File: | build/llvm-toolchain-snapshot-16~++20221003111214+1fa2019828ca/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp |
Warning: | line 11594, column 9 Value stored to 'VectorizedTree' is never read |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | //===- SLPVectorizer.cpp - A bottom up SLP Vectorizer ---------------------===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | // |
9 | // This pass implements the Bottom Up SLP vectorizer. It detects consecutive |
10 | // stores that can be put together into vector-stores. Next, it attempts to |
11 | // construct vectorizable tree using the use-def chains. If a profitable tree |
12 | // was found, the SLP vectorizer performs vectorization on the tree. |
13 | // |
14 | // The pass is inspired by the work described in the paper: |
15 | // "Loop-Aware SLP in GCC" by Ira Rosen, Dorit Nuzman, Ayal Zaks. |
16 | // |
17 | //===----------------------------------------------------------------------===// |
18 | |
19 | #include "llvm/Transforms/Vectorize/SLPVectorizer.h" |
20 | #include "llvm/ADT/DenseMap.h" |
21 | #include "llvm/ADT/DenseSet.h" |
22 | #include "llvm/ADT/Optional.h" |
23 | #include "llvm/ADT/PostOrderIterator.h" |
24 | #include "llvm/ADT/PriorityQueue.h" |
25 | #include "llvm/ADT/STLExtras.h" |
26 | #include "llvm/ADT/SetOperations.h" |
27 | #include "llvm/ADT/SetVector.h" |
28 | #include "llvm/ADT/SmallBitVector.h" |
29 | #include "llvm/ADT/SmallPtrSet.h" |
30 | #include "llvm/ADT/SmallSet.h" |
31 | #include "llvm/ADT/SmallString.h" |
32 | #include "llvm/ADT/Statistic.h" |
33 | #include "llvm/ADT/iterator.h" |
34 | #include "llvm/ADT/iterator_range.h" |
35 | #include "llvm/Analysis/AliasAnalysis.h" |
36 | #include "llvm/Analysis/AssumptionCache.h" |
37 | #include "llvm/Analysis/CodeMetrics.h" |
38 | #include "llvm/Analysis/DemandedBits.h" |
39 | #include "llvm/Analysis/GlobalsModRef.h" |
40 | #include "llvm/Analysis/IVDescriptors.h" |
41 | #include "llvm/Analysis/LoopAccessAnalysis.h" |
42 | #include "llvm/Analysis/LoopInfo.h" |
43 | #include "llvm/Analysis/MemoryLocation.h" |
44 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" |
45 | #include "llvm/Analysis/ScalarEvolution.h" |
46 | #include "llvm/Analysis/ScalarEvolutionExpressions.h" |
47 | #include "llvm/Analysis/TargetLibraryInfo.h" |
48 | #include "llvm/Analysis/TargetTransformInfo.h" |
49 | #include "llvm/Analysis/ValueTracking.h" |
50 | #include "llvm/Analysis/VectorUtils.h" |
51 | #include "llvm/IR/Attributes.h" |
52 | #include "llvm/IR/BasicBlock.h" |
53 | #include "llvm/IR/Constant.h" |
54 | #include "llvm/IR/Constants.h" |
55 | #include "llvm/IR/DataLayout.h" |
56 | #include "llvm/IR/DerivedTypes.h" |
57 | #include "llvm/IR/Dominators.h" |
58 | #include "llvm/IR/Function.h" |
59 | #include "llvm/IR/IRBuilder.h" |
60 | #include "llvm/IR/InstrTypes.h" |
61 | #include "llvm/IR/Instruction.h" |
62 | #include "llvm/IR/Instructions.h" |
63 | #include "llvm/IR/IntrinsicInst.h" |
64 | #include "llvm/IR/Intrinsics.h" |
65 | #include "llvm/IR/Module.h" |
66 | #include "llvm/IR/Operator.h" |
67 | #include "llvm/IR/PatternMatch.h" |
68 | #include "llvm/IR/Type.h" |
69 | #include "llvm/IR/Use.h" |
70 | #include "llvm/IR/User.h" |
71 | #include "llvm/IR/Value.h" |
72 | #include "llvm/IR/ValueHandle.h" |
73 | #ifdef EXPENSIVE_CHECKS |
74 | #include "llvm/IR/Verifier.h" |
75 | #endif |
76 | #include "llvm/Pass.h" |
77 | #include "llvm/Support/Casting.h" |
78 | #include "llvm/Support/CommandLine.h" |
79 | #include "llvm/Support/Compiler.h" |
80 | #include "llvm/Support/DOTGraphTraits.h" |
81 | #include "llvm/Support/Debug.h" |
82 | #include "llvm/Support/ErrorHandling.h" |
83 | #include "llvm/Support/GraphWriter.h" |
84 | #include "llvm/Support/InstructionCost.h" |
85 | #include "llvm/Support/KnownBits.h" |
86 | #include "llvm/Support/MathExtras.h" |
87 | #include "llvm/Support/raw_ostream.h" |
88 | #include "llvm/Transforms/Utils/InjectTLIMappings.h" |
89 | #include "llvm/Transforms/Utils/Local.h" |
90 | #include "llvm/Transforms/Utils/LoopUtils.h" |
91 | #include "llvm/Transforms/Vectorize.h" |
92 | #include <algorithm> |
93 | #include <cassert> |
94 | #include <cstdint> |
95 | #include <iterator> |
96 | #include <memory> |
97 | #include <set> |
98 | #include <string> |
99 | #include <tuple> |
100 | #include <utility> |
101 | #include <vector> |
102 | |
103 | using namespace llvm; |
104 | using namespace llvm::PatternMatch; |
105 | using namespace slpvectorizer; |
106 | |
107 | #define SV_NAME"slp-vectorizer" "slp-vectorizer" |
108 | #define DEBUG_TYPE"SLP" "SLP" |
109 | |
110 | STATISTIC(NumVectorInstructions, "Number of vector instructions generated")static llvm::Statistic NumVectorInstructions = {"SLP", "NumVectorInstructions" , "Number of vector instructions generated"}; |
111 | |
112 | cl::opt<bool> RunSLPVectorization("vectorize-slp", cl::init(true), cl::Hidden, |
113 | cl::desc("Run the SLP vectorization passes")); |
114 | |
115 | static cl::opt<int> |
116 | SLPCostThreshold("slp-threshold", cl::init(0), cl::Hidden, |
117 | cl::desc("Only vectorize if you gain more than this " |
118 | "number ")); |
119 | |
120 | static cl::opt<bool> |
121 | ShouldVectorizeHor("slp-vectorize-hor", cl::init(true), cl::Hidden, |
122 | cl::desc("Attempt to vectorize horizontal reductions")); |
123 | |
124 | static cl::opt<bool> ShouldStartVectorizeHorAtStore( |
125 | "slp-vectorize-hor-store", cl::init(false), cl::Hidden, |
126 | cl::desc( |
127 | "Attempt to vectorize horizontal reductions feeding into a store")); |
128 | |
129 | static cl::opt<int> |
130 | MaxVectorRegSizeOption("slp-max-reg-size", cl::init(128), cl::Hidden, |
131 | cl::desc("Attempt to vectorize for this register size in bits")); |
132 | |
133 | static cl::opt<unsigned> |
134 | MaxVFOption("slp-max-vf", cl::init(0), cl::Hidden, |
135 | cl::desc("Maximum SLP vectorization factor (0=unlimited)")); |
136 | |
137 | static cl::opt<int> |
138 | MaxStoreLookup("slp-max-store-lookup", cl::init(32), cl::Hidden, |
139 | cl::desc("Maximum depth of the lookup for consecutive stores.")); |
140 | |
141 | /// Limits the size of scheduling regions in a block. |
142 | /// It avoid long compile times for _very_ large blocks where vector |
143 | /// instructions are spread over a wide range. |
144 | /// This limit is way higher than needed by real-world functions. |
145 | static cl::opt<int> |
146 | ScheduleRegionSizeBudget("slp-schedule-budget", cl::init(100000), cl::Hidden, |
147 | cl::desc("Limit the size of the SLP scheduling region per block")); |
148 | |
149 | static cl::opt<int> MinVectorRegSizeOption( |
150 | "slp-min-reg-size", cl::init(128), cl::Hidden, |
151 | cl::desc("Attempt to vectorize for this register size in bits")); |
152 | |
153 | static cl::opt<unsigned> RecursionMaxDepth( |
154 | "slp-recursion-max-depth", cl::init(12), cl::Hidden, |
155 | cl::desc("Limit the recursion depth when building a vectorizable tree")); |
156 | |
157 | static cl::opt<unsigned> MinTreeSize( |
158 | "slp-min-tree-size", cl::init(3), cl::Hidden, |
159 | cl::desc("Only vectorize small trees if they are fully vectorizable")); |
160 | |
161 | // The maximum depth that the look-ahead score heuristic will explore. |
162 | // The higher this value, the higher the compilation time overhead. |
163 | static cl::opt<int> LookAheadMaxDepth( |
164 | "slp-max-look-ahead-depth", cl::init(2), cl::Hidden, |
165 | cl::desc("The maximum look-ahead depth for operand reordering scores")); |
166 | |
167 | // The maximum depth that the look-ahead score heuristic will explore |
168 | // when it probing among candidates for vectorization tree roots. |
169 | // The higher this value, the higher the compilation time overhead but unlike |
170 | // similar limit for operands ordering this is less frequently used, hence |
171 | // impact of higher value is less noticeable. |
172 | static cl::opt<int> RootLookAheadMaxDepth( |
173 | "slp-max-root-look-ahead-depth", cl::init(2), cl::Hidden, |
174 | cl::desc("The maximum look-ahead depth for searching best rooting option")); |
175 | |
176 | static cl::opt<bool> |
177 | ViewSLPTree("view-slp-tree", cl::Hidden, |
178 | cl::desc("Display the SLP trees with Graphviz")); |
179 | |
180 | // Limit the number of alias checks. The limit is chosen so that |
181 | // it has no negative effect on the llvm benchmarks. |
182 | static const unsigned AliasedCheckLimit = 10; |
183 | |
184 | // Another limit for the alias checks: The maximum distance between load/store |
185 | // instructions where alias checks are done. |
186 | // This limit is useful for very large basic blocks. |
187 | static const unsigned MaxMemDepDistance = 160; |
188 | |
189 | /// If the ScheduleRegionSizeBudget is exhausted, we allow small scheduling |
190 | /// regions to be handled. |
191 | static const int MinScheduleRegionSize = 16; |
192 | |
193 | /// Predicate for the element types that the SLP vectorizer supports. |
194 | /// |
195 | /// The most important thing to filter here are types which are invalid in LLVM |
196 | /// vectors. We also filter target specific types which have absolutely no |
197 | /// meaningful vectorization path such as x86_fp80 and ppc_f128. This just |
198 | /// avoids spending time checking the cost model and realizing that they will |
199 | /// be inevitably scalarized. |
200 | static bool isValidElementType(Type *Ty) { |
201 | return VectorType::isValidElementType(Ty) && !Ty->isX86_FP80Ty() && |
202 | !Ty->isPPC_FP128Ty(); |
203 | } |
204 | |
205 | /// \returns True if the value is a constant (but not globals/constant |
206 | /// expressions). |
207 | static bool isConstant(Value *V) { |
208 | return isa<Constant>(V) && !isa<ConstantExpr, GlobalValue>(V); |
209 | } |
210 | |
211 | /// Checks if \p V is one of vector-like instructions, i.e. undef, |
212 | /// insertelement/extractelement with constant indices for fixed vector type or |
213 | /// extractvalue instruction. |
214 | static bool isVectorLikeInstWithConstOps(Value *V) { |
215 | if (!isa<InsertElementInst, ExtractElementInst>(V) && |
216 | !isa<ExtractValueInst, UndefValue>(V)) |
217 | return false; |
218 | auto *I = dyn_cast<Instruction>(V); |
219 | if (!I || isa<ExtractValueInst>(I)) |
220 | return true; |
221 | if (!isa<FixedVectorType>(I->getOperand(0)->getType())) |
222 | return false; |
223 | if (isa<ExtractElementInst>(I)) |
224 | return isConstant(I->getOperand(1)); |
225 | 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", 225, __extension__ __PRETTY_FUNCTION__)); |
226 | return isConstant(I->getOperand(2)); |
227 | } |
228 | |
229 | /// \returns true if all of the instructions in \p VL are in the same block or |
230 | /// false otherwise. |
231 | static bool allSameBlock(ArrayRef<Value *> VL) { |
232 | Instruction *I0 = dyn_cast<Instruction>(VL[0]); |
233 | if (!I0) |
234 | return false; |
235 | if (all_of(VL, isVectorLikeInstWithConstOps)) |
236 | return true; |
237 | |
238 | BasicBlock *BB = I0->getParent(); |
239 | for (int I = 1, E = VL.size(); I < E; I++) { |
240 | auto *II = dyn_cast<Instruction>(VL[I]); |
241 | if (!II) |
242 | return false; |
243 | |
244 | if (BB != II->getParent()) |
245 | return false; |
246 | } |
247 | return true; |
248 | } |
249 | |
250 | /// \returns True if all of the values in \p VL are constants (but not |
251 | /// globals/constant expressions). |
252 | static bool allConstant(ArrayRef<Value *> VL) { |
253 | // Constant expressions and globals can't be vectorized like normal integer/FP |
254 | // constants. |
255 | return all_of(VL, isConstant); |
256 | } |
257 | |
258 | /// \returns True if all of the values in \p VL are identical or some of them |
259 | /// are UndefValue. |
260 | static bool isSplat(ArrayRef<Value *> VL) { |
261 | Value *FirstNonUndef = nullptr; |
262 | for (Value *V : VL) { |
263 | if (isa<UndefValue>(V)) |
264 | continue; |
265 | if (!FirstNonUndef) { |
266 | FirstNonUndef = V; |
267 | continue; |
268 | } |
269 | if (V != FirstNonUndef) |
270 | return false; |
271 | } |
272 | return FirstNonUndef != nullptr; |
273 | } |
274 | |
275 | /// \returns True if \p I is commutative, handles CmpInst and BinaryOperator. |
276 | static bool isCommutative(Instruction *I) { |
277 | if (auto *Cmp = dyn_cast<CmpInst>(I)) |
278 | return Cmp->isCommutative(); |
279 | if (auto *BO = dyn_cast<BinaryOperator>(I)) |
280 | return BO->isCommutative(); |
281 | // TODO: This should check for generic Instruction::isCommutative(), but |
282 | // we need to confirm that the caller code correctly handles Intrinsics |
283 | // for example (does not have 2 operands). |
284 | return false; |
285 | } |
286 | |
287 | /// \returns inserting index of InsertElement or InsertValue instruction, |
288 | /// using Offset as base offset for index. |
289 | static Optional<unsigned> getInsertIndex(const Value *InsertInst, |
290 | unsigned Offset = 0) { |
291 | int Index = Offset; |
292 | if (const auto *IE = dyn_cast<InsertElementInst>(InsertInst)) { |
293 | if (const auto *CI = dyn_cast<ConstantInt>(IE->getOperand(2))) { |
294 | auto *VT = cast<FixedVectorType>(IE->getType()); |
295 | if (CI->getValue().uge(VT->getNumElements())) |
296 | return None; |
297 | Index *= VT->getNumElements(); |
298 | Index += CI->getZExtValue(); |
299 | return Index; |
300 | } |
301 | return None; |
302 | } |
303 | |
304 | const auto *IV = cast<InsertValueInst>(InsertInst); |
305 | Type *CurrentType = IV->getType(); |
306 | for (unsigned I : IV->indices()) { |
307 | if (const auto *ST = dyn_cast<StructType>(CurrentType)) { |
308 | Index *= ST->getNumElements(); |
309 | CurrentType = ST->getElementType(I); |
310 | } else if (const auto *AT = dyn_cast<ArrayType>(CurrentType)) { |
311 | Index *= AT->getNumElements(); |
312 | CurrentType = AT->getElementType(); |
313 | } else { |
314 | return None; |
315 | } |
316 | Index += I; |
317 | } |
318 | return Index; |
319 | } |
320 | |
321 | /// Checks if the given value is actually an undefined constant vector. |
322 | /// Also, if the\p ShuffleMask is not empty, tries to check if the non-masked |
323 | /// elements actually mask the insertelement buildvector, if any. |
324 | static bool isUndefVector(const Value *V, ArrayRef<int> ShuffleMask = None) { |
325 | if (isa<UndefValue>(V)) |
326 | return true; |
327 | auto *VecTy = dyn_cast<FixedVectorType>(V->getType()); |
328 | if (!VecTy) |
329 | return false; |
330 | auto *C = dyn_cast<Constant>(V); |
331 | if (!C) { |
332 | if (!ShuffleMask.empty()) { |
333 | const Value *Base = V; |
334 | while (auto *II = dyn_cast<InsertElementInst>(Base)) { |
335 | Base = II->getOperand(0); |
336 | Optional<unsigned> Idx = getInsertIndex(II); |
337 | if (!Idx) |
338 | continue; |
339 | if (*Idx < ShuffleMask.size() && ShuffleMask[*Idx] == UndefMaskElem) |
340 | return false; |
341 | } |
342 | return V != Base && isUndefVector(Base); |
343 | } |
344 | return false; |
345 | } |
346 | for (unsigned I = 0, E = VecTy->getNumElements(); I != E; ++I) { |
347 | if (Constant *Elem = C->getAggregateElement(I)) |
348 | if (!isa<UndefValue>(Elem) && |
349 | (ShuffleMask.empty() || |
350 | (I < ShuffleMask.size() && ShuffleMask[I] == UndefMaskElem))) |
351 | return false; |
352 | } |
353 | return true; |
354 | } |
355 | |
356 | /// Checks if the vector of instructions can be represented as a shuffle, like: |
357 | /// %x0 = extractelement <4 x i8> %x, i32 0 |
358 | /// %x3 = extractelement <4 x i8> %x, i32 3 |
359 | /// %y1 = extractelement <4 x i8> %y, i32 1 |
360 | /// %y2 = extractelement <4 x i8> %y, i32 2 |
361 | /// %x0x0 = mul i8 %x0, %x0 |
362 | /// %x3x3 = mul i8 %x3, %x3 |
363 | /// %y1y1 = mul i8 %y1, %y1 |
364 | /// %y2y2 = mul i8 %y2, %y2 |
365 | /// %ins1 = insertelement <4 x i8> poison, i8 %x0x0, i32 0 |
366 | /// %ins2 = insertelement <4 x i8> %ins1, i8 %x3x3, i32 1 |
367 | /// %ins3 = insertelement <4 x i8> %ins2, i8 %y1y1, i32 2 |
368 | /// %ins4 = insertelement <4 x i8> %ins3, i8 %y2y2, i32 3 |
369 | /// ret <4 x i8> %ins4 |
370 | /// can be transformed into: |
371 | /// %1 = shufflevector <4 x i8> %x, <4 x i8> %y, <4 x i32> <i32 0, i32 3, i32 5, |
372 | /// i32 6> |
373 | /// %2 = mul <4 x i8> %1, %1 |
374 | /// ret <4 x i8> %2 |
375 | /// We convert this initially to something like: |
376 | /// %x0 = extractelement <4 x i8> %x, i32 0 |
377 | /// %x3 = extractelement <4 x i8> %x, i32 3 |
378 | /// %y1 = extractelement <4 x i8> %y, i32 1 |
379 | /// %y2 = extractelement <4 x i8> %y, i32 2 |
380 | /// %1 = insertelement <4 x i8> poison, i8 %x0, i32 0 |
381 | /// %2 = insertelement <4 x i8> %1, i8 %x3, i32 1 |
382 | /// %3 = insertelement <4 x i8> %2, i8 %y1, i32 2 |
383 | /// %4 = insertelement <4 x i8> %3, i8 %y2, i32 3 |
384 | /// %5 = mul <4 x i8> %4, %4 |
385 | /// %6 = extractelement <4 x i8> %5, i32 0 |
386 | /// %ins1 = insertelement <4 x i8> poison, i8 %6, i32 0 |
387 | /// %7 = extractelement <4 x i8> %5, i32 1 |
388 | /// %ins2 = insertelement <4 x i8> %ins1, i8 %7, i32 1 |
389 | /// %8 = extractelement <4 x i8> %5, i32 2 |
390 | /// %ins3 = insertelement <4 x i8> %ins2, i8 %8, i32 2 |
391 | /// %9 = extractelement <4 x i8> %5, i32 3 |
392 | /// %ins4 = insertelement <4 x i8> %ins3, i8 %9, i32 3 |
393 | /// ret <4 x i8> %ins4 |
394 | /// InstCombiner transforms this into a shuffle and vector mul |
395 | /// Mask will return the Shuffle Mask equivalent to the extracted elements. |
396 | /// TODO: Can we split off and reuse the shuffle mask detection from |
397 | /// ShuffleVectorInst/getShuffleCost? |
398 | static Optional<TargetTransformInfo::ShuffleKind> |
399 | isFixedVectorShuffle(ArrayRef<Value *> VL, SmallVectorImpl<int> &Mask) { |
400 | const auto *It = |
401 | find_if(VL, [](Value *V) { return isa<ExtractElementInst>(V); }); |
402 | if (It == VL.end()) |
403 | return None; |
404 | auto *EI0 = cast<ExtractElementInst>(*It); |
405 | if (isa<ScalableVectorType>(EI0->getVectorOperandType())) |
406 | return None; |
407 | unsigned Size = |
408 | cast<FixedVectorType>(EI0->getVectorOperandType())->getNumElements(); |
409 | Value *Vec1 = nullptr; |
410 | Value *Vec2 = nullptr; |
411 | enum ShuffleMode { Unknown, Select, Permute }; |
412 | ShuffleMode CommonShuffleMode = Unknown; |
413 | Mask.assign(VL.size(), UndefMaskElem); |
414 | for (unsigned I = 0, E = VL.size(); I < E; ++I) { |
415 | // Undef can be represented as an undef element in a vector. |
416 | if (isa<UndefValue>(VL[I])) |
417 | continue; |
418 | auto *EI = cast<ExtractElementInst>(VL[I]); |
419 | if (isa<ScalableVectorType>(EI->getVectorOperandType())) |
420 | return None; |
421 | auto *Vec = EI->getVectorOperand(); |
422 | // We can extractelement from undef or poison vector. |
423 | if (isUndefVector(Vec)) |
424 | continue; |
425 | // All vector operands must have the same number of vector elements. |
426 | if (cast<FixedVectorType>(Vec->getType())->getNumElements() != Size) |
427 | return None; |
428 | if (isa<UndefValue>(EI->getIndexOperand())) |
429 | continue; |
430 | auto *Idx = dyn_cast<ConstantInt>(EI->getIndexOperand()); |
431 | if (!Idx) |
432 | return None; |
433 | // Undefined behavior if Idx is negative or >= Size. |
434 | if (Idx->getValue().uge(Size)) |
435 | continue; |
436 | unsigned IntIdx = Idx->getValue().getZExtValue(); |
437 | Mask[I] = IntIdx; |
438 | // For correct shuffling we have to have at most 2 different vector operands |
439 | // in all extractelement instructions. |
440 | if (!Vec1 || Vec1 == Vec) { |
441 | Vec1 = Vec; |
442 | } else if (!Vec2 || Vec2 == Vec) { |
443 | Vec2 = Vec; |
444 | Mask[I] += Size; |
445 | } else { |
446 | return None; |
447 | } |
448 | if (CommonShuffleMode == Permute) |
449 | continue; |
450 | // If the extract index is not the same as the operation number, it is a |
451 | // permutation. |
452 | if (IntIdx != I) { |
453 | CommonShuffleMode = Permute; |
454 | continue; |
455 | } |
456 | CommonShuffleMode = Select; |
457 | } |
458 | // If we're not crossing lanes in different vectors, consider it as blending. |
459 | if (CommonShuffleMode == Select && Vec2) |
460 | return TargetTransformInfo::SK_Select; |
461 | // If Vec2 was never used, we have a permutation of a single vector, otherwise |
462 | // we have permutation of 2 vectors. |
463 | return Vec2 ? TargetTransformInfo::SK_PermuteTwoSrc |
464 | : TargetTransformInfo::SK_PermuteSingleSrc; |
465 | } |
466 | |
467 | namespace { |
468 | |
469 | /// Main data required for vectorization of instructions. |
470 | struct InstructionsState { |
471 | /// The very first instruction in the list with the main opcode. |
472 | Value *OpValue = nullptr; |
473 | |
474 | /// The main/alternate instruction. |
475 | Instruction *MainOp = nullptr; |
476 | Instruction *AltOp = nullptr; |
477 | |
478 | /// The main/alternate opcodes for the list of instructions. |
479 | unsigned getOpcode() const { |
480 | return MainOp ? MainOp->getOpcode() : 0; |
481 | } |
482 | |
483 | unsigned getAltOpcode() const { |
484 | return AltOp ? AltOp->getOpcode() : 0; |
485 | } |
486 | |
487 | /// Some of the instructions in the list have alternate opcodes. |
488 | bool isAltShuffle() const { return AltOp != MainOp; } |
489 | |
490 | bool isOpcodeOrAlt(Instruction *I) const { |
491 | unsigned CheckedOpcode = I->getOpcode(); |
492 | return getOpcode() == CheckedOpcode || getAltOpcode() == CheckedOpcode; |
493 | } |
494 | |
495 | InstructionsState() = delete; |
496 | InstructionsState(Value *OpValue, Instruction *MainOp, Instruction *AltOp) |
497 | : OpValue(OpValue), MainOp(MainOp), AltOp(AltOp) {} |
498 | }; |
499 | |
500 | } // end anonymous namespace |
501 | |
502 | /// Chooses the correct key for scheduling data. If \p Op has the same (or |
503 | /// alternate) opcode as \p OpValue, the key is \p Op. Otherwise the key is \p |
504 | /// OpValue. |
505 | static Value *isOneOf(const InstructionsState &S, Value *Op) { |
506 | auto *I = dyn_cast<Instruction>(Op); |
507 | if (I && S.isOpcodeOrAlt(I)) |
508 | return Op; |
509 | return S.OpValue; |
510 | } |
511 | |
512 | /// \returns true if \p Opcode is allowed as part of of the main/alternate |
513 | /// instruction for SLP vectorization. |
514 | /// |
515 | /// Example of unsupported opcode is SDIV that can potentially cause UB if the |
516 | /// "shuffled out" lane would result in division by zero. |
517 | static bool isValidForAlternation(unsigned Opcode) { |
518 | if (Instruction::isIntDivRem(Opcode)) |
519 | return false; |
520 | |
521 | return true; |
522 | } |
523 | |
524 | static InstructionsState getSameOpcode(ArrayRef<Value *> VL, |
525 | unsigned BaseIndex = 0); |
526 | |
527 | /// Checks if the provided operands of 2 cmp instructions are compatible, i.e. |
528 | /// compatible instructions or constants, or just some other regular values. |
529 | static bool areCompatibleCmpOps(Value *BaseOp0, Value *BaseOp1, Value *Op0, |
530 | Value *Op1) { |
531 | return (isConstant(BaseOp0) && isConstant(Op0)) || |
532 | (isConstant(BaseOp1) && isConstant(Op1)) || |
533 | (!isa<Instruction>(BaseOp0) && !isa<Instruction>(Op0) && |
534 | !isa<Instruction>(BaseOp1) && !isa<Instruction>(Op1)) || |
535 | getSameOpcode({BaseOp0, Op0}).getOpcode() || |
536 | getSameOpcode({BaseOp1, Op1}).getOpcode(); |
537 | } |
538 | |
539 | /// \returns true if a compare instruction \p CI has similar "look" and |
540 | /// same predicate as \p BaseCI, "as is" or with its operands and predicate |
541 | /// swapped, false otherwise. |
542 | static bool isCmpSameOrSwapped(const CmpInst *BaseCI, const CmpInst *CI) { |
543 | 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", 544, __extension__ __PRETTY_FUNCTION__)) |
544 | "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", 544, __extension__ __PRETTY_FUNCTION__)); |
545 | CmpInst::Predicate BasePred = BaseCI->getPredicate(); |
546 | CmpInst::Predicate Pred = CI->getPredicate(); |
547 | CmpInst::Predicate SwappedPred = CmpInst::getSwappedPredicate(Pred); |
548 | |
549 | Value *BaseOp0 = BaseCI->getOperand(0); |
550 | Value *BaseOp1 = BaseCI->getOperand(1); |
551 | Value *Op0 = CI->getOperand(0); |
552 | Value *Op1 = CI->getOperand(1); |
553 | |
554 | return (BasePred == Pred && |
555 | areCompatibleCmpOps(BaseOp0, BaseOp1, Op0, Op1)) || |
556 | (BasePred == SwappedPred && |
557 | areCompatibleCmpOps(BaseOp0, BaseOp1, Op1, Op0)); |
558 | } |
559 | |
560 | /// \returns analysis of the Instructions in \p VL described in |
561 | /// InstructionsState, the Opcode that we suppose the whole list |
562 | /// could be vectorized even if its structure is diverse. |
563 | static InstructionsState getSameOpcode(ArrayRef<Value *> VL, |
564 | unsigned BaseIndex) { |
565 | // Make sure these are all Instructions. |
566 | if (llvm::any_of(VL, [](Value *V) { return !isa<Instruction>(V); })) |
567 | return InstructionsState(VL[BaseIndex], nullptr, nullptr); |
568 | |
569 | bool IsCastOp = isa<CastInst>(VL[BaseIndex]); |
570 | bool IsBinOp = isa<BinaryOperator>(VL[BaseIndex]); |
571 | bool IsCmpOp = isa<CmpInst>(VL[BaseIndex]); |
572 | CmpInst::Predicate BasePred = |
573 | IsCmpOp ? cast<CmpInst>(VL[BaseIndex])->getPredicate() |
574 | : CmpInst::BAD_ICMP_PREDICATE; |
575 | unsigned Opcode = cast<Instruction>(VL[BaseIndex])->getOpcode(); |
576 | unsigned AltOpcode = Opcode; |
577 | unsigned AltIndex = BaseIndex; |
578 | |
579 | // Check for one alternate opcode from another BinaryOperator. |
580 | // TODO - generalize to support all operators (types, calls etc.). |
581 | for (int Cnt = 0, E = VL.size(); Cnt < E; Cnt++) { |
582 | unsigned InstOpcode = cast<Instruction>(VL[Cnt])->getOpcode(); |
583 | if (IsBinOp && isa<BinaryOperator>(VL[Cnt])) { |
584 | if (InstOpcode == Opcode || InstOpcode == AltOpcode) |
585 | continue; |
586 | if (Opcode == AltOpcode && isValidForAlternation(InstOpcode) && |
587 | isValidForAlternation(Opcode)) { |
588 | AltOpcode = InstOpcode; |
589 | AltIndex = Cnt; |
590 | continue; |
591 | } |
592 | } else if (IsCastOp && isa<CastInst>(VL[Cnt])) { |
593 | Type *Ty0 = cast<Instruction>(VL[BaseIndex])->getOperand(0)->getType(); |
594 | Type *Ty1 = cast<Instruction>(VL[Cnt])->getOperand(0)->getType(); |
595 | if (Ty0 == Ty1) { |
596 | if (InstOpcode == Opcode || InstOpcode == AltOpcode) |
597 | continue; |
598 | if (Opcode == AltOpcode) { |
599 | 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", 601, __extension__ __PRETTY_FUNCTION__)) |
600 | 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", 601, __extension__ __PRETTY_FUNCTION__)) |
601 | "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", 601, __extension__ __PRETTY_FUNCTION__)); |
602 | AltOpcode = InstOpcode; |
603 | AltIndex = Cnt; |
604 | continue; |
605 | } |
606 | } |
607 | } else if (auto *Inst = dyn_cast<CmpInst>(VL[Cnt]); Inst && IsCmpOp) { |
608 | auto *BaseInst = cast<CmpInst>(VL[BaseIndex]); |
609 | Type *Ty0 = BaseInst->getOperand(0)->getType(); |
610 | Type *Ty1 = Inst->getOperand(0)->getType(); |
611 | if (Ty0 == Ty1) { |
612 | 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", 612, __extension__ __PRETTY_FUNCTION__)); |
613 | // Check for compatible operands. If the corresponding operands are not |
614 | // compatible - need to perform alternate vectorization. |
615 | CmpInst::Predicate CurrentPred = Inst->getPredicate(); |
616 | CmpInst::Predicate SwappedCurrentPred = |
617 | CmpInst::getSwappedPredicate(CurrentPred); |
618 | |
619 | if (E == 2 && |
620 | (BasePred == CurrentPred || BasePred == SwappedCurrentPred)) |
621 | continue; |
622 | |
623 | if (isCmpSameOrSwapped(BaseInst, Inst)) |
624 | continue; |
625 | auto *AltInst = cast<CmpInst>(VL[AltIndex]); |
626 | if (AltIndex != BaseIndex) { |
627 | if (isCmpSameOrSwapped(AltInst, Inst)) |
628 | continue; |
629 | } else if (BasePred != CurrentPred) { |
630 | 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", 632, __extension__ __PRETTY_FUNCTION__)) |
631 | 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", 632, __extension__ __PRETTY_FUNCTION__)) |
632 | "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", 632, __extension__ __PRETTY_FUNCTION__)); |
633 | AltIndex = Cnt; |
634 | continue; |
635 | } |
636 | CmpInst::Predicate AltPred = AltInst->getPredicate(); |
637 | if (BasePred == CurrentPred || BasePred == SwappedCurrentPred || |
638 | AltPred == CurrentPred || AltPred == SwappedCurrentPred) |
639 | continue; |
640 | } |
641 | } else if (InstOpcode == Opcode || InstOpcode == AltOpcode) |
642 | continue; |
643 | return InstructionsState(VL[BaseIndex], nullptr, nullptr); |
644 | } |
645 | |
646 | return InstructionsState(VL[BaseIndex], cast<Instruction>(VL[BaseIndex]), |
647 | cast<Instruction>(VL[AltIndex])); |
648 | } |
649 | |
650 | /// \returns true if all of the values in \p VL have the same type or false |
651 | /// otherwise. |
652 | static bool allSameType(ArrayRef<Value *> VL) { |
653 | Type *Ty = VL[0]->getType(); |
654 | for (int i = 1, e = VL.size(); i < e; i++) |
655 | if (VL[i]->getType() != Ty) |
656 | return false; |
657 | |
658 | return true; |
659 | } |
660 | |
661 | /// \returns True if Extract{Value,Element} instruction extracts element Idx. |
662 | static Optional<unsigned> getExtractIndex(Instruction *E) { |
663 | unsigned Opcode = E->getOpcode(); |
664 | 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", 666, __extension__ __PRETTY_FUNCTION__)) |
665 | 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", 666, __extension__ __PRETTY_FUNCTION__)) |
666 | "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", 666, __extension__ __PRETTY_FUNCTION__)); |
667 | if (Opcode == Instruction::ExtractElement) { |
668 | auto *CI = dyn_cast<ConstantInt>(E->getOperand(1)); |
669 | if (!CI) |
670 | return None; |
671 | return CI->getZExtValue(); |
672 | } |
673 | ExtractValueInst *EI = cast<ExtractValueInst>(E); |
674 | if (EI->getNumIndices() != 1) |
675 | return None; |
676 | return *EI->idx_begin(); |
677 | } |
678 | |
679 | /// \returns True if in-tree use also needs extract. This refers to |
680 | /// possible scalar operand in vectorized instruction. |
681 | static bool InTreeUserNeedToExtract(Value *Scalar, Instruction *UserInst, |
682 | TargetLibraryInfo *TLI) { |
683 | unsigned Opcode = UserInst->getOpcode(); |
684 | switch (Opcode) { |
685 | case Instruction::Load: { |
686 | LoadInst *LI = cast<LoadInst>(UserInst); |
687 | return (LI->getPointerOperand() == Scalar); |
688 | } |
689 | case Instruction::Store: { |
690 | StoreInst *SI = cast<StoreInst>(UserInst); |
691 | return (SI->getPointerOperand() == Scalar); |
692 | } |
693 | case Instruction::Call: { |
694 | CallInst *CI = cast<CallInst>(UserInst); |
695 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); |
696 | for (unsigned i = 0, e = CI->arg_size(); i != e; ++i) { |
697 | if (isVectorIntrinsicWithScalarOpAtArg(ID, i)) |
698 | return (CI->getArgOperand(i) == Scalar); |
699 | } |
700 | [[fallthrough]]; |
701 | } |
702 | default: |
703 | return false; |
704 | } |
705 | } |
706 | |
707 | /// \returns the AA location that is being access by the instruction. |
708 | static MemoryLocation getLocation(Instruction *I) { |
709 | if (StoreInst *SI = dyn_cast<StoreInst>(I)) |
710 | return MemoryLocation::get(SI); |
711 | if (LoadInst *LI = dyn_cast<LoadInst>(I)) |
712 | return MemoryLocation::get(LI); |
713 | return MemoryLocation(); |
714 | } |
715 | |
716 | /// \returns True if the instruction is not a volatile or atomic load/store. |
717 | static bool isSimple(Instruction *I) { |
718 | if (LoadInst *LI = dyn_cast<LoadInst>(I)) |
719 | return LI->isSimple(); |
720 | if (StoreInst *SI = dyn_cast<StoreInst>(I)) |
721 | return SI->isSimple(); |
722 | if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I)) |
723 | return !MI->isVolatile(); |
724 | return true; |
725 | } |
726 | |
727 | /// Shuffles \p Mask in accordance with the given \p SubMask. |
728 | static void addMask(SmallVectorImpl<int> &Mask, ArrayRef<int> SubMask) { |
729 | if (SubMask.empty()) |
730 | return; |
731 | if (Mask.empty()) { |
732 | Mask.append(SubMask.begin(), SubMask.end()); |
733 | return; |
734 | } |
735 | SmallVector<int> NewMask(SubMask.size(), UndefMaskElem); |
736 | int TermValue = std::min(Mask.size(), SubMask.size()); |
737 | for (int I = 0, E = SubMask.size(); I < E; ++I) { |
738 | if (SubMask[I] >= TermValue || SubMask[I] == UndefMaskElem || |
739 | Mask[SubMask[I]] >= TermValue) |
740 | continue; |
741 | NewMask[I] = Mask[SubMask[I]]; |
742 | } |
743 | Mask.swap(NewMask); |
744 | } |
745 | |
746 | /// Order may have elements assigned special value (size) which is out of |
747 | /// bounds. Such indices only appear on places which correspond to undef values |
748 | /// (see canReuseExtract for details) and used in order to avoid undef values |
749 | /// have effect on operands ordering. |
750 | /// The first loop below simply finds all unused indices and then the next loop |
751 | /// nest assigns these indices for undef values positions. |
752 | /// As an example below Order has two undef positions and they have assigned |
753 | /// values 3 and 7 respectively: |
754 | /// before: 6 9 5 4 9 2 1 0 |
755 | /// after: 6 3 5 4 7 2 1 0 |
756 | static void fixupOrderingIndices(SmallVectorImpl<unsigned> &Order) { |
757 | const unsigned Sz = Order.size(); |
758 | SmallBitVector UnusedIndices(Sz, /*t=*/true); |
759 | SmallBitVector MaskedIndices(Sz); |
760 | for (unsigned I = 0; I < Sz; ++I) { |
761 | if (Order[I] < Sz) |
762 | UnusedIndices.reset(Order[I]); |
763 | else |
764 | MaskedIndices.set(I); |
765 | } |
766 | if (MaskedIndices.none()) |
767 | return; |
768 | 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", 769, __extension__ __PRETTY_FUNCTION__)) |
769 | "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", 769, __extension__ __PRETTY_FUNCTION__)); |
770 | int Idx = UnusedIndices.find_first(); |
771 | int MIdx = MaskedIndices.find_first(); |
772 | while (MIdx >= 0) { |
773 | 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", 773, __extension__ __PRETTY_FUNCTION__)); |
774 | Order[MIdx] = Idx; |
775 | Idx = UnusedIndices.find_next(Idx); |
776 | MIdx = MaskedIndices.find_next(MIdx); |
777 | } |
778 | } |
779 | |
780 | namespace llvm { |
781 | |
782 | static void inversePermutation(ArrayRef<unsigned> Indices, |
783 | SmallVectorImpl<int> &Mask) { |
784 | Mask.clear(); |
785 | const unsigned E = Indices.size(); |
786 | Mask.resize(E, UndefMaskElem); |
787 | for (unsigned I = 0; I < E; ++I) |
788 | Mask[Indices[I]] = I; |
789 | } |
790 | |
791 | /// Reorders the list of scalars in accordance with the given \p Mask. |
792 | static void reorderScalars(SmallVectorImpl<Value *> &Scalars, |
793 | ArrayRef<int> Mask) { |
794 | 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", 794, __extension__ __PRETTY_FUNCTION__)); |
795 | SmallVector<Value *> Prev(Scalars.size(), |
796 | UndefValue::get(Scalars.front()->getType())); |
797 | Prev.swap(Scalars); |
798 | for (unsigned I = 0, E = Prev.size(); I < E; ++I) |
799 | if (Mask[I] != UndefMaskElem) |
800 | Scalars[Mask[I]] = Prev[I]; |
801 | } |
802 | |
803 | /// Checks if the provided value does not require scheduling. It does not |
804 | /// require scheduling if this is not an instruction or it is an instruction |
805 | /// that does not read/write memory and all operands are either not instructions |
806 | /// or phi nodes or instructions from different blocks. |
807 | static bool areAllOperandsNonInsts(Value *V) { |
808 | auto *I = dyn_cast<Instruction>(V); |
809 | if (!I) |
810 | return true; |
811 | return !mayHaveNonDefUseDependency(*I) && |
812 | all_of(I->operands(), [I](Value *V) { |
813 | auto *IO = dyn_cast<Instruction>(V); |
814 | if (!IO) |
815 | return true; |
816 | return isa<PHINode>(IO) || IO->getParent() != I->getParent(); |
817 | }); |
818 | } |
819 | |
820 | /// Checks if the provided value does not require scheduling. It does not |
821 | /// require scheduling if this is not an instruction or it is an instruction |
822 | /// that does not read/write memory and all users are phi nodes or instructions |
823 | /// from the different blocks. |
824 | static bool isUsedOutsideBlock(Value *V) { |
825 | auto *I = dyn_cast<Instruction>(V); |
826 | if (!I) |
827 | return true; |
828 | // Limits the number of uses to save compile time. |
829 | constexpr int UsesLimit = 8; |
830 | return !I->mayReadOrWriteMemory() && !I->hasNUsesOrMore(UsesLimit) && |
831 | all_of(I->users(), [I](User *U) { |
832 | auto *IU = dyn_cast<Instruction>(U); |
833 | if (!IU) |
834 | return true; |
835 | return IU->getParent() != I->getParent() || isa<PHINode>(IU); |
836 | }); |
837 | } |
838 | |
839 | /// Checks if the specified value does not require scheduling. It does not |
840 | /// require scheduling if all operands and all users do not need to be scheduled |
841 | /// in the current basic block. |
842 | static bool doesNotNeedToBeScheduled(Value *V) { |
843 | return areAllOperandsNonInsts(V) && isUsedOutsideBlock(V); |
844 | } |
845 | |
846 | /// Checks if the specified array of instructions does not require scheduling. |
847 | /// It is so if all either instructions have operands that do not require |
848 | /// scheduling or their users do not require scheduling since they are phis or |
849 | /// in other basic blocks. |
850 | static bool doesNotNeedToSchedule(ArrayRef<Value *> VL) { |
851 | return !VL.empty() && |
852 | (all_of(VL, isUsedOutsideBlock) || all_of(VL, areAllOperandsNonInsts)); |
853 | } |
854 | |
855 | namespace slpvectorizer { |
856 | |
857 | /// Bottom Up SLP Vectorizer. |
858 | class BoUpSLP { |
859 | struct TreeEntry; |
860 | struct ScheduleData; |
861 | |
862 | public: |
863 | using ValueList = SmallVector<Value *, 8>; |
864 | using InstrList = SmallVector<Instruction *, 16>; |
865 | using ValueSet = SmallPtrSet<Value *, 16>; |
866 | using StoreList = SmallVector<StoreInst *, 8>; |
867 | using ExtraValueToDebugLocsMap = |
868 | MapVector<Value *, SmallVector<Instruction *, 2>>; |
869 | using OrdersType = SmallVector<unsigned, 4>; |
870 | |
871 | BoUpSLP(Function *Func, ScalarEvolution *Se, TargetTransformInfo *Tti, |
872 | TargetLibraryInfo *TLi, AAResults *Aa, LoopInfo *Li, |
873 | DominatorTree *Dt, AssumptionCache *AC, DemandedBits *DB, |
874 | const DataLayout *DL, OptimizationRemarkEmitter *ORE) |
875 | : BatchAA(*Aa), F(Func), SE(Se), TTI(Tti), TLI(TLi), LI(Li), |
876 | DT(Dt), AC(AC), DB(DB), DL(DL), ORE(ORE), Builder(Se->getContext()) { |
877 | CodeMetrics::collectEphemeralValues(F, AC, EphValues); |
878 | // Use the vector register size specified by the target unless overridden |
879 | // by a command-line option. |
880 | // TODO: It would be better to limit the vectorization factor based on |
881 | // data type rather than just register size. For example, x86 AVX has |
882 | // 256-bit registers, but it does not support integer operations |
883 | // at that width (that requires AVX2). |
884 | if (MaxVectorRegSizeOption.getNumOccurrences()) |
885 | MaxVecRegSize = MaxVectorRegSizeOption; |
886 | else |
887 | MaxVecRegSize = |
888 | TTI->getRegisterBitWidth(TargetTransformInfo::RGK_FixedWidthVector) |
889 | .getFixedSize(); |
890 | |
891 | if (MinVectorRegSizeOption.getNumOccurrences()) |
892 | MinVecRegSize = MinVectorRegSizeOption; |
893 | else |
894 | MinVecRegSize = TTI->getMinVectorRegisterBitWidth(); |
895 | } |
896 | |
897 | /// Vectorize the tree that starts with the elements in \p VL. |
898 | /// Returns the vectorized root. |
899 | Value *vectorizeTree(); |
900 | |
901 | /// Vectorize the tree but with the list of externally used values \p |
902 | /// ExternallyUsedValues. Values in this MapVector can be replaced but the |
903 | /// generated extractvalue instructions. |
904 | Value *vectorizeTree(ExtraValueToDebugLocsMap &ExternallyUsedValues); |
905 | |
906 | /// \returns the cost incurred by unwanted spills and fills, caused by |
907 | /// holding live values over call sites. |
908 | InstructionCost getSpillCost() const; |
909 | |
910 | /// \returns the vectorization cost of the subtree that starts at \p VL. |
911 | /// A negative number means that this is profitable. |
912 | InstructionCost getTreeCost(ArrayRef<Value *> VectorizedVals = None); |
913 | |
914 | /// Construct a vectorizable tree that starts at \p Roots, ignoring users for |
915 | /// the purpose of scheduling and extraction in the \p UserIgnoreLst. |
916 | void buildTree(ArrayRef<Value *> Roots, |
917 | const SmallDenseSet<Value *> &UserIgnoreLst); |
918 | |
919 | /// Construct a vectorizable tree that starts at \p Roots. |
920 | void buildTree(ArrayRef<Value *> Roots); |
921 | |
922 | /// Builds external uses of the vectorized scalars, i.e. the list of |
923 | /// vectorized scalars to be extracted, their lanes and their scalar users. \p |
924 | /// ExternallyUsedValues contains additional list of external uses to handle |
925 | /// vectorization of reductions. |
926 | void |
927 | buildExternalUses(const ExtraValueToDebugLocsMap &ExternallyUsedValues = {}); |
928 | |
929 | /// Clear the internal data structures that are created by 'buildTree'. |
930 | void deleteTree() { |
931 | VectorizableTree.clear(); |
932 | ScalarToTreeEntry.clear(); |
933 | MustGather.clear(); |
934 | ExternalUses.clear(); |
935 | for (auto &Iter : BlocksSchedules) { |
936 | BlockScheduling *BS = Iter.second.get(); |
937 | BS->clear(); |
938 | } |
939 | MinBWs.clear(); |
940 | InstrElementSize.clear(); |
941 | UserIgnoreList = nullptr; |
942 | } |
943 | |
944 | unsigned getTreeSize() const { return VectorizableTree.size(); } |
945 | |
946 | /// Perform LICM and CSE on the newly generated gather sequences. |
947 | void optimizeGatherSequence(); |
948 | |
949 | /// Checks if the specified gather tree entry \p TE can be represented as a |
950 | /// shuffled vector entry + (possibly) permutation with other gathers. It |
951 | /// implements the checks only for possibly ordered scalars (Loads, |
952 | /// ExtractElement, ExtractValue), which can be part of the graph. |
953 | Optional<OrdersType> findReusedOrderedScalars(const TreeEntry &TE); |
954 | |
955 | /// Sort loads into increasing pointers offsets to allow greater clustering. |
956 | Optional<OrdersType> findPartiallyOrderedLoads(const TreeEntry &TE); |
957 | |
958 | /// Gets reordering data for the given tree entry. If the entry is vectorized |
959 | /// - just return ReorderIndices, otherwise check if the scalars can be |
960 | /// reordered and return the most optimal order. |
961 | /// \param TopToBottom If true, include the order of vectorized stores and |
962 | /// insertelement nodes, otherwise skip them. |
963 | Optional<OrdersType> getReorderingData(const TreeEntry &TE, bool TopToBottom); |
964 | |
965 | /// Reorders the current graph to the most profitable order starting from the |
966 | /// root node to the leaf nodes. The best order is chosen only from the nodes |
967 | /// of the same size (vectorization factor). Smaller nodes are considered |
968 | /// parts of subgraph with smaller VF and they are reordered independently. We |
969 | /// can make it because we still need to extend smaller nodes to the wider VF |
970 | /// and we can merge reordering shuffles with the widening shuffles. |
971 | void reorderTopToBottom(); |
972 | |
973 | /// Reorders the current graph to the most profitable order starting from |
974 | /// leaves to the root. It allows to rotate small subgraphs and reduce the |
975 | /// number of reshuffles if the leaf nodes use the same order. In this case we |
976 | /// can merge the orders and just shuffle user node instead of shuffling its |
977 | /// operands. Plus, even the leaf nodes have different orders, it allows to |
978 | /// sink reordering in the graph closer to the root node and merge it later |
979 | /// during analysis. |
980 | void reorderBottomToTop(bool IgnoreReorder = false); |
981 | |
982 | /// \return The vector element size in bits to use when vectorizing the |
983 | /// expression tree ending at \p V. If V is a store, the size is the width of |
984 | /// the stored value. Otherwise, the size is the width of the largest loaded |
985 | /// value reaching V. This method is used by the vectorizer to calculate |
986 | /// vectorization factors. |
987 | unsigned getVectorElementSize(Value *V); |
988 | |
989 | /// Compute the minimum type sizes required to represent the entries in a |
990 | /// vectorizable tree. |
991 | void computeMinimumValueSizes(); |
992 | |
993 | // \returns maximum vector register size as set by TTI or overridden by cl::opt. |
994 | unsigned getMaxVecRegSize() const { |
995 | return MaxVecRegSize; |
996 | } |
997 | |
998 | // \returns minimum vector register size as set by cl::opt. |
999 | unsigned getMinVecRegSize() const { |
1000 | return MinVecRegSize; |
1001 | } |
1002 | |
1003 | unsigned getMinVF(unsigned Sz) const { |
1004 | return std::max(2U, getMinVecRegSize() / Sz); |
1005 | } |
1006 | |
1007 | unsigned getMaximumVF(unsigned ElemWidth, unsigned Opcode) const { |
1008 | unsigned MaxVF = MaxVFOption.getNumOccurrences() ? |
1009 | MaxVFOption : TTI->getMaximumVF(ElemWidth, Opcode); |
1010 | return MaxVF ? MaxVF : UINT_MAX(2147483647 *2U +1U); |
1011 | } |
1012 | |
1013 | /// Check if homogeneous aggregate is isomorphic to some VectorType. |
1014 | /// Accepts homogeneous multidimensional aggregate of scalars/vectors like |
1015 | /// {[4 x i16], [4 x i16]}, { <2 x float>, <2 x float> }, |
1016 | /// {{{i16, i16}, {i16, i16}}, {{i16, i16}, {i16, i16}}} and so on. |
1017 | /// |
1018 | /// \returns number of elements in vector if isomorphism exists, 0 otherwise. |
1019 | unsigned canMapToVector(Type *T, const DataLayout &DL) const; |
1020 | |
1021 | /// \returns True if the VectorizableTree is both tiny and not fully |
1022 | /// vectorizable. We do not vectorize such trees. |
1023 | bool isTreeTinyAndNotFullyVectorizable(bool ForReduction = false) const; |
1024 | |
1025 | /// Assume that a legal-sized 'or'-reduction of shifted/zexted loaded values |
1026 | /// can be load combined in the backend. Load combining may not be allowed in |
1027 | /// the IR optimizer, so we do not want to alter the pattern. For example, |
1028 | /// partially transforming a scalar bswap() pattern into vector code is |
1029 | /// effectively impossible for the backend to undo. |
1030 | /// TODO: If load combining is allowed in the IR optimizer, this analysis |
1031 | /// may not be necessary. |
1032 | bool isLoadCombineReductionCandidate(RecurKind RdxKind) const; |
1033 | |
1034 | /// Assume that a vector of stores of bitwise-or/shifted/zexted loaded values |
1035 | /// can be load combined in the backend. Load combining may not be allowed in |
1036 | /// the IR optimizer, so we do not want to alter the pattern. For example, |
1037 | /// partially transforming a scalar bswap() pattern into vector code is |
1038 | /// effectively impossible for the backend to undo. |
1039 | /// TODO: If load combining is allowed in the IR optimizer, this analysis |
1040 | /// may not be necessary. |
1041 | bool isLoadCombineCandidate() const; |
1042 | |
1043 | OptimizationRemarkEmitter *getORE() { return ORE; } |
1044 | |
1045 | /// This structure holds any data we need about the edges being traversed |
1046 | /// during buildTree_rec(). We keep track of: |
1047 | /// (i) the user TreeEntry index, and |
1048 | /// (ii) the index of the edge. |
1049 | struct EdgeInfo { |
1050 | EdgeInfo() = default; |
1051 | EdgeInfo(TreeEntry *UserTE, unsigned EdgeIdx) |
1052 | : UserTE(UserTE), EdgeIdx(EdgeIdx) {} |
1053 | /// The user TreeEntry. |
1054 | TreeEntry *UserTE = nullptr; |
1055 | /// The operand index of the use. |
1056 | unsigned EdgeIdx = UINT_MAX(2147483647 *2U +1U); |
1057 | #ifndef NDEBUG |
1058 | friend inline raw_ostream &operator<<(raw_ostream &OS, |
1059 | const BoUpSLP::EdgeInfo &EI) { |
1060 | EI.dump(OS); |
1061 | return OS; |
1062 | } |
1063 | /// Debug print. |
1064 | void dump(raw_ostream &OS) const { |
1065 | OS << "{User:" << (UserTE ? std::to_string(UserTE->Idx) : "null") |
1066 | << " EdgeIdx:" << EdgeIdx << "}"; |
1067 | } |
1068 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void dump() const { dump(dbgs()); } |
1069 | #endif |
1070 | }; |
1071 | |
1072 | /// A helper class used for scoring candidates for two consecutive lanes. |
1073 | class LookAheadHeuristics { |
1074 | const DataLayout &DL; |
1075 | ScalarEvolution &SE; |
1076 | const BoUpSLP &R; |
1077 | int NumLanes; // Total number of lanes (aka vectorization factor). |
1078 | int MaxLevel; // The maximum recursion depth for accumulating score. |
1079 | |
1080 | public: |
1081 | LookAheadHeuristics(const DataLayout &DL, ScalarEvolution &SE, |
1082 | const BoUpSLP &R, int NumLanes, int MaxLevel) |
1083 | : DL(DL), SE(SE), R(R), NumLanes(NumLanes), MaxLevel(MaxLevel) {} |
1084 | |
1085 | // The hard-coded scores listed here are not very important, though it shall |
1086 | // be higher for better matches to improve the resulting cost. When |
1087 | // computing the scores of matching one sub-tree with another, we are |
1088 | // basically counting the number of values that are matching. So even if all |
1089 | // scores are set to 1, we would still get a decent matching result. |
1090 | // However, sometimes we have to break ties. For example we may have to |
1091 | // choose between matching loads vs matching opcodes. This is what these |
1092 | // scores are helping us with: they provide the order of preference. Also, |
1093 | // this is important if the scalar is externally used or used in another |
1094 | // tree entry node in the different lane. |
1095 | |
1096 | /// Loads from consecutive memory addresses, e.g. load(A[i]), load(A[i+1]). |
1097 | static const int ScoreConsecutiveLoads = 4; |
1098 | /// The same load multiple times. This should have a better score than |
1099 | /// `ScoreSplat` because it in x86 for a 2-lane vector we can represent it |
1100 | /// with `movddup (%reg), xmm0` which has a throughput of 0.5 versus 0.5 for |
1101 | /// a vector load and 1.0 for a broadcast. |
1102 | static const int ScoreSplatLoads = 3; |
1103 | /// Loads from reversed memory addresses, e.g. load(A[i+1]), load(A[i]). |
1104 | static const int ScoreReversedLoads = 3; |
1105 | /// ExtractElementInst from same vector and consecutive indexes. |
1106 | static const int ScoreConsecutiveExtracts = 4; |
1107 | /// ExtractElementInst from same vector and reversed indices. |
1108 | static const int ScoreReversedExtracts = 3; |
1109 | /// Constants. |
1110 | static const int ScoreConstants = 2; |
1111 | /// Instructions with the same opcode. |
1112 | static const int ScoreSameOpcode = 2; |
1113 | /// Instructions with alt opcodes (e.g, add + sub). |
1114 | static const int ScoreAltOpcodes = 1; |
1115 | /// Identical instructions (a.k.a. splat or broadcast). |
1116 | static const int ScoreSplat = 1; |
1117 | /// Matching with an undef is preferable to failing. |
1118 | static const int ScoreUndef = 1; |
1119 | /// Score for failing to find a decent match. |
1120 | static const int ScoreFail = 0; |
1121 | /// Score if all users are vectorized. |
1122 | static const int ScoreAllUserVectorized = 1; |
1123 | |
1124 | /// \returns the score of placing \p V1 and \p V2 in consecutive lanes. |
1125 | /// \p U1 and \p U2 are the users of \p V1 and \p V2. |
1126 | /// Also, checks if \p V1 and \p V2 are compatible with instructions in \p |
1127 | /// MainAltOps. |
1128 | int getShallowScore(Value *V1, Value *V2, Instruction *U1, Instruction *U2, |
1129 | ArrayRef<Value *> MainAltOps) const { |
1130 | if (V1 == V2) { |
1131 | if (isa<LoadInst>(V1)) { |
1132 | // Retruns true if the users of V1 and V2 won't need to be extracted. |
1133 | auto AllUsersAreInternal = [U1, U2, this](Value *V1, Value *V2) { |
1134 | // Bail out if we have too many uses to save compilation time. |
1135 | static constexpr unsigned Limit = 8; |
1136 | if (V1->hasNUsesOrMore(Limit) || V2->hasNUsesOrMore(Limit)) |
1137 | return false; |
1138 | |
1139 | auto AllUsersVectorized = [U1, U2, this](Value *V) { |
1140 | return llvm::all_of(V->users(), [U1, U2, this](Value *U) { |
1141 | return U == U1 || U == U2 || R.getTreeEntry(U) != nullptr; |
1142 | }); |
1143 | }; |
1144 | return AllUsersVectorized(V1) && AllUsersVectorized(V2); |
1145 | }; |
1146 | // A broadcast of a load can be cheaper on some targets. |
1147 | if (R.TTI->isLegalBroadcastLoad(V1->getType(), |
1148 | ElementCount::getFixed(NumLanes)) && |
1149 | ((int)V1->getNumUses() == NumLanes || |
1150 | AllUsersAreInternal(V1, V2))) |
1151 | return LookAheadHeuristics::ScoreSplatLoads; |
1152 | } |
1153 | return LookAheadHeuristics::ScoreSplat; |
1154 | } |
1155 | |
1156 | auto *LI1 = dyn_cast<LoadInst>(V1); |
1157 | auto *LI2 = dyn_cast<LoadInst>(V2); |
1158 | if (LI1 && LI2) { |
1159 | if (LI1->getParent() != LI2->getParent()) |
1160 | return LookAheadHeuristics::ScoreFail; |
1161 | |
1162 | Optional<int> Dist = getPointersDiff( |
1163 | LI1->getType(), LI1->getPointerOperand(), LI2->getType(), |
1164 | LI2->getPointerOperand(), DL, SE, /*StrictCheck=*/true); |
1165 | if (!Dist || *Dist == 0) |
1166 | return LookAheadHeuristics::ScoreFail; |
1167 | // The distance is too large - still may be profitable to use masked |
1168 | // loads/gathers. |
1169 | if (std::abs(*Dist) > NumLanes / 2) |
1170 | return LookAheadHeuristics::ScoreAltOpcodes; |
1171 | // This still will detect consecutive loads, but we might have "holes" |
1172 | // in some cases. It is ok for non-power-2 vectorization and may produce |
1173 | // better results. It should not affect current vectorization. |
1174 | return (*Dist > 0) ? LookAheadHeuristics::ScoreConsecutiveLoads |
1175 | : LookAheadHeuristics::ScoreReversedLoads; |
1176 | } |
1177 | |
1178 | auto *C1 = dyn_cast<Constant>(V1); |
1179 | auto *C2 = dyn_cast<Constant>(V2); |
1180 | if (C1 && C2) |
1181 | return LookAheadHeuristics::ScoreConstants; |
1182 | |
1183 | // Extracts from consecutive indexes of the same vector better score as |
1184 | // the extracts could be optimized away. |
1185 | Value *EV1; |
1186 | ConstantInt *Ex1Idx; |
1187 | if (match(V1, m_ExtractElt(m_Value(EV1), m_ConstantInt(Ex1Idx)))) { |
1188 | // Undefs are always profitable for extractelements. |
1189 | if (isa<UndefValue>(V2)) |
1190 | return LookAheadHeuristics::ScoreConsecutiveExtracts; |
1191 | Value *EV2 = nullptr; |
1192 | ConstantInt *Ex2Idx = nullptr; |
1193 | if (match(V2, |
1194 | m_ExtractElt(m_Value(EV2), m_CombineOr(m_ConstantInt(Ex2Idx), |
1195 | m_Undef())))) { |
1196 | // Undefs are always profitable for extractelements. |
1197 | if (!Ex2Idx) |
1198 | return LookAheadHeuristics::ScoreConsecutiveExtracts; |
1199 | if (isUndefVector(EV2) && EV2->getType() == EV1->getType()) |
1200 | return LookAheadHeuristics::ScoreConsecutiveExtracts; |
1201 | if (EV2 == EV1) { |
1202 | int Idx1 = Ex1Idx->getZExtValue(); |
1203 | int Idx2 = Ex2Idx->getZExtValue(); |
1204 | int Dist = Idx2 - Idx1; |
1205 | // The distance is too large - still may be profitable to use |
1206 | // shuffles. |
1207 | if (std::abs(Dist) == 0) |
1208 | return LookAheadHeuristics::ScoreSplat; |
1209 | if (std::abs(Dist) > NumLanes / 2) |
1210 | return LookAheadHeuristics::ScoreSameOpcode; |
1211 | return (Dist > 0) ? LookAheadHeuristics::ScoreConsecutiveExtracts |
1212 | : LookAheadHeuristics::ScoreReversedExtracts; |
1213 | } |
1214 | return LookAheadHeuristics::ScoreAltOpcodes; |
1215 | } |
1216 | return LookAheadHeuristics::ScoreFail; |
1217 | } |
1218 | |
1219 | auto *I1 = dyn_cast<Instruction>(V1); |
1220 | auto *I2 = dyn_cast<Instruction>(V2); |
1221 | if (I1 && I2) { |
1222 | if (I1->getParent() != I2->getParent()) |
1223 | return LookAheadHeuristics::ScoreFail; |
1224 | SmallVector<Value *, 4> Ops(MainAltOps.begin(), MainAltOps.end()); |
1225 | Ops.push_back(I1); |
1226 | Ops.push_back(I2); |
1227 | InstructionsState S = getSameOpcode(Ops); |
1228 | // Note: Only consider instructions with <= 2 operands to avoid |
1229 | // complexity explosion. |
1230 | if (S.getOpcode() && |
1231 | (S.MainOp->getNumOperands() <= 2 || !MainAltOps.empty() || |
1232 | !S.isAltShuffle()) && |
1233 | all_of(Ops, [&S](Value *V) { |
1234 | return cast<Instruction>(V)->getNumOperands() == |
1235 | S.MainOp->getNumOperands(); |
1236 | })) |
1237 | return S.isAltShuffle() ? LookAheadHeuristics::ScoreAltOpcodes |
1238 | : LookAheadHeuristics::ScoreSameOpcode; |
1239 | } |
1240 | |
1241 | if (isa<UndefValue>(V2)) |
1242 | return LookAheadHeuristics::ScoreUndef; |
1243 | |
1244 | return LookAheadHeuristics::ScoreFail; |
1245 | } |
1246 | |
1247 | /// Go through the operands of \p LHS and \p RHS recursively until |
1248 | /// MaxLevel, and return the cummulative score. \p U1 and \p U2 are |
1249 | /// the users of \p LHS and \p RHS (that is \p LHS and \p RHS are operands |
1250 | /// of \p U1 and \p U2), except at the beginning of the recursion where |
1251 | /// these are set to nullptr. |
1252 | /// |
1253 | /// For example: |
1254 | /// \verbatim |
1255 | /// A[0] B[0] A[1] B[1] C[0] D[0] B[1] A[1] |
1256 | /// \ / \ / \ / \ / |
1257 | /// + + + + |
1258 | /// G1 G2 G3 G4 |
1259 | /// \endverbatim |
1260 | /// The getScoreAtLevelRec(G1, G2) function will try to match the nodes at |
1261 | /// each level recursively, accumulating the score. It starts from matching |
1262 | /// the additions at level 0, then moves on to the loads (level 1). The |
1263 | /// score of G1 and G2 is higher than G1 and G3, because {A[0],A[1]} and |
1264 | /// {B[0],B[1]} match with LookAheadHeuristics::ScoreConsecutiveLoads, while |
1265 | /// {A[0],C[0]} has a score of LookAheadHeuristics::ScoreFail. |
1266 | /// Please note that the order of the operands does not matter, as we |
1267 | /// evaluate the score of all profitable combinations of operands. In |
1268 | /// other words the score of G1 and G4 is the same as G1 and G2. This |
1269 | /// heuristic is based on ideas described in: |
1270 | /// Look-ahead SLP: Auto-vectorization in the presence of commutative |
1271 | /// operations, CGO 2018 by Vasileios Porpodas, Rodrigo C. O. Rocha, |
1272 | /// Luís F. W. Góes |
1273 | int getScoreAtLevelRec(Value *LHS, Value *RHS, Instruction *U1, |
1274 | Instruction *U2, int CurrLevel, |
1275 | ArrayRef<Value *> MainAltOps) const { |
1276 | |
1277 | // Get the shallow score of V1 and V2. |
1278 | int ShallowScoreAtThisLevel = |
1279 | getShallowScore(LHS, RHS, U1, U2, MainAltOps); |
1280 | |
1281 | // If reached MaxLevel, |
1282 | // or if V1 and V2 are not instructions, |
1283 | // or if they are SPLAT, |
1284 | // or if they are not consecutive, |
1285 | // or if profitable to vectorize loads or extractelements, early return |
1286 | // the current cost. |
1287 | auto *I1 = dyn_cast<Instruction>(LHS); |
1288 | auto *I2 = dyn_cast<Instruction>(RHS); |
1289 | if (CurrLevel == MaxLevel || !(I1 && I2) || I1 == I2 || |
1290 | ShallowScoreAtThisLevel == LookAheadHeuristics::ScoreFail || |
1291 | (((isa<LoadInst>(I1) && isa<LoadInst>(I2)) || |
1292 | (I1->getNumOperands() > 2 && I2->getNumOperands() > 2) || |
1293 | (isa<ExtractElementInst>(I1) && isa<ExtractElementInst>(I2))) && |
1294 | ShallowScoreAtThisLevel)) |
1295 | return ShallowScoreAtThisLevel; |
1296 | 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", 1296, __extension__ __PRETTY_FUNCTION__)); |
1297 | |
1298 | // Contains the I2 operand indexes that got matched with I1 operands. |
1299 | SmallSet<unsigned, 4> Op2Used; |
1300 | |
1301 | // Recursion towards the operands of I1 and I2. We are trying all possible |
1302 | // operand pairs, and keeping track of the best score. |
1303 | for (unsigned OpIdx1 = 0, NumOperands1 = I1->getNumOperands(); |
1304 | OpIdx1 != NumOperands1; ++OpIdx1) { |
1305 | // Try to pair op1I with the best operand of I2. |
1306 | int MaxTmpScore = 0; |
1307 | unsigned MaxOpIdx2 = 0; |
1308 | bool FoundBest = false; |
1309 | // If I2 is commutative try all combinations. |
1310 | unsigned FromIdx = isCommutative(I2) ? 0 : OpIdx1; |
1311 | unsigned ToIdx = isCommutative(I2) |
1312 | ? I2->getNumOperands() |
1313 | : std::min(I2->getNumOperands(), OpIdx1 + 1); |
1314 | 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", 1314, __extension__ __PRETTY_FUNCTION__)); |
1315 | for (unsigned OpIdx2 = FromIdx; OpIdx2 != ToIdx; ++OpIdx2) { |
1316 | // Skip operands already paired with OpIdx1. |
1317 | if (Op2Used.count(OpIdx2)) |
1318 | continue; |
1319 | // Recursively calculate the cost at each level |
1320 | int TmpScore = |
1321 | getScoreAtLevelRec(I1->getOperand(OpIdx1), I2->getOperand(OpIdx2), |
1322 | I1, I2, CurrLevel + 1, None); |
1323 | // Look for the best score. |
1324 | if (TmpScore > LookAheadHeuristics::ScoreFail && |
1325 | TmpScore > MaxTmpScore) { |
1326 | MaxTmpScore = TmpScore; |
1327 | MaxOpIdx2 = OpIdx2; |
1328 | FoundBest = true; |
1329 | } |
1330 | } |
1331 | if (FoundBest) { |
1332 | // Pair {OpIdx1, MaxOpIdx2} was found to be best. Never revisit it. |
1333 | Op2Used.insert(MaxOpIdx2); |
1334 | ShallowScoreAtThisLevel += MaxTmpScore; |
1335 | } |
1336 | } |
1337 | return ShallowScoreAtThisLevel; |
1338 | } |
1339 | }; |
1340 | /// A helper data structure to hold the operands of a vector of instructions. |
1341 | /// This supports a fixed vector length for all operand vectors. |
1342 | class VLOperands { |
1343 | /// For each operand we need (i) the value, and (ii) the opcode that it |
1344 | /// would be attached to if the expression was in a left-linearized form. |
1345 | /// This is required to avoid illegal operand reordering. |
1346 | /// For example: |
1347 | /// \verbatim |
1348 | /// 0 Op1 |
1349 | /// |/ |
1350 | /// Op1 Op2 Linearized + Op2 |
1351 | /// \ / ----------> |/ |
1352 | /// - - |
1353 | /// |
1354 | /// Op1 - Op2 (0 + Op1) - Op2 |
1355 | /// \endverbatim |
1356 | /// |
1357 | /// Value Op1 is attached to a '+' operation, and Op2 to a '-'. |
1358 | /// |
1359 | /// Another way to think of this is to track all the operations across the |
1360 | /// path from the operand all the way to the root of the tree and to |
1361 | /// calculate the operation that corresponds to this path. For example, the |
1362 | /// path from Op2 to the root crosses the RHS of the '-', therefore the |
1363 | /// corresponding operation is a '-' (which matches the one in the |
1364 | /// linearized tree, as shown above). |
1365 | /// |
1366 | /// For lack of a better term, we refer to this operation as Accumulated |
1367 | /// Path Operation (APO). |
1368 | struct OperandData { |
1369 | OperandData() = default; |
1370 | OperandData(Value *V, bool APO, bool IsUsed) |
1371 | : V(V), APO(APO), IsUsed(IsUsed) {} |
1372 | /// The operand value. |
1373 | Value *V = nullptr; |
1374 | /// TreeEntries only allow a single opcode, or an alternate sequence of |
1375 | /// them (e.g, +, -). Therefore, we can safely use a boolean value for the |
1376 | /// APO. It is set to 'true' if 'V' is attached to an inverse operation |
1377 | /// in the left-linearized form (e.g., Sub/Div), and 'false' otherwise |
1378 | /// (e.g., Add/Mul) |
1379 | bool APO = false; |
1380 | /// Helper data for the reordering function. |
1381 | bool IsUsed = false; |
1382 | }; |
1383 | |
1384 | /// During operand reordering, we are trying to select the operand at lane |
1385 | /// that matches best with the operand at the neighboring lane. Our |
1386 | /// selection is based on the type of value we are looking for. For example, |
1387 | /// if the neighboring lane has a load, we need to look for a load that is |
1388 | /// accessing a consecutive address. These strategies are summarized in the |
1389 | /// 'ReorderingMode' enumerator. |
1390 | enum class ReorderingMode { |
1391 | Load, ///< Matching loads to consecutive memory addresses |
1392 | Opcode, ///< Matching instructions based on opcode (same or alternate) |
1393 | Constant, ///< Matching constants |
1394 | Splat, ///< Matching the same instruction multiple times (broadcast) |
1395 | Failed, ///< We failed to create a vectorizable group |
1396 | }; |
1397 | |
1398 | using OperandDataVec = SmallVector<OperandData, 2>; |
1399 | |
1400 | /// A vector of operand vectors. |
1401 | SmallVector<OperandDataVec, 4> OpsVec; |
1402 | |
1403 | const DataLayout &DL; |
1404 | ScalarEvolution &SE; |
1405 | const BoUpSLP &R; |
1406 | |
1407 | /// \returns the operand data at \p OpIdx and \p Lane. |
1408 | OperandData &getData(unsigned OpIdx, unsigned Lane) { |
1409 | return OpsVec[OpIdx][Lane]; |
1410 | } |
1411 | |
1412 | /// \returns the operand data at \p OpIdx and \p Lane. Const version. |
1413 | const OperandData &getData(unsigned OpIdx, unsigned Lane) const { |
1414 | return OpsVec[OpIdx][Lane]; |
1415 | } |
1416 | |
1417 | /// Clears the used flag for all entries. |
1418 | void clearUsed() { |
1419 | for (unsigned OpIdx = 0, NumOperands = getNumOperands(); |
1420 | OpIdx != NumOperands; ++OpIdx) |
1421 | for (unsigned Lane = 0, NumLanes = getNumLanes(); Lane != NumLanes; |
1422 | ++Lane) |
1423 | OpsVec[OpIdx][Lane].IsUsed = false; |
1424 | } |
1425 | |
1426 | /// Swap the operand at \p OpIdx1 with that one at \p OpIdx2. |
1427 | void swap(unsigned OpIdx1, unsigned OpIdx2, unsigned Lane) { |
1428 | std::swap(OpsVec[OpIdx1][Lane], OpsVec[OpIdx2][Lane]); |
1429 | } |
1430 | |
1431 | /// \param Lane lane of the operands under analysis. |
1432 | /// \param OpIdx operand index in \p Lane lane we're looking the best |
1433 | /// candidate for. |
1434 | /// \param Idx operand index of the current candidate value. |
1435 | /// \returns The additional score due to possible broadcasting of the |
1436 | /// elements in the lane. It is more profitable to have power-of-2 unique |
1437 | /// elements in the lane, it will be vectorized with higher probability |
1438 | /// after removing duplicates. Currently the SLP vectorizer supports only |
1439 | /// vectorization of the power-of-2 number of unique scalars. |
1440 | int getSplatScore(unsigned Lane, unsigned OpIdx, unsigned Idx) const { |
1441 | Value *IdxLaneV = getData(Idx, Lane).V; |
1442 | if (!isa<Instruction>(IdxLaneV) || IdxLaneV == getData(OpIdx, Lane).V) |
1443 | return 0; |
1444 | SmallPtrSet<Value *, 4> Uniques; |
1445 | for (unsigned Ln = 0, E = getNumLanes(); Ln < E; ++Ln) { |
1446 | if (Ln == Lane) |
1447 | continue; |
1448 | Value *OpIdxLnV = getData(OpIdx, Ln).V; |
1449 | if (!isa<Instruction>(OpIdxLnV)) |
1450 | return 0; |
1451 | Uniques.insert(OpIdxLnV); |
1452 | } |
1453 | int UniquesCount = Uniques.size(); |
1454 | int UniquesCntWithIdxLaneV = |
1455 | Uniques.contains(IdxLaneV) ? UniquesCount : UniquesCount + 1; |
1456 | Value *OpIdxLaneV = getData(OpIdx, Lane).V; |
1457 | int UniquesCntWithOpIdxLaneV = |
1458 | Uniques.contains(OpIdxLaneV) ? UniquesCount : UniquesCount + 1; |
1459 | if (UniquesCntWithIdxLaneV == UniquesCntWithOpIdxLaneV) |
1460 | return 0; |
1461 | return (PowerOf2Ceil(UniquesCntWithOpIdxLaneV) - |
1462 | UniquesCntWithOpIdxLaneV) - |
1463 | (PowerOf2Ceil(UniquesCntWithIdxLaneV) - UniquesCntWithIdxLaneV); |
1464 | } |
1465 | |
1466 | /// \param Lane lane of the operands under analysis. |
1467 | /// \param OpIdx operand index in \p Lane lane we're looking the best |
1468 | /// candidate for. |
1469 | /// \param Idx operand index of the current candidate value. |
1470 | /// \returns The additional score for the scalar which users are all |
1471 | /// vectorized. |
1472 | int getExternalUseScore(unsigned Lane, unsigned OpIdx, unsigned Idx) const { |
1473 | Value *IdxLaneV = getData(Idx, Lane).V; |
1474 | Value *OpIdxLaneV = getData(OpIdx, Lane).V; |
1475 | // Do not care about number of uses for vector-like instructions |
1476 | // (extractelement/extractvalue with constant indices), they are extracts |
1477 | // themselves and already externally used. Vectorization of such |
1478 | // instructions does not add extra extractelement instruction, just may |
1479 | // remove it. |
1480 | if (isVectorLikeInstWithConstOps(IdxLaneV) && |
1481 | isVectorLikeInstWithConstOps(OpIdxLaneV)) |
1482 | return LookAheadHeuristics::ScoreAllUserVectorized; |
1483 | auto *IdxLaneI = dyn_cast<Instruction>(IdxLaneV); |
1484 | if (!IdxLaneI || !isa<Instruction>(OpIdxLaneV)) |
1485 | return 0; |
1486 | return R.areAllUsersVectorized(IdxLaneI, None) |
1487 | ? LookAheadHeuristics::ScoreAllUserVectorized |
1488 | : 0; |
1489 | } |
1490 | |
1491 | /// Score scaling factor for fully compatible instructions but with |
1492 | /// different number of external uses. Allows better selection of the |
1493 | /// instructions with less external uses. |
1494 | static const int ScoreScaleFactor = 10; |
1495 | |
1496 | /// \Returns the look-ahead score, which tells us how much the sub-trees |
1497 | /// rooted at \p LHS and \p RHS match, the more they match the higher the |
1498 | /// score. This helps break ties in an informed way when we cannot decide on |
1499 | /// the order of the operands by just considering the immediate |
1500 | /// predecessors. |
1501 | int getLookAheadScore(Value *LHS, Value *RHS, ArrayRef<Value *> MainAltOps, |
1502 | int Lane, unsigned OpIdx, unsigned Idx, |
1503 | bool &IsUsed) { |
1504 | LookAheadHeuristics LookAhead(DL, SE, R, getNumLanes(), |
1505 | LookAheadMaxDepth); |
1506 | // Keep track of the instruction stack as we recurse into the operands |
1507 | // during the look-ahead score exploration. |
1508 | int Score = |
1509 | LookAhead.getScoreAtLevelRec(LHS, RHS, /*U1=*/nullptr, /*U2=*/nullptr, |
1510 | /*CurrLevel=*/1, MainAltOps); |
1511 | if (Score) { |
1512 | int SplatScore = getSplatScore(Lane, OpIdx, Idx); |
1513 | if (Score <= -SplatScore) { |
1514 | // Set the minimum score for splat-like sequence to avoid setting |
1515 | // failed state. |
1516 | Score = 1; |
1517 | } else { |
1518 | Score += SplatScore; |
1519 | // Scale score to see the difference between different operands |
1520 | // and similar operands but all vectorized/not all vectorized |
1521 | // uses. It does not affect actual selection of the best |
1522 | // compatible operand in general, just allows to select the |
1523 | // operand with all vectorized uses. |
1524 | Score *= ScoreScaleFactor; |
1525 | Score += getExternalUseScore(Lane, OpIdx, Idx); |
1526 | IsUsed = true; |
1527 | } |
1528 | } |
1529 | return Score; |
1530 | } |
1531 | |
1532 | /// Best defined scores per lanes between the passes. Used to choose the |
1533 | /// best operand (with the highest score) between the passes. |
1534 | /// The key - {Operand Index, Lane}. |
1535 | /// The value - the best score between the passes for the lane and the |
1536 | /// operand. |
1537 | SmallDenseMap<std::pair<unsigned, unsigned>, unsigned, 8> |
1538 | BestScoresPerLanes; |
1539 | |
1540 | // Search all operands in Ops[*][Lane] for the one that matches best |
1541 | // Ops[OpIdx][LastLane] and return its opreand index. |
1542 | // If no good match can be found, return None. |
1543 | Optional<unsigned> getBestOperand(unsigned OpIdx, int Lane, int LastLane, |
1544 | ArrayRef<ReorderingMode> ReorderingModes, |
1545 | ArrayRef<Value *> MainAltOps) { |
1546 | unsigned NumOperands = getNumOperands(); |
1547 | |
1548 | // The operand of the previous lane at OpIdx. |
1549 | Value *OpLastLane = getData(OpIdx, LastLane).V; |
1550 | |
1551 | // Our strategy mode for OpIdx. |
1552 | ReorderingMode RMode = ReorderingModes[OpIdx]; |
1553 | if (RMode == ReorderingMode::Failed) |
1554 | return None; |
1555 | |
1556 | // The linearized opcode of the operand at OpIdx, Lane. |
1557 | bool OpIdxAPO = getData(OpIdx, Lane).APO; |
1558 | |
1559 | // The best operand index and its score. |
1560 | // Sometimes we have more than one option (e.g., Opcode and Undefs), so we |
1561 | // are using the score to differentiate between the two. |
1562 | struct BestOpData { |
1563 | Optional<unsigned> Idx = None; |
1564 | unsigned Score = 0; |
1565 | } BestOp; |
1566 | BestOp.Score = |
1567 | BestScoresPerLanes.try_emplace(std::make_pair(OpIdx, Lane), 0) |
1568 | .first->second; |
1569 | |
1570 | // Track if the operand must be marked as used. If the operand is set to |
1571 | // Score 1 explicitly (because of non power-of-2 unique scalars, we may |
1572 | // want to reestimate the operands again on the following iterations). |
1573 | bool IsUsed = |
1574 | RMode == ReorderingMode::Splat || RMode == ReorderingMode::Constant; |
1575 | // Iterate through all unused operands and look for the best. |
1576 | for (unsigned Idx = 0; Idx != NumOperands; ++Idx) { |
1577 | // Get the operand at Idx and Lane. |
1578 | OperandData &OpData = getData(Idx, Lane); |
1579 | Value *Op = OpData.V; |
1580 | bool OpAPO = OpData.APO; |
1581 | |
1582 | // Skip already selected operands. |
1583 | if (OpData.IsUsed) |
1584 | continue; |
1585 | |
1586 | // Skip if we are trying to move the operand to a position with a |
1587 | // different opcode in the linearized tree form. This would break the |
1588 | // semantics. |
1589 | if (OpAPO != OpIdxAPO) |
1590 | continue; |
1591 | |
1592 | // Look for an operand that matches the current mode. |
1593 | switch (RMode) { |
1594 | case ReorderingMode::Load: |
1595 | case ReorderingMode::Constant: |
1596 | case ReorderingMode::Opcode: { |
1597 | bool LeftToRight = Lane > LastLane; |
1598 | Value *OpLeft = (LeftToRight) ? OpLastLane : Op; |
