool> (Args.size() == 1 && "Expected only single argument." ) ? void (0) : __assert_fail ("Args.size() == 1 && \"Expected only single argument.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 13031, __extension__ __PRETTY_FUNCTION__)); | |||
13032 | ExtraArgs[TreeN] = Args.front(); | |||
13033 | } | |||
13034 | // Add reduction values. The values are sorted for better vectorization | |||
13035 | // results. | |||
13036 | for (Value *V : PossibleRedVals) { | |||
13037 | size_t Key, Idx; | |||
13038 | std::tie(Key, Idx) = generateKeySubkey(V, &TLI, GenerateLoadsSubkey, | |||
13039 | /*AllowAlternate=*/false); | |||
13040 | ++PossibleReducedVals[Key][Idx] | |||
13041 | .insert(std::make_pair(V, 0)) | |||
13042 | .first->second; | |||
13043 | } | |||
13044 | Worklist.append(PossibleReductionOps.rbegin(), | |||
13045 | PossibleReductionOps.rend()); | |||
13046 | } else { | |||
13047 | size_t Key, Idx; | |||
13048 | std::tie(Key, Idx) = generateKeySubkey(TreeN, &TLI, GenerateLoadsSubkey, | |||
13049 | /*AllowAlternate=*/false); | |||
13050 | ++PossibleReducedVals[Key][Idx] | |||
13051 | .insert(std::make_pair(TreeN, 0)) | |||
13052 | .first->second; | |||
13053 | } | |||
13054 | } | |||
13055 | auto PossibleReducedValsVect = PossibleReducedVals.takeVector(); | |||
13056 | // Sort values by the total number of values kinds to start the reduction | |||
13057 | // from the longest possible reduced values sequences. | |||
13058 | for (auto &PossibleReducedVals : PossibleReducedValsVect) { | |||
13059 | auto PossibleRedVals = PossibleReducedVals.second.takeVector(); | |||
13060 | SmallVector<SmallVector<Value *>> PossibleRedValsVect; | |||
13061 | for (auto It = PossibleRedVals.begin(), E = PossibleRedVals.end(); | |||
13062 | It != E; ++It) { | |||
13063 | PossibleRedValsVect.emplace_back(); | |||
13064 | auto RedValsVect = It->second.takeVector(); | |||
13065 | stable_sort(RedValsVect, llvm::less_second()); | |||
13066 | for (const std::pair<Value *, unsigned> &Data : RedValsVect) | |||
13067 | PossibleRedValsVect.back().append(Data.second, Data.first); | |||
13068 | } | |||
13069 | stable_sort(PossibleRedValsVect, [](const auto &P1, const auto &P2) { | |||
13070 | return P1.size() > P2.size(); | |||
13071 | }); | |||
13072 | int NewIdx = -1; | |||
13073 | for (ArrayRef<Value *> Data : PossibleRedValsVect) { | |||
13074 | if (isGoodForReduction(Data) || | |||
13075 | (isa<LoadInst>(Data.front()) && NewIdx >= 0 && | |||
13076 | isa<LoadInst>(ReducedVals[NewIdx].front()) && | |||
13077 | getUnderlyingObject( | |||
13078 | cast<LoadInst>(Data.front())->getPointerOperand()) == | |||
13079 | getUnderlyingObject(cast<LoadInst>(ReducedVals[NewIdx].front()) | |||
13080 | ->getPointerOperand()))) { | |||
13081 | if (NewIdx < 0) { | |||
13082 | NewIdx = ReducedVals.size(); | |||
13083 | ReducedVals.emplace_back(); | |||
13084 | } | |||
13085 | if (DoNotReverseVals.contains(Data.front())) | |||
13086 | ReducedVals[NewIdx].append(Data.begin(), Data.end()); | |||
13087 | else | |||
13088 | ReducedVals[NewIdx].append(Data.rbegin(), Data.rend()); | |||
13089 | } else { | |||
13090 | ReducedVals.emplace_back().append(Data.rbegin(), Data.rend()); | |||
13091 | } | |||
13092 | } | |||
13093 | } | |||
13094 | // Sort the reduced values by number of same/alternate opcode and/or pointer | |||
13095 | // operand. | |||
13096 | stable_sort(ReducedVals, [](ArrayRef<Value *> P1, ArrayRef<Value *> P2) { | |||
13097 | return P1.size() > P2.size(); | |||
13098 | }); | |||
13099 | return true; | |||
13100 | } | |||
13101 | ||||
13102 | /// Attempt to vectorize the tree found by matchAssociativeReduction. | |||
13103 | Value *tryToReduce(BoUpSLP &V, TargetTransformInfo *TTI, | |||
13104 | const TargetLibraryInfo &TLI) { | |||
13105 | constexpr int ReductionLimit = 4; | |||
13106 | constexpr unsigned RegMaxNumber = 4; | |||
13107 | constexpr unsigned RedValsMaxNumber = 128; | |||
13108 | // If there are a sufficient number of reduction values, reduce | |||
13109 | // to a nearby power-of-2. We can safely generate oversized | |||
13110 | // vectors and rely on the backend to split them to legal sizes. | |||
13111 | unsigned NumReducedVals = | |||
13112 | std::accumulate(ReducedVals.begin(), ReducedVals.end(), 0, | |||
13113 | [](unsigned Num, ArrayRef<Value *> Vals) -> unsigned { | |||
13114 | if (!isGoodForReduction(Vals)) | |||
13115 | return Num; | |||
13116 | return Num + Vals.size(); | |||
13117 | }); | |||
13118 | if (NumReducedVals < ReductionLimit && | |||
13119 | (!AllowHorRdxIdenityOptimization || | |||
13120 | all_of(ReducedVals, [](ArrayRef<Value *> RedV) { | |||
13121 | return RedV.size() < 2 || !allConstant(RedV) || !isSplat(RedV); | |||
13122 | }))) { | |||
13123 | for (ReductionOpsType &RdxOps : ReductionOps) | |||
13124 | for (Value *RdxOp : RdxOps) | |||
13125 | V.analyzedReductionRoot(cast<Instruction>(RdxOp)); | |||
13126 | return nullptr; | |||
13127 | } | |||
13128 | ||||
13129 | IRBuilder<> Builder(cast<Instruction>(ReductionRoot)); | |||
13130 | ||||
13131 | // Track the reduced values in case if they are replaced by extractelement | |||
13132 | // because of the vectorization. | |||
13133 | DenseMap<Value *, WeakTrackingVH> TrackedVals( | |||
13134 | ReducedVals.size() * ReducedVals.front().size() + ExtraArgs.size()); | |||
13135 | BoUpSLP::ExtraValueToDebugLocsMap ExternallyUsedValues; | |||
13136 | SmallVector<std::pair<Value *, Value *>> ReplacedExternals; | |||
13137 | ExternallyUsedValues.reserve(ExtraArgs.size() + 1); | |||
13138 | // The same extra argument may be used several times, so log each attempt | |||
13139 | // to use it. | |||
13140 | for (const std::pair<Instruction *, Value *> &Pair : ExtraArgs) { | |||
13141 | 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", 13141, __extension__ __PRETTY_FUNCTION__)); | |||
13142 | ExternallyUsedValues[Pair.second].push_back(Pair.first); | |||
13143 | TrackedVals.try_emplace(Pair.second, Pair.second); | |||
13144 | } | |||
13145 | ||||
13146 | // The compare instruction of a min/max is the insertion point for new | |||
13147 | // instructions and may be replaced with a new compare instruction. | |||
13148 | auto &&GetCmpForMinMaxReduction = [](Instruction *RdxRootInst) { | |||
13149 | 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", 13150, __extension__ __PRETTY_FUNCTION__)) | |||
13150 | "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", 13150, __extension__ __PRETTY_FUNCTION__)); | |||
13151 | Value *ScalarCond = cast<SelectInst>(RdxRootInst)->getCondition(); | |||
13152 | 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", 13153, __extension__ __PRETTY_FUNCTION__)) | |||
13153 | "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", 13153, __extension__ __PRETTY_FUNCTION__)); | |||
13154 | return cast<Instruction>(ScalarCond); | |||
13155 | }; | |||
13156 | ||||
13157 | // Return new VectorizedTree, based on previous value. | |||
13158 | auto GetNewVectorizedTree = [&](Value *VectorizedTree, Value *Res) { | |||
13159 | if (VectorizedTree) { | |||
13160 | // Update the final value in the reduction. | |||
13161 | Builder.SetCurrentDebugLocation( | |||
13162 | cast<Instruction>(ReductionOps.front().front())->getDebugLoc()); | |||
13163 | return createOp(Builder, RdxKind, VectorizedTree, Res, "op.rdx", | |||
13164 | ReductionOps); | |||
13165 | } | |||
13166 | // Initialize the final value in the reduction. | |||
13167 | return Res; | |||
13168 | }; | |||
13169 | // The reduction root is used as the insertion point for new instructions, | |||
13170 | // so set it as externally used to prevent it from being deleted. | |||
13171 | ExternallyUsedValues[ReductionRoot]; | |||
13172 | SmallDenseSet<Value *> IgnoreList(ReductionOps.size() * | |||
13173 | ReductionOps.front().size()); | |||
13174 | for (ReductionOpsType &RdxOps : ReductionOps) | |||
13175 | for (Value *RdxOp : RdxOps) { | |||
13176 | if (!RdxOp) | |||
13177 | continue; | |||
13178 | IgnoreList.insert(RdxOp); | |||
13179 | } | |||
13180 | // Intersect the fast-math-flags from all reduction operations. | |||
13181 | FastMathFlags RdxFMF; | |||
13182 | RdxFMF.set(); | |||
13183 | for (Value *U : IgnoreList) | |||
13184 | if (auto *FPMO = dyn_cast<FPMathOperator>(U)) | |||
13185 | RdxFMF &= FPMO->getFastMathFlags(); | |||
13186 | bool IsCmpSelMinMax = isCmpSelMinMax(cast<Instruction>(ReductionRoot)); | |||
13187 | ||||
13188 | // Need to track reduced vals, they may be changed during vectorization of | |||
13189 | // subvectors. | |||
13190 | for (ArrayRef<Value *> Candidates : ReducedVals) | |||
13191 | for (Value *V : Candidates) | |||
13192 | TrackedVals.try_emplace(V, V); | |||
13193 | ||||
13194 | DenseMap<Value *, unsigned> VectorizedVals(ReducedVals.size()); | |||
13195 | // List of the values that were reduced in other trees as part of gather | |||
13196 | // nodes and thus requiring extract if fully vectorized in other trees. | |||
13197 | SmallPtrSet<Value *, 4> RequiredExtract; | |||
13198 | Value *VectorizedTree = nullptr; | |||
13199 | bool CheckForReusedReductionOps = false; | |||
13200 | // Try to vectorize elements based on their type. | |||
13201 | for (unsigned I = 0, E = ReducedVals.size(); I < E; ++I) { | |||
13202 | ArrayRef<Value *> OrigReducedVals = ReducedVals[I]; | |||
13203 | InstructionsState S = getSameOpcode(OrigReducedVals, TLI); | |||
13204 | SmallVector<Value *> Candidates; | |||
13205 | Candidates.reserve(2 * OrigReducedVals.size()); | |||
13206 | DenseMap<Value *, Value *> TrackedToOrig(2 * OrigReducedVals.size()); | |||
13207 | for (unsigned Cnt = 0, Sz = OrigReducedVals.size(); Cnt < Sz; ++Cnt) { | |||
13208 | Value *RdxVal = TrackedVals.find(OrigReducedVals[Cnt])->second; | |||
13209 | // Check if the reduction value was not overriden by the extractelement | |||
13210 | // instruction because of the vectorization and exclude it, if it is not | |||
13211 | // compatible with other values. | |||
13212 | if (auto *Inst = dyn_cast<Instruction>(RdxVal)) | |||
13213 | if (isVectorLikeInstWithConstOps(Inst) && | |||
13214 | (!S.getOpcode() || !S.isOpcodeOrAlt(Inst))) | |||
13215 | continue; | |||
13216 | Candidates.push_back(RdxVal); | |||
13217 | TrackedToOrig.try_emplace(RdxVal, OrigReducedVals[Cnt]); | |||
13218 | } | |||
13219 | bool ShuffledExtracts = false; | |||
13220 | // Try to handle shuffled extractelements. | |||
13221 | if (S.getOpcode() == Instruction::ExtractElement && !S.isAltShuffle() && | |||
13222 | I + 1 < E) { | |||
13223 | InstructionsState NextS = getSameOpcode(ReducedVals[I + 1], TLI); | |||
13224 | if (NextS.getOpcode() == Instruction::ExtractElement && | |||
13225 | !NextS.isAltShuffle()) { | |||
13226 | SmallVector<Value *> CommonCandidates(Candidates); | |||
13227 | for (Value *RV : ReducedVals[I + 1]) { | |||
13228 | Value *RdxVal = TrackedVals.find(RV)->second; | |||
13229 | // Check if the reduction value was not overriden by the | |||
13230 | // extractelement instruction because of the vectorization and | |||
13231 | // exclude it, if it is not compatible with other values. | |||
13232 | if (auto *Inst = dyn_cast<Instruction>(RdxVal)) | |||
13233 | if (!NextS.getOpcode() || !NextS.isOpcodeOrAlt(Inst)) | |||
13234 | continue; | |||
13235 | CommonCandidates.push_back(RdxVal); | |||
13236 | TrackedToOrig.try_emplace(RdxVal, RV); | |||
13237 | } | |||
13238 | SmallVector<int> Mask; | |||
13239 | if (isFixedVectorShuffle(CommonCandidates, Mask)) { | |||
13240 | ++I; | |||
13241 | Candidates.swap(CommonCandidates); | |||
13242 | ShuffledExtracts = true; | |||
13243 | } | |||
13244 | } | |||
13245 | } | |||
13246 | ||||
13247 | // Emit code for constant values. | |||
13248 | if (AllowHorRdxIdenityOptimization && Candidates.size() > 1 && | |||
13249 | allConstant(Candidates)) { | |||
13250 | Value *Res = Candidates.front(); | |||
13251 | ++VectorizedVals.try_emplace(Candidates.front(), 0).first->getSecond(); | |||
13252 | for (Value *VC : ArrayRef(Candidates).drop_front()) { | |||
13253 | Res = createOp(Builder, RdxKind, Res, VC, "const.rdx", ReductionOps); | |||
13254 | ++VectorizedVals.try_emplace(VC, 0).first->getSecond(); | |||
13255 | if (auto *ResI = dyn_cast<Instruction>(Res)) | |||
13256 | V.analyzedReductionRoot(ResI); | |||
13257 | } | |||
13258 | VectorizedTree = GetNewVectorizedTree(VectorizedTree, Res); | |||
13259 | continue; | |||
13260 | } | |||
13261 | ||||
13262 | unsigned NumReducedVals = Candidates.size(); | |||
13263 | if (NumReducedVals < ReductionLimit && | |||
13264 | (NumReducedVals < 2 || !AllowHorRdxIdenityOptimization || | |||
13265 | !isSplat(Candidates))) | |||
13266 | continue; | |||
13267 | ||||
13268 | // Check if we support repeated scalar values processing (optimization of | |||
13269 | // original scalar identity operations on matched horizontal reductions). | |||
13270 | IsSupportedHorRdxIdentityOp = | |||
13271 | AllowHorRdxIdenityOptimization && RdxKind != RecurKind::Mul && | |||
13272 | RdxKind != RecurKind::FMul && RdxKind != RecurKind::FMulAdd; | |||
13273 | // Gather same values. | |||
13274 | MapVector<Value *, unsigned> SameValuesCounter; | |||
13275 | if (IsSupportedHorRdxIdentityOp) | |||
13276 | for (Value *V : Candidates) | |||
13277 | ++SameValuesCounter.insert(std::make_pair(V, 0)).first->second; | |||
13278 | // Used to check if the reduced values used same number of times. In this | |||
13279 | // case the compiler may produce better code. E.g. if reduced values are | |||
13280 | // aabbccdd (8 x values), then the first node of the tree will have a node | |||
13281 | // for 4 x abcd + shuffle <4 x abcd>, <0, 0, 1, 1, 2, 2, 3, 3>. | |||
13282 | // Plus, the final reduction will be performed on <8 x aabbccdd>. | |||
13283 | // Instead compiler may build <4 x abcd> tree immediately, + reduction (4 | |||
13284 | // x abcd) * 2. | |||
13285 | // Currently it only handles add/fadd/xor. and/or/min/max do not require | |||
13286 | // this analysis, other operations may require an extra estimation of | |||
13287 | // the profitability. | |||
13288 | bool SameScaleFactor = false; | |||
13289 | bool OptReusedScalars = IsSupportedHorRdxIdentityOp && | |||
13290 | SameValuesCounter.size() != Candidates.size(); | |||
13291 | if (OptReusedScalars) { | |||
13292 | SameScaleFactor = | |||
13293 | (RdxKind == RecurKind::Add || RdxKind == RecurKind::FAdd || | |||
13294 | RdxKind == RecurKind::Xor) && | |||
13295 | all_of(drop_begin(SameValuesCounter), | |||
13296 | [&SameValuesCounter](const std::pair<Value *, unsigned> &P) { | |||
13297 | return P.second == SameValuesCounter.front().second; | |||
13298 | }); | |||
13299 | Candidates.resize(SameValuesCounter.size()); | |||
13300 | transform(SameValuesCounter, Candidates.begin(), | |||
13301 | [](const auto &P) { return P.first; }); | |||
13302 | NumReducedVals = Candidates.size(); | |||
13303 | // Have a reduction of the same element. | |||
13304 | if (NumReducedVals == 1) { | |||
13305 | Value *OrigV = TrackedToOrig.find(Candidates.front())->second; | |||
13306 | unsigned Cnt = SameValuesCounter.lookup(OrigV); | |||
13307 | Value *RedVal = | |||
13308 | emitScaleForReusedOps(Candidates.front(), Builder, Cnt); | |||
13309 | VectorizedTree = GetNewVectorizedTree(VectorizedTree, RedVal); | |||
13310 | VectorizedVals.try_emplace(OrigV, Cnt); | |||
13311 | continue; | |||
13312 | } | |||
13313 | } | |||
13314 | ||||
13315 | unsigned MaxVecRegSize = V.getMaxVecRegSize(); | |||
13316 | unsigned EltSize = V.getVectorElementSize(Candidates[0]); | |||
13317 | unsigned MaxElts = | |||
13318 | RegMaxNumber * llvm::bit_floor(MaxVecRegSize / EltSize); | |||
13319 | ||||
13320 | unsigned ReduxWidth = std::min<unsigned>( | |||
13321 | llvm::bit_floor(NumReducedVals), std::max(RedValsMaxNumber, MaxElts)); | |||
13322 | unsigned Start = 0; | |||
13323 | unsigned Pos = Start; | |||
13324 | // Restarts vectorization attempt with lower vector factor. | |||
13325 | unsigned PrevReduxWidth = ReduxWidth; | |||
13326 | bool CheckForReusedReductionOpsLocal = false; | |||
13327 | auto &&AdjustReducedVals = [&Pos, &Start, &ReduxWidth, NumReducedVals, | |||
13328 | &CheckForReusedReductionOpsLocal, | |||
13329 | &PrevReduxWidth, &V, | |||
13330 | &IgnoreList](bool IgnoreVL = false) { | |||
13331 | bool IsAnyRedOpGathered = !IgnoreVL && V.isAnyGathered(IgnoreList); | |||
13332 | if (!CheckForReusedReductionOpsLocal && PrevReduxWidth == ReduxWidth) { | |||
13333 | // Check if any of the reduction ops are gathered. If so, worth | |||
13334 | // trying again with less number of reduction ops. | |||
13335 | CheckForReusedReductionOpsLocal |= IsAnyRedOpGathered; | |||
13336 | } | |||
13337 | ++Pos; | |||
13338 | if (Pos < NumReducedVals - ReduxWidth + 1) | |||
13339 | return IsAnyRedOpGathered; | |||
13340 | Pos = Start; | |||
13341 | ReduxWidth /= 2; | |||
13342 | return IsAnyRedOpGathered; | |||
13343 | }; | |||
13344 | bool AnyVectorized = false; | |||
13345 | while (Pos < NumReducedVals - ReduxWidth + 1 && | |||
13346 | ReduxWidth >= ReductionLimit) { | |||
13347 | // Dependency in tree of the reduction ops - drop this attempt, try | |||
13348 | // later. | |||
13349 | if (CheckForReusedReductionOpsLocal && PrevReduxWidth != ReduxWidth && | |||
13350 | Start == 0) { | |||
13351 | CheckForReusedReductionOps = true; | |||
13352 | break; | |||
13353 | } | |||
13354 | PrevReduxWidth = ReduxWidth; | |||
13355 | ArrayRef<Value *> VL(std::next(Candidates.begin(), Pos), ReduxWidth); | |||
13356 | // Beeing analyzed already - skip. | |||
13357 | if (V.areAnalyzedReductionVals(VL)) { | |||
13358 | (void)AdjustReducedVals(/*IgnoreVL=*/true); | |||
13359 | continue; | |||
13360 | } | |||
13361 | // Early exit if any of the reduction values were deleted during | |||
13362 | // previous vectorization attempts. | |||
13363 | if (any_of(VL, [&V](Value *RedVal) { | |||
13364 | auto *RedValI = dyn_cast<Instruction>(RedVal); | |||
13365 | if (!RedValI) | |||
13366 | return false; | |||
13367 | return V.isDeleted(RedValI); | |||
13368 | })) | |||
13369 | break; | |||
13370 | V.buildTree(VL, IgnoreList); | |||
13371 | if (V.isTreeTinyAndNotFullyVectorizable(/*ForReduction=*/true)) { | |||
13372 | if (!AdjustReducedVals()) | |||
13373 | V.analyzedReductionVals(VL); | |||
13374 | continue; | |||
13375 | } | |||
13376 | if (V.isLoadCombineReductionCandidate(RdxKind)) { | |||
13377 | if (!AdjustReducedVals()) | |||
13378 | V.analyzedReductionVals(VL); | |||
13379 | continue; | |||
13380 | } | |||
13381 | V.reorderTopToBottom(); | |||
13382 | // No need to reorder the root node at all. | |||
13383 | V.reorderBottomToTop(/*IgnoreReorder=*/true); | |||
13384 | // Keep extracted other reduction values, if they are used in the | |||
13385 | // vectorization trees. | |||
13386 | BoUpSLP::ExtraValueToDebugLocsMap LocalExternallyUsedValues( | |||
13387 | ExternallyUsedValues); | |||
13388 | for (unsigned Cnt = 0, Sz = ReducedVals.size(); Cnt < Sz; ++Cnt) { | |||
13389 | if (Cnt == I || (ShuffledExtracts && Cnt == I - 1)) | |||
13390 | continue; | |||
13391 | for_each(ReducedVals[Cnt], | |||
13392 | [&LocalExternallyUsedValues, &TrackedVals](Value *V) { | |||
13393 | if (isa<Instruction>(V)) | |||
13394 | LocalExternallyUsedValues[TrackedVals[V]]; | |||
13395 | }); | |||
13396 | } | |||
13397 | if (!IsSupportedHorRdxIdentityOp) { | |||
13398 | // Number of uses of the candidates in the vector of values. | |||
13399 | assert(SameValuesCounter.empty() &&(static_cast <bool> (SameValuesCounter.empty() && "Reused values counter map is not empty") ? void (0) : __assert_fail ("SameValuesCounter.empty() && \"Reused values counter map is not empty\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 13400, __extension__ __PRETTY_FUNCTION__)) | |||
13400 | "Reused values counter map is not empty")(static_cast <bool> (SameValuesCounter.empty() && "Reused values counter map is not empty") ? void (0) : __assert_fail ("SameValuesCounter.empty() && \"Reused values counter map is not empty\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 13400, __extension__ __PRETTY_FUNCTION__)); | |||
13401 | for (unsigned Cnt = 0; Cnt < NumReducedVals; ++Cnt) { | |||
13402 | if (Cnt >= Pos && Cnt < Pos + ReduxWidth) | |||
13403 | continue; | |||
13404 | Value *V = Candidates[Cnt]; | |||
13405 | Value *OrigV = TrackedToOrig.find(V)->second; | |||
13406 | ++SameValuesCounter[OrigV]; | |||
13407 | } | |||
13408 | } | |||
13409 | SmallPtrSet<Value *, 4> VLScalars(VL.begin(), VL.end()); | |||
13410 | // Gather externally used values. | |||
13411 | SmallPtrSet<Value *, 4> Visited; | |||
13412 | for (unsigned Cnt = 0; Cnt < NumReducedVals; ++Cnt) { | |||
13413 | if (Cnt >= Pos && Cnt < Pos + ReduxWidth) | |||
13414 | continue; | |||
13415 | Value *RdxVal = Candidates[Cnt]; | |||
13416 | if (!Visited.insert(RdxVal).second) | |||
13417 | continue; | |||
13418 | // Check if the scalar was vectorized as part of the vectorization | |||
13419 | // tree but not the top node. | |||
13420 | if (!VLScalars.contains(RdxVal) && V.isVectorized(RdxVal)) { | |||
13421 | LocalExternallyUsedValues[RdxVal]; | |||
13422 | continue; | |||
13423 | } | |||
13424 | Value *OrigV = TrackedToOrig.find(RdxVal)->second; | |||
13425 | unsigned NumOps = | |||
13426 | VectorizedVals.lookup(RdxVal) + SameValuesCounter[OrigV]; | |||
13427 | if (NumOps != ReducedValsToOps.find(OrigV)->second.size()) | |||
13428 | LocalExternallyUsedValues[RdxVal]; | |||
13429 | } | |||
13430 | // Do not need the list of reused scalars in regular mode anymore. | |||
13431 | if (!IsSupportedHorRdxIdentityOp) | |||
13432 | SameValuesCounter.clear(); | |||
13433 | for (Value *RdxVal : VL) | |||
13434 | if (RequiredExtract.contains(RdxVal)) | |||
13435 | LocalExternallyUsedValues[RdxVal]; | |||
13436 | // Update LocalExternallyUsedValues for the scalar, replaced by | |||
13437 | // extractelement instructions. | |||
13438 | for (const std::pair<Value *, Value *> &Pair : ReplacedExternals) { | |||
13439 | auto It = ExternallyUsedValues.find(Pair.first); | |||
13440 | if (It == ExternallyUsedValues.end()) | |||
13441 | continue; | |||
13442 | LocalExternallyUsedValues[Pair.second].append(It->second); | |||
13443 | } | |||
13444 | V.buildExternalUses(LocalExternallyUsedValues); | |||
13445 | ||||
13446 | V.computeMinimumValueSizes(); | |||
13447 | ||||
13448 | // Estimate cost. | |||
13449 | InstructionCost TreeCost = V.getTreeCost(VL); | |||
13450 | InstructionCost ReductionCost = | |||
13451 | getReductionCost(TTI, VL, IsCmpSelMinMax, ReduxWidth, RdxFMF); | |||
13452 | InstructionCost Cost = TreeCost + ReductionCost; | |||
13453 | LLVM_DEBUG(dbgs() << "SLP: Found cost = " << Cost << " for reduction\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Found cost = " << Cost << " for reduction\n"; } } while (false); | |||
13454 | if (!Cost.isValid()) | |||
13455 | return nullptr; | |||
13456 | if (Cost >= -SLPCostThreshold) { | |||
13457 | V.getORE()->emit([&]() { | |||
13458 | return OptimizationRemarkMissed( | |||
13459 | SV_NAME"slp-vectorizer", "HorSLPNotBeneficial", | |||
13460 | ReducedValsToOps.find(VL[0])->second.front()) | |||
13461 | << "Vectorizing horizontal reduction is possible " | |||
13462 | << "but not beneficial with cost " << ore::NV("Cost", Cost) | |||
13463 | << " and threshold " | |||
13464 | << ore::NV("Threshold", -SLPCostThreshold); | |||
13465 | }); | |||
13466 | if (!AdjustReducedVals()) | |||
13467 | V.analyzedReductionVals(VL); | |||
13468 | continue; | |||
13469 | } | |||
13470 | ||||
13471 | 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) | |||
13472 | << Cost << ". (HorRdx)\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Vectorizing horizontal reduction at cost:" << Cost << ". (HorRdx)\n"; } } while (false); | |||
13473 | V.getORE()->emit([&]() { | |||
13474 | return OptimizationRemark( | |||
13475 | SV_NAME"slp-vectorizer", "VectorizedHorizontalReduction", | |||
13476 | ReducedValsToOps.find(VL[0])->second.front()) | |||
13477 | << "Vectorized horizontal reduction with cost " | |||
13478 | << ore::NV("Cost", Cost) << " and with tree size " | |||
13479 | << ore::NV("TreeSize", V.getTreeSize()); | |||
13480 | }); | |||
13481 | ||||
13482 | Builder.setFastMathFlags(RdxFMF); | |||
13483 | ||||
13484 | // Emit a reduction. If the root is a select (min/max idiom), the insert | |||
13485 | // point is the compare condition of that select. | |||
13486 | Instruction *RdxRootInst = cast<Instruction>(ReductionRoot); | |||
13487 | Instruction *InsertPt = RdxRootInst; | |||
13488 | if (IsCmpSelMinMax) | |||
13489 | InsertPt = GetCmpForMinMaxReduction(RdxRootInst); | |||
13490 | ||||
13491 | // Vectorize a tree. | |||
13492 | Value *VectorizedRoot = V.vectorizeTree(LocalExternallyUsedValues, | |||
13493 | ReplacedExternals, InsertPt); | |||
13494 | ||||
13495 | Builder.SetInsertPoint(InsertPt); | |||
13496 | ||||
13497 | // To prevent poison from leaking across what used to be sequential, | |||
13498 | // safe, scalar boolean logic operations, the reduction operand must be | |||
13499 | // frozen. | |||
13500 | if (isBoolLogicOp(RdxRootInst)) | |||
13501 | VectorizedRoot = Builder.CreateFreeze(VectorizedRoot); | |||
13502 | ||||
13503 | // Emit code to correctly handle reused reduced values, if required. | |||
13504 | if (OptReusedScalars && !SameScaleFactor) { | |||
13505 | VectorizedRoot = | |||
13506 | emitReusedOps(VectorizedRoot, Builder, V.getRootNodeScalars(), | |||
13507 | SameValuesCounter, TrackedToOrig); | |||
13508 | } | |||
13509 | ||||
13510 | Value *ReducedSubTree = | |||
13511 | emitReduction(VectorizedRoot, Builder, ReduxWidth, TTI); | |||
13512 | ||||
13513 | // Improved analysis for add/fadd/xor reductions with same scale factor | |||
13514 | // for all operands of reductions. We can emit scalar ops for them | |||
13515 | // instead. | |||
13516 | if (OptReusedScalars && SameScaleFactor) | |||
13517 | ReducedSubTree = emitScaleForReusedOps( | |||
13518 | ReducedSubTree, Builder, SameValuesCounter.front().second); | |||
13519 | ||||
13520 | VectorizedTree = GetNewVectorizedTree(VectorizedTree, ReducedSubTree); | |||
13521 | // Count vectorized reduced values to exclude them from final reduction. | |||
13522 | for (Value *RdxVal : VL) { | |||
13523 | Value *OrigV = TrackedToOrig.find(RdxVal)->second; | |||