1599 | Value *OpRight = (LeftToRight) ? Op : OpLastLane; |
1600 | int Score = getLookAheadScore(OpLeft, OpRight, MainAltOps, Lane, |
1601 | OpIdx, Idx, IsUsed); |
1602 | if (Score > static_cast<int>(BestOp.Score)) { |
1603 | BestOp.Idx = Idx; |
1604 | BestOp.Score = Score; |
1605 | BestScoresPerLanes[std::make_pair(OpIdx, Lane)] = Score; |
1606 | } |
1607 | break; |
1608 | } |
1609 | case ReorderingMode::Splat: |
1610 | if (Op == OpLastLane) |
1611 | BestOp.Idx = Idx; |
1612 | break; |
1613 | case ReorderingMode::Failed: |
1614 | llvm_unreachable("Not expected Failed reordering mode.")::llvm::llvm_unreachable_internal("Not expected Failed reordering mode." , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 1614); |
1615 | } |
1616 | } |
1617 | |
1618 | if (BestOp.Idx) { |
1619 | getData(*BestOp.Idx, Lane).IsUsed = IsUsed; |
1620 | return BestOp.Idx; |
1621 | } |
1622 | // If we could not find a good match return None. |
1623 | return None; |
1624 | } |
1625 | |
1626 | /// Helper for reorderOperandVecs. |
1627 | /// \returns the lane that we should start reordering from. This is the one |
1628 | /// which has the least number of operands that can freely move about or |
1629 | /// less profitable because it already has the most optimal set of operands. |
1630 | unsigned getBestLaneToStartReordering() const { |
1631 | unsigned Min = UINT_MAX(2147483647 *2U +1U); |
1632 | unsigned SameOpNumber = 0; |
1633 | // std::pair<unsigned, unsigned> is used to implement a simple voting |
1634 | // algorithm and choose the lane with the least number of operands that |
1635 | // can freely move about or less profitable because it already has the |
1636 | // most optimal set of operands. The first unsigned is a counter for |
1637 | // voting, the second unsigned is the counter of lanes with instructions |
1638 | // with same/alternate opcodes and same parent basic block. |
1639 | MapVector<unsigned, std::pair<unsigned, unsigned>> HashMap; |
1640 | // Try to be closer to the original results, if we have multiple lanes |
1641 | // with same cost. If 2 lanes have the same cost, use the one with the |
1642 | // lowest index. |
1643 | for (int I = getNumLanes(); I > 0; --I) { |
1644 | unsigned Lane = I - 1; |
1645 | OperandsOrderData NumFreeOpsHash = |
1646 | getMaxNumOperandsThatCanBeReordered(Lane); |
1647 | // Compare the number of operands that can move and choose the one with |
1648 | // the least number. |
1649 | if (NumFreeOpsHash.NumOfAPOs < Min) { |
1650 | Min = NumFreeOpsHash.NumOfAPOs; |
1651 | SameOpNumber = NumFreeOpsHash.NumOpsWithSameOpcodeParent; |
1652 | HashMap.clear(); |
1653 | HashMap[NumFreeOpsHash.Hash] = std::make_pair(1, Lane); |
1654 | } else if (NumFreeOpsHash.NumOfAPOs == Min && |
1655 | NumFreeOpsHash.NumOpsWithSameOpcodeParent < SameOpNumber) { |
1656 | // Select the most optimal lane in terms of number of operands that |
1657 | // should be moved around. |
1658 | SameOpNumber = NumFreeOpsHash.NumOpsWithSameOpcodeParent; |
1659 | HashMap[NumFreeOpsHash.Hash] = std::make_pair(1, Lane); |
1660 | } else if (NumFreeOpsHash.NumOfAPOs == Min && |
1661 | NumFreeOpsHash.NumOpsWithSameOpcodeParent == SameOpNumber) { |
1662 | auto It = HashMap.find(NumFreeOpsHash.Hash); |
1663 | if (It == HashMap.end()) |
1664 | HashMap[NumFreeOpsHash.Hash] = std::make_pair(1, Lane); |
1665 | else |
1666 | ++It->second.first; |
1667 | } |
1668 | } |
1669 | // Select the lane with the minimum counter. |
1670 | unsigned BestLane = 0; |
1671 | unsigned CntMin = UINT_MAX(2147483647 *2U +1U); |
1672 | for (const auto &Data : reverse(HashMap)) { |
1673 | if (Data.second.first < CntMin) { |
1674 | CntMin = Data.second.first; |
1675 | BestLane = Data.second.second; |
1676 | } |
1677 | } |
1678 | return BestLane; |
1679 | } |
1680 | |
1681 | /// Data structure that helps to reorder operands. |
1682 | struct OperandsOrderData { |
1683 | /// The best number of operands with the same APOs, which can be |
1684 | /// reordered. |
1685 | unsigned NumOfAPOs = UINT_MAX(2147483647 *2U +1U); |
1686 | /// Number of operands with the same/alternate instruction opcode and |
1687 | /// parent. |
1688 | unsigned NumOpsWithSameOpcodeParent = 0; |
1689 | /// Hash for the actual operands ordering. |
1690 | /// Used to count operands, actually their position id and opcode |
1691 | /// value. It is used in the voting mechanism to find the lane with the |
1692 | /// least number of operands that can freely move about or less profitable |
1693 | /// because it already has the most optimal set of operands. Can be |
1694 | /// replaced with SmallVector<unsigned> instead but hash code is faster |
1695 | /// and requires less memory. |
1696 | unsigned Hash = 0; |
1697 | }; |
1698 | /// \returns the maximum number of operands that are allowed to be reordered |
1699 | /// for \p Lane and the number of compatible instructions(with the same |
1700 | /// parent/opcode). This is used as a heuristic for selecting the first lane |
1701 | /// to start operand reordering. |
1702 | OperandsOrderData getMaxNumOperandsThatCanBeReordered(unsigned Lane) const { |
1703 | unsigned CntTrue = 0; |
1704 | unsigned NumOperands = getNumOperands(); |
1705 | // Operands with the same APO can be reordered. We therefore need to count |
1706 | // how many of them we have for each APO, like this: Cnt[APO] = x. |
1707 | // Since we only have two APOs, namely true and false, we can avoid using |
1708 | // a map. Instead we can simply count the number of operands that |
1709 | // correspond to one of them (in this case the 'true' APO), and calculate |
1710 | // the other by subtracting it from the total number of operands. |
1711 | // Operands with the same instruction opcode and parent are more |
1712 | // profitable since we don't need to move them in many cases, with a high |
1713 | // probability such lane already can be vectorized effectively. |
1714 | bool AllUndefs = true; |
1715 | unsigned NumOpsWithSameOpcodeParent = 0; |
1716 | Instruction *OpcodeI = nullptr; |
1717 | BasicBlock *Parent = nullptr; |
1718 | unsigned Hash = 0; |
1719 | for (unsigned OpIdx = 0; OpIdx != NumOperands; ++OpIdx) { |
1720 | const OperandData &OpData = getData(OpIdx, Lane); |
1721 | if (OpData.APO) |
1722 | ++CntTrue; |
1723 | // Use Boyer-Moore majority voting for finding the majority opcode and |
1724 | // the number of times it occurs. |
1725 | if (auto *I = dyn_cast<Instruction>(OpData.V)) { |
1726 | if (!OpcodeI || !getSameOpcode({OpcodeI, I}).getOpcode() || |
1727 | I->getParent() != Parent) { |
1728 | if (NumOpsWithSameOpcodeParent == 0) { |
1729 | NumOpsWithSameOpcodeParent = 1; |
1730 | OpcodeI = I; |
1731 | Parent = I->getParent(); |
1732 | } else { |
1733 | --NumOpsWithSameOpcodeParent; |
1734 | } |
1735 | } else { |
1736 | ++NumOpsWithSameOpcodeParent; |
1737 | } |
1738 | } |
1739 | Hash = hash_combine( |
1740 | Hash, hash_value((OpIdx + 1) * (OpData.V->getValueID() + 1))); |
1741 | AllUndefs = AllUndefs && isa<UndefValue>(OpData.V); |
1742 | } |
1743 | if (AllUndefs) |
1744 | return {}; |
1745 | OperandsOrderData Data; |
1746 | Data.NumOfAPOs = std::max(CntTrue, NumOperands - CntTrue); |
1747 | Data.NumOpsWithSameOpcodeParent = NumOpsWithSameOpcodeParent; |
1748 | Data.Hash = Hash; |
1749 | return Data; |
1750 | } |
1751 | |
1752 | /// Go through the instructions in VL and append their operands. |
1753 | void appendOperandsOfVL(ArrayRef<Value *> VL) { |
1754 | 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", 1754, __extension__ __PRETTY_FUNCTION__)); |
1755 | 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", 1756, __extension__ __PRETTY_FUNCTION__)) |
1756 | "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", 1756, __extension__ __PRETTY_FUNCTION__)); |
1757 | 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", 1757, __extension__ __PRETTY_FUNCTION__)); |
1758 | unsigned NumOperands = cast<Instruction>(VL[0])->getNumOperands(); |
1759 | OpsVec.resize(NumOperands); |
1760 | unsigned NumLanes = VL.size(); |
1761 | for (unsigned OpIdx = 0; OpIdx != NumOperands; ++OpIdx) { |
1762 | OpsVec[OpIdx].resize(NumLanes); |
1763 | for (unsigned Lane = 0; Lane != NumLanes; ++Lane) { |
1764 | 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", 1764, __extension__ __PRETTY_FUNCTION__)); |
1765 | // Our tree has just 3 nodes: the root and two operands. |
1766 | // It is therefore trivial to get the APO. We only need to check the |
1767 | // opcode of VL[Lane] and whether the operand at OpIdx is the LHS or |
1768 | // RHS operand. The LHS operand of both add and sub is never attached |
1769 | // to an inversese operation in the linearized form, therefore its APO |
1770 | // is false. The RHS is true only if VL[Lane] is an inverse operation. |
1771 | |
1772 | // Since operand reordering is performed on groups of commutative |
1773 | // operations or alternating sequences (e.g., +, -), we can safely |
1774 | // tell the inverse operations by checking commutativity. |
1775 | bool IsInverseOperation = !isCommutative(cast<Instruction>(VL[Lane])); |
1776 | bool APO = (OpIdx == 0) ? false : IsInverseOperation; |
1777 | OpsVec[OpIdx][Lane] = {cast<Instruction>(VL[Lane])->getOperand(OpIdx), |
1778 | APO, false}; |
1779 | } |
1780 | } |
1781 | } |
1782 | |
1783 | /// \returns the number of operands. |
1784 | unsigned getNumOperands() const { return OpsVec.size(); } |
1785 | |
1786 | /// \returns the number of lanes. |
1787 | unsigned getNumLanes() const { return OpsVec[0].size(); } |
1788 | |
1789 | /// \returns the operand value at \p OpIdx and \p Lane. |
1790 | Value *getValue(unsigned OpIdx, unsigned Lane) const { |
1791 | return getData(OpIdx, Lane).V; |
1792 | } |
1793 | |
1794 | /// \returns true if the data structure is empty. |
1795 | bool empty() const { return OpsVec.empty(); } |
1796 | |
1797 | /// Clears the data. |
1798 | void clear() { OpsVec.clear(); } |
1799 | |
1800 | /// \Returns true if there are enough operands identical to \p Op to fill |
1801 | /// the whole vector. |
1802 | /// Note: This modifies the 'IsUsed' flag, so a cleanUsed() must follow. |
1803 | bool shouldBroadcast(Value *Op, unsigned OpIdx, unsigned Lane) { |
1804 | bool OpAPO = getData(OpIdx, Lane).APO; |
1805 | for (unsigned Ln = 0, Lns = getNumLanes(); Ln != Lns; ++Ln) { |
1806 | if (Ln == Lane) |
1807 | continue; |
1808 | // This is set to true if we found a candidate for broadcast at Lane. |
1809 | bool FoundCandidate = false; |
1810 | for (unsigned OpI = 0, OpE = getNumOperands(); OpI != OpE; ++OpI) { |
1811 | OperandData &Data = getData(OpI, Ln); |
1812 | if (Data.APO != OpAPO || Data.IsUsed) |
1813 | continue; |
1814 | if (Data.V == Op) { |
1815 | FoundCandidate = true; |
1816 | Data.IsUsed = true; |
1817 | break; |
1818 | } |
1819 | } |
1820 | if (!FoundCandidate) |
1821 | return false; |
1822 | } |
1823 | return true; |
1824 | } |
1825 | |
1826 | public: |
1827 | /// Initialize with all the operands of the instruction vector \p RootVL. |
1828 | VLOperands(ArrayRef<Value *> RootVL, const DataLayout &DL, |
1829 | ScalarEvolution &SE, const BoUpSLP &R) |
1830 | : DL(DL), SE(SE), R(R) { |
1831 | // Append all the operands of RootVL. |
1832 | appendOperandsOfVL(RootVL); |
1833 | } |
1834 | |
1835 | /// \Returns a value vector with the operands across all lanes for the |
1836 | /// opearnd at \p OpIdx. |
1837 | ValueList getVL(unsigned OpIdx) const { |
1838 | ValueList OpVL(OpsVec[OpIdx].size()); |
1839 | 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", 1840, __extension__ __PRETTY_FUNCTION__)) |
1840 | "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", 1840, __extension__ __PRETTY_FUNCTION__)); |
1841 | for (unsigned Lane = 0, Lanes = getNumLanes(); Lane != Lanes; ++Lane) |
1842 | OpVL[Lane] = OpsVec[OpIdx][Lane].V; |
1843 | return OpVL; |
1844 | } |
1845 | |
1846 | // Performs operand reordering for 2 or more operands. |
1847 | // The original operands are in OrigOps[OpIdx][Lane]. |
1848 | // The reordered operands are returned in 'SortedOps[OpIdx][Lane]'. |
1849 | void reorder() { |
1850 | unsigned NumOperands = getNumOperands(); |
1851 | unsigned NumLanes = getNumLanes(); |
1852 | // Each operand has its own mode. We are using this mode to help us select |
1853 | // the instructions for each lane, so that they match best with the ones |
1854 | // we have selected so far. |
1855 | SmallVector<ReorderingMode, 2> ReorderingModes(NumOperands); |
1856 | |
1857 | // This is a greedy single-pass algorithm. We are going over each lane |
1858 | // once and deciding on the best order right away with no back-tracking. |
1859 | // However, in order to increase its effectiveness, we start with the lane |
1860 | // that has operands that can move the least. For example, given the |
1861 | // following lanes: |
1862 | // Lane 0 : A[0] = B[0] + C[0] // Visited 3rd |
1863 | // Lane 1 : A[1] = C[1] - B[1] // Visited 1st |
1864 | // Lane 2 : A[2] = B[2] + C[2] // Visited 2nd |
1865 | // Lane 3 : A[3] = C[3] - B[3] // Visited 4th |
1866 | // we will start at Lane 1, since the operands of the subtraction cannot |
1867 | // be reordered. Then we will visit the rest of the lanes in a circular |
1868 | // fashion. That is, Lanes 2, then Lane 0, and finally Lane 3. |
1869 | |
1870 | // Find the first lane that we will start our search from. |
1871 | unsigned FirstLane = getBestLaneToStartReordering(); |
1872 | |
1873 | // Initialize the modes. |
1874 | for (unsigned OpIdx = 0; OpIdx != NumOperands; ++OpIdx) { |
1875 | Value *OpLane0 = getValue(OpIdx, FirstLane); |
1876 | // Keep track if we have instructions with all the same opcode on one |
1877 | // side. |
1878 | if (isa<LoadInst>(OpLane0)) |
1879 | ReorderingModes[OpIdx] = ReorderingMode::Load; |
1880 | else if (isa<Instruction>(OpLane0)) { |
1881 | // Check if OpLane0 should be broadcast. |
1882 | if (shouldBroadcast(OpLane0, OpIdx, FirstLane)) |
1883 | ReorderingModes[OpIdx] = ReorderingMode::Splat; |
1884 | else |
1885 | ReorderingModes[OpIdx] = ReorderingMode::Opcode; |
1886 | } |
1887 | else if (isa<Constant>(OpLane0)) |
1888 | ReorderingModes[OpIdx] = ReorderingMode::Constant; |
1889 | else if (isa<Argument>(OpLane0)) |
1890 | // Our best hope is a Splat. It may save some cost in some cases. |
1891 | ReorderingModes[OpIdx] = ReorderingMode::Splat; |
1892 | else |
1893 | // NOTE: This should be unreachable. |
1894 | ReorderingModes[OpIdx] = ReorderingMode::Failed; |
1895 | } |
1896 | |
1897 | // Check that we don't have same operands. No need to reorder if operands |
1898 | // are just perfect diamond or shuffled diamond match. Do not do it only |
1899 | // for possible broadcasts or non-power of 2 number of scalars (just for |
1900 | // now). |
1901 | auto &&SkipReordering = [this]() { |
1902 | SmallPtrSet<Value *, 4> UniqueValues; |
1903 | ArrayRef<OperandData> Op0 = OpsVec.front(); |
1904 | for (const OperandData &Data : Op0) |
1905 | UniqueValues.insert(Data.V); |
1906 | for (ArrayRef<OperandData> Op : drop_begin(OpsVec, 1)) { |
1907 | if (any_of(Op, [&UniqueValues](const OperandData &Data) { |
1908 | return !UniqueValues.contains(Data.V); |
1909 | })) |
1910 | return false; |
1911 | } |
1912 | // TODO: Check if we can remove a check for non-power-2 number of |
1913 | // scalars after full support of non-power-2 vectorization. |
1914 | return UniqueValues.size() != 2 && isPowerOf2_32(UniqueValues.size()); |
1915 | }; |
1916 | |
1917 | // If the initial strategy fails for any of the operand indexes, then we |
1918 | // perform reordering again in a second pass. This helps avoid assigning |
1919 | // high priority to the failed strategy, and should improve reordering for |
1920 | // the non-failed operand indexes. |
1921 | for (int Pass = 0; Pass != 2; ++Pass) { |
1922 | // Check if no need to reorder operands since they're are perfect or |
1923 | // shuffled diamond match. |
1924 | // Need to to do it to avoid extra external use cost counting for |
1925 | // shuffled matches, which may cause regressions. |
1926 | if (SkipReordering()) |
1927 | break; |
1928 | // Skip the second pass if the first pass did not fail. |
1929 | bool StrategyFailed = false; |
1930 | // Mark all operand data as free to use. |
1931 | clearUsed(); |
1932 | // We keep the original operand order for the FirstLane, so reorder the |
1933 | // rest of the lanes. We are visiting the nodes in a circular fashion, |
1934 | // using FirstLane as the center point and increasing the radius |
1935 | // distance. |
1936 | SmallVector<SmallVector<Value *, 2>> MainAltOps(NumOperands); |
1937 | for (unsigned I = 0; I < NumOperands; ++I) |
1938 | MainAltOps[I].push_back(getData(I, FirstLane).V); |
1939 | |
1940 | for (unsigned Distance = 1; Distance != NumLanes; ++Distance) { |
1941 | // Visit the lane on the right and then the lane on the left. |
1942 | for (int Direction : {+1, -1}) { |
1943 | int Lane = FirstLane + Direction * Distance; |
1944 | if (Lane < 0 || Lane >= (int)NumLanes) |
1945 | continue; |
1946 | int LastLane = Lane - Direction; |
1947 | 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", 1948, __extension__ __PRETTY_FUNCTION__)) |
1948 | "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", 1948, __extension__ __PRETTY_FUNCTION__)); |
1949 | // Look for a good match for each operand. |
1950 | for (unsigned OpIdx = 0; OpIdx != NumOperands; ++OpIdx) { |
1951 | // Search for the operand that matches SortedOps[OpIdx][Lane-1]. |
1952 | Optional<unsigned> BestIdx = getBestOperand( |
1953 | OpIdx, Lane, LastLane, ReorderingModes, MainAltOps[OpIdx]); |
1954 | // By not selecting a value, we allow the operands that follow to |
1955 | // select a better matching value. We will get a non-null value in |
1956 | // the next run of getBestOperand(). |
1957 | if (BestIdx) { |
1958 | // Swap the current operand with the one returned by |
1959 | // getBestOperand(). |
1960 | swap(OpIdx, *BestIdx, Lane); |
1961 | } else { |
1962 | // We failed to find a best operand, set mode to 'Failed'. |
1963 | ReorderingModes[OpIdx] = ReorderingMode::Failed; |
1964 | // Enable the second pass. |
1965 | StrategyFailed = true; |
1966 | } |
1967 | // Try to get the alternate opcode and follow it during analysis. |
1968 | if (MainAltOps[OpIdx].size() != 2) { |
1969 | OperandData &AltOp = getData(OpIdx, Lane); |
1970 | InstructionsState OpS = |
1971 | getSameOpcode({MainAltOps[OpIdx].front(), AltOp.V}); |
1972 | if (OpS.getOpcode() && OpS.isAltShuffle()) |
1973 | MainAltOps[OpIdx].push_back(AltOp.V); |
1974 | } |
1975 | } |
1976 | } |
1977 | } |
1978 | // Skip second pass if the strategy did not fail. |
1979 | if (!StrategyFailed) |
1980 | break; |
1981 | } |
1982 | } |
1983 | |
1984 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) |
1985 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) static StringRef getModeStr(ReorderingMode RMode) { |
1986 | switch (RMode) { |
1987 | case ReorderingMode::Load: |
1988 | return "Load"; |
1989 | case ReorderingMode::Opcode: |
1990 | return "Opcode"; |
1991 | case ReorderingMode::Constant: |
1992 | return "Constant"; |
1993 | case ReorderingMode::Splat: |
1994 | return "Splat"; |
1995 | case ReorderingMode::Failed: |
1996 | return "Failed"; |
1997 | } |
1998 | llvm_unreachable("Unimplemented Reordering Type")::llvm::llvm_unreachable_internal("Unimplemented Reordering Type" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 1998); |
1999 | } |
2000 | |
2001 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) static raw_ostream &printMode(ReorderingMode RMode, |
2002 | raw_ostream &OS) { |
2003 | return OS << getModeStr(RMode); |
2004 | } |
2005 | |
2006 | /// Debug print. |
2007 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) static void dumpMode(ReorderingMode RMode) { |
2008 | printMode(RMode, dbgs()); |
2009 | } |
2010 | |
2011 | friend raw_ostream &operator<<(raw_ostream &OS, ReorderingMode RMode) { |
2012 | return printMode(RMode, OS); |
2013 | } |
2014 | |
2015 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) raw_ostream &print(raw_ostream &OS) const { |
2016 | const unsigned Indent = 2; |
2017 | unsigned Cnt = 0; |
2018 | for (const OperandDataVec &OpDataVec : OpsVec) { |
2019 | OS << "Operand " << Cnt++ << "\n"; |
2020 | for (const OperandData &OpData : OpDataVec) { |
2021 | OS.indent(Indent) << "{"; |
2022 | if (Value *V = OpData.V) |
2023 | OS << *V; |
2024 | else |
2025 | OS << "null"; |
2026 | OS << ", APO:" << OpData.APO << "}\n"; |
2027 | } |
2028 | OS << "\n"; |
2029 | } |
2030 | return OS; |
2031 | } |
2032 | |
2033 | /// Debug print. |
2034 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void dump() const { print(dbgs()); } |
2035 | #endif |
2036 | }; |
2037 | |
2038 | /// Evaluate each pair in \p Candidates and return index into \p Candidates |
2039 | /// for a pair which have highest score deemed to have best chance to form |
2040 | /// root of profitable tree to vectorize. Return None if no candidate scored |
2041 | /// above the LookAheadHeuristics::ScoreFail. |
2042 | /// \param Limit Lower limit of the cost, considered to be good enough score. |
2043 | Optional<int> |
2044 | findBestRootPair(ArrayRef<std::pair<Value *, Value *>> Candidates, |
2045 | int Limit = LookAheadHeuristics::ScoreFail) { |
2046 | LookAheadHeuristics LookAhead(*DL, *SE, *this, /*NumLanes=*/2, |
2047 | RootLookAheadMaxDepth); |
2048 | int BestScore = Limit; |
2049 | Optional<int> Index; |
2050 | for (int I : seq<int>(0, Candidates.size())) { |
2051 | int Score = LookAhead.getScoreAtLevelRec(Candidates[I].first, |
2052 | Candidates[I].second, |
2053 | /*U1=*/nullptr, /*U2=*/nullptr, |
2054 | /*Level=*/1, None); |
2055 | if (Score > BestScore) { |
2056 | BestScore = Score; |
2057 | Index = I; |
2058 | } |
2059 | } |
2060 | return Index; |
2061 | } |
2062 | |
2063 | /// Checks if the instruction is marked for deletion. |
2064 | bool isDeleted(Instruction *I) const { return DeletedInstructions.count(I); } |
2065 | |
2066 | /// Removes an instruction from its block and eventually deletes it. |
2067 | /// It's like Instruction::eraseFromParent() except that the actual deletion |
2068 | /// is delayed until BoUpSLP is destructed. |
2069 | void eraseInstruction(Instruction *I) { |
2070 | DeletedInstructions.insert(I); |
2071 | } |
2072 | |
2073 | /// Checks if the instruction was already analyzed for being possible |
2074 | /// reduction root. |
2075 | bool isAnalyzedReductionRoot(Instruction *I) const { |
2076 | return AnalyzedReductionsRoots.count(I); |
2077 | } |
2078 | /// Register given instruction as already analyzed for being possible |
2079 | /// reduction root. |
2080 | void analyzedReductionRoot(Instruction *I) { |
2081 | AnalyzedReductionsRoots.insert(I); |
2082 | } |
2083 | /// Checks if the provided list of reduced values was checked already for |
2084 | /// vectorization. |
2085 | bool areAnalyzedReductionVals(ArrayRef<Value *> VL) { |
2086 | return AnalyzedReductionVals.contains(hash_value(VL)); |
2087 | } |
2088 | /// Adds the list of reduced values to list of already checked values for the |
2089 | /// vectorization. |
2090 | void analyzedReductionVals(ArrayRef<Value *> VL) { |
2091 | AnalyzedReductionVals.insert(hash_value(VL)); |
2092 | } |
2093 | /// Clear the list of the analyzed reduction root instructions. |
2094 | void clearReductionData() { |
2095 | AnalyzedReductionsRoots.clear(); |
2096 | AnalyzedReductionVals.clear(); |
2097 | } |
2098 | /// Checks if the given value is gathered in one of the nodes. |
2099 | bool isAnyGathered(const SmallDenseSet<Value *> &Vals) const { |
2100 | return any_of(MustGather, [&](Value *V) { return Vals.contains(V); }); |
2101 | } |
2102 | |
2103 | ~BoUpSLP(); |
2104 | |
2105 | private: |
2106 | /// Check if the operands on the edges \p Edges of the \p UserTE allows |
2107 | /// reordering (i.e. the operands can be reordered because they have only one |
2108 | /// user and reordarable). |
2109 | /// \param ReorderableGathers List of all gather nodes that require reordering |
2110 | /// (e.g., gather of extractlements or partially vectorizable loads). |
2111 | /// \param GatherOps List of gather operand nodes for \p UserTE that require |
2112 | /// reordering, subset of \p NonVectorized. |
2113 | bool |
2114 | canReorderOperands(TreeEntry *UserTE, |
2115 | SmallVectorImpl<std::pair<unsigned, TreeEntry *>> &Edges, |
2116 | ArrayRef<TreeEntry *> ReorderableGathers, |
2117 | SmallVectorImpl<TreeEntry *> &GatherOps); |
2118 | |
2119 | /// Checks if the given \p TE is a gather node with clustered reused scalars |
2120 | /// and reorders it per given \p Mask. |
2121 | void reorderNodeWithReuses(TreeEntry &TE, ArrayRef<int> Mask) const; |
2122 | |
2123 | /// Returns vectorized operand \p OpIdx of the node \p UserTE from the graph, |
2124 | /// if any. If it is not vectorized (gather node), returns nullptr. |
2125 | TreeEntry *getVectorizedOperand(TreeEntry *UserTE, unsigned OpIdx) { |
2126 | ArrayRef<Value *> VL = UserTE->getOperand(OpIdx); |
2127 | TreeEntry *TE = nullptr; |
2128 | const auto *It = find_if(VL, [this, &TE](Value *V) { |
2129 | TE = getTreeEntry(V); |
2130 | return TE; |
2131 | }); |
2132 | if (It != VL.end() && TE->isSame(VL)) |
2133 | return TE; |
2134 | return nullptr; |
2135 | } |
2136 | |
2137 | /// Returns vectorized operand \p OpIdx of the node \p UserTE from the graph, |
2138 | /// if any. If it is not vectorized (gather node), returns nullptr. |
2139 | const TreeEntry *getVectorizedOperand(const TreeEntry *UserTE, |
2140 | unsigned OpIdx) const { |
2141 | return const_cast<BoUpSLP *>(this)->getVectorizedOperand( |
2142 | const_cast<TreeEntry *>(UserTE), OpIdx); |
2143 | } |
2144 | |
2145 | /// Checks if all users of \p I are the part of the vectorization tree. |
2146 | bool areAllUsersVectorized(Instruction *I, |
2147 | ArrayRef<Value *> VectorizedVals) const; |
2148 | |
2149 | /// Return information about the vector formed for the specified index |
2150 | /// of a vector of (the same) instruction. |
2151 | TargetTransformInfo::OperandValueInfo getOperandInfo(ArrayRef<Value *> VL, |
2152 | unsigned OpIdx); |
2153 | |
2154 | /// \returns the cost of the vectorizable entry. |
2155 | InstructionCost getEntryCost(const TreeEntry *E, |
2156 | ArrayRef<Value *> VectorizedVals); |
2157 | |
2158 | /// This is the recursive part of buildTree. |
2159 | void buildTree_rec(ArrayRef<Value *> Roots, unsigned Depth, |
2160 | const EdgeInfo &EI); |
2161 | |
2162 | /// \returns true if the ExtractElement/ExtractValue instructions in \p VL can |
2163 | /// be vectorized to use the original vector (or aggregate "bitcast" to a |
2164 | /// vector) and sets \p CurrentOrder to the identity permutation; otherwise |
2165 | /// returns false, setting \p CurrentOrder to either an empty vector or a |
2166 | /// non-identity permutation that allows to reuse extract instructions. |
2167 | bool canReuseExtract(ArrayRef<Value *> VL, Value *OpValue, |
2168 | SmallVectorImpl<unsigned> &CurrentOrder) const; |
2169 | |
2170 | /// Vectorize a single entry in the tree. |
2171 | Value *vectorizeTree(TreeEntry *E); |
2172 | |
2173 | /// Vectorize a single entry in the tree, starting in \p VL. |
2174 | Value *vectorizeTree(ArrayRef<Value *> VL); |
2175 | |
2176 | /// Create a new vector from a list of scalar values. Produces a sequence |
2177 | /// which exploits values reused across lanes, and arranges the inserts |
2178 | /// for ease of later optimization. |
2179 | Value *createBuildVector(ArrayRef<Value *> VL); |
2180 | |
2181 | /// \returns the scalarization cost for this type. Scalarization in this |
2182 | /// context means the creation of vectors from a group of scalars. If \p |
2183 | /// NeedToShuffle is true, need to add a cost of reshuffling some of the |
2184 | /// vector elements. |
2185 | InstructionCost getGatherCost(FixedVectorType *Ty, |
2186 | const APInt &ShuffledIndices, |
2187 | bool NeedToShuffle) const; |
2188 | |
2189 | /// Returns the instruction in the bundle, which can be used as a base point |
2190 | /// for scheduling. Usually it is the last instruction in the bundle, except |
2191 | /// for the case when all operands are external (in this case, it is the first |
2192 | /// instruction in the list). |
2193 | Instruction &getLastInstructionInBundle(const TreeEntry *E); |
2194 | |
2195 | /// Checks if the gathered \p VL can be represented as shuffle(s) of previous |
2196 | /// tree entries. |
2197 | /// \returns ShuffleKind, if gathered values can be represented as shuffles of |
2198 | /// previous tree entries. \p Mask is filled with the shuffle mask. |
2199 | Optional<TargetTransformInfo::ShuffleKind> |
2200 | isGatherShuffledEntry(const TreeEntry *TE, SmallVectorImpl<int> &Mask, |
2201 | SmallVectorImpl<const TreeEntry *> &Entries); |
2202 | |
2203 | /// \returns the scalarization cost for this list of values. Assuming that |
2204 | /// this subtree gets vectorized, we may need to extract the values from the |
2205 | /// roots. This method calculates the cost of extracting the values. |
2206 | InstructionCost getGatherCost(ArrayRef<Value *> VL) const; |
2207 | |
2208 | /// Set the Builder insert point to one after the last instruction in |
2209 | /// the bundle |
2210 | void setInsertPointAfterBundle(const TreeEntry *E); |
2211 | |
2212 | /// \returns a vector from a collection of scalars in \p VL. |
2213 | Value *gather(ArrayRef<Value *> VL); |
2214 | |
2215 | /// \returns whether the VectorizableTree is fully vectorizable and will |
2216 | /// be beneficial even the tree height is tiny. |
2217 | bool isFullyVectorizableTinyTree(bool ForReduction) const; |
2218 | |
2219 | /// Reorder commutative or alt operands to get better probability of |
2220 | /// generating vectorized code. |
2221 | static void reorderInputsAccordingToOpcode(ArrayRef<Value *> VL, |
2222 | SmallVectorImpl<Value *> &Left, |
2223 | SmallVectorImpl<Value *> &Right, |
2224 | const DataLayout &DL, |
2225 | ScalarEvolution &SE, |
2226 | const BoUpSLP &R); |
2227 | |
2228 | /// Helper for `findExternalStoreUsersReorderIndices()`. It iterates over the |
2229 | /// users of \p TE and collects the stores. It returns the map from the store |
2230 | /// pointers to the collected stores. |
2231 | DenseMap<Value *, SmallVector<StoreInst *, 4>> |
2232 | collectUserStores(const BoUpSLP::TreeEntry *TE) const; |
2233 | |
2234 | /// Helper for `findExternalStoreUsersReorderIndices()`. It checks if the |
2235 | /// stores in \p StoresVec can form a vector instruction. If so it returns true |
2236 | /// and populates \p ReorderIndices with the shuffle indices of the the stores |
2237 | /// when compared to the sorted vector. |
2238 | bool canFormVector(const SmallVector<StoreInst *, 4> &StoresVec, |
2239 | OrdersType &ReorderIndices) const; |
2240 | |
2241 | /// Iterates through the users of \p TE, looking for scalar stores that can be |
2242 | /// potentially vectorized in a future SLP-tree. If found, it keeps track of |
2243 | /// their order and builds an order index vector for each store bundle. It |
2244 | /// returns all these order vectors found. |
2245 | /// We run this after the tree has formed, otherwise we may come across user |
2246 | /// instructions that are not yet in the tree. |
2247 | SmallVector<OrdersType, 1> |
2248 | findExternalStoreUsersReorderIndices(TreeEntry *TE) const; |
2249 | |
2250 | struct TreeEntry { |
2251 | using VecTreeTy = SmallVector<std::unique_ptr<TreeEntry>, 8>; |
2252 | TreeEntry(VecTreeTy &Container) : Container(Container) {} |
2253 | |
2254 | /// \returns true if the scalars in VL are equal to this entry. |
2255 | bool isSame(ArrayRef<Value *> VL) const { |
2256 | auto &&IsSame = [VL](ArrayRef<Value *> Scalars, ArrayRef<int> Mask) { |
2257 | if (Mask.size() != VL.size() && VL.size() == Scalars.size()) |
2258 | return std::equal(VL.begin(), VL.end(), Scalars.begin()); |
2259 | return VL.size() == Mask.size() && |
2260 | std::equal(VL.begin(), VL.end(), Mask.begin(), |
2261 | [Scalars](Value *V, int Idx) { |
2262 | return (isa<UndefValue>(V) && |
2263 | Idx == UndefMaskElem) || |
2264 | (Idx != UndefMaskElem && V == Scalars[Idx]); |
2265 | }); |
2266 | }; |
2267 | if (!ReorderIndices.empty()) { |
2268 | // TODO: implement matching if the nodes are just reordered, still can |
2269 | // treat the vector as the same if the list of scalars matches VL |
2270 | // directly, without reordering. |
2271 | SmallVector<int> Mask; |
2272 | inversePermutation(ReorderIndices, Mask); |
2273 | if (VL.size() == Scalars.size()) |
2274 | return IsSame(Scalars, Mask); |
2275 | if (VL.size() == ReuseShuffleIndices.size()) { |
2276 | ::addMask(Mask, ReuseShuffleIndices); |
2277 | return IsSame(Scalars, Mask); |
2278 | } |
2279 | return false; |
2280 | } |
2281 | return IsSame(Scalars, ReuseShuffleIndices); |
2282 | } |
2283 | |
2284 | /// \returns true if current entry has same operands as \p TE. |
2285 | bool hasEqualOperands(const TreeEntry &TE) const { |
2286 | if (TE.getNumOperands() != getNumOperands()) |
2287 | return false; |
2288 | SmallBitVector Used(getNumOperands()); |
2289 | for (unsigned I = 0, E = getNumOperands(); I < E; ++I) { |
2290 | unsigned PrevCount = Used.count(); |
2291 | for (unsigned K = 0; K < E; ++K) { |
2292 | if (Used.test(K)) |
2293 | continue; |
2294 | if (getOperand(K) == TE.getOperand(I)) { |
2295 | Used.set(K); |
2296 | break; |
2297 | } |
2298 | } |
2299 | // Check if we actually found the matching operand. |
2300 | if (PrevCount == Used.count()) |
2301 | return false; |
2302 | } |
2303 | return true; |
2304 | } |
2305 | |
2306 | /// \return Final vectorization factor for the node. Defined by the total |
2307 | /// number of vectorized scalars, including those, used several times in the |
2308 | /// entry and counted in the \a ReuseShuffleIndices, if any. |
2309 | unsigned getVectorFactor() const { |
2310 | if (!ReuseShuffleIndices.empty()) |
2311 | return ReuseShuffleIndices.size(); |
2312 | return Scalars.size(); |
2313 | }; |
2314 | |
2315 | /// A vector of scalars. |
2316 | ValueList Scalars; |
2317 | |
2318 | /// The Scalars are vectorized into this value. It is initialized to Null. |
2319 | Value *VectorizedValue = nullptr; |
2320 | |
2321 | /// Do we need to gather this sequence or vectorize it |
2322 | /// (either with vector instruction or with scatter/gather |
2323 | /// intrinsics for store/load)? |
2324 | enum EntryState { Vectorize, ScatterVectorize, NeedToGather }; |
2325 | EntryState State; |
2326 | |
2327 | /// Does this sequence require some shuffling? |
2328 | SmallVector<int, 4> ReuseShuffleIndices; |
2329 | |
2330 | /// Does this entry require reordering? |
2331 | SmallVector<unsigned, 4> ReorderIndices; |
2332 | |
2333 | /// Points back to the VectorizableTree. |
2334 | /// |
2335 | /// Only used for Graphviz right now. Unfortunately GraphTrait::NodeRef has |
2336 | /// to be a pointer and needs to be able to initialize the child iterator. |
2337 | /// Thus we need a reference back to the container to translate the indices |
2338 | /// to entries. |
2339 | VecTreeTy &Container; |
2340 | |
2341 | /// The TreeEntry index containing the user of this entry. We can actually |
2342 | /// have multiple users so the data structure is not truly a tree. |
2343 | SmallVector<EdgeInfo, 1> UserTreeIndices; |
2344 | |
2345 | /// The index of this treeEntry in VectorizableTree. |
2346 | int Idx = -1; |
2347 | |
2348 | private: |
2349 | /// The operands of each instruction in each lane Operands[op_index][lane]. |
2350 | /// Note: This helps avoid the replication of the code that performs the |
2351 | /// reordering of operands during buildTree_rec() and vectorizeTree(). |
2352 | SmallVector<ValueList, 2> Operands; |
2353 | |
2354 | /// The main/alternate instruction. |
2355 | Instruction *MainOp = nullptr; |
2356 | Instruction *AltOp = nullptr; |
2357 | |
2358 | public: |
2359 | /// Set this bundle's \p OpIdx'th operand to \p OpVL. |
2360 | void setOperand(unsigned OpIdx, ArrayRef<Value *> OpVL) { |
2361 | if (Operands.size() < OpIdx + 1) |
2362 | Operands.resize(OpIdx + 1); |
2363 | 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", 2363, __extension__ __PRETTY_FUNCTION__)); |
2364 | 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", 2365, __extension__ __PRETTY_FUNCTION__)) |
2365 | "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", 2365, __extension__ __PRETTY_FUNCTION__)); |
2366 | Operands[OpIdx].resize(OpVL.size()); |
2367 | copy(OpVL, Operands[OpIdx].begin()); |
2368 | } |
2369 | |
2370 | /// Set the operands of this bundle in their original order. |
2371 | void setOperandsInOrder() { |
2372 | 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", 2372, __extension__ __PRETTY_FUNCTION__)); |
2373 | auto *I0 = cast<Instruction>(Scalars[0]); |
2374 | Operands.resize(I0->getNumOperands()); |
2375 | unsigned NumLanes = Scalars.size(); |
2376 | for (unsigned OpIdx = 0, NumOperands = I0->getNumOperands(); |
2377 | OpIdx != NumOperands; ++OpIdx) { |
2378 | Operands[OpIdx].resize(NumLanes); |
2379 | for (unsigned Lane = 0; Lane != NumLanes; ++Lane) { |
2380 | auto *I = cast<Instruction>(Scalars[Lane]); |
2381 | 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", 2382, __extension__ __PRETTY_FUNCTION__)) |
2382 | "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", 2382, __extension__ __PRETTY_FUNCTION__)); |
2383 | Operands[OpIdx][Lane] = I->getOperand(OpIdx); |
2384 | } |
2385 | } |
2386 | } |
2387 | |
2388 | /// Reorders operands of the node to the given mask \p Mask. |
2389 | void reorderOperands(ArrayRef<int> Mask) { |
2390 | for (ValueList &Operand : Operands) |
2391 | reorderScalars(Operand, Mask); |
2392 | } |
2393 | |
2394 | /// \returns the \p OpIdx operand of this TreeEntry. |
2395 | ValueList &getOperand(unsigned OpIdx) { |
2396 | 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", 2396, __extension__ __PRETTY_FUNCTION__)); |
2397 | return Operands[OpIdx]; |
2398 | } |
2399 | |
2400 | /// \returns the \p OpIdx operand of this TreeEntry. |
2401 | ArrayRef<Value *> getOperand(unsigned OpIdx) const { |
2402 | 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", 2402, __extension__ __PRETTY_FUNCTION__)); |
2403 | return Operands[OpIdx]; |
2404 | } |
2405 | |
2406 | /// \returns the number of operands. |
2407 | unsigned getNumOperands() const { return Operands.size(); } |
2408 | |
2409 | /// \return the single \p OpIdx operand. |
2410 | Value *getSingleOperand(unsigned OpIdx) const { |
2411 | 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", 2411, __extension__ __PRETTY_FUNCTION__)); |
2412 | 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", 2412, __extension__ __PRETTY_FUNCTION__)); |
2413 | return Operands[OpIdx][0]; |
2414 | } |
2415 | |
2416 | /// Some of the instructions in the list have alternate opcodes. |
2417 | bool isAltShuffle() const { return MainOp != AltOp; } |
2418 | |
2419 | bool isOpcodeOrAlt(Instruction *I) const { |
2420 | unsigned CheckedOpcode = I->getOpcode(); |
2421 | return (getOpcode() == CheckedOpcode || |
2422 | getAltOpcode() == CheckedOpcode); |
2423 | } |
2424 | |
2425 | /// Chooses the correct key for scheduling data. If \p Op has the same (or |
2426 | /// alternate) opcode as \p OpValue, the key is \p Op. Otherwise the key is |
2427 | /// \p OpValue. |
2428 | Value *isOneOf(Value *Op) const { |
2429 | auto *I = dyn_cast<Instruction>(Op); |
2430 | if (I && isOpcodeOrAlt(I)) |
2431 | return Op; |
2432 | return MainOp; |
2433 | } |
2434 | |
2435 | void setOperations(const InstructionsState &S) { |
2436 | MainOp = S.MainOp; |
2437 | AltOp = S.AltOp; |
2438 | } |
2439 | |
2440 | Instruction *getMainOp() const { |
2441 | return MainOp; |
2442 | } |
2443 | |
2444 | Instruction *getAltOp() const { |
2445 | return AltOp; |
2446 | } |
2447 | |
2448 | /// The main/alternate opcodes for the list of instructions. |
2449 | unsigned getOpcode() const { |
2450 | return MainOp ? MainOp->getOpcode() : 0; |
2451 | } |
2452 | |
2453 | unsigned getAltOpcode() const { |
2454 | return AltOp ? AltOp->getOpcode() : 0; |
2455 | } |
2456 | |
2457 | /// When ReuseReorderShuffleIndices is empty it just returns position of \p |
2458 | /// V within vector of Scalars. Otherwise, try to remap on its reuse index. |
2459 | int findLaneForValue(Value *V) const { |
2460 | unsigned FoundLane = std::distance(Scalars.begin(), find(Scalars, V)); |
2461 | 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", 2461, __extension__ __PRETTY_FUNCTION__)); |
2462 | if (!ReorderIndices.empty()) |
2463 | FoundLane = ReorderIndices[FoundLane]; |
2464 | 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", 2464, __extension__ __PRETTY_FUNCTION__)); |
2465 | if (!ReuseShuffleIndices.empty()) { |
2466 | FoundLane = std::distance(ReuseShuffleIndices.begin(), |
2467 | find(ReuseShuffleIndices, FoundLane)); |
2468 | } |
2469 | return FoundLane; |
2470 | } |
2471 | |
2472 | #ifndef NDEBUG |
2473 | /// Debug printer. |
2474 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void dump() const { |
2475 | dbgs() << Idx << ".\n"; |
2476 | for (unsigned OpI = 0, OpE = Operands.size(); OpI != OpE; ++OpI) { |
2477 | dbgs() << "Operand " << OpI << ":\n"; |
2478 | for (const Value *V : Operands[OpI]) |
2479 | dbgs().indent(2) << *V << "\n"; |
2480 | } |
2481 | dbgs() << "Scalars: \n"; |
2482 | for (Value *V : Scalars) |
2483 | dbgs().indent(2) << *V << "\n"; |
2484 | dbgs() << "State: "; |
2485 | switch (State) { |
2486 | case Vectorize: |
2487 | dbgs() << "Vectorize\n"; |
2488 | break; |
2489 | case ScatterVectorize: |
2490 | dbgs() << "ScatterVectorize\n"; |
2491 | break; |
2492 | case NeedToGather: |
2493 | dbgs() << "NeedToGather\n"; |
2494 | break; |
2495 | } |
2496 | dbgs() << "MainOp: "; |
2497 | if (MainOp) |
2498 | dbgs() << *MainOp << "\n"; |
2499 | else |
2500 | dbgs() << "NULL\n"; |
2501 | dbgs() << "AltOp: "; |
2502 | if (AltOp) |
2503 | dbgs() << *AltOp << "\n"; |
2504 | else |
2505 | dbgs() << "NULL\n"; |
2506 | dbgs() << "VectorizedValue: "; |
2507 | if (VectorizedValue) |
2508 | dbgs() << *VectorizedValue << "\n"; |
2509 | else |
2510 | dbgs() << "NULL\n"; |
2511 | dbgs() << "ReuseShuffleIndices: "; |
2512 | if (ReuseShuffleIndices.empty()) |
2513 | dbgs() << "Empty"; |
2514 | else |
2515 | for (int ReuseIdx : ReuseShuffleIndices) |
2516 | dbgs() << ReuseIdx << ", "; |
2517 | dbgs() << "\n"; |
2518 | dbgs() << "ReorderIndices: "; |
2519 | for (unsigned ReorderIdx : ReorderIndices) |
2520 | dbgs() << ReorderIdx << ", "; |
2521 | dbgs() << "\n"; |
2522 | dbgs() << "UserTreeIndices: "; |
2523 | for (const auto &EInfo : UserTreeIndices) |
2524 | dbgs() << EInfo << ", "; |
2525 | dbgs() << "\n"; |
2526 | } |
2527 | #endif |
2528 | }; |
2529 | |
2530 | #ifndef NDEBUG |
2531 | void dumpTreeCosts(const TreeEntry *E, InstructionCost ReuseShuffleCost, |
2532 | InstructionCost VecCost, |
2533 | InstructionCost ScalarCost) const { |
2534 | dbgs() << "SLP: Calculated costs for Tree:\n"; E->dump(); |
2535 | dbgs() << "SLP: Costs:\n"; |
2536 | dbgs() << "SLP: ReuseShuffleCost = " << ReuseShuffleCost << "\n"; |
2537 | dbgs() << "SLP: VectorCost = " << VecCost << "\n"; |
2538 | dbgs() << "SLP: ScalarCost = " << ScalarCost << "\n"; |
2539 | dbgs() << "SLP: ReuseShuffleCost + VecCost - ScalarCost = " << |
2540 | ReuseShuffleCost + VecCost - ScalarCost << "\n"; |
2541 | } |
2542 | #endif |
2543 | |
2544 | /// Create a new VectorizableTree entry. |
2545 | TreeEntry *newTreeEntry(ArrayRef<Value *> VL, Optional<ScheduleData *> Bundle, |
2546 | const InstructionsState &S, |
2547 | const EdgeInfo &UserTreeIdx, |
2548 | ArrayRef<int> ReuseShuffleIndices = None, |
2549 | ArrayRef<unsigned> ReorderIndices = None) { |
2550 | TreeEntry::EntryState EntryState = |
2551 | Bundle ? TreeEntry::Vectorize : TreeEntry::NeedToGather; |
2552 | return newTreeEntry(VL, EntryState, Bundle, S, UserTreeIdx, |
2553 | ReuseShuffleIndices, ReorderIndices); |
2554 | } |
2555 | |
2556 | TreeEntry *newTreeEntry(ArrayRef<Value *> VL, |
2557 | TreeEntry::EntryState EntryState, |
2558 | Optional<ScheduleData *> Bundle, |
2559 | const InstructionsState &S, |
2560 | const EdgeInfo &UserTreeIdx, |
2561 | ArrayRef<int> ReuseShuffleIndices = None, |
2562 | ArrayRef<unsigned> ReorderIndices = None) { |
2563 | 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", 2565, __extension__ __PRETTY_FUNCTION__)) |
2564 | (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", 2565, __extension__ __PRETTY_FUNCTION__)) |
2565 | "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", 2565, __extension__ __PRETTY_FUNCTION__)); |
2566 | VectorizableTree.push_back(std::make_unique<TreeEntry>(VectorizableTree)); |
2567 | TreeEntry *Last = VectorizableTree.back().get(); |
2568 | Last->Idx = VectorizableTree.size() - 1; |
2569 | Last->State = EntryState; |
2570 | Last->ReuseShuffleIndices.append(ReuseShuffleIndices.begin(), |
2571 | ReuseShuffleIndices.end()); |
2572 | if (ReorderIndices.empty()) { |
2573 | Last->Scalars.assign(VL.begin(), VL.end()); |
2574 | Last->setOperations(S); |
2575 | } else { |
2576 | // Reorder scalars and build final mask. |
2577 | Last->Scalars.assign(VL.size(), nullptr); |
2578 | transform(ReorderIndices, Last->Scalars.begin(), |
2579 | [VL](unsigned Idx) -> Value * { |
2580 | if (Idx >= VL.size()) |
2581 | return UndefValue::get(VL.front()->getType()); |
2582 | return VL[Idx]; |
2583 | }); |
2584 | InstructionsState S = getSameOpcode(Last->Scalars); |
2585 | Last->setOperations(S); |
2586 | Last->ReorderIndices.append(ReorderIndices.begin(), ReorderIndices.end()); |
2587 | } |
2588 | if (Last->State != TreeEntry::NeedToGather) { |
2589 | for (Value *V : VL) { |
2590 | 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", 2590, __extension__ __PRETTY_FUNCTION__)); |
2591 | ScalarToTreeEntry[V] = Last; |
2592 | } |
2593 | // Update the scheduler bundle to point to this TreeEntry. |
2594 | ScheduleData *BundleMember = *Bundle; |
2595 | 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", 2598, __extension__ __PRETTY_FUNCTION__)) |
2596 | 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", 2598, __extension__ __PRETTY_FUNCTION__)) |
2597 | 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", 2598, __extension__ __PRETTY_FUNCTION__)) |
2598 | "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", 2598, __extension__ __PRETTY_FUNCTION__)); |
2599 | if (BundleMember) { |
2600 | for (Value *V : VL) { |
2601 | if (doesNotNeedToBeScheduled(V)) |
2602 | continue; |
2603 | 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", 2603, __extension__ __PRETTY_FUNCTION__)); |
2604 | BundleMember->TE = Last; |
2605 | BundleMember = BundleMember->NextInBundle; |
2606 | } |
2607 | } |
2608 | 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", 2608, __extension__ __PRETTY_FUNCTION__)); |
2609 | } else { |
2610 | MustGather.insert(VL.begin(), VL.end()); |
2611 | } |
2612 | |
2613 | if (UserTreeIdx.UserTE) |
2614 | Last->UserTreeIndices.push_back(UserTreeIdx); |
2615 | |
2616 | return Last; |
2617 | } |
2618 | |
2619 | /// -- Vectorization State -- |
2620 | /// Holds all of the tree entries. |
2621 | TreeEntry::VecTreeTy VectorizableTree; |
2622 | |
2623 | #ifndef NDEBUG |
2624 | /// Debug printer. |
2625 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void dumpVectorizableTree() const { |
2626 | for (unsigned Id = 0, IdE = VectorizableTree.size(); Id != IdE; ++Id) { |
2627 | VectorizableTree[Id]->dump(); |
2628 | dbgs() << "\n"; |
2629 | } |
2630 | } |
2631 | #endif |
2632 | |
2633 | TreeEntry *getTreeEntry(Value *V) { return ScalarToTreeEntry.lookup(V); } |
2634 | |
2635 | const TreeEntry *getTreeEntry(Value *V) const { |
2636 | return ScalarToTreeEntry.lookup(V); |
2637 | } |
2638 | |
2639 | /// Maps a specific scalar to its tree entry. |
2640 | SmallDenseMap<Value*, TreeEntry *> ScalarToTreeEntry; |
2641 | |
2642 | /// Maps a value to the proposed vectorizable size. |
2643 | SmallDenseMap<Value *, unsigned> InstrElementSize; |
2644 | |
2645 | /// A list of scalars that we found that we need to keep as scalars. |
2646 | ValueSet MustGather; |
2647 | |
2648 | /// This POD struct describes one external user in the vectorized tree. |
2649 | struct ExternalUser { |
2650 | ExternalUser(Value *S, llvm::User *U, int L) |
2651 | : Scalar(S), User(U), Lane(L) {} |
2652 | |
2653 | // Which scalar in our function. |
2654 | Value *Scalar; |
2655 | |
2656 | // Which user that uses the scalar. |
2657 | llvm::User *User; |
2658 | |
2659 | // Which lane does the scalar belong to. |
2660 | int Lane; |
2661 | }; |
2662 | using UserList = SmallVector<ExternalUser, 16>; |
2663 | |
2664 | /// Checks if two instructions may access the same memory. |
2665 | /// |
2666 | /// \p Loc1 is the location of \p Inst1. It is passed explicitly because it |
2667 | /// is invariant in the calling loop. |
2668 | bool isAliased(const MemoryLocation &Loc1, Instruction *Inst1, |
2669 | Instruction *Inst2) { |
2670 | // First check if the result is already in the cache. |
2671 | AliasCacheKey key = std::make_pair(Inst1, Inst2); |
2672 | Optional<bool> &result = AliasCache[key]; |
2673 | if (result) { |
2674 | return result.value(); |
2675 | } |
2676 | bool aliased = true; |
2677 | if (Loc1.Ptr && isSimple(Inst1)) |
2678 | aliased = isModOrRefSet(BatchAA.getModRefInfo(Inst2, Loc1)); |
2679 | // Store the result in the cache. |
2680 | result = aliased; |
2681 | return aliased; |
2682 | } |
2683 | |
2684 | using AliasCacheKey = std::pair<Instruction *, Instruction *>; |
2685 | |
2686 | /// Cache for alias results. |
2687 | /// TODO: consider moving this to the AliasAnalysis itself. |
2688 | DenseMap<AliasCacheKey, Optional<bool>> AliasCache; |
2689 | |
2690 | // Cache for pointerMayBeCaptured calls inside AA. This is preserved |
2691 | // globally through SLP because we don't perform any action which |
2692 | // invalidates capture results. |
2693 | BatchAAResults BatchAA; |
2694 | |
2695 | /// Temporary store for deleted instructions. Instructions will be deleted |
2696 | /// eventually when the BoUpSLP is destructed. The deferral is required to |
2697 | /// ensure that there are no incorrect collisions in the AliasCache, which |
2698 | /// can happen if a new instruction is allocated at the same address as a |
2699 | /// previously deleted instruction. |
2700 | DenseSet<Instruction *> DeletedInstructions; |
2701 | |
2702 | /// Set of the instruction, being analyzed already for reductions. |
2703 | SmallPtrSet<Instruction *, 16> AnalyzedReductionsRoots; |
2704 | |
2705 | /// Set of hashes for the list of reduction values already being analyzed. |
2706 | DenseSet<size_t> AnalyzedReductionVals; |
2707 | |
2708 | /// A list of values that need to extracted out of the tree. |
2709 | /// This list holds pairs of (Internal Scalar : External User). External User |
2710 | /// can be nullptr, it means that this Internal Scalar will be used later, |
2711 | /// after vectorization. |
2712 | UserList ExternalUses; |
2713 | |
2714 | /// Values used only by @llvm.assume calls. |
2715 | SmallPtrSet<const Value *, 32> EphValues; |
2716 | |
2717 | /// Holds all of the instructions that we gathered. |
2718 | SetVector<Instruction *> GatherShuffleSeq; |
2719 | |
2720 | /// A list of blocks that we are going to CSE. |
2721 | SetVector<BasicBlock *> CSEBlocks; |
2722 | |
2723 | /// Contains all scheduling relevant data for an instruction. |
2724 | /// A ScheduleData either represents a single instruction or a member of an |
2725 | /// instruction bundle (= a group of instructions which is combined into a |
2726 | /// vector instruction). |
2727 | struct ScheduleData { |
2728 | // The initial value for the dependency counters. It means that the |
2729 | // dependencies are not calculated yet. |
2730 | enum { InvalidDeps = -1 }; |
2731 | |
2732 | ScheduleData() = default; |
2733 | |
2734 | void init(int BlockSchedulingRegionID, Value *OpVal) { |
2735 | FirstInBundle = this; |
2736 | NextInBundle = nullptr; |
2737 | NextLoadStore = nullptr; |
2738 | IsScheduled = false; |
2739 | SchedulingRegionID = BlockSchedulingRegionID; |
2740 | clearDependencies(); |
2741 | OpValue = OpVal; |
2742 | TE = nullptr; |
2743 | } |
2744 | |
2745 | /// Verify basic self consistency properties |
2746 | void verify() { |
2747 | if (hasValidDependencies()) { |
2748 | assert(UnscheduledDeps <= Dependencies && "invariant")(static_cast <bool> (UnscheduledDeps <= Dependencies && "invariant") ? void (0) : __assert_fail ("UnscheduledDeps <= Dependencies && \"invariant\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2748, __extension__ __PRETTY_FUNCTION__)); |
2749 | } else { |
2750 | assert(UnscheduledDeps == Dependencies && "invariant")(static_cast <bool> (UnscheduledDeps == Dependencies && "invariant") ? void (0) : __assert_fail ("UnscheduledDeps == Dependencies && \"invariant\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2750, __extension__ __PRETTY_FUNCTION__)); |
2751 | } |
2752 | |
2753 | if (IsScheduled) { |
2754 | assert(isSchedulingEntity() &&(static_cast <bool> (isSchedulingEntity() && "unexpected scheduled state" ) ? void (0) : __assert_fail ("isSchedulingEntity() && \"unexpected scheduled state\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2755, __extension__ __PRETTY_FUNCTION__)) |
2755 | "unexpected scheduled state")(static_cast <bool> (isSchedulingEntity() && "unexpected scheduled state" ) ? void (0) : __assert_fail ("isSchedulingEntity() && \"unexpected scheduled state\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 2755, __extension__ __PRETTY_FUNCTION__)); |
2756 | for (const ScheduleData *BundleMember = this; BundleMember; |
2757 | BundleMember = BundleMember->NextInBundle) { |
2758 | 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", 2760, __extension__ __PRETTY_FUNCTION__)) |
2759 | 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", 2760, __extension__ __PRETTY_FUNCTION__)) |
2760 | "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", 2760, __extension__ __PRETTY_FUNCTION__)); |
2761 | 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", 2762, __extension__ __PRETTY_FUNCTION__)) |
2762 | "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", 2762, __extension__ __PRETTY_FUNCTION__)); |
2763 | } |
2764 | } |
2765 | |
2766 | 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", 2767, __extension__ __PRETTY_FUNCTION__)) |
2767 | "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", 2767, __extension__ __PRETTY_FUNCTION__)); |
2768 | } |
2769 | |
2770 | /// Returns true if the dependency information has been calculated. |
2771 | /// Note that depenendency validity can vary between instructions within |
2772 | /// a single bundle. |
2773 | bool hasValidDependencies() const { return Dependencies != InvalidDeps; } |
2774 | |
2775 | /// Returns true for single instructions and for bundle representatives |
2776 | /// (= the head of a bundle). |
2777 | bool isSchedulingEntity() const { return FirstInBundle == this; } |
2778 | |
2779 | /// Returns true if it represents an instruction bundle and not only a |
2780 | /// single instruction. |
2781 | bool isPartOfBundle() const { |
2782 | return NextInBundle != nullptr || FirstInBundle != this || TE; |
2783 | } |
2784 | |
2785 | /// Returns true if it is ready for scheduling, i.e. it has no more |
2786 | /// unscheduled depending instructions/bundles. |
2787 | bool isReady() const { |
2788 | 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", 2789, __extension__ __PRETTY_FUNCTION__)) |
2789 | "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", 2789, __extension__ __PRETTY_FUNCTION__)); |
2790 | return unscheduledDepsInBundle() == 0 && !IsScheduled; |
2791 | } |
2792 | |
2793 | /// Modifies the number of unscheduled dependencies for this instruction, |
2794 | /// and returns the number of remaining dependencies for the containing |
2795 | /// bundle. |
2796 | int incrementUnscheduledDeps(int Incr) { |
2797 | 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", 2798, __extension__ __PRETTY_FUNCTION__)) |
2798 | "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", 2798, __extension__ __PRETTY_FUNCTION__)); |
2799 | UnscheduledDeps += Incr; |
2800 | return FirstInBundle->unscheduledDepsInBundle(); |
2801 | } |
2802 | |
2803 | /// Sets the number of unscheduled dependencies to the number of |
2804 | /// dependencies. |
2805 | void resetUnscheduledDeps() { |
2806 | UnscheduledDeps = Dependencies; |
2807 | } |
2808 | |
2809 | /// Clears all dependency information. |
2810 | void clearDependencies() { |
2811 | Dependencies = InvalidDeps; |
2812 | resetUnscheduledDeps(); |
2813 | MemoryDependencies.clear(); |
2814 | ControlDependencies.clear(); |
2815 | } |
2816 | |
2817 | int unscheduledDepsInBundle() const { |
2818 | 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", 2818, __extension__ __PRETTY_FUNCTION__)); |
2819 | int Sum = 0; |
2820 | for (const ScheduleData *BundleMember = this; BundleMember; |
2821 | BundleMember = BundleMember->NextInBundle) { |
2822 | if (BundleMember->UnscheduledDeps == InvalidDeps) |
2823 | return InvalidDeps; |
2824 | Sum += BundleMember->UnscheduledDeps; |
2825 | } |
2826 | return Sum; |
2827 | } |
2828 | |
2829 | void dump(raw_ostream &os) const { |
2830 | if (!isSchedulingEntity()) { |
2831 | os << "/ " << *Inst; |
2832 | } else if (NextInBundle) { |
2833 | os << '[' << *Inst; |
2834 | ScheduleData *SD = NextInBundle; |
2835 | while (SD) { |
2836 | os << ';' << *SD->Inst; |
2837 | SD = SD->NextInBundle; |
2838 | } |
2839 | os << ']'; |
2840 | } else { |
2841 | os << *Inst; |
2842 | } |
2843 | } |
2844 | |
2845 | Instruction *Inst = nullptr; |
2846 | |
2847 | /// Opcode of the current instruction in the schedule data. |
2848 | Value *OpValue = nullptr; |
2849 | |
2850 | /// The TreeEntry that this instruction corresponds to. |
2851 | TreeEntry *TE = nullptr; |
2852 | |
2853 | /// Points to the head in an instruction bundle (and always to this for |
2854 | /// single instructions). |
2855 | ScheduleData *FirstInBundle = nullptr; |
2856 | |
2857 | /// Single linked list of all instructions in a bundle. Null if it is a |
2858 | /// single instruction. |
2859 | ScheduleData *NextInBundle = nullptr; |
2860 | |
2861 | /// Single linked list of all memory instructions (e.g. load, store, call) |
2862 | /// in the block - until the end of the scheduling region. |
2863 | ScheduleData *NextLoadStore = nullptr; |
2864 | |
2865 | /// The dependent memory instructions. |
2866 | /// This list is derived on demand in calculateDependencies(). |
2867 | SmallVector<ScheduleData *, 4> MemoryDependencies; |
2868 | |
2869 | /// List of instructions which this instruction could be control dependent |
2870 | /// on. Allowing such nodes to be scheduled below this one could introduce |
2871 | /// a runtime fault which didn't exist in the original program. |
2872 | /// ex: this is a load or udiv following a readonly call which inf loops |
2873 | SmallVector<ScheduleData *, 4> ControlDependencies; |
2874 | |
2875 | /// This ScheduleData is in the current scheduling region if this matches |
2876 | /// the current SchedulingRegionID of BlockScheduling. |
2877 | int SchedulingRegionID = 0; |
2878 | |
2879 | /// Used for getting a "good" final ordering of instructions. |
2880 | int SchedulingPriority = 0; |
2881 | |
2882 | /// The number of dependencies. Constitutes of the number of users of the |
2883 | /// instruction plus the number of dependent memory instructions (if any). |
2884 | /// This value is calculated on demand. |
2885 | /// If InvalidDeps, the number of dependencies is not calculated yet. |
2886 | int Dependencies = InvalidDeps; |
2887 | |
2888 | /// The number of dependencies minus the number of dependencies of scheduled |
2889 | /// instructions. As soon as this is zero, the instruction/bundle gets ready |
2890 | /// for scheduling. |
2891 | /// Note that this is negative as long as Dependencies is not calculated. |
2892 | int UnscheduledDeps = InvalidDeps; |
2893 | |
2894 | /// True if this instruction is scheduled (or considered as scheduled in the |
2895 | /// dry-run). |
2896 | bool IsScheduled = false; |
2897 | }; |
2898 | |
2899 | #ifndef NDEBUG |
2900 | friend inline raw_ostream &operator<<(raw_ostream &os, |
2901 | const BoUpSLP::ScheduleData &SD) { |
2902 | SD.dump(os); |
2903 | return os; |
2904 | } |
2905 | #endif |
2906 | |
2907 | friend struct GraphTraits<BoUpSLP *>; |
2908 | friend struct DOTGraphTraits<BoUpSLP *>; |
2909 | |
2910 | /// Contains all scheduling data for a basic block. |
2911 | /// It does not schedules instructions, which are not memory read/write |
2912 | /// instructions and their operands are either constants, or arguments, or |
2913 | /// phis, or instructions from others blocks, or their users are phis or from |
2914 | /// the other blocks. The resulting vector instructions can be placed at the |
2915 | /// beginning of the basic block without scheduling (if operands does not need |
2916 | /// to be scheduled) or at the end of the block (if users are outside of the |
2917 | /// block). It allows to save some compile time and memory used by the |
2918 | /// compiler. |
2919 | /// ScheduleData is assigned for each instruction in between the boundaries of |
2920 | /// the tree entry, even for those, which are not part of the graph. It is |
2921 | /// required to correctly follow the dependencies between the instructions and |
2922 | /// their correct scheduling. The ScheduleData is not allocated for the |
2923 | /// instructions, which do not require scheduling, like phis, nodes with |
2924 | /// extractelements/insertelements only or nodes with instructions, with |
2925 | /// uses/operands outside of the block. |
2926 | struct BlockScheduling { |
2927 | BlockScheduling(BasicBlock *BB) |
2928 | : BB(BB), ChunkSize(BB->size()), ChunkPos(ChunkSize) {} |
2929 | |
2930 | void clear() { |
2931 | ReadyInsts.clear(); |
2932 | ScheduleStart = nullptr; |
2933 | ScheduleEnd = nullptr; |
2934 | FirstLoadStoreInRegion = nullptr; |
2935 | LastLoadStoreInRegion = nullptr; |
2936 | RegionHasStackSave = false; |
2937 | |
2938 | // Reduce the maximum schedule region size by the size of the |
2939 | // previous scheduling run. |
2940 | ScheduleRegionSizeLimit -= ScheduleRegionSize; |
2941 | if (ScheduleRegionSizeLimit < MinScheduleRegionSize) |
2942 | ScheduleRegionSizeLimit = MinScheduleRegionSize; |
2943 | ScheduleRegionSize = 0; |
2944 | |
2945 | // Make a new scheduling region, i.e. all existing ScheduleData is not |
2946 | // in the new region yet. |
2947 | ++SchedulingRegionID; |
2948 | } |
2949 | |
2950 | ScheduleData *getScheduleData(Instruction *I) { |
2951 | if (BB != I->getParent()) |
2952 | // Avoid lookup if can't possibly be in map. |
2953 | return nullptr; |
2954 | ScheduleData *SD = ScheduleDataMap.lookup(I); |
2955 | if (SD && isInSchedulingRegion(SD)) |
2956 | return SD; |
2957 | return nullptr; |
2958 | } |
2959 | |
2960 | ScheduleData *getScheduleData(Value *V) { |
2961 | if (auto *I = dyn_cast<Instruction>(V)) |
2962 | return getScheduleData(I); |
2963 | return nullptr; |
2964 | } |
2965 | |
2966 | ScheduleData *getScheduleData(Value *V, Value *Key) { |
2967 | if (V == Key) |
2968 | return getScheduleData(V); |
2969 | auto I = ExtraScheduleDataMap.find(V); |
2970 | if (I != ExtraScheduleDataMap.end()) { |
2971 | ScheduleData *SD = I->second.lookup(Key); |
2972 | if (SD && isInSchedulingRegion(SD)) |
2973 | return SD; |
2974 | } |
2975 | return nullptr; |
2976 | } |
2977 | |
2978 | bool isInSchedulingRegion(ScheduleData *SD) const { |
2979 | return SD->SchedulingRegionID == SchedulingRegionID; |
2980 | } |
2981 | |
2982 | /// Marks an instruction as scheduled and puts all dependent ready |
2983 | /// instructions into the ready-list. |
2984 | template <typename ReadyListType> |
2985 | void schedule(ScheduleData *SD, ReadyListType &ReadyList) { |
2986 | SD->IsScheduled = true; |
2987 | LLVM_DEBUG(dbgs() << "SLP: schedule " << *SD << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: schedule " << *SD << "\n"; } } while (false); |
2988 | |
2989 | for (ScheduleData *BundleMember = SD; BundleMember; |
2990 | BundleMember = BundleMember->NextInBundle) { |
2991 | if (BundleMember->Inst != BundleMember->OpValue) |
2992 | continue; |
2993 | |
2994 | // Handle the def-use chain dependencies. |
2995 | |
2996 | // Decrement the unscheduled counter and insert to ready list if ready. |
2997 | auto &&DecrUnsched = [this, &ReadyList](Instruction *I) { |
2998 | doForAllOpcodes(I, [&ReadyList](ScheduleData *OpDef) { |
2999 | if (OpDef && OpDef->hasValidDependencies() && |
3000 | OpDef->incrementUnscheduledDeps(-1) == 0) { |
3001 | // There are no more unscheduled dependencies after |
3002 | // decrementing, so we can put the dependent instruction |
3003 | // into the ready list. |
3004 | ScheduleData *DepBundle = OpDef->FirstInBundle; |
3005 | 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", 3006, __extension__ __PRETTY_FUNCTION__)) |
3006 | "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", 3006, __extension__ __PRETTY_FUNCTION__)); |
3007 | ReadyList.insert(DepBundle); |
3008 | LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: gets ready (def): " << *DepBundle << "\n"; } } while (false) |
3009 | << "SLP: gets ready (def): " << *DepBundle << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: gets ready (def): " << *DepBundle << "\n"; } } while (false); |
3010 | } |
3011 | }); |
3012 | }; |
3013 | |
3014 | // If BundleMember is a vector bundle, its operands may have been |
3015 | // reordered during buildTree(). We therefore need to get its operands |
3016 | // through the TreeEntry. |
3017 | if (TreeEntry *TE = BundleMember->TE) { |
3018 | // Need to search for the lane since the tree entry can be reordered. |
3019 | int Lane = std::distance(TE->Scalars.begin(), |
3020 | find(TE->Scalars, BundleMember->Inst)); |
3021 | 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", 3021, __extension__ __PRETTY_FUNCTION__)); |
3022 | |
3023 | // Since vectorization tree is being built recursively this assertion |
3024 | // ensures that the tree entry has all operands set before reaching |
3025 | // this code. Couple of exceptions known at the moment are extracts |
3026 | // where their second (immediate) operand is not added. Since |
3027 | // immediates do not affect scheduler behavior this is considered |
3028 | // okay. |
3029 | auto *In = BundleMember->Inst; |
3030 | 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", 3033, __extension__ __PRETTY_FUNCTION__)) |
3031 | (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", 3033, __extension__ __PRETTY_FUNCTION__)) |
3032 | 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", 3033, __extension__ __PRETTY_FUNCTION__)) |
3033 | "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", 3033, __extension__ __PRETTY_FUNCTION__)); |
3034 | (void)In; // fake use to avoid build failure when assertions disabled |
3035 | |
3036 | for (unsigned OpIdx = 0, NumOperands = TE->getNumOperands(); |
3037 | OpIdx != NumOperands; ++OpIdx) |
3038 | if (auto *I = dyn_cast<Instruction>(TE->getOperand(OpIdx)[Lane])) |
3039 | DecrUnsched(I); |
3040 | } else { |
3041 | // If BundleMember is a stand-alone instruction, no operand reordering |
3042 | // has taken place, so we directly access its operands. |
3043 | for (Use &U : BundleMember->Inst->operands()) |
3044 | if (auto *I = dyn_cast<Instruction>(U.get())) |
3045 | DecrUnsched(I); |
3046 | } |
3047 | // Handle the memory dependencies. |
3048 | for (ScheduleData *MemoryDepSD : BundleMember->MemoryDependencies) { |
3049 | if (MemoryDepSD->hasValidDependencies() && |
3050 | MemoryDepSD->incrementUnscheduledDeps(-1) == 0) { |
3051 | // There are no more unscheduled dependencies after decrementing, |
3052 | // so we can put the dependent instruction into the ready list. |
3053 | ScheduleData *DepBundle = MemoryDepSD->FirstInBundle; |
3054 | 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", 3055, __extension__ __PRETTY_FUNCTION__)) |
3055 | "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", 3055, __extension__ __PRETTY_FUNCTION__)); |
3056 | ReadyList.insert(DepBundle); |
3057 | LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: gets ready (mem): " << *DepBundle << "\n"; } } while (false) |
3058 | << "SLP: gets ready (mem): " << *DepBundle << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: gets ready (mem): " << *DepBundle << "\n"; } } while (false); |
3059 | } |
3060 | } |
3061 | // Handle the control dependencies. |
3062 | for (ScheduleData *DepSD : BundleMember->ControlDependencies) { |
3063 | if (DepSD->incrementUnscheduledDeps(-1) == 0) { |
3064 | // There are no more unscheduled dependencies after decrementing, |
3065 | // so we can put the dependent instruction into the ready list. |
3066 | ScheduleData *DepBundle = DepSD->FirstInBundle; |
3067 | 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", 3068, __extension__ __PRETTY_FUNCTION__)) |
3068 | "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", 3068, __extension__ __PRETTY_FUNCTION__)); |
3069 | ReadyList.insert(DepBundle); |
3070 | LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: gets ready (ctl): " << *DepBundle << "\n"; } } while (false) |
3071 | << "SLP: gets ready (ctl): " << *DepBundle << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: gets ready (ctl): " << *DepBundle << "\n"; } } while (false); |
3072 | } |
3073 | } |
3074 | |
3075 | } |
3076 | } |
3077 | |
3078 | /// Verify basic self consistency properties of the data structure. |
3079 | void verify() { |
3080 | if (!ScheduleStart) |
3081 | return; |
3082 | |
3083 | 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", 3085, __extension__ __PRETTY_FUNCTION__)) |
3084 | 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", 3085, __extension__ __PRETTY_FUNCTION__)) |
3085 | "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", 3085, __extension__ __PRETTY_FUNCTION__)); |
3086 | |
3087 | for (auto *I = ScheduleStart; I != ScheduleEnd; I = I->getNextNode()) { |
3088 | auto *SD = getScheduleData(I); |
3089 | if (!SD) |
3090 | continue; |
3091 | 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", 3092, __extension__ __PRETTY_FUNCTION__)) |
3092 | "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", 3092, __extension__ __PRETTY_FUNCTION__)); |
3093 | 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", 3094, __extension__ __PRETTY_FUNCTION__)) |
3094 | "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", 3094, __extension__ __PRETTY_FUNCTION__)); |
3095 | (void)SD; |
3096 | doForAllOpcodes(I, [](ScheduleData *SD) { SD->verify(); }); |
3097 | } |
3098 | |
3099 | for (auto *SD : ReadyInsts) { |
3100 | 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", 3101, __extension__ __PRETTY_FUNCTION__)) |
3101 | "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", 3101, __extension__ __PRETTY_FUNCTION__)); |
3102 | (void)SD; |
3103 | } |
3104 | } |
3105 | |
3106 | void doForAllOpcodes(Value *V, |
3107 | function_ref<void(ScheduleData *SD)> Action) { |
3108 | if (ScheduleData *SD = getScheduleData(V)) |
3109 | Action(SD); |
3110 | auto I = ExtraScheduleDataMap.find(V); |
3111 | if (I != ExtraScheduleDataMap.end()) |
3112 | for (auto &P : I->second) |
3113 | if (isInSchedulingRegion(P.second)) |
3114 | Action(P.second); |
3115 | } |
3116 | |
3117 | /// Put all instructions into the ReadyList which are ready for scheduling. |
3118 | template <typename ReadyListType> |
3119 | void initialFillReadyList(ReadyListType &ReadyList) { |
3120 | for (auto *I = ScheduleStart; I != ScheduleEnd; I = I->getNextNode()) { |
3121 | doForAllOpcodes(I, [&](ScheduleData *SD) { |
3122 | if (SD->isSchedulingEntity() && SD->hasValidDependencies() && |
3123 | SD->isReady()) { |
3124 | ReadyList.insert(SD); |
3125 | LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: initially in ready list: " << *SD << "\n"; } } while (false) |
3126 | << "SLP: initially in ready list: " << *SD << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: initially in ready list: " << *SD << "\n"; } } while (false); |
3127 | } |
3128 | }); |
3129 | } |
3130 | } |
3131 | |
3132 | /// Build a bundle from the ScheduleData nodes corresponding to the |
3133 | /// scalar instruction for each lane. |
3134 | ScheduleData *buildBundle(ArrayRef<Value *> VL); |
3135 | |
3136 | /// Checks if a bundle of instructions can be scheduled, i.e. has no |
3137 | /// cyclic dependencies. This is only a dry-run, no instructions are |
3138 | /// actually moved at this stage. |
3139 | /// \returns the scheduling bundle. The returned Optional value is non-None |
3140 | /// if \p VL is allowed to be scheduled. |
3141 | Optional<ScheduleData *> |
3142 | tryScheduleBundle(ArrayRef<Value *> VL, BoUpSLP *SLP, |
3143 | const InstructionsState &S); |
3144 | |
3145 | /// Un-bundles a group of instructions. |
3146 | void cancelScheduling(ArrayRef<Value *> VL, Value *OpValue); |
3147 | |
3148 | /// Allocates schedule data chunk. |
3149 | ScheduleData *allocateScheduleDataChunks(); |
3150 | |
3151 | /// Extends the scheduling region so that V is inside the region. |
3152 | /// \returns true if the region size is within the limit. |
3153 | bool extendSchedulingRegion(Value *V, const InstructionsState &S); |
3154 | |
3155 | /// Initialize the ScheduleData structures for new instructions in the |
3156 | /// scheduling region. |
3157 | void initScheduleData(Instruction *FromI, Instruction *ToI, |
3158 | ScheduleData *PrevLoadStore, |
3159 | ScheduleData *NextLoadStore); |
3160 | |
3161 | /// Updates the dependency information of a bundle and of all instructions/ |
3162 | /// bundles which depend on the original bundle. |
3163 | void calculateDependencies(ScheduleData *SD, bool InsertInReadyList, |
3164 | BoUpSLP *SLP); |
3165 | |
3166 | /// Sets all instruction in the scheduling region to un-scheduled. |
3167 | void resetSchedule(); |
3168 | |
3169 | BasicBlock *BB; |
3170 | |
3171 | /// Simple memory allocation for ScheduleData. |
3172 | std::vector<std::unique_ptr<ScheduleData[]>> ScheduleDataChunks; |
3173 | |
3174 | /// The size of a ScheduleData array in ScheduleDataChunks. |
3175 | int ChunkSize; |
3176 | |
3177 | /// The allocator position in the current chunk, which is the last entry |
3178 | /// of ScheduleDataChunks. |
3179 | int ChunkPos; |
3180 | |
3181 | /// Attaches ScheduleData to Instruction. |
3182 | /// Note that the mapping survives during all vectorization iterations, i.e. |
3183 | /// ScheduleData structures are recycled. |
3184 | DenseMap<Instruction *, ScheduleData *> ScheduleDataMap; |
3185 | |
3186 | /// Attaches ScheduleData to Instruction with the leading key. |
3187 | DenseMap<Value *, SmallDenseMap<Value *, ScheduleData *>> |
3188 | ExtraScheduleDataMap; |
3189 | |
3190 | /// The ready-list for scheduling (only used for the dry-run). |
3191 | SetVector<ScheduleData *> ReadyInsts; |
3192 | |
3193 | /// The first instruction of the scheduling region. |
3194 | Instruction *ScheduleStart = nullptr; |
3195 | |
3196 | /// The first instruction _after_ the scheduling region. |
3197 | Instruction *ScheduleEnd = nullptr; |
3198 | |
3199 | /// The first memory accessing instruction in the scheduling region |
3200 | /// (can be null). |
3201 | ScheduleData *FirstLoadStoreInRegion = nullptr; |
3202 | |
3203 | /// The last memory accessing instruction in the scheduling region |
3204 | /// (can be null). |
3205 | ScheduleData *LastLoadStoreInRegion = nullptr; |
3206 | |
3207 | /// Is there an llvm.stacksave or llvm.stackrestore in the scheduling |
3208 | /// region? Used to optimize the dependence calculation for the |
3209 | /// common case where there isn't. |
3210 | bool RegionHasStackSave = false; |
3211 | |
3212 | /// The current size of the scheduling region. |
3213 | int ScheduleRegionSize = 0; |
3214 | |
3215 | /// The maximum size allowed for the scheduling region. |
3216 | int ScheduleRegionSizeLimit = ScheduleRegionSizeBudget; |
3217 | |
3218 | /// The ID of the scheduling region. For a new vectorization iteration this |
3219 | /// is incremented which "removes" all ScheduleData from the region. |
3220 | /// Make sure that the initial SchedulingRegionID is greater than the |
3221 | /// initial SchedulingRegionID in ScheduleData (which is 0). |
3222 | int SchedulingRegionID = 1; |
3223 | }; |
3224 | |
3225 | /// Attaches the BlockScheduling structures to basic blocks. |
3226 | MapVector<BasicBlock *, std::unique_ptr<BlockScheduling>> BlocksSchedules; |
3227 | |
3228 | /// Performs the "real" scheduling. Done before vectorization is actually |
3229 | /// performed in a basic block. |
3230 | void scheduleBlock(BlockScheduling *BS); |
3231 | |
3232 | /// List of users to ignore during scheduling and that don't need extracting. |
3233 | const SmallDenseSet<Value *> *UserIgnoreList = nullptr; |
3234 | |
3235 | /// A DenseMapInfo implementation for holding DenseMaps and DenseSets of |
3236 | /// sorted SmallVectors of unsigned. |
3237 | struct OrdersTypeDenseMapInfo { |
3238 | static OrdersType getEmptyKey() { |
3239 | OrdersType V; |
3240 | V.push_back(~1U); |
3241 | return V; |
3242 | } |
3243 | |
3244 | static OrdersType getTombstoneKey() { |
3245 | OrdersType V; |
3246 | V.push_back(~2U); |
3247 | return V; |
3248 | } |
3249 | |
3250 | static unsigned getHashValue(const OrdersType &V) { |
3251 | return static_cast<unsigned>(hash_combine_range(V.begin(), V.end())); |
3252 | } |
3253 | |
3254 | static bool isEqual(const OrdersType &LHS, const OrdersType &RHS) { |
3255 | return LHS == RHS; |
3256 | } |
3257 | }; |
3258 | |
3259 | // Analysis and block reference. |
3260 | Function *F; |
3261 | ScalarEvolution *SE; |
3262 | TargetTransformInfo *TTI; |
3263 | TargetLibraryInfo *TLI; |
3264 | LoopInfo *LI; |
3265 | DominatorTree *DT; |
3266 | AssumptionCache *AC; |
3267 | DemandedBits *DB; |
3268 | const DataLayout *DL; |
3269 | OptimizationRemarkEmitter *ORE; |
3270 | |
3271 | unsigned MaxVecRegSize; // This is set by TTI or overridden by cl::opt. |
3272 | unsigned MinVecRegSize; // Set by cl::opt (default: 128). |
3273 | |
3274 | /// Instruction builder to construct the vectorized tree. |
3275 | IRBuilder<> Builder; |
3276 | |
3277 | /// A map of scalar integer values to the smallest bit width with which they |
3278 | /// can legally be represented. The values map to (width, signed) pairs, |
3279 | /// where "width" indicates the minimum bit width and "signed" is True if the |
3280 | /// value must be signed-extended, rather than zero-extended, back to its |
3281 | /// original width. |
3282 | MapVector<Value *, std::pair<uint64_t, bool>> MinBWs; |
3283 | }; |
3284 | |
3285 | } // end namespace slpvectorizer |
3286 | |
3287 | template <> struct GraphTraits<BoUpSLP *> { |
3288 | using TreeEntry = BoUpSLP::TreeEntry; |
3289 | |
3290 | /// NodeRef has to be a pointer per the GraphWriter. |
3291 | using NodeRef = TreeEntry *; |
3292 | |
3293 | using ContainerTy = BoUpSLP::TreeEntry::VecTreeTy; |
3294 | |
3295 | /// Add the VectorizableTree to the index iterator to be able to return |
3296 | /// TreeEntry pointers. |
3297 | struct ChildIteratorType |
3298 | : public iterator_adaptor_base< |
3299 | ChildIteratorType, SmallVector<BoUpSLP::EdgeInfo, 1>::iterator> { |
3300 | ContainerTy &VectorizableTree; |
3301 | |
3302 | ChildIteratorType(SmallVector<BoUpSLP::EdgeInfo, 1>::iterator W, |
3303 | ContainerTy &VT) |
3304 | : ChildIteratorType::iterator_adaptor_base(W), VectorizableTree(VT) {} |
3305 | |
3306 | NodeRef operator*() { return I->UserTE; } |
3307 | }; |
3308 | |
3309 | static NodeRef getEntryNode(BoUpSLP &R) { |
3310 | return R.VectorizableTree[0].get(); |
3311 | } |
3312 | |
3313 | static ChildIteratorType child_begin(NodeRef N) { |
3314 | return {N->UserTreeIndices.begin(), N->Container}; |
3315 | } |
3316 | |
3317 | static ChildIteratorType child_end(NodeRef N) { |
3318 | return {N->UserTreeIndices.end(), N->Container}; |
3319 | } |
3320 | |
3321 | /// For the node iterator we just need to turn the TreeEntry iterator into a |
3322 | /// TreeEntry* iterator so that it dereferences to NodeRef. |
3323 | class nodes_iterator { |
3324 | using ItTy = ContainerTy::iterator; |
3325 | ItTy It; |
3326 | |
3327 | public: |
3328 | nodes_iterator(const ItTy &It2) : It(It2) {} |
3329 | NodeRef operator*() { return It->get(); } |
3330 | nodes_iterator operator++() { |
3331 | ++It; |
3332 | return *this; |
3333 | } |
3334 | bool operator!=(const nodes_iterator &N2) const { return N2.It != It; } |
3335 | }; |
3336 | |
3337 | static nodes_iterator nodes_begin(BoUpSLP *R) { |
3338 | return nodes_iterator(R->VectorizableTree.begin()); |
3339 | } |
3340 | |
3341 | static nodes_iterator nodes_end(BoUpSLP *R) { |
3342 | return nodes_iterator(R->VectorizableTree.end()); |
3343 | } |
3344 | |
3345 | static unsigned size(BoUpSLP *R) { return R->VectorizableTree.size(); } |
3346 | }; |
3347 | |
3348 | template <> struct DOTGraphTraits<BoUpSLP *> : public DefaultDOTGraphTraits { |
3349 | using TreeEntry = BoUpSLP::TreeEntry; |
3350 | |
3351 | DOTGraphTraits(bool isSimple = false) : DefaultDOTGraphTraits(isSimple) {} |
3352 | |
3353 | std::string getNodeLabel(const TreeEntry *Entry, const BoUpSLP *R) { |
3354 | std::string Str; |
3355 | raw_string_ostream OS(Str); |
3356 | if (isSplat(Entry->Scalars)) |
3357 | OS << "<splat> "; |
3358 | for (auto *V : Entry->Scalars) { |
3359 | OS << *V; |
3360 | if (llvm::any_of(R->ExternalUses, [&](const BoUpSLP::ExternalUser &EU) { |
3361 | return EU.Scalar == V; |
3362 | })) |
3363 | OS << " <extract>"; |
3364 | OS << "\n"; |
3365 | } |
3366 | return Str; |
3367 | } |
3368 | |
3369 | static std::string getNodeAttributes(const TreeEntry *Entry, |
3370 | const BoUpSLP *) { |
3371 | if (Entry->State == TreeEntry::NeedToGather) |
3372 | return "color=red"; |
3373 | return ""; |
3374 | } |
3375 | }; |
3376 | |
3377 | } // end namespace llvm |
3378 | |
3379 | BoUpSLP::~BoUpSLP() { |
3380 | SmallVector<WeakTrackingVH> DeadInsts; |
3381 | for (auto *I : DeletedInstructions) { |
3382 | for (Use &U : I->operands()) { |
3383 | auto *Op = dyn_cast<Instruction>(U.get()); |
3384 | if (Op && !DeletedInstructions.count(Op) && Op->hasOneUser() && |
3385 | wouldInstructionBeTriviallyDead(Op, TLI)) |
3386 | DeadInsts.emplace_back(Op); |
3387 | } |
3388 | I->dropAllReferences(); |
3389 | } |
3390 | for (auto *I : DeletedInstructions) { |
3391 | 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", 3392, __extension__ __PRETTY_FUNCTION__)) |
3392 | "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", 3392, __extension__ __PRETTY_FUNCTION__)); |
3393 | I->eraseFromParent(); |
3394 | } |
3395 | |
3396 | // Cleanup any dead scalar code feeding the vectorized instructions |
3397 | RecursivelyDeleteTriviallyDeadInstructions(DeadInsts, TLI); |
3398 | |
3399 | #ifdef EXPENSIVE_CHECKS |
3400 | // If we could guarantee that this call is not extremely slow, we could |
3401 | // remove the ifdef limitation (see PR47712). |
3402 | assert(!verifyFunction(*F, &dbgs()))(static_cast <bool> (!verifyFunction(*F, &dbgs())) ? void (0) : __assert_fail ("!verifyFunction(*F, &dbgs())" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 3402, __extension__ __PRETTY_FUNCTION__)); |
3403 | #endif |
3404 | } |
3405 | |
3406 | /// Reorders the given \p Reuses mask according to the given \p Mask. \p Reuses |
3407 | /// contains original mask for the scalars reused in the node. Procedure |
3408 | /// transform this mask in accordance with the given \p Mask. |
3409 | static void reorderReuses(SmallVectorImpl<int> &Reuses, ArrayRef<int> Mask) { |
3410 | 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", 3411, __extension__ __PRETTY_FUNCTION__)) |
3411 | "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", 3411, __extension__ __PRETTY_FUNCTION__)); |
3412 | SmallVector<int> Prev(Reuses.begin(), Reuses.end()); |
3413 | Prev.swap(Reuses); |
3414 | for (unsigned I = 0, E = Prev.size(); I < E; ++I) |
3415 | if (Mask[I] != UndefMaskElem) |
3416 | Reuses[Mask[I]] = Prev[I]; |
3417 | } |
3418 | |
3419 | /// Reorders the given \p Order according to the given \p Mask. \p Order - is |
3420 | /// the original order of the scalars. Procedure transforms the provided order |
3421 | /// in accordance with the given \p Mask. If the resulting \p Order is just an |
3422 | /// identity order, \p Order is cleared. |
3423 | static void reorderOrder(SmallVectorImpl<unsigned> &Order, ArrayRef<int> Mask) { |
3424 | 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", 3424, __extension__ __PRETTY_FUNCTION__)); |
3425 | SmallVector<int> MaskOrder; |
3426 | if (Order.empty()) { |
3427 | MaskOrder.resize(Mask.size()); |
3428 | std::iota(MaskOrder.begin(), MaskOrder.end(), 0); |
3429 | } else { |
3430 | inversePermutation(Order, MaskOrder); |
3431 | } |
3432 | reorderReuses(MaskOrder, Mask); |
3433 | if (ShuffleVectorInst::isIdentityMask(MaskOrder)) { |
3434 | Order.clear(); |
3435 | return; |
3436 | } |
3437 | Order.assign(Mask.size(), Mask.size()); |
3438 | for (unsigned I = 0, E = Mask.size(); I < E; ++I) |
3439 | if (MaskOrder[I] != UndefMaskElem) |
3440 | Order[MaskOrder[I]] = I; |
3441 | fixupOrderingIndices(Order); |
3442 | } |
3443 | |
3444 | Optional<BoUpSLP::OrdersType> |
3445 | BoUpSLP::findReusedOrderedScalars(const BoUpSLP::TreeEntry &TE) { |
3446 | 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", 3446, __extension__ __PRETTY_FUNCTION__)); |
3447 | unsigned NumScalars = TE.Scalars.size(); |
3448 | OrdersType CurrentOrder(NumScalars, NumScalars); |
3449 | SmallVector<int> Positions; |
3450 | SmallBitVector UsedPositions(NumScalars); |
3451 | const TreeEntry *STE = nullptr; |
3452 | // Try to find all gathered scalars that are gets vectorized in other |
3453 | // vectorize node. Here we can have only one single tree vector node to |
3454 | // correctly identify order of the gathered scalars. |
3455 | for (unsigned I = 0; I < NumScalars; ++I) { |
3456 | Value *V = TE.Scalars[I]; |
3457 | if (!isa<LoadInst, ExtractElementInst, ExtractValueInst>(V)) |
3458 | continue; |
3459 | if (const auto *LocalSTE = getTreeEntry(V)) { |
3460 | if (!STE) |
3461 | STE = LocalSTE; |
3462 | else if (STE != LocalSTE) |
3463 | // Take the order only from the single vector node. |
3464 | return None; |
3465 | unsigned Lane = |
3466 | std::distance(STE->Scalars.begin(), find(STE->Scalars, V)); |
3467 | if (Lane >= NumScalars) |
3468 | return None; |
3469 | if (CurrentOrder[Lane] != NumScalars) { |
3470 | if (Lane != I) |
3471 | continue; |
3472 | UsedPositions.reset(CurrentOrder[Lane]); |
3473 | } |
3474 | // The partial identity (where only some elements of the gather node are |
3475 | // in the identity order) is good. |
3476 | CurrentOrder[Lane] = I; |
3477 | UsedPositions.set(I); |
3478 | } |
3479 | } |
3480 | // Need to keep the order if we have a vector entry and at least 2 scalars or |
3481 | // the vectorized entry has just 2 scalars. |
3482 | if (STE && (UsedPositions.count() > 1 || STE->Scalars.size() == 2)) { |
3483 | auto &&IsIdentityOrder = [NumScalars](ArrayRef<unsigned> CurrentOrder) { |
3484 | for (unsigned I = 0; I < NumScalars; ++I) |
3485 | if (CurrentOrder[I] != I && CurrentOrder[I] != NumScalars) |
3486 | return false; |
3487 | return true; |
3488 | }; |
3489 | if (IsIdentityOrder(CurrentOrder)) { |
3490 | CurrentOrder.clear(); |
3491 | return CurrentOrder; |
3492 | } |
3493 | auto *It = CurrentOrder.begin(); |
3494 | for (unsigned I = 0; I < NumScalars;) { |
3495 | if (UsedPositions.test(I)) { |
3496 | ++I; |
3497 | continue; |
3498 | } |
3499 | if (*It == NumScalars) { |
3500 | *It = I; |
3501 | ++I; |
3502 | } |
3503 | ++It; |
3504 | } |
3505 | return CurrentOrder; |
3506 | } |
3507 | return None; |
3508 | } |
3509 | |
3510 | namespace { |
3511 | /// Tracks the state we can represent the loads in the given sequence. |
3512 | enum class LoadsState { Gather, Vectorize, ScatterVectorize }; |
3513 | } // anonymous namespace |
3514 | |
3515 | /// Checks if the given array of loads can be represented as a vectorized, |
3516 | /// scatter or just simple gather. |
3517 | static LoadsState canVectorizeLoads(ArrayRef<Value *> VL, const Value *VL0, |
3518 | const TargetTransformInfo &TTI, |
3519 | const DataLayout &DL, ScalarEvolution &SE, |
3520 | LoopInfo &LI, |
3521 | SmallVectorImpl<unsigned> &Order, |
3522 | SmallVectorImpl<Value *> &PointerOps) { |
3523 | // Check that a vectorized load would load the same memory as a scalar |
3524 | // load. For example, we don't want to vectorize loads that are smaller |
3525 | // than 8-bit. Even though we have a packed struct {<i2, i2, i2, i2>} LLVM |
3526 | // treats loading/storing it as an i8 struct. If we vectorize loads/stores |
3527 | // from such a struct, we read/write packed bits disagreeing with the |
3528 | // unvectorized version. |
3529 | Type *ScalarTy = VL0->getType(); |
3530 | |
3531 | if (DL.getTypeSizeInBits(ScalarTy) != DL.getTypeAllocSizeInBits(ScalarTy)) |
3532 | return LoadsState::Gather; |
3533 | |
3534 | // Make sure all loads in the bundle are simple - we can't vectorize |
3535 | // atomic or volatile loads. |
3536 | PointerOps.clear(); |
3537 | PointerOps.resize(VL.size()); |
3538 | auto *POIter = PointerOps.begin(); |
3539 | for (Value *V : VL) { |
3540 | auto *L = cast<LoadInst>(V); |
3541 | if (!L->isSimple()) |
3542 | return LoadsState::Gather; |
3543 | *POIter = L->getPointerOperand(); |
3544 | ++POIter; |
3545 | } |
3546 | |
3547 | Order.clear(); |
3548 | // Check the order of pointer operands or that all pointers are the same. |
3549 | bool IsSorted = sortPtrAccesses(PointerOps, ScalarTy, DL, SE, Order); |
3550 | if (IsSorted || all_of(PointerOps, [&PointerOps](Value *P) { |
3551 | if (getUnderlyingObject(P) != getUnderlyingObject(PointerOps.front())) |
3552 | return false; |
3553 | auto *GEP = dyn_cast<GetElementPtrInst>(P); |
3554 | if (!GEP) |
3555 | return false; |
3556 | auto *GEP0 = cast<GetElementPtrInst>(PointerOps.front()); |
3557 | return GEP->getNumOperands() == 2 && |
3558 | ((isConstant(GEP->getOperand(1)) && |
3559 | isConstant(GEP0->getOperand(1))) || |
3560 | getSameOpcode({GEP->getOperand(1), GEP0->getOperand(1)}) |
3561 | .getOpcode()); |
3562 | })) { |
3563 | if (IsSorted) { |
3564 | Value *Ptr0; |
3565 | Value *PtrN; |
3566 | if (Order.empty()) { |
3567 | Ptr0 = PointerOps.front(); |
3568 | PtrN = PointerOps.back(); |
3569 | } else { |
3570 | Ptr0 = PointerOps[Order.front()]; |
3571 | PtrN = PointerOps[Order.back()]; |
3572 | } |
3573 | Optional<int> Diff = |
3574 | getPointersDiff(ScalarTy, Ptr0, ScalarTy, PtrN, DL, SE); |
3575 | // Check that the sorted loads are consecutive. |
3576 | if (static_cast<unsigned>(*Diff) == VL.size() - 1) |
3577 | return LoadsState::Vectorize; |
3578 | } |
3579 | // TODO: need to improve analysis of the pointers, if not all of them are |
3580 | // GEPs or have > 2 operands, we end up with a gather node, which just |
3581 | // increases the cost. |
3582 | Loop *L = LI.getLoopFor(cast<LoadInst>(VL0)->getParent()); |
3583 | bool ProfitableGatherPointers = |
3584 | static_cast<unsigned>(count_if(PointerOps, [L](Value *V) { |
3585 | return L && L->isLoopInvariant(V); |
3586 | })) <= VL.size() / 2 && VL.size() > 2; |
3587 | if (ProfitableGatherPointers || all_of(PointerOps, [IsSorted](Value *P) { |
3588 | auto *GEP = dyn_cast<GetElementPtrInst>(P); |
3589 | return (IsSorted && !GEP && doesNotNeedToBeScheduled(P)) || |
3590 | (GEP && GEP->getNumOperands() == 2); |
3591 | })) { |
3592 | Align CommonAlignment = cast<LoadInst>(VL0)->getAlign(); |
3593 | for (Value *V : VL) |
3594 | CommonAlignment = |
3595 | std::min(CommonAlignment, cast<LoadInst>(V)->getAlign()); |
3596 | auto *VecTy = FixedVectorType::get(ScalarTy, VL.size()); |
3597 | if (TTI.isLegalMaskedGather(VecTy, CommonAlignment) && |
3598 | !TTI.forceScalarizeMaskedGather(VecTy, CommonAlignment)) |
3599 | return LoadsState::ScatterVectorize; |
3600 | } |
3601 | } |
3602 | |
3603 | return LoadsState::Gather; |
3604 | } |
3605 | |
3606 | bool clusterSortPtrAccesses(ArrayRef<Value *> VL, Type *ElemTy, |
3607 | const DataLayout &DL, ScalarEvolution &SE, |
3608 | SmallVectorImpl<unsigned> &SortedIndices) { |
3609 | 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", 3611, __extension__ __PRETTY_FUNCTION__)) |
3610 | 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", 3611, __extension__ __PRETTY_FUNCTION__)) |
3611 | "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", 3611, __extension__ __PRETTY_FUNCTION__)); |
3612 | // Map from bases to a vector of (Ptr, Offset, OrigIdx), which we insert each |
3613 | // Ptr into, sort and return the sorted indices with values next to one |
3614 | // another. |
3615 | MapVector<Value *, SmallVector<std::tuple<Value *, int, unsigned>>> Bases; |
3616 | Bases[VL[0]].push_back(std::make_tuple(VL[0], 0U, 0U)); |
3617 | |
3618 | unsigned Cnt = 1; |
3619 | for (Value *Ptr : VL.drop_front()) { |
3620 | bool Found = any_of(Bases, [&](auto &Base) { |
3621 | Optional<int> Diff = |
3622 | getPointersDiff(ElemTy, Base.first, ElemTy, Ptr, DL, SE, |
3623 | /*StrictCheck=*/true); |
3624 | if (!Diff) |
3625 | return false; |
3626 | |
3627 | Base.second.emplace_back(Ptr, *Diff, Cnt++); |
3628 | return true; |
3629 | }); |
3630 | |
3631 | if (!Found) { |
3632 | // If we haven't found enough to usefully cluster, return early. |
3633 | if (Bases.size() > VL.size() / 2 - 1) |
3634 | return false; |
3635 | |
3636 | // Not found already - add a new Base |
3637 | Bases[Ptr].emplace_back(Ptr, 0, Cnt++); |
3638 | } |
3639 | } |
3640 | |
3641 | // For each of the bases sort the pointers by Offset and check if any of the |
3642 | // base become consecutively allocated. |
3643 | bool AnyConsecutive = false; |
3644 | for (auto &Base : Bases) { |
3645 | auto &Vec = Base.second; |
3646 | if (Vec.size() > 1) { |
3647 | llvm::stable_sort(Vec, [](const std::tuple<Value *, int, unsigned> &X, |
3648 | const std::tuple<Value *, int, unsigned> &Y) { |
3649 | return std::get<1>(X) < std::get<1>(Y); |
3650 | }); |
3651 | int InitialOffset = std::get<1>(Vec[0]); |
3652 | AnyConsecutive |= all_of(enumerate(Vec), [InitialOffset](auto &P) { |
3653 | return std::get<1>(P.value()) == int(P.index()) + InitialOffset; |
3654 | }); |
3655 | } |
3656 | } |
3657 | |
3658 | // Fill SortedIndices array only if it looks worth-while to sort the ptrs. |
3659 | SortedIndices.clear(); |
3660 | if (!AnyConsecutive) |
3661 | return false; |
3662 | |
3663 | for (auto &Base : Bases) { |
3664 | for (auto &T : Base.second) |
3665 | SortedIndices.push_back(std::get<2>(T)); |
3666 | } |
3667 | |
3668 | 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", 3669, __extension__ __PRETTY_FUNCTION__)) |
3669 | "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", 3669, __extension__ __PRETTY_FUNCTION__)); |
3670 | return true; |
3671 | } |
3672 | |
3673 | Optional<BoUpSLP::OrdersType> |
3674 | BoUpSLP::findPartiallyOrderedLoads(const BoUpSLP::TreeEntry &TE) { |
3675 | 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", 3675, __extension__ __PRETTY_FUNCTION__)); |
3676 | Type *ScalarTy = TE.Scalars[0]->getType(); |
3677 | |
3678 | SmallVector<Value *> Ptrs; |
3679 | Ptrs.reserve(TE.Scalars.size()); |
3680 | for (Value *V : TE.Scalars) { |
3681 | auto *L = dyn_cast<LoadInst>(V); |
3682 | if (!L || !L->isSimple()) |
3683 | return None; |
3684 | Ptrs.push_back(L->getPointerOperand()); |
3685 | } |
3686 | |
3687 | BoUpSLP::OrdersType Order; |
3688 | if (clusterSortPtrAccesses(Ptrs, ScalarTy, *DL, *SE, Order)) |
3689 | return Order; |
3690 | return None; |
3691 | } |
3692 | |
3693 | Optional<BoUpSLP::OrdersType> BoUpSLP::getReorderingData(const TreeEntry &TE, |
3694 | bool TopToBottom) { |
3695 | // No need to reorder if need to shuffle reuses, still need to shuffle the |
3696 | // node. |
3697 | if (!TE.ReuseShuffleIndices.empty()) { |
3698 | // Check if reuse shuffle indices can be improved by reordering. |
3699 | // For this, check that reuse mask is "clustered", i.e. each scalar values |
3700 | // is used once in each submask of size <number_of_scalars>. |
3701 | // Example: 4 scalar values. |
3702 | // ReuseShuffleIndices mask: 0, 1, 2, 3, 3, 2, 0, 1 - clustered. |
3703 | // 0, 1, 2, 3, 3, 3, 1, 0 - not clustered, because |
3704 | // element 3 is used twice in the second submask. |
3705 | unsigned Sz = TE.Scalars.size(); |
3706 | if (!ShuffleVectorInst::isOneUseSingleSourceMask(TE.ReuseShuffleIndices, |
3707 | Sz)) |
3708 | return None; |
3709 | unsigned VF = TE.getVectorFactor(); |
3710 | // Try build correct order for extractelement instructions. |
3711 | SmallVector<int> ReusedMask(TE.ReuseShuffleIndices.begin(), |
3712 | TE.ReuseShuffleIndices.end()); |
3713 | if (TE.getOpcode() == Instruction::ExtractElement && !TE.isAltShuffle() && |
3714 | all_of(TE.Scalars, [Sz](Value *V) { |
3715 | Optional<unsigned> Idx = getExtractIndex(cast<Instruction>(V)); |
3716 | return Idx && *Idx < Sz; |
3717 | })) { |
3718 | SmallVector<int> ReorderMask(Sz, UndefMaskElem); |
3719 | if (TE.ReorderIndices.empty()) |
3720 | std::iota(ReorderMask.begin(), ReorderMask.end(), 0); |
3721 | else |
3722 | inversePermutation(TE.ReorderIndices, ReorderMask); |
3723 | for (unsigned I = 0; I < VF; ++I) { |
3724 | int &Idx = ReusedMask[I]; |
3725 | if (Idx == UndefMaskElem) |
3726 | continue; |
3727 | Value *V = TE.Scalars[ReorderMask[Idx]]; |
3728 | Optional<unsigned> EI = getExtractIndex(cast<Instruction>(V)); |
3729 | Idx = std::distance(ReorderMask.begin(), find(ReorderMask, *EI)); |
3730 | } |
3731 | } |
3732 | // Build the order of the VF size, need to reorder reuses shuffles, they are |
3733 | // always of VF size. |
3734 | OrdersType ResOrder(VF); |
3735 | std::iota(ResOrder.begin(), ResOrder.end(), 0); |
3736 | auto *It = ResOrder.begin(); |
3737 | for (unsigned K = 0; K < VF; K += Sz) { |
3738 | OrdersType CurrentOrder(TE.ReorderIndices); |
3739 | SmallVector<int> SubMask(makeArrayRef(ReusedMask).slice(K, Sz)); |
3740 | if (SubMask.front() == UndefMaskElem) |
3741 | std::iota(SubMask.begin(), SubMask.end(), 0); |
3742 | reorderOrder(CurrentOrder, SubMask); |
3743 | transform(CurrentOrder, It, [K](unsigned Pos) { return Pos + K; }); |
3744 | std::advance(It, Sz); |
3745 | } |
3746 | if (all_of(enumerate(ResOrder), |
3747 | [](const auto &Data) { return Data.index() == Data.value(); })) |
3748 | return {}; // Use identity order. |
3749 | return ResOrder; |
3750 | } |
3751 | if (TE.State == TreeEntry::Vectorize && |
3752 | (isa<LoadInst, ExtractElementInst, ExtractValueInst>(TE.getMainOp()) || |
3753 | (TopToBottom && isa<StoreInst, InsertElementInst>(TE.getMainOp()))) && |
3754 | !TE.isAltShuffle()) |
3755 | return TE.ReorderIndices; |
3756 | if (TE.State == TreeEntry::NeedToGather) { |
3757 | // TODO: add analysis of other gather nodes with extractelement |
3758 | // instructions and other values/instructions, not only undefs. |
3759 | if (((TE.getOpcode() == Instruction::ExtractElement && |
3760 | !TE.isAltShuffle()) || |
3761 | (all_of(TE.Scalars, |
3762 | [](Value *V) { |
3763 | return isa<UndefValue, ExtractElementInst>(V); |
3764 | }) && |
3765 | any_of(TE.Scalars, |
3766 | [](Value *V) { return isa<ExtractElementInst>(V); }))) && |
3767 | all_of(TE.Scalars, |
3768 | [](Value *V) { |
3769 | auto *EE = dyn_cast<ExtractElementInst>(V); |
3770 | return !EE || isa<FixedVectorType>(EE->getVectorOperandType()); |
3771 | }) && |
3772 | allSameType(TE.Scalars)) { |
3773 | // Check that gather of extractelements can be represented as |
3774 | // just a shuffle of a single vector. |
3775 | OrdersType CurrentOrder; |
3776 | bool Reuse = canReuseExtract(TE.Scalars, TE.getMainOp(), CurrentOrder); |
3777 | if (Reuse || !CurrentOrder.empty()) { |
3778 | if (!CurrentOrder.empty()) |
3779 | fixupOrderingIndices(CurrentOrder); |
3780 | return CurrentOrder; |
3781 | } |
3782 | } |
3783 | if (Optional<OrdersType> CurrentOrder = findReusedOrderedScalars(TE)) |
3784 | return CurrentOrder; |
3785 | if (TE.Scalars.size() >= 4) |
3786 | if (Optional<OrdersType> Order = findPartiallyOrderedLoads(TE)) |
3787 | return Order; |
3788 | } |
3789 | return None; |
3790 | } |
3791 | |
3792 | /// Checks if the given mask is a "clustered" mask with the same clusters of |
3793 | /// size \p Sz, which are not identity submasks. |
3794 | static bool isRepeatedNonIdentityClusteredMask(ArrayRef<int> Mask, |
3795 | unsigned Sz) { |
3796 | ArrayRef<int> FirstCluster = Mask.slice(0, Sz); |
3797 | if (ShuffleVectorInst::isIdentityMask(FirstCluster)) |
3798 | return false; |
3799 | for (unsigned I = 0, E = Mask.size(); I < E; I += Sz) { |
3800 | ArrayRef<int> Cluster = Mask.slice(I, Sz); |
3801 | if (Cluster != FirstCluster) |
3802 | return false; |
3803 | } |
3804 | return true; |
3805 | } |
3806 | |
3807 | void BoUpSLP::reorderNodeWithReuses(TreeEntry &TE, ArrayRef<int> Mask) const { |
3808 | // For vectorized and non-clustered reused - just reorder reuses mask. |
3809 | const unsigned Sz = TE.Scalars.size(); |
3810 | if (TE.State != TreeEntry::NeedToGather || !TE.ReorderIndices.empty() || |
3811 | !ShuffleVectorInst::isOneUseSingleSourceMask(TE.ReuseShuffleIndices, |
3812 | Sz) || |
3813 | !isRepeatedNonIdentityClusteredMask(TE.ReuseShuffleIndices, Sz)) { |
3814 | reorderReuses(TE.ReuseShuffleIndices, Mask); |
3815 | return; |
3816 | } |
3817 | // Try to improve gathered nodes with clustered reuses, if possible. |
3818 | reorderScalars(TE.Scalars, makeArrayRef(TE.ReuseShuffleIndices).slice(0, Sz)); |
3819 | // Fill the reuses mask with the identity submasks. |
3820 | for (auto It = TE.ReuseShuffleIndices.begin(), |
3821 | End = TE.ReuseShuffleIndices.end(); |
3822 | It != End; std::advance(It, Sz)) |
3823 | std::iota(It, std::next(It + Sz), 0); |
3824 | } |
3825 | |
3826 | void BoUpSLP::reorderTopToBottom() { |
3827 | // Maps VF to the graph nodes. |
3828 | DenseMap<unsigned, SetVector<TreeEntry *>> VFToOrderedEntries; |
3829 | // ExtractElement gather nodes which can be vectorized and need to handle |
3830 | // their ordering. |
3831 | DenseMap<const TreeEntry *, OrdersType> GathersToOrders; |
3832 | |
3833 | // AltShuffles can also have a preferred ordering that leads to fewer |
3834 | // instructions, e.g., the addsub instruction in x86. |
3835 | DenseMap<const TreeEntry *, OrdersType> AltShufflesToOrders; |
3836 | |
3837 | // Maps a TreeEntry to the reorder indices of external users. |
3838 | DenseMap<const TreeEntry *, SmallVector<OrdersType, 1>> |
3839 | ExternalUserReorderMap; |
3840 | // FIXME: Workaround for syntax error reported by MSVC buildbots. |
3841 | TargetTransformInfo &TTIRef = *TTI; |
3842 | // Find all reorderable nodes with the given VF. |
3843 | // Currently the are vectorized stores,loads,extracts + some gathering of |
3844 | // extracts. |
3845 | for_each(VectorizableTree, [this, &TTIRef, &VFToOrderedEntries, |
3846 | &GathersToOrders, &ExternalUserReorderMap, |
3847 | &AltShufflesToOrders]( |
3848 | const std::unique_ptr<TreeEntry> &TE) { |
3849 | // Look for external users that will probably be vectorized. |
3850 | SmallVector<OrdersType, 1> ExternalUserReorderIndices = |
3851 | findExternalStoreUsersReorderIndices(TE.get()); |
3852 | if (!ExternalUserReorderIndices.empty()) { |
3853 | VFToOrderedEntries[TE->Scalars.size()].insert(TE.get()); |
3854 | ExternalUserReorderMap.try_emplace(TE.get(), |
3855 | std::move(ExternalUserReorderIndices)); |
3856 | } |
3857 | |
3858 | // Patterns like [fadd,fsub] can be combined into a single instruction in |
3859 | // x86. Reordering them into [fsub,fadd] blocks this pattern. So we need |
3860 | // to take into account their order when looking for the most used order. |
3861 | if (TE->isAltShuffle()) { |
3862 | VectorType *VecTy = |
3863 | FixedVectorType::get(TE->Scalars[0]->getType(), TE->Scalars.size()); |
3864 | unsigned Opcode0 = TE->getOpcode(); |
3865 | unsigned Opcode1 = TE->getAltOpcode(); |
3866 | // The opcode mask selects between the two opcodes. |
3867 | SmallBitVector OpcodeMask(TE->Scalars.size(), false); |
3868 | for (unsigned Lane : seq<unsigned>(0, TE->Scalars.size())) |
3869 | if (cast<Instruction>(TE->Scalars[Lane])->getOpcode() == Opcode1) |
3870 | OpcodeMask.set(Lane); |
3871 | // If this pattern is supported by the target then we consider the order. |
3872 | if (TTIRef.isLegalAltInstr(VecTy, Opcode0, Opcode1, OpcodeMask)) { |
3873 | VFToOrderedEntries[TE->Scalars.size()].insert(TE.get()); |
3874 | AltShufflesToOrders.try_emplace(TE.get(), OrdersType()); |
3875 | } |
3876 | // TODO: Check the reverse order too. |
3877 | } |
3878 | |
3879 | if (Optional<OrdersType> CurrentOrder = |
3880 | getReorderingData(*TE, /*TopToBottom=*/true)) { |
3881 | // Do not include ordering for nodes used in the alt opcode vectorization, |
3882 | // better to reorder them during bottom-to-top stage. If follow the order |
3883 | // here, it causes reordering of the whole graph though actually it is |
3884 | // profitable just to reorder the subgraph that starts from the alternate |
3885 | // opcode vectorization node. Such nodes already end-up with the shuffle |
3886 | // instruction and it is just enough to change this shuffle rather than |
3887 | // rotate the scalars for the whole graph. |
3888 | unsigned Cnt = 0; |
3889 | const TreeEntry *UserTE = TE.get(); |
3890 | while (UserTE && Cnt < RecursionMaxDepth) { |
3891 | if (UserTE->UserTreeIndices.size() != 1) |
3892 | break; |
3893 | if (all_of(UserTE->UserTreeIndices, [](const EdgeInfo &EI) { |
3894 | return EI.UserTE->State == TreeEntry::Vectorize && |
3895 | EI.UserTE->isAltShuffle() && EI.UserTE->Idx != 0; |
3896 | })) |
3897 | return; |
3898 | UserTE = UserTE->UserTreeIndices.back().UserTE; |
3899 | ++Cnt; |
3900 | } |
3901 | VFToOrderedEntries[TE->getVectorFactor()].insert(TE.get()); |
3902 | if (TE->State != TreeEntry::Vectorize || !TE->ReuseShuffleIndices.empty()) |
3903 | GathersToOrders.try_emplace(TE.get(), *CurrentOrder); |
3904 | } |
3905 | }); |
3906 | |
3907 | // Reorder the graph nodes according to their vectorization factor. |
3908 | for (unsigned VF = VectorizableTree.front()->Scalars.size(); VF > 1; |
3909 | VF /= 2) { |
3910 | auto It = VFToOrderedEntries.find(VF); |
3911 | if (It == VFToOrderedEntries.end()) |
3912 | continue; |
3913 | // Try to find the most profitable order. We just are looking for the most |
3914 | // used order and reorder scalar elements in the nodes according to this |
3915 | // mostly used order. |
3916 | ArrayRef<TreeEntry *> OrderedEntries = It->second.getArrayRef(); |
3917 | // All operands are reordered and used only in this node - propagate the |
3918 | // most used order to the user node. |
3919 | MapVector<OrdersType, unsigned, |
3920 | DenseMap<OrdersType, unsigned, OrdersTypeDenseMapInfo>> |
3921 | OrdersUses; |
3922 | SmallPtrSet<const TreeEntry *, 4> VisitedOps; |
3923 | for (const TreeEntry *OpTE : OrderedEntries) { |
3924 | // No need to reorder this nodes, still need to extend and to use shuffle, |
3925 | // just need to merge reordering shuffle and the reuse shuffle. |
3926 | if (!OpTE->ReuseShuffleIndices.empty() && !GathersToOrders.count(OpTE)) |
3927 | continue; |
3928 | // Count number of orders uses. |
3929 | const auto &Order = [OpTE, &GathersToOrders, |
3930 | &AltShufflesToOrders]() -> const OrdersType & { |
3931 | if (OpTE->State == TreeEntry::NeedToGather || |
3932 | !OpTE->ReuseShuffleIndices.empty()) { |
3933 | auto It = GathersToOrders.find(OpTE); |
3934 | if (It != GathersToOrders.end()) |
3935 | return It->second; |
3936 | } |
3937 | if (OpTE->isAltShuffle()) { |
3938 | auto It = AltShufflesToOrders.find(OpTE); |
3939 | if (It != AltShufflesToOrders.end()) |
3940 | return It->second; |
3941 | } |
3942 | return OpTE->ReorderIndices; |
3943 | }(); |
3944 | // First consider the order of the external scalar users. |
3945 | auto It = ExternalUserReorderMap.find(OpTE); |
3946 | if (It != ExternalUserReorderMap.end()) { |
3947 | const auto &ExternalUserReorderIndices = It->second; |
3948 | // If the OpTE vector factor != number of scalars - use natural order, |
3949 | // it is an attempt to reorder node with reused scalars but with |
3950 | // external uses. |
3951 | if (OpTE->getVectorFactor() != OpTE->Scalars.size()) { |
3952 | OrdersUses.insert(std::make_pair(OrdersType(), 0)).first->second += |
3953 | ExternalUserReorderIndices.size(); |
3954 | } else { |
3955 | for (const OrdersType &ExtOrder : ExternalUserReorderIndices) |
3956 | ++OrdersUses.insert(std::make_pair(ExtOrder, 0)).first->second; |
3957 | } |
3958 | // No other useful reorder data in this entry. |
3959 | if (Order.empty()) |
3960 | continue; |
3961 | } |
3962 | // Stores actually store the mask, not the order, need to invert. |
3963 | if (OpTE->State == TreeEntry::Vectorize && !OpTE->isAltShuffle() && |
3964 | OpTE->getOpcode() == Instruction::Store && !Order.empty()) { |
3965 | SmallVector<int> Mask; |
3966 | inversePermutation(Order, Mask); |
3967 | unsigned E = Order.size(); |
3968 | OrdersType CurrentOrder(E, E); |
3969 | transform(Mask, CurrentOrder.begin(), [E](int Idx) { |
3970 | return Idx == UndefMaskElem ? E : static_cast<unsigned>(Idx); |
3971 | }); |
3972 | fixupOrderingIndices(CurrentOrder); |
3973 | ++OrdersUses.insert(std::make_pair(CurrentOrder, 0)).first->second; |
3974 | } else { |
3975 | ++OrdersUses.insert(std::make_pair(Order, 0)).first->second; |
3976 | } |
3977 | } |
3978 | // Set order of the user node. |
3979 | if (OrdersUses.empty()) |
3980 | continue; |
3981 | // Choose the most used order. |
3982 | ArrayRef<unsigned> BestOrder = OrdersUses.front().first; |
3983 | unsigned Cnt = OrdersUses.front().second; |
3984 | for (const auto &Pair : drop_begin(OrdersUses)) { |
3985 | if (Cnt < Pair.second || (Cnt == Pair.second && Pair.first.empty())) { |
3986 | BestOrder = Pair.first; |
3987 | Cnt = Pair.second; |
3988 | } |
3989 | } |
3990 | // Set order of the user node. |
3991 | if (BestOrder.empty()) |
3992 | continue; |
3993 | SmallVector<int> Mask; |
3994 | inversePermutation(BestOrder, Mask); |
3995 | SmallVector<int> MaskOrder(BestOrder.size(), UndefMaskElem); |
3996 | unsigned E = BestOrder.size(); |
3997 | transform(BestOrder, MaskOrder.begin(), [E](unsigned I) { |
3998 | return I < E ? static_cast<int>(I) : UndefMaskElem; |
3999 | }); |
4000 | // Do an actual reordering, if profitable. |
4001 | for (std::unique_ptr<TreeEntry> &TE : VectorizableTree) { |
4002 | // Just do the reordering for the nodes with the given VF. |
4003 | if (TE->Scalars.size() != VF) { |
4004 | if (TE->ReuseShuffleIndices.size() == VF) { |
4005 | // Need to reorder the reuses masks of the operands with smaller VF to |
4006 | // be able to find the match between the graph nodes and scalar |
4007 | // operands of the given node during vectorization/cost estimation. |
4008 | 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", 4014, __extension__ __PRETTY_FUNCTION__)) |
4009 | [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", 4014, __extension__ __PRETTY_FUNCTION__)) |
4010 | 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", 4014, __extension__ __PRETTY_FUNCTION__)) |
4011 | 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", 4014, __extension__ __PRETTY_FUNCTION__)) |
4012 | 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", 4014, __extension__ __PRETTY_FUNCTION__)) |
4013 | }) &&(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", 4014, __extension__ __PRETTY_FUNCTION__)) |
4014 | "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", 4014, __extension__ __PRETTY_FUNCTION__)); |
4015 | // Update ordering of the operands with the smaller VF than the given |
4016 | // one. |
4017 | reorderNodeWithReuses(*TE, Mask); |
4018 | } |
4019 | continue; |
4020 | } |
4021 | if (TE->State == TreeEntry::Vectorize && |
4022 | isa<ExtractElementInst, ExtractValueInst, LoadInst, StoreInst, |
4023 | InsertElementInst>(TE->getMainOp()) && |
4024 | !TE->isAltShuffle()) { |
4025 | // Build correct orders for extract{element,value}, loads and |
4026 | // stores. |
4027 | reorderOrder(TE->ReorderIndices, Mask); |
4028 | if (isa<InsertElementInst, StoreInst>(TE->getMainOp())) |
4029 | TE->reorderOperands(Mask); |
4030 | } else { |
4031 | // Reorder the node and its operands. |
4032 | TE->reorderOperands(Mask); |
4033 | 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", 4034, __extension__ __PRETTY_FUNCTION__)) |
4034 | "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", 4034, __extension__ __PRETTY_FUNCTION__)); |
4035 | reorderScalars(TE->Scalars, Mask); |
4036 | } |
4037 | if (!TE->ReuseShuffleIndices.empty()) { |
4038 | // Apply reversed order to keep the original ordering of the reused |
4039 | // elements to avoid extra reorder indices shuffling. |
4040 | OrdersType CurrentOrder; |
4041 | reorderOrder(CurrentOrder, MaskOrder); |
4042 | SmallVector<int> NewReuses; |
4043 | inversePermutation(CurrentOrder, NewReuses); |
4044 | addMask(NewReuses, TE->ReuseShuffleIndices); |
4045 | TE->ReuseShuffleIndices.swap(NewReuses); |
4046 | } |
4047 | } |
4048 | } |
4049 | } |
4050 | |
4051 | bool BoUpSLP::canReorderOperands( |
4052 | TreeEntry *UserTE, SmallVectorImpl<std::pair<unsigned, TreeEntry *>> &Edges, |
4053 | ArrayRef<TreeEntry *> ReorderableGathers, |
4054 | SmallVectorImpl<TreeEntry *> &GatherOps) { |
4055 | for (unsigned I = 0, E = UserTE->getNumOperands(); I < E; ++I) { |
4056 | if (any_of(Edges, [I](const std::pair<unsigned, TreeEntry *> &OpData) { |
4057 | return OpData.first == I && |
4058 | OpData.second->State == TreeEntry::Vectorize; |
4059 | })) |
4060 | continue; |
4061 | if (TreeEntry *TE = getVectorizedOperand(UserTE, I)) { |
4062 | // Do not reorder if operand node is used by many user nodes. |
4063 | if (any_of(TE->UserTreeIndices, |
4064 | [UserTE](const EdgeInfo &EI) { return EI.UserTE != UserTE; })) |
4065 | return false; |
4066 | // Add the node to the list of the ordered nodes with the identity |
4067 | // order. |
4068 | Edges.emplace_back(I, TE); |
4069 | // Add ScatterVectorize nodes to the list of operands, where just |
4070 | // reordering of the scalars is required. Similar to the gathers, so |
4071 | // simply add to the list of gathered ops. |
4072 | // If there are reused scalars, process this node as a regular vectorize |
4073 | // node, just reorder reuses mask. |
4074 | if (TE->State != TreeEntry::Vectorize && TE->ReuseShuffleIndices.empty()) |
4075 | GatherOps.push_back(TE); |
4076 | continue; |
4077 | } |
4078 | TreeEntry *Gather = nullptr; |
4079 | if (count_if(ReorderableGathers, |
4080 | [&Gather, UserTE, I](TreeEntry *TE) { |
4081 | 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", 4082, __extension__ __PRETTY_FUNCTION__)) |
4082 | "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", 4082, __extension__ __PRETTY_FUNCTION__)); |
4083 | if (any_of(TE->UserTreeIndices, |
4084 | [UserTE, I](const EdgeInfo &EI) { |
4085 | return EI.UserTE == UserTE && EI.EdgeIdx == I; |
4086 | })) { |
4087 | 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", 4088, __extension__ __PRETTY_FUNCTION__)) |
4088 | "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", 4088, __extension__ __PRETTY_FUNCTION__)); |
4089 | Gather = TE; |
4090 | return true; |
4091 | } |
4092 | return false; |
4093 | }) > 1 && |
4094 | !all_of(UserTE->getOperand(I), isConstant)) |
4095 | return false; |
4096 | if (Gather) |
4097 | GatherOps.push_back(Gather); |
4098 | } |
4099 | return true; |
4100 | } |
4101 | |
4102 | void BoUpSLP::reorderBottomToTop(bool IgnoreReorder) { |
4103 | SetVector<TreeEntry *> OrderedEntries; |
4104 | DenseMap<const TreeEntry *, OrdersType> GathersToOrders; |
4105 | // Find all reorderable leaf nodes with the given VF. |
4106 | // Currently the are vectorized loads,extracts without alternate operands + |
4107 | // some gathering of extracts. |
4108 | SmallVector<TreeEntry *> NonVectorized; |
4109 | for_each(VectorizableTree, [this, &OrderedEntries, &GathersToOrders, |
4110 | &NonVectorized]( |
4111 | const std::unique_ptr<TreeEntry> &TE) { |
4112 | if (TE->State != TreeEntry::Vectorize) |
4113 | NonVectorized.push_back(TE.get()); |
4114 | if (Optional<OrdersType> CurrentOrder = |
4115 | getReorderingData(*TE, /*TopToBottom=*/false)) { |
4116 | OrderedEntries.insert(TE.get()); |
4117 | if (TE->State != TreeEntry::Vectorize || !TE->ReuseShuffleIndices.empty()) |
4118 | GathersToOrders.try_emplace(TE.get(), *CurrentOrder); |
4119 | } |
4120 | }); |
4121 | |
4122 | // 1. Propagate order to the graph nodes, which use only reordered nodes. |
4123 | // I.e., if the node has operands, that are reordered, try to make at least |
4124 | // one operand order in the natural order and reorder others + reorder the |
4125 | // user node itself. |
4126 | SmallPtrSet<const TreeEntry *, 4> Visited; |
4127 | while (!OrderedEntries.empty()) { |
4128 | // 1. Filter out only reordered nodes. |
4129 | // 2. If the entry has multiple uses - skip it and jump to the next node. |
4130 | DenseMap<TreeEntry *, SmallVector<std::pair<unsigned, TreeEntry *>>> Users; |
4131 | SmallVector<TreeEntry *> Filtered; |
4132 | for (TreeEntry *TE : OrderedEntries) { |
4133 | if (!(TE->State == TreeEntry::Vectorize || |
4134 | (TE->State == TreeEntry::NeedToGather && |
4135 | GathersToOrders.count(TE))) || |
4136 | TE->UserTreeIndices.empty() || !TE->ReuseShuffleIndices.empty() || |
4137 | !all_of(drop_begin(TE->UserTreeIndices), |
4138 | [TE](const EdgeInfo &EI) { |
4139 | return EI.UserTE == TE->UserTreeIndices.front().UserTE; |
4140 | }) || |
4141 | !Visited.insert(TE).second) { |
4142 | Filtered.push_back(TE); |
4143 | continue; |
4144 | } |
4145 | // Build a map between user nodes and their operands order to speedup |
4146 | // search. The graph currently does not provide this dependency directly. |
4147 | for (EdgeInfo &EI : TE->UserTreeIndices) { |
4148 | TreeEntry *UserTE = EI.UserTE; |
4149 | auto It = Users.find(UserTE); |
4150 | if (It == Users.end()) |
4151 | It = Users.insert({UserTE, {}}).first; |
4152 | It->second.emplace_back(EI.EdgeIdx, TE); |
4153 | } |
4154 | } |
4155 | // Erase filtered entries. |
4156 | for_each(Filtered, |
4157 | [&OrderedEntries](TreeEntry *TE) { OrderedEntries.remove(TE); }); |
4158 | SmallVector< |
4159 | std::pair<TreeEntry *, SmallVector<std::pair<unsigned, TreeEntry *>>>> |
4160 | UsersVec(Users.begin(), Users.end()); |
4161 | sort(UsersVec, [](const auto &Data1, const auto &Data2) { |
4162 | return Data1.first->Idx > Data2.first->Idx; |
4163 | }); |
4164 | for (auto &Data : UsersVec) { |
4165 | // Check that operands are used only in the User node. |
4166 | SmallVector<TreeEntry *> GatherOps; |
4167 | if (!canReorderOperands(Data.first, Data.second, NonVectorized, |
4168 | GatherOps)) { |
4169 | for_each(Data.second, |
4170 | [&OrderedEntries](const std::pair<unsigned, TreeEntry *> &Op) { |
4171 | OrderedEntries.remove(Op.second); |
4172 | }); |
4173 | continue; |
4174 | } |
4175 | // All operands are reordered and used only in this node - propagate the |
4176 | // most used order to the user node. |
4177 | MapVector<OrdersType, unsigned, |
4178 | DenseMap<OrdersType, unsigned, OrdersTypeDenseMapInfo>> |
4179 | OrdersUses; |
4180 | // Do the analysis for each tree entry only once, otherwise the order of |
4181 | // the same node my be considered several times, though might be not |
4182 | // profitable. |
4183 | SmallPtrSet<const TreeEntry *, 4> VisitedOps; |
4184 | SmallPtrSet<const TreeEntry *, 4> VisitedUsers; |
4185 | for (const auto &Op : Data.second) { |
4186 | TreeEntry *OpTE = Op.second; |
4187 | if (!VisitedOps.insert(OpTE).second) |
4188 | continue; |
4189 | if (!OpTE->ReuseShuffleIndices.empty() && !GathersToOrders.count(OpTE)) |
4190 | continue; |
4191 | const auto &Order = [OpTE, &GathersToOrders]() -> const OrdersType & { |
4192 | if (OpTE->State == TreeEntry::NeedToGather || |
4193 | !OpTE->ReuseShuffleIndices.empty()) |
4194 | return GathersToOrders.find(OpTE)->second; |
4195 | return OpTE->ReorderIndices; |
4196 | }(); |
4197 | unsigned NumOps = count_if( |
4198 | Data.second, [OpTE](const std::pair<unsigned, TreeEntry *> &P) { |
4199 | return P.second == OpTE; |
4200 | }); |
4201 | // Stores actually store the mask, not the order, need to invert. |
4202 | if (OpTE->State == TreeEntry::Vectorize && !OpTE->isAltShuffle() && |
4203 | OpTE->getOpcode() == Instruction::Store && !Order.empty()) { |
4204 | SmallVector<int> Mask; |
4205 | inversePermutation(Order, Mask); |
4206 | unsigned E = Order.size(); |
4207 | OrdersType CurrentOrder(E, E); |
4208 | transform(Mask, CurrentOrder.begin(), [E](int Idx) { |
4209 | return Idx == UndefMaskElem ? E : static_cast<unsigned>(Idx); |
4210 | }); |
4211 | fixupOrderingIndices(CurrentOrder); |
4212 | OrdersUses.insert(std::make_pair(CurrentOrder, 0)).first->second += |
4213 | NumOps; |
4214 | } else { |
4215 | OrdersUses.insert(std::make_pair(Order, 0)).first->second += NumOps; |
4216 | } |
4217 | auto Res = OrdersUses.insert(std::make_pair(OrdersType(), 0)); |
4218 | const auto &&AllowsReordering = [IgnoreReorder, &GathersToOrders]( |
4219 | const TreeEntry *TE) { |
4220 | if (!TE->ReorderIndices.empty() || !TE->ReuseShuffleIndices.empty() || |
4221 | (TE->State == TreeEntry::Vectorize && TE->isAltShuffle()) || |
4222 | (IgnoreReorder && TE->Idx == 0)) |
4223 | return true; |
4224 | if (TE->State == TreeEntry::NeedToGather) { |
4225 | auto It = GathersToOrders.find(TE); |
4226 | if (It != GathersToOrders.end()) |
4227 | return !It->second.empty(); |
4228 | return true; |
4229 | } |
4230 | return false; |
4231 | }; |
4232 | for (const EdgeInfo &EI : OpTE->UserTreeIndices) { |
4233 | TreeEntry *UserTE = EI.UserTE; |
4234 | if (!VisitedUsers.insert(UserTE).second) |
4235 | continue; |
4236 | // May reorder user node if it requires reordering, has reused |
4237 | // scalars, is an alternate op vectorize node or its op nodes require |
4238 | // reordering. |
4239 | if (AllowsReordering(UserTE)) |
4240 | continue; |
4241 | // Check if users allow reordering. |
4242 | // Currently look up just 1 level of operands to avoid increase of |
4243 | // the compile time. |
4244 | // Profitable to reorder if definitely more operands allow |
4245 | // reordering rather than those with natural order. |
4246 | ArrayRef<std::pair<unsigned, TreeEntry *>> Ops = Users[UserTE]; |
4247 | if (static_cast<unsigned>(count_if( |
4248 | Ops, [UserTE, &AllowsReordering]( |
4249 | const std::pair<unsigned, TreeEntry *> &Op) { |
4250 | return AllowsReordering(Op.second) && |
4251 | all_of(Op.second->UserTreeIndices, |
4252 | [UserTE](const EdgeInfo &EI) { |
4253 | return EI.UserTE == UserTE; |
4254 | }); |
4255 | })) <= Ops.size() / 2) |
4256 | ++Res.first->second; |
4257 | } |
4258 | } |
4259 | // If no orders - skip current nodes and jump to the next one, if any. |
4260 | if (OrdersUses.empty()) { |
4261 | for_each(Data.second, |
4262 | [&OrderedEntries](const std::pair<unsigned, TreeEntry *> &Op) { |
4263 | OrderedEntries.remove(Op.second); |
4264 | }); |
4265 | continue; |
4266 | } |
4267 | // Choose the best order. |
4268 | ArrayRef<unsigned> BestOrder = OrdersUses.front().first; |
4269 | unsigned Cnt = OrdersUses.front().second; |
4270 | for (const auto &Pair : drop_begin(OrdersUses)) { |
4271 | if (Cnt < Pair.second || (Cnt == Pair.second && Pair.first.empty())) { |
4272 | BestOrder = Pair.first; |
4273 | Cnt = Pair.second; |
4274 | } |
4275 | } |
4276 | // Set order of the user node (reordering of operands and user nodes). |
4277 | if (BestOrder.empty()) { |
4278 | for_each(Data.second, |
4279 | [&OrderedEntries](const std::pair<unsigned, TreeEntry *> &Op) { |
4280 | OrderedEntries.remove(Op.second); |
4281 | }); |
4282 | continue; |
4283 | } |
4284 | // Erase operands from OrderedEntries list and adjust their orders. |
4285 | VisitedOps.clear(); |
4286 | SmallVector<int> Mask; |
4287 | inversePermutation(BestOrder, Mask); |
4288 | SmallVector<int> MaskOrder(BestOrder.size(), UndefMaskElem); |
4289 | unsigned E = BestOrder.size(); |
4290 | transform(BestOrder, MaskOrder.begin(), [E](unsigned I) { |
4291 | return I < E ? static_cast<int>(I) : UndefMaskElem; |
4292 | }); |
4293 | for (const std::pair<unsigned, TreeEntry *> &Op : Data.second) { |
4294 | TreeEntry *TE = Op.second; |
4295 | OrderedEntries.remove(TE); |
4296 | if (!VisitedOps.insert(TE).second) |
4297 | continue; |
4298 | if (TE->ReuseShuffleIndices.size() == BestOrder.size()) { |
4299 | reorderNodeWithReuses(*TE, Mask); |
4300 | continue; |
4301 | } |
4302 | // Gathers are processed separately. |
4303 | if (TE->State != TreeEntry::Vectorize) |
4304 | continue; |
4305 | 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", 4307, __extension__ __PRETTY_FUNCTION__)) |
4306 | 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", 4307, __extension__ __PRETTY_FUNCTION__)) |
4307 | "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", 4307, __extension__ __PRETTY_FUNCTION__)); |
4308 | reorderOrder(TE->ReorderIndices, Mask); |
4309 | if (IgnoreReorder && TE == VectorizableTree.front().get()) |
4310 | IgnoreReorder = false; |
4311 | } |
4312 | // For gathers just need to reorder its scalars. |
4313 | for (TreeEntry *Gather : GatherOps) { |
4314 | 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", 4315, __extension__ __PRETTY_FUNCTION__)) |
4315 | "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", 4315, __extension__ __PRETTY_FUNCTION__)); |
4316 | if (!Gather->ReuseShuffleIndices.empty()) { |
4317 | // Just reorder reuses indices. |
4318 | reorderReuses(Gather->ReuseShuffleIndices, Mask); |
4319 | continue; |
4320 | } |
4321 | reorderScalars(Gather->Scalars, Mask); |
4322 | OrderedEntries.remove(Gather); |
4323 | } |
4324 | // Reorder operands of the user node and set the ordering for the user |
4325 | // node itself. |
4326 | if (Data.first->State != TreeEntry::Vectorize || |
4327 | !isa<ExtractElementInst, ExtractValueInst, LoadInst>( |
4328 | Data.first->getMainOp()) || |
4329 | Data.first->isAltShuffle()) |
4330 | Data.first->reorderOperands(Mask); |
4331 | if (!isa<InsertElementInst, StoreInst>(Data.first->getMainOp()) || |
4332 | Data.first->isAltShuffle()) { |
4333 | reorderScalars(Data.first->Scalars, Mask); |
4334 | reorderOrder(Data.first->ReorderIndices, MaskOrder); |
4335 | if (Data.first->ReuseShuffleIndices.empty() && |
4336 | !Data.first->ReorderIndices.empty() && |
4337 | !Data.first->isAltShuffle()) { |
4338 | // Insert user node to the list to try to sink reordering deeper in |
4339 | // the graph. |
4340 | OrderedEntries.insert(Data.first); |
4341 | } |
4342 | } else { |
4343 | reorderOrder(Data.first->ReorderIndices, Mask); |
4344 | } |
4345 | } |
4346 | } |
4347 | // If the reordering is unnecessary, just remove the reorder. |
4348 | if (IgnoreReorder && !VectorizableTree.front()->ReorderIndices.empty() && |
4349 | VectorizableTree.front()->ReuseShuffleIndices.empty()) |
4350 | VectorizableTree.front()->ReorderIndices.clear(); |
4351 | } |
4352 | |
4353 | void BoUpSLP::buildExternalUses( |
4354 | const ExtraValueToDebugLocsMap &ExternallyUsedValues) { |
4355 | // Collect the values that we need to extract from the tree. |
4356 | for (auto &TEPtr : VectorizableTree) { |
4357 | TreeEntry *Entry = TEPtr.get(); |
4358 | |
4359 | // No need to handle users of gathered values. |
4360 | if (Entry->State == TreeEntry::NeedToGather) |
4361 | continue; |
4362 | |
4363 | // For each lane: |
4364 | for (int Lane = 0, LE = Entry->Scalars.size(); Lane != LE; ++Lane) { |
4365 | Value *Scalar = Entry->Scalars[Lane]; |
4366 | int FoundLane = Entry->findLaneForValue(Scalar); |
4367 | |
4368 | // Check if the scalar is externally used as an extra arg. |
4369 | auto ExtI = ExternallyUsedValues.find(Scalar); |
4370 | if (ExtI != ExternallyUsedValues.end()) { |
4371 | 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) |
4372 | << 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); |
4373 | ExternalUses.emplace_back(Scalar, nullptr, FoundLane); |
4374 | } |
4375 | for (User *U : Scalar->users()) { |
4376 | LLVM_DEBUG(dbgs() << "SLP: Checking user:" << *U << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Checking user:" << *U << ".\n"; } } while (false); |
4377 | |
4378 | Instruction *UserInst = dyn_cast<Instruction>(U); |
4379 | if (!UserInst) |
4380 | continue; |
4381 | |
4382 | if (isDeleted(UserInst)) |
4383 | continue; |
4384 | |
4385 | // Skip in-tree scalars that become vectors |
4386 | if (TreeEntry *UseEntry = getTreeEntry(U)) { |
4387 | Value *UseScalar = UseEntry->Scalars[0]; |
4388 | // Some in-tree scalars will remain as scalar in vectorized |
4389 | // instructions. If that is the case, the one in Lane 0 will |
4390 | // be used. |
4391 | if (UseScalar != U || |
4392 | UseEntry->State == TreeEntry::ScatterVectorize || |
4393 | !InTreeUserNeedToExtract(Scalar, UserInst, TLI)) { |
4394 | 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) |
4395 | << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: \tInternal user will be removed:" << *U << ".\n"; } } while (false); |
4396 | 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", 4396, __extension__ __PRETTY_FUNCTION__)); |
4397 | continue; |
4398 | } |
4399 | } |
4400 | |
4401 | // Ignore users in the user ignore list. |
4402 | if (UserIgnoreList && UserIgnoreList->contains(UserInst)) |
4403 | continue; |
4404 | |
4405 | 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) |
4406 | << 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); |
4407 | ExternalUses.push_back(ExternalUser(Scalar, U, FoundLane)); |
4408 | } |
4409 | } |
4410 | } |
4411 | } |
4412 | |
4413 | DenseMap<Value *, SmallVector<StoreInst *, 4>> |
4414 | BoUpSLP::collectUserStores(const BoUpSLP::TreeEntry *TE) const { |
4415 | DenseMap<Value *, SmallVector<StoreInst *, 4>> PtrToStoresMap; |
4416 | for (unsigned Lane : seq<unsigned>(0, TE->Scalars.size())) { |
4417 | Value *V = TE->Scalars[Lane]; |
4418 | // To save compilation time we don't visit if we have too many users. |
4419 | static constexpr unsigned UsersLimit = 4; |
4420 | if (V->hasNUsesOrMore(UsersLimit)) |
4421 | break; |
4422 | |
4423 | // Collect stores per pointer object. |
4424 | for (User *U : V->users()) { |
4425 | auto *SI = dyn_cast<StoreInst>(U); |
4426 | if (SI == nullptr || !SI->isSimple() || |
4427 | !isValidElementType(SI->getValueOperand()->getType())) |
4428 | continue; |
4429 | // Skip entry if already |
4430 | if (getTreeEntry(U)) |
4431 | continue; |
4432 | |
4433 | Value *Ptr = getUnderlyingObject(SI->getPointerOperand()); |
4434 | auto &StoresVec = PtrToStoresMap[Ptr]; |
4435 | // For now just keep one store per pointer object per lane. |
4436 | // TODO: Extend this to support multiple stores per pointer per lane |
4437 | if (StoresVec.size() > Lane) |
4438 | continue; |
4439 | // Skip if in different BBs. |
4440 | if (!StoresVec.empty() && |
4441 | SI->getParent() != StoresVec.back()->getParent()) |
4442 | continue; |
4443 | // Make sure that the stores are of the same type. |
4444 | if (!StoresVec.empty() && |
4445 | SI->getValueOperand()->getType() != |
4446 | StoresVec.back()->getValueOperand()->getType()) |
4447 | continue; |
4448 | StoresVec.push_back(SI); |
4449 | } |
4450 | } |
4451 | return PtrToStoresMap; |
4452 | } |
4453 | |
4454 | bool BoUpSLP::canFormVector(const SmallVector<StoreInst *, 4> &StoresVec, |
4455 | OrdersType &ReorderIndices) const { |
4456 | // We check whether the stores in StoreVec can form a vector by sorting them |
4457 | // and checking whether they are consecutive. |
4458 | |
4459 | // To avoid calling getPointersDiff() while sorting we create a vector of |
4460 | // pairs {store, offset from first} and sort this instead. |
4461 | SmallVector<std::pair<StoreInst *, int>, 4> StoreOffsetVec(StoresVec.size()); |
4462 | StoreInst *S0 = StoresVec[0]; |
4463 | StoreOffsetVec[0] = {S0, 0}; |
4464 | Type *S0Ty = S0->getValueOperand()->getType(); |
4465 | Value *S0Ptr = S0->getPointerOperand(); |
4466 | for (unsigned Idx : seq<unsigned>(1, StoresVec.size())) { |
4467 | StoreInst *SI = StoresVec[Idx]; |
4468 | Optional<int> Diff = |
4469 | getPointersDiff(S0Ty, S0Ptr, SI->getValueOperand()->getType(), |
4470 | SI->getPointerOperand(), *DL, *SE, |
4471 | /*StrictCheck=*/true); |
4472 | // We failed to compare the pointers so just abandon this StoresVec. |
4473 | if (!Diff) |
4474 | return false; |
4475 | StoreOffsetVec[Idx] = {StoresVec[Idx], *Diff}; |
4476 | } |
4477 | |
4478 | // Sort the vector based on the pointers. We create a copy because we may |
4479 | // need the original later for calculating the reorder (shuffle) indices. |
4480 | stable_sort(StoreOffsetVec, [](const std::pair<StoreInst *, int> &Pair1, |
4481 | const std::pair<StoreInst *, int> &Pair2) { |
4482 | int Offset1 = Pair1.second; |
4483 | int Offset2 = Pair2.second; |
4484 | return Offset1 < Offset2; |
4485 | }); |
4486 | |
4487 | // Check if the stores are consecutive by checking if their difference is 1. |
4488 | for (unsigned Idx : seq<unsigned>(1, StoreOffsetVec.size())) |
4489 | if (StoreOffsetVec[Idx].second != StoreOffsetVec[Idx-1].second + 1) |
4490 | return false; |
4491 | |
4492 | // Calculate the shuffle indices according to their offset against the sorted |
4493 | // StoreOffsetVec. |
4494 | ReorderIndices.reserve(StoresVec.size()); |
4495 | for (StoreInst *SI : StoresVec) { |
4496 | unsigned Idx = find_if(StoreOffsetVec, |
4497 | [SI](const std::pair<StoreInst *, int> &Pair) { |
4498 | return Pair.first == SI; |
4499 | }) - |
4500 | StoreOffsetVec.begin(); |
4501 | ReorderIndices.push_back(Idx); |
4502 | } |
4503 | // Identity order (e.g., {0,1,2,3}) is modeled as an empty OrdersType in |
4504 | // reorderTopToBottom() and reorderBottomToTop(), so we are following the |
4505 | // same convention here. |
4506 | auto IsIdentityOrder = [](const OrdersType &Order) { |
4507 | for (unsigned Idx : seq<unsigned>(0, Order.size())) |
4508 | if (Idx != Order[Idx]) |
4509 | return false; |
4510 | return true; |
4511 | }; |
4512 | if (IsIdentityOrder(ReorderIndices)) |
4513 | ReorderIndices.clear(); |
4514 | |
4515 | return true; |
4516 | } |
4517 | |
4518 | #ifndef NDEBUG |
4519 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) static void dumpOrder(const BoUpSLP::OrdersType &Order) { |
4520 | for (unsigned Idx : Order) |
4521 | dbgs() << Idx << ", "; |
4522 | dbgs() << "\n"; |
4523 | } |
4524 | #endif |
4525 | |
4526 | SmallVector<BoUpSLP::OrdersType, 1> |
4527 | BoUpSLP::findExternalStoreUsersReorderIndices(TreeEntry *TE) const { |
4528 | unsigned NumLanes = TE->Scalars.size(); |
4529 | |
4530 | DenseMap<Value *, SmallVector<StoreInst *, 4>> PtrToStoresMap = |
4531 | collectUserStores(TE); |
4532 | |
4533 | // Holds the reorder indices for each candidate store vector that is a user of |
4534 | // the current TreeEntry. |
4535 | SmallVector<OrdersType, 1> ExternalReorderIndices; |
4536 | |
4537 | // Now inspect the stores collected per pointer and look for vectorization |
4538 | // candidates. For each candidate calculate the reorder index vector and push |
4539 | // it into `ExternalReorderIndices` |
4540 | for (const auto &Pair : PtrToStoresMap) { |
4541 | auto &StoresVec = Pair.second; |
4542 | // If we have fewer than NumLanes stores, then we can't form a vector. |
4543 | if (StoresVec.size() != NumLanes) |
4544 | continue; |
4545 | |
4546 | // If the stores are not consecutive then abandon this StoresVec. |
4547 | OrdersType ReorderIndices; |
4548 | if (!canFormVector(StoresVec, ReorderIndices)) |
4549 | continue; |
4550 | |
4551 | // We now know that the scalars in StoresVec can form a vector instruction, |
4552 | // so set the reorder indices. |
4553 | ExternalReorderIndices.push_back(ReorderIndices); |
4554 | } |
4555 | return ExternalReorderIndices; |
4556 | } |
4557 | |
4558 | void BoUpSLP::buildTree(ArrayRef<Value *> Roots, |
4559 | const SmallDenseSet<Value *> &UserIgnoreLst) { |
4560 | deleteTree(); |
4561 | UserIgnoreList = &UserIgnoreLst; |
4562 | if (!allSameType(Roots)) |
4563 | return; |
4564 | buildTree_rec(Roots, 0, EdgeInfo()); |
4565 | } |
4566 | |
4567 | void BoUpSLP::buildTree(ArrayRef<Value *> Roots) { |
4568 | deleteTree(); |
4569 | if (!allSameType(Roots)) |
4570 | return; |
4571 | buildTree_rec(Roots, 0, EdgeInfo()); |
4572 | } |
4573 | |
4574 | /// \return true if the specified list of values has only one instruction that |
4575 | /// requires scheduling, false otherwise. |
4576 | #ifndef NDEBUG |
4577 | static bool needToScheduleSingleInstruction(ArrayRef<Value *> VL) { |
4578 | Value *NeedsScheduling = nullptr; |
4579 | for (Value *V : VL) { |
4580 | if (doesNotNeedToBeScheduled(V)) |
4581 | continue; |
4582 | if (!NeedsScheduling) { |
4583 | NeedsScheduling = V; |
4584 | continue; |
4585 | } |
4586 | return false; |
4587 | } |
4588 | return NeedsScheduling; |
4589 | } |
4590 | #endif |
4591 | |
4592 | /// Generates key/subkey pair for the given value to provide effective sorting |
4593 | /// of the values and better detection of the vectorizable values sequences. The |
4594 | /// keys/subkeys can be used for better sorting of the values themselves (keys) |
4595 | /// and in values subgroups (subkeys). |
4596 | static std::pair<size_t, size_t> generateKeySubkey( |
4597 | Value *V, const TargetLibraryInfo *TLI, |
4598 | function_ref<hash_code(size_t, LoadInst *)> LoadsSubkeyGenerator, |
4599 | bool AllowAlternate) { |
4600 | hash_code Key = hash_value(V->getValueID() + 2); |
4601 | hash_code SubKey = hash_value(0); |
4602 | // Sort the loads by the distance between the pointers. |
4603 | if (auto *LI = dyn_cast<LoadInst>(V)) { |
4604 | Key = hash_combine(hash_value(Instruction::Load), Key); |
4605 | if (LI->isSimple()) |
4606 | SubKey = hash_value(LoadsSubkeyGenerator(Key, LI)); |
4607 | else |
4608 | SubKey = hash_value(LI); |
4609 | } else if (isVectorLikeInstWithConstOps(V)) { |
4610 | // Sort extracts by the vector operands. |
4611 | if (isa<ExtractElementInst, UndefValue>(V)) |
4612 | Key = hash_value(Value::UndefValueVal + 1); |
4613 | if (auto *EI = dyn_cast<ExtractElementInst>(V)) { |
4614 | if (!isUndefVector(EI->getVectorOperand()) && |
4615 | !isa<UndefValue>(EI->getIndexOperand())) |
4616 | SubKey = hash_value(EI->getVectorOperand()); |
4617 | } |
4618 | } else if (auto *I = dyn_cast<Instruction>(V)) { |
4619 | // Sort other instructions just by the opcodes except for CMPInst. |
4620 | // For CMP also sort by the predicate kind. |
4621 | if ((isa<BinaryOperator, CastInst>(I)) && |
4622 | isValidForAlternation(I->getOpcode())) { |
4623 | if (AllowAlternate) |
4624 | Key = hash_value(isa<BinaryOperator>(I) ? 1 : 0); |
4625 | else |
4626 | Key = hash_combine(hash_value(I->getOpcode()), Key); |
4627 | SubKey = hash_combine( |
4628 | hash_value(I->getOpcode()), hash_value(I->getType()), |
4629 | hash_value(isa<BinaryOperator>(I) |
4630 | ? I->getType() |
4631 | : cast<CastInst>(I)->getOperand(0)->getType())); |
4632 | // For casts, look through the only operand to improve compile time. |
4633 | if (isa<CastInst>(I)) { |
4634 | std::pair<size_t, size_t> OpVals = |
4635 | generateKeySubkey(I->getOperand(0), TLI, LoadsSubkeyGenerator, |
4636 | /*=AllowAlternate*/ true); |
4637 | Key = hash_combine(OpVals.first, Key); |
4638 | SubKey = hash_combine(OpVals.first, SubKey); |
4639 | } |
4640 | } else if (auto *CI = dyn_cast<CmpInst>(I)) { |
4641 | CmpInst::Predicate Pred = CI->getPredicate(); |
4642 | if (CI->isCommutative()) |
4643 | Pred = std::min(Pred, CmpInst::getInversePredicate(Pred)); |
4644 | CmpInst::Predicate SwapPred = CmpInst::getSwappedPredicate(Pred); |
4645 | SubKey = hash_combine(hash_value(I->getOpcode()), hash_value(Pred), |
4646 | hash_value(SwapPred), |
4647 | hash_value(CI->getOperand(0)->getType())); |
4648 | } else if (auto *Call = dyn_cast<CallInst>(I)) { |
4649 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(Call, TLI); |
4650 | if (isTriviallyVectorizable(ID)) { |
4651 | SubKey = hash_combine(hash_value(I->getOpcode()), hash_value(ID)); |
4652 | } else if (!VFDatabase(*Call).getMappings(*Call).empty()) { |
4653 | SubKey = hash_combine(hash_value(I->getOpcode()), |
4654 | hash_value(Call->getCalledFunction())); |
4655 | } else { |
4656 | Key = hash_combine(hash_value(Call), Key); |
4657 | SubKey = hash_combine(hash_value(I->getOpcode()), hash_value(Call)); |
4658 | } |
4659 | for (const CallBase::BundleOpInfo &Op : Call->bundle_op_infos()) |
4660 | SubKey = hash_combine(hash_value(Op.Begin), hash_value(Op.End), |
4661 | hash_value(Op.Tag), SubKey); |
4662 | } else if (auto *Gep = dyn_cast<GetElementPtrInst>(I)) { |
4663 | if (Gep->getNumOperands() == 2 && isa<ConstantInt>(Gep->getOperand(1))) |
4664 | SubKey = hash_value(Gep->getPointerOperand()); |
4665 | else |
4666 | SubKey = hash_value(Gep); |
4667 | } else if (BinaryOperator::isIntDivRem(I->getOpcode()) && |
4668 | !isa<ConstantInt>(I->getOperand(1))) { |
4669 | // Do not try to vectorize instructions with potentially high cost. |
4670 | SubKey = hash_value(I); |
4671 | } else { |
4672 | SubKey = hash_value(I->getOpcode()); |
4673 | } |
4674 | Key = hash_combine(hash_value(I->getParent()), Key); |
4675 | } |
4676 | return std::make_pair(Key, SubKey); |
4677 | } |
4678 | |
4679 | /// Checks if the specified instruction \p I is an alternate operation for |
4680 | /// the given \p MainOp and \p AltOp instructions. |
4681 | static bool isAlternateInstruction(const Instruction *I, |
4682 | const Instruction *MainOp, |
4683 | const Instruction *AltOp); |
4684 | |
4685 | void BoUpSLP::buildTree_rec(ArrayRef<Value *> VL, unsigned Depth, |
4686 | const EdgeInfo &UserTreeIdx) { |
4687 | 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", 4687, __extension__ __PRETTY_FUNCTION__)); |
4688 | |
4689 | SmallVector<int> ReuseShuffleIndicies; |
4690 | SmallVector<Value *> UniqueValues; |
4691 | auto &&TryToFindDuplicates = [&VL, &ReuseShuffleIndicies, &UniqueValues, |
4692 | &UserTreeIdx, |
4693 | this](const InstructionsState &S) { |
4694 | // Check that every instruction appears once in this bundle. |
4695 | DenseMap<Value *, unsigned> UniquePositions; |
4696 | for (Value *V : VL) { |
4697 | if (isConstant(V)) { |
4698 | ReuseShuffleIndicies.emplace_back( |
4699 | isa<UndefValue>(V) ? UndefMaskElem : UniqueValues.size()); |
4700 | UniqueValues.emplace_back(V); |
4701 | continue; |
4702 | } |
4703 | auto Res = UniquePositions.try_emplace(V, UniqueValues.size()); |
4704 | ReuseShuffleIndicies.emplace_back(Res.first->second); |
4705 | if (Res.second) |
4706 | UniqueValues.emplace_back(V); |
4707 | } |
4708 | size_t NumUniqueScalarValues = UniqueValues.size(); |
4709 | if (NumUniqueScalarValues == VL.size()) { |
4710 | ReuseShuffleIndicies.clear(); |
4711 | } else { |
4712 | 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); |
4713 | if (NumUniqueScalarValues <= 1 || |
4714 | (UniquePositions.size() == 1 && all_of(UniqueValues, |
4715 | [](Value *V) { |
4716 | return isa<UndefValue>(V) || |
4717 | !isConstant(V); |
4718 | })) || |
4719 | !llvm::isPowerOf2_32(NumUniqueScalarValues)) { |
4720 | 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); |
4721 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx); |
4722 | return false; |
4723 | } |
4724 | VL = UniqueValues; |
4725 | } |
4726 | return true; |
4727 | }; |
4728 | |
4729 | InstructionsState S = getSameOpcode(VL); |
4730 | |
4731 | // Gather if we hit the RecursionMaxDepth, unless this is a load (or z/sext of |
4732 | // a load), in which case peek through to include it in the tree, without |
4733 | // ballooning over-budget. |
4734 | if (Depth >= RecursionMaxDepth && |
4735 | !(S.MainOp && isa<Instruction>(S.MainOp) && S.MainOp == S.AltOp && |
4736 | VL.size() >= 4 && |
4737 | (match(S.MainOp, m_Load(m_Value())) || all_of(VL, [&S](const Value *I) { |
4738 | return match(I, |
4739 | m_OneUse(m_ZExtOrSExt(m_OneUse(m_Load(m_Value()))))) && |
4740 | cast<Instruction>(I)->getOpcode() == |
4741 | cast<Instruction>(S.MainOp)->getOpcode(); |
4742 | })))) { |
4743 | 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); |
4744 | if (TryToFindDuplicates(S)) |
4745 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, |
4746 | ReuseShuffleIndicies); |
4747 | return; |
4748 | } |
4749 | |
4750 | // Don't handle scalable vectors |
4751 | if (S.getOpcode() == Instruction::ExtractElement && |
4752 | isa<ScalableVectorType>( |
4753 | cast<ExtractElementInst>(S.OpValue)->getVectorOperandType())) { |
4754 | 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); |
4755 | if (TryToFindDuplicates(S)) |
4756 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, |
4757 | ReuseShuffleIndicies); |
4758 | return; |
4759 | } |
4760 | |
4761 | // Don't handle vectors. |
4762 | if (S.OpValue->getType()->isVectorTy() && |
4763 | !isa<InsertElementInst>(S.OpValue)) { |
4764 | 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); |
4765 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx); |
4766 | return; |
4767 | } |
4768 | |
4769 | if (StoreInst *SI = dyn_cast<StoreInst>(S.OpValue)) |
4770 | if (SI->getValueOperand()->getType()->isVectorTy()) { |
4771 | 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); |
4772 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx); |
4773 | return; |
4774 | } |
4775 | |
4776 | // If all of the operands are identical or constant we have a simple solution. |
4777 | // If we deal with insert/extract instructions, they all must have constant |
4778 | // indices, otherwise we should gather them, not try to vectorize. |
4779 | // If alternate op node with 2 elements with gathered operands - do not |
4780 | // vectorize. |
4781 | auto &&NotProfitableForVectorization = [&S, this, |
4782 | Depth](ArrayRef<Value *> VL) { |
4783 | if (!S.getOpcode() || !S.isAltShuffle() || VL.size() > 2) |
4784 | return false; |
4785 | if (VectorizableTree.size() < MinTreeSize) |
4786 | return false; |
4787 | if (Depth >= RecursionMaxDepth - 1) |
4788 | return true; |
4789 | // Check if all operands are extracts, part of vector node or can build a |
4790 | // regular vectorize node. |
4791 | SmallVector<unsigned, 2> InstsCount(VL.size(), 0); |
4792 | for (Value *V : VL) { |
4793 | auto *I = cast<Instruction>(V); |
4794 | InstsCount.push_back(count_if(I->operand_values(), [](Value *Op) { |
4795 | return isa<Instruction>(Op) || isVectorLikeInstWithConstOps(Op); |
4796 | })); |
4797 | } |
4798 | bool IsCommutative = isCommutative(S.MainOp) || isCommutative(S.AltOp); |