13524 | if (IsSupportedHorRdxIdentityOp) { | |||
13525 | VectorizedVals.try_emplace(OrigV, SameValuesCounter[RdxVal]); | |||
13526 | continue; | |||
13527 | } | |||
13528 | ++VectorizedVals.try_emplace(OrigV, 0).first->getSecond(); | |||
13529 | if (!V.isVectorized(RdxVal)) | |||
13530 | RequiredExtract.insert(RdxVal); | |||
13531 | } | |||
13532 | Pos += ReduxWidth; | |||
13533 | Start = Pos; | |||
13534 | ReduxWidth = llvm::bit_floor(NumReducedVals - Pos); | |||
13535 | AnyVectorized = true; | |||
13536 | } | |||
13537 | if (OptReusedScalars && !AnyVectorized) { | |||
13538 | for (const std::pair<Value *, unsigned> &P : SameValuesCounter) { | |||
13539 | Value *RedVal = emitScaleForReusedOps(P.first, Builder, P.second); | |||
13540 | VectorizedTree = GetNewVectorizedTree(VectorizedTree, RedVal); | |||
13541 | Value *OrigV = TrackedToOrig.find(P.first)->second; | |||
13542 | VectorizedVals.try_emplace(OrigV, P.second); | |||
13543 | } | |||
13544 | continue; | |||
13545 | } | |||
13546 | } | |||
13547 | if (VectorizedTree) { | |||
13548 | // Reorder operands of bool logical op in the natural order to avoid | |||
13549 | // possible problem with poison propagation. If not possible to reorder | |||
13550 | // (both operands are originally RHS), emit an extra freeze instruction | |||
13551 | // for the LHS operand. | |||
13552 | //I.e., if we have original code like this: | |||
13553 | // RedOp1 = select i1 ?, i1 LHS, i1 false | |||
13554 | // RedOp2 = select i1 RHS, i1 ?, i1 false | |||
13555 | ||||
13556 | // Then, we swap LHS/RHS to create a new op that matches the poison | |||
13557 | // semantics of the original code. | |||
13558 | ||||
13559 | // If we have original code like this and both values could be poison: | |||
13560 | // RedOp1 = select i1 ?, i1 LHS, i1 false | |||
13561 | // RedOp2 = select i1 ?, i1 RHS, i1 false | |||
13562 | ||||
13563 | // Then, we must freeze LHS in the new op. | |||
13564 | auto &&FixBoolLogicalOps = | |||
13565 | [&Builder, VectorizedTree](Value *&LHS, Value *&RHS, | |||
13566 | Instruction *RedOp1, Instruction *RedOp2) { | |||
13567 | if (!isBoolLogicOp(RedOp1)) | |||
13568 | return; | |||
13569 | if (LHS == VectorizedTree || getRdxOperand(RedOp1, 0) == LHS || | |||
13570 | isGuaranteedNotToBePoison(LHS)) | |||
13571 | return; | |||
13572 | if (!isBoolLogicOp(RedOp2)) | |||
13573 | return; | |||
13574 | if (RHS == VectorizedTree || getRdxOperand(RedOp2, 0) == RHS || | |||
13575 | isGuaranteedNotToBePoison(RHS)) { | |||
13576 | std::swap(LHS, RHS); | |||
13577 | return; | |||
13578 | } | |||
13579 | LHS = Builder.CreateFreeze(LHS); | |||
13580 | }; | |||
13581 | // Finish the reduction. | |||
13582 | // Need to add extra arguments and not vectorized possible reduction | |||
13583 | // values. | |||
13584 | // Try to avoid dependencies between the scalar remainders after | |||
13585 | // reductions. | |||
13586 | auto &&FinalGen = | |||
13587 | [this, &Builder, &TrackedVals, &FixBoolLogicalOps]( | |||
13588 | ArrayRef<std::pair<Instruction *, Value *>> InstVals) { | |||
13589 | unsigned Sz = InstVals.size(); | |||
13590 | SmallVector<std::pair<Instruction *, Value *>> ExtraReds(Sz / 2 + | |||
13591 | Sz % 2); | |||
13592 | for (unsigned I = 0, E = (Sz / 2) * 2; I < E; I += 2) { | |||
13593 | Instruction *RedOp = InstVals[I + 1].first; | |||
13594 | Builder.SetCurrentDebugLocation(RedOp->getDebugLoc()); | |||
13595 | Value *RdxVal1 = InstVals[I].second; | |||
13596 | Value *StableRdxVal1 = RdxVal1; | |||
13597 | auto It1 = TrackedVals.find(RdxVal1); | |||
13598 | if (It1 != TrackedVals.end()) | |||
13599 | StableRdxVal1 = It1->second; | |||
13600 | Value *RdxVal2 = InstVals[I + 1].second; | |||
13601 | Value *StableRdxVal2 = RdxVal2; | |||
13602 | auto It2 = TrackedVals.find(RdxVal2); | |||
13603 | if (It2 != TrackedVals.end()) | |||
13604 | StableRdxVal2 = It2->second; | |||
13605 | // To prevent poison from leaking across what used to be | |||
13606 | // sequential, safe, scalar boolean logic operations, the | |||
13607 | // reduction operand must be frozen. | |||
13608 | FixBoolLogicalOps(StableRdxVal1, StableRdxVal2, InstVals[I].first, | |||
13609 | RedOp); | |||
13610 | Value *ExtraRed = createOp(Builder, RdxKind, StableRdxVal1, | |||
13611 | StableRdxVal2, "op.rdx", ReductionOps); | |||
13612 | ExtraReds[I / 2] = std::make_pair(InstVals[I].first, ExtraRed); | |||
13613 | } | |||
13614 | if (Sz % 2 == 1) | |||
13615 | ExtraReds[Sz / 2] = InstVals.back(); | |||
13616 | return ExtraReds; | |||
13617 | }; | |||
13618 | SmallVector<std::pair<Instruction *, Value *>> ExtraReductions; | |||
13619 | ExtraReductions.emplace_back(cast<Instruction>(ReductionRoot), | |||
13620 | VectorizedTree); | |||
13621 | SmallPtrSet<Value *, 8> Visited; | |||
13622 | for (ArrayRef<Value *> Candidates : ReducedVals) { | |||
13623 | for (Value *RdxVal : Candidates) { | |||
13624 | if (!Visited.insert(RdxVal).second) | |||
13625 | continue; | |||
13626 | unsigned NumOps = VectorizedVals.lookup(RdxVal); | |||
13627 | for (Instruction *RedOp : | |||
13628 | ArrayRef(ReducedValsToOps.find(RdxVal)->second) | |||
13629 | .drop_back(NumOps)) | |||
13630 | ExtraReductions.emplace_back(RedOp, RdxVal); | |||
13631 | } | |||
13632 | } | |||
13633 | for (auto &Pair : ExternallyUsedValues) { | |||
13634 | // Add each externally used value to the final reduction. | |||
13635 | for (auto *I : Pair.second) | |||
13636 | ExtraReductions.emplace_back(I, Pair.first); | |||
13637 | } | |||
13638 | // Iterate through all not-vectorized reduction values/extra arguments. | |||
13639 | while (ExtraReductions.size() > 1) { | |||
13640 | VectorizedTree = ExtraReductions.front().second; | |||
13641 | SmallVector<std::pair<Instruction *, Value *>> NewReds = | |||
13642 | FinalGen(ExtraReductions); | |||
13643 | ExtraReductions.swap(NewReds); | |||
13644 | } | |||
13645 | VectorizedTree = ExtraReductions.front().second; | |||
13646 | ||||
13647 | ReductionRoot->replaceAllUsesWith(VectorizedTree); | |||
13648 | ||||
13649 | // The original scalar reduction is expected to have no remaining | |||
13650 | // uses outside the reduction tree itself. Assert that we got this | |||
13651 | // correct, replace internal uses with undef, and mark for eventual | |||
13652 | // deletion. | |||
13653 | #ifndef NDEBUG | |||
13654 | SmallSet<Value *, 4> IgnoreSet; | |||
13655 | for (ArrayRef<Value *> RdxOps : ReductionOps) | |||
13656 | IgnoreSet.insert(RdxOps.begin(), RdxOps.end()); | |||
13657 | #endif | |||
13658 | for (ArrayRef<Value *> RdxOps : ReductionOps) { | |||
13659 | for (Value *Ignore : RdxOps) { | |||
13660 | if (!Ignore) | |||
13661 | continue; | |||
13662 | #ifndef NDEBUG | |||
13663 | for (auto *U : Ignore->users()) { | |||
13664 | assert(IgnoreSet.count(U) &&(static_cast <bool> (IgnoreSet.count(U) && "All users must be either in the reduction ops list." ) ? void (0) : __assert_fail ("IgnoreSet.count(U) && \"All users must be either in the reduction ops list.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 13665, __extension__ __PRETTY_FUNCTION__)) | |||
13665 | "All users must be either in the reduction ops list.")(static_cast <bool> (IgnoreSet.count(U) && "All users must be either in the reduction ops list." ) ? void (0) : __assert_fail ("IgnoreSet.count(U) && \"All users must be either in the reduction ops list.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 13665, __extension__ __PRETTY_FUNCTION__)); | |||
13666 | } | |||
13667 | #endif | |||
13668 | if (!Ignore->use_empty()) { | |||
13669 | Value *Undef = UndefValue::get(Ignore->getType()); | |||
13670 | Ignore->replaceAllUsesWith(Undef); | |||
13671 | } | |||
13672 | V.eraseInstruction(cast<Instruction>(Ignore)); | |||
13673 | } | |||
13674 | } | |||
13675 | } else if (!CheckForReusedReductionOps) { | |||
13676 | for (ReductionOpsType &RdxOps : ReductionOps) | |||
13677 | for (Value *RdxOp : RdxOps) | |||
13678 | V.analyzedReductionRoot(cast<Instruction>(RdxOp)); | |||
13679 | } | |||
13680 | return VectorizedTree; | |||
13681 | } | |||
13682 | ||||
13683 | private: | |||
13684 | /// Calculate the cost of a reduction. | |||
13685 | InstructionCost getReductionCost(TargetTransformInfo *TTI, | |||
13686 | ArrayRef<Value *> ReducedVals, | |||
13687 | bool IsCmpSelMinMax, unsigned ReduxWidth, | |||
13688 | FastMathFlags FMF) { | |||
13689 | TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput; | |||
13690 | Value *FirstReducedVal = ReducedVals.front(); | |||
13691 | Type *ScalarTy = FirstReducedVal->getType(); | |||
13692 | FixedVectorType *VectorTy = FixedVectorType::get(ScalarTy, ReduxWidth); | |||
13693 | InstructionCost VectorCost = 0, ScalarCost; | |||
13694 | // If all of the reduced values are constant, the vector cost is 0, since | |||
13695 | // the reduction value can be calculated at the compile time. | |||
13696 | bool AllConsts = allConstant(ReducedVals); | |||
13697 | auto EvaluateScalarCost = [&](function_ref<InstructionCost()> GenCostFn) { | |||
13698 | InstructionCost Cost = 0; | |||
13699 | // Scalar cost is repeated for N-1 elements. | |||
13700 | int Cnt = ReducedVals.size(); | |||
13701 | for (Value *RdxVal : ReducedVals) { | |||
13702 | if (Cnt == 1) | |||
13703 | break; | |||
13704 | --Cnt; | |||
13705 | if (RdxVal->hasNUsesOrMore(IsCmpSelMinMax ? 3 : 2)) { | |||
13706 | Cost += GenCostFn(); | |||
13707 | continue; | |||
13708 | } | |||
13709 | InstructionCost ScalarCost = 0; | |||
13710 | for (User *U : RdxVal->users()) { | |||
13711 | auto *RdxOp = cast<Instruction>(U); | |||
13712 | if (hasRequiredNumberOfUses(IsCmpSelMinMax, RdxOp)) { | |||
13713 | ScalarCost += TTI->getInstructionCost(RdxOp, CostKind); | |||
13714 | continue; | |||
13715 | } | |||
13716 | ScalarCost = InstructionCost::getInvalid(); | |||
13717 | break; | |||
13718 | } | |||
13719 | if (ScalarCost.isValid()) | |||
13720 | Cost += ScalarCost; | |||
13721 | else | |||
13722 | Cost += GenCostFn(); | |||
13723 | } | |||
13724 | return Cost; | |||
13725 | }; | |||
13726 | switch (RdxKind) { | |||
13727 | case RecurKind::Add: | |||
13728 | case RecurKind::Mul: | |||
13729 | case RecurKind::Or: | |||
13730 | case RecurKind::And: | |||
13731 | case RecurKind::Xor: | |||
13732 | case RecurKind::FAdd: | |||
13733 | case RecurKind::FMul: { | |||
13734 | unsigned RdxOpcode = RecurrenceDescriptor::getOpcode(RdxKind); | |||
13735 | if (!AllConsts) | |||
13736 | VectorCost = | |||
13737 | TTI->getArithmeticReductionCost(RdxOpcode, VectorTy, FMF, CostKind); | |||
13738 | ScalarCost = EvaluateScalarCost([&]() { | |||
13739 | return TTI->getArithmeticInstrCost(RdxOpcode, ScalarTy, CostKind); | |||
13740 | }); | |||
13741 | break; | |||
13742 | } | |||
13743 | case RecurKind::FMax: | |||
13744 | case RecurKind::FMin: | |||
13745 | case RecurKind::SMax: | |||
13746 | case RecurKind::SMin: | |||
13747 | case RecurKind::UMax: | |||
13748 | case RecurKind::UMin: { | |||
13749 | if (!AllConsts) { | |||
13750 | auto *VecCondTy = | |||
13751 | cast<VectorType>(CmpInst::makeCmpResultType(VectorTy)); | |||
13752 | bool IsUnsigned = | |||
13753 | RdxKind == RecurKind::UMax || RdxKind == RecurKind::UMin; | |||
13754 | VectorCost = TTI->getMinMaxReductionCost(VectorTy, VecCondTy, | |||
13755 | IsUnsigned, FMF, CostKind); | |||
13756 | } | |||
13757 | Intrinsic::ID Id = getMinMaxReductionIntrinsicOp(RdxKind); | |||
13758 | ScalarCost = EvaluateScalarCost([&]() { | |||
13759 | IntrinsicCostAttributes ICA(Id, ScalarTy, {ScalarTy, ScalarTy}, FMF); | |||
13760 | return TTI->getIntrinsicInstrCost(ICA, CostKind); | |||
13761 | }); | |||
13762 | break; | |||
13763 | } | |||
13764 | default: | |||
13765 | 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", 13765); | |||
13766 | } | |||
13767 | ||||
13768 | 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) | |||
13769 | << " 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) | |||
13770 | << " (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); | |||
13771 | return VectorCost - ScalarCost; | |||
13772 | } | |||
13773 | ||||
13774 | /// Emit a horizontal reduction of the vectorized value. | |||
13775 | Value *emitReduction(Value *VectorizedValue, IRBuilder<> &Builder, | |||
13776 | unsigned ReduxWidth, const TargetTransformInfo *TTI) { | |||
13777 | 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", 13777, __extension__ __PRETTY_FUNCTION__)); | |||
13778 | 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", 13779, __extension__ __PRETTY_FUNCTION__)) | |||
13779 | "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", 13779, __extension__ __PRETTY_FUNCTION__)); | |||
13780 | 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", 13781, __extension__ __PRETTY_FUNCTION__)) | |||
13781 | "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", 13781, __extension__ __PRETTY_FUNCTION__)); | |||
13782 | ||||
13783 | ++NumVectorInstructions; | |||
13784 | return createSimpleTargetReduction(Builder, TTI, VectorizedValue, RdxKind); | |||
13785 | } | |||
13786 | ||||
13787 | /// Emits optimized code for unique scalar value reused \p Cnt times. | |||