4799 | if ((IsCommutative && |
4800 | std::accumulate(InstsCount.begin(), InstsCount.end(), 0) < 2) || |
4801 | (!IsCommutative && |
4802 | all_of(InstsCount, [](unsigned ICnt) { return ICnt < 2; }))) |
4803 | return true; |
4804 | 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", 4804, __extension__ __PRETTY_FUNCTION__)); |
4805 | SmallVector<SmallVector<std::pair<Value *, Value *>>> Candidates; |
4806 | auto *I1 = cast<Instruction>(VL.front()); |
4807 | auto *I2 = cast<Instruction>(VL.back()); |
4808 | for (int Op = 0, E = S.MainOp->getNumOperands(); Op < E; ++Op) |
4809 | Candidates.emplace_back().emplace_back(I1->getOperand(Op), |
4810 | I2->getOperand(Op)); |
4811 | if (static_cast<unsigned>(count_if( |
4812 | Candidates, [this](ArrayRef<std::pair<Value *, Value *>> Cand) { |
4813 | return findBestRootPair(Cand, LookAheadHeuristics::ScoreSplat); |
4814 | })) >= S.MainOp->getNumOperands() / 2) |
4815 | return false; |
4816 | if (S.MainOp->getNumOperands() > 2) |
4817 | return true; |
4818 | if (IsCommutative) { |
4819 | // Check permuted operands. |
4820 | Candidates.clear(); |
4821 | for (int Op = 0, E = S.MainOp->getNumOperands(); Op < E; ++Op) |
4822 | Candidates.emplace_back().emplace_back(I1->getOperand(Op), |
4823 | I2->getOperand((Op + 1) % E)); |
4824 | if (any_of( |
4825 | Candidates, [this](ArrayRef<std::pair<Value *, Value *>> Cand) { |
4826 | return findBestRootPair(Cand, LookAheadHeuristics::ScoreSplat); |
4827 | })) |
4828 | return false; |
4829 | } |
4830 | return true; |
4831 | }; |
4832 | SmallVector<unsigned> SortedIndices; |
4833 | BasicBlock *BB = nullptr; |
4834 | bool IsScatterVectorizeUserTE = |
4835 | UserTreeIdx.UserTE && |
4836 | UserTreeIdx.UserTE->State == TreeEntry::ScatterVectorize; |
4837 | bool AreAllSameInsts = |
4838 | (S.getOpcode() && allSameBlock(VL)) || |
4839 | (S.OpValue->getType()->isPointerTy() && IsScatterVectorizeUserTE && |
4840 | VL.size() > 2 && |
4841 | all_of(VL, |
4842 | [&BB](Value *V) { |
4843 | auto *I = dyn_cast<GetElementPtrInst>(V); |
4844 | if (!I) |
4845 | return doesNotNeedToBeScheduled(V); |
4846 | if (!BB) |
4847 | BB = I->getParent(); |
4848 | return BB == I->getParent() && I->getNumOperands() == 2; |
4849 | }) && |
4850 | BB && |
4851 | sortPtrAccesses(VL, UserTreeIdx.UserTE->getMainOp()->getType(), *DL, *SE, |
4852 | SortedIndices)); |
4853 | if (allConstant(VL) || isSplat(VL) || !AreAllSameInsts || |
4854 | (isa<InsertElementInst, ExtractValueInst, ExtractElementInst>( |
4855 | S.OpValue) && |
4856 | !all_of(VL, isVectorLikeInstWithConstOps)) || |
4857 | NotProfitableForVectorization(VL)) { |
4858 | 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); |
4859 | if (TryToFindDuplicates(S)) |
4860 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, |
4861 | ReuseShuffleIndicies); |
4862 | return; |
4863 | } |
4864 | |
4865 | // We now know that this is a vector of instructions of the same type from |
4866 | // the same block. |
4867 | |
4868 | // Don't vectorize ephemeral values. |
4869 | if (!EphValues.empty()) { |
4870 | for (Value *V : VL) { |
4871 | if (EphValues.count(V)) { |
4872 | LLVM_DEBUG(dbgs() << "SLP: The instruction (" << *Vdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: The instruction (" << * V << ") is ephemeral.\n"; } } while (false) |
4873 | << ") is ephemeral.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: The instruction (" << * V << ") is ephemeral.\n"; } } while (false); |
4874 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx); |
4875 | return; |
4876 | } |
4877 | } |
4878 | } |
4879 | |
4880 | // Check if this is a duplicate of another entry. |
4881 | if (TreeEntry *E = getTreeEntry(S.OpValue)) { |
4882 | 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); |
4883 | if (!E->isSame(VL)) { |
4884 | 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); |
4885 | if (TryToFindDuplicates(S)) |
4886 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, |
4887 | ReuseShuffleIndicies); |
4888 | return; |
4889 | } |
4890 | // Record the reuse of the tree node. FIXME, currently this is only used to |
4891 | // properly draw the graph rather than for the actual vectorization. |
4892 | E->UserTreeIndices.push_back(UserTreeIdx); |
4893 | 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) |
4894 | << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Perfect diamond merge at " << *S.OpValue << ".\n"; } } while (false); |
4895 | return; |
4896 | } |
4897 | |
4898 | // Check that none of the instructions in the bundle are already in the tree. |
4899 | for (Value *V : VL) { |
4900 | if (!IsScatterVectorizeUserTE && !isa<Instruction>(V)) |
4901 | continue; |
4902 | if (getTreeEntry(V)) { |
4903 | 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) |
4904 | << ") is already in tree.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: The instruction (" << * V << ") is already in tree.\n"; } } while (false); |
4905 | if (TryToFindDuplicates(S)) |
4906 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, |
4907 | ReuseShuffleIndicies); |
4908 | return; |
4909 | } |
4910 | } |
4911 | |
4912 | // The reduction nodes (stored in UserIgnoreList) also should stay scalar. |
4913 | if (UserIgnoreList && !UserIgnoreList->empty()) { |
4914 | for (Value *V : VL) { |
4915 | if (UserIgnoreList && UserIgnoreList->contains(V)) { |
4916 | 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); |
4917 | if (TryToFindDuplicates(S)) |
4918 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, |
4919 | ReuseShuffleIndicies); |
4920 | return; |
4921 | } |
4922 | } |
4923 | } |
4924 | |
4925 | // Special processing for sorted pointers for ScatterVectorize node with |
4926 | // constant indeces only. |
4927 | if (AreAllSameInsts && !(S.getOpcode() && allSameBlock(VL)) && |
4928 | UserTreeIdx.UserTE && |
4929 | UserTreeIdx.UserTE->State == TreeEntry::ScatterVectorize) { |
4930 | 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", 4933, __extension__ __PRETTY_FUNCTION__)) |
4931 | 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", 4933, __extension__ __PRETTY_FUNCTION__)) |
4932 | 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", 4933, __extension__ __PRETTY_FUNCTION__)) |
4933 | "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", 4933, __extension__ __PRETTY_FUNCTION__)); |
4934 | // Reset S to make it GetElementPtr kind of node. |
4935 | const auto *It = find_if(VL, [](Value *V) { return isa<GetElementPtrInst>(V); }); |
4936 | 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", 4936, __extension__ __PRETTY_FUNCTION__)); |
4937 | S = getSameOpcode(*It); |
4938 | } |
4939 | |
4940 | // Check that all of the users of the scalars that we want to vectorize are |
4941 | // schedulable. |
4942 | auto *VL0 = cast<Instruction>(S.OpValue); |
4943 | BB = VL0->getParent(); |
4944 | |
4945 | if (!DT->isReachableFromEntry(BB)) { |
4946 | // Don't go into unreachable blocks. They may contain instructions with |
4947 | // dependency cycles which confuse the final scheduling. |
4948 | 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); |
4949 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx); |
4950 | return; |
4951 | } |
4952 | |
4953 | // Don't go into catchswitch blocks, which can happen with PHIs. |
4954 | // Such blocks can only have PHIs and the catchswitch. There is no |
4955 | // place to insert a shuffle if we need to, so just avoid that issue. |
4956 | if (isa<CatchSwitchInst>(BB->getTerminator())) { |
4957 | 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); |
4958 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx); |
4959 | return; |
4960 | } |
4961 | |
4962 | // Check that every instruction appears once in this bundle. |
4963 | if (!TryToFindDuplicates(S)) |
4964 | return; |
4965 | |
4966 | auto &BSRef = BlocksSchedules[BB]; |
4967 | if (!BSRef) |
4968 | BSRef = std::make_unique<BlockScheduling>(BB); |
4969 | |
4970 | BlockScheduling &BS = *BSRef; |
4971 | |
4972 | Optional<ScheduleData *> Bundle = BS.tryScheduleBundle(VL, this, S); |
4973 | #ifdef EXPENSIVE_CHECKS |
4974 | // Make sure we didn't break any internal invariants |
4975 | BS.verify(); |
4976 | #endif |
4977 | if (!Bundle) { |
4978 | 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); |
4979 | 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", 4981, __extension__ __PRETTY_FUNCTION__)) |
4980 | !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", 4981, __extension__ __PRETTY_FUNCTION__)) |
4981 | "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", 4981, __extension__ __PRETTY_FUNCTION__)); |
4982 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, |
4983 | ReuseShuffleIndicies); |
4984 | return; |
4985 | } |
4986 | 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); |
4987 | |
4988 | unsigned ShuffleOrOp = S.isAltShuffle() ? |
4989 | (unsigned) Instruction::ShuffleVector : S.getOpcode(); |
4990 | switch (ShuffleOrOp) { |
4991 | case Instruction::PHI: { |
4992 | auto *PH = cast<PHINode>(VL0); |
4993 | |
4994 | // Check for terminator values (e.g. invoke). |
4995 | for (Value *V : VL) |
4996 | for (Value *Incoming : cast<PHINode>(V)->incoming_values()) { |
4997 | Instruction *Term = dyn_cast<Instruction>(Incoming); |
4998 | if (Term && Term->isTerminator()) { |
4999 | LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Need to swizzle PHINodes (terminator use).\n" ; } } while (false) |
5000 | << "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); |
5001 | BS.cancelScheduling(VL, VL0); |
5002 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, |
5003 | ReuseShuffleIndicies); |
5004 | return; |
5005 | } |
5006 | } |
5007 | |
5008 | TreeEntry *TE = |
5009 | newTreeEntry(VL, Bundle, S, UserTreeIdx, ReuseShuffleIndicies); |
5010 | 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); |
5011 | |
5012 | // Keeps the reordered operands to avoid code duplication. |
5013 | SmallVector<ValueList, 2> OperandsVec; |
5014 | for (unsigned I = 0, E = PH->getNumIncomingValues(); I < E; ++I) { |
5015 | if (!DT->isReachableFromEntry(PH->getIncomingBlock(I))) { |
5016 | ValueList Operands(VL.size(), PoisonValue::get(PH->getType())); |
5017 | TE->setOperand(I, Operands); |
5018 | OperandsVec.push_back(Operands); |
5019 | continue; |
5020 | } |
5021 | ValueList Operands; |
5022 | // Prepare the operand vector. |
5023 | for (Value *V : VL) |
5024 | Operands.push_back(cast<PHINode>(V)->getIncomingValueForBlock( |
5025 | PH->getIncomingBlock(I))); |
5026 | TE->setOperand(I, Operands); |
5027 | OperandsVec.push_back(Operands); |
5028 | } |
5029 | for (unsigned OpIdx = 0, OpE = OperandsVec.size(); OpIdx != OpE; ++OpIdx) |
5030 | buildTree_rec(OperandsVec[OpIdx], Depth + 1, {TE, OpIdx}); |
5031 | return; |
5032 | } |
5033 | case Instruction::ExtractValue: |
5034 | case Instruction::ExtractElement: { |
5035 | OrdersType CurrentOrder; |
5036 | bool Reuse = canReuseExtract(VL, VL0, CurrentOrder); |
5037 | if (Reuse) { |
5038 | 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); |
5039 | newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, |
5040 | ReuseShuffleIndicies); |
5041 | // This is a special case, as it does not gather, but at the same time |
5042 | // we are not extending buildTree_rec() towards the operands. |
5043 | ValueList Op0; |
5044 | Op0.assign(VL.size(), VL0->getOperand(0)); |
5045 | VectorizableTree.back()->setOperand(0, Op0); |
5046 | return; |
5047 | } |
5048 | if (!CurrentOrder.empty()) { |
5049 | 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) |
5050 | 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) |
5051 | "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) |
5052 | 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) |
5053 | 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) |
5054 | 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) |
5055 | })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); |
5056 | fixupOrderingIndices(CurrentOrder); |
5057 | // Insert new order with initial value 0, if it does not exist, |
5058 | // otherwise return the iterator to the existing one. |
5059 | newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, |
5060 | ReuseShuffleIndicies, CurrentOrder); |
5061 | // This is a special case, as it does not gather, but at the same time |
5062 | // we are not extending buildTree_rec() towards the operands. |
5063 | ValueList Op0; |
5064 | Op0.assign(VL.size(), VL0->getOperand(0)); |
5065 | VectorizableTree.back()->setOperand(0, Op0); |
5066 | return; |
5067 | } |
5068 | LLVM_DEBUG(dbgs() << "SLP: Gather extract sequence.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gather extract sequence.\n"; } } while (false); |
5069 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, |
5070 | ReuseShuffleIndicies); |
5071 | BS.cancelScheduling(VL, VL0); |
5072 | return; |
5073 | } |
5074 | case Instruction::InsertElement: { |
5075 | 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", 5075, __extension__ __PRETTY_FUNCTION__)); |
5076 | |
5077 | // Check that we have a buildvector and not a shuffle of 2 or more |
5078 | // different vectors. |
5079 | ValueSet SourceVectors; |
5080 | for (Value *V : VL) { |
5081 | SourceVectors.insert(cast<Instruction>(V)->getOperand(0)); |
5082 | assert(getInsertIndex(V) != None && "Non-constant or undef index?")(static_cast <bool> (getInsertIndex(V) != None && "Non-constant or undef index?") ? void (0) : __assert_fail ( "getInsertIndex(V) != None && \"Non-constant or undef index?\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 5082, __extension__ __PRETTY_FUNCTION__)); |
5083 | } |
5084 | |
5085 | if (count_if(VL, [&SourceVectors](Value *V) { |
5086 | return !SourceVectors.contains(V); |
5087 | }) >= 2) { |
5088 | // Found 2nd source vector - cancel. |
5089 | 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) |
5090 | "different source vectors.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gather of insertelement vectors with " "different source vectors.\n"; } } while (false); |
5091 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx); |
5092 | BS.cancelScheduling(VL, VL0); |
5093 | return; |
5094 | } |
5095 | |
5096 | auto OrdCompare = [](const std::pair<int, int> &P1, |
5097 | const std::pair<int, int> &P2) { |
5098 | return P1.first > P2.first; |
5099 | }; |
5100 | PriorityQueue<std::pair<int, int>, SmallVector<std::pair<int, int>>, |
5101 | decltype(OrdCompare)> |
5102 | Indices(OrdCompare); |
5103 | for (int I = 0, E = VL.size(); I < E; ++I) { |
5104 | unsigned Idx = *getInsertIndex(VL[I]); |
5105 | Indices.emplace(Idx, I); |
5106 | } |
5107 | OrdersType CurrentOrder(VL.size(), VL.size()); |
5108 | bool IsIdentity = true; |
5109 | for (int I = 0, E = VL.size(); I < E; ++I) { |
5110 | CurrentOrder[Indices.top().second] = I; |
5111 | IsIdentity &= Indices.top().second == I; |
5112 | Indices.pop(); |
5113 | } |
5114 | if (IsIdentity) |
5115 | CurrentOrder.clear(); |
5116 | TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, |
5117 | None, CurrentOrder); |
5118 | LLVM_DEBUG(dbgs() << "SLP: added inserts bundle.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: added inserts bundle.\n"; } } while (false); |
5119 | |
5120 | constexpr int NumOps = 2; |
5121 | ValueList VectorOperands[NumOps]; |
5122 | for (int I = 0; I < NumOps; ++I) { |
5123 | for (Value *V : VL) |
5124 | VectorOperands[I].push_back(cast<Instruction>(V)->getOperand(I)); |
5125 | |
5126 | TE->setOperand(I, VectorOperands[I]); |
5127 | } |
5128 | buildTree_rec(VectorOperands[NumOps - 1], Depth + 1, {TE, NumOps - 1}); |
5129 | return; |
5130 | } |
5131 | case Instruction::Load: { |
5132 | // Check that a vectorized load would load the same memory as a scalar |
5133 | // load. For example, we don't want to vectorize loads that are smaller |
5134 | // than 8-bit. Even though we have a packed struct {<i2, i2, i2, i2>} LLVM |
5135 | // treats loading/storing it as an i8 struct. If we vectorize loads/stores |
5136 | // from such a struct, we read/write packed bits disagreeing with the |
5137 | // unvectorized version. |
5138 | SmallVector<Value *> PointerOps; |
5139 | OrdersType CurrentOrder; |
5140 | TreeEntry *TE = nullptr; |
5141 | switch (canVectorizeLoads(VL, VL0, *TTI, *DL, *SE, *LI, CurrentOrder, |
5142 | PointerOps)) { |
5143 | case LoadsState::Vectorize: |
5144 | if (CurrentOrder.empty()) { |
5145 | // Original loads are consecutive and does not require reordering. |
5146 | TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, |
5147 | ReuseShuffleIndicies); |
5148 | 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); |
5149 | } else { |
5150 | fixupOrderingIndices(CurrentOrder); |
5151 | // Need to reorder. |
5152 | TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, |
5153 | ReuseShuffleIndicies, CurrentOrder); |
5154 | 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); |
5155 | } |
5156 | TE->setOperandsInOrder(); |
5157 | break; |
5158 | case LoadsState::ScatterVectorize: |
5159 | // Vectorizing non-consecutive loads with `llvm.masked.gather`. |
5160 | TE = newTreeEntry(VL, TreeEntry::ScatterVectorize, Bundle, S, |
5161 | UserTreeIdx, ReuseShuffleIndicies); |
5162 | TE->setOperandsInOrder(); |
5163 | buildTree_rec(PointerOps, Depth + 1, {TE, 0}); |
5164 | 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); |
5165 | break; |
5166 | case LoadsState::Gather: |
5167 | BS.cancelScheduling(VL, VL0); |
5168 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, |
5169 | ReuseShuffleIndicies); |
5170 | #ifndef NDEBUG |
5171 | Type *ScalarTy = VL0->getType(); |
5172 | if (DL->getTypeSizeInBits(ScalarTy) != |
5173 | DL->getTypeAllocSizeInBits(ScalarTy)) |
5174 | 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); |
5175 | else if (any_of(VL, [](Value *V) { |
5176 | return !cast<LoadInst>(V)->isSimple(); |
5177 | })) |
5178 | 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); |
5179 | else |
5180 | 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); |
5181 | #endif // NDEBUG |
5182 | break; |
5183 | } |
5184 | return; |
5185 | } |
5186 | case Instruction::ZExt: |
5187 | case Instruction::SExt: |
5188 | case Instruction::FPToUI: |
5189 | case Instruction::FPToSI: |
5190 | case Instruction::FPExt: |
5191 | case Instruction::PtrToInt: |
5192 | case Instruction::IntToPtr: |
5193 | case Instruction::SIToFP: |
5194 | case Instruction::UIToFP: |
5195 | case Instruction::Trunc: |
5196 | case Instruction::FPTrunc: |
5197 | case Instruction::BitCast: { |
5198 | Type *SrcTy = VL0->getOperand(0)->getType(); |
5199 | for (Value *V : VL) { |
5200 | Type *Ty = cast<Instruction>(V)->getOperand(0)->getType(); |
5201 | if (Ty != SrcTy || !isValidElementType(Ty)) { |
5202 | BS.cancelScheduling(VL, VL0); |
5203 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, |
5204 | ReuseShuffleIndicies); |
5205 | LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering casts with different src types.\n" ; } } while (false) |
5206 | << "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); |
5207 | return; |
5208 | } |
5209 | } |
5210 | TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, |
5211 | ReuseShuffleIndicies); |
5212 | 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); |
5213 | |
5214 | TE->setOperandsInOrder(); |
5215 | for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) { |
5216 | ValueList Operands; |
5217 | // Prepare the operand vector. |
5218 | for (Value *V : VL) |
5219 | Operands.push_back(cast<Instruction>(V)->getOperand(i)); |
5220 | |
5221 | buildTree_rec(Operands, Depth + 1, {TE, i}); |
5222 | } |
5223 | return; |
5224 | } |
5225 | case Instruction::ICmp: |
5226 | case Instruction::FCmp: { |
5227 | // Check that all of the compares have the same predicate. |
5228 | CmpInst::Predicate P0 = cast<CmpInst>(VL0)->getPredicate(); |
5229 | CmpInst::Predicate SwapP0 = CmpInst::getSwappedPredicate(P0); |
5230 | Type *ComparedTy = VL0->getOperand(0)->getType(); |
5231 | for (Value *V : VL) { |
5232 | CmpInst *Cmp = cast<CmpInst>(V); |
5233 | if ((Cmp->getPredicate() != P0 && Cmp->getPredicate() != SwapP0) || |
5234 | Cmp->getOperand(0)->getType() != ComparedTy) { |
5235 | BS.cancelScheduling(VL, VL0); |
5236 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, |
5237 | ReuseShuffleIndicies); |
5238 | LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering cmp with different predicate.\n" ; } } while (false) |
5239 | << "SLP: Gathering cmp with different predicate.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering cmp with different predicate.\n" ; } } while (false); |
5240 | return; |
5241 | } |
5242 | } |
5243 | |
5244 | TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, |
5245 | ReuseShuffleIndicies); |
5246 | 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); |
5247 | |
5248 | ValueList Left, Right; |
5249 | if (cast<CmpInst>(VL0)->isCommutative()) { |
5250 | // Commutative predicate - collect + sort operands of the instructions |
5251 | // so that each side is more likely to have the same opcode. |
5252 | 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", 5252, __extension__ __PRETTY_FUNCTION__)); |
5253 | reorderInputsAccordingToOpcode(VL, Left, Right, *DL, *SE, *this); |
5254 | } else { |
5255 | // Collect operands - commute if it uses the swapped predicate. |
5256 | for (Value *V : VL) { |
5257 | auto *Cmp = cast<CmpInst>(V); |
5258 | Value *LHS = Cmp->getOperand(0); |
5259 | Value *RHS = Cmp->getOperand(1); |
5260 | if (Cmp->getPredicate() != P0) |
5261 | std::swap(LHS, RHS); |
5262 | Left.push_back(LHS); |
5263 | Right.push_back(RHS); |
5264 | } |
5265 | } |
5266 | TE->setOperand(0, Left); |
5267 | TE->setOperand(1, Right); |
5268 | buildTree_rec(Left, Depth + 1, {TE, 0}); |
5269 | buildTree_rec(Right, Depth + 1, {TE, 1}); |
5270 | return; |
5271 | } |
5272 | case Instruction::Select: |
5273 | case Instruction::FNeg: |
5274 | case Instruction::Add: |
5275 | case Instruction::FAdd: |
5276 | case Instruction::Sub: |
5277 | case Instruction::FSub: |
5278 | case Instruction::Mul: |
5279 | case Instruction::FMul: |
5280 | case Instruction::UDiv: |
5281 | case Instruction::SDiv: |
5282 | case Instruction::FDiv: |
5283 | case Instruction::URem: |
5284 | case Instruction::SRem: |
5285 | case Instruction::FRem: |
5286 | case Instruction::Shl: |
5287 | case Instruction::LShr: |
5288 | case Instruction::AShr: |
5289 | case Instruction::And: |
5290 | case Instruction::Or: |
5291 | case Instruction::Xor: { |
5292 | TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, |
5293 | ReuseShuffleIndicies); |
5294 | 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); |
5295 | |
5296 | // Sort operands of the instructions so that each side is more likely to |
5297 | // have the same opcode. |
5298 | if (isa<BinaryOperator>(VL0) && VL0->isCommutative()) { |
5299 | ValueList Left, Right; |
5300 | reorderInputsAccordingToOpcode(VL, Left, Right, *DL, *SE, *this); |
5301 | TE->setOperand(0, Left); |
5302 | TE->setOperand(1, Right); |
5303 | buildTree_rec(Left, Depth + 1, {TE, 0}); |
5304 | buildTree_rec(Right, Depth + 1, {TE, 1}); |
5305 | return; |
5306 | } |
5307 | |
5308 | TE->setOperandsInOrder(); |
5309 | for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) { |
5310 | ValueList Operands; |
5311 | // Prepare the operand vector. |
5312 | for (Value *V : VL) |
5313 | Operands.push_back(cast<Instruction>(V)->getOperand(i)); |
5314 | |
5315 | buildTree_rec(Operands, Depth + 1, {TE, i}); |
5316 | } |
5317 | return; |
5318 | } |
5319 | case Instruction::GetElementPtr: { |
5320 | // We don't combine GEPs with complicated (nested) indexing. |
5321 | for (Value *V : VL) { |
5322 | auto *I = dyn_cast<GetElementPtrInst>(V); |
5323 | if (!I) |
5324 | continue; |
5325 | if (I->getNumOperands() != 2) { |
5326 | 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); |
5327 | BS.cancelScheduling(VL, VL0); |
5328 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, |
5329 | ReuseShuffleIndicies); |
5330 | return; |
5331 | } |
5332 | } |
5333 | |
5334 | // We can't combine several GEPs into one vector if they operate on |
5335 | // different types. |
5336 | Type *Ty0 = cast<GEPOperator>(VL0)->getSourceElementType(); |
5337 | for (Value *V : VL) { |
5338 | auto *GEP = dyn_cast<GEPOperator>(V); |
5339 | if (!GEP) |
5340 | continue; |
5341 | Type *CurTy = GEP->getSourceElementType(); |
5342 | if (Ty0 != CurTy) { |
5343 | LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: not-vectorizable GEP (different types).\n" ; } } while (false) |
5344 | << "SLP: not-vectorizable GEP (different types).\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: not-vectorizable GEP (different types).\n" ; } } while (false); |
5345 | BS.cancelScheduling(VL, VL0); |
5346 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, |
5347 | ReuseShuffleIndicies); |
5348 | return; |
5349 | } |
5350 | } |
5351 | |
5352 | // We don't combine GEPs with non-constant indexes. |
5353 | Type *Ty1 = VL0->getOperand(1)->getType(); |
5354 | for (Value *V : VL) { |
5355 | auto *I = dyn_cast<GetElementPtrInst>(V); |
5356 | if (!I) |
5357 | continue; |
5358 | auto *Op = I->getOperand(1); |
5359 | if ((!IsScatterVectorizeUserTE && !isa<ConstantInt>(Op)) || |
5360 | (Op->getType() != Ty1 && |
5361 | ((IsScatterVectorizeUserTE && !isa<ConstantInt>(Op)) || |
5362 | Op->getType()->getScalarSizeInBits() > |
5363 | DL->getIndexSizeInBits( |
5364 | V->getType()->getPointerAddressSpace())))) { |
5365 | LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: not-vectorizable GEP (non-constant indexes).\n" ; } } while (false) |
5366 | << "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); |
5367 | BS.cancelScheduling(VL, VL0); |
5368 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, |
5369 | ReuseShuffleIndicies); |
5370 | return; |
5371 | } |
5372 | } |
5373 | |
5374 | TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, |
5375 | ReuseShuffleIndicies); |
5376 | 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); |
5377 | SmallVector<ValueList, 2> Operands(2); |
5378 | // Prepare the operand vector for pointer operands. |
5379 | for (Value *V : VL) { |
5380 | auto *GEP = dyn_cast<GetElementPtrInst>(V); |
5381 | if (!GEP) { |
5382 | Operands.front().push_back(V); |
5383 | continue; |
5384 | } |
5385 | Operands.front().push_back(GEP->getPointerOperand()); |
5386 | } |
5387 | TE->setOperand(0, Operands.front()); |
5388 | // Need to cast all indices to the same type before vectorization to |
5389 | // avoid crash. |
5390 | // Required to be able to find correct matches between different gather |
5391 | // nodes and reuse the vectorized values rather than trying to gather them |
5392 | // again. |
5393 | int IndexIdx = 1; |
5394 | Type *VL0Ty = VL0->getOperand(IndexIdx)->getType(); |
5395 | Type *Ty = all_of(VL, |
5396 | [VL0Ty, IndexIdx](Value *V) { |
5397 | auto *GEP = dyn_cast<GetElementPtrInst>(V); |
5398 | if (!GEP) |
5399 | return true; |
5400 | return VL0Ty == GEP->getOperand(IndexIdx)->getType(); |
5401 | }) |
5402 | ? VL0Ty |
5403 | : DL->getIndexType(cast<GetElementPtrInst>(VL0) |
5404 | ->getPointerOperandType() |
5405 | ->getScalarType()); |
5406 | // Prepare the operand vector. |
5407 | for (Value *V : VL) { |
5408 | auto *I = dyn_cast<GetElementPtrInst>(V); |
5409 | if (!I) { |
5410 | Operands.back().push_back( |
5411 | ConstantInt::get(Ty, 0, /*isSigned=*/false)); |
5412 | continue; |
5413 | } |
5414 | auto *Op = I->getOperand(IndexIdx); |
5415 | auto *CI = dyn_cast<ConstantInt>(Op); |
5416 | if (!CI) |
5417 | Operands.back().push_back(Op); |
5418 | else |
5419 | Operands.back().push_back(ConstantExpr::getIntegerCast( |
5420 | CI, Ty, CI->getValue().isSignBitSet())); |
5421 | } |
5422 | TE->setOperand(IndexIdx, Operands.back()); |
5423 | |
5424 | for (unsigned I = 0, Ops = Operands.size(); I < Ops; ++I) |
5425 | buildTree_rec(Operands[I], Depth + 1, {TE, I}); |
5426 | return; |
5427 | } |
5428 | case Instruction::Store: { |
5429 | // Check if the stores are consecutive or if we need to swizzle them. |
5430 | llvm::Type *ScalarTy = cast<StoreInst>(VL0)->getValueOperand()->getType(); |
5431 | // Avoid types that are padded when being allocated as scalars, while |
5432 | // being packed together in a vector (such as i1). |
5433 | if (DL->getTypeSizeInBits(ScalarTy) != |
5434 | DL->getTypeAllocSizeInBits(ScalarTy)) { |
5435 | BS.cancelScheduling(VL, VL0); |
5436 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, |
5437 | ReuseShuffleIndicies); |
5438 | 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); |
5439 | return; |
5440 | } |
5441 | // Make sure all stores in the bundle are simple - we can't vectorize |
5442 | // atomic or volatile stores. |
5443 | SmallVector<Value *, 4> PointerOps(VL.size()); |
5444 | ValueList Operands(VL.size()); |
5445 | auto POIter = PointerOps.begin(); |
5446 | auto OIter = Operands.begin(); |
5447 | for (Value *V : VL) { |
5448 | auto *SI = cast<StoreInst>(V); |
5449 | if (!SI->isSimple()) { |
5450 | BS.cancelScheduling(VL, VL0); |
5451 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, |
5452 | ReuseShuffleIndicies); |
5453 | 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); |
5454 | return; |
5455 | } |
5456 | *POIter = SI->getPointerOperand(); |
5457 | *OIter = SI->getValueOperand(); |
5458 | ++POIter; |
5459 | ++OIter; |
5460 | } |
5461 | |
5462 | OrdersType CurrentOrder; |
5463 | // Check the order of pointer operands. |
5464 | if (llvm::sortPtrAccesses(PointerOps, ScalarTy, *DL, *SE, CurrentOrder)) { |
5465 | Value *Ptr0; |
5466 | Value *PtrN; |
5467 | if (CurrentOrder.empty()) { |
5468 | Ptr0 = PointerOps.front(); |
5469 | PtrN = PointerOps.back(); |
5470 | } else { |
5471 | Ptr0 = PointerOps[CurrentOrder.front()]; |
5472 | PtrN = PointerOps[CurrentOrder.back()]; |
5473 | } |
5474 | Optional<int> Dist = |
5475 | getPointersDiff(ScalarTy, Ptr0, ScalarTy, PtrN, *DL, *SE); |
5476 | // Check that the sorted pointer operands are consecutive. |
5477 | if (static_cast<unsigned>(*Dist) == VL.size() - 1) { |
5478 | if (CurrentOrder.empty()) { |
5479 | // Original stores are consecutive and does not require reordering. |
5480 | TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, |
5481 | UserTreeIdx, ReuseShuffleIndicies); |
5482 | TE->setOperandsInOrder(); |
5483 | buildTree_rec(Operands, Depth + 1, {TE, 0}); |
5484 | 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); |
5485 | } else { |
5486 | fixupOrderingIndices(CurrentOrder); |
5487 | TreeEntry *TE = |
5488 | newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, |
5489 | ReuseShuffleIndicies, CurrentOrder); |
5490 | TE->setOperandsInOrder(); |
5491 | buildTree_rec(Operands, Depth + 1, {TE, 0}); |
5492 | 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); |
5493 | } |
5494 | return; |
5495 | } |
5496 | } |
5497 | |
5498 | BS.cancelScheduling(VL, VL0); |
5499 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, |
5500 | ReuseShuffleIndicies); |
5501 | LLVM_DEBUG(dbgs() << "SLP: Non-consecutive store.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Non-consecutive store.\n"; } } while (false); |
5502 | return; |
5503 | } |
5504 | case Instruction::Call: { |
5505 | // Check if the calls are all to the same vectorizable intrinsic or |
5506 | // library function. |
5507 | CallInst *CI = cast<CallInst>(VL0); |
5508 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); |
5509 | |
5510 | VFShape Shape = VFShape::get( |
5511 | *CI, ElementCount::getFixed(static_cast<unsigned int>(VL.size())), |
5512 | false /*HasGlobalPred*/); |
5513 | Function *VecFunc = VFDatabase(*CI).getVectorizedFunction(Shape); |
5514 | |
5515 | if (!VecFunc && !isTriviallyVectorizable(ID)) { |
5516 | BS.cancelScheduling(VL, VL0); |
5517 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, |
5518 | ReuseShuffleIndicies); |
5519 | LLVM_DEBUG(dbgs() << "SLP: Non-vectorizable call.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Non-vectorizable call.\n"; } } while (false); |
5520 | return; |
5521 | } |
5522 | Function *F = CI->getCalledFunction(); |
5523 | unsigned NumArgs = CI->arg_size(); |
5524 | SmallVector<Value*, 4> ScalarArgs(NumArgs, nullptr); |
5525 | for (unsigned j = 0; j != NumArgs; ++j) |
5526 | if (isVectorIntrinsicWithScalarOpAtArg(ID, j)) |
5527 | ScalarArgs[j] = CI->getArgOperand(j); |
5528 | for (Value *V : VL) { |
5529 | CallInst *CI2 = dyn_cast<CallInst>(V); |
5530 | if (!CI2 || CI2->getCalledFunction() != F || |
5531 | getVectorIntrinsicIDForCall(CI2, TLI) != ID || |
5532 | (VecFunc && |
5533 | VecFunc != VFDatabase(*CI2).getVectorizedFunction(Shape)) || |
5534 | !CI->hasIdenticalOperandBundleSchema(*CI2)) { |
5535 | BS.cancelScheduling(VL, VL0); |
5536 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, |
5537 | ReuseShuffleIndicies); |
5538 | LLVM_DEBUG(dbgs() << "SLP: mismatched calls:" << *CI << "!=" << *Vdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: mismatched calls:" << * CI << "!=" << *V << "\n"; } } while (false) |
5539 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: mismatched calls:" << * CI << "!=" << *V << "\n"; } } while (false); |
5540 | return; |
5541 | } |
5542 | // Some intrinsics have scalar arguments and should be same in order for |
5543 | // them to be vectorized. |
5544 | for (unsigned j = 0; j != NumArgs; ++j) { |
5545 | if (isVectorIntrinsicWithScalarOpAtArg(ID, j)) { |
5546 | Value *A1J = CI2->getArgOperand(j); |
5547 | if (ScalarArgs[j] != A1J) { |
5548 | BS.cancelScheduling(VL, VL0); |
5549 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, |
5550 | ReuseShuffleIndicies); |
5551 | 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) |
5552 | << " argument " << ScalarArgs[j] << "!=" << A1Jdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: mismatched arguments in call:" << *CI << " argument " << ScalarArgs[j] << "!=" << A1J << "\n"; } } while (false) |
5553 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: mismatched arguments in call:" << *CI << " argument " << ScalarArgs[j] << "!=" << A1J << "\n"; } } while (false); |
5554 | return; |
5555 | } |
5556 | } |
5557 | } |
5558 | // Verify that the bundle operands are identical between the two calls. |
5559 | if (CI->hasOperandBundles() && |
5560 | !std::equal(CI->op_begin() + CI->getBundleOperandsStartIndex(), |
5561 | CI->op_begin() + CI->getBundleOperandsEndIndex(), |
5562 | CI2->op_begin() + CI2->getBundleOperandsStartIndex())) { |
5563 | BS.cancelScheduling(VL, VL0); |
5564 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, |
5565 | ReuseShuffleIndicies); |
5566 | 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) |
5567 | << *CI << "!=" << *V << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: mismatched bundle operands in calls:" << *CI << "!=" << *V << '\n'; } } while (false); |
5568 | return; |
5569 | } |
5570 | } |
5571 | |
5572 | TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, |
5573 | ReuseShuffleIndicies); |
5574 | TE->setOperandsInOrder(); |
5575 | for (unsigned i = 0, e = CI->arg_size(); i != e; ++i) { |
5576 | // For scalar operands no need to to create an entry since no need to |
5577 | // vectorize it. |
5578 | if (isVectorIntrinsicWithScalarOpAtArg(ID, i)) |
5579 | continue; |
5580 | ValueList Operands; |
5581 | // Prepare the operand vector. |
5582 | for (Value *V : VL) { |
5583 | auto *CI2 = cast<CallInst>(V); |
5584 | Operands.push_back(CI2->getArgOperand(i)); |
5585 | } |
5586 | buildTree_rec(Operands, Depth + 1, {TE, i}); |
5587 | } |
5588 | return; |
5589 | } |
5590 | case Instruction::ShuffleVector: { |
5591 | // If this is not an alternate sequence of opcode like add-sub |
5592 | // then do not vectorize this instruction. |
5593 | if (!S.isAltShuffle()) { |
5594 | BS.cancelScheduling(VL, VL0); |
5595 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, |
5596 | ReuseShuffleIndicies); |
5597 | 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); |
5598 | return; |
5599 | } |
5600 | TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, |
5601 | ReuseShuffleIndicies); |
5602 | 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); |
5603 | |
5604 | // Reorder operands if reordering would enable vectorization. |
5605 | auto *CI = dyn_cast<CmpInst>(VL0); |
5606 | if (isa<BinaryOperator>(VL0) || CI) { |
5607 | ValueList Left, Right; |
5608 | if (!CI || all_of(VL, [](Value *V) { |
5609 | return cast<CmpInst>(V)->isCommutative(); |
5610 | })) { |
5611 | reorderInputsAccordingToOpcode(VL, Left, Right, *DL, *SE, *this); |
5612 | } else { |
5613 | auto *MainCI = cast<CmpInst>(S.MainOp); |
5614 | auto *AltCI = cast<CmpInst>(S.AltOp); |
5615 | CmpInst::Predicate MainP = MainCI->getPredicate(); |
5616 | CmpInst::Predicate AltP = AltCI->getPredicate(); |
5617 | 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", 5618, __extension__ __PRETTY_FUNCTION__)) |
5618 | "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", 5618, __extension__ __PRETTY_FUNCTION__)); |
5619 | // Collect operands - commute if it uses the swapped predicate or |
5620 | // alternate operation. |
5621 | for (Value *V : VL) { |
5622 | auto *Cmp = cast<CmpInst>(V); |
5623 | Value *LHS = Cmp->getOperand(0); |
5624 | Value *RHS = Cmp->getOperand(1); |
5625 | |
5626 | if (isAlternateInstruction(Cmp, MainCI, AltCI)) { |
5627 | if (AltP == CmpInst::getSwappedPredicate(Cmp->getPredicate())) |
5628 | std::swap(LHS, RHS); |
5629 | } else { |
5630 | if (MainP == CmpInst::getSwappedPredicate(Cmp->getPredicate())) |
5631 | std::swap(LHS, RHS); |
5632 | } |
5633 | Left.push_back(LHS); |
5634 | Right.push_back(RHS); |
5635 | } |
5636 | } |
5637 | TE->setOperand(0, Left); |
5638 | TE->setOperand(1, Right); |
5639 | buildTree_rec(Left, Depth + 1, {TE, 0}); |
5640 | buildTree_rec(Right, Depth + 1, {TE, 1}); |
5641 | return; |
5642 | } |
5643 | |
5644 | TE->setOperandsInOrder(); |
5645 | for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) { |
5646 | ValueList Operands; |
5647 | // Prepare the operand vector. |
5648 | for (Value *V : VL) |
5649 | Operands.push_back(cast<Instruction>(V)->getOperand(i)); |
5650 | |
5651 | buildTree_rec(Operands, Depth + 1, {TE, i}); |
5652 | } |
5653 | return; |
5654 | } |
5655 | default: |
5656 | BS.cancelScheduling(VL, VL0); |
5657 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, |
5658 | ReuseShuffleIndicies); |
5659 | LLVM_DEBUG(dbgs() << "SLP: Gathering unknown instruction.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering unknown instruction.\n" ; } } while (false); |
5660 | return; |
5661 | } |
5662 | } |
5663 | |
5664 | unsigned BoUpSLP::canMapToVector(Type *T, const DataLayout &DL) const { |
5665 | unsigned N = 1; |
5666 | Type *EltTy = T; |
5667 | |
5668 | while (isa<StructType, ArrayType, VectorType>(EltTy)) { |
5669 | if (auto *ST = dyn_cast<StructType>(EltTy)) { |
5670 | // Check that struct is homogeneous. |
5671 | for (const auto *Ty : ST->elements()) |
5672 | if (Ty != *ST->element_begin()) |
5673 | return 0; |
5674 | N *= ST->getNumElements(); |
5675 | EltTy = *ST->element_begin(); |
5676 | } else if (auto *AT = dyn_cast<ArrayType>(EltTy)) { |
5677 | N *= AT->getNumElements(); |
5678 | EltTy = AT->getElementType(); |
5679 | } else { |
5680 | auto *VT = cast<FixedVectorType>(EltTy); |
5681 | N *= VT->getNumElements(); |
5682 | EltTy = VT->getElementType(); |
5683 | } |
5684 | } |
5685 | |
5686 | if (!isValidElementType(EltTy)) |
5687 | return 0; |
5688 | uint64_t VTSize = DL.getTypeStoreSizeInBits(FixedVectorType::get(EltTy, N)); |
5689 | if (VTSize < MinVecRegSize || VTSize > MaxVecRegSize || VTSize != DL.getTypeStoreSizeInBits(T)) |
5690 | return 0; |
5691 | return N; |
5692 | } |
5693 | |
5694 | bool BoUpSLP::canReuseExtract(ArrayRef<Value *> VL, Value *OpValue, |
5695 | SmallVectorImpl<unsigned> &CurrentOrder) const { |
5696 | const auto *It = find_if(VL, [](Value *V) { |
5697 | return isa<ExtractElementInst, ExtractValueInst>(V); |
5698 | }); |
5699 | 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", 5699, __extension__ __PRETTY_FUNCTION__)); |
5700 | auto *E0 = cast<Instruction>(*It); |
5701 | 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", 5706, __extension__ __PRETTY_FUNCTION__)) |
5702 | [](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", 5706, __extension__ __PRETTY_FUNCTION__)) |
5703 | 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", 5706, __extension__ __PRETTY_FUNCTION__)) |
5704 | 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", 5706, __extension__ __PRETTY_FUNCTION__)) |
5705 | }) &&(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", 5706, __extension__ __PRETTY_FUNCTION__)) |
5706 | "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", 5706, __extension__ __PRETTY_FUNCTION__)); |
5707 | // Check if all of the extracts come from the same vector and from the |
5708 | // correct offset. |
5709 | Value *Vec = E0->getOperand(0); |
5710 | |
5711 | CurrentOrder.clear(); |
5712 | |
5713 | // We have to extract from a vector/aggregate with the same number of elements. |
5714 | unsigned NElts; |
5715 | if (E0->getOpcode() == Instruction::ExtractValue) { |
5716 | const DataLayout &DL = E0->getModule()->getDataLayout(); |
5717 | NElts = canMapToVector(Vec->getType(), DL); |
5718 | if (!NElts) |
5719 | return false; |
5720 | // Check if load can be rewritten as load of vector. |
5721 | LoadInst *LI = dyn_cast<LoadInst>(Vec); |
5722 | if (!LI || !LI->isSimple() || !LI->hasNUses(VL.size())) |
5723 | return false; |
5724 | } else { |
5725 | NElts = cast<FixedVectorType>(Vec->getType())->getNumElements(); |
5726 | } |
5727 | |
5728 | if (NElts != VL.size()) |
5729 | return false; |
5730 | |
5731 | // Check that all of the indices extract from the correct offset. |
5732 | bool ShouldKeepOrder = true; |
5733 | unsigned E = VL.size(); |
5734 | // Assign to all items the initial value E + 1 so we can check if the extract |
5735 | // instruction index was used already. |
5736 | // Also, later we can check that all the indices are used and we have a |
5737 | // consecutive access in the extract instructions, by checking that no |
5738 | // element of CurrentOrder still has value E + 1. |
5739 | CurrentOrder.assign(E, E); |
5740 | unsigned I = 0; |
5741 | for (; I < E; ++I) { |
5742 | auto *Inst = dyn_cast<Instruction>(VL[I]); |
5743 | if (!Inst) |
5744 | continue; |
5745 | if (Inst->getOperand(0) != Vec) |
5746 | break; |
5747 | if (auto *EE = dyn_cast<ExtractElementInst>(Inst)) |
5748 | if (isa<UndefValue>(EE->getIndexOperand())) |
5749 | continue; |
5750 | Optional<unsigned> Idx = getExtractIndex(Inst); |
5751 | if (!Idx) |
5752 | break; |
5753 | const unsigned ExtIdx = *Idx; |
5754 | if (ExtIdx != I) { |
5755 | if (ExtIdx >= E || CurrentOrder[ExtIdx] != E) |
5756 | break; |
5757 | ShouldKeepOrder = false; |
5758 | CurrentOrder[ExtIdx] = I; |
5759 | } else { |
5760 | if (CurrentOrder[I] != E) |
5761 | break; |
5762 | CurrentOrder[I] = I; |
5763 | } |
5764 | } |
5765 | if (I < E) { |
5766 | CurrentOrder.clear(); |
5767 | return false; |
5768 | } |
5769 | if (ShouldKeepOrder) |
5770 | CurrentOrder.clear(); |
5771 | |
5772 | return ShouldKeepOrder; |
5773 | } |
5774 | |
5775 | bool BoUpSLP::areAllUsersVectorized(Instruction *I, |
5776 | ArrayRef<Value *> VectorizedVals) const { |
5777 | return (I->hasOneUse() && is_contained(VectorizedVals, I)) || |
5778 | all_of(I->users(), [this](User *U) { |
5779 | return ScalarToTreeEntry.count(U) > 0 || |
5780 | isVectorLikeInstWithConstOps(U) || |
5781 | (isa<ExtractElementInst>(U) && MustGather.contains(U)); |
5782 | }); |
5783 | } |
5784 | |
5785 | static std::pair<InstructionCost, InstructionCost> |
5786 | getVectorCallCosts(CallInst *CI, FixedVectorType *VecTy, |
5787 | TargetTransformInfo *TTI, TargetLibraryInfo *TLI) { |
5788 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); |
5789 | |
5790 | // Calculate the cost of the scalar and vector calls. |
5791 | SmallVector<Type *, 4> VecTys; |
5792 | for (Use &Arg : CI->args()) |
5793 | VecTys.push_back( |
5794 | FixedVectorType::get(Arg->getType(), VecTy->getNumElements())); |
5795 | FastMathFlags FMF; |
5796 | if (auto *FPCI = dyn_cast<FPMathOperator>(CI)) |
5797 | FMF = FPCI->getFastMathFlags(); |
5798 | SmallVector<const Value *> Arguments(CI->args()); |
5799 | IntrinsicCostAttributes CostAttrs(ID, VecTy, Arguments, VecTys, FMF, |
5800 | dyn_cast<IntrinsicInst>(CI)); |
5801 | auto IntrinsicCost = |
5802 | TTI->getIntrinsicInstrCost(CostAttrs, TTI::TCK_RecipThroughput); |
5803 | |
5804 | auto Shape = VFShape::get(*CI, ElementCount::getFixed(static_cast<unsigned>( |
5805 | VecTy->getNumElements())), |
5806 | false /*HasGlobalPred*/); |
5807 | Function *VecFunc = VFDatabase(*CI).getVectorizedFunction(Shape); |
5808 | auto LibCost = IntrinsicCost; |
5809 | if (!CI->isNoBuiltin() && VecFunc) { |
5810 | // Calculate the cost of the vector library call. |
5811 | // If the corresponding vector call is cheaper, return its cost. |
5812 | LibCost = TTI->getCallInstrCost(nullptr, VecTy, VecTys, |
5813 | TTI::TCK_RecipThroughput); |
5814 | } |
5815 | return {IntrinsicCost, LibCost}; |
5816 | } |
5817 | |
5818 | /// Compute the cost of creating a vector of type \p VecTy containing the |
5819 | /// extracted values from \p VL. |
5820 | static InstructionCost |
5821 | computeExtractCost(ArrayRef<Value *> VL, FixedVectorType *VecTy, |
5822 | TargetTransformInfo::ShuffleKind ShuffleKind, |
5823 | ArrayRef<int> Mask, TargetTransformInfo &TTI) { |
5824 | unsigned NumOfParts = TTI.getNumberOfParts(VecTy); |
5825 | |
5826 | if (ShuffleKind != TargetTransformInfo::SK_PermuteSingleSrc || !NumOfParts || |
5827 | VecTy->getNumElements() < NumOfParts) |
5828 | return TTI.getShuffleCost(ShuffleKind, VecTy, Mask); |
5829 | |
5830 | bool AllConsecutive = true; |
5831 | unsigned EltsPerVector = VecTy->getNumElements() / NumOfParts; |
5832 | unsigned Idx = -1; |
5833 | InstructionCost Cost = 0; |
5834 | |
5835 | // Process extracts in blocks of EltsPerVector to check if the source vector |
5836 | // operand can be re-used directly. If not, add the cost of creating a shuffle |
5837 | // to extract the values into a vector register. |
5838 | SmallVector<int> RegMask(EltsPerVector, UndefMaskElem); |
5839 | for (auto *V : VL) { |
5840 | ++Idx; |
5841 | |
5842 | // Reached the start of a new vector registers. |
5843 | if (Idx % EltsPerVector == 0) { |
5844 | RegMask.assign(EltsPerVector, UndefMaskElem); |
5845 | AllConsecutive = true; |
5846 | continue; |
5847 | } |
5848 | |
5849 | // Need to exclude undefs from analysis. |
5850 | if (isa<UndefValue>(V) || Mask[Idx] == UndefMaskElem) |
5851 | continue; |
5852 | |
5853 | // Check all extracts for a vector register on the target directly |
5854 | // extract values in order. |
5855 | unsigned CurrentIdx = *getExtractIndex(cast<Instruction>(V)); |
5856 | if (!isa<UndefValue>(VL[Idx - 1]) && Mask[Idx - 1] != UndefMaskElem) { |
5857 | unsigned PrevIdx = *getExtractIndex(cast<Instruction>(VL[Idx - 1])); |
5858 | AllConsecutive &= PrevIdx + 1 == CurrentIdx && |
5859 | CurrentIdx % EltsPerVector == Idx % EltsPerVector; |
5860 | RegMask[Idx % EltsPerVector] = CurrentIdx % EltsPerVector; |
5861 | } |
5862 | |
5863 | if (AllConsecutive) |
5864 | continue; |
5865 | |
5866 | // Skip all indices, except for the last index per vector block. |
5867 | if ((Idx + 1) % EltsPerVector != 0 && Idx + 1 != VL.size()) |
5868 | continue; |
5869 | |
5870 | // If we have a series of extracts which are not consecutive and hence |
5871 | // cannot re-use the source vector register directly, compute the shuffle |
5872 | // cost to extract the vector with EltsPerVector elements. |
5873 | Cost += TTI.getShuffleCost( |
5874 | TargetTransformInfo::SK_PermuteSingleSrc, |
5875 | FixedVectorType::get(VecTy->getElementType(), EltsPerVector), RegMask); |
5876 | } |
5877 | return Cost; |
5878 | } |
5879 | |
5880 | /// Build shuffle mask for shuffle graph entries and lists of main and alternate |
5881 | /// operations operands. |
5882 | static void |
5883 | buildShuffleEntryMask(ArrayRef<Value *> VL, ArrayRef<unsigned> ReorderIndices, |
5884 | ArrayRef<int> ReusesIndices, |
5885 | const function_ref<bool(Instruction *)> IsAltOp, |
5886 | SmallVectorImpl<int> &Mask, |
5887 | SmallVectorImpl<Value *> *OpScalars = nullptr, |
5888 | SmallVectorImpl<Value *> *AltScalars = nullptr) { |
5889 | unsigned Sz = VL.size(); |
5890 | Mask.assign(Sz, UndefMaskElem); |
5891 | SmallVector<int> OrderMask; |
5892 | if (!ReorderIndices.empty()) |
5893 | inversePermutation(ReorderIndices, OrderMask); |
5894 | for (unsigned I = 0; I < Sz; ++I) { |
5895 | unsigned Idx = I; |
5896 | if (!ReorderIndices.empty()) |
5897 | Idx = OrderMask[I]; |
5898 | auto *OpInst = cast<Instruction>(VL[Idx]); |
5899 | if (IsAltOp(OpInst)) { |
5900 | Mask[I] = Sz + Idx; |
5901 | if (AltScalars) |
5902 | AltScalars->push_back(OpInst); |
5903 | } else { |
5904 | Mask[I] = Idx; |
5905 | if (OpScalars) |
5906 | OpScalars->push_back(OpInst); |
5907 | } |
5908 | } |
5909 | if (!ReusesIndices.empty()) { |
5910 | SmallVector<int> NewMask(ReusesIndices.size(), UndefMaskElem); |
5911 | transform(ReusesIndices, NewMask.begin(), [&Mask](int Idx) { |
5912 | return Idx != UndefMaskElem ? Mask[Idx] : UndefMaskElem; |
5913 | }); |
5914 | Mask.swap(NewMask); |
5915 | } |
5916 | } |
5917 | |
5918 | static bool isAlternateInstruction(const Instruction *I, |
5919 | const Instruction *MainOp, |
5920 | const Instruction *AltOp) { |
5921 | if (auto *MainCI = dyn_cast<CmpInst>(MainOp)) { |
5922 | auto *AltCI = cast<CmpInst>(AltOp); |
5923 | CmpInst::Predicate MainP = MainCI->getPredicate(); |
5924 | CmpInst::Predicate AltP = AltCI->getPredicate(); |
5925 | 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", 5925, __extension__ __PRETTY_FUNCTION__)); |
5926 | auto *CI = cast<CmpInst>(I); |
5927 | if (isCmpSameOrSwapped(MainCI, CI)) |
5928 | return false; |
5929 | if (isCmpSameOrSwapped(AltCI, CI)) |
5930 | return true; |
5931 | CmpInst::Predicate P = CI->getPredicate(); |
5932 | CmpInst::Predicate SwappedP = CmpInst::getSwappedPredicate(P); |
5933 | |
5934 | 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", 5936, __extension__ __PRETTY_FUNCTION__)) |
5935 | "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", 5936, __extension__ __PRETTY_FUNCTION__)) |
5936 | "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", 5936, __extension__ __PRETTY_FUNCTION__)); |
5937 | (void)AltP; |
5938 | return MainP != P && MainP != SwappedP; |
5939 | } |
5940 | return I->getOpcode() == AltOp->getOpcode(); |
5941 | } |
5942 | |
5943 | TTI::OperandValueInfo BoUpSLP::getOperandInfo(ArrayRef<Value *> VL, |
5944 | unsigned OpIdx) { |
5945 | assert(!VL.empty())(static_cast <bool> (!VL.empty()) ? void (0) : __assert_fail ("!VL.empty()", "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5945, __extension__ __PRETTY_FUNCTION__)); |
5946 | const auto *Op0 = cast<Instruction>(VL.front())->getOperand(OpIdx); |
5947 | |
5948 | const bool IsConstant = all_of(VL, [&](Value *V) { |
5949 | // TODO: We should allow undef elements here |
5950 | auto *Op = cast<Instruction>(V)->getOperand(OpIdx); |
5951 | return isConstant(Op) && !isa<UndefValue>(Op); |
5952 | }); |
5953 | const bool IsUniform = all_of(VL, [&](Value *V) { |
5954 | // TODO: We should allow undef elements here |
5955 | return cast<Instruction>(V)->getOperand(OpIdx) == Op0; |
5956 | }); |
5957 | const bool IsPowerOfTwo = all_of(VL, [&](Value *V) { |
5958 | // TODO: We should allow undef elements here |
5959 | auto *Op = cast<Instruction>(V)->getOperand(OpIdx); |
5960 | if (auto *CI = dyn_cast<ConstantInt>(Op)) |
5961 | return CI->getValue().isPowerOf2(); |
5962 | return false; |
5963 | }); |
5964 | const bool IsNegatedPowerOfTwo = all_of(VL, [&](Value *V) { |
5965 | // TODO: We should allow undef elements here |
5966 | auto *Op = cast<Instruction>(V)->getOperand(OpIdx); |
5967 | if (auto *CI = dyn_cast<ConstantInt>(Op)) |
5968 | return CI->getValue().isNegatedPowerOf2(); |
5969 | return false; |
5970 | }); |
5971 | |
5972 | TTI::OperandValueKind VK = TTI::OK_AnyValue; |
5973 | if (IsConstant && IsUniform) |
5974 | VK = TTI::OK_UniformConstantValue; |
5975 | else if (IsConstant) |
5976 | VK = TTI::OK_NonUniformConstantValue; |
5977 | else if (IsUniform) |
5978 | VK = TTI::OK_UniformValue; |
5979 | |
5980 | TTI::OperandValueProperties VP = TTI::OP_None; |
5981 | VP = IsPowerOfTwo ? TTI::OP_PowerOf2 : VP; |
5982 | VP = IsNegatedPowerOfTwo ? TTI::OP_NegatedPowerOf2 : VP; |
5983 | |
5984 | return {VK, VP}; |
5985 | } |
5986 | |
5987 | InstructionCost BoUpSLP::getEntryCost(const TreeEntry *E, |
5988 | ArrayRef<Value *> VectorizedVals) { |
5989 | ArrayRef<Value*> VL = E->Scalars; |
5990 | |
5991 | Type *ScalarTy = VL[0]->getType(); |
5992 | if (StoreInst *SI = dyn_cast<StoreInst>(VL[0])) |
5993 | ScalarTy = SI->getValueOperand()->getType(); |
5994 | else if (CmpInst *CI = dyn_cast<CmpInst>(VL[0])) |
5995 | ScalarTy = CI->getOperand(0)->getType(); |
5996 | else if (auto *IE = dyn_cast<InsertElementInst>(VL[0])) |
5997 | ScalarTy = IE->getOperand(1)->getType(); |
5998 | auto *VecTy = FixedVectorType::get(ScalarTy, VL.size()); |
5999 | TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput; |
6000 | |
6001 | // If we have computed a smaller type for the expression, update VecTy so |
6002 | // that the costs will be accurate. |
6003 | if (MinBWs.count(VL[0])) |
6004 | VecTy = FixedVectorType::get( |
6005 | IntegerType::get(F->getContext(), MinBWs[VL[0]].first), VL.size()); |
6006 | unsigned EntryVF = E->getVectorFactor(); |
6007 | auto *FinalVecTy = FixedVectorType::get(VecTy->getElementType(), EntryVF); |
6008 | |
6009 | bool NeedToShuffleReuses = !E->ReuseShuffleIndices.empty(); |
6010 | // FIXME: it tries to fix a problem with MSVC buildbots. |
6011 | TargetTransformInfo &TTIRef = *TTI; |
6012 | auto &&AdjustExtractsCost = [this, &TTIRef, CostKind, VL, VecTy, |
6013 | VectorizedVals, E](InstructionCost &Cost) { |
6014 | DenseMap<Value *, int> ExtractVectorsTys; |
6015 | SmallPtrSet<Value *, 4> CheckedExtracts; |
6016 | for (auto *V : VL) { |
6017 | if (isa<UndefValue>(V)) |
6018 | continue; |
6019 | // If all users of instruction are going to be vectorized and this |
6020 | // instruction itself is not going to be vectorized, consider this |
6021 | // instruction as dead and remove its cost from the final cost of the |
6022 | // vectorized tree. |
6023 | // Also, avoid adjusting the cost for extractelements with multiple uses |
6024 | // in different graph entries. |
6025 | const TreeEntry *VE = getTreeEntry(V); |
6026 | if (!CheckedExtracts.insert(V).second || |
6027 | !areAllUsersVectorized(cast<Instruction>(V), VectorizedVals) || |
6028 | (VE && VE != E)) |
6029 | continue; |
6030 | auto *EE = cast<ExtractElementInst>(V); |
6031 | Optional<unsigned> EEIdx = getExtractIndex(EE); |
6032 | if (!EEIdx) |
6033 | continue; |
6034 | unsigned Idx = *EEIdx; |
6035 | if (TTIRef.getNumberOfParts(VecTy) != |
6036 | TTIRef.getNumberOfParts(EE->getVectorOperandType())) { |
6037 | auto It = |
6038 | ExtractVectorsTys.try_emplace(EE->getVectorOperand(), Idx).first; |
6039 | It->getSecond() = std::min<int>(It->second, Idx); |
6040 | } |
6041 | // Take credit for instruction that will become dead. |
6042 | if (EE->hasOneUse()) { |
6043 | Instruction *Ext = EE->user_back(); |
6044 | if (isa<SExtInst, ZExtInst>(Ext) && all_of(Ext->users(), [](User *U) { |
6045 | return isa<GetElementPtrInst>(U); |
6046 | })) { |
6047 | // Use getExtractWithExtendCost() to calculate the cost of |
6048 | // extractelement/ext pair. |
6049 | Cost -= |
6050 | TTIRef.getExtractWithExtendCost(Ext->getOpcode(), Ext->getType(), |
6051 | EE->getVectorOperandType(), Idx); |
6052 | // Add back the cost of s|zext which is subtracted separately. |
6053 | Cost += TTIRef.getCastInstrCost( |
6054 | Ext->getOpcode(), Ext->getType(), EE->getType(), |
6055 | TTI::getCastContextHint(Ext), CostKind, Ext); |
6056 | continue; |
6057 | } |
6058 | } |
6059 | Cost -= TTIRef.getVectorInstrCost(*EE, EE->getVectorOperandType(), Idx); |
6060 | } |
6061 | // Add a cost for subvector extracts/inserts if required. |
6062 | for (const auto &Data : ExtractVectorsTys) { |
6063 | auto *EEVTy = cast<FixedVectorType>(Data.first->getType()); |
6064 | unsigned NumElts = VecTy->getNumElements(); |
6065 | if (Data.second % NumElts == 0) |
6066 | continue; |
6067 | if (TTIRef.getNumberOfParts(EEVTy) > TTIRef.getNumberOfParts(VecTy)) { |
6068 | unsigned Idx = (Data.second / NumElts) * NumElts; |
6069 | unsigned EENumElts = EEVTy->getNumElements(); |
6070 | if (Idx + NumElts <= EENumElts) { |
6071 | Cost += |
6072 | TTIRef.getShuffleCost(TargetTransformInfo::SK_ExtractSubvector, |
6073 | EEVTy, None, CostKind, Idx, VecTy); |
6074 | } else { |
6075 | // Need to round up the subvector type vectorization factor to avoid a |
6076 | // crash in cost model functions. Make SubVT so that Idx + VF of SubVT |
6077 | // <= EENumElts. |
6078 | auto *SubVT = |
6079 | FixedVectorType::get(VecTy->getElementType(), EENumElts - Idx); |
6080 | Cost += |
6081 | TTIRef.getShuffleCost(TargetTransformInfo::SK_ExtractSubvector, |
6082 | EEVTy, None, CostKind, Idx, SubVT); |
6083 | } |
6084 | } else { |
6085 | Cost += TTIRef.getShuffleCost(TargetTransformInfo::SK_InsertSubvector, |
6086 | VecTy, None, CostKind, 0, EEVTy); |
6087 | } |
6088 | } |
6089 | }; |
6090 | if (E->State == TreeEntry::NeedToGather) { |
6091 | if (allConstant(VL)) |
6092 | return 0; |
6093 | if (isa<InsertElementInst>(VL[0])) |
6094 | return InstructionCost::getInvalid(); |
6095 | SmallVector<int> Mask; |
6096 | SmallVector<const TreeEntry *> Entries; |
6097 | Optional<TargetTransformInfo::ShuffleKind> Shuffle = |
6098 | isGatherShuffledEntry(E, Mask, Entries); |
6099 | if (Shuffle) { |
6100 | InstructionCost GatherCost = 0; |
6101 | if (ShuffleVectorInst::isIdentityMask(Mask)) { |
6102 | // Perfect match in the graph, will reuse the previously vectorized |
6103 | // node. Cost is 0. |
6104 | 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) |
6105 | dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: perfect diamond match for gather bundle that starts with " << *VL.front() << ".\n"; } } while (false) |
6106 | << "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) |
6107 | << *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); |
6108 | if (NeedToShuffleReuses) |
6109 | GatherCost = |
6110 | TTI->getShuffleCost(TargetTransformInfo::SK_PermuteSingleSrc, |
6111 | FinalVecTy, E->ReuseShuffleIndices); |
6112 | } else { |
6113 | 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) |
6114 | << " 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) |
6115 | << *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); |
6116 | // Detected that instead of gather we can emit a shuffle of single/two |
6117 | // previously vectorized nodes. Add the cost of the permutation rather |
6118 | // than gather. |
6119 | ::addMask(Mask, E->ReuseShuffleIndices); |
6120 | GatherCost = TTI->getShuffleCost(*Shuffle, FinalVecTy, Mask); |
6121 | } |
6122 | return GatherCost; |
6123 | } |
6124 | if ((E->getOpcode() == Instruction::ExtractElement || |
6125 | all_of(E->Scalars, |
6126 | [](Value *V) { |
6127 | return isa<ExtractElementInst, UndefValue>(V); |
6128 | })) && |
6129 | allSameType(VL)) { |
6130 | // Check that gather of extractelements can be represented as just a |
6131 | // shuffle of a single/two vectors the scalars are extracted from. |
6132 | SmallVector<int> Mask; |
6133 | Optional<TargetTransformInfo::ShuffleKind> ShuffleKind = |
6134 | isFixedVectorShuffle(VL, Mask); |
6135 | if (ShuffleKind) { |
6136 | // Found the bunch of extractelement instructions that must be gathered |
6137 | // into a vector and can be represented as a permutation elements in a |
6138 | // single input vector or of 2 input vectors. |
6139 | InstructionCost Cost = |
6140 | computeExtractCost(VL, VecTy, *ShuffleKind, Mask, *TTI); |
6141 | AdjustExtractsCost(Cost); |
6142 | if (NeedToShuffleReuses) |
6143 | Cost += TTI->getShuffleCost(TargetTransformInfo::SK_PermuteSingleSrc, |
6144 | FinalVecTy, E->ReuseShuffleIndices); |
6145 | return Cost; |
6146 | } |
6147 | } |
6148 | if (isSplat(VL)) { |
6149 | // Found the broadcasting of the single scalar, calculate the cost as the |
6150 | // broadcast. |
6151 | assert(VecTy == FinalVecTy &&(static_cast <bool> (VecTy == FinalVecTy && "No reused scalars expected for broadcast." ) ? void (0) : __assert_fail ("VecTy == FinalVecTy && \"No reused scalars expected for broadcast.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 6152, __extension__ __PRETTY_FUNCTION__)) |
6152 | "No reused scalars expected for broadcast.")(static_cast <bool> (VecTy == FinalVecTy && "No reused scalars expected for broadcast." ) ? void (0) : __assert_fail ("VecTy == FinalVecTy && \"No reused scalars expected for broadcast.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 6152, __extension__ __PRETTY_FUNCTION__)); |
6153 | return TTI->getShuffleCost(TargetTransformInfo::SK_Broadcast, VecTy, |
6154 | /*Mask=*/None, CostKind, /*Index=*/0, |
6155 | /*SubTp=*/nullptr, /*Args=*/VL[0]); |
6156 | } |
6157 | InstructionCost ReuseShuffleCost = 0; |
6158 | if (NeedToShuffleReuses) |
6159 | ReuseShuffleCost = TTI->getShuffleCost( |
6160 | TTI::SK_PermuteSingleSrc, FinalVecTy, E->ReuseShuffleIndices); |
6161 | // Improve gather cost for gather of loads, if we can group some of the |
6162 | // loads into vector loads. |
6163 | if (VL.size() > 2 && E->getOpcode() == Instruction::Load && |
6164 | !E->isAltShuffle()) { |
6165 | BoUpSLP::ValueSet VectorizedLoads; |
6166 | unsigned StartIdx = 0; |
6167 | unsigned VF = VL.size() / 2; |
6168 | unsigned VectorizedCnt = 0; |
6169 | unsigned ScatterVectorizeCnt = 0; |
6170 | const unsigned Sz = DL->getTypeSizeInBits(E->getMainOp()->getType()); |
6171 | for (unsigned MinVF = getMinVF(2 * Sz); VF >= MinVF; VF /= 2) { |
6172 | for (unsigned Cnt = StartIdx, End = VL.size(); Cnt + VF <= End; |
6173 | Cnt += VF) { |
6174 | ArrayRef<Value *> Slice = VL.slice(Cnt, VF); |
6175 | if (!VectorizedLoads.count(Slice.front()) && |
6176 | !VectorizedLoads.count(Slice.back()) && allSameBlock(Slice)) { |
6177 | SmallVector<Value *> PointerOps; |
6178 | OrdersType CurrentOrder; |
6179 | LoadsState LS = |
6180 | canVectorizeLoads(Slice, Slice.front(), *TTI, *DL, *SE, *LI, |
6181 | CurrentOrder, PointerOps); |
6182 | switch (LS) { |
6183 | case LoadsState::Vectorize: |
6184 | case LoadsState::ScatterVectorize: |
6185 | // Mark the vectorized loads so that we don't vectorize them |
6186 | // again. |
6187 | if (LS == LoadsState::Vectorize) |
6188 | ++VectorizedCnt; |
6189 | else |
6190 | ++ScatterVectorizeCnt; |
6191 | VectorizedLoads.insert(Slice.begin(), Slice.end()); |
6192 | // If we vectorized initial block, no need to try to vectorize it |
6193 | // again. |
6194 | if (Cnt == StartIdx) |
6195 | StartIdx += VF; |
6196 | break; |
6197 | case LoadsState::Gather: |
6198 | break; |
6199 | } |
6200 | } |
6201 | } |
6202 | // Check if the whole array was vectorized already - exit. |
6203 | if (StartIdx >= VL.size()) |
6204 | break; |
6205 | // Found vectorizable parts - exit. |
6206 | if (!VectorizedLoads.empty()) |
6207 | break; |
6208 | } |
6209 | if (!VectorizedLoads.empty()) { |
6210 | InstructionCost GatherCost = 0; |
6211 | unsigned NumParts = TTI->getNumberOfParts(VecTy); |
6212 | bool NeedInsertSubvectorAnalysis = |
6213 | !NumParts || (VL.size() / VF) > NumParts; |
6214 | // Get the cost for gathered loads. |
6215 | for (unsigned I = 0, End = VL.size(); I < End; I += VF) { |
6216 | if (VectorizedLoads.contains(VL[I])) |
6217 | continue; |
6218 | GatherCost += getGatherCost(VL.slice(I, VF)); |
6219 | } |
6220 | // The cost for vectorized loads. |
6221 | InstructionCost ScalarsCost = 0; |
6222 | for (Value *V : VectorizedLoads) { |
6223 | auto *LI = cast<LoadInst>(V); |
6224 | ScalarsCost += TTI->getMemoryOpCost( |
6225 | Instruction::Load, LI->getType(), LI->getAlign(), |
6226 | LI->getPointerAddressSpace(), CostKind, |
6227 | {TTI::OK_AnyValue, TTI::OP_None}, LI); |
6228 | } |
6229 | auto *LI = cast<LoadInst>(E->getMainOp()); |
6230 | auto *LoadTy = FixedVectorType::get(LI->getType(), VF); |
6231 | Align Alignment = LI->getAlign(); |
6232 | GatherCost += VectorizedCnt * |
6233 | TTI->getMemoryOpCost(Instruction::Load, LoadTy, Alignment, |
6234 | LI->getPointerAddressSpace(), |
6235 | CostKind, {TTI::OK_AnyValue, |
6236 | TTI::OP_None}, LI); |
6237 | GatherCost += ScatterVectorizeCnt * |
6238 | TTI->getGatherScatterOpCost( |
6239 | Instruction::Load, LoadTy, LI->getPointerOperand(), |
6240 | /*VariableMask=*/false, Alignment, CostKind, LI); |
6241 | if (NeedInsertSubvectorAnalysis) { |
6242 | // Add the cost for the subvectors insert. |
6243 | for (int I = VF, E = VL.size(); I < E; I += VF) |
6244 | GatherCost += TTI->getShuffleCost(TTI::SK_InsertSubvector, VecTy, |
6245 | None, CostKind, I, LoadTy); |
6246 | } |
6247 | return ReuseShuffleCost + GatherCost - ScalarsCost; |
6248 | } |
6249 | } |
6250 | return ReuseShuffleCost + getGatherCost(VL); |
6251 | } |
6252 | InstructionCost CommonCost = 0; |
6253 | SmallVector<int> Mask; |
6254 | if (!E->ReorderIndices.empty()) { |
6255 | SmallVector<int> NewMask; |
6256 | if (E->getOpcode() == Instruction::Store) { |
6257 | // For stores the order is actually a mask. |
6258 | NewMask.resize(E->ReorderIndices.size()); |
6259 | copy(E->ReorderIndices, NewMask.begin()); |
6260 | } else { |
6261 | inversePermutation(E->ReorderIndices, NewMask); |
6262 | } |
6263 | ::addMask(Mask, NewMask); |
6264 | } |
6265 | if (NeedToShuffleReuses) |
6266 | ::addMask(Mask, E->ReuseShuffleIndices); |
6267 | if (!Mask.empty() && !ShuffleVectorInst::isIdentityMask(Mask)) |
6268 | CommonCost = |
6269 | TTI->getShuffleCost(TTI::SK_PermuteSingleSrc, FinalVecTy, Mask); |
6270 | 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", 6272, __extension__ __PRETTY_FUNCTION__)) |
6271 | 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", 6272, __extension__ __PRETTY_FUNCTION__)) |
6272 | "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", 6272, __extension__ __PRETTY_FUNCTION__)); |
6273 | 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", 6277, __extension__ __PRETTY_FUNCTION__)) |
6274 | ((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", 6277, __extension__ __PRETTY_FUNCTION__)) |
6275 | (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", 6277, __extension__ __PRETTY_FUNCTION__)) |
6276 | 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", 6277, __extension__ __PRETTY_FUNCTION__)) |
6277 | "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", 6277, __extension__ __PRETTY_FUNCTION__)); |
6278 | Instruction *VL0 = E->getMainOp(); |
6279 | unsigned ShuffleOrOp = |
6280 | E->isAltShuffle() ? (unsigned)Instruction::ShuffleVector : E->getOpcode(); |
6281 | switch (ShuffleOrOp) { |
6282 | case Instruction::PHI: |
6283 | return 0; |
6284 | |
6285 | case Instruction::ExtractValue: |
6286 | case Instruction::ExtractElement: { |
6287 | // The common cost of removal ExtractElement/ExtractValue instructions + |
6288 | // the cost of shuffles, if required to resuffle the original vector. |
6289 | if (NeedToShuffleReuses) { |
6290 | unsigned Idx = 0; |
6291 | for (unsigned I : E->ReuseShuffleIndices) { |
6292 | if (ShuffleOrOp == Instruction::ExtractElement) { |
6293 | auto *EE = cast<ExtractElementInst>(VL[I]); |
6294 | CommonCost -= TTI->getVectorInstrCost( |
6295 | *EE, EE->getVectorOperandType(), *getExtractIndex(EE)); |
6296 | } else { |
6297 | CommonCost -= TTI->getVectorInstrCost(Instruction::ExtractElement, |
6298 | VecTy, Idx); |
6299 | ++Idx; |
6300 | } |
6301 | } |
6302 | Idx = EntryVF; |
6303 | for (Value *V : VL) { |
6304 | if (ShuffleOrOp == Instruction::ExtractElement) { |
6305 | auto *EE = cast<ExtractElementInst>(V); |
6306 | CommonCost += TTI->getVectorInstrCost( |
6307 | *EE, EE->getVectorOperandType(), *getExtractIndex(EE)); |
6308 | } else { |
6309 | --Idx; |
6310 | CommonCost += TTI->getVectorInstrCost(Instruction::ExtractElement, |
6311 | VecTy, Idx); |
6312 | } |
6313 | } |
6314 | } |
6315 | if (ShuffleOrOp == Instruction::ExtractValue) { |
6316 | for (unsigned I = 0, E = VL.size(); I < E; ++I) { |
6317 | auto *EI = cast<Instruction>(VL[I]); |
6318 | // Take credit for instruction that will become dead. |
6319 | if (EI->hasOneUse()) { |
6320 | Instruction *Ext = EI->user_back(); |
6321 | if (isa<SExtInst, ZExtInst>(Ext) && |
6322 | all_of(Ext->users(), |
6323 | [](User *U) { return isa<GetElementPtrInst>(U); })) { |
6324 | // Use getExtractWithExtendCost() to calculate the cost of |
6325 | // extractelement/ext pair. |
6326 | CommonCost -= TTI->getExtractWithExtendCost( |
6327 | Ext->getOpcode(), Ext->getType(), VecTy, I); |
6328 | // Add back the cost of s|zext which is subtracted separately. |
6329 | CommonCost += TTI->getCastInstrCost( |
6330 | Ext->getOpcode(), Ext->getType(), EI->getType(), |
6331 | TTI::getCastContextHint(Ext), CostKind, Ext); |
6332 | continue; |
6333 | } |
6334 | } |
6335 | CommonCost -= |
6336 | TTI->getVectorInstrCost(Instruction::ExtractElement, VecTy, I); |
6337 | } |
6338 | } else { |
6339 | AdjustExtractsCost(CommonCost); |
6340 | } |
6341 | return CommonCost; |
6342 | } |
6343 | case Instruction::InsertElement: { |
6344 | 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", 6345, __extension__ __PRETTY_FUNCTION__)) |
6345 | "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", 6345, __extension__ __PRETTY_FUNCTION__)); |
6346 | auto *SrcVecTy = cast<FixedVectorType>(VL0->getType()); |
6347 | unsigned const NumElts = SrcVecTy->getNumElements(); |
6348 | unsigned const NumScalars = VL.size(); |
6349 | |
6350 | unsigned NumOfParts = TTI->getNumberOfParts(SrcVecTy); |
6351 | |
6352 | SmallVector<int> InsertMask(NumElts, UndefMaskElem); |
6353 | unsigned OffsetBeg = *getInsertIndex(VL.front()); |
6354 | unsigned OffsetEnd = OffsetBeg; |
6355 | InsertMask[OffsetBeg] = 0; |
6356 | for (auto [I, V] : enumerate(VL.drop_front())) { |
6357 | unsigned Idx = *getInsertIndex(V); |
6358 | if (OffsetBeg > Idx) |
6359 | OffsetBeg = Idx; |
6360 | else if (OffsetEnd < Idx) |
6361 | OffsetEnd = Idx; |
6362 | InsertMask[Idx] = I + 1; |
6363 | } |
6364 | unsigned VecScalarsSz = PowerOf2Ceil(NumElts); |
6365 | if (NumOfParts > 0) |
6366 | VecScalarsSz = PowerOf2Ceil((NumElts + NumOfParts - 1) / NumOfParts); |
6367 | unsigned VecSz = |
6368 | (1 + OffsetEnd / VecScalarsSz - OffsetBeg / VecScalarsSz) * |
6369 | VecScalarsSz; |
6370 | unsigned Offset = VecScalarsSz * (OffsetBeg / VecScalarsSz); |
6371 | unsigned InsertVecSz = std::min<unsigned>( |
6372 | PowerOf2Ceil(OffsetEnd - OffsetBeg + 1), |
6373 | ((OffsetEnd - OffsetBeg + VecScalarsSz) / VecScalarsSz) * |
6374 | VecScalarsSz); |
6375 | bool IsWholeSubvector = |
6376 | OffsetBeg == Offset && ((OffsetEnd + 1) % VecScalarsSz == 0); |
6377 | // Check if we can safely insert a subvector. If it is not possible, just |
6378 | // generate a whole-sized vector and shuffle the source vector and the new |
6379 | // subvector. |
6380 | if (OffsetBeg + InsertVecSz > VecSz) { |
6381 | // Align OffsetBeg to generate correct mask. |
6382 | OffsetBeg = alignDown(OffsetBeg, VecSz, Offset); |
6383 | InsertVecSz = VecSz; |
6384 | } |
6385 | |
6386 | APInt DemandedElts = APInt::getZero(NumElts); |
6387 | // TODO: Add support for Instruction::InsertValue. |
6388 | SmallVector<int> Mask; |
6389 | if (!E->ReorderIndices.empty()) { |
6390 | inversePermutation(E->ReorderIndices, Mask); |
6391 | Mask.append(InsertVecSz - Mask.size(), UndefMaskElem); |
6392 | } else { |
6393 | Mask.assign(VecSz, UndefMaskElem); |
6394 | std::iota(Mask.begin(), std::next(Mask.begin(), InsertVecSz), 0); |
6395 | } |
6396 | bool IsIdentity = true; |
6397 | SmallVector<int> PrevMask(InsertVecSz, UndefMaskElem); |
6398 | Mask.swap(PrevMask); |
6399 | for (unsigned I = 0; I < NumScalars; ++I) { |
6400 | unsigned InsertIdx = *getInsertIndex(VL[PrevMask[I]]); |
6401 | DemandedElts.setBit(InsertIdx); |
6402 | IsIdentity &= InsertIdx - OffsetBeg == I; |
6403 | Mask[InsertIdx - OffsetBeg] = I; |
6404 | } |
6405 | 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", 6405, __extension__ __PRETTY_FUNCTION__)); |
6406 | |
6407 | InstructionCost Cost = 0; |
6408 | Cost -= TTI->getScalarizationOverhead(SrcVecTy, DemandedElts, |
6409 | /*Insert*/ true, /*Extract*/ false); |
6410 | |
6411 | // First cost - resize to actual vector size if not identity shuffle or |
6412 | // need to shift the vector. |
6413 | // Do not calculate the cost if the actual size is the register size and |
6414 | // we can merge this shuffle with the following SK_Select. |
6415 | auto *InsertVecTy = |
6416 | FixedVectorType::get(SrcVecTy->getElementType(), InsertVecSz); |
6417 | if (!IsIdentity) |
6418 | Cost += TTI->getShuffleCost(TargetTransformInfo::SK_PermuteSingleSrc, |
6419 | InsertVecTy, Mask); |
6420 | auto *FirstInsert = cast<Instruction>(*find_if(E->Scalars, [E](Value *V) { |
6421 | return !is_contained(E->Scalars, cast<Instruction>(V)->getOperand(0)); |
6422 | })); |
6423 | // Second cost - permutation with subvector, if some elements are from the |