13788 | Value *emitScaleForReusedOps(Value *VectorizedValue, IRBuilderBase &Builder, | |||
13789 | unsigned Cnt) { | |||
13790 | assert(IsSupportedHorRdxIdentityOp &&(static_cast <bool> (IsSupportedHorRdxIdentityOp && "The optimization of matched scalar identity horizontal reductions " "must be supported.") ? void (0) : __assert_fail ("IsSupportedHorRdxIdentityOp && \"The optimization of matched scalar identity horizontal reductions \" \"must be supported.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 13792, __extension__ __PRETTY_FUNCTION__)) | |||
13791 | "The optimization of matched scalar identity horizontal reductions "(static_cast <bool> (IsSupportedHorRdxIdentityOp && "The optimization of matched scalar identity horizontal reductions " "must be supported.") ? void (0) : __assert_fail ("IsSupportedHorRdxIdentityOp && \"The optimization of matched scalar identity horizontal reductions \" \"must be supported.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 13792, __extension__ __PRETTY_FUNCTION__)) | |||
13792 | "must be supported.")(static_cast <bool> (IsSupportedHorRdxIdentityOp && "The optimization of matched scalar identity horizontal reductions " "must be supported.") ? void (0) : __assert_fail ("IsSupportedHorRdxIdentityOp && \"The optimization of matched scalar identity horizontal reductions \" \"must be supported.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 13792, __extension__ __PRETTY_FUNCTION__)); | |||
13793 | switch (RdxKind) { | |||
13794 | case RecurKind::Add: { | |||
13795 | // res = mul vv, n | |||
13796 | Value *Scale = ConstantInt::get(VectorizedValue->getType(), Cnt); | |||
13797 | LLVM_DEBUG(dbgs() << "SLP: Add (to-mul) " << Cnt << "of "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Add (to-mul) " << Cnt << "of " << VectorizedValue << ". (HorRdx)\n"; } } while (false) | |||
13798 | << VectorizedValue << ". (HorRdx)\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Add (to-mul) " << Cnt << "of " << VectorizedValue << ". (HorRdx)\n"; } } while (false); | |||
13799 | return Builder.CreateMul(VectorizedValue, Scale); | |||
13800 | } | |||
13801 | case RecurKind::Xor: { | |||
13802 | // res = n % 2 ? 0 : vv | |||
13803 | LLVM_DEBUG(dbgs() << "SLP: Xor " << Cnt << "of " << VectorizedValuedo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Xor " << Cnt << "of " << VectorizedValue << ". (HorRdx)\n"; } } while ( false) | |||
13804 | << ". (HorRdx)\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Xor " << Cnt << "of " << VectorizedValue << ". (HorRdx)\n"; } } while ( false); | |||
13805 | if (Cnt % 2 == 0) | |||
13806 | return Constant::getNullValue(VectorizedValue->getType()); | |||
13807 | return VectorizedValue; | |||
13808 | } | |||
13809 | case RecurKind::FAdd: { | |||
13810 | // res = fmul v, n | |||
13811 | Value *Scale = ConstantFP::get(VectorizedValue->getType(), Cnt); | |||
13812 | LLVM_DEBUG(dbgs() << "SLP: FAdd (to-fmul) " << Cnt << "of "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: FAdd (to-fmul) " << Cnt << "of " << VectorizedValue << ". (HorRdx)\n" ; } } while (false) | |||
13813 | << VectorizedValue << ". (HorRdx)\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: FAdd (to-fmul) " << Cnt << "of " << VectorizedValue << ". (HorRdx)\n" ; } } while (false); | |||
13814 | return Builder.CreateFMul(VectorizedValue, Scale); | |||
13815 | } | |||
13816 | case RecurKind::And: | |||
13817 | case RecurKind::Or: | |||
13818 | case RecurKind::SMax: | |||
13819 | case RecurKind::SMin: | |||
13820 | case RecurKind::UMax: | |||
13821 | case RecurKind::UMin: | |||
13822 | case RecurKind::FMax: | |||
13823 | case RecurKind::FMin: | |||
13824 | // res = vv | |||
13825 | return VectorizedValue; | |||
13826 | case RecurKind::Mul: | |||
13827 | case RecurKind::FMul: | |||
13828 | case RecurKind::FMulAdd: | |||
13829 | case RecurKind::SelectICmp: | |||
13830 | case RecurKind::SelectFCmp: | |||
13831 | case RecurKind::None: | |||
13832 | llvm_unreachable("Unexpected reduction kind for repeated scalar.")::llvm::llvm_unreachable_internal("Unexpected reduction kind for repeated scalar." , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 13832); | |||
13833 | } | |||
13834 | return nullptr; | |||
13835 | } | |||
13836 | ||||
13837 | /// Emits actual operation for the scalar identity values, found during | |||
13838 | /// horizontal reduction analysis. | |||
13839 | Value *emitReusedOps(Value *VectorizedValue, IRBuilderBase &Builder, | |||
13840 | ArrayRef<Value *> VL, | |||
13841 | const MapVector<Value *, unsigned> &SameValuesCounter, | |||
13842 | const DenseMap<Value *, Value *> &TrackedToOrig) { | |||
13843 | assert(IsSupportedHorRdxIdentityOp &&(static_cast <bool> (IsSupportedHorRdxIdentityOp && "The optimization of matched scalar identity horizontal reductions " "must be supported.") ? void (0) : __assert_fail ("IsSupportedHorRdxIdentityOp && \"The optimization of matched scalar identity horizontal reductions \" \"must be supported.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 13845, __extension__ __PRETTY_FUNCTION__)) | |||
13844 | "The optimization of matched scalar identity horizontal reductions "(static_cast <bool> (IsSupportedHorRdxIdentityOp && "The optimization of matched scalar identity horizontal reductions " "must be supported.") ? void (0) : __assert_fail ("IsSupportedHorRdxIdentityOp && \"The optimization of matched scalar identity horizontal reductions \" \"must be supported.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 13845, __extension__ __PRETTY_FUNCTION__)) | |||
13845 | "must be supported.")(static_cast <bool> (IsSupportedHorRdxIdentityOp && "The optimization of matched scalar identity horizontal reductions " "must be supported.") ? void (0) : __assert_fail ("IsSupportedHorRdxIdentityOp && \"The optimization of matched scalar identity horizontal reductions \" \"must be supported.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 13845, __extension__ __PRETTY_FUNCTION__)); | |||
13846 | switch (RdxKind) { | |||
13847 | case RecurKind::Add: { | |||
13848 | // root = mul prev_root, <1, 1, n, 1> | |||
13849 | SmallVector<Constant *> Vals; | |||
13850 | for (Value *V : VL) { | |||
13851 | unsigned Cnt = SameValuesCounter.lookup(TrackedToOrig.find(V)->second); | |||
13852 | Vals.push_back(ConstantInt::get(V->getType(), Cnt, /*IsSigned=*/false)); | |||
13853 | } | |||
13854 | auto *Scale = ConstantVector::get(Vals); | |||
13855 | LLVM_DEBUG(dbgs() << "SLP: Add (to-mul) " << Scale << "of "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Add (to-mul) " << Scale << "of " << VectorizedValue << ". (HorRdx)\n" ; } } while (false) | |||
13856 | << VectorizedValue << ". (HorRdx)\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Add (to-mul) " << Scale << "of " << VectorizedValue << ". (HorRdx)\n" ; } } while (false); | |||
13857 | return Builder.CreateMul(VectorizedValue, Scale); | |||
13858 | } | |||
13859 | case RecurKind::And: | |||
13860 | case RecurKind::Or: | |||
13861 | // No need for multiple or/and(s). | |||
13862 | LLVM_DEBUG(dbgs() << "SLP: And/or of same " << VectorizedValuedo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: And/or of same " << VectorizedValue << ". (HorRdx)\n"; } } while (false) | |||
13863 | << ". (HorRdx)\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: And/or of same " << VectorizedValue << ". (HorRdx)\n"; } } while (false); | |||
13864 | return VectorizedValue; | |||
13865 | case RecurKind::SMax: | |||
13866 | case RecurKind::SMin: | |||
13867 | case RecurKind::UMax: | |||
13868 | case RecurKind::UMin: | |||
13869 | case RecurKind::FMax: | |||
13870 | case RecurKind::FMin: | |||
13871 | // No need for multiple min/max(s) of the same value. | |||
13872 | LLVM_DEBUG(dbgs() << "SLP: Max/min of same " << VectorizedValuedo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Max/min of same " << VectorizedValue << ". (HorRdx)\n"; } } while (false) | |||
13873 | << ". (HorRdx)\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Max/min of same " << VectorizedValue << ". (HorRdx)\n"; } } while (false); | |||
13874 | return VectorizedValue; | |||
13875 | case RecurKind::Xor: { | |||
13876 | // Replace values with even number of repeats with 0, since | |||
13877 | // x xor x = 0. | |||
13878 | // root = shuffle prev_root, zeroinitalizer, <0, 1, 2, vf, 4, vf, 5, 6, | |||
13879 | // 7>, if elements 4th and 6th elements have even number of repeats. | |||
13880 | SmallVector<int> Mask( | |||
13881 | cast<FixedVectorType>(VectorizedValue->getType())->getNumElements(), | |||
13882 | PoisonMaskElem); | |||
13883 | std::iota(Mask.begin(), Mask.end(), 0); | |||
13884 | bool NeedShuffle = false; | |||
13885 | for (unsigned I = 0, VF = VL.size(); I < VF; ++I) { | |||
13886 | Value *V = VL[I]; | |||
13887 | unsigned Cnt = SameValuesCounter.lookup(TrackedToOrig.find(V)->second); | |||
13888 | if (Cnt % 2 == 0) { | |||
13889 | Mask[I] = VF; | |||
13890 | NeedShuffle = true; | |||
13891 | } | |||
13892 | } | |||
13893 | LLVM_DEBUG(dbgs() << "SLP: Xor <"; for (int Ido { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Xor <"; for (int I : Mask ) dbgs() << I << " "; dbgs() << "> of " << VectorizedValue << ". (HorRdx)\n"; } } while (false) | |||
13894 | : Mask) dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Xor <"; for (int I : Mask ) dbgs() << I << " "; dbgs() << "> of " << VectorizedValue << ". (HorRdx)\n"; } } while (false) | |||
13895 | << I << " ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Xor <"; for (int I : Mask ) dbgs() << I << " "; dbgs() << "> of " << VectorizedValue << ". (HorRdx)\n"; } } while (false) | |||
13896 | dbgs() << "> of " << VectorizedValue << ". (HorRdx)\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Xor <"; for (int I : Mask ) dbgs() << I << " "; dbgs() << "> of " << VectorizedValue << ". (HorRdx)\n"; } } while (false); | |||
13897 | if (NeedShuffle) | |||
13898 | VectorizedValue = Builder.CreateShuffleVector( | |||
13899 | VectorizedValue, | |||
13900 | ConstantVector::getNullValue(VectorizedValue->getType()), Mask); | |||
13901 | return VectorizedValue; | |||
13902 | } | |||
13903 | case RecurKind::FAdd: { | |||
13904 | // root = fmul prev_root, <1.0, 1.0, n.0, 1.0> | |||
13905 | SmallVector<Constant *> Vals; | |||
13906 | for (Value *V : VL) { | |||
13907 | unsigned Cnt = SameValuesCounter.lookup(TrackedToOrig.find(V)->second); | |||
13908 | Vals.push_back(ConstantFP::get(V->getType(), Cnt)); | |||
13909 | } | |||
13910 | auto *Scale = ConstantVector::get(Vals); | |||
13911 | return Builder.CreateFMul(VectorizedValue, Scale); | |||
13912 | } | |||
13913 | case RecurKind::Mul: | |||
13914 | case RecurKind::FMul: | |||
13915 | case RecurKind::FMulAdd: | |||
13916 | case RecurKind::SelectICmp: | |||
13917 | case RecurKind::SelectFCmp: | |||
13918 | case RecurKind::None: | |||
13919 | llvm_unreachable("Unexpected reduction kind for reused scalars.")::llvm::llvm_unreachable_internal("Unexpected reduction kind for reused scalars." , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 13919); | |||
13920 | } | |||
13921 | return nullptr; | |||
13922 | } | |||
13923 | }; | |||
13924 | } // end anonymous namespace | |||
13925 | ||||
13926 | static std::optional<unsigned> getAggregateSize(Instruction *InsertInst) { | |||
13927 | if (auto *IE = dyn_cast<InsertElementInst>(InsertInst)) | |||
13928 | return cast<FixedVectorType>(IE->getType())->getNumElements(); | |||
13929 | ||||
13930 | unsigned AggregateSize = 1; | |||
13931 | auto *IV = cast<InsertValueInst>(InsertInst); | |||
13932 | Type *CurrentType = IV->getType(); | |||
13933 | do { | |||
13934 | if (auto *ST = dyn_cast<StructType>(CurrentType)) { | |||
13935 | for (auto *Elt : ST->elements()) | |||
13936 | if (Elt != ST->getElementType(0)) // check homogeneity | |||
13937 | return std::nullopt; | |||
13938 | AggregateSize *= ST->getNumElements(); | |||
13939 | CurrentType = ST->getElementType(0); | |||
13940 | } else if (auto *AT = dyn_cast<ArrayType>(CurrentType)) { | |||
13941 | AggregateSize *= AT->getNumElements(); | |||
13942 | CurrentType = AT->getElementType(); | |||
13943 | } else if (auto *VT = dyn_cast<FixedVectorType>(CurrentType)) { | |||
13944 | AggregateSize *= VT->getNumElements(); | |||
13945 | return AggregateSize; | |||
13946 | } else if (CurrentType->isSingleValueType()) { | |||
13947 | return AggregateSize; | |||
13948 | } else { | |||
13949 | return std::nullopt; | |||
13950 | } | |||
13951 | } while (true); | |||
13952 | } | |||
13953 | ||||
13954 | static void findBuildAggregate_rec(Instruction *LastInsertInst, | |||
13955 | TargetTransformInfo *TTI, | |||
13956 | SmallVectorImpl<Value *> &BuildVectorOpds, | |||
13957 | SmallVectorImpl<Value *> &InsertElts, | |||
13958 | unsigned OperandOffset) { | |||
13959 | do { | |||
13960 | Value *InsertedOperand = LastInsertInst->getOperand(1); | |||
13961 | std::optional<unsigned> OperandIndex = | |||
13962 | getInsertIndex(LastInsertInst, OperandOffset); | |||
13963 | if (!OperandIndex) | |||
13964 | return; | |||
13965 | if (isa<InsertElementInst, InsertValueInst>(InsertedOperand)) { | |||