6424 | // initial vector or inserting a subvector. |
6425 | // TODO: Implement the analysis of the FirstInsert->getOperand(0) |
6426 | // subvector of ActualVecTy. |
6427 | if (!isUndefVector(FirstInsert->getOperand(0), InsertMask) && |
6428 | NumScalars != NumElts && !IsWholeSubvector) { |
6429 | if (InsertVecSz != VecSz) { |
6430 | auto *ActualVecTy = |
6431 | FixedVectorType::get(SrcVecTy->getElementType(), VecSz); |
6432 | Cost += |
6433 | TTI->getShuffleCost(TTI::SK_InsertSubvector, ActualVecTy, None, |
6434 | CostKind, OffsetBeg - Offset, InsertVecTy); |
6435 | } else { |
6436 | for (unsigned I = 0, End = OffsetBeg - Offset; I < End; ++I) |
6437 | Mask[I] = I; |
6438 | for (unsigned I = OffsetBeg - Offset, End = OffsetEnd - Offset; |
6439 | I <= End; ++I) |
6440 | if (Mask[I] != UndefMaskElem) |
6441 | Mask[I] = I + VecSz; |
6442 | for (unsigned I = OffsetEnd + 1 - Offset; I < VecSz; ++I) |
6443 | Mask[I] = I; |
6444 | Cost += TTI->getShuffleCost(TTI::SK_PermuteTwoSrc, InsertVecTy, Mask); |
6445 | } |
6446 | } |
6447 | return Cost; |
6448 | } |
6449 | case Instruction::ZExt: |
6450 | case Instruction::SExt: |
6451 | case Instruction::FPToUI: |
6452 | case Instruction::FPToSI: |
6453 | case Instruction::FPExt: |
6454 | case Instruction::PtrToInt: |
6455 | case Instruction::IntToPtr: |
6456 | case Instruction::SIToFP: |
6457 | case Instruction::UIToFP: |
6458 | case Instruction::Trunc: |
6459 | case Instruction::FPTrunc: |
6460 | case Instruction::BitCast: { |
6461 | Type *SrcTy = VL0->getOperand(0)->getType(); |
6462 | InstructionCost ScalarEltCost = |
6463 | TTI->getCastInstrCost(E->getOpcode(), ScalarTy, SrcTy, |
6464 | TTI::getCastContextHint(VL0), CostKind, VL0); |
6465 | if (NeedToShuffleReuses) { |
6466 | CommonCost -= (EntryVF - VL.size()) * ScalarEltCost; |
6467 | } |
6468 | |
6469 | // Calculate the cost of this instruction. |
6470 | InstructionCost ScalarCost = VL.size() * ScalarEltCost; |
6471 | |
6472 | auto *SrcVecTy = FixedVectorType::get(SrcTy, VL.size()); |
6473 | InstructionCost VecCost = 0; |
6474 | // Check if the values are candidates to demote. |
6475 | if (!MinBWs.count(VL0) || VecTy != SrcVecTy) { |
6476 | VecCost = CommonCost + TTI->getCastInstrCost( |
6477 | E->getOpcode(), VecTy, SrcVecTy, |
6478 | TTI::getCastContextHint(VL0), CostKind, VL0); |
6479 | } |
6480 | LLVM_DEBUG(dumpTreeCosts(E, CommonCost, VecCost, ScalarCost))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dumpTreeCosts(E, CommonCost, VecCost, ScalarCost); } } while (false); |
6481 | return VecCost - ScalarCost; |
6482 | } |
6483 | case Instruction::FCmp: |
6484 | case Instruction::ICmp: |
6485 | case Instruction::Select: { |
6486 | // Calculate the cost of this instruction. |
6487 | InstructionCost ScalarEltCost = |
6488 | TTI->getCmpSelInstrCost(E->getOpcode(), ScalarTy, Builder.getInt1Ty(), |
6489 | CmpInst::BAD_ICMP_PREDICATE, CostKind, VL0); |
6490 | if (NeedToShuffleReuses) { |
6491 | CommonCost -= (EntryVF - VL.size()) * ScalarEltCost; |
6492 | } |
6493 | auto *MaskTy = FixedVectorType::get(Builder.getInt1Ty(), VL.size()); |
6494 | InstructionCost ScalarCost = VecTy->getNumElements() * ScalarEltCost; |
6495 | |
6496 | // Check if all entries in VL are either compares or selects with compares |
6497 | // as condition that have the same predicates. |
6498 | CmpInst::Predicate VecPred = CmpInst::BAD_ICMP_PREDICATE; |
6499 | bool First = true; |
6500 | for (auto *V : VL) { |
6501 | CmpInst::Predicate CurrentPred; |
6502 | auto MatchCmp = m_Cmp(CurrentPred, m_Value(), m_Value()); |
6503 | if ((!match(V, m_Select(MatchCmp, m_Value(), m_Value())) && |
6504 | !match(V, MatchCmp)) || |
6505 | (!First && VecPred != CurrentPred)) { |
6506 | VecPred = CmpInst::BAD_ICMP_PREDICATE; |
6507 | break; |
6508 | } |
6509 | First = false; |
6510 | VecPred = CurrentPred; |
6511 | } |
6512 | |
6513 | InstructionCost VecCost = TTI->getCmpSelInstrCost( |
6514 | E->getOpcode(), VecTy, MaskTy, VecPred, CostKind, VL0); |
6515 | // Check if it is possible and profitable to use min/max for selects in |
6516 | // VL. |
6517 | // |
6518 | auto IntrinsicAndUse = canConvertToMinOrMaxIntrinsic(VL); |
6519 | if (IntrinsicAndUse.first != Intrinsic::not_intrinsic) { |
6520 | IntrinsicCostAttributes CostAttrs(IntrinsicAndUse.first, VecTy, |
6521 | {VecTy, VecTy}); |
6522 | InstructionCost IntrinsicCost = |
6523 | TTI->getIntrinsicInstrCost(CostAttrs, CostKind); |
6524 | // If the selects are the only uses of the compares, they will be dead |
6525 | // and we can adjust the cost by removing their cost. |
6526 | if (IntrinsicAndUse.second) |
6527 | IntrinsicCost -= TTI->getCmpSelInstrCost(Instruction::ICmp, VecTy, |
6528 | MaskTy, VecPred, CostKind); |
6529 | VecCost = std::min(VecCost, IntrinsicCost); |
6530 | } |
6531 | LLVM_DEBUG(dumpTreeCosts(E, CommonCost, VecCost, ScalarCost))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dumpTreeCosts(E, CommonCost, VecCost, ScalarCost); } } while (false); |
6532 | return CommonCost + VecCost - ScalarCost; |
6533 | } |
6534 | case Instruction::FNeg: |
6535 | case Instruction::Add: |
6536 | case Instruction::FAdd: |
6537 | case Instruction::Sub: |
6538 | case Instruction::FSub: |
6539 | case Instruction::Mul: |
6540 | case Instruction::FMul: |
6541 | case Instruction::UDiv: |
6542 | case Instruction::SDiv: |
6543 | case Instruction::FDiv: |
6544 | case Instruction::URem: |
6545 | case Instruction::SRem: |
6546 | case Instruction::FRem: |
6547 | case Instruction::Shl: |
6548 | case Instruction::LShr: |
6549 | case Instruction::AShr: |
6550 | case Instruction::And: |
6551 | case Instruction::Or: |
6552 | case Instruction::Xor: { |
6553 | const unsigned OpIdx = isa<BinaryOperator>(VL0) ? 1 : 0; |
6554 | |
6555 | InstructionCost ScalarCost = 0; |
6556 | for (auto *V : VL) { |
6557 | auto *VI = cast<Instruction>(V); |
6558 | TTI::OperandValueInfo Op1Info = TTI::getOperandInfo(VI->getOperand(0)); |
6559 | TTI::OperandValueInfo Op2Info = TTI::getOperandInfo(VI->getOperand(OpIdx)); |
6560 | SmallVector<const Value *, 4> Operands(VI->operand_values()); |
6561 | ScalarCost += |
6562 | TTI->getArithmeticInstrCost(E->getOpcode(), ScalarTy, CostKind, |
6563 | Op1Info, Op2Info, Operands, VI); |
6564 | } |
6565 | if (NeedToShuffleReuses) { |
6566 | CommonCost -= (EntryVF - VL.size()) * ScalarCost/VL.size(); |
6567 | } |
6568 | TTI::OperandValueInfo Op1Info = getOperandInfo(VL, 0); |
6569 | TTI::OperandValueInfo Op2Info = getOperandInfo(VL, OpIdx); |
6570 | InstructionCost VecCost = |
6571 | TTI->getArithmeticInstrCost(E->getOpcode(), VecTy, CostKind, |
6572 | Op1Info, Op2Info); |
6573 | LLVM_DEBUG(dumpTreeCosts(E, CommonCost, VecCost, ScalarCost))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dumpTreeCosts(E, CommonCost, VecCost, ScalarCost); } } while (false); |
6574 | return CommonCost + VecCost - ScalarCost; |
6575 | } |
6576 | case Instruction::GetElementPtr: { |
6577 | TargetTransformInfo::OperandValueKind Op1VK = |
6578 | TargetTransformInfo::OK_AnyValue; |
6579 | TargetTransformInfo::OperandValueKind Op2VK = |
6580 | any_of(VL, |
6581 | [](Value *V) { |
6582 | return isa<GetElementPtrInst>(V) && |
6583 | !isConstant( |
6584 | cast<GetElementPtrInst>(V)->getOperand(1)); |
6585 | }) |
6586 | ? TargetTransformInfo::OK_AnyValue |
6587 | : TargetTransformInfo::OK_UniformConstantValue; |
6588 | |
6589 | InstructionCost ScalarEltCost = TTI->getArithmeticInstrCost( |
6590 | Instruction::Add, ScalarTy, CostKind, |
6591 | {Op1VK, TargetTransformInfo::OP_None}, |
6592 | {Op2VK, TargetTransformInfo::OP_None}); |
6593 | if (NeedToShuffleReuses) { |
6594 | CommonCost -= (EntryVF - VL.size()) * ScalarEltCost; |
6595 | } |
6596 | InstructionCost ScalarCost = VecTy->getNumElements() * ScalarEltCost; |
6597 | InstructionCost VecCost = TTI->getArithmeticInstrCost( |
6598 | Instruction::Add, VecTy, CostKind, |
6599 | {Op1VK, TargetTransformInfo::OP_None}, |
6600 | {Op2VK, TargetTransformInfo::OP_None}); |
6601 | LLVM_DEBUG(dumpTreeCosts(E, CommonCost, VecCost, ScalarCost))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dumpTreeCosts(E, CommonCost, VecCost, ScalarCost); } } while (false); |
6602 | return CommonCost + VecCost - ScalarCost; |
6603 | } |
6604 | case Instruction::Load: { |
6605 | // Cost of wide load - cost of scalar loads. |
6606 | Align Alignment = cast<LoadInst>(VL0)->getAlign(); |
6607 | InstructionCost ScalarEltCost = |
6608 | TTI->getMemoryOpCost(Instruction::Load, ScalarTy, Alignment, 0, |
6609 | CostKind, {TTI::OK_AnyValue, TTI::OP_None}, VL0); |
6610 | if (NeedToShuffleReuses) { |
6611 | CommonCost -= (EntryVF - VL.size()) * ScalarEltCost; |
6612 | } |
6613 | InstructionCost ScalarLdCost = VecTy->getNumElements() * ScalarEltCost; |
6614 | InstructionCost VecLdCost; |
6615 | if (E->State == TreeEntry::Vectorize) { |
6616 | VecLdCost = TTI->getMemoryOpCost(Instruction::Load, VecTy, Alignment, 0, |
6617 | CostKind, {TTI::OK_AnyValue, TTI::OP_None}, VL0); |
6618 | } else { |
6619 | 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", 6619, __extension__ __PRETTY_FUNCTION__)); |
6620 | Align CommonAlignment = Alignment; |
6621 | for (Value *V : VL) |
6622 | CommonAlignment = |
6623 | std::min(CommonAlignment, cast<LoadInst>(V)->getAlign()); |
6624 | VecLdCost = TTI->getGatherScatterOpCost( |
6625 | Instruction::Load, VecTy, cast<LoadInst>(VL0)->getPointerOperand(), |
6626 | /*VariableMask=*/false, CommonAlignment, CostKind, VL0); |
6627 | } |
6628 | LLVM_DEBUG(dumpTreeCosts(E, CommonCost, VecLdCost, ScalarLdCost))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dumpTreeCosts(E, CommonCost, VecLdCost, ScalarLdCost ); } } while (false); |
6629 | return CommonCost + VecLdCost - ScalarLdCost; |
6630 | } |
6631 | case Instruction::Store: { |
6632 | // We know that we can merge the stores. Calculate the cost. |
6633 | bool IsReorder = !E->ReorderIndices.empty(); |
6634 | auto *SI = |
6635 | cast<StoreInst>(IsReorder ? VL[E->ReorderIndices.front()] : VL0); |
6636 | Align Alignment = SI->getAlign(); |
6637 | InstructionCost ScalarStCost = 0; |
6638 | for (auto *V : VL) { |
6639 | auto *VI = cast<Instruction>(V); |
6640 | TTI::OperandValueInfo OpInfo = TTI::getOperandInfo(VI->getOperand(0)); |
6641 | ScalarStCost += |
6642 | TTI->getMemoryOpCost(Instruction::Store, ScalarTy, Alignment, 0, |
6643 | CostKind, OpInfo, VI); |
6644 | } |
6645 | TTI::OperandValueInfo OpInfo = getOperandInfo(VL, 0); |
6646 | InstructionCost VecStCost = |
6647 | TTI->getMemoryOpCost(Instruction::Store, VecTy, Alignment, 0, CostKind, |
6648 | OpInfo); |
6649 | LLVM_DEBUG(dumpTreeCosts(E, CommonCost, VecStCost, ScalarStCost))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dumpTreeCosts(E, CommonCost, VecStCost, ScalarStCost ); } } while (false); |
6650 | return CommonCost + VecStCost - ScalarStCost; |
6651 | } |
6652 | case Instruction::Call: { |
6653 | CallInst *CI = cast<CallInst>(VL0); |
6654 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); |
6655 | |
6656 | // Calculate the cost of the scalar and vector calls. |
6657 | IntrinsicCostAttributes CostAttrs(ID, *CI, 1); |
6658 | InstructionCost ScalarEltCost = |
6659 | TTI->getIntrinsicInstrCost(CostAttrs, CostKind); |
6660 | if (NeedToShuffleReuses) { |
6661 | CommonCost -= (EntryVF - VL.size()) * ScalarEltCost; |
6662 | } |
6663 | InstructionCost ScalarCallCost = VecTy->getNumElements() * ScalarEltCost; |
6664 | |
6665 | auto VecCallCosts = getVectorCallCosts(CI, VecTy, TTI, TLI); |
6666 | InstructionCost VecCallCost = |
6667 | std::min(VecCallCosts.first, VecCallCosts.second); |
6668 | |
6669 | LLVM_DEBUG(dbgs() << "SLP: Call cost " << VecCallCost - ScalarCallCostdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Call cost " << VecCallCost - ScalarCallCost << " (" << VecCallCost << "-" << ScalarCallCost << ")" << " for " << *CI << "\n"; } } while (false) |
6670 | << " (" << VecCallCost << "-" << ScalarCallCost << ")"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Call cost " << VecCallCost - ScalarCallCost << " (" << VecCallCost << "-" << ScalarCallCost << ")" << " for " << *CI << "\n"; } } while (false) |
6671 | << " for " << *CI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Call cost " << VecCallCost - ScalarCallCost << " (" << VecCallCost << "-" << ScalarCallCost << ")" << " for " << *CI << "\n"; } } while (false); |
6672 | |
6673 | return CommonCost + VecCallCost - ScalarCallCost; |
6674 | } |
6675 | case Instruction::ShuffleVector: { |
6676 | 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", 6682, __extension__ __PRETTY_FUNCTION__)) |
6677 | ((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", 6682, __extension__ __PRETTY_FUNCTION__)) |
6678 | 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", 6682, __extension__ __PRETTY_FUNCTION__)) |
6679 | (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", 6682, __extension__ __PRETTY_FUNCTION__)) |
6680 | 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", 6682, __extension__ __PRETTY_FUNCTION__)) |
6681 | (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", 6682, __extension__ __PRETTY_FUNCTION__)) |
6682 | "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", 6682, __extension__ __PRETTY_FUNCTION__)); |
6683 | InstructionCost ScalarCost = 0; |
6684 | if (NeedToShuffleReuses) { |
6685 | for (unsigned Idx : E->ReuseShuffleIndices) { |
6686 | Instruction *I = cast<Instruction>(VL[Idx]); |
6687 | CommonCost -= TTI->getInstructionCost(I, CostKind); |
6688 | } |
6689 | for (Value *V : VL) { |
6690 | Instruction *I = cast<Instruction>(V); |
6691 | CommonCost += TTI->getInstructionCost(I, CostKind); |
6692 | } |
6693 | } |
6694 | for (Value *V : VL) { |
6695 | Instruction *I = cast<Instruction>(V); |
6696 | 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", 6696, __extension__ __PRETTY_FUNCTION__)); |
6697 | ScalarCost += TTI->getInstructionCost(I, CostKind); |
6698 | } |
6699 | // VecCost is equal to sum of the cost of creating 2 vectors |
6700 | // and the cost of creating shuffle. |
6701 | InstructionCost VecCost = 0; |
6702 | // Try to find the previous shuffle node with the same operands and same |
6703 | // main/alternate ops. |
6704 | auto &&TryFindNodeWithEqualOperands = [this, E]() { |
6705 | for (const std::unique_ptr<TreeEntry> &TE : VectorizableTree) { |
6706 | if (TE.get() == E) |
6707 | break; |
6708 | if (TE->isAltShuffle() && |
6709 | ((TE->getOpcode() == E->getOpcode() && |
6710 | TE->getAltOpcode() == E->getAltOpcode()) || |
6711 | (TE->getOpcode() == E->getAltOpcode() && |
6712 | TE->getAltOpcode() == E->getOpcode())) && |
6713 | TE->hasEqualOperands(*E)) |
6714 | return true; |
6715 | } |
6716 | return false; |
6717 | }; |
6718 | if (TryFindNodeWithEqualOperands()) { |
6719 | LLVM_DEBUG({do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { { dbgs() << "SLP: diamond match for alternate node found.\n" ; E->dump(); }; } } while (false) |
6720 | 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) |
6721 | E->dump();do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { { dbgs() << "SLP: diamond match for alternate node found.\n" ; E->dump(); }; } } while (false) |
6722 | })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { { dbgs() << "SLP: diamond match for alternate node found.\n" ; E->dump(); }; } } while (false); |
6723 | // No need to add new vector costs here since we're going to reuse |
6724 | // same main/alternate vector ops, just do different shuffling. |
6725 | } else if (Instruction::isBinaryOp(E->getOpcode())) { |
6726 | VecCost = TTI->getArithmeticInstrCost(E->getOpcode(), VecTy, CostKind); |
6727 | VecCost += TTI->getArithmeticInstrCost(E->getAltOpcode(), VecTy, |
6728 | CostKind); |
6729 | } else if (auto *CI0 = dyn_cast<CmpInst>(VL0)) { |
6730 | VecCost = TTI->getCmpSelInstrCost(E->getOpcode(), ScalarTy, |
6731 | Builder.getInt1Ty(), |
6732 | CI0->getPredicate(), CostKind, VL0); |
6733 | VecCost += TTI->getCmpSelInstrCost( |
6734 | E->getOpcode(), ScalarTy, Builder.getInt1Ty(), |
6735 | cast<CmpInst>(E->getAltOp())->getPredicate(), CostKind, |
6736 | E->getAltOp()); |
6737 | } else { |
6738 | Type *Src0SclTy = E->getMainOp()->getOperand(0)->getType(); |
6739 | Type *Src1SclTy = E->getAltOp()->getOperand(0)->getType(); |
6740 | auto *Src0Ty = FixedVectorType::get(Src0SclTy, VL.size()); |
6741 | auto *Src1Ty = FixedVectorType::get(Src1SclTy, VL.size()); |
6742 | VecCost = TTI->getCastInstrCost(E->getOpcode(), VecTy, Src0Ty, |
6743 | TTI::CastContextHint::None, CostKind); |
6744 | VecCost += TTI->getCastInstrCost(E->getAltOpcode(), VecTy, Src1Ty, |
6745 | TTI::CastContextHint::None, CostKind); |
6746 | } |
6747 | |
6748 | if (E->ReuseShuffleIndices.empty()) { |
6749 | CommonCost = |
6750 | TTI->getShuffleCost(TargetTransformInfo::SK_Select, FinalVecTy); |
6751 | } else { |
6752 | SmallVector<int> Mask; |
6753 | buildShuffleEntryMask( |
6754 | E->Scalars, E->ReorderIndices, E->ReuseShuffleIndices, |
6755 | [E](Instruction *I) { |
6756 | 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", 6756, __extension__ __PRETTY_FUNCTION__)); |
6757 | return I->getOpcode() == E->getAltOpcode(); |
6758 | }, |
6759 | Mask); |
6760 | CommonCost = TTI->getShuffleCost(TargetTransformInfo::SK_PermuteTwoSrc, |
6761 | FinalVecTy, Mask); |
6762 | } |
6763 | LLVM_DEBUG(dumpTreeCosts(E, CommonCost, VecCost, ScalarCost))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dumpTreeCosts(E, CommonCost, VecCost, ScalarCost); } } while (false); |
6764 | return CommonCost + VecCost - ScalarCost; |
6765 | } |
6766 | default: |
6767 | llvm_unreachable("Unknown instruction")::llvm::llvm_unreachable_internal("Unknown instruction", "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 6767); |
6768 | } |
6769 | } |
6770 | |
6771 | bool BoUpSLP::isFullyVectorizableTinyTree(bool ForReduction) const { |
6772 | 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) |
6773 | << 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); |
6774 | |
6775 | auto &&AreVectorizableGathers = [this](const TreeEntry *TE, unsigned Limit) { |
6776 | SmallVector<int> Mask; |
6777 | return TE->State == TreeEntry::NeedToGather && |
6778 | !any_of(TE->Scalars, |
6779 | [this](Value *V) { return EphValues.contains(V); }) && |
6780 | (allConstant(TE->Scalars) || isSplat(TE->Scalars) || |
6781 | TE->Scalars.size() < Limit || |
6782 | ((TE->getOpcode() == Instruction::ExtractElement || |
6783 | all_of(TE->Scalars, |
6784 | [](Value *V) { |
6785 | return isa<ExtractElementInst, UndefValue>(V); |
6786 | })) && |
6787 | isFixedVectorShuffle(TE->Scalars, Mask)) || |
6788 | (TE->State == TreeEntry::NeedToGather && |
6789 | TE->getOpcode() == Instruction::Load && !TE->isAltShuffle())); |
6790 | }; |
6791 | |
6792 | // We only handle trees of heights 1 and 2. |
6793 | if (VectorizableTree.size() == 1 && |
6794 | (VectorizableTree[0]->State == TreeEntry::Vectorize || |
6795 | (ForReduction && |
6796 | AreVectorizableGathers(VectorizableTree[0].get(), |
6797 | VectorizableTree[0]->Scalars.size()) && |
6798 | VectorizableTree[0]->getVectorFactor() > 2))) |
6799 | return true; |
6800 | |
6801 | if (VectorizableTree.size() != 2) |
6802 | return false; |
6803 | |
6804 | // Handle splat and all-constants stores. Also try to vectorize tiny trees |
6805 | // with the second gather nodes if they have less scalar operands rather than |
6806 | // the initial tree element (may be profitable to shuffle the second gather) |
6807 | // or they are extractelements, which form shuffle. |
6808 | SmallVector<int> Mask; |
6809 | if (VectorizableTree[0]->State == TreeEntry::Vectorize && |
6810 | AreVectorizableGathers(VectorizableTree[1].get(), |
6811 | VectorizableTree[0]->Scalars.size())) |
6812 | return true; |
6813 | |
6814 | // Gathering cost would be too much for tiny trees. |
6815 | if (VectorizableTree[0]->State == TreeEntry::NeedToGather || |
6816 | (VectorizableTree[1]->State == TreeEntry::NeedToGather && |
6817 | VectorizableTree[0]->State != TreeEntry::ScatterVectorize)) |
6818 | return false; |
6819 | |
6820 | return true; |
6821 | } |
6822 | |
6823 | static bool isLoadCombineCandidateImpl(Value *Root, unsigned NumElts, |
6824 | TargetTransformInfo *TTI, |
6825 | bool MustMatchOrInst) { |
6826 | // Look past the root to find a source value. Arbitrarily follow the |
6827 | // path through operand 0 of any 'or'. Also, peek through optional |
6828 | // shift-left-by-multiple-of-8-bits. |
6829 | Value *ZextLoad = Root; |
6830 | const APInt *ShAmtC; |
6831 | bool FoundOr = false; |
6832 | while (!isa<ConstantExpr>(ZextLoad) && |
6833 | (match(ZextLoad, m_Or(m_Value(), m_Value())) || |
6834 | (match(ZextLoad, m_Shl(m_Value(), m_APInt(ShAmtC))) && |
6835 | ShAmtC->urem(8) == 0))) { |
6836 | auto *BinOp = cast<BinaryOperator>(ZextLoad); |
6837 | ZextLoad = BinOp->getOperand(0); |
6838 | if (BinOp->getOpcode() == Instruction::Or) |
6839 | FoundOr = true; |
6840 | } |
6841 | // Check if the input is an extended load of the required or/shift expression. |
6842 | Value *Load; |
6843 | if ((MustMatchOrInst && !FoundOr) || ZextLoad == Root || |
6844 | !match(ZextLoad, m_ZExt(m_Value(Load))) || !isa<LoadInst>(Load)) |
6845 | return false; |
6846 | |
6847 | // Require that the total load bit width is a legal integer type. |
6848 | // For example, <8 x i8> --> i64 is a legal integer on a 64-bit target. |
6849 | // But <16 x i8> --> i128 is not, so the backend probably can't reduce it. |
6850 | Type *SrcTy = Load->getType(); |
6851 | unsigned LoadBitWidth = SrcTy->getIntegerBitWidth() * NumElts; |
6852 | if (!TTI->isTypeLegal(IntegerType::get(Root->getContext(), LoadBitWidth))) |
6853 | return false; |
6854 | |
6855 | // Everything matched - assume that we can fold the whole sequence using |
6856 | // load combining. |
6857 | 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) |
6858 | << *(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); |
6859 | |
6860 | return true; |
6861 | } |
6862 | |
6863 | bool BoUpSLP::isLoadCombineReductionCandidate(RecurKind RdxKind) const { |
6864 | if (RdxKind != RecurKind::Or) |
6865 | return false; |
6866 | |
6867 | unsigned NumElts = VectorizableTree[0]->Scalars.size(); |
6868 | Value *FirstReduced = VectorizableTree[0]->Scalars[0]; |
6869 | return isLoadCombineCandidateImpl(FirstReduced, NumElts, TTI, |
6870 | /* MatchOr */ false); |
6871 | } |
6872 | |
6873 | bool BoUpSLP::isLoadCombineCandidate() const { |
6874 | // Peek through a final sequence of stores and check if all operations are |
6875 | // likely to be load-combined. |
6876 | unsigned NumElts = VectorizableTree[0]->Scalars.size(); |
6877 | for (Value *Scalar : VectorizableTree[0]->Scalars) { |
6878 | Value *X; |
6879 | if (!match(Scalar, m_Store(m_Value(X), m_Value())) || |
6880 | !isLoadCombineCandidateImpl(X, NumElts, TTI, /* MatchOr */ true)) |
6881 | return false; |
6882 | } |
6883 | return true; |
6884 | } |
6885 | |
6886 | bool BoUpSLP::isTreeTinyAndNotFullyVectorizable(bool ForReduction) const { |
6887 | // No need to vectorize inserts of gathered values. |
6888 | if (VectorizableTree.size() == 2 && |
6889 | isa<InsertElementInst>(VectorizableTree[0]->Scalars[0]) && |
6890 | VectorizableTree[1]->State == TreeEntry::NeedToGather && |
6891 | (VectorizableTree[1]->getVectorFactor() <= 2 || |
6892 | !(isSplat(VectorizableTree[1]->Scalars) || |
6893 | allConstant(VectorizableTree[1]->Scalars)))) |
6894 | return true; |
6895 | |
6896 | // We can vectorize the tree if its size is greater than or equal to the |
6897 | // minimum size specified by the MinTreeSize command line option. |
6898 | if (VectorizableTree.size() >= MinTreeSize) |
6899 | return false; |
6900 | |
6901 | // If we have a tiny tree (a tree whose size is less than MinTreeSize), we |
6902 | // can vectorize it if we can prove it fully vectorizable. |
6903 | if (isFullyVectorizableTinyTree(ForReduction)) |
6904 | return false; |
6905 | |
6906 | 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", 6908, __extension__ __PRETTY_FUNCTION__)) |
6907 | ? 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", 6908, __extension__ __PRETTY_FUNCTION__)) |
6908 | : 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", 6908, __extension__ __PRETTY_FUNCTION__)); |
6909 | |
6910 | // Otherwise, we can't vectorize the tree. It is both tiny and not fully |
6911 | // vectorizable. |
6912 | return true; |
6913 | } |
6914 | |
6915 | InstructionCost BoUpSLP::getSpillCost() const { |
6916 | // Walk from the bottom of the tree to the top, tracking which values are |
6917 | // live. When we see a call instruction that is not part of our tree, |
6918 | // query TTI to see if there is a cost to keeping values live over it |
6919 | // (for example, if spills and fills are required). |
6920 | unsigned BundleWidth = VectorizableTree.front()->Scalars.size(); |
6921 | InstructionCost Cost = 0; |
6922 | |
6923 | SmallPtrSet<Instruction*, 4> LiveValues; |
6924 | Instruction *PrevInst = nullptr; |
6925 | |
6926 | // The entries in VectorizableTree are not necessarily ordered by their |
6927 | // position in basic blocks. Collect them and order them by dominance so later |
6928 | // instructions are guaranteed to be visited first. For instructions in |
6929 | // different basic blocks, we only scan to the beginning of the block, so |
6930 | // their order does not matter, as long as all instructions in a basic block |
6931 | // are grouped together. Using dominance ensures a deterministic order. |
6932 | SmallVector<Instruction *, 16> OrderedScalars; |
6933 | for (const auto &TEPtr : VectorizableTree) { |
6934 | Instruction *Inst = dyn_cast<Instruction>(TEPtr->Scalars[0]); |
6935 | if (!Inst) |
6936 | continue; |
6937 | OrderedScalars.push_back(Inst); |
6938 | } |
6939 | llvm::sort(OrderedScalars, [&](Instruction *A, Instruction *B) { |
6940 | auto *NodeA = DT->getNode(A->getParent()); |
6941 | auto *NodeB = DT->getNode(B->getParent()); |
6942 | 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", 6942, __extension__ __PRETTY_FUNCTION__)); |
6943 | 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", 6943, __extension__ __PRETTY_FUNCTION__)); |
6944 | 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", 6945, __extension__ __PRETTY_FUNCTION__)) |
6945 | "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", 6945, __extension__ __PRETTY_FUNCTION__)); |
6946 | if (NodeA != NodeB) |
6947 | return NodeA->getDFSNumIn() < NodeB->getDFSNumIn(); |
6948 | return B->comesBefore(A); |
6949 | }); |
6950 | |
6951 | for (Instruction *Inst : OrderedScalars) { |
6952 | if (!PrevInst) { |
6953 | PrevInst = Inst; |
6954 | continue; |
6955 | } |
6956 | |
6957 | // Update LiveValues. |
6958 | LiveValues.erase(PrevInst); |
6959 | for (auto &J : PrevInst->operands()) { |
6960 | if (isa<Instruction>(&*J) && getTreeEntry(&*J)) |
6961 | LiveValues.insert(cast<Instruction>(&*J)); |
6962 | } |
6963 | |
6964 | 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) |
6965 | 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) |
6966 | 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) |
6967 | 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) |
6968 | 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) |
6969 | 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) |
6970 | })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); |
6971 | |
6972 | // Now find the sequence of instructions between PrevInst and Inst. |
6973 | unsigned NumCalls = 0; |
6974 | BasicBlock::reverse_iterator InstIt = ++Inst->getIterator().getReverse(), |
6975 | PrevInstIt = |
6976 | PrevInst->getIterator().getReverse(); |
6977 | while (InstIt != PrevInstIt) { |
6978 | if (PrevInstIt == PrevInst->getParent()->rend()) { |
6979 | PrevInstIt = Inst->getParent()->rbegin(); |
6980 | continue; |
6981 | } |
6982 | |
6983 | // Debug information does not impact spill cost. |
6984 | if ((isa<CallInst>(&*PrevInstIt) && |
6985 | !isa<DbgInfoIntrinsic>(&*PrevInstIt)) && |
6986 | &*PrevInstIt != PrevInst) |
6987 | NumCalls++; |
6988 | |
6989 | ++PrevInstIt; |
6990 | } |
6991 | |
6992 | if (NumCalls) { |
6993 | SmallVector<Type*, 4> V; |
6994 | for (auto *II : LiveValues) { |
6995 | auto *ScalarTy = II->getType(); |
6996 | if (auto *VectorTy = dyn_cast<FixedVectorType>(ScalarTy)) |
6997 | ScalarTy = VectorTy->getElementType(); |
6998 | V.push_back(FixedVectorType::get(ScalarTy, BundleWidth)); |
6999 | } |
7000 | Cost += NumCalls * TTI->getCostOfKeepingLiveOverCall(V); |
7001 | } |
7002 | |
7003 | PrevInst = Inst; |
7004 | } |
7005 | |
7006 | return Cost; |
7007 | } |
7008 | |
7009 | /// Check if two insertelement instructions are from the same buildvector. |
7010 | static bool areTwoInsertFromSameBuildVector( |
7011 | InsertElementInst *VU, InsertElementInst *V, |
7012 | function_ref<Value *(InsertElementInst *)> GetBaseOperand) { |
7013 | // Instructions must be from the same basic blocks. |
7014 | if (VU->getParent() != V->getParent()) |
7015 | return false; |
7016 | // Checks if 2 insertelements are from the same buildvector. |
7017 | if (VU->getType() != V->getType()) |
7018 | return false; |
7019 | // Multiple used inserts are separate nodes. |
7020 | if (!VU->hasOneUse() && !V->hasOneUse()) |
7021 | return false; |
7022 | auto *IE1 = VU; |
7023 | auto *IE2 = V; |
7024 | unsigned Idx1 = *getInsertIndex(IE1); |
7025 | unsigned Idx2 = *getInsertIndex(IE2); |
7026 | // Go through the vector operand of insertelement instructions trying to find |
7027 | // either VU as the original vector for IE2 or V as the original vector for |
7028 | // IE1. |
7029 | do { |
7030 | if (IE2 == VU) |
7031 | return VU->hasOneUse(); |
7032 | if (IE1 == V) |
7033 | return V->hasOneUse(); |
7034 | if (IE1) { |
7035 | if ((IE1 != VU && !IE1->hasOneUse()) || |
7036 | getInsertIndex(IE1).value_or(Idx2) == Idx2) |
7037 | IE1 = nullptr; |
7038 | else |
7039 | IE1 = dyn_cast_or_null<InsertElementInst>(GetBaseOperand(IE1)); |
7040 | } |
7041 | if (IE2) { |
7042 | if ((IE2 != V && !IE2->hasOneUse()) || |
7043 | getInsertIndex(IE2).value_or(Idx1) == Idx1) |
7044 | IE2 = nullptr; |
7045 | else |
7046 | IE2 = dyn_cast_or_null<InsertElementInst>(GetBaseOperand(IE2)); |
7047 | } |
7048 | } while (IE1 || IE2); |
7049 | return false; |
7050 | } |
7051 | |
7052 | /// Checks if the \p IE1 instructions is followed by \p IE2 instruction in the |
7053 | /// buildvector sequence. |
7054 | static bool isFirstInsertElement(const InsertElementInst *IE1, |
7055 | const InsertElementInst *IE2) { |
7056 | if (IE1 == IE2) |
7057 | return false; |
7058 | const auto *I1 = IE1; |
7059 | const auto *I2 = IE2; |
7060 | const InsertElementInst *PrevI1; |
7061 | const InsertElementInst *PrevI2; |
7062 | unsigned Idx1 = *getInsertIndex(IE1); |
7063 | unsigned Idx2 = *getInsertIndex(IE2); |
7064 | do { |
7065 | if (I2 == IE1) |
7066 | return true; |
7067 | if (I1 == IE2) |
7068 | return false; |
7069 | PrevI1 = I1; |
7070 | PrevI2 = I2; |
7071 | if (I1 && (I1 == IE1 || I1->hasOneUse()) && |
7072 | getInsertIndex(I1).value_or(Idx2) != Idx2) |
7073 | I1 = dyn_cast<InsertElementInst>(I1->getOperand(0)); |
7074 | if (I2 && ((I2 == IE2 || I2->hasOneUse())) && |
7075 | getInsertIndex(I2).value_or(Idx1) != Idx1) |
7076 | I2 = dyn_cast<InsertElementInst>(I2->getOperand(0)); |
7077 | } while ((I1 && PrevI1 != I1) || (I2 && PrevI2 != I2)); |
7078 | llvm_unreachable("Two different buildvectors not expected.")::llvm::llvm_unreachable_internal("Two different buildvectors not expected." , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7078); |
7079 | } |
7080 | |
7081 | namespace { |
7082 | /// Returns incoming Value *, if the requested type is Value * too, or a default |
7083 | /// value, otherwise. |
7084 | struct ValueSelect { |
7085 | template <typename U> |
7086 | static std::enable_if_t<std::is_same<Value *, U>::value, Value *> |
7087 | get(Value *V) { |
7088 | return V; |
7089 | } |
7090 | template <typename U> |
7091 | static std::enable_if_t<!std::is_same<Value *, U>::value, U> get(Value *) { |
7092 | return U(); |
7093 | } |
7094 | }; |
7095 | } // namespace |
7096 | |
7097 | /// Does the analysis of the provided shuffle masks and performs the requested |
7098 | /// actions on the vectors with the given shuffle masks. It tries to do it in |
7099 | /// several steps. |
7100 | /// 1. If the Base vector is not undef vector, resizing the very first mask to |
7101 | /// have common VF and perform action for 2 input vectors (including non-undef |
7102 | /// Base). Other shuffle masks are combined with the resulting after the 1 stage |
7103 | /// and processed as a shuffle of 2 elements. |
7104 | /// 2. If the Base is undef vector and have only 1 shuffle mask, perform the |
7105 | /// action only for 1 vector with the given mask, if it is not the identity |
7106 | /// mask. |
7107 | /// 3. If > 2 masks are used, perform the remaining shuffle actions for 2 |
7108 | /// vectors, combing the masks properly between the steps. |
7109 | template <typename T> |
7110 | static T *performExtractsShuffleAction( |
7111 | MutableArrayRef<std::pair<T *, SmallVector<int>>> ShuffleMask, Value *Base, |
7112 | function_ref<unsigned(T *)> GetVF, |
7113 | function_ref<std::pair<T *, bool>(T *, ArrayRef<int>, bool)> ResizeAction, |
7114 | function_ref<T *(ArrayRef<int>, ArrayRef<T *>)> Action) { |
7115 | 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", 7115, __extension__ __PRETTY_FUNCTION__)); |
7116 | SmallVector<int> Mask(ShuffleMask.begin()->second); |
7117 | auto VMIt = std::next(ShuffleMask.begin()); |
7118 | T *Prev = nullptr; |
7119 | bool IsBaseNotUndef = !isUndefVector(Base, Mask); |
7120 | if (IsBaseNotUndef) { |
7121 | // Base is not undef, need to combine it with the next subvectors. |
7122 | std::pair<T *, bool> Res = |
7123 | ResizeAction(ShuffleMask.begin()->first, Mask, /*ForSingleMask=*/false); |
7124 | for (unsigned Idx = 0, VF = Mask.size(); Idx < VF; ++Idx) { |
7125 | if (Mask[Idx] == UndefMaskElem) |
7126 | Mask[Idx] = Idx; |
7127 | else |
7128 | Mask[Idx] = (Res.second ? Idx : Mask[Idx]) + VF; |
7129 | } |
7130 | auto *V = ValueSelect::get<T *>(Base); |
7131 | (void)V; |
7132 | 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", 7133, __extension__ __PRETTY_FUNCTION__)) |
7133 | "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", 7133, __extension__ __PRETTY_FUNCTION__)); |
7134 | Prev = Action(Mask, {nullptr, Res.first}); |
7135 | } else if (ShuffleMask.size() == 1) { |
7136 | // Base is undef and only 1 vector is shuffled - perform the action only for |
7137 | // single vector, if the mask is not the identity mask. |
7138 | std::pair<T *, bool> Res = ResizeAction(ShuffleMask.begin()->first, Mask, |
7139 | /*ForSingleMask=*/true); |
7140 | if (Res.second) |
7141 | // Identity mask is found. |
7142 | Prev = Res.first; |
7143 | else |
7144 | Prev = Action(Mask, {ShuffleMask.begin()->first}); |
7145 | } else { |
7146 | // Base is undef and at least 2 input vectors shuffled - perform 2 vectors |
7147 | // shuffles step by step, combining shuffle between the steps. |
7148 | unsigned Vec1VF = GetVF(ShuffleMask.begin()->first); |
7149 | unsigned Vec2VF = GetVF(VMIt->first); |
7150 | if (Vec1VF == Vec2VF) { |
7151 | // No need to resize the input vectors since they are of the same size, we |
7152 | // can shuffle them directly. |
7153 | ArrayRef<int> SecMask = VMIt->second; |
7154 | for (unsigned I = 0, VF = Mask.size(); I < VF; ++I) { |
7155 | if (SecMask[I] != UndefMaskElem) { |
7156 | assert(Mask[I] == UndefMaskElem && "Multiple uses of scalars.")(static_cast <bool> (Mask[I] == UndefMaskElem && "Multiple uses of scalars.") ? void (0) : __assert_fail ("Mask[I] == UndefMaskElem && \"Multiple uses of scalars.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7156, __extension__ __PRETTY_FUNCTION__)); |
7157 | Mask[I] = SecMask[I] + Vec1VF; |
7158 | } |
7159 | } |
7160 | Prev = Action(Mask, {ShuffleMask.begin()->first, VMIt->first}); |
7161 | } else { |
7162 | // Vectors of different sizes - resize and reshuffle. |
7163 | std::pair<T *, bool> Res1 = ResizeAction(ShuffleMask.begin()->first, Mask, |
7164 | /*ForSingleMask=*/false); |
7165 | std::pair<T *, bool> Res2 = |
7166 | ResizeAction(VMIt->first, VMIt->second, /*ForSingleMask=*/false); |
7167 | ArrayRef<int> SecMask = VMIt->second; |
7168 | for (unsigned I = 0, VF = Mask.size(); I < VF; ++I) { |
7169 | if (Mask[I] != UndefMaskElem) { |
7170 | assert(SecMask[I] == UndefMaskElem && "Multiple uses of scalars.")(static_cast <bool> (SecMask[I] == UndefMaskElem && "Multiple uses of scalars.") ? void (0) : __assert_fail ("SecMask[I] == UndefMaskElem && \"Multiple uses of scalars.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7170, __extension__ __PRETTY_FUNCTION__)); |
7171 | if (Res1.second) |
7172 | Mask[I] = I; |
7173 | } else if (SecMask[I] != UndefMaskElem) { |
7174 | assert(Mask[I] == UndefMaskElem && "Multiple uses of scalars.")(static_cast <bool> (Mask[I] == UndefMaskElem && "Multiple uses of scalars.") ? void (0) : __assert_fail ("Mask[I] == UndefMaskElem && \"Multiple uses of scalars.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7174, __extension__ __PRETTY_FUNCTION__)); |
7175 | Mask[I] = (Res2.second ? I : SecMask[I]) + VF; |
7176 | } |
7177 | } |
7178 | Prev = Action(Mask, {Res1.first, Res2.first}); |
7179 | } |
7180 | VMIt = std::next(VMIt); |
7181 | } |
7182 | // Perform requested actions for the remaining masks/vectors. |
7183 | for (auto E = ShuffleMask.end(); VMIt != E; ++VMIt) { |
7184 | // Shuffle other input vectors, if any. |
7185 | std::pair<T *, bool> Res = |
7186 | ResizeAction(VMIt->first, VMIt->second, /*ForSingleMask=*/false); |
7187 | ArrayRef<int> SecMask = VMIt->second; |
7188 | for (unsigned I = 0, VF = Mask.size(); I < VF; ++I) { |
7189 | if (SecMask[I] != UndefMaskElem) { |
7190 | assert((Mask[I] == UndefMaskElem || IsBaseNotUndef) &&(static_cast <bool> ((Mask[I] == UndefMaskElem || IsBaseNotUndef ) && "Multiple uses of scalars.") ? void (0) : __assert_fail ("(Mask[I] == UndefMaskElem || IsBaseNotUndef) && \"Multiple uses of scalars.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7191, __extension__ __PRETTY_FUNCTION__)) |
7191 | "Multiple uses of scalars.")(static_cast <bool> ((Mask[I] == UndefMaskElem || IsBaseNotUndef ) && "Multiple uses of scalars.") ? void (0) : __assert_fail ("(Mask[I] == UndefMaskElem || IsBaseNotUndef) && \"Multiple uses of scalars.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7191, __extension__ __PRETTY_FUNCTION__)); |
7192 | Mask[I] = (Res.second ? I : SecMask[I]) + VF; |
7193 | } else if (Mask[I] != UndefMaskElem) { |
7194 | Mask[I] = I; |
7195 | } |
7196 | } |
7197 | Prev = Action(Mask, {Prev, Res.first}); |
7198 | } |
7199 | return Prev; |
7200 | } |
7201 | |
7202 | InstructionCost BoUpSLP::getTreeCost(ArrayRef<Value *> VectorizedVals) { |
7203 | InstructionCost Cost = 0; |
7204 | 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) |
7205 | << VectorizableTree.size() << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Calculating cost for tree of size " << VectorizableTree.size() << ".\n"; } } while ( false); |
7206 | |
7207 | unsigned BundleWidth = VectorizableTree[0]->Scalars.size(); |
7208 | |
7209 | for (unsigned I = 0, E = VectorizableTree.size(); I < E; ++I) { |
7210 | TreeEntry &TE = *VectorizableTree[I]; |
7211 | if (TE.State == TreeEntry::NeedToGather) { |
7212 | if (const TreeEntry *E = getTreeEntry(TE.getMainOp()); |
7213 | E && E->getVectorFactor() == TE.getVectorFactor() && |
7214 | E->isSame(TE.Scalars)) { |
7215 | // Some gather nodes might be absolutely the same as some vectorizable |
7216 | // nodes after reordering, need to handle it. |
7217 | 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) |
7218 | << *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) |
7219 | << "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); |
7220 | continue; |
7221 | } |
7222 | } |
7223 | |
7224 | InstructionCost C = getEntryCost(&TE, VectorizedVals); |
7225 | Cost += C; |
7226 | 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) |
7227 | << " 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) |
7228 | << ".\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) |
7229 | << "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); |
7230 | } |
7231 | |
7232 | SmallPtrSet<Value *, 16> ExtractCostCalculated; |
7233 | InstructionCost ExtractCost = 0; |
7234 | SmallVector<MapVector<const TreeEntry *, SmallVector<int>>> ShuffleMasks; |
7235 | SmallVector<std::pair<Value *, const TreeEntry *>> FirstUsers; |
7236 | SmallVector<APInt> DemandedElts; |
7237 | for (ExternalUser &EU : ExternalUses) { |
7238 | // We only add extract cost once for the same scalar. |
7239 | if (!isa_and_nonnull<InsertElementInst>(EU.User) && |
7240 | !ExtractCostCalculated.insert(EU.Scalar).second) |
7241 | continue; |
7242 | |
7243 | // Uses by ephemeral values are free (because the ephemeral value will be |
7244 | // removed prior to code generation, and so the extraction will be |
7245 | // removed as well). |
7246 | if (EphValues.count(EU.User)) |
7247 | continue; |
7248 | |
7249 | // No extract cost for vector "scalar" |
7250 | if (isa<FixedVectorType>(EU.Scalar->getType())) |
7251 | continue; |
7252 | |
7253 | // Already counted the cost for external uses when tried to adjust the cost |
7254 | // for extractelements, no need to add it again. |
7255 | if (isa<ExtractElementInst>(EU.Scalar)) |
7256 | continue; |
7257 | |
7258 | // If found user is an insertelement, do not calculate extract cost but try |
7259 | // to detect it as a final shuffled/identity match. |
7260 | if (auto *VU = dyn_cast_or_null<InsertElementInst>(EU.User)) { |
7261 | if (auto *FTy = dyn_cast<FixedVectorType>(VU->getType())) { |
7262 | Optional<unsigned> InsertIdx = getInsertIndex(VU); |
7263 | if (InsertIdx) { |
7264 | const TreeEntry *ScalarTE = getTreeEntry(EU.Scalar); |
7265 | auto *It = find_if( |
7266 | FirstUsers, |
7267 | [this, VU](const std::pair<Value *, const TreeEntry *> &Pair) { |
7268 | return areTwoInsertFromSameBuildVector( |
7269 | VU, cast<InsertElementInst>(Pair.first), |
7270 | [this](InsertElementInst *II) -> Value * { |
7271 | Value *Op0 = II->getOperand(0); |
7272 | if (getTreeEntry(II) && !getTreeEntry(Op0)) |
7273 | return nullptr; |
7274 | return Op0; |
7275 | }); |
7276 | }); |
7277 | int VecId = -1; |
7278 | if (It == FirstUsers.end()) { |
7279 | (void)ShuffleMasks.emplace_back(); |
7280 | SmallVectorImpl<int> &Mask = ShuffleMasks.back()[ScalarTE]; |
7281 | if (Mask.empty()) |
7282 | Mask.assign(FTy->getNumElements(), UndefMaskElem); |
7283 | // Find the insertvector, vectorized in tree, if any. |
7284 | Value *Base = VU; |
7285 | while (auto *IEBase = dyn_cast<InsertElementInst>(Base)) { |
7286 | if (IEBase != EU.User && |
7287 | (!IEBase->hasOneUse() || |
7288 | getInsertIndex(IEBase).value_or(*InsertIdx) == *InsertIdx)) |
7289 | break; |
7290 | // Build the mask for the vectorized insertelement instructions. |
7291 | if (const TreeEntry *E = getTreeEntry(IEBase)) { |
7292 | VU = IEBase; |
7293 | do { |
7294 | IEBase = cast<InsertElementInst>(Base); |
7295 | int Idx = *getInsertIndex(IEBase); |
7296 | assert(Mask[Idx] == UndefMaskElem &&(static_cast <bool> (Mask[Idx] == UndefMaskElem && "InsertElementInstruction used already.") ? void (0) : __assert_fail ("Mask[Idx] == UndefMaskElem && \"InsertElementInstruction used already.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7297, __extension__ __PRETTY_FUNCTION__)) |
7297 | "InsertElementInstruction used already.")(static_cast <bool> (Mask[Idx] == UndefMaskElem && "InsertElementInstruction used already.") ? void (0) : __assert_fail ("Mask[Idx] == UndefMaskElem && \"InsertElementInstruction used already.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7297, __extension__ __PRETTY_FUNCTION__)); |
7298 | Mask[Idx] = Idx; |
7299 | Base = IEBase->getOperand(0); |
7300 | } while (E == getTreeEntry(Base)); |
7301 | break; |
7302 | } |
7303 | Base = cast<InsertElementInst>(Base)->getOperand(0); |
7304 | } |
7305 | FirstUsers.emplace_back(VU, ScalarTE); |
7306 | DemandedElts.push_back(APInt::getZero(FTy->getNumElements())); |
7307 | VecId = FirstUsers.size() - 1; |
7308 | } else { |
7309 | if (isFirstInsertElement(VU, cast<InsertElementInst>(It->first))) |
7310 | It->first = VU; |
7311 | VecId = std::distance(FirstUsers.begin(), It); |
7312 | } |
7313 | int InIdx = *InsertIdx; |
7314 | SmallVectorImpl<int> &Mask = ShuffleMasks[VecId][ScalarTE]; |
7315 | if (Mask.empty()) |
7316 | Mask.assign(FTy->getNumElements(), UndefMaskElem); |
7317 | Mask[InIdx] = EU.Lane; |
7318 | DemandedElts[VecId].setBit(InIdx); |
7319 | continue; |
7320 | } |
7321 | } |
7322 | } |
7323 | |
7324 | // If we plan to rewrite the tree in a smaller type, we will need to sign |
7325 | // extend the extracted value back to the original type. Here, we account |
7326 | // for the extract and the added cost of the sign extend if needed. |
7327 | auto *VecTy = FixedVectorType::get(EU.Scalar->getType(), BundleWidth); |
7328 | auto *ScalarRoot = VectorizableTree[0]->Scalars[0]; |
7329 | if (MinBWs.count(ScalarRoot)) { |
7330 | auto *MinTy = IntegerType::get(F->getContext(), MinBWs[ScalarRoot].first); |
7331 | auto Extend = |
7332 | MinBWs[ScalarRoot].second ? Instruction::SExt : Instruction::ZExt; |
7333 | VecTy = FixedVectorType::get(MinTy, BundleWidth); |
7334 | ExtractCost += TTI->getExtractWithExtendCost(Extend, EU.Scalar->getType(), |
7335 | VecTy, EU.Lane); |
7336 | } else { |
7337 | ExtractCost += |
7338 | TTI->getVectorInstrCost(Instruction::ExtractElement, VecTy, EU.Lane); |
7339 | } |
7340 | } |
7341 | |
7342 | InstructionCost SpillCost = getSpillCost(); |
7343 | Cost += SpillCost + ExtractCost; |
7344 | auto &&ResizeToVF = [this, &Cost](const TreeEntry *TE, ArrayRef<int> Mask, |
7345 | bool) { |
7346 | InstructionCost C = 0; |
7347 | unsigned VF = Mask.size(); |
7348 | unsigned VecVF = TE->getVectorFactor(); |
7349 | if (VF != VecVF && |
7350 | (any_of(Mask, [VF](int Idx) { return Idx >= static_cast<int>(VF); }) || |
7351 | (all_of(Mask, |
7352 | [VF](int Idx) { return Idx < 2 * static_cast<int>(VF); }) && |
7353 | !ShuffleVectorInst::isIdentityMask(Mask)))) { |
7354 | SmallVector<int> OrigMask(VecVF, UndefMaskElem); |
7355 | std::copy(Mask.begin(), std::next(Mask.begin(), std::min(VF, VecVF)), |
7356 | OrigMask.begin()); |
7357 | C = TTI->getShuffleCost( |
7358 | TTI::SK_PermuteSingleSrc, |
7359 | FixedVectorType::get(TE->getMainOp()->getType(), VecVF), OrigMask); |
7360 | 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) |
7361 | 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) |
7362 | << " 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) |
7363 | 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); |
7364 | Cost += C; |
7365 | return std::make_pair(TE, true); |
7366 | } |
7367 | return std::make_pair(TE, false); |
7368 | }; |
7369 | // Calculate the cost of the reshuffled vectors, if any. |
7370 | for (int I = 0, E = FirstUsers.size(); I < E; ++I) { |
7371 | Value *Base = cast<Instruction>(FirstUsers[I].first)->getOperand(0); |
7372 | unsigned VF = ShuffleMasks[I].begin()->second.size(); |
7373 | auto *FTy = FixedVectorType::get( |
7374 | cast<VectorType>(FirstUsers[I].first->getType())->getElementType(), VF); |
7375 | auto Vector = ShuffleMasks[I].takeVector(); |
7376 | auto &&EstimateShufflesCost = [this, FTy, |
7377 | &Cost](ArrayRef<int> Mask, |
7378 | ArrayRef<const TreeEntry *> TEs) { |
7379 | 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", 7380, __extension__ __PRETTY_FUNCTION__)) |
7380 | "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", 7380, __extension__ __PRETTY_FUNCTION__)); |
7381 | if (TEs.size() == 1) { |
7382 | int Limit = 2 * Mask.size(); |
7383 | if (!all_of(Mask, [Limit](int Idx) { return Idx < Limit; }) || |
7384 | !ShuffleVectorInst::isIdentityMask(Mask)) { |
7385 | InstructionCost C = |
7386 | TTI->getShuffleCost(TTI::SK_PermuteSingleSrc, FTy, Mask); |
7387 | 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) |
7388 | << " 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) |
7389 | "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) |
7390 | 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) |
7391 | 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); |
7392 | Cost += C; |
7393 | } |
7394 | } else { |
7395 | InstructionCost C = |
7396 | TTI->getShuffleCost(TTI::SK_PermuteTwoSrc, FTy, Mask); |
7397 | 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) |
7398 | << " 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) |
7399 | "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) |
7400 | 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) |
7401 | 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); |
7402 | Cost += C; |
7403 | } |
7404 | return TEs.back(); |
7405 | }; |
7406 | (void)performExtractsShuffleAction<const TreeEntry>( |
7407 | makeMutableArrayRef(Vector.data(), Vector.size()), Base, |
7408 | [](const TreeEntry *E) { return E->getVectorFactor(); }, ResizeToVF, |
7409 | EstimateShufflesCost); |
7410 | InstructionCost InsertCost = TTI->getScalarizationOverhead( |
7411 | cast<FixedVectorType>(FirstUsers[I].first->getType()), DemandedElts[I], |
7412 | /*Insert*/ true, /*Extract*/ false); |
7413 | Cost -= InsertCost; |
7414 | } |
7415 | |
7416 | #ifndef NDEBUG |
7417 | SmallString<256> Str; |
7418 | { |
7419 | raw_svector_ostream OS(Str); |
7420 | OS << "SLP: Spill Cost = " << SpillCost << ".\n" |
7421 | << "SLP: Extract Cost = " << ExtractCost << ".\n" |
7422 | << "SLP: Total Cost = " << Cost << ".\n"; |
7423 | } |
7424 | LLVM_DEBUG(dbgs() << Str)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << Str; } } while (false); |
7425 | if (ViewSLPTree) |
7426 | ViewGraph(this, "SLP" + F->getName(), false, Str); |
7427 | #endif |
7428 | |
7429 | return Cost; |
7430 | } |
7431 | |
7432 | Optional<TargetTransformInfo::ShuffleKind> |
7433 | BoUpSLP::isGatherShuffledEntry(const TreeEntry *TE, SmallVectorImpl<int> &Mask, |
7434 | SmallVectorImpl<const TreeEntry *> &Entries) { |
7435 | // TODO: currently checking only for Scalars in the tree entry, need to count |
7436 | // reused elements too for better cost estimation. |
7437 | Mask.assign(TE->Scalars.size(), UndefMaskElem); |
7438 | Entries.clear(); |
7439 | // Build a lists of values to tree entries. |
7440 | DenseMap<Value *, SmallPtrSet<const TreeEntry *, 4>> ValueToTEs; |
7441 | for (const std::unique_ptr<TreeEntry> &EntryPtr : VectorizableTree) { |
7442 | if (EntryPtr.get() == TE) |
7443 | break; |
7444 | if (EntryPtr->State != TreeEntry::NeedToGather) |
7445 | continue; |
7446 | for (Value *V : EntryPtr->Scalars) |
7447 | ValueToTEs.try_emplace(V).first->getSecond().insert(EntryPtr.get()); |
7448 | } |
7449 | // Find all tree entries used by the gathered values. If no common entries |
7450 | // found - not a shuffle. |
7451 | // Here we build a set of tree nodes for each gathered value and trying to |
7452 | // find the intersection between these sets. If we have at least one common |
7453 | // tree node for each gathered value - we have just a permutation of the |
7454 | // single vector. If we have 2 different sets, we're in situation where we |
7455 | // have a permutation of 2 input vectors. |
7456 | SmallVector<SmallPtrSet<const TreeEntry *, 4>> UsedTEs; |
7457 | DenseMap<Value *, int> UsedValuesEntry; |
7458 | for (Value *V : TE->Scalars) { |
7459 | if (isa<UndefValue>(V)) |
7460 | continue; |
7461 | // Build a list of tree entries where V is used. |
7462 | SmallPtrSet<const TreeEntry *, 4> VToTEs; |
7463 | auto It = ValueToTEs.find(V); |
7464 | if (It != ValueToTEs.end()) |
7465 | VToTEs = It->second; |
7466 | if (const TreeEntry *VTE = getTreeEntry(V)) |
7467 | VToTEs.insert(VTE); |
7468 | if (VToTEs.empty()) |
7469 | return None; |
7470 | if (UsedTEs.empty()) { |
7471 | // The first iteration, just insert the list of nodes to vector. |
7472 | UsedTEs.push_back(VToTEs); |
7473 | } else { |
7474 | // Need to check if there are any previously used tree nodes which use V. |
7475 | // If there are no such nodes, consider that we have another one input |
7476 | // vector. |
7477 | SmallPtrSet<const TreeEntry *, 4> SavedVToTEs(VToTEs); |
7478 | unsigned Idx = 0; |
7479 | for (SmallPtrSet<const TreeEntry *, 4> &Set : UsedTEs) { |
7480 | // Do we have a non-empty intersection of previously listed tree entries |
7481 | // and tree entries using current V? |
7482 | set_intersect(VToTEs, Set); |
7483 | if (!VToTEs.empty()) { |
7484 | // Yes, write the new subset and continue analysis for the next |
7485 | // scalar. |
7486 | Set.swap(VToTEs); |
7487 | break; |
7488 | } |
7489 | VToTEs = SavedVToTEs; |
7490 | ++Idx; |
7491 | } |
7492 | // No non-empty intersection found - need to add a second set of possible |
7493 | // source vectors. |
7494 | if (Idx == UsedTEs.size()) { |
7495 | // If the number of input vectors is greater than 2 - not a permutation, |
7496 | // fallback to the regular gather. |
7497 | if (UsedTEs.size() == 2) |
7498 | return None; |
7499 | UsedTEs.push_back(SavedVToTEs); |
7500 | Idx = UsedTEs.size() - 1; |
7501 | } |
7502 | UsedValuesEntry.try_emplace(V, Idx); |
7503 | } |
7504 | } |
7505 | |
7506 | if (UsedTEs.empty()) { |
7507 | assert(all_of(TE->Scalars, UndefValue::classof) &&(static_cast <bool> (all_of(TE->Scalars, UndefValue:: classof) && "Expected vector of undefs only.") ? void (0) : __assert_fail ("all_of(TE->Scalars, UndefValue::classof) && \"Expected vector of undefs only.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7508, __extension__ __PRETTY_FUNCTION__)) |
7508 | "Expected vector of undefs only.")(static_cast <bool> (all_of(TE->Scalars, UndefValue:: classof) && "Expected vector of undefs only.") ? void (0) : __assert_fail ("all_of(TE->Scalars, UndefValue::classof) && \"Expected vector of undefs only.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7508, __extension__ __PRETTY_FUNCTION__)); |
7509 | return None; |
7510 | } |
7511 | |
7512 | unsigned VF = 0; |
7513 | if (UsedTEs.size() == 1) { |
7514 | // Try to find the perfect match in another gather node at first. |
7515 | auto It = find_if(UsedTEs.front(), [TE](const TreeEntry *EntryPtr) { |
7516 | return EntryPtr->isSame(TE->Scalars); |
7517 | }); |
7518 | if (It != UsedTEs.front().end()) { |
7519 | Entries.push_back(*It); |
7520 | std::iota(Mask.begin(), Mask.end(), 0); |
7521 | return TargetTransformInfo::SK_PermuteSingleSrc; |
7522 | } |
7523 | // No perfect match, just shuffle, so choose the first tree node. |
7524 | Entries.push_back(*UsedTEs.front().begin()); |
7525 | } else { |
7526 | // Try to find nodes with the same vector factor. |
7527 | 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", 7527, __extension__ __PRETTY_FUNCTION__)); |
7528 | DenseMap<int, const TreeEntry *> VFToTE; |
7529 | for (const TreeEntry *TE : UsedTEs.front()) |
7530 | VFToTE.try_emplace(TE->getVectorFactor(), TE); |
7531 | for (const TreeEntry *TE : UsedTEs.back()) { |
7532 | auto It = VFToTE.find(TE->getVectorFactor()); |
7533 | if (It != VFToTE.end()) { |
7534 | VF = It->first; |
7535 | Entries.push_back(It->second); |
7536 | Entries.push_back(TE); |
7537 | break; |
7538 | } |
7539 | } |
7540 | // No 2 source vectors with the same vector factor - give up and do regular |
7541 | // gather. |
7542 | if (Entries.empty()) |
7543 | return None; |
7544 | } |
7545 | |
7546 | // Build a shuffle mask for better cost estimation and vector emission. |
7547 | for (int I = 0, E = TE->Scalars.size(); I < E; ++I) { |
7548 | Value *V = TE->Scalars[I]; |
7549 | if (isa<UndefValue>(V)) |
7550 | continue; |
7551 | unsigned Idx = UsedValuesEntry.lookup(V); |
7552 | const TreeEntry *VTE = Entries[Idx]; |
7553 | int FoundLane = VTE->findLaneForValue(V); |
7554 | Mask[I] = Idx * VF + FoundLane; |
7555 | // Extra check required by isSingleSourceMaskImpl function (called by |
7556 | // ShuffleVectorInst::isSingleSourceMask). |
7557 | if (Mask[I] >= 2 * E) |
7558 | return None; |
7559 | } |
7560 | switch (Entries.size()) { |
7561 | case 1: |
7562 | return TargetTransformInfo::SK_PermuteSingleSrc; |
7563 | case 2: |
7564 | return TargetTransformInfo::SK_PermuteTwoSrc; |
7565 | default: |
7566 | break; |
7567 | } |
7568 | return None; |
7569 | } |
7570 | |
7571 | InstructionCost BoUpSLP::getGatherCost(FixedVectorType *Ty, |
7572 | const APInt &ShuffledIndices, |
7573 | bool NeedToShuffle) const { |
7574 | InstructionCost Cost = |
7575 | TTI->getScalarizationOverhead(Ty, ~ShuffledIndices, /*Insert*/ true, |
7576 | /*Extract*/ false); |
7577 | if (NeedToShuffle) |
7578 | Cost += TTI->getShuffleCost(TargetTransformInfo::SK_PermuteSingleSrc, Ty); |
7579 | return Cost; |
7580 | } |
7581 | |
7582 | InstructionCost BoUpSLP::getGatherCost(ArrayRef<Value *> VL) const { |
7583 | // Find the type of the operands in VL. |
7584 | Type *ScalarTy = VL[0]->getType(); |
7585 | if (StoreInst *SI = dyn_cast<StoreInst>(VL[0])) |
7586 | ScalarTy = SI->getValueOperand()->getType(); |
7587 | auto *VecTy = FixedVectorType::get(ScalarTy, VL.size()); |
7588 | bool DuplicateNonConst = false; |
7589 | // Find the cost of inserting/extracting values from the vector. |
7590 | // Check if the same elements are inserted several times and count them as |
7591 | // shuffle candidates. |
7592 | APInt ShuffledElements = APInt::getZero(VL.size()); |
7593 | DenseSet<Value *> UniqueElements; |
7594 | // Iterate in reverse order to consider insert elements with the high cost. |
7595 | for (unsigned I = VL.size(); I > 0; --I) { |
7596 | unsigned Idx = I - 1; |
7597 | // No need to shuffle duplicates for constants. |
7598 | if (isConstant(VL[Idx])) { |
7599 | ShuffledElements.setBit(Idx); |
7600 | continue; |
7601 | } |
7602 | if (!UniqueElements.insert(VL[Idx]).second) { |
7603 | DuplicateNonConst = true; |
7604 | ShuffledElements.setBit(Idx); |
7605 | } |
7606 | } |
7607 | return getGatherCost(VecTy, ShuffledElements, DuplicateNonConst); |
7608 | } |
7609 | |
7610 | // Perform operand reordering on the instructions in VL and return the reordered |
7611 | // operands in Left and Right. |
7612 | void BoUpSLP::reorderInputsAccordingToOpcode(ArrayRef<Value *> VL, |
7613 | SmallVectorImpl<Value *> &Left, |
7614 | SmallVectorImpl<Value *> &Right, |
7615 | const DataLayout &DL, |
7616 | ScalarEvolution &SE, |
7617 | const BoUpSLP &R) { |
7618 | if (VL.empty()) |
7619 | return; |
7620 | VLOperands Ops(VL, DL, SE, R); |
7621 | // Reorder the operands in place. |
7622 | Ops.reorder(); |
7623 | Left = Ops.getVL(0); |
7624 | Right = Ops.getVL(1); |
7625 | } |
7626 | |
7627 | Instruction &BoUpSLP::getLastInstructionInBundle(const TreeEntry *E) { |
7628 | // Get the basic block this bundle is in. All instructions in the bundle |
7629 | // should be in this block (except for extractelement-like instructions with |
7630 | // constant indeces). |
7631 | auto *Front = E->getMainOp(); |
7632 | auto *BB = Front->getParent(); |
7633 | 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", 7640, __extension__ __PRETTY_FUNCTION__)) |
7634 | 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", 7640, __extension__ __PRETTY_FUNCTION__)) |
7635 | !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", 7640, __extension__ __PRETTY_FUNCTION__)) |
7636 | 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", 7640, __extension__ __PRETTY_FUNCTION__)) |
7637 | 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", 7640, __extension__ __PRETTY_FUNCTION__)) |
7638 | 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", 7640, __extension__ __PRETTY_FUNCTION__)) |
7639 | 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", 7640, __extension__ __PRETTY_FUNCTION__)) |
7640 | }))(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", 7640, __extension__ __PRETTY_FUNCTION__)); |
7641 | |
7642 | auto &&FindLastInst = [E, Front, this, &BB]() { |
7643 | Instruction *LastInst = Front; |
7644 | for (Value *V : E->Scalars) { |
7645 | auto *I = dyn_cast<Instruction>(V); |
7646 | if (!I) |
7647 | continue; |
7648 | if (LastInst->getParent() == I->getParent()) { |
7649 | if (LastInst->comesBefore(I)) |
7650 | LastInst = I; |
7651 | continue; |
7652 | } |
7653 | assert(isVectorLikeInstWithConstOps(LastInst) &&(static_cast <bool> (isVectorLikeInstWithConstOps(LastInst ) && isVectorLikeInstWithConstOps(I) && "Expected vector-like insts only." ) ? void (0) : __assert_fail ("isVectorLikeInstWithConstOps(LastInst) && isVectorLikeInstWithConstOps(I) && \"Expected vector-like insts only.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7655, __extension__ __PRETTY_FUNCTION__)) |
7654 | isVectorLikeInstWithConstOps(I) &&(static_cast <bool> (isVectorLikeInstWithConstOps(LastInst ) && isVectorLikeInstWithConstOps(I) && "Expected vector-like insts only." ) ? void (0) : __assert_fail ("isVectorLikeInstWithConstOps(LastInst) && isVectorLikeInstWithConstOps(I) && \"Expected vector-like insts only.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7655, __extension__ __PRETTY_FUNCTION__)) |
7655 | "Expected vector-like insts only.")(static_cast <bool> (isVectorLikeInstWithConstOps(LastInst ) && isVectorLikeInstWithConstOps(I) && "Expected vector-like insts only." ) ? void (0) : __assert_fail ("isVectorLikeInstWithConstOps(LastInst) && isVectorLikeInstWithConstOps(I) && \"Expected vector-like insts only.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7655, __extension__ __PRETTY_FUNCTION__)); |
7656 | if (!DT->isReachableFromEntry(LastInst->getParent())) { |
7657 | LastInst = I; |
7658 | continue; |
7659 | } |
7660 | if (!DT->isReachableFromEntry(I->getParent())) |
7661 | continue; |
7662 | auto *NodeA = DT->getNode(LastInst->getParent()); |
7663 | auto *NodeB = DT->getNode(I->getParent()); |
7664 | 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", 7664, __extension__ __PRETTY_FUNCTION__)); |
7665 | 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", 7665, __extension__ __PRETTY_FUNCTION__)); |
7666 | 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", 7668, __extension__ __PRETTY_FUNCTION__)) |
7667 | (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", 7668, __extension__ __PRETTY_FUNCTION__)) |
7668 | "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", 7668, __extension__ __PRETTY_FUNCTION__)); |
7669 | if (NodeA->getDFSNumIn() < NodeB->getDFSNumIn()) |
7670 | LastInst = I; |
7671 | } |
7672 | BB = LastInst->getParent(); |
7673 | return LastInst; |
7674 | }; |
7675 | |
7676 | auto &&FindFirstInst = [E, Front]() { |
7677 | Instruction *FirstInst = Front; |
7678 | for (Value *V : E->Scalars) { |
7679 | auto *I = dyn_cast<Instruction>(V); |
7680 | if (!I) |
7681 | continue; |
7682 | if (I->comesBefore(FirstInst)) |
7683 | FirstInst = I; |
7684 | } |
7685 | return FirstInst; |
7686 | }; |
7687 | |
7688 | // Set the insert point to the beginning of the basic block if the entry |
7689 | // should not be scheduled. |
7690 | if (E->State != TreeEntry::NeedToGather && |
7691 | doesNotNeedToSchedule(E->Scalars)) { |
7692 | Instruction *InsertInst; |
7693 | if (all_of(E->Scalars, isUsedOutsideBlock)) |
7694 | InsertInst = FindLastInst(); |
7695 | else |
7696 | InsertInst = FindFirstInst(); |
7697 | return *InsertInst; |
7698 | } |
7699 | |
7700 | // The last instruction in the bundle in program order. |
7701 | Instruction *LastInst = nullptr; |
7702 | |
7703 | // Find the last instruction. The common case should be that BB has been |
7704 | // scheduled, and the last instruction is VL.back(). So we start with |
7705 | // VL.back() and iterate over schedule data until we reach the end of the |
7706 | // bundle. The end of the bundle is marked by null ScheduleData. |
7707 | if (BlocksSchedules.count(BB)) { |
7708 | Value *V = E->isOneOf(E->Scalars.back()); |
7709 | if (doesNotNeedToBeScheduled(V)) |
7710 | V = *find_if_not(E->Scalars, doesNotNeedToBeScheduled); |
7711 | auto *Bundle = BlocksSchedules[BB]->getScheduleData(V); |
7712 | if (Bundle && Bundle->isPartOfBundle()) |
7713 | for (; Bundle; Bundle = Bundle->NextInBundle) |
7714 | if (Bundle->OpValue == Bundle->Inst) |
7715 | LastInst = Bundle->Inst; |
7716 | } |
7717 | |
7718 | // LastInst can still be null at this point if there's either not an entry |
7719 | // for BB in BlocksSchedules or there's no ScheduleData available for |
7720 | // VL.back(). This can be the case if buildTree_rec aborts for various |
7721 | // reasons (e.g., the maximum recursion depth is reached, the maximum region |
7722 | // size is reached, etc.). ScheduleData is initialized in the scheduling |
7723 | // "dry-run". |
7724 | // |
7725 | // If this happens, we can still find the last instruction by brute force. We |
7726 | // iterate forwards from Front (inclusive) until we either see all |
7727 | // instructions in the bundle or reach the end of the block. If Front is the |
7728 | // last instruction in program order, LastInst will be set to Front, and we |
7729 | // will visit all the remaining instructions in the block. |
7730 | // |
7731 | // One of the reasons we exit early from buildTree_rec is to place an upper |
7732 | // bound on compile-time. Thus, taking an additional compile-time hit here is |
7733 | // not ideal. However, this should be exceedingly rare since it requires that |
7734 | // we both exit early from buildTree_rec and that the bundle be out-of-order |
7735 | // (causing us to iterate all the way to the end of the block). |
7736 | if (!LastInst) |
7737 | LastInst = FindLastInst(); |
7738 | 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", 7738, __extension__ __PRETTY_FUNCTION__)); |
7739 | return *LastInst; |
7740 | } |
7741 | |
7742 | void BoUpSLP::setInsertPointAfterBundle(const TreeEntry *E) { |
7743 | auto *Front = E->getMainOp(); |
7744 | Instruction *LastInst = &getLastInstructionInBundle(E); |
7745 | 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", 7745, __extension__ __PRETTY_FUNCTION__)); |
7746 | // If the instruction is PHI, set the insert point after all the PHIs. |
7747 | bool IsPHI = isa<PHINode>(LastInst); |
7748 | if (IsPHI) |
7749 | LastInst = LastInst->getParent()->getFirstNonPHI(); |
7750 | if (IsPHI || (E->State != TreeEntry::NeedToGather && |
7751 | doesNotNeedToSchedule(E->Scalars))) { |
7752 | Builder.SetInsertPoint(LastInst); |
7753 | } else { |
7754 | // Set the insertion point after the last instruction in the bundle. Set the |
7755 | // debug location to Front. |
7756 | Builder.SetInsertPoint(LastInst->getParent(), |
7757 | std::next(LastInst->getIterator())); |
7758 | } |
7759 | Builder.SetCurrentDebugLocation(Front->getDebugLoc()); |
7760 | } |
7761 | |
7762 | Value *BoUpSLP::gather(ArrayRef<Value *> VL) { |
7763 | // List of instructions/lanes from current block and/or the blocks which are |
7764 | // part of the current loop. These instructions will be inserted at the end to |
7765 | // make it possible to optimize loops and hoist invariant instructions out of |
7766 | // the loops body with better chances for success. |
7767 | SmallVector<std::pair<Value *, unsigned>, 4> PostponedInsts; |
7768 | SmallSet<int, 4> PostponedIndices; |
7769 | Loop *L = LI->getLoopFor(Builder.GetInsertBlock()); |
7770 | auto &&CheckPredecessor = [](BasicBlock *InstBB, BasicBlock *InsertBB) { |
7771 | SmallPtrSet<BasicBlock *, 4> Visited; |
7772 | while (InsertBB && InsertBB != InstBB && Visited.insert(InsertBB).second) |
7773 | InsertBB = InsertBB->getSinglePredecessor(); |
7774 | return InsertBB && InsertBB == InstBB; |
7775 | }; |
7776 | for (int I = 0, E = VL.size(); I < E; ++I) { |
7777 | if (auto *Inst = dyn_cast<Instruction>(VL[I])) |
7778 | if ((CheckPredecessor(Inst->getParent(), Builder.GetInsertBlock()) || |
7779 | getTreeEntry(Inst) || (L && (L->contains(Inst)))) && |
7780 | PostponedIndices.insert(I).second) |
7781 | PostponedInsts.emplace_back(Inst, I); |
7782 | } |
7783 | |
7784 | auto &&CreateInsertElement = [this](Value *Vec, Value *V, unsigned Pos) { |
7785 | Vec = Builder.CreateInsertElement(Vec, V, Builder.getInt32(Pos)); |
7786 | auto *InsElt = dyn_cast<InsertElementInst>(Vec); |
7787 | if (!InsElt) |
7788 | return Vec; |
7789 | GatherShuffleSeq.insert(InsElt); |
7790 | CSEBlocks.insert(InsElt->getParent()); |
7791 | // Add to our 'need-to-extract' list. |
7792 | if (TreeEntry *Entry = getTreeEntry(V)) { |
7793 | // Find which lane we need to extract. |
7794 | unsigned FoundLane = Entry->findLaneForValue(V); |
7795 | ExternalUses.emplace_back(V, InsElt, FoundLane); |
7796 | } |
7797 | return Vec; |
7798 | }; |
7799 | Value *Val0 = |
7800 | isa<StoreInst>(VL[0]) ? cast<StoreInst>(VL[0])->getValueOperand() : VL[0]; |
7801 | FixedVectorType *VecTy = FixedVectorType::get(Val0->getType(), VL.size()); |
7802 | Value *Vec = PoisonValue::get(VecTy); |
7803 | SmallVector<int> NonConsts; |
7804 | // Insert constant values at first. |
7805 | for (int I = 0, E = VL.size(); I < E; ++I) { |
7806 | if (PostponedIndices.contains(I)) |
7807 | continue; |
7808 | if (!isConstant(VL[I])) { |
7809 | NonConsts.push_back(I); |
7810 | continue; |
7811 | } |
7812 | Vec = CreateInsertElement(Vec, VL[I], I); |
7813 | } |
7814 | // Insert non-constant values. |
7815 | for (int I : NonConsts) |
7816 | Vec = CreateInsertElement(Vec, VL[I], I); |
7817 | // Append instructions, which are/may be part of the loop, in the end to make |
7818 | // it possible to hoist non-loop-based instructions. |
7819 | for (const std::pair<Value *, unsigned> &Pair : PostponedInsts) |
7820 | Vec = CreateInsertElement(Vec, Pair.first, Pair.second); |
7821 | |
7822 | return Vec; |
7823 | } |
7824 | |
7825 | namespace { |
7826 | /// Merges shuffle masks and emits final shuffle instruction, if required. |
7827 | class ShuffleInstructionBuilder { |
7828 | IRBuilderBase &Builder; |
7829 | const unsigned VF = 0; |
7830 | bool IsFinalized = false; |
7831 | SmallVector<int, 4> Mask; |
7832 | /// Holds all of the instructions that we gathered. |
7833 | SetVector<Instruction *> &GatherShuffleSeq; |
7834 | /// A list of blocks that we are going to CSE. |
7835 | SetVector<BasicBlock *> &CSEBlocks; |
7836 | |
7837 | public: |
7838 | ShuffleInstructionBuilder(IRBuilderBase &Builder, unsigned VF, |
7839 | SetVector<Instruction *> &GatherShuffleSeq, |
7840 | SetVector<BasicBlock *> &CSEBlocks) |
7841 | : Builder(Builder), VF(VF), GatherShuffleSeq(GatherShuffleSeq), |
7842 | CSEBlocks(CSEBlocks) {} |
7843 | |
7844 | /// Adds a mask, inverting it before applying. |
7845 | void addInversedMask(ArrayRef<unsigned> SubMask) { |
7846 | if (SubMask.empty()) |
7847 | return; |
7848 | SmallVector<int, 4> NewMask; |
7849 | inversePermutation(SubMask, NewMask); |
7850 | addMask(NewMask); |
7851 | } |
7852 | |
7853 | /// Functions adds masks, merging them into single one. |
7854 | void addMask(ArrayRef<unsigned> SubMask) { |
7855 | SmallVector<int, 4> NewMask(SubMask); |
7856 | addMask(NewMask); |
7857 | } |
7858 | |
7859 | void addMask(ArrayRef<int> SubMask) { ::addMask(Mask, SubMask); } |
7860 | |
7861 | Value *finalize(Value *V) { |
7862 | IsFinalized = true; |
7863 | unsigned ValueVF = cast<FixedVectorType>(V->getType())->getNumElements(); |
7864 | if (VF == ValueVF && Mask.empty()) |
7865 | return V; |
7866 | SmallVector<int, 4> NormalizedMask(VF, UndefMaskElem); |
7867 | std::iota(NormalizedMask.begin(), NormalizedMask.end(), 0); |
7868 | addMask(NormalizedMask); |
7869 | |
7870 | if (VF == ValueVF && ShuffleVectorInst::isIdentityMask(Mask)) |
7871 | return V; |
7872 | Value *Vec = Builder.CreateShuffleVector(V, Mask, "shuffle"); |
7873 | if (auto *I = dyn_cast<Instruction>(Vec)) { |
7874 | GatherShuffleSeq.insert(I); |
7875 | CSEBlocks.insert(I->getParent()); |
7876 | } |
7877 | return Vec; |
7878 | } |
7879 | |
7880 | ~ShuffleInstructionBuilder() { |
7881 | assert((IsFinalized || Mask.empty()) &&(static_cast <bool> ((IsFinalized || Mask.empty()) && "Shuffle construction must be finalized.") ? void (0) : __assert_fail ("(IsFinalized || Mask.empty()) && \"Shuffle construction must be finalized.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7882, __extension__ __PRETTY_FUNCTION__)) |
7882 | "Shuffle construction must be finalized.")(static_cast <bool> ((IsFinalized || Mask.empty()) && "Shuffle construction must be finalized.") ? void (0) : __assert_fail ("(IsFinalized || Mask.empty()) && \"Shuffle construction must be finalized.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7882, __extension__ __PRETTY_FUNCTION__)); |
7883 | } |
7884 | }; |
7885 | } // namespace |
7886 | |
7887 | Value *BoUpSLP::vectorizeTree(ArrayRef<Value *> VL) { |
7888 | const unsigned VF = VL.size(); |
7889 | InstructionsState S = getSameOpcode(VL); |
7890 | // Special processing for GEPs bundle, which may include non-gep values. |
7891 | if (!S.getOpcode() && VL.front()->getType()->isPointerTy()) { |
7892 | const auto *It = |
7893 | find_if(VL, [](Value *V) { return isa<GetElementPtrInst>(V); }); |
7894 | if (It != VL.end()) |
7895 | S = getSameOpcode(*It); |
7896 | } |
7897 | if (S.getOpcode()) { |
7898 | if (TreeEntry *E = getTreeEntry(S.OpValue)) |
7899 | if (E->isSame(VL)) { |
7900 | Value *V = vectorizeTree(E); |
7901 | if (VF != cast<FixedVectorType>(V->getType())->getNumElements()) { |
7902 | if (!E->ReuseShuffleIndices.empty()) { |