13966 | findBuildAggregate_rec(cast<Instruction>(InsertedOperand), TTI, | |||
13967 | BuildVectorOpds, InsertElts, *OperandIndex); | |||
13968 | ||||
13969 | } else { | |||
13970 | BuildVectorOpds[*OperandIndex] = InsertedOperand; | |||
13971 | InsertElts[*OperandIndex] = LastInsertInst; | |||
13972 | } | |||
13973 | LastInsertInst = dyn_cast<Instruction>(LastInsertInst->getOperand(0)); | |||
13974 | } while (LastInsertInst != nullptr && | |||
13975 | isa<InsertValueInst, InsertElementInst>(LastInsertInst) && | |||
13976 | LastInsertInst->hasOneUse()); | |||
13977 | } | |||
13978 | ||||
13979 | /// Recognize construction of vectors like | |||
13980 | /// %ra = insertelement <4 x float> poison, float %s0, i32 0 | |||
13981 | /// %rb = insertelement <4 x float> %ra, float %s1, i32 1 | |||
13982 | /// %rc = insertelement <4 x float> %rb, float %s2, i32 2 | |||
13983 | /// %rd = insertelement <4 x float> %rc, float %s3, i32 3 | |||
13984 | /// starting from the last insertelement or insertvalue instruction. | |||
13985 | /// | |||
13986 | /// Also recognize homogeneous aggregates like {<2 x float>, <2 x float>}, | |||
13987 | /// {{float, float}, {float, float}}, [2 x {float, float}] and so on. | |||
13988 | /// See llvm/test/Transforms/SLPVectorizer/X86/pr42022.ll for examples. | |||
13989 | /// | |||
13990 | /// Assume LastInsertInst is of InsertElementInst or InsertValueInst type. | |||
13991 | /// | |||
13992 | /// \return true if it matches. | |||
13993 | static bool findBuildAggregate(Instruction *LastInsertInst, | |||
13994 | TargetTransformInfo *TTI, | |||
13995 | SmallVectorImpl<Value *> &BuildVectorOpds, | |||
13996 | SmallVectorImpl<Value *> &InsertElts) { | |||
13997 | ||||
13998 | 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", 14000, __extension__ __PRETTY_FUNCTION__)) | |||
13999 | 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", 14000, __extension__ __PRETTY_FUNCTION__)) | |||
14000 | "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", 14000, __extension__ __PRETTY_FUNCTION__)); | |||
14001 | ||||
14002 | 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", 14003, __extension__ __PRETTY_FUNCTION__)) | |||
14003 | "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", 14003, __extension__ __PRETTY_FUNCTION__)); | |||
14004 | ||||
14005 | std::optional<unsigned> AggregateSize = getAggregateSize(LastInsertInst); | |||
14006 | if (!AggregateSize) | |||
14007 | return false; | |||
14008 | BuildVectorOpds.resize(*AggregateSize); | |||
14009 | InsertElts.resize(*AggregateSize); | |||
14010 | ||||
14011 | findBuildAggregate_rec(LastInsertInst, TTI, BuildVectorOpds, InsertElts, 0); | |||
14012 | llvm::erase_value(BuildVectorOpds, nullptr); | |||
14013 | llvm::erase_value(InsertElts, nullptr); | |||
14014 | if (BuildVectorOpds.size() >= 2) | |||
14015 | return true; | |||
14016 | ||||
14017 | return false; | |||
14018 | } | |||
14019 | ||||
14020 | /// Try and get a reduction instruction from a phi node. | |||
14021 | /// | |||
14022 | /// Given a phi node \p P in a block \p ParentBB, consider possible reductions | |||
14023 | /// if they come from either \p ParentBB or a containing loop latch. | |||
14024 | /// | |||
14025 | /// \returns A candidate reduction value if possible, or \code nullptr \endcode | |||
14026 | /// if not possible. | |||
14027 | static Instruction *getReductionInstr(const DominatorTree *DT, PHINode *P, | |||
14028 | BasicBlock *ParentBB, LoopInfo *LI) { | |||
14029 | // There are situations where the reduction value is not dominated by the | |||
14030 | // reduction phi. Vectorizing such cases has been reported to cause | |||
14031 | // miscompiles. See PR25787. | |||
14032 | auto DominatedReduxValue = [&](Value *R) { | |||
14033 | return isa<Instruction>(R) && | |||
14034 | DT->dominates(P->getParent(), cast<Instruction>(R)->getParent()); | |||
14035 | }; | |||
14036 | ||||
14037 | Instruction *Rdx = nullptr; | |||
14038 | ||||
14039 | // Return the incoming value if it comes from the same BB as the phi node. | |||
14040 | if (P->getIncomingBlock(0) == ParentBB) { | |||
14041 | Rdx = dyn_cast<Instruction>(P->getIncomingValue(0)); | |||
14042 | } else if (P->getIncomingBlock(1) == ParentBB) { | |||
14043 | Rdx = dyn_cast<Instruction>(P->getIncomingValue(1)); | |||
14044 | } | |||
14045 | ||||
14046 | if (Rdx && DominatedReduxValue(Rdx)) | |||
14047 | return Rdx; | |||
14048 | ||||
14049 | // Otherwise, check whether we have a loop latch to look at. | |||
14050 | Loop *BBL = LI->getLoopFor(ParentBB); | |||
14051 | if (!BBL) | |||
14052 | return nullptr; | |||
14053 | BasicBlock *BBLatch = BBL->getLoopLatch(); | |||
14054 | if (!BBLatch) | |||
14055 | return nullptr; | |||
14056 | ||||
14057 | // There is a loop latch, return the incoming value if it comes from | |||
14058 | // that. This reduction pattern occasionally turns up. | |||
14059 | if (P->getIncomingBlock(0) == BBLatch) { | |||
14060 | Rdx = dyn_cast<Instruction>(P->getIncomingValue(0)); | |||
14061 | } else if (P->getIncomingBlock(1) == BBLatch) { | |||
14062 | Rdx = dyn_cast<Instruction>(P->getIncomingValue(1)); | |||
14063 | } | |||
14064 | ||||
14065 | if (Rdx && DominatedReduxValue(Rdx)) | |||
14066 | return Rdx; | |||
14067 | ||||
14068 | return nullptr; | |||
14069 | } | |||
14070 | ||||
14071 | static bool matchRdxBop(Instruction *I, Value *&V0, Value *&V1) { | |||
14072 | if (match(I, m_BinOp(m_Value(V0), m_Value(V1)))) | |||
14073 | return true; | |||
14074 | if (match(I, m_Intrinsic<Intrinsic::maxnum>(m_Value(V0), m_Value(V1)))) | |||
14075 | return true; | |||
14076 | if (match(I, m_Intrinsic<Intrinsic::minnum>(m_Value(V0), m_Value(V1)))) | |||
14077 | return true; | |||
14078 | if (match(I, m_Intrinsic<Intrinsic::smax>(m_Value(V0), m_Value(V1)))) | |||
14079 | return true; | |||
14080 | if (match(I, m_Intrinsic<Intrinsic::smin>(m_Value(V0), m_Value(V1)))) | |||
14081 | return true; | |||
14082 | if (match(I, m_Intrinsic<Intrinsic::umax>(m_Value(V0), m_Value(V1)))) | |||
14083 | return true; | |||
14084 | if (match(I, m_Intrinsic<Intrinsic::umin>(m_Value(V0), m_Value(V1)))) | |||
14085 | return true; | |||
14086 | return false; | |||
14087 | } | |||
14088 | ||||
14089 | /// We could have an initial reduction that is not an add. | |||
14090 | /// r *= v1 + v2 + v3 + v4 | |||
14091 | /// In such a case start looking for a tree rooted in the first '+'. | |||
14092 | /// \Returns the new root if found, which may be nullptr if not an instruction. | |||
14093 | static Instruction *tryGetSecondaryReductionRoot(PHINode *Phi, | |||
14094 | Instruction *Root) { | |||
14095 | assert((isa<BinaryOperator>(Root) || isa<SelectInst>(Root) ||(static_cast <bool> ((isa<BinaryOperator>(Root) || isa<SelectInst>(Root) || isa<IntrinsicInst>(Root )) && "Expected binop, select, or intrinsic for reduction matching" ) ? void (0) : __assert_fail ("(isa<BinaryOperator>(Root) || isa<SelectInst>(Root) || isa<IntrinsicInst>(Root)) && \"Expected binop, select, or intrinsic for reduction matching\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 14097, __extension__ __PRETTY_FUNCTION__)) | |||
14096 | isa<IntrinsicInst>(Root)) &&(static_cast <bool> ((isa<BinaryOperator>(Root) || isa<SelectInst>(Root) || isa<IntrinsicInst>(Root )) && "Expected binop, select, or intrinsic for reduction matching" ) ? void (0) : __assert_fail ("(isa<BinaryOperator>(Root) || isa<SelectInst>(Root) || isa<IntrinsicInst>(Root)) && \"Expected binop, select, or intrinsic for reduction matching\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 14097, __extension__ __PRETTY_FUNCTION__)) | |||
14097 | "Expected binop, select, or intrinsic for reduction matching")(static_cast <bool> ((isa<BinaryOperator>(Root) || isa<SelectInst>(Root) || isa<IntrinsicInst>(Root )) && "Expected binop, select, or intrinsic for reduction matching" ) ? void (0) : __assert_fail ("(isa<BinaryOperator>(Root) || isa<SelectInst>(Root) || isa<IntrinsicInst>(Root)) && \"Expected binop, select, or intrinsic for reduction matching\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 14097, __extension__ __PRETTY_FUNCTION__)); | |||
14098 | Value *LHS = | |||
14099 | Root->getOperand(HorizontalReduction::getFirstOperandIndex(Root)); | |||
14100 | Value *RHS = | |||
14101 | Root->getOperand(HorizontalReduction::getFirstOperandIndex(Root) + 1); | |||
14102 | if (LHS == Phi) | |||
14103 | return dyn_cast<Instruction>(RHS); | |||
14104 | if (RHS == Phi) | |||
14105 | return dyn_cast<Instruction>(LHS); | |||
14106 | return nullptr; | |||
14107 | } | |||
14108 | ||||
14109 | /// \p Returns the first operand of \p I that does not match \p Phi. If | |||
14110 | /// operand is not an instruction it returns nullptr. | |||
14111 | static Instruction *getNonPhiOperand(Instruction *I, PHINode *Phi) { | |||
14112 | Value *Op0 = nullptr; | |||
14113 | Value *Op1 = nullptr; | |||
14114 | if (!matchRdxBop(I, Op0, Op1)) | |||
14115 | return nullptr; | |||
14116 | return dyn_cast<Instruction>(Op0 == Phi ? Op1 : Op0); | |||
14117 | } | |||
14118 | ||||
14119 | /// \Returns true if \p I is a candidate instruction for reduction vectorization. | |||
14120 | static bool isReductionCandidate(Instruction *I) { | |||
14121 | bool IsSelect = match(I, m_Select(m_Value(), m_Value(), m_Value())); | |||
14122 | Value *B0 = nullptr, *B1 = nullptr; | |||
14123 | bool IsBinop = matchRdxBop(I, B0, B1); | |||
14124 | return IsBinop || IsSelect; | |||
14125 | } | |||
14126 | ||||
14127 | bool SLPVectorizerPass::vectorizeHorReduction( | |||
14128 | PHINode *P, Instruction *Root, BasicBlock *BB, BoUpSLP &R, TargetTransformInfo *TTI, | |||
14129 | SmallVectorImpl<WeakTrackingVH> &PostponedInsts) { | |||
14130 | if (!ShouldVectorizeHor) | |||
14131 | return false; | |||
14132 | bool TryOperandsAsNewSeeds = P && isa<BinaryOperator>(Root); | |||
14133 | ||||
14134 | if (Root->getParent() != BB || isa<PHINode>(Root)) | |||
14135 | return false; | |||
14136 | ||||
14137 | // If we can find a secondary reduction root, use that instead. | |||
14138 | auto SelectRoot = [&]() { | |||
14139 | if (TryOperandsAsNewSeeds && isReductionCandidate(Root) && | |||
14140 | HorizontalReduction::getRdxKind(Root) != RecurKind::None) | |||
14141 | if (Instruction *NewRoot = tryGetSecondaryReductionRoot(P, Root)) | |||
14142 | return NewRoot; | |||
14143 | return Root; | |||
14144 | }; | |||
14145 | ||||
14146 | // Start analysis starting from Root instruction. If horizontal reduction is | |||
14147 | // found, try to vectorize it. If it is not a horizontal reduction or | |||
14148 | // vectorization is not possible or not effective, and currently analyzed | |||
14149 | // instruction is a binary operation, try to vectorize the operands, using | |||
14150 | // pre-order DFS traversal order. If the operands were not vectorized, repeat | |||
14151 | // the same procedure considering each operand as a possible root of the | |||
14152 | // horizontal reduction. | |||
14153 | // Interrupt the process if the Root instruction itself was vectorized or all | |||
14154 | // sub-trees not higher that RecursionMaxDepth were analyzed/vectorized. | |||
14155 | // If a horizintal reduction was not matched or vectorized we collect | |||
14156 | // instructions for possible later attempts for vectorization. | |||
14157 | std::queue<std::pair<Instruction *, unsigned>> Stack; | |||
14158 | Stack.emplace(SelectRoot(), 0); | |||
14159 | SmallPtrSet<Value *, 8> VisitedInstrs; | |||
14160 | bool Res = false; | |||
14161 | auto &&TryToReduce = [this, TTI, &R](Instruction *Inst) -> Value * { | |||
14162 | if (R.isAnalyzedReductionRoot(Inst)) | |||
14163 | return nullptr; | |||
14164 | if (!isReductionCandidate(Inst)) | |||
14165 | return nullptr; | |||
14166 | HorizontalReduction HorRdx; | |||
14167 | if (!HorRdx.matchAssociativeReduction(R, Inst, *SE, *DL, *TLI)) | |||
14168 | return nullptr; | |||
14169 | return HorRdx.tryToReduce(R, TTI, *TLI); | |||
14170 | }; | |||
14171 | auto TryAppendToPostponedInsts = [&](Instruction *FutureSeed) { | |||
14172 | if (TryOperandsAsNewSeeds && FutureSeed == Root) { | |||
14173 | FutureSeed = getNonPhiOperand(Root, P); | |||
14174 | if (!FutureSeed) | |||
14175 | return false; | |||
14176 | } | |||
14177 | // Do not collect CmpInst or InsertElementInst/InsertValueInst as their | |||
14178 | // analysis is done separately. | |||
14179 | if (!isa<CmpInst, InsertElementInst, InsertValueInst>(FutureSeed)) | |||
14180 | PostponedInsts.push_back(FutureSeed); | |||
14181 | return true; | |||
14182 | }; | |||
14183 | ||||
14184 | while (!Stack.empty()) { | |||
14185 | Instruction *Inst; | |||
14186 | unsigned Level; | |||
14187 | std::tie(Inst, Level) = Stack.front(); | |||
14188 | Stack.pop(); | |||
14189 | // Do not try to analyze instruction that has already been vectorized. | |||
14190 | // This may happen when we vectorize instruction operands on a previous | |||
14191 | // iteration while stack was populated before that happened. | |||
14192 | if (R.isDeleted(Inst)) | |||
14193 | continue; | |||
14194 | if (Value *VectorizedV = TryToReduce(Inst)) { | |||
14195 | Res = true; | |||