7903 | // Reshuffle to get only unique values. |
7904 | // If some of the scalars are duplicated in the vectorization tree |
7905 | // entry, we do not vectorize them but instead generate a mask for |
7906 | // the reuses. But if there are several users of the same entry, |
7907 | // they may have different vectorization factors. This is especially |
7908 | // important for PHI nodes. In this case, we need to adapt the |
7909 | // resulting instruction for the user vectorization factor and have |
7910 | // to reshuffle it again to take only unique elements of the vector. |
7911 | // Without this code the function incorrectly returns reduced vector |
7912 | // instruction with the same elements, not with the unique ones. |
7913 | |
7914 | // block: |
7915 | // %phi = phi <2 x > { .., %entry} {%shuffle, %block} |
7916 | // %2 = shuffle <2 x > %phi, poison, <4 x > <1, 1, 0, 0> |
7917 | // ... (use %2) |
7918 | // %shuffle = shuffle <2 x> %2, poison, <2 x> {2, 0} |
7919 | // br %block |
7920 | SmallVector<int> UniqueIdxs(VF, UndefMaskElem); |
7921 | SmallSet<int, 4> UsedIdxs; |
7922 | int Pos = 0; |
7923 | int Sz = VL.size(); |
7924 | for (int Idx : E->ReuseShuffleIndices) { |
7925 | if (Idx != Sz && Idx != UndefMaskElem && |
7926 | UsedIdxs.insert(Idx).second) |
7927 | UniqueIdxs[Idx] = Pos; |
7928 | ++Pos; |
7929 | } |
7930 | 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", 7931, __extension__ __PRETTY_FUNCTION__)) |
7931 | "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", 7931, __extension__ __PRETTY_FUNCTION__)); |
7932 | UniqueIdxs.append(VF - UsedIdxs.size(), UndefMaskElem); |
7933 | V = Builder.CreateShuffleVector(V, UniqueIdxs, "shrink.shuffle"); |
7934 | } else { |
7935 | 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", 7937, __extension__ __PRETTY_FUNCTION__)) |
7936 | "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", 7937, __extension__ __PRETTY_FUNCTION__)) |
7937 | "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", 7937, __extension__ __PRETTY_FUNCTION__)); |
7938 | SmallVector<int> UniformMask(VF, 0); |
7939 | std::iota(UniformMask.begin(), UniformMask.end(), 0); |
7940 | V = Builder.CreateShuffleVector(V, UniformMask, "shrink.shuffle"); |
7941 | } |
7942 | if (auto *I = dyn_cast<Instruction>(V)) { |
7943 | GatherShuffleSeq.insert(I); |
7944 | CSEBlocks.insert(I->getParent()); |
7945 | } |
7946 | } |
7947 | return V; |
7948 | } |
7949 | } |
7950 | |
7951 | // Can't vectorize this, so simply build a new vector with each lane |
7952 | // corresponding to the requested value. |
7953 | return createBuildVector(VL); |
7954 | } |
7955 | Value *BoUpSLP::createBuildVector(ArrayRef<Value *> VL) { |
7956 | assert(any_of(VectorizableTree,(static_cast <bool> (any_of(VectorizableTree, [VL](const std::unique_ptr<TreeEntry> &TE) { return TE->State == TreeEntry::NeedToGather && TE->isSame(VL); }) && "Non-matching gather node.") ? void (0) : __assert_fail ("any_of(VectorizableTree, [VL](const std::unique_ptr<TreeEntry> &TE) { return TE->State == TreeEntry::NeedToGather && TE->isSame(VL); }) && \"Non-matching gather node.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7960, __extension__ __PRETTY_FUNCTION__)) |
7957 | [VL](const std::unique_ptr<TreeEntry> &TE) {(static_cast <bool> (any_of(VectorizableTree, [VL](const std::unique_ptr<TreeEntry> &TE) { return TE->State == TreeEntry::NeedToGather && TE->isSame(VL); }) && "Non-matching gather node.") ? void (0) : __assert_fail ("any_of(VectorizableTree, [VL](const std::unique_ptr<TreeEntry> &TE) { return TE->State == TreeEntry::NeedToGather && TE->isSame(VL); }) && \"Non-matching gather node.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7960, __extension__ __PRETTY_FUNCTION__)) |
7958 | return TE->State == TreeEntry::NeedToGather && TE->isSame(VL);(static_cast <bool> (any_of(VectorizableTree, [VL](const std::unique_ptr<TreeEntry> &TE) { return TE->State == TreeEntry::NeedToGather && TE->isSame(VL); }) && "Non-matching gather node.") ? void (0) : __assert_fail ("any_of(VectorizableTree, [VL](const std::unique_ptr<TreeEntry> &TE) { return TE->State == TreeEntry::NeedToGather && TE->isSame(VL); }) && \"Non-matching gather node.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7960, __extension__ __PRETTY_FUNCTION__)) |
7959 | }) &&(static_cast <bool> (any_of(VectorizableTree, [VL](const std::unique_ptr<TreeEntry> &TE) { return TE->State == TreeEntry::NeedToGather && TE->isSame(VL); }) && "Non-matching gather node.") ? void (0) : __assert_fail ("any_of(VectorizableTree, [VL](const std::unique_ptr<TreeEntry> &TE) { return TE->State == TreeEntry::NeedToGather && TE->isSame(VL); }) && \"Non-matching gather node.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7960, __extension__ __PRETTY_FUNCTION__)) |
7960 | "Non-matching gather node.")(static_cast <bool> (any_of(VectorizableTree, [VL](const std::unique_ptr<TreeEntry> &TE) { return TE->State == TreeEntry::NeedToGather && TE->isSame(VL); }) && "Non-matching gather node.") ? void (0) : __assert_fail ("any_of(VectorizableTree, [VL](const std::unique_ptr<TreeEntry> &TE) { return TE->State == TreeEntry::NeedToGather && TE->isSame(VL); }) && \"Non-matching gather node.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7960, __extension__ __PRETTY_FUNCTION__)); |
7961 | unsigned VF = VL.size(); |
7962 | // Exploit possible reuse of values across lanes. |
7963 | SmallVector<int> ReuseShuffleIndicies; |
7964 | SmallVector<Value *> UniqueValues; |
7965 | if (VL.size() > 2) { |
7966 | DenseMap<Value *, unsigned> UniquePositions; |
7967 | unsigned NumValues = |
7968 | std::distance(VL.begin(), find_if(reverse(VL), [](Value *V) { |
7969 | return !isa<UndefValue>(V); |
7970 | }).base()); |
7971 | VF = std::max<unsigned>(VF, PowerOf2Ceil(NumValues)); |
7972 | int UniqueVals = 0; |
7973 | for (Value *V : VL.drop_back(VL.size() - VF)) { |
7974 | if (isa<UndefValue>(V)) { |
7975 | ReuseShuffleIndicies.emplace_back(UndefMaskElem); |
7976 | continue; |
7977 | } |
7978 | if (isConstant(V)) { |
7979 | ReuseShuffleIndicies.emplace_back(UniqueValues.size()); |
7980 | UniqueValues.emplace_back(V); |
7981 | continue; |
7982 | } |
7983 | auto Res = UniquePositions.try_emplace(V, UniqueValues.size()); |
7984 | ReuseShuffleIndicies.emplace_back(Res.first->second); |
7985 | if (Res.second) { |
7986 | UniqueValues.emplace_back(V); |
7987 | ++UniqueVals; |
7988 | } |
7989 | } |
7990 | if (UniqueVals == 1 && UniqueValues.size() == 1) { |
7991 | // Emit pure splat vector. |
7992 | ReuseShuffleIndicies.append(VF - ReuseShuffleIndicies.size(), |
7993 | UndefMaskElem); |
7994 | } else if (UniqueValues.size() >= VF - 1 || UniqueValues.size() <= 1) { |
7995 | if (UniqueValues.empty()) { |
7996 | assert(all_of(VL, UndefValue::classof) && "Expected list of undefs.")(static_cast <bool> (all_of(VL, UndefValue::classof) && "Expected list of undefs.") ? void (0) : __assert_fail ("all_of(VL, UndefValue::classof) && \"Expected list of undefs.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 7996, __extension__ __PRETTY_FUNCTION__)); |
7997 | NumValues = VF; |
7998 | } |
7999 | ReuseShuffleIndicies.clear(); |
8000 | UniqueValues.clear(); |
8001 | UniqueValues.append(VL.begin(), std::next(VL.begin(), NumValues)); |
8002 | } |
8003 | UniqueValues.append(VF - UniqueValues.size(), |
8004 | PoisonValue::get(VL[0]->getType())); |
8005 | VL = UniqueValues; |
8006 | } |
8007 | |
8008 | ShuffleInstructionBuilder ShuffleBuilder(Builder, VF, GatherShuffleSeq, |
8009 | CSEBlocks); |
8010 | Value *Vec = gather(VL); |
8011 | if (!ReuseShuffleIndicies.empty()) { |
8012 | ShuffleBuilder.addMask(ReuseShuffleIndicies); |
8013 | Vec = ShuffleBuilder.finalize(Vec); |
8014 | } |
8015 | return Vec; |
8016 | } |
8017 | |
8018 | Value *BoUpSLP::vectorizeTree(TreeEntry *E) { |
8019 | IRBuilder<>::InsertPointGuard Guard(Builder); |
8020 | |
8021 | if (E->VectorizedValue) { |
8022 | 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); |
8023 | return E->VectorizedValue; |
8024 | } |
8025 | |
8026 | bool NeedToShuffleReuses = !E->ReuseShuffleIndices.empty(); |
8027 | unsigned VF = E->getVectorFactor(); |
8028 | ShuffleInstructionBuilder ShuffleBuilder(Builder, VF, GatherShuffleSeq, |
8029 | CSEBlocks); |
8030 | if (E->State == TreeEntry::NeedToGather) { |
8031 | if (E->getMainOp()) |
8032 | setInsertPointAfterBundle(E); |
8033 | Value *Vec; |
8034 | SmallVector<int> Mask; |
8035 | SmallVector<const TreeEntry *> Entries; |
8036 | Optional<TargetTransformInfo::ShuffleKind> Shuffle = |
8037 | isGatherShuffledEntry(E, Mask, Entries); |
8038 | if (Shuffle) { |
8039 | 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", 8040, __extension__ __PRETTY_FUNCTION__)) |
8040 | "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", 8040, __extension__ __PRETTY_FUNCTION__)); |
8041 | Vec = Builder.CreateShuffleVector(Entries.front()->VectorizedValue, |
8042 | Entries.back()->VectorizedValue, Mask); |
8043 | if (auto *I = dyn_cast<Instruction>(Vec)) { |
8044 | GatherShuffleSeq.insert(I); |
8045 | CSEBlocks.insert(I->getParent()); |
8046 | } |
8047 | } else { |
8048 | Vec = gather(E->Scalars); |
8049 | } |
8050 | if (NeedToShuffleReuses) { |
8051 | ShuffleBuilder.addMask(E->ReuseShuffleIndices); |
8052 | Vec = ShuffleBuilder.finalize(Vec); |
8053 | } |
8054 | E->VectorizedValue = Vec; |
8055 | return Vec; |
8056 | } |
8057 | |
8058 | 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", 8060, __extension__ __PRETTY_FUNCTION__)) |
8059 | 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", 8060, __extension__ __PRETTY_FUNCTION__)) |
8060 | "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", 8060, __extension__ __PRETTY_FUNCTION__)); |
8061 | unsigned ShuffleOrOp = |
8062 | E->isAltShuffle() ? (unsigned)Instruction::ShuffleVector : E->getOpcode(); |
8063 | Instruction *VL0 = E->getMainOp(); |
8064 | Type *ScalarTy = VL0->getType(); |
8065 | if (auto *Store = dyn_cast<StoreInst>(VL0)) |
8066 | ScalarTy = Store->getValueOperand()->getType(); |
8067 | else if (auto *IE = dyn_cast<InsertElementInst>(VL0)) |
8068 | ScalarTy = IE->getOperand(1)->getType(); |
8069 | auto *VecTy = FixedVectorType::get(ScalarTy, E->Scalars.size()); |
8070 | switch (ShuffleOrOp) { |
8071 | case Instruction::PHI: { |
8072 | 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", 8075, __extension__ __PRETTY_FUNCTION__)) |
8073 | 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", 8075, __extension__ __PRETTY_FUNCTION__)) |
8074 | !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", 8075, __extension__ __PRETTY_FUNCTION__)) |
8075 | "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", 8075, __extension__ __PRETTY_FUNCTION__)); |
8076 | auto *PH = cast<PHINode>(VL0); |
8077 | Builder.SetInsertPoint(PH->getParent()->getFirstNonPHI()); |
8078 | Builder.SetCurrentDebugLocation(PH->getDebugLoc()); |
8079 | PHINode *NewPhi = Builder.CreatePHI(VecTy, PH->getNumIncomingValues()); |
8080 | Value *V = NewPhi; |
8081 | |
8082 | // Adjust insertion point once all PHI's have been generated. |
8083 | Builder.SetInsertPoint(&*PH->getParent()->getFirstInsertionPt()); |
8084 | Builder.SetCurrentDebugLocation(PH->getDebugLoc()); |
8085 | |
8086 | ShuffleBuilder.addInversedMask(E->ReorderIndices); |
8087 | ShuffleBuilder.addMask(E->ReuseShuffleIndices); |
8088 | V = ShuffleBuilder.finalize(V); |
8089 | |
8090 | E->VectorizedValue = V; |
8091 | |
8092 | // PHINodes may have multiple entries from the same block. We want to |
8093 | // visit every block once. |
8094 | SmallPtrSet<BasicBlock*, 4> VisitedBBs; |
8095 | |
8096 | for (unsigned i = 0, e = PH->getNumIncomingValues(); i < e; ++i) { |
8097 | ValueList Operands; |
8098 | BasicBlock *IBB = PH->getIncomingBlock(i); |
8099 | |
8100 | if (!VisitedBBs.insert(IBB).second) { |
8101 | NewPhi->addIncoming(NewPhi->getIncomingValueForBlock(IBB), IBB); |
8102 | continue; |
8103 | } |
8104 | |
8105 | Builder.SetInsertPoint(IBB->getTerminator()); |
8106 | Builder.SetCurrentDebugLocation(PH->getDebugLoc()); |
8107 | Value *Vec = vectorizeTree(E->getOperand(i)); |
8108 | NewPhi->addIncoming(Vec, IBB); |
8109 | } |
8110 | |
8111 | 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", 8112, __extension__ __PRETTY_FUNCTION__)) |
8112 | "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", 8112, __extension__ __PRETTY_FUNCTION__)); |
8113 | return V; |
8114 | } |
8115 | |
8116 | case Instruction::ExtractElement: { |
8117 | Value *V = E->getSingleOperand(0); |
8118 | Builder.SetInsertPoint(VL0); |
8119 | ShuffleBuilder.addInversedMask(E->ReorderIndices); |
8120 | ShuffleBuilder.addMask(E->ReuseShuffleIndices); |
8121 | V = ShuffleBuilder.finalize(V); |
8122 | E->VectorizedValue = V; |
8123 | return V; |
8124 | } |
8125 | case Instruction::ExtractValue: { |
8126 | auto *LI = cast<LoadInst>(E->getSingleOperand(0)); |
8127 | Builder.SetInsertPoint(LI); |
8128 | auto *PtrTy = PointerType::get(VecTy, LI->getPointerAddressSpace()); |
8129 | Value *Ptr = Builder.CreateBitCast(LI->getOperand(0), PtrTy); |
8130 | LoadInst *V = Builder.CreateAlignedLoad(VecTy, Ptr, LI->getAlign()); |
8131 | Value *NewV = propagateMetadata(V, E->Scalars); |
8132 | ShuffleBuilder.addInversedMask(E->ReorderIndices); |
8133 | ShuffleBuilder.addMask(E->ReuseShuffleIndices); |
8134 | NewV = ShuffleBuilder.finalize(NewV); |
8135 | E->VectorizedValue = NewV; |
8136 | return NewV; |
8137 | } |
8138 | case Instruction::InsertElement: { |
8139 | 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", 8139, __extension__ __PRETTY_FUNCTION__)); |
8140 | Builder.SetInsertPoint(cast<Instruction>(E->Scalars.back())); |
8141 | Value *V = vectorizeTree(E->getOperand(1)); |
8142 | |
8143 | // Create InsertVector shuffle if necessary |
8144 | auto *FirstInsert = cast<Instruction>(*find_if(E->Scalars, [E](Value *V) { |
8145 | return !is_contained(E->Scalars, cast<Instruction>(V)->getOperand(0)); |
8146 | })); |
8147 | const unsigned NumElts = |
8148 | cast<FixedVectorType>(FirstInsert->getType())->getNumElements(); |
8149 | const unsigned NumScalars = E->Scalars.size(); |
8150 | |
8151 | unsigned Offset = *getInsertIndex(VL0); |
8152 | 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", 8152, __extension__ __PRETTY_FUNCTION__)); |
8153 | |
8154 | // Create shuffle to resize vector |
8155 | SmallVector<int> Mask; |
8156 | if (!E->ReorderIndices.empty()) { |
8157 | inversePermutation(E->ReorderIndices, Mask); |
8158 | Mask.append(NumElts - NumScalars, UndefMaskElem); |
8159 | } else { |
8160 | Mask.assign(NumElts, UndefMaskElem); |
8161 | std::iota(Mask.begin(), std::next(Mask.begin(), NumScalars), 0); |
8162 | } |
8163 | // Create InsertVector shuffle if necessary |
8164 | bool IsIdentity = true; |
8165 | SmallVector<int> PrevMask(NumElts, UndefMaskElem); |
8166 | Mask.swap(PrevMask); |
8167 | for (unsigned I = 0; I < NumScalars; ++I) { |
8168 | Value *Scalar = E->Scalars[PrevMask[I]]; |
8169 | unsigned InsertIdx = *getInsertIndex(Scalar); |
8170 | IsIdentity &= InsertIdx - Offset == I; |
8171 | Mask[InsertIdx - Offset] = I; |
8172 | } |
8173 | if (!IsIdentity || NumElts != NumScalars) { |
8174 | V = Builder.CreateShuffleVector(V, Mask); |
8175 | if (auto *I = dyn_cast<Instruction>(V)) { |
8176 | GatherShuffleSeq.insert(I); |
8177 | CSEBlocks.insert(I->getParent()); |
8178 | } |
8179 | } |
8180 | |
8181 | SmallVector<int> InsertMask(NumElts, UndefMaskElem); |
8182 | for (unsigned I = 0; I < NumElts; I++) { |
8183 | if (Mask[I] != UndefMaskElem) |
8184 | InsertMask[Offset + I] = NumElts + I; |
8185 | } |
8186 | if (Offset != 0 || |
8187 | !isUndefVector(FirstInsert->getOperand(0), InsertMask)) { |
8188 | for (unsigned I = 0; I < NumElts; I++) { |
8189 | if (InsertMask[I] == UndefMaskElem) |
8190 | InsertMask[I] = I; |
8191 | } |
8192 | |
8193 | V = Builder.CreateShuffleVector( |
8194 | FirstInsert->getOperand(0), V, InsertMask, |
8195 | cast<Instruction>(E->Scalars.back())->getName()); |
8196 | if (auto *I = dyn_cast<Instruction>(V)) { |
8197 | GatherShuffleSeq.insert(I); |
8198 | CSEBlocks.insert(I->getParent()); |
8199 | } |
8200 | } |
8201 | |
8202 | ++NumVectorInstructions; |
8203 | E->VectorizedValue = V; |
8204 | return V; |
8205 | } |
8206 | case Instruction::ZExt: |
8207 | case Instruction::SExt: |
8208 | case Instruction::FPToUI: |
8209 | case Instruction::FPToSI: |
8210 | case Instruction::FPExt: |
8211 | case Instruction::PtrToInt: |
8212 | case Instruction::IntToPtr: |
8213 | case Instruction::SIToFP: |
8214 | case Instruction::UIToFP: |
8215 | case Instruction::Trunc: |
8216 | case Instruction::FPTrunc: |
8217 | case Instruction::BitCast: { |
8218 | setInsertPointAfterBundle(E); |
8219 | |
8220 | Value *InVec = vectorizeTree(E->getOperand(0)); |
8221 | |
8222 | if (E->VectorizedValue) { |
8223 | 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); |
8224 | return E->VectorizedValue; |
8225 | } |
8226 | |
8227 | auto *CI = cast<CastInst>(VL0); |
8228 | Value *V = Builder.CreateCast(CI->getOpcode(), InVec, VecTy); |
8229 | ShuffleBuilder.addInversedMask(E->ReorderIndices); |
8230 | ShuffleBuilder.addMask(E->ReuseShuffleIndices); |
8231 | V = ShuffleBuilder.finalize(V); |
8232 | |
8233 | E->VectorizedValue = V; |
8234 | ++NumVectorInstructions; |
8235 | return V; |
8236 | } |
8237 | case Instruction::FCmp: |
8238 | case Instruction::ICmp: { |
8239 | setInsertPointAfterBundle(E); |
8240 | |
8241 | Value *L = vectorizeTree(E->getOperand(0)); |
8242 | Value *R = vectorizeTree(E->getOperand(1)); |
8243 | |
8244 | if (E->VectorizedValue) { |
8245 | 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); |
8246 | return E->VectorizedValue; |
8247 | } |
8248 | |
8249 | CmpInst::Predicate P0 = cast<CmpInst>(VL0)->getPredicate(); |
8250 | Value *V = Builder.CreateCmp(P0, L, R); |
8251 | propagateIRFlags(V, E->Scalars, VL0); |
8252 | ShuffleBuilder.addInversedMask(E->ReorderIndices); |
8253 | ShuffleBuilder.addMask(E->ReuseShuffleIndices); |
8254 | V = ShuffleBuilder.finalize(V); |
8255 | |
8256 | E->VectorizedValue = V; |
8257 | ++NumVectorInstructions; |
8258 | return V; |
8259 | } |
8260 | case Instruction::Select: { |
8261 | setInsertPointAfterBundle(E); |
8262 | |
8263 | Value *Cond = vectorizeTree(E->getOperand(0)); |
8264 | Value *True = vectorizeTree(E->getOperand(1)); |
8265 | Value *False = vectorizeTree(E->getOperand(2)); |
8266 | |
8267 | if (E->VectorizedValue) { |
8268 | 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); |
8269 | return E->VectorizedValue; |
8270 | } |
8271 | |
8272 | Value *V = Builder.CreateSelect(Cond, True, False); |
8273 | ShuffleBuilder.addInversedMask(E->ReorderIndices); |
8274 | ShuffleBuilder.addMask(E->ReuseShuffleIndices); |
8275 | V = ShuffleBuilder.finalize(V); |
8276 | |
8277 | E->VectorizedValue = V; |
8278 | ++NumVectorInstructions; |
8279 | return V; |
8280 | } |
8281 | case Instruction::FNeg: { |
8282 | setInsertPointAfterBundle(E); |
8283 | |
8284 | Value *Op = vectorizeTree(E->getOperand(0)); |
8285 | |
8286 | if (E->VectorizedValue) { |
8287 | 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); |
8288 | return E->VectorizedValue; |
8289 | } |
8290 | |
8291 | Value *V = Builder.CreateUnOp( |
8292 | static_cast<Instruction::UnaryOps>(E->getOpcode()), Op); |
8293 | propagateIRFlags(V, E->Scalars, VL0); |
8294 | if (auto *I = dyn_cast<Instruction>(V)) |
8295 | V = propagateMetadata(I, E->Scalars); |
8296 | |
8297 | ShuffleBuilder.addInversedMask(E->ReorderIndices); |
8298 | ShuffleBuilder.addMask(E->ReuseShuffleIndices); |
8299 | V = ShuffleBuilder.finalize(V); |
8300 | |
8301 | E->VectorizedValue = V; |
8302 | ++NumVectorInstructions; |
8303 | |
8304 | return V; |
8305 | } |
8306 | case Instruction::Add: |
8307 | case Instruction::FAdd: |
8308 | case Instruction::Sub: |
8309 | case Instruction::FSub: |
8310 | case Instruction::Mul: |
8311 | case Instruction::FMul: |
8312 | case Instruction::UDiv: |
8313 | case Instruction::SDiv: |
8314 | case Instruction::FDiv: |
8315 | case Instruction::URem: |
8316 | case Instruction::SRem: |
8317 | case Instruction::FRem: |
8318 | case Instruction::Shl: |
8319 | case Instruction::LShr: |
8320 | case Instruction::AShr: |
8321 | case Instruction::And: |
8322 | case Instruction::Or: |
8323 | case Instruction::Xor: { |
8324 | setInsertPointAfterBundle(E); |
8325 | |
8326 | Value *LHS = vectorizeTree(E->getOperand(0)); |
8327 | Value *RHS = vectorizeTree(E->getOperand(1)); |
8328 | |
8329 | if (E->VectorizedValue) { |
8330 | 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); |
8331 | return E->VectorizedValue; |
8332 | } |
8333 | |
8334 | Value *V = Builder.CreateBinOp( |
8335 | static_cast<Instruction::BinaryOps>(E->getOpcode()), LHS, |
8336 | RHS); |
8337 | propagateIRFlags(V, E->Scalars, VL0); |
8338 | if (auto *I = dyn_cast<Instruction>(V)) |
8339 | V = propagateMetadata(I, E->Scalars); |
8340 | |
8341 | ShuffleBuilder.addInversedMask(E->ReorderIndices); |
8342 | ShuffleBuilder.addMask(E->ReuseShuffleIndices); |
8343 | V = ShuffleBuilder.finalize(V); |
8344 | |
8345 | E->VectorizedValue = V; |
8346 | ++NumVectorInstructions; |
8347 | |
8348 | return V; |
8349 | } |
8350 | case Instruction::Load: { |
8351 | // Loads are inserted at the head of the tree because we don't want to |
8352 | // sink them all the way down past store instructions. |
8353 | setInsertPointAfterBundle(E); |
8354 | |
8355 | LoadInst *LI = cast<LoadInst>(VL0); |
8356 | Instruction *NewLI; |
8357 | unsigned AS = LI->getPointerAddressSpace(); |
8358 | Value *PO = LI->getPointerOperand(); |
8359 | if (E->State == TreeEntry::Vectorize) { |
8360 | Value *VecPtr = Builder.CreateBitCast(PO, VecTy->getPointerTo(AS)); |
8361 | NewLI = Builder.CreateAlignedLoad(VecTy, VecPtr, LI->getAlign()); |
8362 | |
8363 | // The pointer operand uses an in-tree scalar so we add the new BitCast |
8364 | // or LoadInst to ExternalUses list to make sure that an extract will |
8365 | // be generated in the future. |
8366 | if (TreeEntry *Entry = getTreeEntry(PO)) { |
8367 | // Find which lane we need to extract. |
8368 | unsigned FoundLane = Entry->findLaneForValue(PO); |
8369 | ExternalUses.emplace_back( |
8370 | PO, PO != VecPtr ? cast<User>(VecPtr) : NewLI, FoundLane); |
8371 | } |
8372 | } else { |
8373 | 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", 8373, __extension__ __PRETTY_FUNCTION__)); |
8374 | Value *VecPtr = vectorizeTree(E->getOperand(0)); |
8375 | // Use the minimum alignment of the gathered loads. |
8376 | Align CommonAlignment = LI->getAlign(); |
8377 | for (Value *V : E->Scalars) |
8378 | CommonAlignment = |
8379 | std::min(CommonAlignment, cast<LoadInst>(V)->getAlign()); |
8380 | NewLI = Builder.CreateMaskedGather(VecTy, VecPtr, CommonAlignment); |
8381 | } |
8382 | Value *V = propagateMetadata(NewLI, E->Scalars); |
8383 | |
8384 | ShuffleBuilder.addInversedMask(E->ReorderIndices); |
8385 | ShuffleBuilder.addMask(E->ReuseShuffleIndices); |
8386 | V = ShuffleBuilder.finalize(V); |
8387 | E->VectorizedValue = V; |
8388 | ++NumVectorInstructions; |
8389 | return V; |
8390 | } |
8391 | case Instruction::Store: { |
8392 | auto *SI = cast<StoreInst>(VL0); |
8393 | unsigned AS = SI->getPointerAddressSpace(); |
8394 | |
8395 | setInsertPointAfterBundle(E); |
8396 | |
8397 | Value *VecValue = vectorizeTree(E->getOperand(0)); |
8398 | ShuffleBuilder.addMask(E->ReorderIndices); |
8399 | VecValue = ShuffleBuilder.finalize(VecValue); |
8400 | |
8401 | Value *ScalarPtr = SI->getPointerOperand(); |
8402 | Value *VecPtr = Builder.CreateBitCast( |
8403 | ScalarPtr, VecValue->getType()->getPointerTo(AS)); |
8404 | StoreInst *ST = |
8405 | Builder.CreateAlignedStore(VecValue, VecPtr, SI->getAlign()); |
8406 | |
8407 | // The pointer operand uses an in-tree scalar, so add the new BitCast or |
8408 | // StoreInst to ExternalUses to make sure that an extract will be |
8409 | // generated in the future. |
8410 | if (TreeEntry *Entry = getTreeEntry(ScalarPtr)) { |
8411 | // Find which lane we need to extract. |
8412 | unsigned FoundLane = Entry->findLaneForValue(ScalarPtr); |
8413 | ExternalUses.push_back(ExternalUser( |
8414 | ScalarPtr, ScalarPtr != VecPtr ? cast<User>(VecPtr) : ST, |
8415 | FoundLane)); |
8416 | } |
8417 | |
8418 | Value *V = propagateMetadata(ST, E->Scalars); |
8419 | |
8420 | E->VectorizedValue = V; |
8421 | ++NumVectorInstructions; |
8422 | return V; |
8423 | } |
8424 | case Instruction::GetElementPtr: { |
8425 | auto *GEP0 = cast<GetElementPtrInst>(VL0); |
8426 | setInsertPointAfterBundle(E); |
8427 | |
8428 | Value *Op0 = vectorizeTree(E->getOperand(0)); |
8429 | |
8430 | SmallVector<Value *> OpVecs; |
8431 | for (int J = 1, N = GEP0->getNumOperands(); J < N; ++J) { |
8432 | Value *OpVec = vectorizeTree(E->getOperand(J)); |
8433 | OpVecs.push_back(OpVec); |
8434 | } |
8435 | |
8436 | Value *V = Builder.CreateGEP(GEP0->getSourceElementType(), Op0, OpVecs); |
8437 | if (Instruction *I = dyn_cast<GetElementPtrInst>(V)) { |
8438 | SmallVector<Value *> GEPs; |
8439 | for (Value *V : E->Scalars) { |
8440 | if (isa<GetElementPtrInst>(V)) |
8441 | GEPs.push_back(V); |
8442 | } |
8443 | V = propagateMetadata(I, GEPs); |
8444 | } |
8445 | |
8446 | ShuffleBuilder.addInversedMask(E->ReorderIndices); |
8447 | ShuffleBuilder.addMask(E->ReuseShuffleIndices); |
8448 | V = ShuffleBuilder.finalize(V); |
8449 | |
8450 | E->VectorizedValue = V; |
8451 | ++NumVectorInstructions; |
8452 | |
8453 | return V; |
8454 | } |
8455 | case Instruction::Call: { |
8456 | CallInst *CI = cast<CallInst>(VL0); |
8457 | setInsertPointAfterBundle(E); |
8458 | |
8459 | Intrinsic::ID IID = Intrinsic::not_intrinsic; |
8460 | if (Function *FI = CI->getCalledFunction()) |
8461 | IID = FI->getIntrinsicID(); |
8462 | |
8463 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); |
8464 | |
8465 | auto VecCallCosts = getVectorCallCosts(CI, VecTy, TTI, TLI); |
8466 | bool UseIntrinsic = ID != Intrinsic::not_intrinsic && |
8467 | VecCallCosts.first <= VecCallCosts.second; |
8468 | |
8469 | Value *ScalarArg = nullptr; |
8470 | std::vector<Value *> OpVecs; |
8471 | SmallVector<Type *, 2> TysForDecl = |
8472 | {FixedVectorType::get(CI->getType(), E->Scalars.size())}; |
8473 | for (int j = 0, e = CI->arg_size(); j < e; ++j) { |
8474 | ValueList OpVL; |
8475 | // Some intrinsics have scalar arguments. This argument should not be |
8476 | // vectorized. |
8477 | if (UseIntrinsic && isVectorIntrinsicWithScalarOpAtArg(IID, j)) { |
8478 | CallInst *CEI = cast<CallInst>(VL0); |
8479 | ScalarArg = CEI->getArgOperand(j); |
8480 | OpVecs.push_back(CEI->getArgOperand(j)); |
8481 | if (isVectorIntrinsicWithOverloadTypeAtArg(IID, j)) |
8482 | TysForDecl.push_back(ScalarArg->getType()); |
8483 | continue; |
8484 | } |
8485 | |
8486 | Value *OpVec = vectorizeTree(E->getOperand(j)); |
8487 | LLVM_DEBUG(dbgs() << "SLP: OpVec[" << j << "]: " << *OpVec << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: OpVec[" << j << "]: " << *OpVec << "\n"; } } while (false); |
8488 | OpVecs.push_back(OpVec); |
8489 | if (isVectorIntrinsicWithOverloadTypeAtArg(IID, j)) |
8490 | TysForDecl.push_back(OpVec->getType()); |
8491 | } |
8492 | |
8493 | Function *CF; |
8494 | if (!UseIntrinsic) { |
8495 | VFShape Shape = |
8496 | VFShape::get(*CI, ElementCount::getFixed(static_cast<unsigned>( |
8497 | VecTy->getNumElements())), |
8498 | false /*HasGlobalPred*/); |
8499 | CF = VFDatabase(*CI).getVectorizedFunction(Shape); |
8500 | } else { |
8501 | CF = Intrinsic::getDeclaration(F->getParent(), ID, TysForDecl); |
8502 | } |
8503 | |
8504 | SmallVector<OperandBundleDef, 1> OpBundles; |
8505 | CI->getOperandBundlesAsDefs(OpBundles); |
8506 | Value *V = Builder.CreateCall(CF, OpVecs, OpBundles); |
8507 | |
8508 | // The scalar argument uses an in-tree scalar so we add the new vectorized |
8509 | // call to ExternalUses list to make sure that an extract will be |
8510 | // generated in the future. |
8511 | if (ScalarArg) { |
8512 | if (TreeEntry *Entry = getTreeEntry(ScalarArg)) { |
8513 | // Find which lane we need to extract. |
8514 | unsigned FoundLane = Entry->findLaneForValue(ScalarArg); |
8515 | ExternalUses.push_back( |
8516 | ExternalUser(ScalarArg, cast<User>(V), FoundLane)); |
8517 | } |
8518 | } |
8519 | |
8520 | propagateIRFlags(V, E->Scalars, VL0); |
8521 | ShuffleBuilder.addInversedMask(E->ReorderIndices); |
8522 | ShuffleBuilder.addMask(E->ReuseShuffleIndices); |
8523 | V = ShuffleBuilder.finalize(V); |
8524 | |
8525 | E->VectorizedValue = V; |
8526 | ++NumVectorInstructions; |
8527 | return V; |
8528 | } |
8529 | case Instruction::ShuffleVector: { |
8530 | 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", 8536, __extension__ __PRETTY_FUNCTION__)) |
8531 | ((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", 8536, __extension__ __PRETTY_FUNCTION__)) |
8532 | 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", 8536, __extension__ __PRETTY_FUNCTION__)) |
8533 | (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", 8536, __extension__ __PRETTY_FUNCTION__)) |
8534 | 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", 8536, __extension__ __PRETTY_FUNCTION__)) |
8535 | (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", 8536, __extension__ __PRETTY_FUNCTION__)) |
8536 | "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", 8536, __extension__ __PRETTY_FUNCTION__)); |
8537 | |
8538 | Value *LHS = nullptr, *RHS = nullptr; |
8539 | if (Instruction::isBinaryOp(E->getOpcode()) || isa<CmpInst>(VL0)) { |
8540 | setInsertPointAfterBundle(E); |
8541 | LHS = vectorizeTree(E->getOperand(0)); |
8542 | RHS = vectorizeTree(E->getOperand(1)); |
8543 | } else { |
8544 | setInsertPointAfterBundle(E); |
8545 | LHS = vectorizeTree(E->getOperand(0)); |
8546 | } |
8547 | |
8548 | if (E->VectorizedValue) { |
8549 | 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); |
8550 | return E->VectorizedValue; |
8551 | } |
8552 | |
8553 | Value *V0, *V1; |
8554 | if (Instruction::isBinaryOp(E->getOpcode())) { |
8555 | V0 = Builder.CreateBinOp( |
8556 | static_cast<Instruction::BinaryOps>(E->getOpcode()), LHS, RHS); |
8557 | V1 = Builder.CreateBinOp( |
8558 | static_cast<Instruction::BinaryOps>(E->getAltOpcode()), LHS, RHS); |
8559 | } else if (auto *CI0 = dyn_cast<CmpInst>(VL0)) { |
8560 | V0 = Builder.CreateCmp(CI0->getPredicate(), LHS, RHS); |
8561 | auto *AltCI = cast<CmpInst>(E->getAltOp()); |
8562 | CmpInst::Predicate AltPred = AltCI->getPredicate(); |
8563 | V1 = Builder.CreateCmp(AltPred, LHS, RHS); |
8564 | } else { |
8565 | V0 = Builder.CreateCast( |
8566 | static_cast<Instruction::CastOps>(E->getOpcode()), LHS, VecTy); |
8567 | V1 = Builder.CreateCast( |
8568 | static_cast<Instruction::CastOps>(E->getAltOpcode()), LHS, VecTy); |
8569 | } |
8570 | // Add V0 and V1 to later analysis to try to find and remove matching |
8571 | // instruction, if any. |
8572 | for (Value *V : {V0, V1}) { |
8573 | if (auto *I = dyn_cast<Instruction>(V)) { |
8574 | GatherShuffleSeq.insert(I); |
8575 | CSEBlocks.insert(I->getParent()); |
8576 | } |
8577 | } |
8578 | |
8579 | // Create shuffle to take alternate operations from the vector. |
8580 | // Also, gather up main and alt scalar ops to propagate IR flags to |
8581 | // each vector operation. |
8582 | ValueList OpScalars, AltScalars; |
8583 | SmallVector<int> Mask; |
8584 | buildShuffleEntryMask( |
8585 | E->Scalars, E->ReorderIndices, E->ReuseShuffleIndices, |
8586 | [E](Instruction *I) { |
8587 | 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", 8587, __extension__ __PRETTY_FUNCTION__)); |
8588 | return isAlternateInstruction(I, E->getMainOp(), E->getAltOp()); |
8589 | }, |
8590 | Mask, &OpScalars, &AltScalars); |
8591 | |
8592 | propagateIRFlags(V0, OpScalars); |
8593 | propagateIRFlags(V1, AltScalars); |
8594 | |
8595 | Value *V = Builder.CreateShuffleVector(V0, V1, Mask); |
8596 | if (auto *I = dyn_cast<Instruction>(V)) { |
8597 | V = propagateMetadata(I, E->Scalars); |
8598 | GatherShuffleSeq.insert(I); |
8599 | CSEBlocks.insert(I->getParent()); |
8600 | } |
8601 | V = ShuffleBuilder.finalize(V); |
8602 | |
8603 | E->VectorizedValue = V; |
8604 | ++NumVectorInstructions; |
8605 | |
8606 | return V; |
8607 | } |
8608 | default: |
8609 | llvm_unreachable("unknown inst")::llvm::llvm_unreachable_internal("unknown inst", "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 8609); |
8610 | } |
8611 | return nullptr; |
8612 | } |
8613 | |
8614 | Value *BoUpSLP::vectorizeTree() { |
8615 | ExtraValueToDebugLocsMap ExternallyUsedValues; |
8616 | return vectorizeTree(ExternallyUsedValues); |
8617 | } |
8618 | |
8619 | namespace { |
8620 | /// Data type for handling buildvector sequences with the reused scalars from |
8621 | /// other tree entries. |
8622 | struct ShuffledInsertData { |
8623 | /// List of insertelements to be replaced by shuffles. |
8624 | SmallVector<InsertElementInst *> InsertElements; |
8625 | /// The parent vectors and shuffle mask for the given list of inserts. |
8626 | MapVector<Value *, SmallVector<int>> ValueMasks; |
8627 | }; |
8628 | } // namespace |
8629 | |
8630 | Value * |
8631 | BoUpSLP::vectorizeTree(ExtraValueToDebugLocsMap &ExternallyUsedValues) { |
8632 | // All blocks must be scheduled before any instructions are inserted. |
8633 | for (auto &BSIter : BlocksSchedules) { |
8634 | scheduleBlock(BSIter.second.get()); |
8635 | } |
8636 | |
8637 | Builder.SetInsertPoint(&F->getEntryBlock().front()); |
8638 | auto *VectorRoot = vectorizeTree(VectorizableTree[0].get()); |
8639 | |
8640 | // If the vectorized tree can be rewritten in a smaller type, we truncate the |
8641 | // vectorized root. InstCombine will then rewrite the entire expression. We |
8642 | // sign extend the extracted values below. |
8643 | auto *ScalarRoot = VectorizableTree[0]->Scalars[0]; |
8644 | if (MinBWs.count(ScalarRoot)) { |
8645 | if (auto *I = dyn_cast<Instruction>(VectorRoot)) { |
8646 | // If current instr is a phi and not the last phi, insert it after the |
8647 | // last phi node. |
8648 | if (isa<PHINode>(I)) |
8649 | Builder.SetInsertPoint(&*I->getParent()->getFirstInsertionPt()); |
8650 | else |
8651 | Builder.SetInsertPoint(&*++BasicBlock::iterator(I)); |
8652 | } |
8653 | auto BundleWidth = VectorizableTree[0]->Scalars.size(); |
8654 | auto *MinTy = IntegerType::get(F->getContext(), MinBWs[ScalarRoot].first); |
8655 | auto *VecTy = FixedVectorType::get(MinTy, BundleWidth); |
8656 | auto *Trunc = Builder.CreateTrunc(VectorRoot, VecTy); |
8657 | VectorizableTree[0]->VectorizedValue = Trunc; |
8658 | } |
8659 | |
8660 | LLVM_DEBUG(dbgs() << "SLP: Extracting " << ExternalUses.size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Extracting " << ExternalUses .size() << " values .\n"; } } while (false) |
8661 | << " values .\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Extracting " << ExternalUses .size() << " values .\n"; } } while (false); |
8662 | |
8663 | SmallVector<ShuffledInsertData> ShuffledInserts; |
8664 | // Maps vector instruction to original insertelement instruction |
8665 | DenseMap<Value *, InsertElementInst *> VectorToInsertElement; |
8666 | // Extract all of the elements with the external uses. |
8667 | for (const auto &ExternalUse : ExternalUses) { |
8668 | Value *Scalar = ExternalUse.Scalar; |
8669 | llvm::User *User = ExternalUse.User; |
8670 | |
8671 | // Skip users that we already RAUW. This happens when one instruction |
8672 | // has multiple uses of the same value. |
8673 | if (User && !is_contained(Scalar->users(), User)) |
8674 | continue; |
8675 | TreeEntry *E = getTreeEntry(Scalar); |
8676 | assert(E && "Invalid scalar")(static_cast <bool> (E && "Invalid scalar") ? void (0) : __assert_fail ("E && \"Invalid scalar\"", "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 8676, __extension__ __PRETTY_FUNCTION__)); |
8677 | 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", 8678, __extension__ __PRETTY_FUNCTION__)) |
8678 | "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", 8678, __extension__ __PRETTY_FUNCTION__)); |
8679 | // Non-instruction pointers are not deleted, just skip them. |
8680 | if (E->getOpcode() == Instruction::GetElementPtr && |
8681 | !isa<GetElementPtrInst>(Scalar)) |
8682 | continue; |
8683 | |
8684 | Value *Vec = E->VectorizedValue; |
8685 | 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", 8685, __extension__ __PRETTY_FUNCTION__)); |
8686 | |
8687 | Value *Lane = Builder.getInt32(ExternalUse.Lane); |
8688 | auto ExtractAndExtendIfNeeded = [&](Value *Vec) { |
8689 | if (Scalar->getType() != Vec->getType()) { |
8690 | Value *Ex; |
8691 | // "Reuse" the existing extract to improve final codegen. |
8692 | if (auto *ES = dyn_cast<ExtractElementInst>(Scalar)) { |
8693 | Ex = Builder.CreateExtractElement(ES->getOperand(0), |
8694 | ES->getOperand(1)); |
8695 | } else { |
8696 | Ex = Builder.CreateExtractElement(Vec, Lane); |
8697 | } |
8698 | // If necessary, sign-extend or zero-extend ScalarRoot |
8699 | // to the larger type. |
8700 | if (!MinBWs.count(ScalarRoot)) |
8701 | return Ex; |
8702 | if (MinBWs[ScalarRoot].second) |
8703 | return Builder.CreateSExt(Ex, Scalar->getType()); |
8704 | return Builder.CreateZExt(Ex, Scalar->getType()); |
8705 | } |
8706 | 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", 8708, __extension__ __PRETTY_FUNCTION__)) |
8707 | 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", 8708, __extension__ __PRETTY_FUNCTION__)) |
8708 | "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", 8708, __extension__ __PRETTY_FUNCTION__)); |
8709 | auto *IE = cast<InsertElementInst>(Scalar); |
8710 | VectorToInsertElement.try_emplace(Vec, IE); |
8711 | return Vec; |
8712 | }; |
8713 | // If User == nullptr, the Scalar is used as extra arg. Generate |
8714 | // ExtractElement instruction and update the record for this scalar in |
8715 | // ExternallyUsedValues. |
8716 | if (!User) { |
8717 | 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", 8719, __extension__ __PRETTY_FUNCTION__)) |
8718 | "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", 8719, __extension__ __PRETTY_FUNCTION__)) |
8719 | "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", 8719, __extension__ __PRETTY_FUNCTION__)); |
8720 | if (auto *VecI = dyn_cast<Instruction>(Vec)) { |
8721 | Builder.SetInsertPoint(VecI->getParent(), |
8722 | std::next(VecI->getIterator())); |
8723 | } else { |
8724 | Builder.SetInsertPoint(&F->getEntryBlock().front()); |
8725 | } |
8726 | Value *NewInst = ExtractAndExtendIfNeeded(Vec); |
8727 | CSEBlocks.insert(cast<Instruction>(Scalar)->getParent()); |
8728 | auto &NewInstLocs = ExternallyUsedValues[NewInst]; |
8729 | auto It = ExternallyUsedValues.find(Scalar); |
8730 | assert(It != ExternallyUsedValues.end() &&(static_cast <bool> (It != ExternallyUsedValues.end() && "Externally used scalar is not found in ExternallyUsedValues" ) ? void (0) : __assert_fail ("It != ExternallyUsedValues.end() && \"Externally used scalar is not found in ExternallyUsedValues\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8731, __extension__ __PRETTY_FUNCTION__)) |
8731 | "Externally used scalar is not found in ExternallyUsedValues")(static_cast <bool> (It != ExternallyUsedValues.end() && "Externally used scalar is not found in ExternallyUsedValues" ) ? void (0) : __assert_fail ("It != ExternallyUsedValues.end() && \"Externally used scalar is not found in ExternallyUsedValues\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8731, __extension__ __PRETTY_FUNCTION__)); |
8732 | NewInstLocs.append(It->second); |
8733 | ExternallyUsedValues.erase(Scalar); |
8734 | // Required to update internally referenced instructions. |
8735 | Scalar->replaceAllUsesWith(NewInst); |
8736 | continue; |
8737 | } |
8738 | |
8739 | if (auto *VU = dyn_cast<InsertElementInst>(User)) { |
8740 | // Skip if the scalar is another vector op or Vec is not an instruction. |
8741 | if (!Scalar->getType()->isVectorTy() && isa<Instruction>(Vec)) { |
8742 | if (auto *FTy = dyn_cast<FixedVectorType>(User->getType())) { |
8743 | Optional<unsigned> InsertIdx = getInsertIndex(VU); |
8744 | if (InsertIdx) { |
8745 | // Need to use original vector, if the root is truncated. |
8746 | if (MinBWs.count(Scalar) && |
8747 | VectorizableTree[0]->VectorizedValue == Vec) |
8748 | Vec = VectorRoot; |
8749 | auto *It = |
8750 | find_if(ShuffledInserts, [VU](const ShuffledInsertData &Data) { |
8751 | // Checks if 2 insertelements are from the same buildvector. |
8752 | InsertElementInst *VecInsert = Data.InsertElements.front(); |
8753 | return areTwoInsertFromSameBuildVector( |
8754 | VU, VecInsert, |
8755 | [](InsertElementInst *II) { return II->getOperand(0); }); |
8756 | }); |
8757 | unsigned Idx = *InsertIdx; |
8758 | if (It == ShuffledInserts.end()) { |
8759 | (void)ShuffledInserts.emplace_back(); |
8760 | It = std::next(ShuffledInserts.begin(), |
8761 | ShuffledInserts.size() - 1); |
8762 | SmallVectorImpl<int> &Mask = It->ValueMasks[Vec]; |
8763 | if (Mask.empty()) |
8764 | Mask.assign(FTy->getNumElements(), UndefMaskElem); |
8765 | // Find the insertvector, vectorized in tree, if any. |
8766 | Value *Base = VU; |
8767 | while (auto *IEBase = dyn_cast<InsertElementInst>(Base)) { |
8768 | if (IEBase != User && |
8769 | (!IEBase->hasOneUse() || |
8770 | getInsertIndex(IEBase).value_or(Idx) == Idx)) |
8771 | break; |
8772 | // Build the mask for the vectorized insertelement instructions. |
8773 | if (const TreeEntry *E = getTreeEntry(IEBase)) { |
8774 | do { |
8775 | IEBase = cast<InsertElementInst>(Base); |
8776 | int IEIdx = *getInsertIndex(IEBase); |
8777 | assert(Mask[Idx] == UndefMaskElem &&(static_cast <bool> (Mask[Idx] == UndefMaskElem && "InsertElementInstruction used already.") ? void (0) : __assert_fail ("Mask[Idx] == UndefMaskElem && \"InsertElementInstruction used already.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8778, __extension__ __PRETTY_FUNCTION__)) |
8778 | "InsertElementInstruction used already.")(static_cast <bool> (Mask[Idx] == UndefMaskElem && "InsertElementInstruction used already.") ? void (0) : __assert_fail ("Mask[Idx] == UndefMaskElem && \"InsertElementInstruction used already.\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8778, __extension__ __PRETTY_FUNCTION__)); |
8779 | Mask[IEIdx] = IEIdx; |
8780 | Base = IEBase->getOperand(0); |
8781 | } while (E == getTreeEntry(Base)); |
8782 | break; |
8783 | } |
8784 | Base = cast<InsertElementInst>(Base)->getOperand(0); |
8785 | // After the vectorization the def-use chain has changed, need |
8786 | // to look through original insertelement instructions, if they |
8787 | // get replaced by vector instructions. |
8788 | auto It = VectorToInsertElement.find(Base); |
8789 | if (It != VectorToInsertElement.end()) |
8790 | Base = It->second; |
8791 | } |
8792 | } |
8793 | SmallVectorImpl<int> &Mask = It->ValueMasks[Vec]; |
8794 | if (Mask.empty()) |
8795 | Mask.assign(FTy->getNumElements(), UndefMaskElem); |
8796 | Mask[Idx] = ExternalUse.Lane; |
8797 | It->InsertElements.push_back(cast<InsertElementInst>(User)); |
8798 | continue; |
8799 | } |
8800 | } |
8801 | } |
8802 | } |
8803 | |
8804 | // Generate extracts for out-of-tree users. |
8805 | // Find the insertion point for the extractelement lane. |
8806 | if (auto *VecI = dyn_cast<Instruction>(Vec)) { |
8807 | if (PHINode *PH = dyn_cast<PHINode>(User)) { |
8808 | for (int i = 0, e = PH->getNumIncomingValues(); i != e; ++i) { |
8809 | if (PH->getIncomingValue(i) == Scalar) { |
8810 | Instruction *IncomingTerminator = |
8811 | PH->getIncomingBlock(i)->getTerminator(); |
8812 | if (isa<CatchSwitchInst>(IncomingTerminator)) { |
8813 | Builder.SetInsertPoint(VecI->getParent(), |
8814 | std::next(VecI->getIterator())); |
8815 | } else { |
8816 | Builder.SetInsertPoint(PH->getIncomingBlock(i)->getTerminator()); |
8817 | } |
8818 | Value *NewInst = ExtractAndExtendIfNeeded(Vec); |
8819 | CSEBlocks.insert(PH->getIncomingBlock(i)); |
8820 | PH->setOperand(i, NewInst); |
8821 | } |
8822 | } |
8823 | } else { |
8824 | Builder.SetInsertPoint(cast<Instruction>(User)); |
8825 | Value *NewInst = ExtractAndExtendIfNeeded(Vec); |
8826 | CSEBlocks.insert(cast<Instruction>(User)->getParent()); |
8827 | User->replaceUsesOfWith(Scalar, NewInst); |
8828 | } |
8829 | } else { |
8830 | Builder.SetInsertPoint(&F->getEntryBlock().front()); |
8831 | Value *NewInst = ExtractAndExtendIfNeeded(Vec); |
8832 | CSEBlocks.insert(&F->getEntryBlock()); |
8833 | User->replaceUsesOfWith(Scalar, NewInst); |
8834 | } |
8835 | |
8836 | LLVM_DEBUG(dbgs() << "SLP: Replaced:" << *User << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Replaced:" << *User << ".\n"; } } while (false); |
8837 | } |
8838 | |
8839 | // Checks if the mask is an identity mask. |
8840 | auto &&IsIdentityMask = [](ArrayRef<int> Mask, FixedVectorType *VecTy) { |
8841 | int Limit = Mask.size(); |
8842 | return VecTy->getNumElements() == Mask.size() && |
8843 | all_of(Mask, [Limit](int Idx) { return Idx < Limit; }) && |
8844 | ShuffleVectorInst::isIdentityMask(Mask); |
8845 | }; |
8846 | // Tries to combine 2 different masks into single one. |
8847 | auto &&CombineMasks = [](SmallVectorImpl<int> &Mask, ArrayRef<int> ExtMask) { |
8848 | SmallVector<int> NewMask(ExtMask.size(), UndefMaskElem); |
8849 | for (int I = 0, Sz = ExtMask.size(); I < Sz; ++I) { |
8850 | if (ExtMask[I] == UndefMaskElem) |
8851 | continue; |
8852 | NewMask[I] = Mask[ExtMask[I]]; |
8853 | } |
8854 | Mask.swap(NewMask); |
8855 | }; |
8856 | // Peek through shuffles, trying to simplify the final shuffle code. |
8857 | auto &&PeekThroughShuffles = |
8858 | [&IsIdentityMask, &CombineMasks](Value *&V, SmallVectorImpl<int> &Mask, |
8859 | bool CheckForLengthChange = false) { |
8860 | while (auto *SV = dyn_cast<ShuffleVectorInst>(V)) { |
8861 | // Exit if not a fixed vector type or changing size shuffle. |
8862 | if (!isa<FixedVectorType>(SV->getType()) || |
8863 | (CheckForLengthChange && SV->changesLength())) |
8864 | break; |
8865 | // Exit if the identity or broadcast mask is found. |
8866 | if (IsIdentityMask(Mask, cast<FixedVectorType>(SV->getType())) || |
8867 | SV->isZeroEltSplat()) |
8868 | break; |
8869 | bool IsOp1Undef = isUndefVector(SV->getOperand(0), Mask); |
8870 | bool IsOp2Undef = isUndefVector(SV->getOperand(1), Mask); |
8871 | if (!IsOp1Undef && !IsOp2Undef) |
8872 | break; |
8873 | SmallVector<int> ShuffleMask(SV->getShuffleMask().begin(), |
8874 | SV->getShuffleMask().end()); |
8875 | CombineMasks(ShuffleMask, Mask); |
8876 | Mask.swap(ShuffleMask); |
8877 | if (IsOp2Undef) |
8878 | V = SV->getOperand(0); |
8879 | else |
8880 | V = SV->getOperand(1); |
8881 | } |
8882 | }; |
8883 | // Smart shuffle instruction emission, walks through shuffles trees and |
8884 | // tries to find the best matching vector for the actual shuffle |
8885 | // instruction. |
8886 | auto &&CreateShuffle = [this, &IsIdentityMask, &PeekThroughShuffles, |
8887 | &CombineMasks](Value *V1, Value *V2, |
8888 | ArrayRef<int> Mask) -> Value * { |
8889 | 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", 8889, __extension__ __PRETTY_FUNCTION__)); |
8890 | if (V2 && !isUndefVector(V2, Mask)) { |
8891 | // Peek through shuffles. |
8892 | Value *Op1 = V1; |
8893 | Value *Op2 = V2; |
8894 | int VF = |
8895 | cast<VectorType>(V1->getType())->getElementCount().getKnownMinValue(); |
8896 | SmallVector<int> CombinedMask1(Mask.size(), UndefMaskElem); |
8897 | SmallVector<int> CombinedMask2(Mask.size(), UndefMaskElem); |
8898 | for (int I = 0, E = Mask.size(); I < E; ++I) { |
8899 | if (Mask[I] < VF) |
8900 | CombinedMask1[I] = Mask[I]; |
8901 | else |
8902 | CombinedMask2[I] = Mask[I] - VF; |
8903 | } |
8904 | Value *PrevOp1; |
8905 | Value *PrevOp2; |
8906 | do { |
8907 | PrevOp1 = Op1; |
8908 | PrevOp2 = Op2; |
8909 | PeekThroughShuffles(Op1, CombinedMask1, /*CheckForLengthChange=*/true); |
8910 | PeekThroughShuffles(Op2, CombinedMask2, /*CheckForLengthChange=*/true); |
8911 | // Check if we have 2 resizing shuffles - need to peek through operands |
8912 | // again. |
8913 | if (auto *SV1 = dyn_cast<ShuffleVectorInst>(Op1)) |
8914 | if (auto *SV2 = dyn_cast<ShuffleVectorInst>(Op2)) |
8915 | if (SV1->getOperand(0)->getType() == |
8916 | SV2->getOperand(0)->getType() && |
8917 | SV1->getOperand(0)->getType() != SV1->getType() && |
8918 | isUndefVector(SV1->getOperand(1), CombinedMask1) && |
8919 | isUndefVector(SV2->getOperand(1), CombinedMask2)) { |
8920 | Op1 = SV1->getOperand(0); |
8921 | Op2 = SV2->getOperand(0); |
8922 | SmallVector<int> ShuffleMask1(SV1->getShuffleMask().begin(), |
8923 | SV1->getShuffleMask().end()); |
8924 | CombineMasks(ShuffleMask1, CombinedMask1); |
8925 | CombinedMask1.swap(ShuffleMask1); |
8926 | SmallVector<int> ShuffleMask2(SV2->getShuffleMask().begin(), |
8927 | SV2->getShuffleMask().end()); |
8928 | CombineMasks(ShuffleMask2, CombinedMask2); |
8929 | CombinedMask2.swap(ShuffleMask2); |
8930 | } |
8931 | } while (PrevOp1 != Op1 || PrevOp2 != Op2); |
8932 | VF = cast<VectorType>(Op1->getType()) |
8933 | ->getElementCount() |
8934 | .getKnownMinValue(); |
8935 | for (int I = 0, E = Mask.size(); I < E; ++I) { |
8936 | if (CombinedMask2[I] != UndefMaskElem) { |
8937 | assert(CombinedMask1[I] == UndefMaskElem &&(static_cast <bool> (CombinedMask1[I] == UndefMaskElem && "Expected undefined mask element") ? void (0) : __assert_fail ("CombinedMask1[I] == UndefMaskElem && \"Expected undefined mask element\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8938, __extension__ __PRETTY_FUNCTION__)) |
8938 | "Expected undefined mask element")(static_cast <bool> (CombinedMask1[I] == UndefMaskElem && "Expected undefined mask element") ? void (0) : __assert_fail ("CombinedMask1[I] == UndefMaskElem && \"Expected undefined mask element\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 8938, __extension__ __PRETTY_FUNCTION__)); |
8939 | CombinedMask1[I] = CombinedMask2[I] + (Op1 == Op2 ? 0 : VF); |
8940 | } |
8941 | } |
8942 | Value *Vec = Builder.CreateShuffleVector( |
8943 | Op1, Op1 == Op2 ? PoisonValue::get(Op1->getType()) : Op2, |
8944 | CombinedMask1); |
8945 | if (auto *I = dyn_cast<Instruction>(Vec)) { |
8946 | GatherShuffleSeq.insert(I); |
8947 | CSEBlocks.insert(I->getParent()); |
8948 | } |
8949 | return Vec; |
8950 | } |
8951 | if (isa<PoisonValue>(V1)) |
8952 | return PoisonValue::get(FixedVectorType::get( |
8953 | cast<VectorType>(V1->getType())->getElementType(), Mask.size())); |
8954 | Value *Op = V1; |
8955 | SmallVector<int> CombinedMask(Mask); |
8956 | PeekThroughShuffles(Op, CombinedMask); |
8957 | if (!isa<FixedVectorType>(Op->getType()) || |
8958 | !IsIdentityMask(CombinedMask, cast<FixedVectorType>(Op->getType()))) { |
8959 | Value *Vec = Builder.CreateShuffleVector(Op, CombinedMask); |
8960 | if (auto *I = dyn_cast<Instruction>(Vec)) { |
8961 | GatherShuffleSeq.insert(I); |
8962 | CSEBlocks.insert(I->getParent()); |
8963 | } |
8964 | return Vec; |
8965 | } |
8966 | return Op; |
8967 | }; |
8968 | |
8969 | auto &&ResizeToVF = [&CreateShuffle](Value *Vec, ArrayRef<int> Mask, |
8970 | bool ForSingleMask) { |
8971 | unsigned VF = Mask.size(); |
8972 | unsigned VecVF = cast<FixedVectorType>(Vec->getType())->getNumElements(); |
8973 | if (VF != VecVF) { |
8974 | if (any_of(Mask, [VF](int Idx) { return Idx >= static_cast<int>(VF); })) { |
8975 | Vec = CreateShuffle(Vec, nullptr, Mask); |
8976 | return std::make_pair(Vec, true); |
8977 | } |
8978 | if (!ForSingleMask) { |
8979 | SmallVector<int> ResizeMask(VF, UndefMaskElem); |
8980 | for (unsigned I = 0; I < VF; ++I) { |
8981 | if (Mask[I] != UndefMaskElem) |
8982 | ResizeMask[Mask[I]] = Mask[I]; |
8983 | } |
8984 | Vec = CreateShuffle(Vec, nullptr, ResizeMask); |
8985 | } |
8986 | } |
8987 | |
8988 | return std::make_pair(Vec, false); |
8989 | }; |
8990 | // Perform shuffling of the vectorize tree entries for better handling of |
8991 | // external extracts. |
8992 | for (int I = 0, E = ShuffledInserts.size(); I < E; ++I) { |
8993 | // Find the first and the last instruction in the list of insertelements. |
8994 | sort(ShuffledInserts[I].InsertElements, isFirstInsertElement); |
8995 | InsertElementInst *FirstInsert = ShuffledInserts[I].InsertElements.front(); |
8996 | InsertElementInst *LastInsert = ShuffledInserts[I].InsertElements.back(); |
8997 | Builder.SetInsertPoint(LastInsert); |
8998 | auto Vector = ShuffledInserts[I].ValueMasks.takeVector(); |
8999 | Value *NewInst = performExtractsShuffleAction<Value>( |
9000 | makeMutableArrayRef(Vector.data(), Vector.size()), |
9001 | FirstInsert->getOperand(0), |
9002 | [](Value *Vec) { |
9003 | return cast<VectorType>(Vec->getType()) |
9004 | ->getElementCount() |
9005 | .getKnownMinValue(); |
9006 | }, |
9007 | ResizeToVF, |
9008 | [FirstInsert, &CreateShuffle](ArrayRef<int> Mask, |
9009 | ArrayRef<Value *> Vals) { |
9010 | 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", 9011, __extension__ __PRETTY_FUNCTION__)) |
9011 | "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", 9011, __extension__ __PRETTY_FUNCTION__)); |
9012 | if (Vals.size() == 1) { |
9013 | // Do not create shuffle if the mask is a simple identity |
9014 | // non-resizing mask. |
9015 | if (Mask.size() != cast<FixedVectorType>(Vals.front()->getType()) |
9016 | ->getNumElements() || |
9017 | !ShuffleVectorInst::isIdentityMask(Mask)) |
9018 | return CreateShuffle(Vals.front(), nullptr, Mask); |
9019 | return Vals.front(); |
9020 | } |
9021 | return CreateShuffle(Vals.front() ? Vals.front() |
9022 | : FirstInsert->getOperand(0), |
9023 | Vals.back(), Mask); |
9024 | }); |
9025 | auto It = ShuffledInserts[I].InsertElements.rbegin(); |
9026 | // Rebuild buildvector chain. |
9027 | InsertElementInst *II = nullptr; |
9028 | if (It != ShuffledInserts[I].InsertElements.rend()) |
9029 | II = *It; |
9030 | SmallVector<Instruction *> Inserts; |
9031 | while (It != ShuffledInserts[I].InsertElements.rend()) { |
9032 | 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", 9032, __extension__ __PRETTY_FUNCTION__)); |
9033 | if (*It == II) |
9034 | ++It; |
9035 | else |
9036 | Inserts.push_back(cast<Instruction>(II)); |
9037 | II = dyn_cast<InsertElementInst>(II->getOperand(0)); |
9038 | } |
9039 | for (Instruction *II : reverse(Inserts)) { |
9040 | II->replaceUsesOfWith(II->getOperand(0), NewInst); |
9041 | if (auto *NewI = dyn_cast<Instruction>(NewInst)) |
9042 | if (II->getParent() == NewI->getParent() && II->comesBefore(NewI)) |
9043 | II->moveAfter(NewI); |
9044 | NewInst = II; |
9045 | } |
9046 | LastInsert->replaceAllUsesWith(NewInst); |
9047 | for (InsertElementInst *IE : reverse(ShuffledInserts[I].InsertElements)) { |
9048 | IE->replaceUsesOfWith(IE->getOperand(0), |
9049 | PoisonValue::get(IE->getOperand(0)->getType())); |
9050 | IE->replaceUsesOfWith(IE->getOperand(1), |
9051 | PoisonValue::get(IE->getOperand(1)->getType())); |
9052 | eraseInstruction(IE); |
9053 | } |
9054 | CSEBlocks.insert(LastInsert->getParent()); |
9055 | } |
9056 | |
9057 | // For each vectorized value: |
9058 | for (auto &TEPtr : VectorizableTree) { |
9059 | TreeEntry *Entry = TEPtr.get(); |
9060 | |
9061 | // No need to handle users of gathered values. |
9062 | if (Entry->State == TreeEntry::NeedToGather) |
9063 | continue; |
9064 | |
9065 | 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", 9065, __extension__ __PRETTY_FUNCTION__)); |
9066 | |
9067 | // For each lane: |
9068 | for (int Lane = 0, LE = Entry->Scalars.size(); Lane != LE; ++Lane) { |
9069 | Value *Scalar = Entry->Scalars[Lane]; |
9070 | |
9071 | if (Entry->getOpcode() == Instruction::GetElementPtr && |
9072 | !isa<GetElementPtrInst>(Scalar)) |
9073 | continue; |
9074 | #ifndef NDEBUG |
9075 | Type *Ty = Scalar->getType(); |
9076 | if (!Ty->isVoidTy()) { |
9077 | for (User *U : Scalar->users()) { |
9078 | LLVM_DEBUG(dbgs() << "SLP: \tvalidating user:" << *U << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: \tvalidating user:" << *U << ".\n"; } } while (false); |
9079 | |
9080 | // It is legal to delete users in the ignorelist. |
9081 | 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", 9085, __extension__ __PRETTY_FUNCTION__)) |
9082 | (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", 9085, __extension__ __PRETTY_FUNCTION__)) |
9083 | (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", 9085, __extension__ __PRETTY_FUNCTION__)) |
9084 | 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", 9085, __extension__ __PRETTY_FUNCTION__)) |
9085 | "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", 9085, __extension__ __PRETTY_FUNCTION__)); |
9086 | } |
9087 | } |
9088 | #endif |
9089 | LLVM_DEBUG(dbgs() << "SLP: \tErasing scalar:" << *Scalar << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: \tErasing scalar:" << * Scalar << ".\n"; } } while (false); |
9090 | eraseInstruction(cast<Instruction>(Scalar)); |
9091 | } |
9092 | } |
9093 | |
9094 | Builder.ClearInsertionPoint(); |
9095 | InstrElementSize.clear(); |
9096 | |
9097 | return VectorizableTree[0]->VectorizedValue; |
9098 | } |
9099 | |
9100 | void BoUpSLP::optimizeGatherSequence() { |
9101 | LLVM_DEBUG(dbgs() << "SLP: Optimizing " << GatherShuffleSeq.size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Optimizing " << GatherShuffleSeq .size() << " gather sequences instructions.\n"; } } while (false) |
9102 | << " gather sequences instructions.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Optimizing " << GatherShuffleSeq .size() << " gather sequences instructions.\n"; } } while (false); |
9103 | // LICM InsertElementInst sequences. |
9104 | for (Instruction *I : GatherShuffleSeq) { |
9105 | if (isDeleted(I)) |
9106 | continue; |
9107 | |
9108 | // Check if this block is inside a loop. |
9109 | Loop *L = LI->getLoopFor(I->getParent()); |
9110 | if (!L) |
9111 | continue; |
9112 | |
9113 | // Check if it has a preheader. |
9114 | BasicBlock *PreHeader = L->getLoopPreheader(); |
9115 | if (!PreHeader) |
9116 | continue; |
9117 | |
9118 | // If the vector or the element that we insert into it are |
9119 | // instructions that are defined in this basic block then we can't |
9120 | // hoist this instruction. |
9121 | if (any_of(I->operands(), [L](Value *V) { |
9122 | auto *OpI = dyn_cast<Instruction>(V); |
9123 | return OpI && L->contains(OpI); |
9124 | })) |
9125 | continue; |
9126 | |
9127 | // We can hoist this instruction. Move it to the pre-header. |
9128 | I->moveBefore(PreHeader->getTerminator()); |
9129 | CSEBlocks.insert(PreHeader); |
9130 | } |
9131 | |
9132 | // Make a list of all reachable blocks in our CSE queue. |
9133 | SmallVector<const DomTreeNode *, 8> CSEWorkList; |
9134 | CSEWorkList.reserve(CSEBlocks.size()); |
9135 | for (BasicBlock *BB : CSEBlocks) |
9136 | if (DomTreeNode *N = DT->getNode(BB)) { |
9137 | assert(DT->isReachableFromEntry(N))(static_cast <bool> (DT->isReachableFromEntry(N)) ? void (0) : __assert_fail ("DT->isReachableFromEntry(N)", "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 9137, __extension__ __PRETTY_FUNCTION__)); |
9138 | CSEWorkList.push_back(N); |
9139 | } |
9140 | |
9141 | // Sort blocks by domination. This ensures we visit a block after all blocks |
9142 | // dominating it are visited. |
9143 | llvm::sort(CSEWorkList, [](const DomTreeNode *A, const DomTreeNode *B) { |
9144 | 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", 9145, __extension__ __PRETTY_FUNCTION__)) |
9145 | "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", 9145, __extension__ __PRETTY_FUNCTION__)); |
9146 | return A->getDFSNumIn() < B->getDFSNumIn(); |
9147 | }); |
9148 | |
9149 | // Less defined shuffles can be replaced by the more defined copies. |
9150 | // Between two shuffles one is less defined if it has the same vector operands |
9151 | // and its mask indeces are the same as in the first one or undefs. E.g. |
9152 | // shuffle %0, poison, <0, 0, 0, undef> is less defined than shuffle %0, |
9153 | // poison, <0, 0, 0, 0>. |
9154 | auto &&IsIdenticalOrLessDefined = [this](Instruction *I1, Instruction *I2, |
9155 | SmallVectorImpl<int> &NewMask) { |
9156 | if (I1->getType() != I2->getType()) |
9157 | return false; |
9158 | auto *SI1 = dyn_cast<ShuffleVectorInst>(I1); |
9159 | auto *SI2 = dyn_cast<ShuffleVectorInst>(I2); |
9160 | if (!SI1 || !SI2) |
9161 | return I1->isIdenticalTo(I2); |
9162 | if (SI1->isIdenticalTo(SI2)) |
9163 | return true; |
9164 | for (int I = 0, E = SI1->getNumOperands(); I < E; ++I) |
9165 | if (SI1->getOperand(I) != SI2->getOperand(I)) |
9166 | return false; |
9167 | // Check if the second instruction is more defined than the first one. |
9168 | NewMask.assign(SI2->getShuffleMask().begin(), SI2->getShuffleMask().end()); |
9169 | ArrayRef<int> SM1 = SI1->getShuffleMask(); |
9170 | // Count trailing undefs in the mask to check the final number of used |
9171 | // registers. |
9172 | unsigned LastUndefsCnt = 0; |
9173 | for (int I = 0, E = NewMask.size(); I < E; ++I) { |
9174 | if (SM1[I] == UndefMaskElem) |
9175 | ++LastUndefsCnt; |
9176 | else |
9177 | LastUndefsCnt = 0; |
9178 | if (NewMask[I] != UndefMaskElem && SM1[I] != UndefMaskElem && |
9179 | NewMask[I] != SM1[I]) |
9180 | return false; |
9181 | if (NewMask[I] == UndefMaskElem) |
9182 | NewMask[I] = SM1[I]; |
9183 | } |
9184 | // Check if the last undefs actually change the final number of used vector |
9185 | // registers. |
9186 | return SM1.size() - LastUndefsCnt > 1 && |
9187 | TTI->getNumberOfParts(SI1->getType()) == |
9188 | TTI->getNumberOfParts( |
9189 | FixedVectorType::get(SI1->getType()->getElementType(), |
9190 | SM1.size() - LastUndefsCnt)); |
9191 | }; |
9192 | // Perform O(N^2) search over the gather/shuffle sequences and merge identical |
9193 | // instructions. TODO: We can further optimize this scan if we split the |
9194 | // instructions into different buckets based on the insert lane. |
9195 | SmallVector<Instruction *, 16> Visited; |
9196 | for (auto I = CSEWorkList.begin(), E = CSEWorkList.end(); I != E; ++I) { |
9197 | 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", 9199, __extension__ __PRETTY_FUNCTION__)) |
9198 | (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", 9199, __extension__ __PRETTY_FUNCTION__)) |
9199 | "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", 9199, __extension__ __PRETTY_FUNCTION__)); |
9200 | BasicBlock *BB = (*I)->getBlock(); |
9201 | // For all instructions in blocks containing gather sequences: |
9202 | for (Instruction &In : llvm::make_early_inc_range(*BB)) { |
9203 | if (isDeleted(&In)) |
9204 | continue; |
9205 | if (!isa<InsertElementInst, ExtractElementInst, ShuffleVectorInst>(&In) && |
9206 | !GatherShuffleSeq.contains(&In)) |
9207 | continue; |
9208 | |
9209 | // Check if we can replace this instruction with any of the |
9210 | // visited instructions. |
9211 | bool Replaced = false; |
9212 | for (Instruction *&V : Visited) { |
9213 | SmallVector<int> NewMask; |
9214 | if (IsIdenticalOrLessDefined(&In, V, NewMask) && |
9215 | DT->dominates(V->getParent(), In.getParent())) { |
9216 | In.replaceAllUsesWith(V); |
9217 | eraseInstruction(&In); |
9218 | if (auto *SI = dyn_cast<ShuffleVectorInst>(V)) |
9219 | if (!NewMask.empty()) |
9220 | SI->setShuffleMask(NewMask); |
9221 | Replaced = true; |
9222 | break; |
9223 | } |
9224 | if (isa<ShuffleVectorInst>(In) && isa<ShuffleVectorInst>(V) && |
9225 | GatherShuffleSeq.contains(V) && |
9226 | IsIdenticalOrLessDefined(V, &In, NewMask) && |
9227 | DT->dominates(In.getParent(), V->getParent())) { |
9228 | In.moveAfter(V); |
9229 | V->replaceAllUsesWith(&In); |
9230 | eraseInstruction(V); |
9231 | if (auto *SI = dyn_cast<ShuffleVectorInst>(&In)) |
9232 | if (!NewMask.empty()) |
9233 | SI->setShuffleMask(NewMask); |
9234 | V = &In; |
9235 | Replaced = true; |
9236 | break; |
9237 | } |
9238 | } |
9239 | if (!Replaced) { |
9240 | 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", 9240, __extension__ __PRETTY_FUNCTION__)); |
9241 | Visited.push_back(&In); |
9242 | } |
9243 | } |
9244 | } |
9245 | CSEBlocks.clear(); |
9246 | GatherShuffleSeq.clear(); |
9247 | } |
9248 | |
9249 | BoUpSLP::ScheduleData * |
9250 | BoUpSLP::BlockScheduling::buildBundle(ArrayRef<Value *> VL) { |
9251 | ScheduleData *Bundle = nullptr; |
9252 | ScheduleData *PrevInBundle = nullptr; |
9253 | for (Value *V : VL) { |
9254 | if (doesNotNeedToBeScheduled(V)) |
9255 | continue; |
9256 | ScheduleData *BundleMember = getScheduleData(V); |
9257 | 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", 9259, __extension__ __PRETTY_FUNCTION__)) |
9258 | "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", 9259, __extension__ __PRETTY_FUNCTION__)) |
9259 | "(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", 9259, __extension__ __PRETTY_FUNCTION__)); |
9260 | 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", 9261, __extension__ __PRETTY_FUNCTION__)) |
9261 | "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", 9261, __extension__ __PRETTY_FUNCTION__)); |
9262 | if (PrevInBundle) { |
9263 | PrevInBundle->NextInBundle = BundleMember; |
9264 | } else { |
9265 | Bundle = BundleMember; |
9266 | } |
9267 | |
9268 | // Group the instructions to a bundle. |
9269 | BundleMember->FirstInBundle = Bundle; |
9270 | PrevInBundle = BundleMember; |
9271 | } |
9272 | 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", 9272, __extension__ __PRETTY_FUNCTION__)); |
9273 | return Bundle; |
9274 | } |
9275 | |
9276 | // Groups the instructions to a bundle (which is then a single scheduling entity) |
9277 | // and schedules instructions until the bundle gets ready. |
9278 | Optional<BoUpSLP::ScheduleData *> |
9279 | BoUpSLP::BlockScheduling::tryScheduleBundle(ArrayRef<Value *> VL, BoUpSLP *SLP, |
9280 | const InstructionsState &S) { |
9281 | // No need to schedule PHIs, insertelement, extractelement and extractvalue |
9282 | // instructions. |
9283 | if (isa<PHINode>(S.OpValue) || isVectorLikeInstWithConstOps(S.OpValue) || |
9284 | doesNotNeedToSchedule(VL)) |
9285 | return nullptr; |
9286 | |
9287 | // Initialize the instruction bundle. |
9288 | Instruction *OldScheduleEnd = ScheduleEnd; |
9289 | LLVM_DEBUG(dbgs() << "SLP: bundle: " << *S.OpValue << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: bundle: " << *S.OpValue << "\n"; } } while (false); |
9290 | |
9291 | auto TryScheduleBundleImpl = [this, OldScheduleEnd, SLP](bool ReSchedule, |
9292 | ScheduleData *Bundle) { |
9293 | // The scheduling region got new instructions at the lower end (or it is a |
9294 | // new region for the first bundle). This makes it necessary to |
9295 | // recalculate all dependencies. |
9296 | // It is seldom that this needs to be done a second time after adding the |
9297 | // initial bundle to the region. |
9298 | if (ScheduleEnd != OldScheduleEnd) { |
9299 | for (auto *I = ScheduleStart; I != ScheduleEnd; I = I->getNextNode()) |
9300 | doForAllOpcodes(I, [](ScheduleData *SD) { SD->clearDependencies(); }); |
9301 | ReSchedule = true; |
9302 | } |
9303 | if (Bundle) { |
9304 | 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) |
9305 | << " 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); |
9306 | calculateDependencies(Bundle, /*InsertInReadyList=*/true, SLP); |
9307 | } |
9308 | |
9309 | if (ReSchedule) { |
9310 | resetSchedule(); |
9311 | initialFillReadyList(ReadyInsts); |
9312 | } |
9313 | |
9314 | // Now try to schedule the new bundle or (if no bundle) just calculate |
9315 | // dependencies. As soon as the bundle is "ready" it means that there are no |
9316 | // cyclic dependencies and we can schedule it. Note that's important that we |
9317 | // don't "schedule" the bundle yet (see cancelScheduling). |
9318 | while (((!Bundle && ReSchedule) || (Bundle && !Bundle->isReady())) && |
9319 | !ReadyInsts.empty()) { |
9320 | ScheduleData *Picked = ReadyInsts.pop_back_val(); |
9321 | 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", 9322, __extension__ __PRETTY_FUNCTION__)) |
9322 | "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", 9322, __extension__ __PRETTY_FUNCTION__)); |
9323 | schedule(Picked, ReadyInsts); |
9324 | } |
9325 | }; |
9326 | |
9327 | // Make sure that the scheduling region contains all |
9328 | // instructions of the bundle. |
9329 | for (Value *V : VL) { |
9330 | if (doesNotNeedToBeScheduled(V)) |
9331 | continue; |
9332 | if (!extendSchedulingRegion(V, S)) { |
9333 | // If the scheduling region got new instructions at the lower end (or it |
9334 | // is a new region for the first bundle). This makes it necessary to |
9335 | // recalculate all dependencies. |
9336 | // Otherwise the compiler may crash trying to incorrectly calculate |
9337 | // dependencies and emit instruction in the wrong order at the actual |
9338 | // scheduling. |
9339 | TryScheduleBundleImpl(/*ReSchedule=*/false, nullptr); |
9340 | return None; |
9341 | } |
9342 | } |
9343 | |
9344 | bool ReSchedule = false; |
9345 | for (Value *V : VL) { |
9346 | if (doesNotNeedToBeScheduled(V)) |
9347 | continue; |
9348 | ScheduleData *BundleMember = getScheduleData(V); |
9349 | 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", 9350, __extension__ __PRETTY_FUNCTION__)) |
9350 | "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", 9350, __extension__ __PRETTY_FUNCTION__)); |
9351 | |
9352 | // Make sure we don't leave the pieces of the bundle in the ready list when |
9353 | // whole bundle might not be ready. |
9354 | ReadyInsts.remove(BundleMember); |
9355 | |
9356 | if (!BundleMember->IsScheduled) |
9357 | continue; |
9358 | // A bundle member was scheduled as single instruction before and now |
9359 | // needs to be scheduled as part of the bundle. We just get rid of the |
9360 | // existing schedule. |
9361 | 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) |
9362 | << " was already scheduled\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: reset schedule because " << *BundleMember << " was already scheduled\n"; } } while (false); |
9363 | ReSchedule = true; |
9364 | } |
9365 | |
9366 | auto *Bundle = buildBundle(VL); |
9367 | TryScheduleBundleImpl(ReSchedule, Bundle); |
9368 | if (!Bundle->isReady()) { |
9369 | cancelScheduling(VL, S.OpValue); |
9370 | return None; |
9371 | } |
9372 | return Bundle; |
9373 | } |
9374 | |
9375 | void BoUpSLP::BlockScheduling::cancelScheduling(ArrayRef<Value *> VL, |
9376 | Value *OpValue) { |
9377 | if (isa<PHINode>(OpValue) || isVectorLikeInstWithConstOps(OpValue) || |
9378 | doesNotNeedToSchedule(VL)) |
9379 | return; |
9380 | |
9381 | if (doesNotNeedToBeScheduled(OpValue)) |
9382 | OpValue = *find_if_not(VL, doesNotNeedToBeScheduled); |
9383 | ScheduleData *Bundle = getScheduleData(OpValue); |
9384 | 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); |
9385 | 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", 9386, __extension__ __PRETTY_FUNCTION__)) |
9386 | "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", 9386, __extension__ __PRETTY_FUNCTION__)); |
9387 | 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", 9389, __extension__ __PRETTY_FUNCTION__)) |
9388 | (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", 9389, __extension__ __PRETTY_FUNCTION__)) |
9389 | "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", 9389, __extension__ __PRETTY_FUNCTION__)); |
9390 | |
9391 | // Remove the bundle from the ready list. |
9392 | if (Bundle->isReady()) |
9393 | ReadyInsts.remove(Bundle); |
9394 | |
9395 | // Un-bundle: make single instructions out of the bundle. |
9396 | ScheduleData *BundleMember = Bundle; |
9397 | while (BundleMember) { |
9398 | 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", 9398, __extension__ __PRETTY_FUNCTION__)); |
9399 | BundleMember->FirstInBundle = BundleMember; |
9400 | ScheduleData *Next = BundleMember->NextInBundle; |
9401 | BundleMember->NextInBundle = nullptr; |
9402 | BundleMember->TE = nullptr; |
9403 | if (BundleMember->unscheduledDepsInBundle() == 0) { |
9404 | ReadyInsts.insert(BundleMember); |
9405 | } |
9406 | BundleMember = Next; |
9407 | } |
9408 | } |
9409 | |
9410 | BoUpSLP::ScheduleData *BoUpSLP::BlockScheduling::allocateScheduleDataChunks() { |
9411 | // Allocate a new ScheduleData for the instruction. |
9412 | if (ChunkPos >= ChunkSize) { |
9413 | ScheduleDataChunks.push_back(std::make_unique<ScheduleData[]>(ChunkSize)); |
9414 | ChunkPos = 0; |
9415 | } |
9416 | return &(ScheduleDataChunks.back()[ChunkPos++]); |
9417 | } |
9418 | |
9419 | bool BoUpSLP::BlockScheduling::extendSchedulingRegion(Value *V, |
9420 | const InstructionsState &S) { |
9421 | if (getScheduleData(V, isOneOf(S, V))) |
9422 | return true; |
9423 | Instruction *I = dyn_cast<Instruction>(V); |
9424 | 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", 9424, __extension__ __PRETTY_FUNCTION__)); |
9425 | 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", 9428, __extension__ __PRETTY_FUNCTION__)) |
9426 | !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", 9428, __extension__ __PRETTY_FUNCTION__)) |
9427 | "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", 9428, __extension__ __PRETTY_FUNCTION__)) |
9428 | "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", 9428, __extension__ __PRETTY_FUNCTION__)); |
9429 | auto &&CheckScheduleForI = [this, &S](Instruction *I) -> bool { |
9430 | ScheduleData *ISD = getScheduleData(I); |
9431 | if (!ISD) |
9432 | return false; |
9433 | 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", 9434, __extension__ __PRETTY_FUNCTION__)) |
9434 | "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", 9434, __extension__ __PRETTY_FUNCTION__)); |
9435 | ScheduleData *SD = allocateScheduleDataChunks(); |
9436 | SD->Inst = I; |
9437 | SD->init(SchedulingRegionID, S.OpValue); |
9438 | ExtraScheduleDataMap[I][S.OpValue] = SD; |
9439 | return true; |
9440 | }; |
9441 | if (CheckScheduleForI(I)) |
9442 | return true; |
9443 | if (!ScheduleStart) { |
9444 | // It's the first instruction in the new region. |
9445 | initScheduleData(I, I->getNextNode(), nullptr, nullptr); |
9446 | ScheduleStart = I; |
9447 | ScheduleEnd = I->getNextNode(); |
9448 | if (isOneOf(S, I) != I) |
9449 | CheckScheduleForI(I); |
9450 | 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", 9450, __extension__ __PRETTY_FUNCTION__)); |
9451 | 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); |
9452 | return true; |
9453 | } |
9454 | // Search up and down at the same time, because we don't know if the new |
9455 | // instruction is above or below the existing scheduling region. |
9456 | BasicBlock::reverse_iterator UpIter = |
9457 | ++ScheduleStart->getIterator().getReverse(); |
9458 | BasicBlock::reverse_iterator UpperEnd = BB->rend(); |
9459 | BasicBlock::iterator DownIter = ScheduleEnd->getIterator(); |
9460 | BasicBlock::iterator LowerEnd = BB->end(); |
9461 | while (UpIter != UpperEnd && DownIter != LowerEnd && &*UpIter != I && |
9462 | &*DownIter != I) { |
9463 | if (++ScheduleRegionSize > ScheduleRegionSizeLimit) { |
9464 | 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); |
9465 | return false; |
9466 | } |
9467 | |
9468 | ++UpIter; |
9469 | ++DownIter; |
9470 | } |
9471 | if (DownIter == LowerEnd || (UpIter != UpperEnd && &*UpIter == I)) { |
9472 | 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", 9473, __extension__ __PRETTY_FUNCTION__)) |
9473 | "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", 9473, __extension__ __PRETTY_FUNCTION__)); |
9474 | initScheduleData(I, ScheduleStart, nullptr, FirstLoadStoreInRegion); |
9475 | ScheduleStart = I; |
9476 | if (isOneOf(S, I) != I) |
9477 | CheckScheduleForI(I); |
9478 | 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) |
9479 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: extend schedule region start to " << *I << "\n"; } } while (false); |
9480 | return true; |
9481 | } |
9482 | 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", 9484, __extension__ __PRETTY_FUNCTION__)) |
9483 | "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", 9484, __extension__ __PRETTY_FUNCTION__)) |
9484 | "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", 9484, __extension__ __PRETTY_FUNCTION__)); |
9485 | 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", 9486, __extension__ __PRETTY_FUNCTION__)) |
9486 | "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", 9486, __extension__ __PRETTY_FUNCTION__)); |
9487 | initScheduleData(ScheduleEnd, I->getNextNode(), LastLoadStoreInRegion, |
9488 | nullptr); |
9489 | ScheduleEnd = I->getNextNode(); |
9490 | if (isOneOf(S, I) != I) |
9491 | CheckScheduleForI(I); |
9492 | 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", 9492, __extension__ __PRETTY_FUNCTION__)); |
9493 | 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); |
9494 | return true; |
9495 | } |
9496 | |
9497 | void BoUpSLP::BlockScheduling::initScheduleData(Instruction *FromI, |
9498 | Instruction *ToI, |
9499 | ScheduleData *PrevLoadStore, |
9500 | ScheduleData *NextLoadStore) { |
9501 | ScheduleData *CurrentLoadStore = PrevLoadStore; |
9502 | for (Instruction *I = FromI; I != ToI; I = I->getNextNode()) { |
9503 | // No need to allocate data for non-schedulable instructions. |
9504 | if (doesNotNeedToBeScheduled(I)) |
9505 | continue; |
9506 | ScheduleData *SD = ScheduleDataMap.lookup(I); |
9507 | if (!SD) { |
9508 | SD = allocateScheduleDataChunks(); |
9509 | ScheduleDataMap[I] = SD; |
9510 | SD->Inst = I; |
9511 | } |
9512 | 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", 9513, __extension__ __PRETTY_FUNCTION__)) |
9513 | "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", 9513, __extension__ __PRETTY_FUNCTION__)); |
9514 | SD->init(SchedulingRegionID, I); |
9515 | |
9516 | if (I->mayReadOrWriteMemory() && |
9517 | (!isa<IntrinsicInst>(I) || |
9518 | (cast<IntrinsicInst>(I)->getIntrinsicID() != Intrinsic::sideeffect && |
9519 | cast<IntrinsicInst>(I)->getIntrinsicID() != |
9520 | Intrinsic::pseudoprobe))) { |
9521 | // Update the linked list of memory accessing instructions. |
9522 | if (CurrentLoadStore) { |
9523 | CurrentLoadStore->NextLoadStore = SD; |
9524 | } else { |
9525 | FirstLoadStoreInRegion = SD; |
9526 | } |
9527 | CurrentLoadStore = SD; |
9528 | } |
9529 | |
9530 | if (match(I, m_Intrinsic<Intrinsic::stacksave>()) || |
9531 | match(I, m_Intrinsic<Intrinsic::stackrestore>())) |
9532 | RegionHasStackSave = true; |
9533 | } |
9534 | if (NextLoadStore) { |
9535 | if (CurrentLoadStore) |
9536 | CurrentLoadStore->NextLoadStore = NextLoadStore; |
9537 | } else { |
9538 | LastLoadStoreInRegion = CurrentLoadStore; |
9539 | } |
9540 | } |
9541 | |
9542 | void BoUpSLP::BlockScheduling::calculateDependencies(ScheduleData *SD, |
9543 | bool InsertInReadyList, |
9544 | BoUpSLP *SLP) { |
9545 | assert(SD->isSchedulingEntity())(static_cast <bool> (SD->isSchedulingEntity()) ? void (0) : __assert_fail ("SD->isSchedulingEntity()", "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 9545, __extension__ __PRETTY_FUNCTION__)); |
9546 | |
9547 | SmallVector<ScheduleData *, 10> WorkList; |
9548 | WorkList.push_back(SD); |
9549 | |
9550 | while (!WorkList.empty()) { |
9551 | ScheduleData *SD = WorkList.pop_back_val(); |
9552 | for (ScheduleData *BundleMember = SD; BundleMember; |
9553 | BundleMember = BundleMember->NextInBundle) { |
9554 | assert(isInSchedulingRegion(BundleMember))(static_cast <bool> (isInSchedulingRegion(BundleMember) ) ? void (0) : __assert_fail ("isInSchedulingRegion(BundleMember)" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 9554, __extension__ __PRETTY_FUNCTION__)); |
9555 | if (BundleMember->hasValidDependencies()) |
9556 | continue; |
9557 | |
9558 | LLVM_DEBUG(dbgs() << "SLP: update deps of " << *BundleMemberdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: update deps of " << *BundleMember << "\n"; } } while (false) |
9559 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: update deps of " << *BundleMember << "\n"; } } while (false); |
9560 | BundleMember->Dependencies = 0; |
9561 | BundleMember->resetUnscheduledDeps(); |
9562 | |
9563 | // Handle def-use chain dependencies. |
9564 | if (BundleMember->OpValue != BundleMember->Inst) { |
9565 | if (ScheduleData *UseSD = getScheduleData(BundleMember->Inst)) { |
9566 | BundleMember->Dependencies++; |
9567 | ScheduleData *DestBundle = UseSD->FirstInBundle; |
9568 | if (!DestBundle->IsScheduled) |
9569 | BundleMember->incrementUnscheduledDeps(1); |
9570 | if (!DestBundle->hasValidDependencies()) |
9571 | WorkList.push_back(DestBundle); |
9572 | } |
9573 | } else { |
9574 | for (User *U : BundleMember->Inst->users()) { |
9575 | if (ScheduleData *UseSD = getScheduleData(cast<Instruction>(U))) { |
9576 | BundleMember->Dependencies++; |
9577 | ScheduleData *DestBundle = UseSD->FirstInBundle; |
9578 | if (!DestBundle->IsScheduled) |
9579 | BundleMember->incrementUnscheduledDeps(1); |
9580 | if (!DestBundle->hasValidDependencies()) |
9581 | WorkList.push_back(DestBundle); |
9582 | } |
9583 | } |
9584 | } |
9585 | |
9586 | auto makeControlDependent = [&](Instruction *I) { |
9587 | auto *DepDest = getScheduleData(I); |
9588 | 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", 9588, __extension__ __PRETTY_FUNCTION__)); |
9589 | DepDest->ControlDependencies.push_back(BundleMember); |
9590 | BundleMember->Dependencies++; |
9591 | ScheduleData *DestBundle = DepDest->FirstInBundle; |
9592 | if (!DestBundle->IsScheduled) |
9593 | BundleMember->incrementUnscheduledDeps(1); |
9594 | if (!DestBundle->hasValidDependencies()) |
9595 | WorkList.push_back(DestBundle); |
9596 | }; |
9597 | |
9598 | // Any instruction which isn't safe to speculate at the beginning of the |
9599 | // block is control dependend on any early exit or non-willreturn call |
9600 | // which proceeds it. |
9601 | if (!isGuaranteedToTransferExecutionToSuccessor(BundleMember->Inst)) { |
9602 | for (Instruction *I = BundleMember->Inst->getNextNode(); |
9603 | I != ScheduleEnd; I = I->getNextNode()) { |
9604 | if (isSafeToSpeculativelyExecute(I, &*BB->begin(), SLP->AC)) |
9605 | continue; |
9606 | |
9607 | // Add the dependency |
9608 | makeControlDependent(I); |
9609 | |
9610 | if (!isGuaranteedToTransferExecutionToSuccessor(I)) |
9611 | // Everything past here must be control dependent on I. |
9612 | break; |
9613 | } |
9614 | } |
9615 | |
9616 | if (RegionHasStackSave) { |
9617 | // If we have an inalloc alloca instruction, it needs to be scheduled |
9618 | // after any preceeding stacksave. We also need to prevent any alloca |
9619 | // from reordering above a preceeding stackrestore. |
9620 | if (match(BundleMember->Inst, m_Intrinsic<Intrinsic::stacksave>()) || |
9621 | match(BundleMember->Inst, m_Intrinsic<Intrinsic::stackrestore>())) { |
9622 | for (Instruction *I = BundleMember->Inst->getNextNode(); |
9623 | I != ScheduleEnd; I = I->getNextNode()) { |
9624 | if (match(I, m_Intrinsic<Intrinsic::stacksave>()) || |
9625 | match(I, m_Intrinsic<Intrinsic::stackrestore>())) |
9626 | // Any allocas past here must be control dependent on I, and I |
9627 | // must be memory dependend on BundleMember->Inst. |
9628 | break; |
9629 | |
9630 | if (!isa<AllocaInst>(I)) |
9631 | continue; |
9632 | |
9633 | // Add the dependency |
9634 | makeControlDependent(I); |
9635 | } |
9636 | } |
9637 | |
9638 | // In addition to the cases handle just above, we need to prevent |
9639 | // allocas from moving below a stacksave. The stackrestore case |
9640 | // is currently thought to be conservatism. |
9641 | if (isa<AllocaInst>(BundleMember->Inst)) { |
9642 | for (Instruction *I = BundleMember->Inst->getNextNode(); |
9643 | I != ScheduleEnd; I = I->getNextNode()) { |
9644 | if (!match(I, m_Intrinsic<Intrinsic::stacksave>()) && |
9645 | !match(I, m_Intrinsic<Intrinsic::stackrestore>())) |
9646 | continue; |
9647 | |
9648 | // Add the dependency |
9649 | makeControlDependent(I); |
9650 | break; |
9651 | } |
9652 | } |
9653 | } |
9654 | |
9655 | // Handle the memory dependencies (if any). |
9656 | ScheduleData *DepDest = BundleMember->NextLoadStore; |
9657 | if (!DepDest) |
9658 | continue; |
9659 | Instruction *SrcInst = BundleMember->Inst; |
9660 | 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", 9661, __extension__ __PRETTY_FUNCTION__)) |
9661 | "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", 9661, __extension__ __PRETTY_FUNCTION__)); |
9662 | MemoryLocation SrcLoc = getLocation(SrcInst); |
9663 | bool SrcMayWrite = BundleMember->Inst->mayWriteToMemory(); |
9664 | unsigned numAliased = 0; |
9665 | unsigned DistToSrc = 1; |
9666 | |
9667 | for ( ; DepDest; DepDest = DepDest->NextLoadStore) { |
9668 | assert(isInSchedulingRegion(DepDest))(static_cast <bool> (isInSchedulingRegion(DepDest)) ? void (0) : __assert_fail ("isInSchedulingRegion(DepDest)", "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 9668, __extension__ __PRETTY_FUNCTION__)); |
9669 | |
9670 | // We have two limits to reduce the complexity: |
9671 | // 1) AliasedCheckLimit: It's a small limit to reduce calls to |
9672 | // SLP->isAliased (which is the expensive part in this loop). |
9673 | // 2) MaxMemDepDistance: It's for very large blocks and it aborts |
9674 | // the whole loop (even if the loop is fast, it's quadratic). |
9675 | // It's important for the loop break condition (see below) to |
9676 | // check this limit even between two read-only instructions. |
9677 | if (DistToSrc >= MaxMemDepDistance || |
9678 | ((SrcMayWrite || DepDest->Inst->mayWriteToMemory()) && |
9679 | (numAliased >= AliasedCheckLimit || |
9680 | SLP->isAliased(SrcLoc, SrcInst, DepDest->Inst)))) { |
9681 | |
9682 | // We increment the counter only if the locations are aliased |
9683 | // (instead of counting all alias checks). This gives a better |
9684 | // balance between reduced runtime and accurate dependencies. |
9685 | numAliased++; |
9686 | |
9687 | DepDest->MemoryDependencies.push_back(BundleMember); |
9688 | BundleMember->Dependencies++; |
9689 | ScheduleData *DestBundle = DepDest->FirstInBundle; |
9690 | if (!DestBundle->IsScheduled) { |
9691 | BundleMember->incrementUnscheduledDeps(1); |
9692 | } |
9693 | if (!DestBundle->hasValidDependencies()) { |
9694 | WorkList.push_back(DestBundle); |
9695 | } |
9696 | } |
9697 | |
9698 | // Example, explaining the loop break condition: Let's assume our |
9699 | // starting instruction is i0 and MaxMemDepDistance = 3. |
9700 | // |
9701 | // +--------v--v--v |
9702 | // i0,i1,i2,i3,i4,i5,i6,i7,i8 |
9703 | // +--------^--^--^ |
9704 | // |
9705 | // MaxMemDepDistance let us stop alias-checking at i3 and we add |
9706 | // dependencies from i0 to i3,i4,.. (even if they are not aliased). |
9707 | // Previously we already added dependencies from i3 to i6,i7,i8 |
9708 | // (because of MaxMemDepDistance). As we added a dependency from |
9709 | // i0 to i3, we have transitive dependencies from i0 to i6,i7,i8 |
9710 | // and we can abort this loop at i6. |
9711 | if (DistToSrc >= 2 * MaxMemDepDistance) |
9712 | break; |
9713 | DistToSrc++; |
9714 | } |
9715 | } |
9716 | if (InsertInReadyList && SD->isReady()) { |
9717 | ReadyInsts.insert(SD); |
9718 | 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) |
9719 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: gets ready on update: " << *SD->Inst << "\n"; } } while (false); |
9720 | } |
9721 | } |
9722 | } |
9723 | |
9724 | void BoUpSLP::BlockScheduling::resetSchedule() { |
9725 | 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", 9726, __extension__ __PRETTY_FUNCTION__)) |
9726 | "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", 9726, __extension__ __PRETTY_FUNCTION__)); |
9727 | for (Instruction *I = ScheduleStart; I != ScheduleEnd; I = I->getNextNode()) { |
9728 | doForAllOpcodes(I, [&](ScheduleData *SD) { |
9729 | 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", 9730, __extension__ __PRETTY_FUNCTION__)) |
9730 | "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", 9730, __extension__ __PRETTY_FUNCTION__)); |
9731 | SD->IsScheduled = false; |
9732 | SD->resetUnscheduledDeps(); |
9733 | }); |
9734 | } |
9735 | ReadyInsts.clear(); |
9736 | } |
9737 | |
9738 | void BoUpSLP::scheduleBlock(BlockScheduling *BS) { |
9739 | if (!BS->ScheduleStart) |
9740 | return; |
9741 | |
9742 | 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); |
9743 | |
9744 | // A key point - if we got here, pre-scheduling was able to find a valid |
9745 | // scheduling of the sub-graph of the scheduling window which consists |
9746 | // of all vector bundles and their transitive users. As such, we do not |
9747 | // need to reschedule anything *outside of* that subgraph. |
9748 | |
9749 | BS->resetSchedule(); |
9750 | |
9751 | // For the real scheduling we use a more sophisticated ready-list: it is |
9752 | // sorted by the original instruction location. This lets the final schedule |
9753 | // be as close as possible to the original instruction order. |
9754 | // WARNING: If changing this order causes a correctness issue, that means |
9755 | // there is some missing dependence edge in the schedule data graph. |
9756 | struct ScheduleDataCompare { |
9757 | bool operator()(ScheduleData *SD1, ScheduleData *SD2) const { |
9758 | return SD2->SchedulingPriority < SD1->SchedulingPriority; |
9759 | } |
9760 | }; |
9761 | std::set<ScheduleData *, ScheduleDataCompare> ReadyInsts; |
9762 | |
9763 | // Ensure that all dependency data is updated (for nodes in the sub-graph) |
9764 | // and fill the ready-list with initial instructions. |
9765 | int Idx = 0; |
9766 | for (auto *I = BS->ScheduleStart; I != BS->ScheduleEnd; |
9767 | I = I->getNextNode()) { |
9768 | BS->doForAllOpcodes(I, [this, &Idx, BS](ScheduleData *SD) { |
9769 | TreeEntry *SDTE = getTreeEntry(SD->Inst); |
9770 | (void)SDTE; |
9771 | 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", 9774, __extension__ __PRETTY_FUNCTION__)) |
9772 | 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", 9774, __extension__ __PRETTY_FUNCTION__)) |
9773 | (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", 9774, __extension__ __PRETTY_FUNCTION__)) |
9774 | "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", 9774, __extension__ __PRETTY_FUNCTION__)); |
9775 | SD->FirstInBundle->SchedulingPriority = Idx++; |
9776 | |
9777 | if (SD->isSchedulingEntity() && SD->isPartOfBundle()) |
9778 | BS->calculateDependencies(SD, false, this); |
9779 | }); |
9780 | } |
9781 | BS->initialFillReadyList(ReadyInsts); |
9782 | |
9783 | Instruction *LastScheduledInst = BS->ScheduleEnd; |
9784 | |
9785 | // Do the "real" scheduling. |
9786 | while (!ReadyInsts.empty()) { |
9787 | ScheduleData *picked = *ReadyInsts.begin(); |
9788 | ReadyInsts.erase(ReadyInsts.begin()); |
9789 | |
9790 | // Move the scheduled instruction(s) to their dedicated places, if not |
9791 | // there yet. |
9792 | for (ScheduleData *BundleMember = picked; BundleMember; |
9793 | BundleMember = BundleMember->NextInBundle) { |
9794 | Instruction *pickedInst = BundleMember->Inst; |
9795 | if (pickedInst->getNextNode() != LastScheduledInst) |
9796 | pickedInst->moveBefore(LastScheduledInst); |
9797 | LastScheduledInst = pickedInst; |
9798 | } |
9799 | |
9800 | BS->schedule(picked, ReadyInsts); |
9801 | } |
9802 | |
9803 | // Check that we didn't break any of our invariants. |
9804 | #ifdef EXPENSIVE_CHECKS |
9805 | BS->verify(); |
9806 | #endif |
9807 | |
9808 | #if !defined(NDEBUG) || defined(EXPENSIVE_CHECKS) |
9809 | // Check that all schedulable entities got scheduled |
9810 | for (auto *I = BS->ScheduleStart; I != BS->ScheduleEnd; I = I->getNextNode()) { |
9811 | BS->doForAllOpcodes(I, [&](ScheduleData *SD) { |
9812 | if (SD->isSchedulingEntity() && SD->hasValidDependencies()) { |
9813 | 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", 9813, __extension__ __PRETTY_FUNCTION__)); |
9814 | } |
9815 | }); |
9816 | } |
9817 | #endif |
9818 | |
9819 | // Avoid duplicate scheduling of the block. |
9820 | BS->ScheduleStart = nullptr; |
9821 | } |
9822 | |
9823 | unsigned BoUpSLP::getVectorElementSize(Value *V) { |
9824 | // If V is a store, just return the width of the stored value (or value |
9825 | // truncated just before storing) without traversing the expression tree. |
9826 | // This is the common case. |
9827 | if (auto *Store = dyn_cast<StoreInst>(V)) |
9828 | return DL->getTypeSizeInBits(Store->getValueOperand()->getType()); |
9829 | |
9830 | if (auto *IEI = dyn_cast<InsertElementInst>(V)) |
9831 | return getVectorElementSize(IEI->getOperand(1)); |
9832 | |
9833 | auto E = InstrElementSize.find(V); |
9834 | if (E != InstrElementSize.end()) |
9835 | return E->second; |
9836 | |
9837 | // If V is not a store, we can traverse the expression tree to find loads |
9838 | // that feed it. The type of the loaded value may indicate a more suitable |
9839 | // width than V's type. We want to base the vector element size on the width |
9840 | // of memory operations where possible. |
9841 | SmallVector<std::pair<Instruction *, BasicBlock *>, 16> Worklist; |
9842 | SmallPtrSet<Instruction *, 16> Visited; |
9843 | if (auto *I = dyn_cast<Instruction>(V)) { |
9844 | Worklist.emplace_back(I, I->getParent()); |
9845 | Visited.insert(I); |
9846 | } |
9847 | |
9848 | // Traverse the expression tree in bottom-up order looking for loads. If we |
9849 | // encounter an instruction we don't yet handle, we give up. |
9850 | auto Width = 0u; |
9851 | while (!Worklist.empty()) { |
9852 | Instruction *I; |
9853 | BasicBlock *Parent; |
9854 | std::tie(I, Parent) = Worklist.pop_back_val(); |
9855 | |
9856 | // We should only be looking at scalar instructions here. If the current |
9857 | // instruction has a vector type, skip. |
9858 | auto *Ty = I->getType(); |
9859 | if (isa<VectorType>(Ty)) |
9860 | continue; |
9861 | |
9862 | // If the current instruction is a load, update MaxWidth to reflect the |
9863 | // width of the loaded value. |
9864 | if (isa<LoadInst, ExtractElementInst, ExtractValueInst>(I)) |
9865 | Width = std::max<unsigned>(Width, DL->getTypeSizeInBits(Ty)); |
9866 | |
9867 | // Otherwise, we need to visit the operands of the instruction. We only |
9868 | // handle the interesting cases from buildTree here. If an operand is an |
9869 | // instruction we haven't yet visited and from the same basic block as the |
9870 | // user or the use is a PHI node, we add it to the worklist. |
9871 | else if (isa<PHINode, CastInst, GetElementPtrInst, CmpInst, SelectInst, |
9872 | BinaryOperator, UnaryOperator>(I)) { |
9873 | for (Use &U : I->operands()) |
9874 | if (auto *J = dyn_cast<Instruction>(U.get())) |
9875 | if (Visited.insert(J).second && |
9876 | (isa<PHINode>(I) || J->getParent() == Parent)) |
9877 | Worklist.emplace_back(J, J->getParent()); |
9878 | } else { |
9879 | break; |
9880 | } |
9881 | } |
9882 | |
9883 | // If we didn't encounter a memory access in the expression tree, or if we |
9884 | // gave up for some reason, just return the width of V. Otherwise, return the |
9885 | // maximum width we found. |
9886 | if (!Width) { |
9887 | if (auto *CI = dyn_cast<CmpInst>(V)) |
9888 | V = CI->getOperand(0); |
9889 | Width = DL->getTypeSizeInBits(V->getType()); |
9890 | } |
9891 | |
9892 | for (Instruction *I : Visited) |
9893 | InstrElementSize[I] = Width; |
9894 | |
9895 | return Width; |
9896 | } |
9897 | |
9898 | // Determine if a value V in a vectorizable expression Expr can be demoted to a |
9899 | // smaller type with a truncation. We collect the values that will be demoted |
9900 | // in ToDemote and additional roots that require investigating in Roots. |
9901 | static bool collectValuesToDemote(Value *V, SmallPtrSetImpl<Value *> &Expr, |
9902 | SmallVectorImpl<Value *> &ToDemote, |
9903 | SmallVectorImpl<Value *> &Roots) { |
9904 | // We can always demote constants. |
9905 | if (isa<Constant>(V)) { |
9906 | ToDemote.push_back(V); |
9907 | return true; |
9908 | } |
9909 | |
9910 | // If the value is not an instruction in the expression with only one use, it |
9911 | // cannot be demoted. |
9912 | auto *I = dyn_cast<Instruction>(V); |
9913 | if (!I || !I->hasOneUse() || !Expr.count(I)) |
9914 | return false; |
9915 | |
9916 | switch (I->getOpcode()) { |
9917 | |
9918 | // We can always demote truncations and extensions. Since truncations can |
9919 | // seed additional demotion, we save the truncated value. |
9920 | case Instruction::Trunc: |
9921 | Roots.push_back(I->getOperand(0)); |
9922 | break; |
9923 | case Instruction::ZExt: |
9924 | case Instruction::SExt: |
9925 | if (isa<ExtractElementInst, InsertElementInst>(I->getOperand(0))) |
9926 | return false; |
9927 | break; |
9928 | |
9929 | // We can demote certain binary operations if we can demote both of their |
9930 | // operands. |
9931 | case Instruction::Add: |
9932 | case Instruction::Sub: |
9933 | case Instruction::Mul: |
9934 | case Instruction::And: |
9935 | case Instruction::Or: |
9936 | case Instruction::Xor: |
9937 | if (!collectValuesToDemote(I->getOperand(0), Expr, ToDemote, Roots) || |
9938 | !collectValuesToDemote(I->getOperand(1), Expr, ToDemote, Roots)) |
9939 | return false; |
9940 | break; |
9941 | |
9942 | // We can demote selects if we can demote their true and false values. |
9943 | case Instruction::Select: { |
9944 | SelectInst *SI = cast<SelectInst>(I); |
9945 | if (!collectValuesToDemote(SI->getTrueValue(), Expr, ToDemote, Roots) || |
9946 | !collectValuesToDemote(SI->getFalseValue(), Expr, ToDemote, Roots)) |
9947 | return false; |
9948 | break; |
9949 | } |
9950 | |
9951 | // We can demote phis if we can demote all their incoming operands. Note that |
9952 | // we don't need to worry about cycles since we ensure single use above. |
9953 | case Instruction::PHI: { |
9954 | PHINode *PN = cast<PHINode>(I); |
9955 | for (Value *IncValue : PN->incoming_values()) |
9956 | if (!collectValuesToDemote(IncValue, Expr, ToDemote, Roots)) |
9957 | return false; |
9958 | break; |
9959 | } |
9960 | |
9961 | // Otherwise, conservatively give up. |
9962 | default: |
9963 | return false; |
9964 | } |
9965 | |
9966 | // Record the value that we can demote. |
9967 | ToDemote.push_back(V); |
9968 | return true; |
9969 | } |
9970 | |
9971 | void BoUpSLP::computeMinimumValueSizes() { |
9972 | // If there are no external uses, the expression tree must be rooted by a |
9973 | // store. We can't demote in-memory values, so there is nothing to do here. |
9974 | if (ExternalUses.empty()) |
9975 | return; |
9976 | |
9977 | // We only attempt to truncate integer expressions. |
9978 | auto &TreeRoot = VectorizableTree[0]->Scalars; |
9979 | auto *TreeRootIT = dyn_cast<IntegerType>(TreeRoot[0]->getType()); |
9980 | if (!TreeRootIT) |
9981 | return; |
9982 | |
9983 | // If the expression is not rooted by a store, these roots should have |
9984 | // external uses. We will rely on InstCombine to rewrite the expression in |
9985 | // the narrower type. However, InstCombine only rewrites single-use values. |
9986 | // This means that if a tree entry other than a root is used externally, it |
9987 | // must have multiple uses and InstCombine will not rewrite it. The code |
9988 | // below ensures that only the roots are used externally. |
9989 | SmallPtrSet<Value *, 32> Expr(TreeRoot.begin(), TreeRoot.end()); |
9990 | for (auto &EU : ExternalUses) |
9991 | if (!Expr.erase(EU.Scalar)) |
9992 | return; |
9993 | if (!Expr.empty()) |
9994 | return; |
9995 | |
9996 | // Collect the scalar values of the vectorizable expression. We will use this |
9997 | // context to determine which values can be demoted. If we see a truncation, |
9998 | // we mark it as seeding another demotion. |
9999 | for (auto &EntryPtr : VectorizableTree) |
10000 | Expr.insert(EntryPtr->Scalars.begin(), EntryPtr->Scalars.end()); |
10001 | |
10002 | // Ensure the roots of the vectorizable tree don't form a cycle. They must |
10003 | // have a single external user that is not in the vectorizable tree. |
10004 | for (auto *Root : TreeRoot) |
10005 | if (!Root->hasOneUse() || Expr.count(*Root->user_begin())) |
10006 | return; |
10007 | |
10008 | // Conservatively determine if we can actually truncate the roots of the |
10009 | // expression. Collect the values that can be demoted in ToDemote and |
10010 | // additional roots that require investigating in Roots. |
10011 | SmallVector<Value *, 32> ToDemote; |
10012 | SmallVector<Value *, 4> Roots; |
10013 | for (auto *Root : TreeRoot) |
10014 | if (!collectValuesToDemote(Root, Expr, ToDemote, Roots)) |
10015 | return; |
10016 | |
10017 | // The maximum bit width required to represent all the values that can be |
10018 | // demoted without loss of precision. It would be safe to truncate the roots |
10019 | // of the expression to this width. |
10020 | auto MaxBitWidth = 8u; |
10021 | |
10022 | // We first check if all the bits of the roots are demanded. If they're not, |
10023 | // we can truncate the roots to this narrower type. |
10024 | for (auto *Root : TreeRoot) { |
10025 | auto Mask = DB->getDemandedBits(cast<Instruction>(Root)); |
10026 | MaxBitWidth = std::max<unsigned>( |
10027 | Mask.getBitWidth() - Mask.countLeadingZeros(), MaxBitWidth); |
10028 | } |
10029 | |
10030 | // True if the roots can be zero-extended back to their original type, rather |
10031 | // than sign-extended. We know that if the leading bits are not demanded, we |
10032 | // can safely zero-extend. So we initialize IsKnownPositive to True. |
10033 | bool IsKnownPositive = true; |
10034 | |
10035 | // If all the bits of the roots are demanded, we can try a little harder to |
10036 | // compute a narrower type. This can happen, for example, if the roots are |
10037 | // getelementptr indices. InstCombine promotes these indices to the pointer |
10038 | // width. Thus, all their bits are technically demanded even though the |
10039 | // address computation might be vectorized in a smaller type. |
10040 | // |
10041 | // We start by looking at each entry that can be demoted. We compute the |
10042 | // maximum bit width required to store the scalar by using ValueTracking to |
10043 | // compute the number of high-order bits we can truncate. |
10044 | if (MaxBitWidth == DL->getTypeSizeInBits(TreeRoot[0]->getType()) && |
10045 | llvm::all_of(TreeRoot, [](Value *R) { |
10046 | 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", 10046, __extension__ __PRETTY_FUNCTION__)); |
10047 | return isa<GetElementPtrInst>(R->user_back()); |
10048 | })) { |
10049 | MaxBitWidth = 8u; |
10050 | |
10051 | // Determine if the sign bit of all the roots is known to be zero. If not, |
10052 | // IsKnownPositive is set to False. |
10053 | IsKnownPositive = llvm::all_of(TreeRoot, [&](Value *R) { |
10054 | KnownBits Known = computeKnownBits(R, *DL); |
10055 | return Known.isNonNegative(); |
10056 | }); |
10057 | |
10058 | // Determine the maximum number of bits required to store the scalar |
10059 | // values. |
10060 | for (auto *Scalar : ToDemote) { |
10061 | auto NumSignBits = ComputeNumSignBits(Scalar, *DL, 0, AC, nullptr, DT); |
10062 | auto NumTypeBits = DL->getTypeSizeInBits(Scalar->getType()); |
10063 | MaxBitWidth = std::max<unsigned>(NumTypeBits - NumSignBits, MaxBitWidth); |
10064 | } |
10065 | |
10066 | // If we can't prove that the sign bit is zero, we must add one to the |
10067 | // maximum bit width to account for the unknown sign bit. This preserves |
10068 | // the existing sign bit so we can safely sign-extend the root back to the |
10069 | // original type. Otherwise, if we know the sign bit is zero, we will |
10070 | // zero-extend the root instead. |
10071 | // |
10072 | // FIXME: This is somewhat suboptimal, as there will be cases where adding |
10073 | // one to the maximum bit width will yield a larger-than-necessary |
10074 | // type. In general, we need to add an extra bit only if we can't |
10075 | // prove that the upper bit of the original type is equal to the |
10076 | // upper bit of the proposed smaller type. If these two bits are the |
10077 | // same (either zero or one) we know that sign-extending from the |
10078 | // smaller type will result in the same value. Here, since we can't |
10079 | // yet prove this, we are just making the proposed smaller type |
10080 | // larger to ensure correctness. |
10081 | if (!IsKnownPositive) |
10082 | ++MaxBitWidth; |
10083 | } |
10084 | |
10085 | // Round MaxBitWidth up to the next power-of-two. |
10086 | if (!isPowerOf2_64(MaxBitWidth)) |
10087 | MaxBitWidth = NextPowerOf2(MaxBitWidth); |
10088 | |
10089 | // If the maximum bit width we compute is less than the with of the roots' |
10090 | // type, we can proceed with the narrowing. Otherwise, do nothing. |
10091 | if (MaxBitWidth >= TreeRootIT->getBitWidth()) |
10092 | return; |
10093 | |
10094 | // If we can truncate the root, we must collect additional values that might |
10095 | // be demoted as a result. That is, those seeded by truncations we will |
10096 | // modify. |
10097 | while (!Roots.empty()) |
10098 | collectValuesToDemote(Roots.pop_back_val(), Expr, ToDemote, Roots); |
10099 | |
10100 | // Finally, map the values we can demote to the maximum bit with we computed. |
10101 | for (auto *Scalar : ToDemote) |
10102 | MinBWs[Scalar] = std::make_pair(MaxBitWidth, !IsKnownPositive); |
10103 | } |
10104 | |
10105 | namespace { |
10106 | |
10107 | /// The SLPVectorizer Pass. |
10108 | struct SLPVectorizer : public FunctionPass { |
10109 | SLPVectorizerPass Impl; |
10110 | |
10111 | /// Pass identification, replacement for typeid |
10112 | static char ID; |
10113 | |
10114 | explicit SLPVectorizer() : FunctionPass(ID) { |
10115 | initializeSLPVectorizerPass(*PassRegistry::getPassRegistry()); |
10116 | } |
10117 | |
10118 | bool doInitialization(Module &M) override { return false; } |
10119 | |
10120 | bool runOnFunction(Function &F) override { |
10121 | if (skipFunction(F)) |
10122 | return false; |
10123 | |
10124 | auto *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE(); |
10125 | auto *TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F); |
10126 | auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>(); |
10127 | auto *TLI = TLIP ? &TLIP->getTLI(F) : nullptr; |
10128 | auto *AA = &getAnalysis<AAResultsWrapperPass>().getAAResults(); |
10129 | auto *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); |
10130 | auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); |
10131 | auto *AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F); |
10132 | auto *DB = &getAnalysis<DemandedBitsWrapperPass>().getDemandedBits(); |
10133 | auto *ORE = &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE(); |
10134 | |
10135 | return Impl.runImpl(F, SE, TTI, TLI, AA, LI, DT, AC, DB, ORE); |
10136 | } |
10137 | |
10138 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
10139 | FunctionPass::getAnalysisUsage(AU); |
10140 | AU.addRequired<AssumptionCacheTracker>(); |
10141 | AU.addRequired<ScalarEvolutionWrapperPass>(); |
10142 | AU.addRequired<AAResultsWrapperPass>(); |
10143 | AU.addRequired<TargetTransformInfoWrapperPass>(); |
10144 | AU.addRequired<LoopInfoWrapperPass>(); |
10145 | AU.addRequired<DominatorTreeWrapperPass>(); |
10146 | AU.addRequired<DemandedBitsWrapperPass>(); |
10147 | AU.addRequired<OptimizationRemarkEmitterWrapperPass>(); |
10148 | AU.addRequired<InjectTLIMappingsLegacy>(); |
10149 | AU.addPreserved<LoopInfoWrapperPass>(); |
10150 | AU.addPreserved<DominatorTreeWrapperPass>(); |
10151 | AU.addPreserved<AAResultsWrapperPass>(); |
10152 | AU.addPreserved<GlobalsAAWrapperPass>(); |
10153 | AU.setPreservesCFG(); |
10154 | } |
10155 | }; |
10156 | |
10157 | } // end anonymous namespace |
10158 | |
10159 | PreservedAnalyses SLPVectorizerPass::run(Function &F, FunctionAnalysisManager &AM) { |
10160 | auto *SE = &AM.getResult<ScalarEvolutionAnalysis>(F); |
10161 | auto *TTI = &AM.getResult<TargetIRAnalysis>(F); |
10162 | auto *TLI = AM.getCachedResult<TargetLibraryAnalysis>(F); |
10163 | auto *AA = &AM.getResult<AAManager>(F); |
10164 | auto *LI = &AM.getResult<LoopAnalysis>(F); |
10165 | auto *DT = &AM.getResult<DominatorTreeAnalysis>(F); |
10166 | auto *AC = &AM.getResult<AssumptionAnalysis>(F); |
10167 | auto *DB = &AM.getResult<DemandedBitsAnalysis>(F); |
10168 | auto *ORE = &AM.getResult<OptimizationRemarkEmitterAnalysis>(F); |
10169 | |
10170 | bool Changed = runImpl(F, SE, TTI, TLI, AA, LI, DT, AC, DB, ORE); |
10171 | if (!Changed) |
10172 | return PreservedAnalyses::all(); |
10173 | |
10174 | PreservedAnalyses PA; |
10175 | PA.preserveSet<CFGAnalyses>(); |
10176 | return PA; |
10177 | } |
10178 | |
10179 | bool SLPVectorizerPass::runImpl(Function &F, ScalarEvolution *SE_, |
10180 | TargetTransformInfo *TTI_, |
10181 | TargetLibraryInfo *TLI_, AAResults *AA_, |
10182 | LoopInfo *LI_, DominatorTree *DT_, |
10183 | AssumptionCache *AC_, DemandedBits *DB_, |
10184 | OptimizationRemarkEmitter *ORE_) { |
10185 | if (!RunSLPVectorization) |
10186 | return false; |
10187 | SE = SE_; |
10188 | TTI = TTI_; |
10189 | TLI = TLI_; |
10190 | AA = AA_; |
10191 | LI = LI_; |
10192 | DT = DT_; |
10193 | AC = AC_; |
10194 | DB = DB_; |
10195 | DL = &F.getParent()->getDataLayout(); |
10196 | |
10197 | Stores.clear(); |
10198 | GEPs.clear(); |
10199 | bool Changed = false; |
10200 | |
10201 | // If the target claims to have no vector registers don't attempt |
10202 | // vectorization. |
10203 | if (!TTI->getNumberOfRegisters(TTI->getRegisterClassForType(true))) { |
10204 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Didn't find any vector registers for target, abort.\n" ; } } while (false) |
10205 | 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); |
10206 | return false; |
10207 | } |
10208 | |
10209 | // Don't vectorize when the attribute NoImplicitFloat is used. |
10210 | if (F.hasFnAttribute(Attribute::NoImplicitFloat)) |
10211 | return false; |
10212 | |
10213 | 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); |
10214 | |
10215 | // Use the bottom up slp vectorizer to construct chains that start with |
10216 | // store instructions. |
10217 | BoUpSLP R(&F, SE, TTI, TLI, AA, LI, DT, AC, DB, DL, ORE_); |
10218 | |
10219 | // A general note: the vectorizer must use BoUpSLP::eraseInstruction() to |
10220 | // delete instructions. |
10221 | |
10222 | // Update DFS numbers now so that we can use them for ordering. |
10223 | DT->updateDFSNumbers(); |
10224 | |
10225 | // Scan the blocks in the function in post order. |
10226 | for (auto *BB : post_order(&F.getEntryBlock())) { |
10227 | // Start new block - clear the list of reduction roots. |
10228 | R.clearReductionData(); |
10229 | collectSeedInstructions(BB); |
10230 | |
10231 | // Vectorize trees that end at stores. |
10232 | if (!Stores.empty()) { |
10233 | 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) |
10234 | << " underlying objects.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Found stores for " << Stores .size() << " underlying objects.\n"; } } while (false); |
10235 | Changed |= vectorizeStoreChains(R); |
10236 | } |
10237 | |
10238 | // Vectorize trees that end at reductions. |
10239 | Changed |= vectorizeChainsInBlock(BB, R); |
10240 | |
10241 | // Vectorize the index computations of getelementptr instructions. This |
10242 | // is primarily intended to catch gather-like idioms ending at |
10243 | // non-consecutive loads. |
10244 | if (!GEPs.empty()) { |
10245 | 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) |
10246 | << " underlying objects.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Found GEPs for " << GEPs .size() << " underlying objects.\n"; } } while (false); |
10247 | Changed |= vectorizeGEPIndices(BB, R); |
10248 | } |
10249 | } |
10250 | |
10251 | if (Changed) { |
10252 | R.optimizeGatherSequence(); |
10253 | LLVM_DEBUG(dbgs() << "SLP: vectorized \"" << F.getName() << "\"\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: vectorized \"" << F.getName () << "\"\n"; } } while (false); |
10254 | } |
10255 | return Changed; |
10256 | } |
10257 | |
10258 | bool SLPVectorizerPass::vectorizeStoreChain(ArrayRef<Value *> Chain, BoUpSLP &R, |
10259 | unsigned Idx, unsigned MinVF) { |
10260 | 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) |
10261 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing a store chain of length " << Chain.size() << "\n"; } } while (false); |
10262 | const unsigned Sz = R.getVectorElementSize(Chain[0]); |
10263 | unsigned VF = Chain.size(); |
10264 | |
10265 | if (!isPowerOf2_32(Sz) || !isPowerOf2_32(VF) || VF < 2 || VF < MinVF) |
10266 | return false; |
10267 | |
10268 | 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) |
10269 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing " << VF << " stores at offset " << Idx << "\n"; } } while ( false); |
10270 | |
10271 | R.buildTree(Chain); |
10272 | if (R.isTreeTinyAndNotFullyVectorizable()) |
10273 | return false; |
10274 | if (R.isLoadCombineCandidate()) |
10275 | return false; |
10276 | R.reorderTopToBottom(); |
10277 | R.reorderBottomToTop(); |
10278 | R.buildExternalUses(); |
10279 | |
10280 | R.computeMinimumValueSizes(); |
10281 | |
10282 | InstructionCost Cost = R.getTreeCost(); |
10283 | |
10284 | 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 ); |
10285 | if (Cost < -SLPCostThreshold) { |
10286 | 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); |
10287 | |
10288 | using namespace ore; |
10289 | |
10290 | R.getORE()->emit(OptimizationRemark(SV_NAME"slp-vectorizer", "StoresVectorized", |
10291 | cast<StoreInst>(Chain[0])) |
10292 | << "Stores SLP vectorized with cost " << NV("Cost", Cost) |
10293 | << " and with tree size " |
10294 | << NV("TreeSize", R.getTreeSize())); |
10295 | |
10296 | R.vectorizeTree(); |
10297 | return true; |
10298 | } |
10299 | |
10300 | return false; |
10301 | } |
10302 | |
10303 | bool SLPVectorizerPass::vectorizeStores(ArrayRef<StoreInst *> Stores, |
10304 | BoUpSLP &R) { |
10305 | // We may run into multiple chains that merge into a single chain. We mark the |
10306 | // stores that we vectorized so that we don't visit the same store twice. |
10307 | BoUpSLP::ValueSet VectorizedStores; |
10308 | bool Changed = false; |
10309 | |
10310 | int E = Stores.size(); |
10311 | SmallBitVector Tails(E, false); |
10312 | int MaxIter = MaxStoreLookup.getValue(); |
10313 | SmallVector<std::pair<int, int>, 16> ConsecutiveChain( |
10314 | E, std::make_pair(E, INT_MAX2147483647)); |
10315 | SmallVector<SmallBitVector, 4> CheckedPairs(E, SmallBitVector(E, false)); |
10316 | int IterCnt; |
10317 | auto &&FindConsecutiveAccess = [this, &Stores, &Tails, &IterCnt, MaxIter, |
10318 | &CheckedPairs, |
10319 | &ConsecutiveChain](int K, int Idx) { |
10320 | if (IterCnt >= MaxIter) |
10321 | return true; |
10322 | if (CheckedPairs[Idx].test(K)) |
10323 | return ConsecutiveChain[K].second == 1 && |
10324 | ConsecutiveChain[K].first == Idx; |
10325 | ++IterCnt; |
10326 | CheckedPairs[Idx].set(K); |
10327 | CheckedPairs[K].set(Idx); |
10328 | Optional<int> Diff = getPointersDiff( |
10329 | Stores[K]->getValueOperand()->getType(), Stores[K]->getPointerOperand(), |
10330 | Stores[Idx]->getValueOperand()->getType(), |
10331 | Stores[Idx]->getPointerOperand(), *DL, *SE, /*StrictCheck=*/true); |
10332 | if (!Diff || *Diff == 0) |
10333 | return false; |
10334 | int Val = *Diff; |
10335 | if (Val < 0) { |
10336 | if (ConsecutiveChain[Idx].second > -Val) { |
10337 | Tails.set(K); |
10338 | ConsecutiveChain[Idx] = std::make_pair(K, -Val); |
10339 | } |
10340 | return false; |
10341 | } |
10342 | if (ConsecutiveChain[K].second <= Val) |
10343 | return false; |
10344 | |
10345 | Tails.set(Idx); |
10346 | ConsecutiveChain[K] = std::make_pair(Idx, Val); |
10347 | return Val == 1; |
10348 | }; |
10349 | // Do a quadratic search on all of the given stores in reverse order and find |
10350 | // all of the pairs of stores that follow each other. |
10351 | for (int Idx = E - 1; Idx >= 0; --Idx) { |
10352 | // If a store has multiple consecutive store candidates, search according |
10353 | // to the sequence: Idx-1, Idx+1, Idx-2, Idx+2, ... |
10354 | // This is because usually pairing with immediate succeeding or preceding |
10355 | // candidate create the best chance to find slp vectorization opportunity. |
10356 | const int MaxLookDepth = std::max(E - Idx, Idx + 1); |
10357 | IterCnt = 0; |
10358 | for (int Offset = 1, F = MaxLookDepth; Offset < F; ++Offset) |
10359 | if ((Idx >= Offset && FindConsecutiveAccess(Idx - Offset, Idx)) || |
10360 | (Idx + Offset < E && FindConsecutiveAccess(Idx + Offset, Idx))) |
10361 | break; |
10362 | } |
10363 | |
10364 | // Tracks if we tried to vectorize stores starting from the given tail |
10365 | // already. |
10366 | SmallBitVector TriedTails(E, false); |
10367 | // For stores that start but don't end a link in the chain: |
10368 | for (int Cnt = E; Cnt > 0; --Cnt) { |
10369 | int I = Cnt - 1; |
10370 | if (ConsecutiveChain[I].first == E || Tails.test(I)) |
10371 | continue; |
10372 | // We found a store instr that starts a chain. Now follow the chain and try |
10373 | // to vectorize it. |
10374 | BoUpSLP::ValueList Operands; |
10375 | // Collect the chain into a list. |
10376 | while (I != E && !VectorizedStores.count(Stores[I])) { |
10377 | Operands.push_back(Stores[I]); |
10378 | Tails.set(I); |
10379 | if (ConsecutiveChain[I].second != 1) { |
10380 | // Mark the new end in the chain and go back, if required. It might be |
10381 | // required if the original stores come in reversed order, for example. |
10382 | if (ConsecutiveChain[I].first != E && |
10383 | Tails.test(ConsecutiveChain[I].first) && !TriedTails.test(I) && |
10384 | !VectorizedStores.count(Stores[ConsecutiveChain[I].first])) { |
10385 | TriedTails.set(I); |
10386 | Tails.reset(ConsecutiveChain[I].first); |
10387 | if (Cnt < ConsecutiveChain[I].first + 2) |
10388 | Cnt = ConsecutiveChain[I].first + 2; |
10389 | } |
10390 | break; |
10391 | } |
10392 | // Move to the next value in the chain. |
10393 | I = ConsecutiveChain[I].first; |
10394 | } |
10395 | 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", 10395, __extension__ __PRETTY_FUNCTION__)); |
10396 | |
10397 | unsigned MaxVecRegSize = R.getMaxVecRegSize(); |
10398 | unsigned EltSize = R.getVectorElementSize(Operands[0]); |
10399 | unsigned MaxElts = llvm::PowerOf2Floor(MaxVecRegSize / EltSize); |
10400 | |
10401 | unsigned MaxVF = std::min(R.getMaximumVF(EltSize, Instruction::Store), |
10402 | MaxElts); |
10403 | auto *Store = cast<StoreInst>(Operands[0]); |
10404 | Type *StoreTy = Store->getValueOperand()->getType(); |
10405 | Type *ValueTy = StoreTy; |
10406 | if (auto *Trunc = dyn_cast<TruncInst>(Store->getValueOperand())) |
10407 | ValueTy = Trunc->getSrcTy(); |
10408 | unsigned MinVF = TTI->getStoreMinimumVF( |
10409 | R.getMinVF(DL->getTypeSizeInBits(ValueTy)), StoreTy, ValueTy); |
10410 | |
10411 | if (MaxVF <= MinVF) { |
10412 | 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) |
10413 | << "MinVF (" << MinVF << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Vectorization infeasible as MaxVF (" << MaxVF << ") <= " << "MinVF (" << MinVF << ")\n"; } } while (false); |
10414 | } |
10415 | |
10416 | // FIXME: Is division-by-2 the correct step? Should we assert that the |
10417 | // register size is a power-of-2? |
10418 | unsigned StartIdx = 0; |
10419 | for (unsigned Size = MaxVF; Size >= MinVF; Size /= 2) { |
10420 | for (unsigned Cnt = StartIdx, E = Operands.size(); Cnt + Size <= E;) { |
10421 | ArrayRef<Value *> Slice = makeArrayRef(Operands).slice(Cnt, Size); |
10422 | if (!VectorizedStores.count(Slice.front()) && |
10423 | !VectorizedStores.count(Slice.back()) && |
10424 | vectorizeStoreChain(Slice, R, Cnt, MinVF)) { |
10425 | // Mark the vectorized stores so that we don't vectorize them again. |
10426 | VectorizedStores.insert(Slice.begin(), Slice.end()); |
10427 | Changed = true; |
10428 | // If we vectorized initial block, no need to try to vectorize it |
10429 | // again. |
10430 | if (Cnt == StartIdx) |
10431 | StartIdx += Size; |
10432 | Cnt += Size; |
10433 | continue; |
10434 | } |
10435 | ++Cnt; |
10436 | } |
10437 | // Check if the whole array was vectorized already - exit. |
10438 | if (StartIdx >= Operands.size()) |
10439 | break; |
10440 | } |
10441 | } |
10442 | |
10443 | return Changed; |
10444 | } |
10445 | |
10446 | void SLPVectorizerPass::collectSeedInstructions(BasicBlock *BB) { |
10447 | // Initialize the collections. We will make a single pass over the block. |
10448 | Stores.clear(); |
10449 | GEPs.clear(); |
10450 | |
10451 | // Visit the store and getelementptr instructions in BB and organize them in |
10452 | // Stores and GEPs according to the underlying objects of their pointer |
10453 | // operands. |
10454 | for (Instruction &I : *BB) { |
10455 | // Ignore store instructions that are volatile or have a pointer operand |
10456 | // that doesn't point to a scalar type. |
10457 | if (auto *SI = dyn_cast<StoreInst>(&I)) { |
10458 | if (!SI->isSimple()) |
10459 | continue; |
10460 | if (!isValidElementType(SI->getValueOperand()->getType())) |
10461 | continue; |
10462 | Stores[getUnderlyingObject(SI->getPointerOperand())].push_back(SI); |
10463 | } |
10464 | |
10465 | // Ignore getelementptr instructions that have more than one index, a |
10466 | // constant index, or a pointer operand that doesn't point to a scalar |
10467 | // type. |
10468 | else if (auto *GEP = dyn_cast<GetElementPtrInst>(&I)) { |
10469 | auto Idx = GEP->idx_begin()->get(); |
10470 | if (GEP->getNumIndices() > 1 || isa<Constant>(Idx)) |
10471 | continue; |
10472 | if (!isValidElementType(Idx->getType())) |
10473 | continue; |
10474 | if (GEP->getType()->isVectorTy()) |
10475 | continue; |
10476 | GEPs[GEP->getPointerOperand()].push_back(GEP); |
10477 | } |
10478 | } |
10479 | } |
10480 | |
10481 | bool SLPVectorizerPass::tryToVectorizePair(Value *A, Value *B, BoUpSLP &R) { |
10482 | if (!A || !B) |
10483 | return false; |
10484 | if (isa<InsertElementInst>(A) || isa<InsertElementInst>(B)) |
10485 | return false; |
10486 | Value *VL[] = {A, B}; |
10487 | return tryToVectorizeList(VL, R); |
10488 | } |
10489 | |
10490 | bool SLPVectorizerPass::tryToVectorizeList(ArrayRef<Value *> VL, BoUpSLP &R, |
10491 | bool LimitForRegisterSize) { |
10492 | if (VL.size() < 2) |
10493 | return false; |
10494 | |
10495 | 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) |
10496 | << VL.size() << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Trying to vectorize a list of length = " << VL.size() << ".\n"; } } while (false); |
10497 | |
10498 | // Check that all of the parts are instructions of the same type, |
10499 | // we permit an alternate opcode via InstructionsState. |
10500 | InstructionsState S = getSameOpcode(VL); |
10501 | if (!S.getOpcode()) |
10502 | return false; |
10503 | |
10504 | Instruction *I0 = cast<Instruction>(S.OpValue); |
10505 | // Make sure invalid types (including vector type) are rejected before |
10506 | // determining vectorization factor for scalar instructions. |
10507 | for (Value *V : VL) { |
10508 | Type *Ty = V->getType(); |
10509 | if (!isa<InsertElementInst>(V) && !isValidElementType(Ty)) { |
10510 | // NOTE: the following will give user internal llvm type name, which may |
10511 | // not be useful. |
10512 | R.getORE()->emit([&]() { |
10513 | std::string type_str; |
10514 | llvm::raw_string_ostream rso(type_str); |
10515 | Ty->print(rso); |
10516 | return OptimizationRemarkMissed(SV_NAME"slp-vectorizer", "UnsupportedType", I0) |
10517 | << "Cannot SLP vectorize list: type " |
10518 | << rso.str() + " is unsupported by vectorizer"; |
10519 | }); |
10520 | return false; |
10521 | } |
10522 | } |
10523 | |
10524 | unsigned Sz = R.getVectorElementSize(I0); |
10525 | unsigned MinVF = R.getMinVF(Sz); |
10526 | unsigned MaxVF = std::max<unsigned>(PowerOf2Floor(VL.size()), MinVF); |
10527 | MaxVF = std::min(R.getMaximumVF(Sz, S.getOpcode()), MaxVF); |
10528 | if (MaxVF < 2) { |
10529 | R.getORE()->emit([&]() { |
10530 | return OptimizationRemarkMissed(SV_NAME"slp-vectorizer", "SmallVF", I0) |
10531 | << "Cannot SLP vectorize list: vectorization factor " |
10532 | << "less than 2 is not supported"; |
10533 | }); |
10534 | return false; |
10535 | } |
10536 | |
10537 | bool Changed = false; |
10538 | bool CandidateFound = false; |
10539 | InstructionCost MinCost = SLPCostThreshold.getValue(); |
10540 | Type *ScalarTy = VL[0]->getType(); |
10541 | if (auto *IE = dyn_cast<InsertElementInst>(VL[0])) |
10542 | ScalarTy = IE->getOperand(1)->getType(); |
10543 | |
10544 | unsigned NextInst = 0, MaxInst = VL.size(); |
10545 | for (unsigned VF = MaxVF; NextInst + 1 < MaxInst && VF >= MinVF; VF /= 2) { |
10546 | // No actual vectorization should happen, if number of parts is the same as |
10547 | // provided vectorization factor (i.e. the scalar type is used for vector |
10548 | // code during codegen). |
10549 | auto *VecTy = FixedVectorType::get(ScalarTy, VF); |
10550 | if (TTI->getNumberOfParts(VecTy) == VF) |
10551 | continue; |
10552 | for (unsigned I = NextInst; I < MaxInst; ++I) { |
10553 | unsigned OpsWidth = 0; |
10554 | |
10555 | if (I + VF > MaxInst) |
10556 | OpsWidth = MaxInst - I; |
10557 | else |
10558 | OpsWidth = VF; |
10559 | |
10560 | if (!isPowerOf2_32(OpsWidth)) |
10561 | continue; |
10562 | |
10563 | if ((LimitForRegisterSize && OpsWidth < MaxVF) || |
10564 | (VF > MinVF && OpsWidth <= VF / 2) || (VF == MinVF && OpsWidth < 2)) |
10565 | break; |
10566 | |
10567 | ArrayRef<Value *> Ops = VL.slice(I, OpsWidth); |
10568 | // Check that a previous iteration of this loop did not delete the Value. |
10569 | if (llvm::any_of(Ops, [&R](Value *V) { |
10570 | auto *I = dyn_cast<Instruction>(V); |
10571 | return I && R.isDeleted(I); |
10572 | })) |
10573 | continue; |
10574 | |
10575 | LLVM_DEBUG(dbgs() << "SLP: Analyzing " << OpsWidth << " operations "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing " << OpsWidth << " operations " << "\n"; } } while (false) |
10576 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing " << OpsWidth << " operations " << "\n"; } } while (false); |
10577 | |
10578 | R.buildTree(Ops); |
10579 | if (R.isTreeTinyAndNotFullyVectorizable()) |
10580 | continue; |
10581 | R.reorderTopToBottom(); |
10582 | R.reorderBottomToTop(!isa<InsertElementInst>(Ops.front())); |
10583 | R.buildExternalUses(); |
10584 | |
10585 | R.computeMinimumValueSizes(); |
10586 | InstructionCost Cost = R.getTreeCost(); |
10587 | CandidateFound = true; |
10588 | MinCost = std::min(MinCost, Cost); |
10589 | |
10590 | 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 ); |
10591 | if (Cost < -SLPCostThreshold) { |
10592 | 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); |
10593 | R.getORE()->emit(OptimizationRemark(SV_NAME"slp-vectorizer", "VectorizedList", |
10594 | cast<Instruction>(Ops[0])) |
10595 | << "SLP vectorized with cost " << ore::NV("Cost", Cost) |
10596 | << " and with tree size " |
10597 | << ore::NV("TreeSize", R.getTreeSize())); |
10598 | |
10599 | R.vectorizeTree(); |
10600 | // Move to the next bundle. |
10601 | I += VF - 1; |
10602 | NextInst = I + 1; |
10603 | Changed = true; |
10604 | } |
10605 | } |
10606 | } |
10607 | |
10608 | if (!Changed && CandidateFound) { |
10609 | R.getORE()->emit([&]() { |
10610 | return OptimizationRemarkMissed(SV_NAME"slp-vectorizer", "NotBeneficial", I0) |
10611 | << "List vectorization was possible but not beneficial with cost " |
10612 | << ore::NV("Cost", MinCost) << " >= " |
10613 | << ore::NV("Treshold", -SLPCostThreshold); |
10614 | }); |
10615 | } else if (!Changed) { |
10616 | R.getORE()->emit([&]() { |
10617 | return OptimizationRemarkMissed(SV_NAME"slp-vectorizer", "NotPossible", I0) |
10618 | << "Cannot SLP vectorize list: vectorization was impossible" |
10619 | << " with available vectorization factors"; |
10620 | }); |
10621 | } |
10622 | return Changed; |
10623 | } |
10624 | |
10625 | bool SLPVectorizerPass::tryToVectorize(Instruction *I, BoUpSLP &R) { |
10626 | if (!I) |
10627 | return false; |
10628 | |
10629 | if (!isa<BinaryOperator, CmpInst>(I) || isa<VectorType>(I->getType())) |
10630 | return false; |
10631 | |
10632 | Value *P = I->getParent(); |
10633 | |
10634 | // Vectorize in current basic block only. |
10635 | auto *Op0 = dyn_cast<Instruction>(I->getOperand(0)); |
10636 | auto *Op1 = dyn_cast<Instruction>(I->getOperand(1)); |
10637 | if (!Op0 || !Op1 || Op0->getParent() != P || Op1->getParent() != P) |
10638 | return false; |
10639 | |
10640 | // First collect all possible candidates |
10641 | SmallVector<std::pair<Value *, Value *>, 4> Candidates; |
10642 | Candidates.emplace_back(Op0, Op1); |
10643 | |
10644 | auto *A = dyn_cast<BinaryOperator>(Op0); |
10645 | auto *B = dyn_cast<BinaryOperator>(Op1); |
10646 | // Try to skip B. |
10647 | if (A && B && B->hasOneUse()) { |
10648 | auto *B0 = dyn_cast<BinaryOperator>(B->getOperand(0)); |
10649 | auto *B1 = dyn_cast<BinaryOperator>(B->getOperand(1)); |
10650 | if (B0 && B0->getParent() == P) |
10651 | Candidates.emplace_back(A, B0); |
10652 | if (B1 && B1->getParent() == P) |
10653 | Candidates.emplace_back(A, B1); |
10654 | } |
10655 | // Try to skip A. |
10656 | if (B && A && A->hasOneUse()) { |
10657 | auto *A0 = dyn_cast<BinaryOperator>(A->getOperand(0)); |
10658 | auto *A1 = dyn_cast<BinaryOperator>(A->getOperand(1)); |
10659 | if (A0 && A0->getParent() == P) |
10660 | Candidates.emplace_back(A0, B); |
10661 | if (A1 && A1->getParent() == P) |
10662 | Candidates.emplace_back(A1, B); |
10663 | } |
10664 | |
10665 | if (Candidates.size() == 1) |
10666 | return tryToVectorizePair(Op0, Op1, R); |
10667 | |
10668 | // We have multiple options. Try to pick the single best. |
10669 | Optional<int> BestCandidate = R.findBestRootPair(Candidates); |
10670 | if (!BestCandidate) |
10671 | return false; |
10672 | return tryToVectorizePair(Candidates[*BestCandidate].first, |
10673 | Candidates[*BestCandidate].second, R); |
10674 | } |
10675 | |
10676 | namespace { |
10677 | |
10678 | /// Model horizontal reductions. |
10679 | /// |
10680 | /// A horizontal reduction is a tree of reduction instructions that has values |
10681 | /// that can be put into a vector as its leaves. For example: |
10682 | /// |
10683 | /// mul mul mul mul |
10684 | /// \ / \ / |
10685 | /// + + |
10686 | /// \ / |
10687 | /// + |
10688 | /// This tree has "mul" as its leaf values and "+" as its reduction |
10689 | /// instructions. A reduction can feed into a store or a binary operation |
10690 | /// feeding a phi. |
10691 | /// ... |
10692 | /// \ / |
10693 | /// + |
10694 | /// | |
10695 | /// phi += |
10696 | /// |
10697 | /// Or: |
10698 | /// ... |
10699 | /// \ / |
10700 | /// + |
10701 | /// | |
10702 | /// *p = |
10703 | /// |
10704 | class HorizontalReduction { |
10705 | using ReductionOpsType = SmallVector<Value *, 16>; |
10706 | using ReductionOpsListType = SmallVector<ReductionOpsType, 2>; |
10707 | ReductionOpsListType ReductionOps; |
10708 | /// List of possibly reduced values. |
10709 | SmallVector<SmallVector<Value *>> ReducedVals; |
10710 | /// Maps reduced value to the corresponding reduction operation. |
10711 | DenseMap<Value *, SmallVector<Instruction *>> ReducedValsToOps; |
10712 | // Use map vector to make stable output. |
10713 | MapVector<Instruction *, Value *> ExtraArgs; |
10714 | WeakTrackingVH ReductionRoot; |
10715 | /// The type of reduction operation. |
10716 | RecurKind RdxKind; |
10717 | |
10718 | static bool isCmpSelMinMax(Instruction *I) { |
10719 | return match(I, m_Select(m_Cmp(), m_Value(), m_Value())) && |
10720 | RecurrenceDescriptor::isMinMaxRecurrenceKind(getRdxKind(I)); |
10721 | } |
10722 | |
10723 | // And/or are potentially poison-safe logical patterns like: |
10724 | // select x, y, false |
10725 | // select x, true, y |
10726 | static bool isBoolLogicOp(Instruction *I) { |
10727 | return isa<SelectInst>(I) && |
10728 | (match(I, m_LogicalAnd()) || match(I, m_LogicalOr())); |
10729 | } |
10730 | |
10731 | /// Checks if instruction is associative and can be vectorized. |
10732 | static bool isVectorizable(RecurKind Kind, Instruction *I) { |
10733 | if (Kind == RecurKind::None) |
10734 | return false; |
10735 | |
10736 | // Integer ops that map to select instructions or intrinsics are fine. |
10737 | if (RecurrenceDescriptor::isIntMinMaxRecurrenceKind(Kind) || |
10738 | isBoolLogicOp(I)) |
10739 | return true; |
10740 | |
10741 | if (Kind == RecurKind::FMax || Kind == RecurKind::FMin) { |
10742 | // FP min/max are associative except for NaN and -0.0. We do not |
10743 | // have to rule out -0.0 here because the intrinsic semantics do not |
10744 | // specify a fixed result for it. |
10745 | return I->getFastMathFlags().noNaNs(); |
10746 | } |
10747 | |
10748 | return I->isAssociative(); |
10749 | } |
10750 | |
10751 | static Value *getRdxOperand(Instruction *I, unsigned Index) { |
10752 | // Poison-safe 'or' takes the form: select X, true, Y |
10753 | // To make that work with the normal operand processing, we skip the |
10754 | // true value operand. |
10755 | // TODO: Change the code and data structures to handle this without a hack. |
10756 | if (getRdxKind(I) == RecurKind::Or && isa<SelectInst>(I) && Index == 1) |
10757 | return I->getOperand(2); |
10758 | return I->getOperand(Index); |
10759 | } |
10760 | |
10761 | /// Creates reduction operation with the current opcode. |
10762 | static Value *createOp(IRBuilder<> &Builder, RecurKind Kind, Value *LHS, |
10763 | Value *RHS, const Twine &Name, bool UseSelect) { |
10764 | unsigned RdxOpcode = RecurrenceDescriptor::getOpcode(Kind); |
10765 | switch (Kind) { |
10766 | case RecurKind::Or: |
10767 | if (UseSelect && |
10768 | LHS->getType() == CmpInst::makeCmpResultType(LHS->getType())) |
10769 | return Builder.CreateSelect(LHS, Builder.getTrue(), RHS, Name); |
10770 | return Builder.CreateBinOp((Instruction::BinaryOps)RdxOpcode, LHS, RHS, |
10771 | Name); |
10772 | case RecurKind::And: |
10773 | if (UseSelect && |
10774 | LHS->getType() == CmpInst::makeCmpResultType(LHS->getType())) |
10775 | return Builder.CreateSelect(LHS, RHS, Builder.getFalse(), Name); |
10776 | return Builder.CreateBinOp((Instruction::BinaryOps)RdxOpcode, LHS, RHS, |
10777 | Name); |
10778 | case RecurKind::Add: |
10779 | case RecurKind::Mul: |
10780 | case RecurKind::Xor: |
10781 | case RecurKind::FAdd: |
10782 | case RecurKind::FMul: |
10783 | return Builder.CreateBinOp((Instruction::BinaryOps)RdxOpcode, LHS, RHS, |
10784 | Name); |
10785 | case RecurKind::FMax: |
10786 | return Builder.CreateBinaryIntrinsic(Intrinsic::maxnum, LHS, RHS); |
10787 | case RecurKind::FMin: |
10788 | return Builder.CreateBinaryIntrinsic(Intrinsic::minnum, LHS, RHS); |
10789 | case RecurKind::SMax: |
10790 | if (UseSelect) { |
10791 | Value *Cmp = Builder.CreateICmpSGT(LHS, RHS, Name); |
10792 | return Builder.CreateSelect(Cmp, LHS, RHS, Name); |
10793 | } |
10794 | return Builder.CreateBinaryIntrinsic(Intrinsic::smax, LHS, RHS); |
10795 | case RecurKind::SMin: |
10796 | if (UseSelect) { |
10797 | Value *Cmp = Builder.CreateICmpSLT(LHS, RHS, Name); |
10798 | return Builder.CreateSelect(Cmp, LHS, RHS, Name); |
10799 | } |
10800 | return Builder.CreateBinaryIntrinsic(Intrinsic::smin, LHS, RHS); |
10801 | case RecurKind::UMax: |
10802 | if (UseSelect) { |
10803 | Value *Cmp = Builder.CreateICmpUGT(LHS, RHS, Name); |
10804 | return Builder.CreateSelect(Cmp, LHS, RHS, Name); |
10805 | } |
10806 | return Builder.CreateBinaryIntrinsic(Intrinsic::umax, LHS, RHS); |
10807 | case RecurKind::UMin: |
10808 | if (UseSelect) { |
10809 | Value *Cmp = Builder.CreateICmpULT(LHS, RHS, Name); |
10810 | return Builder.CreateSelect(Cmp, LHS, RHS, Name); |
10811 | } |
10812 | return Builder.CreateBinaryIntrinsic(Intrinsic::umin, LHS, RHS); |
10813 | default: |
10814 | llvm_unreachable("Unknown reduction operation.")::llvm::llvm_unreachable_internal("Unknown reduction operation." , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 10814); |
10815 | } |
10816 | } |
10817 | |
10818 | /// Creates reduction operation with the current opcode with the IR flags |
10819 | /// from \p ReductionOps, dropping nuw/nsw flags. |
10820 | static Value *createOp(IRBuilder<> &Builder, RecurKind RdxKind, Value *LHS, |
10821 | Value *RHS, const Twine &Name, |
10822 | const ReductionOpsListType &ReductionOps) { |
10823 | bool UseSelect = ReductionOps.size() == 2 || |
10824 | // Logical or/and. |
10825 | (ReductionOps.size() == 1 && |
10826 | isa<SelectInst>(ReductionOps.front().front())); |
10827 | 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", 10829, __extension__ __PRETTY_FUNCTION__)) |
10828 | 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", 10829, __extension__ __PRETTY_FUNCTION__)) |
10829 | "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", 10829, __extension__ __PRETTY_FUNCTION__)); |
10830 | Value *Op = createOp(Builder, RdxKind, LHS, RHS, Name, UseSelect); |
10831 | if (RecurrenceDescriptor::isIntMinMaxRecurrenceKind(RdxKind)) { |
10832 | if (auto *Sel = dyn_cast<SelectInst>(Op)) { |
10833 | propagateIRFlags(Sel->getCondition(), ReductionOps[0], nullptr, |
10834 | /*IncludeWrapFlags=*/false); |
10835 | propagateIRFlags(Op, ReductionOps[1], nullptr, |
10836 | /*IncludeWrapFlags=*/false); |
10837 | return Op; |
10838 | } |
10839 | } |
10840 | propagateIRFlags(Op, ReductionOps[0], nullptr, /*IncludeWrapFlags=*/false); |
10841 | return Op; |
10842 | } |
10843 | |
10844 | static RecurKind getRdxKind(Value *V) { |
10845 | auto *I = dyn_cast<Instruction>(V); |
10846 | if (!I) |
10847 | return RecurKind::None; |
10848 | if (match(I, m_Add(m_Value(), m_Value()))) |
10849 | return RecurKind::Add; |
10850 | if (match(I, m_Mul(m_Value(), m_Value()))) |
10851 | return RecurKind::Mul; |
10852 | if (match(I, m_And(m_Value(), m_Value())) || |
10853 | match(I, m_LogicalAnd(m_Value(), m_Value()))) |
10854 | return RecurKind::And; |
10855 | if (match(I, m_Or(m_Value(), m_Value())) || |
10856 | match(I, m_LogicalOr(m_Value(), m_Value()))) |
10857 | return RecurKind::Or; |
10858 | if (match(I, m_Xor(m_Value(), m_Value()))) |
10859 | return RecurKind::Xor; |
10860 | if (match(I, m_FAdd(m_Value(), m_Value()))) |
10861 | return RecurKind::FAdd; |
10862 | if (match(I, m_FMul(m_Value(), m_Value()))) |
10863 | return RecurKind::FMul; |
10864 | |
10865 | if (match(I, m_Intrinsic<Intrinsic::maxnum>(m_Value(), m_Value()))) |
10866 | return RecurKind::FMax; |
10867 | if (match(I, m_Intrinsic<Intrinsic::minnum>(m_Value(), m_Value()))) |
10868 | return RecurKind::FMin; |
10869 | |
10870 | // This matches either cmp+select or intrinsics. SLP is expected to handle |
10871 | // either form. |
10872 | // TODO: If we are canonicalizing to intrinsics, we can remove several |
10873 | // special-case paths that deal with selects. |
10874 | if (match(I, m_SMax(m_Value(), m_Value()))) |
10875 | return RecurKind::SMax; |
10876 | if (match(I, m_SMin(m_Value(), m_Value()))) |
10877 | return RecurKind::SMin; |
10878 | if (match(I, m_UMax(m_Value(), m_Value()))) |
10879 | return RecurKind::UMax; |
10880 | if (match(I, m_UMin(m_Value(), m_Value()))) |
10881 | return RecurKind::UMin; |
10882 | |
10883 | if (auto *Select = dyn_cast<SelectInst>(I)) { |
10884 | // Try harder: look for min/max pattern based on instructions producing |
10885 | // same values such as: select ((cmp Inst1, Inst2), Inst1, Inst2). |
10886 | // During the intermediate stages of SLP, it's very common to have |
10887 | // pattern like this (since optimizeGatherSequence is run only once |
10888 | // at the end): |
10889 | // %1 = extractelement <2 x i32> %a, i32 0 |
10890 | // %2 = extractelement <2 x i32> %a, i32 1 |
10891 | // %cond = icmp sgt i32 %1, %2 |
10892 | // %3 = extractelement <2 x i32> %a, i32 0 |
10893 | // %4 = extractelement <2 x i32> %a, i32 1 |
10894 | // %select = select i1 %cond, i32 %3, i32 %4 |
10895 | CmpInst::Predicate Pred; |
10896 | Instruction *L1; |
10897 | Instruction *L2; |
10898 | |
10899 | Value *LHS = Select->getTrueValue(); |
10900 | Value *RHS = Select->getFalseValue(); |
10901 | Value *Cond = Select->getCondition(); |
10902 | |
10903 | // TODO: Support inverse predicates. |
10904 | if (match(Cond, m_Cmp(Pred, m_Specific(LHS), m_Instruction(L2)))) { |
10905 | if (!isa<ExtractElementInst>(RHS) || |
10906 | !L2->isIdenticalTo(cast<Instruction>(RHS))) |
10907 | return RecurKind::None; |
10908 | } else if (match(Cond, m_Cmp(Pred, m_Instruction(L1), m_Specific(RHS)))) { |
10909 | if (!isa<ExtractElementInst>(LHS) || |
10910 | !L1->isIdenticalTo(cast<Instruction>(LHS))) |
10911 | return RecurKind::None; |
10912 | } else { |
10913 | if (!isa<ExtractElementInst>(LHS) || !isa<ExtractElementInst>(RHS)) |
10914 | return RecurKind::None; |
10915 | if (!match(Cond, m_Cmp(Pred, m_Instruction(L1), m_Instruction(L2))) || |
10916 | !L1->isIdenticalTo(cast<Instruction>(LHS)) || |
10917 | !L2->isIdenticalTo(cast<Instruction>(RHS))) |
10918 | return RecurKind::None; |
10919 | } |
10920 | |
10921 | switch (Pred) { |
10922 | default: |
10923 | return RecurKind::None; |
10924 | case CmpInst::ICMP_SGT: |
10925 | case CmpInst::ICMP_SGE: |
10926 | return RecurKind::SMax; |
10927 | case CmpInst::ICMP_SLT: |
10928 | case CmpInst::ICMP_SLE: |
10929 | return RecurKind::SMin; |
10930 | case CmpInst::ICMP_UGT: |
10931 | case CmpInst::ICMP_UGE: |
10932 | return RecurKind::UMax; |
10933 | case CmpInst::ICMP_ULT: |
10934 | case CmpInst::ICMP_ULE: |
10935 | return RecurKind::UMin; |
10936 | } |
10937 | } |
10938 | return RecurKind::None; |
10939 | } |
10940 | |
10941 | /// Get the index of the first operand. |
10942 | static unsigned getFirstOperandIndex(Instruction *I) { |
10943 | return isCmpSelMinMax(I) ? 1 : 0; |
10944 | } |
10945 | |
10946 | /// Total number of operands in the reduction operation. |
10947 | static unsigned getNumberOfOperands(Instruction *I) { |
10948 | return isCmpSelMinMax(I) ? 3 : 2; |
10949 | } |
10950 | |
10951 | /// Checks if the instruction is in basic block \p BB. |
10952 | /// For a cmp+sel min/max reduction check that both ops are in \p BB. |
10953 | static bool hasSameParent(Instruction *I, BasicBlock *BB) { |
10954 | if (isCmpSelMinMax(I) || isBoolLogicOp(I)) { |
10955 | auto *Sel = cast<SelectInst>(I); |
10956 | auto *Cmp = dyn_cast<Instruction>(Sel->getCondition()); |
10957 | return Sel->getParent() == BB && Cmp && Cmp->getParent() == BB; |
10958 | } |
10959 | return I->getParent() == BB; |
10960 | } |
10961 | |
10962 | /// Expected number of uses for reduction operations/reduced values. |
10963 | static bool hasRequiredNumberOfUses(bool IsCmpSelMinMax, Instruction *I) { |
10964 | if (IsCmpSelMinMax) { |
10965 | // SelectInst must be used twice while the condition op must have single |
10966 | // use only. |
10967 | if (auto *Sel = dyn_cast<SelectInst>(I)) |
10968 | return Sel->hasNUses(2) && Sel->getCondition()->hasOneUse(); |
10969 | return I->hasNUses(2); |
10970 | } |
10971 | |
10972 | // Arithmetic reduction operation must be used once only. |
10973 | return I->hasOneUse(); |
10974 | } |
10975 | |
10976 | /// Initializes the list of reduction operations. |
10977 | void initReductionOps(Instruction *I) { |
10978 | if (isCmpSelMinMax(I)) |
10979 | ReductionOps.assign(2, ReductionOpsType()); |
10980 | else |
10981 | ReductionOps.assign(1, ReductionOpsType()); |
10982 | } |
10983 | |
10984 | /// Add all reduction operations for the reduction instruction \p I. |
10985 | void addReductionOps(Instruction *I) { |
10986 | if (isCmpSelMinMax(I)) { |
10987 | ReductionOps[0].emplace_back(cast<SelectInst>(I)->getCondition()); |
10988 | ReductionOps[1].emplace_back(I); |
10989 | } else { |
10990 | ReductionOps[0].emplace_back(I); |
10991 | } |
10992 | } |
10993 | |
10994 | static Value *getLHS(RecurKind Kind, Instruction *I) { |
10995 | if (Kind == RecurKind::None) |
10996 | return nullptr; |
10997 | return I->getOperand(getFirstOperandIndex(I)); |
10998 | } |
10999 | static Value *getRHS(RecurKind Kind, Instruction *I) { |
11000 | if (Kind == RecurKind::None) |
11001 | return nullptr; |
11002 | return I->getOperand(getFirstOperandIndex(I) + 1); |
11003 | } |
11004 | |
11005 | public: |
11006 | HorizontalReduction() = default; |
11007 | |
11008 | /// Try to find a reduction tree. |
11009 | bool matchAssociativeReduction(PHINode *Phi, Instruction *Inst, |
11010 | ScalarEvolution &SE, const DataLayout &DL, |
11011 | const TargetLibraryInfo &TLI) { |
11012 | assert((!Phi || is_contained(Phi->operands(), Inst)) &&(static_cast <bool> ((!Phi || is_contained(Phi->operands (), Inst)) && "Phi needs to use the binary operator") ? void (0) : __assert_fail ("(!Phi || is_contained(Phi->operands(), Inst)) && \"Phi needs to use the binary operator\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 11013, __extension__ __PRETTY_FUNCTION__)) |
11013 | "Phi needs to use the binary operator")(static_cast <bool> ((!Phi || is_contained(Phi->operands (), Inst)) && "Phi needs to use the binary operator") ? void (0) : __assert_fail ("(!Phi || is_contained(Phi->operands(), Inst)) && \"Phi needs to use the binary operator\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 11013, __extension__ __PRETTY_FUNCTION__)); |
11014 | assert((isa<BinaryOperator>(Inst) || isa<SelectInst>(Inst) ||(static_cast <bool> ((isa<BinaryOperator>(Inst) || isa<SelectInst>(Inst) || isa<IntrinsicInst>(Inst )) && "Expected binop, select, or intrinsic for reduction matching" ) ? void (0) : __assert_fail ("(isa<BinaryOperator>(Inst) || isa<SelectInst>(Inst) || isa<IntrinsicInst>(Inst)) && \"Expected binop, select, or intrinsic for reduction matching\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 11016, __extension__ __PRETTY_FUNCTION__)) |
11015 | isa<IntrinsicInst>(Inst)) &&(static_cast <bool> ((isa<BinaryOperator>(Inst) || isa<SelectInst>(Inst) || isa<IntrinsicInst>(Inst )) && "Expected binop, select, or intrinsic for reduction matching" ) ? void (0) : __assert_fail ("(isa<BinaryOperator>(Inst) || isa<SelectInst>(Inst) || isa<IntrinsicInst>(Inst)) && \"Expected binop, select, or intrinsic for reduction matching\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 11016, __extension__ __PRETTY_FUNCTION__)) |
11016 | "Expected binop, select, or intrinsic for reduction matching")(static_cast <bool> ((isa<BinaryOperator>(Inst) || isa<SelectInst>(Inst) || isa<IntrinsicInst>(Inst )) && "Expected binop, select, or intrinsic for reduction matching" ) ? void (0) : __assert_fail ("(isa<BinaryOperator>(Inst) || isa<SelectInst>(Inst) || isa<IntrinsicInst>(Inst)) && \"Expected binop, select, or intrinsic for reduction matching\"" , "llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp", 11016, __extension__ __PRETTY_FUNCTION__)); |