14196 | if (auto *I = dyn_cast<Instruction>(VectorizedV)) { | |||
14197 | // Try to find another reduction. | |||
14198 | Stack.emplace(I, Level); | |||
14199 | continue; | |||
14200 | } | |||
14201 | } else { | |||
14202 | // We could not vectorize `Inst` so try to use it as a future seed. | |||
14203 | if (!TryAppendToPostponedInsts(Inst)) { | |||
14204 | assert(Stack.empty() && "Expected empty stack")(static_cast <bool> (Stack.empty() && "Expected empty stack" ) ? void (0) : __assert_fail ("Stack.empty() && \"Expected empty stack\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 14204, __extension__ __PRETTY_FUNCTION__)); | |||
14205 | break; | |||
14206 | } | |||
14207 | } | |||
14208 | ||||
14209 | // Try to vectorize operands. | |||
14210 | // Continue analysis for the instruction from the same basic block only to | |||
14211 | // save compile time. | |||
14212 | if (++Level < RecursionMaxDepth) | |||
14213 | for (auto *Op : Inst->operand_values()) | |||
14214 | if (VisitedInstrs.insert(Op).second) | |||
14215 | if (auto *I = dyn_cast<Instruction>(Op)) | |||
14216 | // Do not try to vectorize CmpInst operands, this is done | |||
14217 | // separately. | |||
14218 | if (!isa<PHINode, CmpInst, InsertElementInst, InsertValueInst>(I) && | |||
14219 | !R.isDeleted(I) && I->getParent() == BB) | |||
14220 | Stack.emplace(I, Level); | |||
14221 | } | |||
14222 | return Res; | |||
14223 | } | |||
14224 | ||||
14225 | bool SLPVectorizerPass::vectorizeRootInstruction(PHINode *P, Instruction *Root, | |||
14226 | BasicBlock *BB, BoUpSLP &R, | |||
14227 | TargetTransformInfo *TTI) { | |||
14228 | SmallVector<WeakTrackingVH> PostponedInsts; | |||
14229 | bool Res = vectorizeHorReduction(P, Root, BB, R, TTI, PostponedInsts); | |||
14230 | Res |= tryToVectorize(PostponedInsts, R); | |||
14231 | return Res; | |||
14232 | } | |||
14233 | ||||
14234 | bool SLPVectorizerPass::tryToVectorize(ArrayRef<WeakTrackingVH> Insts, | |||
14235 | BoUpSLP &R) { | |||
14236 | bool Res = false; | |||
14237 | for (Value *V : Insts) | |||
14238 | if (auto *Inst = dyn_cast<Instruction>(V); Inst && !R.isDeleted(Inst)) | |||
14239 | Res |= tryToVectorize(Inst, R); | |||
14240 | return Res; | |||
14241 | } | |||
14242 | ||||
14243 | bool SLPVectorizerPass::vectorizeInsertValueInst(InsertValueInst *IVI, | |||
14244 | BasicBlock *BB, BoUpSLP &R) { | |||
14245 | const DataLayout &DL = BB->getModule()->getDataLayout(); | |||
14246 | if (!R.canMapToVector(IVI->getType(), DL)) | |||
14247 | return false; | |||
14248 | ||||
14249 | SmallVector<Value *, 16> BuildVectorOpds; | |||
14250 | SmallVector<Value *, 16> BuildVectorInsts; | |||
14251 | if (!findBuildAggregate(IVI, TTI, BuildVectorOpds, BuildVectorInsts)) | |||
14252 | return false; | |||
14253 | ||||
14254 | 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); | |||
14255 | // Aggregate value is unlikely to be processed in vector register. | |||
14256 | return tryToVectorizeList(BuildVectorOpds, R); | |||
14257 | } | |||
14258 | ||||
14259 | bool SLPVectorizerPass::vectorizeInsertElementInst(InsertElementInst *IEI, | |||
14260 | BasicBlock *BB, BoUpSLP &R) { | |||
14261 | SmallVector<Value *, 16> BuildVectorInsts; | |||
14262 | SmallVector<Value *, 16> BuildVectorOpds; | |||
14263 | SmallVector<int> Mask; | |||
14264 | if (!findBuildAggregate(IEI, TTI, BuildVectorOpds, BuildVectorInsts) || | |||
14265 | (llvm::all_of( | |||
14266 | BuildVectorOpds, | |||
14267 | [](Value *V) { return isa<ExtractElementInst, UndefValue>(V); }) && | |||
14268 | isFixedVectorShuffle(BuildVectorOpds, Mask))) | |||
14269 | return false; | |||
14270 | ||||
14271 | 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); | |||
14272 | return tryToVectorizeList(BuildVectorInsts, R); | |||
14273 | } | |||
14274 | ||||
14275 | template <typename T> | |||
14276 | static bool tryToVectorizeSequence( | |||
14277 | SmallVectorImpl<T *> &Incoming, function_ref<bool(T *, T *)> Comparator, | |||
14278 | function_ref<bool(T *, T *)> AreCompatible, | |||
14279 | function_ref<bool(ArrayRef<T *>, bool)> TryToVectorizeHelper, | |||
14280 | bool LimitForRegisterSize, BoUpSLP &R) { | |||
14281 | bool Changed = false; | |||
14282 | // Sort by type, parent, operands. | |||
14283 | stable_sort(Incoming, Comparator); | |||
14284 | ||||
14285 | // Try to vectorize elements base on their type. | |||
14286 | SmallVector<T *> Candidates; | |||
14287 | for (auto *IncIt = Incoming.begin(), *E = Incoming.end(); IncIt != E;) { | |||
14288 | // Look for the next elements with the same type, parent and operand | |||
14289 | // kinds. | |||
14290 | auto *SameTypeIt = IncIt; | |||
14291 | while (SameTypeIt != E && AreCompatible(*SameTypeIt, *IncIt)) | |||
14292 | ++SameTypeIt; | |||
14293 | ||||
14294 | // Try to vectorize them. | |||
14295 | unsigned NumElts = (SameTypeIt - IncIt); | |||
14296 | 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) | |||
14297 | << NumElts << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Trying to vectorize starting at nodes (" << NumElts << ")\n"; } } while (false); | |||
14298 | // The vectorization is a 3-state attempt: | |||
14299 | // 1. Try to vectorize instructions with the same/alternate opcodes with the | |||
14300 | // size of maximal register at first. | |||
14301 | // 2. Try to vectorize remaining instructions with the same type, if | |||
14302 | // possible. This may result in the better vectorization results rather than | |||
14303 | // if we try just to vectorize instructions with the same/alternate opcodes. | |||
14304 | // 3. Final attempt to try to vectorize all instructions with the | |||
14305 | // same/alternate ops only, this may result in some extra final | |||
14306 | // vectorization. | |||
14307 | if (NumElts > 1 && | |||
14308 | TryToVectorizeHelper(ArrayRef(IncIt, NumElts), LimitForRegisterSize)) { | |||
14309 | // Success start over because instructions might have been changed. | |||
14310 | Changed = true; | |||
14311 | } else { | |||
14312 | /// \Returns the minimum number of elements that we will attempt to | |||
14313 | /// vectorize. | |||
14314 | auto GetMinNumElements = [&R](Value *V) { | |||
14315 | unsigned EltSize = R.getVectorElementSize(V); | |||
14316 | return std::max(2U, R.getMaxVecRegSize() / EltSize); | |||
14317 | }; | |||
14318 | if (NumElts < GetMinNumElements(*IncIt) && | |||
14319 | (Candidates.empty() || | |||
14320 | Candidates.front()->getType() == (*IncIt)->getType())) { | |||
14321 | Candidates.append(IncIt, std::next(IncIt, NumElts)); | |||
14322 | } | |||
14323 | } | |||
14324 | // Final attempt to vectorize instructions with the same types. | |||
14325 | if (Candidates.size() > 1 && | |||
14326 | (SameTypeIt == E || (*SameTypeIt)->getType() != (*IncIt)->getType())) { | |||
14327 | if (TryToVectorizeHelper(Candidates, /*LimitForRegisterSize=*/false)) { | |||
14328 | // Success start over because instructions might have been changed. | |||
14329 | Changed = true; | |||
14330 | } else if (LimitForRegisterSize) { | |||
14331 | // Try to vectorize using small vectors. | |||
14332 | for (auto *It = Candidates.begin(), *End = Candidates.end(); | |||
14333 | It != End;) { | |||
14334 | auto *SameTypeIt = It; | |||
14335 | while (SameTypeIt != End && AreCompatible(*SameTypeIt, *It)) | |||
14336 | ++SameTypeIt; | |||
14337 | unsigned NumElts = (SameTypeIt - It); | |||
14338 | if (NumElts > 1 && | |||
14339 | TryToVectorizeHelper(ArrayRef(It, NumElts), | |||
14340 | /*LimitForRegisterSize=*/false)) | |||
14341 | Changed = true; | |||
14342 | It = SameTypeIt; | |||
14343 | } | |||
14344 | } | |||
14345 | Candidates.clear(); | |||
14346 | } | |||
14347 | ||||
14348 | // Start over at the next instruction of a different type (or the end). | |||
14349 | IncIt = SameTypeIt; | |||
14350 | } | |||
14351 | return Changed; | |||
14352 | } | |||
14353 | ||||
14354 | /// Compare two cmp instructions. If IsCompatibility is true, function returns | |||
14355 | /// true if 2 cmps have same/swapped predicates and mos compatible corresponding | |||
14356 | /// operands. If IsCompatibility is false, function implements strict weak | |||
14357 | /// ordering relation between two cmp instructions, returning true if the first | |||
14358 | /// instruction is "less" than the second, i.e. its predicate is less than the | |||
14359 | /// predicate of the second or the operands IDs are less than the operands IDs | |||
14360 | /// of the second cmp instruction. | |||
14361 | template <bool IsCompatibility> | |||
14362 | static bool compareCmp(Value *V, Value *V2, TargetLibraryInfo &TLI, | |||
14363 | function_ref<bool(Instruction *)> IsDeleted) { | |||
14364 | auto *CI1 = cast<CmpInst>(V); | |||
14365 | auto *CI2 = cast<CmpInst>(V2); | |||
14366 | if (IsDeleted(CI2) || !isValidElementType(CI2->getType())) | |||
14367 | return false; | |||
14368 | if (CI1->getOperand(0)->getType()->getTypeID() < | |||
14369 | CI2->getOperand(0)->getType()->getTypeID()) | |||
14370 | return !IsCompatibility; | |||
14371 | if (CI1->getOperand(0)->getType()->getTypeID() > | |||
14372 | CI2->getOperand(0)->getType()->getTypeID()) | |||
14373 | return false; | |||
14374 | CmpInst::Predicate Pred1 = CI1->getPredicate(); | |||
14375 | CmpInst::Predicate Pred2 = CI2->getPredicate(); | |||
14376 | CmpInst::Predicate SwapPred1 = CmpInst::getSwappedPredicate(Pred1); | |||
14377 | CmpInst::Predicate SwapPred2 = CmpInst::getSwappedPredicate(Pred2); | |||
14378 | CmpInst::Predicate BasePred1 = std::min(Pred1, SwapPred1); | |||
14379 | CmpInst::Predicate BasePred2 = std::min(Pred2, SwapPred2); | |||
14380 | if (BasePred1 < BasePred2) | |||
14381 | return !IsCompatibility; | |||
14382 | if (BasePred1 > BasePred2) | |||
14383 | return false; | |||
14384 | // Compare operands. | |||
14385 | bool LEPreds = Pred1 <= Pred2; | |||
14386 | bool GEPreds = Pred1 >= Pred2; | |||
14387 | for (int I = 0, E = CI1->getNumOperands(); I < E; ++I) { | |||
14388 | auto *Op1 = CI1->getOperand(LEPreds ? I : E - I - 1); | |||
14389 | auto *Op2 = CI2->getOperand(GEPreds ? I : E - I - 1); | |||
14390 | if (Op1->getValueID() < Op2->getValueID()) | |||
14391 | return !IsCompatibility; | |||
14392 | if (Op1->getValueID() > Op2->getValueID()) | |||
14393 | return false; | |||
14394 | if (auto *I1 = dyn_cast<Instruction>(Op1)) | |||
14395 | if (auto *I2 = dyn_cast<Instruction>(Op2)) { | |||
14396 | if (I1->getParent() != I2->getParent()) | |||
14397 | return false; | |||
14398 | InstructionsState S = getSameOpcode({I1, I2}, TLI); | |||
14399 | if (S.getOpcode()) | |||
14400 | continue; | |||
14401 | return false; | |||
14402 | } | |||
14403 | } | |||
14404 | return IsCompatibility; | |||
14405 | } | |||
14406 | ||||
14407 | bool SLPVectorizerPass::vectorizeCmpInsts(ArrayRef<CmpInst *> CmpInsts, | |||
14408 | BasicBlock *BB, BoUpSLP &R) { | |||
14409 | bool Changed = false; | |||
14410 | // Try to find reductions first. | |||
14411 | for (Instruction *I : CmpInsts) { | |||
14412 | if (R.isDeleted(I)) | |||
14413 | continue; | |||
14414 | for (Value *Op : I->operands()) | |||
14415 | if (auto *RootOp = dyn_cast<Instruction>(Op)) | |||
14416 | Changed |= vectorizeRootInstruction(nullptr, RootOp, BB, R, TTI); | |||
14417 | } | |||
14418 | // Try to vectorize operands as vector bundles. | |||
14419 | for (Instruction *I : CmpInsts) { | |||
14420 | if (R.isDeleted(I)) | |||
14421 | continue; | |||
14422 | Changed |= tryToVectorize(I, R); | |||
14423 | } | |||
14424 | // Try to vectorize list of compares. | |||
14425 | // Sort by type, compare predicate, etc. | |||
14426 | auto CompareSorter = [&](Value *V, Value *V2) { | |||
14427 | return compareCmp<false>(V, V2, *TLI, | |||
14428 | [&R](Instruction *I) { return R.isDeleted(I); }); | |||
14429 | }; | |||
14430 | ||||
14431 | auto AreCompatibleCompares = [&](Value *V1, Value *V2) { | |||
14432 | if (V1 == V2) | |||
14433 | return true; | |||
14434 | return compareCmp<true>(V1, V2, *TLI, | |||
14435 | [&R](Instruction *I) { return R.isDeleted(I); }); | |||
14436 | }; | |||
14437 | ||||
14438 | SmallVector<Value *> Vals(CmpInsts.begin(), CmpInsts.end()); | |||
14439 | Changed |= tryToVectorizeSequence<Value>( | |||
14440 | Vals, CompareSorter, AreCompatibleCompares, | |||
14441 | [this, &R](ArrayRef<Value *> Candidates, bool LimitForRegisterSize) { | |||
14442 | // Exclude possible reductions from other blocks. | |||
14443 | bool ArePossiblyReducedInOtherBlock = any_of(Candidates, [](Value *V) { | |||
14444 | return any_of(V->users(), [V](User *U) { | |||
14445 | auto *Select = dyn_cast<SelectInst>(U); | |||
14446 | return Select && | |||
14447 | Select->getParent() != cast<Instruction>(V)->getParent(); | |||
14448 | }); | |||
14449 | }); | |||
14450 | if (ArePossiblyReducedInOtherBlock) | |||
14451 | return false; | |||
14452 | return tryToVectorizeList(Candidates, R, LimitForRegisterSize); | |||
14453 | }, | |||
14454 | /*LimitForRegisterSize=*/true, R); | |||
14455 | return Changed; | |||
14456 | } | |||
14457 | ||||
14458 | bool SLPVectorizerPass::vectorizeSimpleInstructions(InstSetVector &Instructions, | |||
14459 | BasicBlock *BB, BoUpSLP &R, | |||
14460 | bool AtTerminator) { | |||
14461 | assert(all_of(Instructions,(static_cast <bool> (all_of(Instructions, [](auto *I) { return isa<CmpInst, InsertElementInst, InsertValueInst> (I); }) && "This function only accepts Cmp and Insert instructions" ) ? void (0) : __assert_fail ("all_of(Instructions, [](auto *I) { return isa<CmpInst, InsertElementInst, InsertValueInst>(I); }) && \"This function only accepts Cmp and Insert instructions\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 14465, __extension__ __PRETTY_FUNCTION__)) | |||
14462 | [](auto *I) {(static_cast <bool> (all_of(Instructions, [](auto *I) { return isa<CmpInst, InsertElementInst, InsertValueInst> (I); }) && "This function only accepts Cmp and Insert instructions" ) ? void (0) : __assert_fail ("all_of(Instructions, [](auto *I) { return isa<CmpInst, InsertElementInst, InsertValueInst>(I); }) && \"This function only accepts Cmp and Insert instructions\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 14465, __extension__ __PRETTY_FUNCTION__)) | |||
14463 | return isa<CmpInst, InsertElementInst, InsertValueInst>(I);(static_cast <bool> (all_of(Instructions, [](auto *I) { return isa<CmpInst, InsertElementInst, InsertValueInst> (I); }) && "This function only accepts Cmp and Insert instructions" ) ? void (0) : __assert_fail ("all_of(Instructions, [](auto *I) { return isa<CmpInst, InsertElementInst, InsertValueInst>(I); }) && \"This function only accepts Cmp and Insert instructions\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 14465, __extension__ __PRETTY_FUNCTION__)) | |||
14464 | }) &&(static_cast <bool> (all_of(Instructions, [](auto *I) { return isa<CmpInst, InsertElementInst, InsertValueInst> (I); }) && "This function only accepts Cmp and Insert instructions" ) ? void (0) : __assert_fail ("all_of(Instructions, [](auto *I) { return isa<CmpInst, InsertElementInst, InsertValueInst>(I); }) && \"This function only accepts Cmp and Insert instructions\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 14465, __extension__ __PRETTY_FUNCTION__)) | |||
14465 | "This function only accepts Cmp and Insert instructions")(static_cast <bool> (all_of(Instructions, [](auto *I) { return isa<CmpInst, InsertElementInst, InsertValueInst> (I); }) && "This function only accepts Cmp and Insert instructions" ) ? void (0) : __assert_fail ("all_of(Instructions, [](auto *I) { return isa<CmpInst, InsertElementInst, InsertValueInst>(I); }) && \"This function only accepts Cmp and Insert instructions\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 14465, __extension__ __PRETTY_FUNCTION__)); | |||
14466 | bool OpsChanged = false; | |||
14467 | SmallVector<CmpInst *, 4> PostponedCmps; | |||
14468 | SmallVector<WeakTrackingVH> PostponedInsts; | |||
14469 | // pass1 - try to vectorize reductions only | |||
14470 | for (auto *I : reverse(Instructions)) { | |||
14471 | if (R.isDeleted(I)) | |||
14472 | continue; | |||
14473 | if (isa<CmpInst>(I)) { | |||
14474 | PostponedCmps.push_back(cast<CmpInst>(I)); | |||
14475 | continue; | |||
14476 | } | |||
14477 | OpsChanged |= vectorizeHorReduction(nullptr, I, BB, R, TTI, PostponedInsts); | |||
14478 | } | |||
14479 | // pass2 - try to match and vectorize a buildvector sequence. | |||
14480 | for (auto *I : reverse(Instructions)) { | |||
14481 | if (R.isDeleted(I) || isa<CmpInst>(I)) | |||
14482 | continue; | |||
14483 | if (auto *LastInsertValue = dyn_cast<InsertValueInst>(I)) { | |||
14484 | OpsChanged |= vectorizeInsertValueInst(LastInsertValue, BB, R); | |||
14485 | } else if (auto *LastInsertElem = dyn_cast<InsertElementInst>(I)) { | |||
14486 | OpsChanged |= vectorizeInsertElementInst(LastInsertElem, BB, R); | |||
14487 | } | |||
14488 | } | |||
14489 | // Now try to vectorize postponed instructions. | |||
14490 | OpsChanged |= tryToVectorize(PostponedInsts, R); | |||
14491 | ||||
14492 | if (AtTerminator) { | |||
14493 | OpsChanged |= vectorizeCmpInsts(PostponedCmps, BB, R); | |||
14494 | Instructions.clear(); | |||
14495 | } else { | |||
14496 | Instructions.clear(); | |||
14497 | // Insert in reverse order since the PostponedCmps vector was filled in | |||
14498 | // reverse order. | |||
14499 | Instructions.insert(PostponedCmps.rbegin(), PostponedCmps.rend()); | |||
14500 | } | |||
14501 | return OpsChanged; | |||
14502 | } | |||
14503 | ||||
14504 | bool SLPVectorizerPass::vectorizeChainsInBlock(BasicBlock *BB, BoUpSLP &R) { | |||
14505 | bool Changed = false; | |||
14506 | SmallVector<Value *, 4> Incoming; | |||
14507 | SmallPtrSet<Value *, 16> VisitedInstrs; | |||
14508 | // Maps phi nodes to the non-phi nodes found in the use tree for each phi | |||
14509 | // node. Allows better to identify the chains that can be vectorized in the | |||
14510 | // better way. | |||
14511 | DenseMap<Value *, SmallVector<Value *, 4>> PHIToOpcodes; | |||
14512 | auto PHICompare = [this, &PHIToOpcodes](Value *V1, Value *V2) { | |||
14513 | 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", 14515, __extension__ __PRETTY_FUNCTION__)) | |||
14514 | 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", 14515, __extension__ __PRETTY_FUNCTION__)) | |||
14515 | "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", 14515, __extension__ __PRETTY_FUNCTION__)); | |||
14516 | // It is fine to compare type IDs here, since we expect only vectorizable | |||
14517 | // types, like ints, floats and pointers, we don't care about other type. | |||
14518 | if (V1->getType()->getTypeID() < V2->getType()->getTypeID()) | |||
14519 | return true; | |||
14520 | if (V1->getType()->getTypeID() > V2->getType()->getTypeID()) | |||
14521 | return false; | |||
14522 | ArrayRef<Value *> Opcodes1 = PHIToOpcodes[V1]; | |||
14523 | ArrayRef<Value *> Opcodes2 = PHIToOpcodes[V2]; | |||
14524 | if (Opcodes1.size() < Opcodes2.size()) | |||
14525 | return true; | |||
14526 | if (Opcodes1.size() > Opcodes2.size()) | |||
14527 | return false; | |||
14528 | std::optional<bool> ConstOrder; | |||
14529 | for (int I = 0, E = Opcodes1.size(); I < E; ++I) { | |||
14530 | // Undefs are compatible with any other value. | |||
14531 | if (isa<UndefValue>(Opcodes1[I]) || isa<UndefValue>(Opcodes2[I])) { | |||
14532 | if (!ConstOrder) | |||
14533 | ConstOrder = | |||
14534 | !isa<UndefValue>(Opcodes1[I]) && isa<UndefValue>(Opcodes2[I]); | |||
14535 | continue; | |||
14536 | } | |||
14537 | if (auto *I1 = dyn_cast<Instruction>(Opcodes1[I])) | |||
14538 | if (auto *I2 = dyn_cast<Instruction>(Opcodes2[I])) { | |||
14539 | DomTreeNodeBase<BasicBlock> *NodeI1 = DT->getNode(I1->getParent()); | |||
14540 | DomTreeNodeBase<BasicBlock> *NodeI2 = DT->getNode(I2->getParent()); | |||
14541 | if (!NodeI1) | |||
14542 | return NodeI2 != nullptr; | |||
14543 | if (!NodeI2) | |||
14544 | return false; | |||
14545 | 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", 14547, __extension__ __PRETTY_FUNCTION__)) | |||
14546 | (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", 14547, __extension__ __PRETTY_FUNCTION__)) | |||
14547 | "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", 14547, __extension__ __PRETTY_FUNCTION__)); | |||
14548 | if (NodeI1 != NodeI2) | |||
14549 | return NodeI1->getDFSNumIn() < NodeI2->getDFSNumIn(); | |||
14550 | InstructionsState S = getSameOpcode({I1, I2}, *TLI); | |||
14551 | if (S.getOpcode()) | |||
14552 | continue; | |||
14553 | return I1->getOpcode() < I2->getOpcode(); | |||
14554 | } | |||
14555 | if (isa<Constant>(Opcodes1[I]) && isa<Constant>(Opcodes2[I])) { | |||
14556 | if (!ConstOrder) | |||
14557 | ConstOrder = Opcodes1[I]->getValueID() < Opcodes2[I]->getValueID(); | |||
14558 | continue; | |||
14559 | } | |||
14560 | if (Opcodes1[I]->getValueID() < Opcodes2[I]->getValueID()) | |||
14561 | return true; | |||
14562 | if (Opcodes1[I]->getValueID() > Opcodes2[I]->getValueID()) | |||
14563 | return false; | |||
14564 | } | |||
14565 | return ConstOrder && *ConstOrder; | |||
14566 | }; | |||
14567 | auto AreCompatiblePHIs = [&PHIToOpcodes, this](Value *V1, Value *V2) { | |||
14568 | if (V1 == V2) | |||
14569 | return true; | |||
14570 | if (V1->getType() != V2->getType()) | |||
14571 | return false; | |||
14572 | ArrayRef<Value *> Opcodes1 = PHIToOpcodes[V1]; | |||
14573 | ArrayRef<Value *> Opcodes2 = PHIToOpcodes[V2]; | |||
14574 | if (Opcodes1.size() != Opcodes2.size()) | |||
14575 | return false; | |||
14576 | for (int I = 0, E = Opcodes1.size(); I < E; ++I) { | |||
14577 | // Undefs are compatible with any other value. | |||
14578 | if (isa<UndefValue>(Opcodes1[I]) || isa<UndefValue>(Opcodes2[I])) | |||
14579 | continue; | |||
14580 | if (auto *I1 = dyn_cast<Instruction>(Opcodes1[I])) | |||
14581 | if (auto *I2 = dyn_cast<Instruction>(Opcodes2[I])) { | |||
14582 | if (I1->getParent() != I2->getParent()) | |||
14583 | return false; | |||
14584 | InstructionsState S = getSameOpcode({I1, I2}, *TLI); | |||
14585 | if (S.getOpcode()) | |||
14586 | continue; | |||
14587 | return false; | |||
14588 | } | |||
14589 | if (isa<Constant>(Opcodes1[I]) && isa<Constant>(Opcodes2[I])) | |||
14590 | continue; | |||
14591 | if (Opcodes1[I]->getValueID() != Opcodes2[I]->getValueID()) | |||
14592 | return false; | |||
14593 | } | |||
14594 | return true; | |||
14595 | }; | |||
14596 | ||||
14597 | bool HaveVectorizedPhiNodes = false; | |||
14598 | do { | |||
14599 | // Collect the incoming values from the PHIs. | |||
14600 | Incoming.clear(); | |||
14601 | for (Instruction &I : *BB) { | |||
14602 | PHINode *P = dyn_cast<PHINode>(&I); | |||
14603 | if (!P) | |||
14604 | break; | |||
14605 | ||||
14606 | // No need to analyze deleted, vectorized and non-vectorizable | |||
14607 | // instructions. | |||
14608 | if (!VisitedInstrs.count(P) && !R.isDeleted(P) && | |||
14609 | isValidElementType(P->getType())) | |||
14610 | Incoming.push_back(P); | |||
14611 | } | |||
14612 | ||||
14613 | // Find the corresponding non-phi nodes for better matching when trying to | |||
14614 | // build the tree. | |||
14615 | for (Value *V : Incoming) { | |||
14616 | SmallVectorImpl<Value *> &Opcodes = | |||
14617 | PHIToOpcodes.try_emplace(V).first->getSecond(); | |||
14618 | if (!Opcodes.empty()) | |||
14619 | continue; | |||
14620 | SmallVector<Value *, 4> Nodes(1, V); | |||
14621 | SmallPtrSet<Value *, 4> Visited; | |||
14622 | while (!Nodes.empty()) { | |||
14623 | auto *PHI = cast<PHINode>(Nodes.pop_back_val()); | |||
14624 | if (!Visited.insert(PHI).second) | |||
14625 | continue; | |||
14626 | for (Value *V : PHI->incoming_values()) { | |||
14627 | if (auto *PHI1 = dyn_cast<PHINode>((V))) { | |||
14628 | Nodes.push_back(PHI1); | |||
14629 | continue; | |||
14630 | } | |||
14631 | Opcodes.emplace_back(V); | |||
14632 | } | |||
14633 | } | |||
14634 | } | |||
14635 | ||||
14636 | HaveVectorizedPhiNodes = tryToVectorizeSequence<Value>( | |||
14637 | Incoming, PHICompare, AreCompatiblePHIs, | |||
14638 | [this, &R](ArrayRef<Value *> Candidates, bool LimitForRegisterSize) { | |||
14639 | return tryToVectorizeList(Candidates, R, LimitForRegisterSize); | |||
14640 | }, | |||
14641 | /*LimitForRegisterSize=*/true, R); | |||
14642 | Changed |= HaveVectorizedPhiNodes; | |||
14643 | VisitedInstrs.insert(Incoming.begin(), Incoming.end()); | |||
14644 | } while (HaveVectorizedPhiNodes); | |||
14645 | ||||
14646 | VisitedInstrs.clear(); | |||
14647 | ||||
14648 | InstSetVector PostProcessInstructions; | |||
14649 | SmallDenseSet<Instruction *, 4> KeyNodes; | |||
14650 | for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) { | |||
14651 | // Skip instructions with scalable type. The num of elements is unknown at | |||
14652 | // compile-time for scalable type. | |||
14653 | if (isa<ScalableVectorType>(it->getType())) | |||
14654 | continue; | |||
14655 | ||||
14656 | // Skip instructions marked for the deletion. | |||
14657 | if (R.isDeleted(&*it)) | |||
14658 | continue; | |||
14659 | // We may go through BB multiple times so skip the one we have checked. | |||
14660 | if (!VisitedInstrs.insert(&*it).second) { | |||
14661 | if (it->use_empty() && KeyNodes.contains(&*it) && | |||
14662 | vectorizeSimpleInstructions(PostProcessInstructions, BB, R, | |||
14663 | it->isTerminator())) { | |||
14664 | // We would like to start over since some instructions are deleted | |||
14665 | // and the iterator may become invalid value. | |||
14666 | Changed = true; | |||
14667 | it = BB->begin(); | |||
14668 | e = BB->end(); | |||
14669 | } | |||
14670 | continue; | |||
14671 | } | |||
14672 | ||||
14673 | if (isa<DbgInfoIntrinsic>(it)) | |||
14674 | continue; | |||
14675 | ||||
14676 | // Try to vectorize reductions that use PHINodes. | |||
14677 | if (PHINode *P = dyn_cast<PHINode>(it)) { | |||
14678 | // Check that the PHI is a reduction PHI. | |||
14679 | if (P->getNumIncomingValues() == 2) { | |||
14680 | // Try to match and vectorize a horizontal reduction. | |||
14681 | Instruction *Root = getReductionInstr(DT, P, BB, LI); | |||
14682 | if (Root && vectorizeRootInstruction(P, Root, BB, R, TTI)) { | |||
14683 | Changed = true; | |||
14684 | it = BB->begin(); | |||
14685 | e = BB->end(); | |||
14686 | continue; | |||
14687 | } | |||
14688 | } | |||
14689 | // Try to vectorize the incoming values of the PHI, to catch reductions | |||
14690 | // that feed into PHIs. | |||
14691 | for (unsigned I = 0, E = P->getNumIncomingValues(); I != E; I++) { | |||
14692 | // Skip if the incoming block is the current BB for now. Also, bypass | |||
14693 | // unreachable IR for efficiency and to avoid crashing. | |||
14694 | // TODO: Collect the skipped incoming values and try to vectorize them | |||
14695 | // after processing BB. | |||
14696 | if (BB == P->getIncomingBlock(I) || | |||
14697 | !DT->isReachableFromEntry(P->getIncomingBlock(I))) | |||
14698 | continue; | |||
14699 | ||||
14700 | // Postponed instructions should not be vectorized here, delay their | |||
14701 | // vectorization. | |||
14702 | if (auto *PI = dyn_cast<Instruction>(P->getIncomingValue(I)); | |||
14703 | PI && !PostProcessInstructions.contains(PI)) | |||
14704 | Changed |= vectorizeRootInstruction(nullptr, PI, | |||
14705 | P->getIncomingBlock(I), R, TTI); | |||
14706 | } | |||
14707 | continue; | |||
14708 | } | |||
14709 | ||||
14710 | // Ran into an instruction without users, like terminator, or function call | |||
14711 | // with ignored return value, store. Ignore unused instructions (basing on | |||
14712 | // instruction type, except for CallInst and InvokeInst). | |||
14713 | if (it->use_empty() && | |||
14714 | (it->getType()->isVoidTy() || isa<CallInst, InvokeInst>(it))) { | |||
14715 | KeyNodes.insert(&*it); | |||
14716 | bool OpsChanged = false; | |||
14717 | auto *SI = dyn_cast<StoreInst>(it); | |||
14718 | bool TryToVectorizeRoot = ShouldStartVectorizeHorAtStore || !SI; | |||
14719 | if (SI) { | |||
14720 | auto I = Stores.find(getUnderlyingObject(SI->getPointerOperand())); | |||
14721 | // Try to vectorize chain in store, if this is the only store to the | |||
14722 | // address in the block. | |||
14723 | // TODO: This is just a temporarily solution to save compile time. Need | |||
14724 | // to investigate if we can safely turn on slp-vectorize-hor-store | |||
14725 | // instead to allow lookup for reduction chains in all non-vectorized | |||
14726 | // stores (need to check side effects and compile time). | |||
14727 | TryToVectorizeRoot = (I == Stores.end() || I->second.size() == 1) && | |||
14728 | SI->getValueOperand()->hasOneUse(); | |||
14729 | } | |||
14730 | if (TryToVectorizeRoot) { | |||
14731 | for (auto *V : it->operand_values()) { | |||
14732 | // Postponed instructions should not be vectorized here, delay their | |||
14733 | // vectorization. | |||
14734 | if (auto *VI = dyn_cast<Instruction>(V); | |||
14735 | VI && !PostProcessInstructions.contains(VI)) | |||
14736 | // Try to match and vectorize a horizontal reduction. | |||
14737 | OpsChanged |= vectorizeRootInstruction(nullptr, VI, BB, R, TTI); | |||
14738 | } | |||
14739 | } | |||
14740 | // Start vectorization of post-process list of instructions from the | |||
14741 | // top-tree instructions to try to vectorize as many instructions as | |||
14742 | // possible. | |||
14743 | OpsChanged |= vectorizeSimpleInstructions(PostProcessInstructions, BB, R, | |||
14744 | it->isTerminator()); | |||
14745 | if (OpsChanged) { | |||
14746 | // We would like to start over since some instructions are deleted | |||
14747 | // and the iterator may become invalid value. | |||
14748 | Changed = true; | |||
14749 | it = BB->begin(); | |||
14750 | e = BB->end(); | |||
14751 | continue; | |||
14752 | } | |||
14753 | } | |||
14754 | ||||
14755 | if (isa<CmpInst, InsertElementInst, InsertValueInst>(it)) | |||
14756 | PostProcessInstructions.insert(&*it); | |||
14757 | } | |||
14758 | ||||
14759 | return Changed; | |||
14760 | } | |||
14761 | ||||
14762 | bool SLPVectorizerPass::vectorizeGEPIndices(BasicBlock *BB, BoUpSLP &R) { | |||
14763 | auto Changed = false; | |||
14764 | for (auto &Entry : GEPs) { | |||
14765 | // If the getelementptr list has fewer than two elements, there's nothing | |||
14766 | // to do. | |||
14767 | if (Entry.second.size() < 2) | |||
14768 | continue; | |||
14769 | ||||
14770 | 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 ) | |||
14771 | << 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 ); | |||
14772 | ||||
14773 | // Process the GEP list in chunks suitable for the target's supported | |||
14774 | // vector size. If a vector register can't hold 1 element, we are done. We | |||
14775 | // are trying to vectorize the index computations, so the maximum number of | |||
14776 | // elements is based on the size of the index expression, rather than the | |||
14777 | // size of the GEP itself (the target's pointer size). | |||
14778 | unsigned MaxVecRegSize = R.getMaxVecRegSize(); | |||
14779 | unsigned EltSize = R.getVectorElementSize(*Entry.second[0]->idx_begin()); | |||
14780 | if (MaxVecRegSize < EltSize) | |||
14781 | continue; | |||
14782 | ||||
14783 | unsigned MaxElts = MaxVecRegSize / EltSize; | |||
14784 | for (unsigned BI = 0, BE = Entry.second.size(); BI < BE; BI += MaxElts) { | |||
14785 | auto Len = std::min<unsigned>(BE - BI, MaxElts); | |||
14786 | ArrayRef<GetElementPtrInst *> GEPList(&Entry.second[BI], Len); | |||
14787 | ||||
14788 | // Initialize a set a candidate getelementptrs. Note that we use a | |||
14789 | // SetVector here to preserve program order. If the index computations | |||
14790 | // are vectorizable and begin with loads, we want to minimize the chance | |||
14791 | // of having to reorder them later. | |||
14792 | SetVector<Value *> Candidates(GEPList.begin(), GEPList.end()); | |||
14793 | ||||
14794 | // Some of the candidates may have already been vectorized after we | |||
14795 | // initially collected them. If so, they are marked as deleted, so remove | |||
14796 | // them from the set of candidates. | |||
14797 | Candidates.remove_if( | |||
14798 | [&R](Value *I) { return R.isDeleted(cast<Instruction>(I)); }); | |||
14799 | ||||
14800 | // Remove from the set of candidates all pairs of getelementptrs with | |||
14801 | // constant differences. Such getelementptrs are likely not good | |||
14802 | // candidates for vectorization in a bottom-up phase since one can be | |||
14803 | // computed from the other. We also ensure all candidate getelementptr | |||
14804 | // indices are unique. | |||
14805 | for (int I = 0, E = GEPList.size(); I < E && Candidates.size() > 1; ++I) { | |||
14806 | auto *GEPI = GEPList[I]; | |||
14807 | if (!Candidates.count(GEPI)) | |||
14808 | continue; | |||
14809 | auto *SCEVI = SE->getSCEV(GEPList[I]); | |||
14810 | for (int J = I + 1; J < E && Candidates.size() > 1; ++J) { | |||
14811 | auto *GEPJ = GEPList[J]; | |||
14812 | auto *SCEVJ = SE->getSCEV(GEPList[J]); | |||
14813 | if (isa<SCEVConstant>(SE->getMinusSCEV(SCEVI, SCEVJ))) { | |||
14814 | Candidates.remove(GEPI); | |||
14815 | Candidates.remove(GEPJ); | |||
14816 | } else if (GEPI->idx_begin()->get() == GEPJ->idx_begin()->get()) { | |||
14817 | Candidates.remove(GEPJ); | |||
14818 | } | |||
14819 | } | |||
14820 | } | |||
14821 | ||||
14822 | // We break out of the above computation as soon as we know there are | |||
14823 | // fewer than two candidates remaining. | |||
14824 | if (Candidates.size() < 2) | |||
14825 | continue; | |||
14826 | ||||
14827 | // Add the single, non-constant index of each candidate to the bundle. We | |||
14828 | // ensured the indices met these constraints when we originally collected | |||
14829 | // the getelementptrs. | |||
14830 | SmallVector<Value *, 16> Bundle(Candidates.size()); | |||
14831 | auto BundleIndex = 0u; | |||
14832 | for (auto *V : Candidates) { | |||
14833 | auto *GEP = cast<GetElementPtrInst>(V); | |||
14834 | auto *GEPIdx = GEP->idx_begin()->get(); | |||
14835 | 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", 14835, __extension__ __PRETTY_FUNCTION__)); | |||
14836 | Bundle[BundleIndex++] = GEPIdx; | |||
14837 | } | |||
14838 | ||||
14839 | // Try and vectorize the indices. We are currently only interested in | |||
14840 | // gather-like cases of the form: | |||
14841 | // | |||
14842 | // ... = g[a[0] - b[0]] + g[a[1] - b[1]] + ... | |||
14843 | // | |||
14844 | // where the loads of "a", the loads of "b", and the subtractions can be | |||
14845 | // performed in parallel. It's likely that detecting this pattern in a | |||
14846 | // bottom-up phase will be simpler and less costly than building a | |||
14847 | // full-blown top-down phase beginning at the consecutive loads. | |||
14848 | Changed |= tryToVectorizeList(Bundle, R); | |||
14849 | } | |||
14850 | } | |||
14851 | return Changed; | |||
14852 | } | |||
14853 | ||||
14854 | bool SLPVectorizerPass::vectorizeStoreChains(BoUpSLP &R) { | |||
14855 | bool Changed = false; | |||
14856 | // Sort by type, base pointers and values operand. Value operands must be | |||
14857 | // compatible (have the same opcode, same parent), otherwise it is | |||
14858 | // definitely not profitable to try to vectorize them. | |||
14859 | auto &&StoreSorter = [this](StoreInst *V, StoreInst *V2) { | |||
14860 | if (V->getPointerOperandType()->getTypeID() < | |||
14861 | V2->getPointerOperandType()->getTypeID()) | |||
14862 | return true; | |||
14863 | if (V->getPointerOperandType()->getTypeID() > | |||
14864 | V2->getPointerOperandType()->getTypeID()) | |||
14865 | return false; | |||
14866 | // UndefValues are compatible with all other values. | |||
14867 | if (isa<UndefValue>(V->getValueOperand()) || | |||
14868 | isa<UndefValue>(V2->getValueOperand())) | |||
14869 | return false; | |||
14870 | if (auto *I1 = dyn_cast<Instruction>(V->getValueOperand())) | |||
14871 | if (auto *I2 = dyn_cast<Instruction>(V2->getValueOperand())) { | |||
14872 | DomTreeNodeBase<llvm::BasicBlock> *NodeI1 = | |||
14873 | DT->getNode(I1->getParent()); | |||
14874 | DomTreeNodeBase<llvm::BasicBlock> *NodeI2 = | |||
14875 | DT->getNode(I2->getParent()); | |||
14876 | 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", 14876, __extension__ __PRETTY_FUNCTION__)); | |||
14877 | assert(NodeI2 && "Should only process reachable instructions")(static_cast <bool> (NodeI2 && "Should only process reachable instructions" ) ? void (0) : __assert_fail ("NodeI2 && \"Should only process reachable instructions\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 14877, __extension__ __PRETTY_FUNCTION__)); | |||
14878 | 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", 14880, __extension__ __PRETTY_FUNCTION__)) | |||
14879 | (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", 14880, __extension__ __PRETTY_FUNCTION__)) | |||
14880 | "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", 14880, __extension__ __PRETTY_FUNCTION__)); | |||
14881 | if (NodeI1 != NodeI2) | |||
14882 | return NodeI1->getDFSNumIn() < NodeI2->getDFSNumIn(); | |||
14883 | InstructionsState S = getSameOpcode({I1, I2}, *TLI); | |||
14884 | if (S.getOpcode()) | |||
14885 | return false; | |||
14886 | return I1->getOpcode() < I2->getOpcode(); | |||
14887 | } | |||
14888 | if (isa<Constant>(V->getValueOperand()) && | |||
14889 | isa<Constant>(V2->getValueOperand())) | |||
14890 | return false; | |||
14891 | return V->getValueOperand()->getValueID() < | |||
14892 | V2->getValueOperand()->getValueID(); | |||
14893 | }; | |||
14894 | ||||
14895 | auto &&AreCompatibleStores = [this](StoreInst *V1, StoreInst *V2) { | |||
14896 | if (V1 == V2) | |||
14897 | return true; | |||
14898 | if (V1->getPointerOperandType() != V2->getPointerOperandType()) | |||
14899 | return false; | |||
14900 | // Undefs are compatible with any other value. | |||
14901 | if (isa<UndefValue>(V1->getValueOperand()) || | |||
14902 | isa<UndefValue>(V2->getValueOperand())) | |||
14903 | return true; | |||
14904 | if (auto *I1 = dyn_cast<Instruction>(V1->getValueOperand())) | |||
14905 | if (auto *I2 = dyn_cast<Instruction>(V2->getValueOperand())) { | |||
14906 | if (I1->getParent() != I2->getParent()) | |||
14907 | return false; | |||
14908 | InstructionsState S = getSameOpcode({I1, I2}, *TLI); | |||
14909 | return S.getOpcode() > 0; | |||
14910 | } | |||
14911 | if (isa<Constant>(V1->getValueOperand()) && | |||
14912 | isa<Constant>(V2->getValueOperand())) | |||
14913 | return true; | |||
14914 | return V1->getValueOperand()->getValueID() == | |||
14915 | V2->getValueOperand()->getValueID(); | |||
14916 | }; | |||
14917 | ||||
14918 | // Attempt to sort and vectorize each of the store-groups. | |||
14919 | for (auto &Pair : Stores) { | |||
14920 | if (Pair.second.size() < 2) | |||
14921 | continue; | |||
14922 | ||||
14923 | 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 ) | |||
14924 | << 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 ); | |||
14925 | ||||
14926 | if (!isValidElementType(Pair.second.front()->getValueOperand()->getType())) | |||
14927 | continue; | |||
14928 | ||||
14929 | Changed |= tryToVectorizeSequence<StoreInst>( | |||
14930 | Pair.second, StoreSorter, AreCompatibleStores, | |||
14931 | [this, &R](ArrayRef<StoreInst *> Candidates, bool) { | |||
14932 | return vectorizeStores(Candidates, R); | |||
14933 | }, | |||
14934 | /*LimitForRegisterSize=*/false, R); | |||
14935 | } | |||
14936 | return Changed; | |||
14937 | } |