11017 | RdxKind = getRdxKind(Inst); |
11018 | |
11019 | // We could have a initial reductions that is not an add. |
11020 | // r *= v1 + v2 + v3 + v4 |
11021 | // In such a case start looking for a tree rooted in the first '+'. |
11022 | if (Phi) { |
11023 | if (getLHS(RdxKind, Inst) == Phi) { |
11024 | Phi = nullptr; |
11025 | Inst = dyn_cast<Instruction>(getRHS(RdxKind, Inst)); |
11026 | if (!Inst) |
11027 | return false; |
11028 | RdxKind = getRdxKind(Inst); |
11029 | } else if (getRHS(RdxKind, Inst) == Phi) { |
11030 | Phi = nullptr; |
11031 | Inst = dyn_cast<Instruction>(getLHS(RdxKind, Inst)); |
11032 | if (!Inst) |
11033 | return false; |
11034 | RdxKind = getRdxKind(Inst); |
11035 | } |
11036 | } |
11037 | |
11038 | if (!isVectorizable(RdxKind, Inst)) |
11039 | return false; |
11040 | |
11041 | // Analyze "regular" integer/FP types for reductions - no target-specific |
11042 | // types or pointers. |
11043 | Type *Ty = Inst->getType(); |
11044 | if (!isValidElementType(Ty) || Ty->isPointerTy()) |
11045 | return false; |
11046 | |
11047 | // Though the ultimate reduction may have multiple uses, its condition must |
11048 | // have only single use. |
11049 | if (auto *Sel = dyn_cast<SelectInst>(Inst)) |
11050 | if (!Sel->getCondition()->hasOneUse()) |
11051 | return false; |
11052 | |
11053 | ReductionRoot = Inst; |
11054 | |
11055 | // Iterate through all the operands of the possible reduction tree and |
11056 | // gather all the reduced values, sorting them by their value id. |
11057 | BasicBlock *BB = Inst->getParent(); |
11058 | bool IsCmpSelMinMax = isCmpSelMinMax(Inst); |
11059 | SmallVector<Instruction *> Worklist(1, Inst); |
11060 | // Checks if the operands of the \p TreeN instruction are also reduction |
11061 | // operations or should be treated as reduced values or an extra argument, |
11062 | // which is not part of the reduction. |
11063 | auto &&CheckOperands = [this, IsCmpSelMinMax, |
11064 | BB](Instruction *TreeN, |
11065 | SmallVectorImpl<Value *> &ExtraArgs, |
11066 | SmallVectorImpl<Value *> &PossibleReducedVals, |
11067 | SmallVectorImpl<Instruction *> &ReductionOps) { |
11068 | for (int I = getFirstOperandIndex(TreeN), |
11069 | End = getNumberOfOperands(TreeN); |
11070 | I < End; ++I) { |
11071 | Value *EdgeVal = getRdxOperand(TreeN, I); |
11072 | ReducedValsToOps[EdgeVal].push_back(TreeN); |
11073 | auto *EdgeInst = dyn_cast<Instruction>(EdgeVal); |
11074 | // Edge has wrong parent - mark as an extra argument. |
11075 | if (EdgeInst && !isVectorLikeInstWithConstOps(EdgeInst) && |
11076 | !hasSameParent(EdgeInst, BB)) { |
11077 | ExtraArgs.push_back(EdgeVal); |
11078 | continue; |
11079 | } |
11080 | // If the edge is not an instruction, or it is different from the main |
11081 | // reduction opcode or has too many uses - possible reduced value. |
11082 | if (!EdgeInst || getRdxKind(EdgeInst) != RdxKind || |
11083 | IsCmpSelMinMax != isCmpSelMinMax(EdgeInst) || |
11084 | !hasRequiredNumberOfUses(IsCmpSelMinMax, EdgeInst) || |
11085 | !isVectorizable(getRdxKind(EdgeInst), EdgeInst)) { |
11086 | PossibleReducedVals.push_back(EdgeVal); |
11087 | continue; |
11088 | } |
11089 | ReductionOps.push_back(EdgeInst); |
11090 | } |
11091 | }; |
11092 | // Try to regroup reduced values so that it gets more profitable to try to |
11093 | // reduce them. Values are grouped by their value ids, instructions - by |
11094 | // instruction op id and/or alternate op id, plus do extra analysis for |
11095 | // loads (grouping them by the distabce between pointers) and cmp |
11096 | // instructions (grouping them by the predicate). |
11097 | MapVector<size_t, MapVector<size_t, MapVector<Value *, unsigned>>> |
11098 | PossibleReducedVals; |
11099 | initReductionOps(Inst); |
11100 | while (!Worklist.empty()) { |
11101 | Instruction *TreeN = Worklist.pop_back_val(); |
11102 | SmallVector<Value *> Args; |
11103 | SmallVector<Value *> PossibleRedVals; |
11104 | SmallVector<Instruction *> PossibleReductionOps; |
11105 | CheckOperands(TreeN, Args, PossibleRedVals, PossibleReductionOps); |
11106 | // If too many extra args - mark the instruction itself as a reduction |
11107 | // value, not a reduction operation. |
11108 | if (Args.size() < 2) { |
11109 | addReductionOps(TreeN); |
11110 | // Add extra args. |
11111 | if (!Args.empty()) { |
11112 | 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", 11112, __extension__ __PRETTY_FUNCTION__)); |
11113 | ExtraArgs[TreeN] = Args.front(); |
11114 | } |
11115 | // Add reduction values. The values are sorted for better vectorization |
11116 | // results. |
11117 | for (Value *V : PossibleRedVals) { |
11118 | size_t Key, Idx; |
11119 | std::tie(Key, Idx) = generateKeySubkey( |
11120 | V, &TLI, |
11121 | [&PossibleReducedVals, &DL, &SE](size_t Key, LoadInst *LI) { |
11122 | auto It = PossibleReducedVals.find(Key); |
11123 | if (It != PossibleReducedVals.end()) { |
11124 | for (const auto &LoadData : It->second) { |
11125 | auto *RLI = cast<LoadInst>(LoadData.second.front().first); |
11126 | if (getPointersDiff(RLI->getType(), |
11127 | RLI->getPointerOperand(), LI->getType(), |
11128 | LI->getPointerOperand(), DL, SE, |
11129 | /*StrictCheck=*/true)) |
11130 | return hash_value(RLI->getPointerOperand()); |
11131 | } |
11132 | } |
11133 | return hash_value(LI->getPointerOperand()); |
11134 | }, |
11135 | /*AllowAlternate=*/false); |
11136 | ++PossibleReducedVals[Key][Idx] |
11137 | .insert(std::make_pair(V, 0)) |
11138 | .first->second; |
11139 | } |
11140 | Worklist.append(PossibleReductionOps.rbegin(), |
11141 | PossibleReductionOps.rend()); |
11142 | } else { |
11143 | size_t Key, Idx; |
11144 | std::tie(Key, Idx) = generateKeySubkey( |
11145 | TreeN, &TLI, |
11146 | [&PossibleReducedVals, &DL, &SE](size_t Key, LoadInst *LI) { |
11147 | auto It = PossibleReducedVals.find(Key); |
11148 | if (It != PossibleReducedVals.end()) { |
11149 | for (const auto &LoadData : It->second) { |
11150 | auto *RLI = cast<LoadInst>(LoadData.second.front().first); |
11151 | if (getPointersDiff(RLI->getType(), RLI->getPointerOperand(), |
11152 | LI->getType(), LI->getPointerOperand(), |
11153 | DL, SE, /*StrictCheck=*/true)) |
11154 | return hash_value(RLI->getPointerOperand()); |
11155 | } |
11156 | } |
11157 | return hash_value(LI->getPointerOperand()); |
11158 | }, |
11159 | /*AllowAlternate=*/false); |
11160 | ++PossibleReducedVals[Key][Idx] |
11161 | .insert(std::make_pair(TreeN, 0)) |
11162 | .first->second; |
11163 | } |
11164 | } |
11165 | auto PossibleReducedValsVect = PossibleReducedVals.takeVector(); |
11166 | // Sort values by the total number of values kinds to start the reduction |
11167 | // from the longest possible reduced values sequences. |
11168 | for (auto &PossibleReducedVals : PossibleReducedValsVect) { |
11169 | auto PossibleRedVals = PossibleReducedVals.second.takeVector(); |
11170 | SmallVector<SmallVector<Value *>> PossibleRedValsVect; |
11171 | for (auto It = PossibleRedVals.begin(), E = PossibleRedVals.end(); |
11172 | It != E; ++It) { |
11173 | PossibleRedValsVect.emplace_back(); |
11174 | auto RedValsVect = It->second.takeVector(); |
11175 | stable_sort(RedValsVect, llvm::less_second()); |
11176 | for (const std::pair<Value *, unsigned> &Data : RedValsVect) |
11177 | PossibleRedValsVect.back().append(Data.second, Data.first); |
11178 | } |
11179 | stable_sort(PossibleRedValsVect, [](const auto &P1, const auto &P2) { |
11180 | return P1.size() > P2.size(); |
11181 | }); |
11182 | ReducedVals.emplace_back(); |
11183 | for (ArrayRef<Value *> Data : PossibleRedValsVect) |
11184 | ReducedVals.back().append(Data.rbegin(), Data.rend()); |
11185 | } |
11186 | // Sort the reduced values by number of same/alternate opcode and/or pointer |
11187 | // operand. |
11188 | stable_sort(ReducedVals, [](ArrayRef<Value *> P1, ArrayRef<Value *> P2) { |
11189 | return P1.size() > P2.size(); |
11190 | }); |
11191 | return true; |
11192 | } |
11193 | |
11194 | /// Attempt to vectorize the tree found by matchAssociativeReduction. |
11195 | Value *tryToReduce(BoUpSLP &V, TargetTransformInfo *TTI) { |
11196 | constexpr int ReductionLimit = 4; |
11197 | constexpr unsigned RegMaxNumber = 4; |
11198 | constexpr unsigned RedValsMaxNumber = 128; |
11199 | // If there are a sufficient number of reduction values, reduce |
11200 | // to a nearby power-of-2. We can safely generate oversized |
11201 | // vectors and rely on the backend to split them to legal sizes. |
11202 | unsigned NumReducedVals = std::accumulate( |
11203 | ReducedVals.begin(), ReducedVals.end(), 0, |
11204 | [](int Num, ArrayRef<Value *> Vals) { return Num + Vals.size(); }); |
11205 | if (NumReducedVals < ReductionLimit) |
11206 | return nullptr; |
11207 | |
11208 | IRBuilder<> Builder(cast<Instruction>(ReductionRoot)); |
11209 | |
11210 | // Track the reduced values in case if they are replaced by extractelement |
11211 | // because of the vectorization. |
11212 | DenseMap<Value *, WeakTrackingVH> TrackedVals; |
11213 | BoUpSLP::ExtraValueToDebugLocsMap ExternallyUsedValues; |
11214 | // The same extra argument may be used several times, so log each attempt |
11215 | // to use it. |
11216 | for (const std::pair<Instruction *, Value *> &Pair : ExtraArgs) { |
11217 | 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", 11217, __extension__ __PRETTY_FUNCTION__)); |
11218 | ExternallyUsedValues[Pair.second].push_back(Pair.first); |
11219 | TrackedVals.try_emplace(Pair.second, Pair.second); |
11220 | } |
11221 | |
11222 | // The compare instruction of a min/max is the insertion point for new |
11223 | // instructions and may be replaced with a new compare instruction. |
11224 | auto &&GetCmpForMinMaxReduction = [](Instruction *RdxRootInst) { |
11225 | 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", 11226, __extension__ __PRETTY_FUNCTION__)) |
11226 | "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", 11226, __extension__ __PRETTY_FUNCTION__)); |
11227 | Value *ScalarCond = cast<SelectInst>(RdxRootInst)->getCondition(); |
11228 | 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", 11229, __extension__ __PRETTY_FUNCTION__)) |
11229 | "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", 11229, __extension__ __PRETTY_FUNCTION__)); |
11230 | return cast<Instruction>(ScalarCond); |
11231 | }; |
11232 | |
11233 | // The reduction root is used as the insertion point for new instructions, |
11234 | // so set it as externally used to prevent it from being deleted. |
11235 | ExternallyUsedValues[ReductionRoot]; |
11236 | SmallDenseSet<Value *> IgnoreList; |
11237 | for (ReductionOpsType &RdxOps : ReductionOps) |
11238 | for (Value *RdxOp : RdxOps) { |
11239 | if (!RdxOp) |
11240 | continue; |
11241 | IgnoreList.insert(RdxOp); |
11242 | } |
11243 | bool IsCmpSelMinMax = isCmpSelMinMax(cast<Instruction>(ReductionRoot)); |
11244 | |
11245 | // Need to track reduced vals, they may be changed during vectorization of |
11246 | // subvectors. |
11247 | for (ArrayRef<Value *> Candidates : ReducedVals) |
11248 | for (Value *V : Candidates) |
11249 | TrackedVals.try_emplace(V, V); |
11250 | |
11251 | DenseMap<Value *, unsigned> VectorizedVals; |
11252 | Value *VectorizedTree = nullptr; |
11253 | bool CheckForReusedReductionOps = false; |
11254 | // Try to vectorize elements based on their type. |
11255 | for (unsigned I = 0, E = ReducedVals.size(); I < E; ++I) { |
11256 | ArrayRef<Value *> OrigReducedVals = ReducedVals[I]; |
11257 | InstructionsState S = getSameOpcode(OrigReducedVals); |
11258 | SmallVector<Value *> Candidates; |
11259 | DenseMap<Value *, Value *> TrackedToOrig; |
11260 | for (unsigned Cnt = 0, Sz = OrigReducedVals.size(); Cnt < Sz; ++Cnt) { |
11261 | Value *RdxVal = TrackedVals.find(OrigReducedVals[Cnt])->second; |
11262 | // Check if the reduction value was not overriden by the extractelement |
11263 | // instruction because of the vectorization and exclude it, if it is not |
11264 | // compatible with other values. |
11265 | if (auto *Inst = dyn_cast<Instruction>(RdxVal)) |
11266 | if (isVectorLikeInstWithConstOps(Inst) && |
11267 | (!S.getOpcode() || !S.isOpcodeOrAlt(Inst))) |
11268 | continue; |
11269 | Candidates.push_back(RdxVal); |
11270 | TrackedToOrig.try_emplace(RdxVal, OrigReducedVals[Cnt]); |
11271 | } |
11272 | bool ShuffledExtracts = false; |
11273 | // Try to handle shuffled extractelements. |
11274 | if (S.getOpcode() == Instruction::ExtractElement && !S.isAltShuffle() && |
11275 | I + 1 < E) { |
11276 | InstructionsState NextS = getSameOpcode(ReducedVals[I + 1]); |
11277 | if (NextS.getOpcode() == Instruction::ExtractElement && |
11278 | !NextS.isAltShuffle()) { |
11279 | SmallVector<Value *> CommonCandidates(Candidates); |
11280 | for (Value *RV : ReducedVals[I + 1]) { |
11281 | Value *RdxVal = TrackedVals.find(RV)->second; |
11282 | // Check if the reduction value was not overriden by the |
11283 | // extractelement instruction because of the vectorization and |
11284 | // exclude it, if it is not compatible with other values. |
11285 | if (auto *Inst = dyn_cast<Instruction>(RdxVal)) |
11286 | if (!NextS.getOpcode() || !NextS.isOpcodeOrAlt(Inst)) |
11287 | continue; |
11288 | CommonCandidates.push_back(RdxVal); |
11289 | TrackedToOrig.try_emplace(RdxVal, RV); |
11290 | } |
11291 | SmallVector<int> Mask; |
11292 | if (isFixedVectorShuffle(CommonCandidates, Mask)) { |
11293 | ++I; |
11294 | Candidates.swap(CommonCandidates); |
11295 | ShuffledExtracts = true; |
11296 | } |
11297 | } |
11298 | } |
11299 | unsigned NumReducedVals = Candidates.size(); |
11300 | if (NumReducedVals < ReductionLimit) |
11301 | continue; |
11302 | |
11303 | unsigned MaxVecRegSize = V.getMaxVecRegSize(); |
11304 | unsigned EltSize = V.getVectorElementSize(Candidates[0]); |
11305 | unsigned MaxElts = RegMaxNumber * PowerOf2Floor(MaxVecRegSize / EltSize); |
11306 | |
11307 | unsigned ReduxWidth = std::min<unsigned>( |
11308 | PowerOf2Floor(NumReducedVals), std::max(RedValsMaxNumber, MaxElts)); |
11309 | unsigned Start = 0; |
11310 | unsigned Pos = Start; |
11311 | // Restarts vectorization attempt with lower vector factor. |
11312 | unsigned PrevReduxWidth = ReduxWidth; |
11313 | bool CheckForReusedReductionOpsLocal = false; |
11314 | auto &&AdjustReducedVals = [&Pos, &Start, &ReduxWidth, NumReducedVals, |
11315 | &CheckForReusedReductionOpsLocal, |
11316 | &PrevReduxWidth, &V, |
11317 | &IgnoreList](bool IgnoreVL = false) { |
11318 | bool IsAnyRedOpGathered = !IgnoreVL && V.isAnyGathered(IgnoreList); |
11319 | if (!CheckForReusedReductionOpsLocal && PrevReduxWidth == ReduxWidth) { |
11320 | // Check if any of the reduction ops are gathered. If so, worth |
11321 | // trying again with less number of reduction ops. |
11322 | CheckForReusedReductionOpsLocal |= IsAnyRedOpGathered; |
11323 | } |
11324 | ++Pos; |
11325 | if (Pos < NumReducedVals - ReduxWidth + 1) |
11326 | return IsAnyRedOpGathered; |
11327 | Pos = Start; |
11328 | ReduxWidth /= 2; |
11329 | return IsAnyRedOpGathered; |
11330 | }; |
11331 | while (Pos < NumReducedVals - ReduxWidth + 1 && |
11332 | ReduxWidth >= ReductionLimit) { |
11333 | // Dependency in tree of the reduction ops - drop this attempt, try |
11334 | // later. |
11335 | if (CheckForReusedReductionOpsLocal && PrevReduxWidth != ReduxWidth && |
11336 | Start == 0) { |
11337 | CheckForReusedReductionOps = true; |
11338 | break; |
11339 | } |
11340 | PrevReduxWidth = ReduxWidth; |
11341 | ArrayRef<Value *> VL(std::next(Candidates.begin(), Pos), ReduxWidth); |
11342 | // Beeing analyzed already - skip. |
11343 | if (V.areAnalyzedReductionVals(VL)) { |
11344 | (void)AdjustReducedVals(/*IgnoreVL=*/true); |
11345 | continue; |
11346 | } |
11347 | // Early exit if any of the reduction values were deleted during |
11348 | // previous vectorization attempts. |
11349 | if (any_of(VL, [&V](Value *RedVal) { |
11350 | auto *RedValI = dyn_cast<Instruction>(RedVal); |
11351 | if (!RedValI) |
11352 | return false; |
11353 | return V.isDeleted(RedValI); |
11354 | })) |
11355 | break; |
11356 | V.buildTree(VL, IgnoreList); |
11357 | if (V.isTreeTinyAndNotFullyVectorizable(/*ForReduction=*/true)) { |
11358 | if (!AdjustReducedVals()) |
11359 | V.analyzedReductionVals(VL); |
11360 | continue; |
11361 | } |
11362 | if (V.isLoadCombineReductionCandidate(RdxKind)) { |
11363 | if (!AdjustReducedVals()) |
11364 | V.analyzedReductionVals(VL); |
11365 | continue; |
11366 | } |
11367 | V.reorderTopToBottom(); |
11368 | // No need to reorder the root node at all. |
11369 | V.reorderBottomToTop(/*IgnoreReorder=*/true); |
11370 | // Keep extracted other reduction values, if they are used in the |
11371 | // vectorization trees. |
11372 | BoUpSLP::ExtraValueToDebugLocsMap LocalExternallyUsedValues( |
11373 | ExternallyUsedValues); |
11374 | for (unsigned Cnt = 0, Sz = ReducedVals.size(); Cnt < Sz; ++Cnt) { |
11375 | if (Cnt == I || (ShuffledExtracts && Cnt == I - 1)) |
11376 | continue; |
11377 | for_each(ReducedVals[Cnt], |
11378 | [&LocalExternallyUsedValues, &TrackedVals](Value *V) { |
11379 | if (isa<Instruction>(V)) |
11380 | LocalExternallyUsedValues[TrackedVals[V]]; |
11381 | }); |
11382 | } |
11383 | // Number of uses of the candidates in the vector of values. |
11384 | SmallDenseMap<Value *, unsigned> NumUses; |
11385 | for (unsigned Cnt = 0; Cnt < Pos; ++Cnt) { |
11386 | Value *V = Candidates[Cnt]; |
11387 | if (NumUses.count(V) > 0) |
11388 | continue; |
11389 | NumUses[V] = std::count(VL.begin(), VL.end(), V); |
11390 | } |
11391 | for (unsigned Cnt = Pos + ReduxWidth; Cnt < NumReducedVals; ++Cnt) { |
11392 | Value *V = Candidates[Cnt]; |
11393 | if (NumUses.count(V) > 0) |
11394 | continue; |
11395 | NumUses[V] = std::count(VL.begin(), VL.end(), V); |
11396 | } |
11397 | // Gather externally used values. |
11398 | SmallPtrSet<Value *, 4> Visited; |
11399 | for (unsigned Cnt = 0; Cnt < Pos; ++Cnt) { |
11400 | Value *V = Candidates[Cnt]; |
11401 | if (!Visited.insert(V).second) |
11402 | continue; |
11403 | unsigned NumOps = VectorizedVals.lookup(V) + NumUses[V]; |
11404 | if (NumOps != ReducedValsToOps.find(V)->second.size()) |
11405 | LocalExternallyUsedValues[V]; |
11406 | } |
11407 | for (unsigned Cnt = Pos + ReduxWidth; Cnt < NumReducedVals; ++Cnt) { |
11408 | Value *V = Candidates[Cnt]; |
11409 | if (!Visited.insert(V).second) |
11410 | continue; |
11411 | unsigned NumOps = VectorizedVals.lookup(V) + NumUses[V]; |
11412 | if (NumOps != ReducedValsToOps.find(V)->second.size()) |
11413 | LocalExternallyUsedValues[V]; |
11414 | } |
11415 | V.buildExternalUses(LocalExternallyUsedValues); |
11416 | |
11417 | V.computeMinimumValueSizes(); |
11418 | |
11419 | // Intersect the fast-math-flags from all reduction operations. |
11420 | FastMathFlags RdxFMF; |
11421 | RdxFMF.set(); |
11422 | for (Value *U : IgnoreList) |
11423 | if (auto *FPMO = dyn_cast<FPMathOperator>(U)) |
11424 | RdxFMF &= FPMO->getFastMathFlags(); |
11425 | // Estimate cost. |
11426 | InstructionCost TreeCost = V.getTreeCost(VL); |
11427 | InstructionCost ReductionCost = |
11428 | getReductionCost(TTI, VL, ReduxWidth, RdxFMF); |
11429 | InstructionCost Cost = TreeCost + ReductionCost; |
11430 | 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); |
11431 | if (!Cost.isValid()) { |
11432 | return nullptr; |
11433 | } |
11434 | if (Cost >= -SLPCostThreshold) { |
11435 | V.getORE()->emit([&]() { |
11436 | return OptimizationRemarkMissed( |
11437 | SV_NAME"slp-vectorizer", "HorSLPNotBeneficial", |
11438 | ReducedValsToOps.find(VL[0])->second.front()) |
11439 | << "Vectorizing horizontal reduction is possible " |
11440 | << "but not beneficial with cost " << ore::NV("Cost", Cost) |
11441 | << " and threshold " |
11442 | << ore::NV("Threshold", -SLPCostThreshold); |
11443 | }); |
11444 | if (!AdjustReducedVals()) |
11445 | V.analyzedReductionVals(VL); |
11446 | continue; |
11447 | } |
11448 | |
11449 | 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) |
11450 | << Cost << ". (HorRdx)\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Vectorizing horizontal reduction at cost:" << Cost << ". (HorRdx)\n"; } } while (false); |
11451 | V.getORE()->emit([&]() { |
11452 | return OptimizationRemark( |
11453 | SV_NAME"slp-vectorizer", "VectorizedHorizontalReduction", |
11454 | ReducedValsToOps.find(VL[0])->second.front()) |
11455 | << "Vectorized horizontal reduction with cost " |
11456 | << ore::NV("Cost", Cost) << " and with tree size " |
11457 | << ore::NV("TreeSize", V.getTreeSize()); |
11458 | }); |
11459 | |
11460 | Builder.setFastMathFlags(RdxFMF); |
11461 | |
11462 | // Vectorize a tree. |
11463 | Value *VectorizedRoot = V.vectorizeTree(LocalExternallyUsedValues); |
11464 | |
11465 | // Emit a reduction. If the root is a select (min/max idiom), the insert |
11466 | // point is the compare condition of that select. |
11467 | Instruction *RdxRootInst = cast<Instruction>(ReductionRoot); |
11468 | if (IsCmpSelMinMax) |
11469 | Builder.SetInsertPoint(GetCmpForMinMaxReduction(RdxRootInst)); |
11470 | else |
11471 | Builder.SetInsertPoint(RdxRootInst); |
11472 | |
11473 | // To prevent poison from leaking across what used to be sequential, |
11474 | // safe, scalar boolean logic operations, the reduction operand must be |
11475 | // frozen. |
11476 | if (isBoolLogicOp(RdxRootInst)) |
11477 | VectorizedRoot = Builder.CreateFreeze(VectorizedRoot); |
11478 | |
11479 | Value *ReducedSubTree = |
11480 | emitReduction(VectorizedRoot, Builder, ReduxWidth, TTI); |
11481 | |
11482 | if (!VectorizedTree) { |
11483 | // Initialize the final value in the reduction. |
11484 | VectorizedTree = ReducedSubTree; |
11485 | } else { |
11486 | // Update the final value in the reduction. |
11487 | Builder.SetCurrentDebugLocation( |
11488 | cast<Instruction>(ReductionOps.front().front())->getDebugLoc()); |
11489 | VectorizedTree = createOp(Builder, RdxKind, VectorizedTree, |
11490 | ReducedSubTree, "op.rdx", ReductionOps); |
11491 | } |
11492 | // Count vectorized reduced values to exclude them from final reduction. |
11493 | for (Value *V : VL) |
11494 | ++VectorizedVals.try_emplace(TrackedToOrig.find(V)->second, 0) |
11495 | .first->getSecond(); |
11496 | Pos += ReduxWidth; |
11497 | Start = Pos; |
11498 | ReduxWidth = PowerOf2Floor(NumReducedVals - Pos); |
11499 | } |
11500 | } |
11501 | if (VectorizedTree) { |
11502 | // Reorder operands of bool logical op in the natural order to avoid |
11503 | // possible problem with poison propagation. If not possible to reorder |
11504 | // (both operands are originally RHS), emit an extra freeze instruction |
11505 | // for the LHS operand. |
11506 | //I.e., if we have original code like this: |
11507 | // RedOp1 = select i1 ?, i1 LHS, i1 false |
11508 | // RedOp2 = select i1 RHS, i1 ?, i1 false |
11509 | |
11510 | // Then, we swap LHS/RHS to create a new op that matches the poison |
11511 | // semantics of the original code. |
11512 | |
11513 | // If we have original code like this and both values could be poison: |
11514 | // RedOp1 = select i1 ?, i1 LHS, i1 false |
11515 | // RedOp2 = select i1 ?, i1 RHS, i1 false |
11516 | |
11517 | // Then, we must freeze LHS in the new op. |
11518 | auto &&FixBoolLogicalOps = |
11519 | [&Builder, VectorizedTree](Value *&LHS, Value *&RHS, |
11520 | Instruction *RedOp1, Instruction *RedOp2) { |
11521 | if (!isBoolLogicOp(RedOp1)) |
11522 | return; |
11523 | if (LHS == VectorizedTree || getRdxOperand(RedOp1, 0) == LHS || |
11524 | isGuaranteedNotToBePoison(LHS)) |
11525 | return; |
11526 | if (!isBoolLogicOp(RedOp2)) |
11527 | return; |
11528 | if (RHS == VectorizedTree || getRdxOperand(RedOp2, 0) == RHS || |
11529 | isGuaranteedNotToBePoison(RHS)) { |
11530 | std::swap(LHS, RHS); |
11531 | return; |
11532 | } |
11533 | LHS = Builder.CreateFreeze(LHS); |
11534 | }; |
11535 | // Finish the reduction. |
11536 | // Need to add extra arguments and not vectorized possible reduction |
11537 | // values. |
11538 | // Try to avoid dependencies between the scalar remainders after |
11539 | // reductions. |
11540 | auto &&FinalGen = |
11541 | [this, &Builder, &TrackedVals, &FixBoolLogicalOps]( |
11542 | ArrayRef<std::pair<Instruction *, Value *>> InstVals) { |
11543 | unsigned Sz = InstVals.size(); |
11544 | SmallVector<std::pair<Instruction *, Value *>> ExtraReds(Sz / 2 + |
11545 | Sz % 2); |
11546 | for (unsigned I = 0, E = (Sz / 2) * 2; I < E; I += 2) { |
11547 | Instruction *RedOp = InstVals[I + 1].first; |
11548 | Builder.SetCurrentDebugLocation(RedOp->getDebugLoc()); |
11549 | Value *RdxVal1 = InstVals[I].second; |
11550 | Value *StableRdxVal1 = RdxVal1; |
11551 | auto It1 = TrackedVals.find(RdxVal1); |
11552 | if (It1 != TrackedVals.end()) |
11553 | StableRdxVal1 = It1->second; |
11554 | Value *RdxVal2 = InstVals[I + 1].second; |
11555 | Value *StableRdxVal2 = RdxVal2; |
11556 | auto It2 = TrackedVals.find(RdxVal2); |
11557 | if (It2 != TrackedVals.end()) |
11558 | StableRdxVal2 = It2->second; |
11559 | // To prevent poison from leaking across what used to be |
11560 | // sequential, safe, scalar boolean logic operations, the |
11561 | // reduction operand must be frozen. |
11562 | FixBoolLogicalOps(StableRdxVal1, StableRdxVal2, InstVals[I].first, |
11563 | RedOp); |
11564 | Value *ExtraRed = createOp(Builder, RdxKind, StableRdxVal1, |
11565 | StableRdxVal2, "op.rdx", ReductionOps); |
11566 | ExtraReds[I / 2] = std::make_pair(InstVals[I].first, ExtraRed); |
11567 | } |
11568 | if (Sz % 2 == 1) |
11569 | ExtraReds[Sz / 2] = InstVals.back(); |
11570 | return ExtraReds; |
11571 | }; |
11572 | SmallVector<std::pair<Instruction *, Value *>> ExtraReductions; |
11573 | ExtraReductions.emplace_back(cast<Instruction>(ReductionRoot), |
11574 | VectorizedTree); |
11575 | SmallPtrSet<Value *, 8> Visited; |
11576 | for (ArrayRef<Value *> Candidates : ReducedVals) { |
11577 | for (Value *RdxVal : Candidates) { |
11578 | if (!Visited.insert(RdxVal).second) |
11579 | continue; |
11580 | unsigned NumOps = VectorizedVals.lookup(RdxVal); |
11581 | for (Instruction *RedOp : |
11582 | makeArrayRef(ReducedValsToOps.find(RdxVal)->second) |
11583 | .drop_back(NumOps)) |
11584 | ExtraReductions.emplace_back(RedOp, RdxVal); |
11585 | } |
11586 | } |
11587 | for (auto &Pair : ExternallyUsedValues) { |
11588 | // Add each externally used value to the final reduction. |
11589 | for (auto *I : Pair.second) |
11590 | ExtraReductions.emplace_back(I, Pair.first); |
11591 | } |
11592 | // Iterate through all not-vectorized reduction values/extra arguments. |
11593 | while (ExtraReductions.size() > 1) { |
11594 | VectorizedTree = ExtraReductions.front().second; |
Value stored to 'VectorizedTree' is never read | |
11595 | SmallVector<std::pair<Instruction *, Value *>> NewReds = |
11596 | FinalGen(ExtraReductions); |
11597 | ExtraReductions.swap(NewReds); |
11598 | } |
11599 | VectorizedTree = ExtraReductions.front().second; |
11600 | |
11601 | ReductionRoot->replaceAllUsesWith(VectorizedTree); |
11602 | |
11603 | // The original scalar reduction is expected to have no remaining |
11604 | // uses outside the reduction tree itself. Assert that we got this |
11605 | // correct, replace internal uses with undef, and mark for eventual |
11606 | // deletion. |
11607 | #ifndef NDEBUG |
11608 | SmallSet<Value *, 4> IgnoreSet; |
11609 | for (ArrayRef<Value *> RdxOps : ReductionOps) |
11610 | IgnoreSet.insert(RdxOps.begin(), RdxOps.end()); |
11611 | #endif |
11612 | for (ArrayRef<Value *> RdxOps : ReductionOps) { |
11613 | for (Value *Ignore : RdxOps) { |
11614 | if (!Ignore) |
11615 | continue; |
11616 | #ifndef NDEBUG |
11617 | for (auto *U : Ignore->users()) { |
11618 | 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", 11619, __extension__ __PRETTY_FUNCTION__)) |
11619 | "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", 11619, __extension__ __PRETTY_FUNCTION__)); |
11620 | } |
11621 | #endif |
11622 | if (!Ignore->use_empty()) { |
11623 | Value *Undef = UndefValue::get(Ignore->getType()); |
11624 | Ignore->replaceAllUsesWith(Undef); |
11625 | } |
11626 | V.eraseInstruction(cast<Instruction>(Ignore)); |
11627 | } |
11628 | } |
11629 | } else if (!CheckForReusedReductionOps) { |
11630 | for (ReductionOpsType &RdxOps : ReductionOps) |
11631 | for (Value *RdxOp : RdxOps) |
11632 | V.analyzedReductionRoot(cast<Instruction>(RdxOp)); |
11633 | } |
11634 | return VectorizedTree; |
11635 | } |
11636 | |
11637 | private: |
11638 | /// Calculate the cost of a reduction. |
11639 | InstructionCost getReductionCost(TargetTransformInfo *TTI, |
11640 | ArrayRef<Value *> ReducedVals, |
11641 | unsigned ReduxWidth, FastMathFlags FMF) { |
11642 | TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput; |
11643 | Value *FirstReducedVal = ReducedVals.front(); |
11644 | Type *ScalarTy = FirstReducedVal->getType(); |
11645 | FixedVectorType *VectorTy = FixedVectorType::get(ScalarTy, ReduxWidth); |
11646 | InstructionCost VectorCost = 0, ScalarCost; |
11647 | // If all of the reduced values are constant, the vector cost is 0, since |
11648 | // the reduction value can be calculated at the compile time. |
11649 | bool AllConsts = all_of(ReducedVals, isConstant); |
11650 | switch (RdxKind) { |
11651 | case RecurKind::Add: |
11652 | case RecurKind::Mul: |
11653 | case RecurKind::Or: |
11654 | case RecurKind::And: |
11655 | case RecurKind::Xor: |
11656 | case RecurKind::FAdd: |
11657 | case RecurKind::FMul: { |
11658 | unsigned RdxOpcode = RecurrenceDescriptor::getOpcode(RdxKind); |
11659 | if (!AllConsts) |
11660 | VectorCost = |
11661 | TTI->getArithmeticReductionCost(RdxOpcode, VectorTy, FMF, CostKind); |
11662 | ScalarCost = TTI->getArithmeticInstrCost(RdxOpcode, ScalarTy, CostKind); |
11663 | break; |
11664 | } |
11665 | case RecurKind::FMax: |
11666 | case RecurKind::FMin: { |
11667 | auto *SclCondTy = CmpInst::makeCmpResultType(ScalarTy); |
11668 | if (!AllConsts) { |
11669 | auto *VecCondTy = |
11670 | cast<VectorType>(CmpInst::makeCmpResultType(VectorTy)); |
11671 | VectorCost = |
11672 | TTI->getMinMaxReductionCost(VectorTy, VecCondTy, |
11673 | /*IsUnsigned=*/false, CostKind); |
11674 | } |
11675 | CmpInst::Predicate RdxPred = getMinMaxReductionPredicate(RdxKind); |
11676 | ScalarCost = TTI->getCmpSelInstrCost(Instruction::FCmp, ScalarTy, |
11677 | SclCondTy, RdxPred, CostKind) + |
11678 | TTI->getCmpSelInstrCost(Instruction::Select, ScalarTy, |
11679 | SclCondTy, RdxPred, CostKind); |
11680 | break; |
11681 | } |
11682 | case RecurKind::SMax: |
11683 | case RecurKind::SMin: |
11684 | case RecurKind::UMax: |
11685 | case RecurKind::UMin: { |
11686 | auto *SclCondTy = CmpInst::makeCmpResultType(ScalarTy); |
11687 | if (!AllConsts) { |
11688 | auto *VecCondTy = |
11689 | cast<VectorType>(CmpInst::makeCmpResultType(VectorTy)); |
11690 | bool IsUnsigned = |
11691 | RdxKind == RecurKind::UMax || RdxKind == RecurKind::UMin; |
11692 | VectorCost = TTI->getMinMaxReductionCost(VectorTy, VecCondTy, |
11693 | IsUnsigned, CostKind); |
11694 | } |
11695 | CmpInst::Predicate RdxPred = getMinMaxReductionPredicate(RdxKind); |
11696 | ScalarCost = TTI->getCmpSelInstrCost(Instruction::ICmp, ScalarTy, |
11697 | SclCondTy, RdxPred, CostKind) + |
11698 | TTI->getCmpSelInstrCost(Instruction::Select, ScalarTy, |
11699 | SclCondTy, RdxPred, CostKind); |
11700 | break; |
11701 | } |
11702 | default: |
11703 | 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", 11703); |
11704 | } |
11705 | |
11706 | // Scalar cost is repeated for N-1 elements. |
11707 | ScalarCost *= (ReduxWidth - 1); |
11708 | 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) |
11709 | << " 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) |
11710 | << " (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); |
11711 | return VectorCost - ScalarCost; |
11712 | } |
11713 | |
11714 | /// Emit a horizontal reduction of the vectorized value. |
11715 | Value *emitReduction(Value *VectorizedValue, IRBuilder<> &Builder, |
11716 | unsigned ReduxWidth, const TargetTransformInfo *TTI) { |
11717 | 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", 11717, __extension__ __PRETTY_FUNCTION__)); |
11718 | 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", 11719, __extension__ __PRETTY_FUNCTION__)) |
11719 | "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", 11719, __extension__ __PRETTY_FUNCTION__)); |
11720 | 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", 11721, __extension__ __PRETTY_FUNCTION__)) |
11721 | "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", 11721, __extension__ __PRETTY_FUNCTION__)); |
11722 | |
11723 | ++NumVectorInstructions; |
11724 | return createSimpleTargetReduction(Builder, TTI, VectorizedValue, RdxKind); |
11725 | } |
11726 | }; |
11727 | |
11728 | } // end anonymous namespace |
11729 | |
11730 | static Optional<unsigned> getAggregateSize(Instruction *InsertInst) { |
11731 | if (auto *IE = dyn_cast<InsertElementInst>(InsertInst)) |
11732 | return cast<FixedVectorType>(IE->getType())->getNumElements(); |
11733 | |
11734 | unsigned AggregateSize = 1; |
11735 | auto *IV = cast<InsertValueInst>(InsertInst); |
11736 | Type *CurrentType = IV->getType(); |
11737 | do { |
11738 | if (auto *ST = dyn_cast<StructType>(CurrentType)) { |
11739 | for (auto *Elt : ST->elements()) |
11740 | if (Elt != ST->getElementType(0)) // check homogeneity |
11741 | return None; |
11742 | AggregateSize *= ST->getNumElements(); |
11743 | CurrentType = ST->getElementType(0); |
11744 | } else if (auto *AT = dyn_cast<ArrayType>(CurrentType)) { |
11745 | AggregateSize *= AT->getNumElements(); |
11746 | CurrentType = AT->getElementType(); |
11747 | } else if (auto *VT = dyn_cast<FixedVectorType>(CurrentType)) { |
11748 | AggregateSize *= VT->getNumElements(); |
11749 | return AggregateSize; |
11750 | } else if (CurrentType->isSingleValueType()) { |
11751 | return AggregateSize; |
11752 | } else { |
11753 | return None; |
11754 | } |
11755 | } while (true); |
11756 | } |
11757 | |
11758 | static void findBuildAggregate_rec(Instruction *LastInsertInst, |
11759 | TargetTransformInfo *TTI, |
11760 | SmallVectorImpl<Value *> &BuildVectorOpds, |
11761 | SmallVectorImpl<Value *> &InsertElts, |
11762 | unsigned OperandOffset) { |
11763 | do { |
11764 | Value *InsertedOperand = LastInsertInst->getOperand(1); |
11765 | Optional<unsigned> OperandIndex = |
11766 | getInsertIndex(LastInsertInst, OperandOffset); |
11767 | if (!OperandIndex) |
11768 | return; |
11769 | if (isa<InsertElementInst, InsertValueInst>(InsertedOperand)) { |
11770 | findBuildAggregate_rec(cast<Instruction>(InsertedOperand), TTI, |
11771 | BuildVectorOpds, InsertElts, *OperandIndex); |
11772 | |
11773 | } else { |
11774 | BuildVectorOpds[*OperandIndex] = InsertedOperand; |
11775 | InsertElts[*OperandIndex] = LastInsertInst; |
11776 | } |
11777 | LastInsertInst = dyn_cast<Instruction>(LastInsertInst->getOperand(0)); |
11778 | } while (LastInsertInst != nullptr && |
11779 | isa<InsertValueInst, InsertElementInst>(LastInsertInst) && |
11780 | LastInsertInst->hasOneUse()); |
11781 | } |
11782 | |
11783 | /// Recognize construction of vectors like |
11784 | /// %ra = insertelement <4 x float> poison, float %s0, i32 0 |
11785 | /// %rb = insertelement <4 x float> %ra, float %s1, i32 1 |
11786 | /// %rc = insertelement <4 x float> %rb, float %s2, i32 2 |
11787 | /// %rd = insertelement <4 x float> %rc, float %s3, i32 3 |
11788 | /// starting from the last insertelement or insertvalue instruction. |
11789 | /// |
11790 | /// Also recognize homogeneous aggregates like {<2 x float>, <2 x float>}, |
11791 | /// {{float, float}, {float, float}}, [2 x {float, float}] and so on. |
11792 | /// See llvm/test/Transforms/SLPVectorizer/X86/pr42022.ll for examples. |
11793 | /// |
11794 | /// Assume LastInsertInst is of InsertElementInst or InsertValueInst type. |
11795 | /// |
11796 | /// \return true if it matches. |
11797 | static bool findBuildAggregate(Instruction *LastInsertInst, |
11798 | TargetTransformInfo *TTI, |
11799 | SmallVectorImpl<Value *> &BuildVectorOpds, |
11800 | SmallVectorImpl<Value *> &InsertElts) { |
11801 | |
11802 | 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", 11804, __extension__ __PRETTY_FUNCTION__)) |
11803 | 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", 11804, __extension__ __PRETTY_FUNCTION__)) |
11804 | "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", 11804, __extension__ __PRETTY_FUNCTION__)); |
11805 | |
11806 | 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", 11807, __extension__ __PRETTY_FUNCTION__)) |
11807 | "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", 11807, __extension__ __PRETTY_FUNCTION__)); |
11808 | |
11809 | Optional<unsigned> AggregateSize = getAggregateSize(LastInsertInst); |
11810 | if (!AggregateSize) |
11811 | return false; |
11812 | BuildVectorOpds.resize(*AggregateSize); |
11813 | InsertElts.resize(*AggregateSize); |
11814 | |
11815 | findBuildAggregate_rec(LastInsertInst, TTI, BuildVectorOpds, InsertElts, 0); |
11816 | llvm::erase_value(BuildVectorOpds, nullptr); |
11817 | llvm::erase_value(InsertElts, nullptr); |
11818 | if (BuildVectorOpds.size() >= 2) |
11819 | return true; |
11820 | |
11821 | return false; |
11822 | } |
11823 | |
11824 | /// Try and get a reduction value from a phi node. |
11825 | /// |
11826 | /// Given a phi node \p P in a block \p ParentBB, consider possible reductions |
11827 | /// if they come from either \p ParentBB or a containing loop latch. |
11828 | /// |
11829 | /// \returns A candidate reduction value if possible, or \code nullptr \endcode |
11830 | /// if not possible. |
11831 | static Value *getReductionValue(const DominatorTree *DT, PHINode *P, |
11832 | BasicBlock *ParentBB, LoopInfo *LI) { |
11833 | // There are situations where the reduction value is not dominated by the |
11834 | // reduction phi. Vectorizing such cases has been reported to cause |
11835 | // miscompiles. See PR25787. |
11836 | auto DominatedReduxValue = [&](Value *R) { |
11837 | return isa<Instruction>(R) && |
11838 | DT->dominates(P->getParent(), cast<Instruction>(R)->getParent()); |
11839 | }; |
11840 | |
11841 | Value *Rdx = nullptr; |
11842 | |
11843 | // Return the incoming value if it comes from the same BB as the phi node. |
11844 | if (P->getIncomingBlock(0) == ParentBB) { |
11845 | Rdx = P->getIncomingValue(0); |
11846 | } else if (P->getIncomingBlock(1) == ParentBB) { |
11847 | Rdx = P->getIncomingValue(1); |
11848 | } |
11849 | |
11850 | if (Rdx && DominatedReduxValue(Rdx)) |
11851 | return Rdx; |
11852 | |
11853 | // Otherwise, check whether we have a loop latch to look at. |
11854 | Loop *BBL = LI->getLoopFor(ParentBB); |
11855 | if (!BBL) |
11856 | return nullptr; |
11857 | BasicBlock *BBLatch = BBL->getLoopLatch(); |
11858 | if (!BBLatch) |
11859 | return nullptr; |
11860 | |
11861 | // There is a loop latch, return the incoming value if it comes from |
11862 | // that. This reduction pattern occasionally turns up. |
11863 | if (P->getIncomingBlock(0) == BBLatch) { |
11864 | Rdx = P->getIncomingValue(0); |
11865 | } else if (P->getIncomingBlock(1) == BBLatch) { |
11866 | Rdx = P->getIncomingValue(1); |
11867 | } |
11868 | |
11869 | if (Rdx && DominatedReduxValue(Rdx)) |
11870 | return Rdx; |
11871 | |
11872 | return nullptr; |
11873 | } |
11874 | |
11875 | static bool matchRdxBop(Instruction *I, Value *&V0, Value *&V1) { |
11876 | if (match(I, m_BinOp(m_Value(V0), m_Value(V1)))) |
11877 | return true; |
11878 | if (match(I, m_Intrinsic<Intrinsic::maxnum>(m_Value(V0), m_Value(V1)))) |
11879 | return true; |
11880 | if (match(I, m_Intrinsic<Intrinsic::minnum>(m_Value(V0), m_Value(V1)))) |
11881 | return true; |
11882 | if (match(I, m_Intrinsic<Intrinsic::smax>(m_Value(V0), m_Value(V1)))) |
11883 | return true; |
11884 | if (match(I, m_Intrinsic<Intrinsic::smin>(m_Value(V0), m_Value(V1)))) |
11885 | return true; |
11886 | if (match(I, m_Intrinsic<Intrinsic::umax>(m_Value(V0), m_Value(V1)))) |
11887 | return true; |
11888 | if (match(I, m_Intrinsic<Intrinsic::umin>(m_Value(V0), m_Value(V1)))) |
11889 | return true; |
11890 | return false; |
11891 | } |
11892 | |
11893 | bool SLPVectorizerPass::vectorizeHorReduction( |
11894 | PHINode *P, Value *V, BasicBlock *BB, BoUpSLP &R, TargetTransformInfo *TTI, |
11895 | SmallVectorImpl<WeakTrackingVH> &PostponedInsts) { |
11896 | if (!ShouldVectorizeHor) |
11897 | return false; |
11898 | |
11899 | auto *Root = dyn_cast_or_null<Instruction>(V); |
11900 | if (!Root) |
11901 | return false; |
11902 | |
11903 | if (!isa<BinaryOperator>(Root)) |
11904 | P = nullptr; |
11905 | |
11906 | if (Root->getParent() != BB || isa<PHINode>(Root)) |
11907 | return false; |
11908 | // Start analysis starting from Root instruction. If horizontal reduction is |
11909 | // found, try to vectorize it. If it is not a horizontal reduction or |
11910 | // vectorization is not possible or not effective, and currently analyzed |
11911 | // instruction is a binary operation, try to vectorize the operands, using |
11912 | // pre-order DFS traversal order. If the operands were not vectorized, repeat |
11913 | // the same procedure considering each operand as a possible root of the |
11914 | // horizontal reduction. |
11915 | // Interrupt the process if the Root instruction itself was vectorized or all |
11916 | // sub-trees not higher that RecursionMaxDepth were analyzed/vectorized. |
11917 | // If a horizintal reduction was not matched or vectorized we collect |
11918 | // instructions for possible later attempts for vectorization. |
11919 | std::queue<std::pair<Instruction *, unsigned>> Stack; |
11920 | Stack.emplace(Root, 0); |
11921 | SmallPtrSet<Value *, 8> VisitedInstrs; |
11922 | bool Res = false; |
11923 | auto &&TryToReduce = [this, TTI, &P, &R](Instruction *Inst, Value *&B0, |
11924 | Value *&B1) -> Value * { |
11925 | if (R.isAnalyzedReductionRoot(Inst)) |
11926 | return nullptr; |
11927 | bool IsBinop = matchRdxBop(Inst, B0, B1); |
11928 | bool IsSelect = match(Inst, m_Select(m_Value(), m_Value(), m_Value())); |
11929 | if (IsBinop || IsSelect) { |
11930 | HorizontalReduction HorRdx; |
11931 | if (HorRdx.matchAssociativeReduction(P, Inst, *SE, *DL, *TLI)) |
11932 | return HorRdx.tryToReduce(R, TTI); |
11933 | } |
11934 | return nullptr; |
11935 | }; |
11936 | while (!Stack.empty()) { |
11937 | Instruction *Inst; |
11938 | unsigned Level; |
11939 | std::tie(Inst, Level) = Stack.front(); |
11940 | Stack.pop(); |
11941 | // Do not try to analyze instruction that has already been vectorized. |
11942 | // This may happen when we vectorize instruction operands on a previous |
11943 | // iteration while stack was populated before that happened. |
11944 | if (R.isDeleted(Inst)) |
11945 | continue; |
11946 | Value *B0 = nullptr, *B1 = nullptr; |
11947 | if (Value *V = TryToReduce(Inst, B0, B1)) { |
11948 | Res = true; |
11949 | // Set P to nullptr to avoid re-analysis of phi node in |
11950 | // matchAssociativeReduction function unless this is the root node. |
11951 | P = nullptr; |
11952 | if (auto *I = dyn_cast<Instruction>(V)) { |
11953 | // Try to find another reduction. |
11954 | Stack.emplace(I, Level); |
11955 | continue; |
11956 | } |
11957 | } else { |
11958 | bool IsBinop = B0 && B1; |
11959 | if (P && IsBinop) { |
11960 | Inst = dyn_cast<Instruction>(B0); |
11961 | if (Inst == P) |
11962 | Inst = dyn_cast<Instruction>(B1); |
11963 | if (!Inst) { |
11964 | // Set P to nullptr to avoid re-analysis of phi node in |
11965 | // matchAssociativeReduction function unless this is the root node. |
11966 | P = nullptr; |
11967 | continue; |
11968 | } |
11969 | } |
11970 | // Set P to nullptr to avoid re-analysis of phi node in |
11971 | // matchAssociativeReduction function unless this is the root node. |
11972 | P = nullptr; |
11973 | // Do not collect CmpInst or InsertElementInst/InsertValueInst as their |
11974 | // analysis is done separately. |
11975 | if (!isa<CmpInst, InsertElementInst, InsertValueInst>(Inst)) |
11976 | PostponedInsts.push_back(Inst); |
11977 | } |
11978 | |
11979 | // Try to vectorize operands. |
11980 | // Continue analysis for the instruction from the same basic block only to |
11981 | // save compile time. |
11982 | if (++Level < RecursionMaxDepth) |
11983 | for (auto *Op : Inst->operand_values()) |
11984 | if (VisitedInstrs.insert(Op).second) |
11985 | if (auto *I = dyn_cast<Instruction>(Op)) |
11986 | // Do not try to vectorize CmpInst operands, this is done |
11987 | // separately. |
11988 | if (!isa<PHINode, CmpInst, InsertElementInst, InsertValueInst>(I) && |
11989 | !R.isDeleted(I) && I->getParent() == BB) |
11990 | Stack.emplace(I, Level); |
11991 | } |
11992 | return Res; |
11993 | } |
11994 | |
11995 | bool SLPVectorizerPass::vectorizeRootInstruction(PHINode *P, Value *V, |
11996 | BasicBlock *BB, BoUpSLP &R, |
11997 | TargetTransformInfo *TTI) { |
11998 | SmallVector<WeakTrackingVH> PostponedInsts; |
11999 | bool Res = vectorizeHorReduction(P, V, BB, R, TTI, PostponedInsts); |
12000 | Res |= tryToVectorize(PostponedInsts, R); |
12001 | return Res; |
12002 | } |
12003 | |
12004 | bool SLPVectorizerPass::tryToVectorize(ArrayRef<WeakTrackingVH> Insts, |
12005 | BoUpSLP &R) { |
12006 | bool Res = false; |
12007 | for (Value *V : Insts) |
12008 | if (auto *Inst = dyn_cast<Instruction>(V); Inst && !R.isDeleted(Inst)) |
12009 | Res |= tryToVectorize(Inst, R); |
12010 | return Res; |
12011 | } |
12012 | |
12013 | bool SLPVectorizerPass::vectorizeInsertValueInst(InsertValueInst *IVI, |
12014 | BasicBlock *BB, BoUpSLP &R) { |
12015 | const DataLayout &DL = BB->getModule()->getDataLayout(); |
12016 | if (!R.canMapToVector(IVI->getType(), DL)) |
12017 | return false; |
12018 | |
12019 | SmallVector<Value *, 16> BuildVectorOpds; |
12020 | SmallVector<Value *, 16> BuildVectorInsts; |
12021 | if (!findBuildAggregate(IVI, TTI, BuildVectorOpds, BuildVectorInsts)) |
12022 | return false; |
12023 | |
12024 | 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); |
12025 | // Aggregate value is unlikely to be processed in vector register. |
12026 | return tryToVectorizeList(BuildVectorOpds, R); |
12027 | } |
12028 | |
12029 | bool SLPVectorizerPass::vectorizeInsertElementInst(InsertElementInst *IEI, |
12030 | BasicBlock *BB, BoUpSLP &R) { |
12031 | SmallVector<Value *, 16> BuildVectorInsts; |
12032 | SmallVector<Value *, 16> BuildVectorOpds; |
12033 | SmallVector<int> Mask; |
12034 | if (!findBuildAggregate(IEI, TTI, BuildVectorOpds, BuildVectorInsts) || |
12035 | (llvm::all_of( |
12036 | BuildVectorOpds, |
12037 | [](Value *V) { return isa<ExtractElementInst, UndefValue>(V); }) && |
12038 | isFixedVectorShuffle(BuildVectorOpds, Mask))) |
12039 | return false; |
12040 | |
12041 | 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); |
12042 | return tryToVectorizeList(BuildVectorInsts, R); |
12043 | } |
12044 | |
12045 | template <typename T> |
12046 | static bool |
12047 | tryToVectorizeSequence(SmallVectorImpl<T *> &Incoming, |
12048 | function_ref<unsigned(T *)> Limit, |
12049 | function_ref<bool(T *, T *)> Comparator, |
12050 | function_ref<bool(T *, T *)> AreCompatible, |
12051 | function_ref<bool(ArrayRef<T *>, bool)> TryToVectorizeHelper, |
12052 | bool LimitForRegisterSize) { |
12053 | bool Changed = false; |
12054 | // Sort by type, parent, operands. |
12055 | stable_sort(Incoming, Comparator); |
12056 | |
12057 | // Try to vectorize elements base on their type. |
12058 | SmallVector<T *> Candidates; |
12059 | for (auto *IncIt = Incoming.begin(), *E = Incoming.end(); IncIt != E;) { |
12060 | // Look for the next elements with the same type, parent and operand |
12061 | // kinds. |
12062 | auto *SameTypeIt = IncIt; |
12063 | while (SameTypeIt != E && AreCompatible(*SameTypeIt, *IncIt)) |
12064 | ++SameTypeIt; |
12065 | |
12066 | // Try to vectorize them. |
12067 | unsigned NumElts = (SameTypeIt - IncIt); |
12068 | 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) |
12069 | << NumElts << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Trying to vectorize starting at nodes (" << NumElts << ")\n"; } } while (false); |
12070 | // The vectorization is a 3-state attempt: |
12071 | // 1. Try to vectorize instructions with the same/alternate opcodes with the |
12072 | // size of maximal register at first. |
12073 | // 2. Try to vectorize remaining instructions with the same type, if |
12074 | // possible. This may result in the better vectorization results rather than |
12075 | // if we try just to vectorize instructions with the same/alternate opcodes. |
12076 | // 3. Final attempt to try to vectorize all instructions with the |
12077 | // same/alternate ops only, this may result in some extra final |
12078 | // vectorization. |
12079 | if (NumElts > 1 && |
12080 | TryToVectorizeHelper(makeArrayRef(IncIt, NumElts), LimitForRegisterSize)) { |
12081 | // Success start over because instructions might have been changed. |
12082 | Changed = true; |
12083 | } else if (NumElts < Limit(*IncIt) && |
12084 | (Candidates.empty() || |
12085 | Candidates.front()->getType() == (*IncIt)->getType())) { |
12086 | Candidates.append(IncIt, std::next(IncIt, NumElts)); |
12087 | } |
12088 | // Final attempt to vectorize instructions with the same types. |
12089 | if (Candidates.size() > 1 && |
12090 | (SameTypeIt == E || (*SameTypeIt)->getType() != (*IncIt)->getType())) { |
12091 | if (TryToVectorizeHelper(Candidates, /*LimitForRegisterSize=*/false)) { |
12092 | // Success start over because instructions might have been changed. |
12093 | Changed = true; |
12094 | } else if (LimitForRegisterSize) { |
12095 | // Try to vectorize using small vectors. |
12096 | for (auto *It = Candidates.begin(), *End = Candidates.end(); |
12097 | It != End;) { |
12098 | auto *SameTypeIt = It; |
12099 | while (SameTypeIt != End && AreCompatible(*SameTypeIt, *It)) |
12100 | ++SameTypeIt; |
12101 | unsigned NumElts = (SameTypeIt - It); |
12102 | if (NumElts > 1 && TryToVectorizeHelper(makeArrayRef(It, NumElts), |
12103 | /*LimitForRegisterSize=*/false)) |
12104 | Changed = true; |
12105 | It = SameTypeIt; |
12106 | } |
12107 | } |
12108 | Candidates.clear(); |
12109 | } |
12110 | |
12111 | // Start over at the next instruction of a different type (or the end). |
12112 | IncIt = SameTypeIt; |
12113 | } |
12114 | return Changed; |
12115 | } |
12116 | |
12117 | /// Compare two cmp instructions. If IsCompatibility is true, function returns |
12118 | /// true if 2 cmps have same/swapped predicates and mos compatible corresponding |
12119 | /// operands. If IsCompatibility is false, function implements strict weak |
12120 | /// ordering relation between two cmp instructions, returning true if the first |
12121 | /// instruction is "less" than the second, i.e. its predicate is less than the |
12122 | /// predicate of the second or the operands IDs are less than the operands IDs |
12123 | /// of the second cmp instruction. |
12124 | template <bool IsCompatibility> |
12125 | static bool compareCmp(Value *V, Value *V2, |
12126 | function_ref<bool(Instruction *)> IsDeleted) { |
12127 | auto *CI1 = cast<CmpInst>(V); |
12128 | auto *CI2 = cast<CmpInst>(V2); |
12129 | if (IsDeleted(CI2) || !isValidElementType(CI2->getType())) |
12130 | return false; |
12131 | if (CI1->getOperand(0)->getType()->getTypeID() < |
12132 | CI2->getOperand(0)->getType()->getTypeID()) |
12133 | return !IsCompatibility; |
12134 | if (CI1->getOperand(0)->getType()->getTypeID() > |
12135 | CI2->getOperand(0)->getType()->getTypeID()) |
12136 | return false; |
12137 | CmpInst::Predicate Pred1 = CI1->getPredicate(); |
12138 | CmpInst::Predicate Pred2 = CI2->getPredicate(); |
12139 | CmpInst::Predicate SwapPred1 = CmpInst::getSwappedPredicate(Pred1); |
12140 | CmpInst::Predicate SwapPred2 = CmpInst::getSwappedPredicate(Pred2); |
12141 | CmpInst::Predicate BasePred1 = std::min(Pred1, SwapPred1); |
12142 | CmpInst::Predicate BasePred2 = std::min(Pred2, SwapPred2); |
12143 | if (BasePred1 < BasePred2) |
12144 | return !IsCompatibility; |
12145 | if (BasePred1 > BasePred2) |
12146 | return false; |
12147 | // Compare operands. |
12148 | bool LEPreds = Pred1 <= Pred2; |
12149 | bool GEPreds = Pred1 >= Pred2; |
12150 | for (int I = 0, E = CI1->getNumOperands(); I < E; ++I) { |
12151 | auto *Op1 = CI1->getOperand(LEPreds ? I : E - I - 1); |
12152 | auto *Op2 = CI2->getOperand(GEPreds ? I : E - I - 1); |
12153 | if (Op1->getValueID() < Op2->getValueID()) |
12154 | return !IsCompatibility; |
12155 | if (Op1->getValueID() > Op2->getValueID()) |
12156 | return false; |
12157 | if (auto *I1 = dyn_cast<Instruction>(Op1)) |
12158 | if (auto *I2 = dyn_cast<Instruction>(Op2)) { |
12159 | if (I1->getParent() != I2->getParent()) |
12160 | return false; |
12161 | InstructionsState S = getSameOpcode({I1, I2}); |
12162 | if (S.getOpcode()) |
12163 | continue; |
12164 | return false; |
12165 | } |
12166 | } |
12167 | return IsCompatibility; |
12168 | } |
12169 | |
12170 | bool SLPVectorizerPass::vectorizeSimpleInstructions(InstSetVector &Instructions, |
12171 | BasicBlock *BB, BoUpSLP &R, |
12172 | bool AtTerminator) { |
12173 | bool OpsChanged = false; |
12174 | SmallVector<Instruction *, 4> PostponedCmps; |
12175 | SmallVector<WeakTrackingVH> PostponedInsts; |
12176 | // pass1 - try to vectorize reductions only |
12177 | for (auto *I : reverse(Instructions)) { |
12178 | if (R.isDeleted(I)) |
12179 | continue; |
12180 | if (isa<CmpInst>(I)) { |
12181 | PostponedCmps.push_back(I); |
12182 | continue; |
12183 | } |
12184 | OpsChanged |= vectorizeHorReduction(nullptr, I, BB, R, TTI, PostponedInsts); |
12185 | } |
12186 | // pass2 - try to match and vectorize a buildvector sequence. |
12187 | for (auto *I : reverse(Instructions)) { |
12188 | if (R.isDeleted(I) || isa<CmpInst>(I)) |
12189 | continue; |
12190 | if (auto *LastInsertValue = dyn_cast<InsertValueInst>(I)) { |
12191 | OpsChanged |= vectorizeInsertValueInst(LastInsertValue, BB, R); |
12192 | } else if (auto *LastInsertElem = dyn_cast<InsertElementInst>(I)) { |
12193 | OpsChanged |= vectorizeInsertElementInst(LastInsertElem, BB, R); |
12194 | } |
12195 | } |
12196 | // Now try to vectorize postponed instructions. |
12197 | OpsChanged |= tryToVectorize(PostponedInsts, R); |
12198 | |
12199 | if (AtTerminator) { |
12200 | // Try to find reductions first. |
12201 | for (Instruction *I : PostponedCmps) { |
12202 | if (R.isDeleted(I)) |
12203 | continue; |
12204 | for (Value *Op : I->operands()) |
12205 | OpsChanged |= vectorizeRootInstruction(nullptr, Op, BB, R, TTI); |
12206 | } |
12207 | // Try to vectorize operands as vector bundles. |
12208 | for (Instruction *I : PostponedCmps) { |
12209 | if (R.isDeleted(I)) |
12210 | continue; |
12211 | OpsChanged |= tryToVectorize(I, R); |
12212 | } |
12213 | // Try to vectorize list of compares. |
12214 | // Sort by type, compare predicate, etc. |
12215 | auto &&CompareSorter = [&R](Value *V, Value *V2) { |
12216 | return compareCmp<false>(V, V2, |
12217 | [&R](Instruction *I) { return R.isDeleted(I); }); |
12218 | }; |
12219 | |
12220 | auto &&AreCompatibleCompares = [&R](Value *V1, Value *V2) { |
12221 | if (V1 == V2) |
12222 | return true; |
12223 | return compareCmp<true>(V1, V2, |
12224 | [&R](Instruction *I) { return R.isDeleted(I); }); |
12225 | }; |
12226 | auto Limit = [&R](Value *V) { |
12227 | unsigned EltSize = R.getVectorElementSize(V); |
12228 | return std::max(2U, R.getMaxVecRegSize() / EltSize); |
12229 | }; |
12230 | |
12231 | SmallVector<Value *> Vals(PostponedCmps.begin(), PostponedCmps.end()); |
12232 | OpsChanged |= tryToVectorizeSequence<Value>( |
12233 | Vals, Limit, CompareSorter, AreCompatibleCompares, |
12234 | [this, &R](ArrayRef<Value *> Candidates, bool LimitForRegisterSize) { |
12235 | // Exclude possible reductions from other blocks. |
12236 | bool ArePossiblyReducedInOtherBlock = |
12237 | any_of(Candidates, [](Value *V) { |
12238 | return any_of(V->users(), [V](User *U) { |
12239 | return isa<SelectInst>(U) && |
12240 | cast<SelectInst>(U)->getParent() != |
12241 | cast<Instruction>(V)->getParent(); |
12242 | }); |
12243 | }); |
12244 | if (ArePossiblyReducedInOtherBlock) |
12245 | return false; |
12246 | return tryToVectorizeList(Candidates, R, LimitForRegisterSize); |
12247 | }, |
12248 | /*LimitForRegisterSize=*/true); |
12249 | Instructions.clear(); |
12250 | } else { |
12251 | Instructions.clear(); |
12252 | // Insert in reverse order since the PostponedCmps vector was filled in |
12253 | // reverse order. |
12254 | Instructions.insert(PostponedCmps.rbegin(), PostponedCmps.rend()); |
12255 | } |
12256 | return OpsChanged; |
12257 | } |
12258 | |
12259 | bool SLPVectorizerPass::vectorizeChainsInBlock(BasicBlock *BB, BoUpSLP &R) { |
12260 | bool Changed = false; |
12261 | SmallVector<Value *, 4> Incoming; |
12262 | SmallPtrSet<Value *, 16> VisitedInstrs; |
12263 | // Maps phi nodes to the non-phi nodes found in the use tree for each phi |
12264 | // node. Allows better to identify the chains that can be vectorized in the |
12265 | // better way. |
12266 | DenseMap<Value *, SmallVector<Value *, 4>> PHIToOpcodes; |
12267 | auto PHICompare = [this, &PHIToOpcodes](Value *V1, Value *V2) { |
12268 | 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", 12270, __extension__ __PRETTY_FUNCTION__)) |
12269 | 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", 12270, __extension__ __PRETTY_FUNCTION__)) |
12270 | "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", 12270, __extension__ __PRETTY_FUNCTION__)); |
12271 | // It is fine to compare type IDs here, since we expect only vectorizable |
12272 | // types, like ints, floats and pointers, we don't care about other type. |
12273 | if (V1->getType()->getTypeID() < V2->getType()->getTypeID()) |
12274 | return true; |
12275 | if (V1->getType()->getTypeID() > V2->getType()->getTypeID()) |
12276 | return false; |
12277 | ArrayRef<Value *> Opcodes1 = PHIToOpcodes[V1]; |
12278 | ArrayRef<Value *> Opcodes2 = PHIToOpcodes[V2]; |
12279 | if (Opcodes1.size() < Opcodes2.size()) |
12280 | return true; |
12281 | if (Opcodes1.size() > Opcodes2.size()) |
12282 | return false; |
12283 | Optional<bool> ConstOrder; |
12284 | for (int I = 0, E = Opcodes1.size(); I < E; ++I) { |
12285 | // Undefs are compatible with any other value. |
12286 | if (isa<UndefValue>(Opcodes1[I]) || isa<UndefValue>(Opcodes2[I])) { |
12287 | if (!ConstOrder) |
12288 | ConstOrder = |
12289 | !isa<UndefValue>(Opcodes1[I]) && isa<UndefValue>(Opcodes2[I]); |
12290 | continue; |
12291 | } |
12292 | if (auto *I1 = dyn_cast<Instruction>(Opcodes1[I])) |
12293 | if (auto *I2 = dyn_cast<Instruction>(Opcodes2[I])) { |
12294 | DomTreeNodeBase<BasicBlock> *NodeI1 = DT->getNode(I1->getParent()); |
12295 | DomTreeNodeBase<BasicBlock> *NodeI2 = DT->getNode(I2->getParent()); |
12296 | if (!NodeI1) |
12297 | return NodeI2 != nullptr; |
12298 | if (!NodeI2) |
12299 | return false; |
12300 | 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", 12302, __extension__ __PRETTY_FUNCTION__)) |
12301 | (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", 12302, __extension__ __PRETTY_FUNCTION__)) |
12302 | "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", 12302, __extension__ __PRETTY_FUNCTION__)); |
12303 | if (NodeI1 != NodeI2) |
12304 | return NodeI1->getDFSNumIn() < NodeI2->getDFSNumIn(); |
12305 | InstructionsState S = getSameOpcode({I1, I2}); |
12306 | if (S.getOpcode()) |
12307 | continue; |
12308 | return I1->getOpcode() < I2->getOpcode(); |
12309 | } |
12310 | if (isa<Constant>(Opcodes1[I]) && isa<Constant>(Opcodes2[I])) { |
12311 | if (!ConstOrder) |
12312 | ConstOrder = Opcodes1[I]->getValueID() < Opcodes2[I]->getValueID(); |
12313 | continue; |
12314 | } |
12315 | if (Opcodes1[I]->getValueID() < Opcodes2[I]->getValueID()) |
12316 | return true; |
12317 | if (Opcodes1[I]->getValueID() > Opcodes2[I]->getValueID()) |
12318 | return false; |
12319 | } |
12320 | return ConstOrder && *ConstOrder; |
12321 | }; |
12322 | auto AreCompatiblePHIs = [&PHIToOpcodes](Value *V1, Value *V2) { |
12323 | if (V1 == V2) |
12324 | return true; |
12325 | if (V1->getType() != V2->getType()) |
12326 | return false; |
12327 | ArrayRef<Value *> Opcodes1 = PHIToOpcodes[V1]; |
12328 | ArrayRef<Value *> Opcodes2 = PHIToOpcodes[V2]; |
12329 | if (Opcodes1.size() != Opcodes2.size()) |
12330 | return false; |
12331 | for (int I = 0, E = Opcodes1.size(); I < E; ++I) { |
12332 | // Undefs are compatible with any other value. |
12333 | if (isa<UndefValue>(Opcodes1[I]) || isa<UndefValue>(Opcodes2[I])) |
12334 | continue; |
12335 | if (auto *I1 = dyn_cast<Instruction>(Opcodes1[I])) |
12336 | if (auto *I2 = dyn_cast<Instruction>(Opcodes2[I])) { |
12337 | if (I1->getParent() != I2->getParent()) |
12338 | return false; |
12339 | InstructionsState S = getSameOpcode({I1, I2}); |
12340 | if (S.getOpcode()) |
12341 | continue; |
12342 | return false; |
12343 | } |
12344 | if (isa<Constant>(Opcodes1[I]) && isa<Constant>(Opcodes2[I])) |
12345 | continue; |
12346 | if (Opcodes1[I]->getValueID() != Opcodes2[I]->getValueID()) |
12347 | return false; |
12348 | } |
12349 | return true; |
12350 | }; |
12351 | auto Limit = [&R](Value *V) { |
12352 | unsigned EltSize = R.getVectorElementSize(V); |
12353 | return std::max(2U, R.getMaxVecRegSize() / EltSize); |
12354 | }; |
12355 | |
12356 | bool HaveVectorizedPhiNodes = false; |
12357 | do { |
12358 | // Collect the incoming values from the PHIs. |
12359 | Incoming.clear(); |
12360 | for (Instruction &I : *BB) { |
12361 | PHINode *P = dyn_cast<PHINode>(&I); |
12362 | if (!P) |
12363 | break; |
12364 | |
12365 | // No need to analyze deleted, vectorized and non-vectorizable |
12366 | // instructions. |
12367 | if (!VisitedInstrs.count(P) && !R.isDeleted(P) && |
12368 | isValidElementType(P->getType())) |
12369 | Incoming.push_back(P); |
12370 | } |
12371 | |
12372 | // Find the corresponding non-phi nodes for better matching when trying to |
12373 | // build the tree. |
12374 | for (Value *V : Incoming) { |
12375 | SmallVectorImpl<Value *> &Opcodes = |
12376 | PHIToOpcodes.try_emplace(V).first->getSecond(); |
12377 | if (!Opcodes.empty()) |
12378 | continue; |
12379 | SmallVector<Value *, 4> Nodes(1, V); |
12380 | SmallPtrSet<Value *, 4> Visited; |
12381 | while (!Nodes.empty()) { |
12382 | auto *PHI = cast<PHINode>(Nodes.pop_back_val()); |
12383 | if (!Visited.insert(PHI).second) |
12384 | continue; |
12385 | for (Value *V : PHI->incoming_values()) { |
12386 | if (auto *PHI1 = dyn_cast<PHINode>((V))) { |
12387 | Nodes.push_back(PHI1); |
12388 | continue; |
12389 | } |
12390 | Opcodes.emplace_back(V); |
12391 | } |
12392 | } |
12393 | } |
12394 | |
12395 | HaveVectorizedPhiNodes = tryToVectorizeSequence<Value>( |
12396 | Incoming, Limit, PHICompare, AreCompatiblePHIs, |
12397 | [this, &R](ArrayRef<Value *> Candidates, bool LimitForRegisterSize) { |
12398 | return tryToVectorizeList(Candidates, R, LimitForRegisterSize); |
12399 | }, |
12400 | /*LimitForRegisterSize=*/true); |
12401 | Changed |= HaveVectorizedPhiNodes; |
12402 | VisitedInstrs.insert(Incoming.begin(), Incoming.end()); |
12403 | } while (HaveVectorizedPhiNodes); |
12404 | |
12405 | VisitedInstrs.clear(); |
12406 | |
12407 | InstSetVector PostProcessInstructions; |
12408 | SmallDenseSet<Instruction *, 4> KeyNodes; |
12409 | for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) { |
12410 | // Skip instructions with scalable type. The num of elements is unknown at |
12411 | // compile-time for scalable type. |
12412 | if (isa<ScalableVectorType>(it->getType())) |
12413 | continue; |
12414 | |
12415 | // Skip instructions marked for the deletion. |
12416 | if (R.isDeleted(&*it)) |
12417 | continue; |
12418 | // We may go through BB multiple times so skip the one we have checked. |
12419 | if (!VisitedInstrs.insert(&*it).second) { |
12420 | if (it->use_empty() && KeyNodes.contains(&*it) && |
12421 | vectorizeSimpleInstructions(PostProcessInstructions, BB, R, |
12422 | it->isTerminator())) { |
12423 | // We would like to start over since some instructions are deleted |
12424 | // and the iterator may become invalid value. |
12425 | Changed = true; |
12426 | it = BB->begin(); |
12427 | e = BB->end(); |
12428 | } |
12429 | continue; |
12430 | } |
12431 | |
12432 | if (isa<DbgInfoIntrinsic>(it)) |
12433 | continue; |
12434 | |
12435 | // Try to vectorize reductions that use PHINodes. |
12436 | if (PHINode *P = dyn_cast<PHINode>(it)) { |
12437 | // Check that the PHI is a reduction PHI. |
12438 | if (P->getNumIncomingValues() == 2) { |
12439 | // Try to match and vectorize a horizontal reduction. |
12440 | if (vectorizeRootInstruction(P, getReductionValue(DT, P, BB, LI), BB, R, |
12441 | TTI)) { |
12442 | Changed = true; |
12443 | it = BB->begin(); |
12444 | e = BB->end(); |
12445 | continue; |
12446 | } |
12447 | } |
12448 | // Try to vectorize the incoming values of the PHI, to catch reductions |
12449 | // that feed into PHIs. |
12450 | for (unsigned I = 0, E = P->getNumIncomingValues(); I != E; I++) { |
12451 | // Skip if the incoming block is the current BB for now. Also, bypass |
12452 | // unreachable IR for efficiency and to avoid crashing. |
12453 | // TODO: Collect the skipped incoming values and try to vectorize them |
12454 | // after processing BB. |
12455 | if (BB == P->getIncomingBlock(I) || |
12456 | !DT->isReachableFromEntry(P->getIncomingBlock(I))) |
12457 | continue; |
12458 | |
12459 | // Postponed instructions should not be vectorized here, delay their |
12460 | // vectorization. |
12461 | if (auto *PI = dyn_cast<Instruction>(P->getIncomingValue(I)); |
12462 | PI && !PostProcessInstructions.contains(PI)) |
12463 | Changed |= vectorizeRootInstruction(nullptr, P->getIncomingValue(I), |
12464 | P->getIncomingBlock(I), R, TTI); |
12465 | } |
12466 | continue; |
12467 | } |
12468 | |
12469 | // Ran into an instruction without users, like terminator, or function call |
12470 | // with ignored return value, store. Ignore unused instructions (basing on |
12471 | // instruction type, except for CallInst and InvokeInst). |
12472 | if (it->use_empty() && |
12473 | (it->getType()->isVoidTy() || isa<CallInst, InvokeInst>(it))) { |
12474 | KeyNodes.insert(&*it); |
12475 | bool OpsChanged = false; |
12476 | auto *SI = dyn_cast<StoreInst>(it); |
12477 | bool TryToVectorizeRoot = ShouldStartVectorizeHorAtStore || !SI; |
12478 | if (SI) { |
12479 | auto I = Stores.find(getUnderlyingObject(SI->getPointerOperand())); |
12480 | // Try to vectorize chain in store, if this is the only store to the |
12481 | // address in the block. |
12482 | // TODO: This is just a temporarily solution to save compile time. Need |
12483 | // to investigate if we can safely turn on slp-vectorize-hor-store |
12484 | // instead to allow lookup for reduction chains in all non-vectorized |
12485 | // stores (need to check side effects and compile time). |
12486 | TryToVectorizeRoot = (I == Stores.end() || I->second.size() == 1) && |
12487 | SI->getValueOperand()->hasOneUse(); |
12488 | } |
12489 | if (TryToVectorizeRoot) { |
12490 | for (auto *V : it->operand_values()) { |
12491 | // Postponed instructions should not be vectorized here, delay their |
12492 | // vectorization. |
12493 | if (auto *VI = dyn_cast<Instruction>(V); |
12494 | VI && !PostProcessInstructions.contains(VI)) |
12495 | // Try to match and vectorize a horizontal reduction. |
12496 | OpsChanged |= vectorizeRootInstruction(nullptr, V, BB, R, TTI); |
12497 | } |
12498 | } |
12499 | // Start vectorization of post-process list of instructions from the |
12500 | // top-tree instructions to try to vectorize as many instructions as |
12501 | // possible. |
12502 | OpsChanged |= vectorizeSimpleInstructions(PostProcessInstructions, BB, R, |
12503 | it->isTerminator()); |
12504 | if (OpsChanged) { |
12505 | // We would like to start over since some instructions are deleted |
12506 | // and the iterator may become invalid value. |
12507 | Changed = true; |
12508 | it = BB->begin(); |
12509 | e = BB->end(); |
12510 | continue; |
12511 | } |
12512 | } |
12513 | |
12514 | if (isa<CmpInst, InsertElementInst, InsertValueInst>(it)) |
12515 | PostProcessInstructions.insert(&*it); |
12516 | } |
12517 | |
12518 | return Changed; |
12519 | } |
12520 | |
12521 | bool SLPVectorizerPass::vectorizeGEPIndices(BasicBlock *BB, BoUpSLP &R) { |
12522 | auto Changed = false; |
12523 | for (auto &Entry : GEPs) { |
12524 | // If the getelementptr list has fewer than two elements, there's nothing |
12525 | // to do. |
12526 | if (Entry.second.size() < 2) |
12527 | continue; |
12528 | |
12529 | 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 ) |
12530 | << 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 ); |
12531 | |
12532 | // Process the GEP list in chunks suitable for the target's supported |
12533 | // vector size. If a vector register can't hold 1 element, we are done. We |
12534 | // are trying to vectorize the index computations, so the maximum number of |
12535 | // elements is based on the size of the index expression, rather than the |
12536 | // size of the GEP itself (the target's pointer size). |
12537 | unsigned MaxVecRegSize = R.getMaxVecRegSize(); |
12538 | unsigned EltSize = R.getVectorElementSize(*Entry.second[0]->idx_begin()); |
12539 | if (MaxVecRegSize < EltSize) |
12540 | continue; |
12541 | |
12542 | unsigned MaxElts = MaxVecRegSize / EltSize; |
12543 | for (unsigned BI = 0, BE = Entry.second.size(); BI < BE; BI += MaxElts) { |
12544 | auto Len = std::min<unsigned>(BE - BI, MaxElts); |
12545 | ArrayRef<GetElementPtrInst *> GEPList(&Entry.second[BI], Len); |
12546 | |
12547 | // Initialize a set a candidate getelementptrs. Note that we use a |
12548 | // SetVector here to preserve program order. If the index computations |
12549 | // are vectorizable and begin with loads, we want to minimize the chance |
12550 | // of having to reorder them later. |
12551 | SetVector<Value *> Candidates(GEPList.begin(), GEPList.end()); |
12552 | |
12553 | // Some of the candidates may have already been vectorized after we |
12554 | // initially collected them. If so, they are marked as deleted, so remove |
12555 | // them from the set of candidates. |
12556 | Candidates.remove_if( |
12557 | [&R](Value *I) { return R.isDeleted(cast<Instruction>(I)); }); |
12558 | |
12559 | // Remove from the set of candidates all pairs of getelementptrs with |
12560 | // constant differences. Such getelementptrs are likely not good |
12561 | // candidates for vectorization in a bottom-up phase since one can be |
12562 | // computed from the other. We also ensure all candidate getelementptr |
12563 | // indices are unique. |
12564 | for (int I = 0, E = GEPList.size(); I < E && Candidates.size() > 1; ++I) { |
12565 | auto *GEPI = GEPList[I]; |
12566 | if (!Candidates.count(GEPI)) |
12567 | continue; |
12568 | auto *SCEVI = SE->getSCEV(GEPList[I]); |
12569 | for (int J = I + 1; J < E && Candidates.size() > 1; ++J) { |
12570 | auto *GEPJ = GEPList[J]; |
12571 | auto *SCEVJ = SE->getSCEV(GEPList[J]); |
12572 | if (isa<SCEVConstant>(SE->getMinusSCEV(SCEVI, SCEVJ))) { |
12573 | Candidates.remove(GEPI); |
12574 | Candidates.remove(GEPJ); |
12575 | } else if (GEPI->idx_begin()->get() == GEPJ->idx_begin()->get()) { |
12576 | Candidates.remove(GEPJ); |
12577 | } |
12578 | } |
12579 | } |
12580 | |
12581 | // We break out of the above computation as soon as we know there are |
12582 | // fewer than two candidates remaining. |
12583 | if (Candidates.size() < 2) |
12584 | continue; |
12585 | |
12586 | // Add the single, non-constant index of each candidate to the bundle. We |
12587 | // ensured the indices met these constraints when we originally collected |
12588 | // the getelementptrs. |
12589 | SmallVector<Value *, 16> Bundle(Candidates.size()); |
12590 | auto BundleIndex = 0u; |
12591 | for (auto *V : Candidates) { |
12592 | auto *GEP = cast<GetElementPtrInst>(V); |
12593 | auto *GEPIdx = GEP->idx_begin()->get(); |
12594 | 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", 12594, __extension__ __PRETTY_FUNCTION__)); |
12595 | Bundle[BundleIndex++] = GEPIdx; |
12596 | } |
12597 | |
12598 | // Try and vectorize the indices. We are currently only interested in |
12599 | // gather-like cases of the form: |
12600 | // |
12601 | // ... = g[a[0] - b[0]] + g[a[1] - b[1]] + ... |
12602 | // |
12603 | // where the loads of "a", the loads of "b", and the subtractions can be |
12604 | // performed in parallel. It's likely that detecting this pattern in a |
12605 | // bottom-up phase will be simpler and less costly than building a |
12606 | // full-blown top-down phase beginning at the consecutive loads. |
12607 | Changed |= tryToVectorizeList(Bundle, R); |
12608 | } |
12609 | } |
12610 | return Changed; |
12611 | } |
12612 | |
12613 | bool SLPVectorizerPass::vectorizeStoreChains(BoUpSLP &R) { |
12614 | bool Changed = false; |
12615 | // Sort by type, base pointers and values operand. Value operands must be |
12616 | // compatible (have the same opcode, same parent), otherwise it is |
12617 | // definitely not profitable to try to vectorize them. |
12618 | auto &&StoreSorter = [this](StoreInst *V, StoreInst *V2) { |
12619 | if (V->getPointerOperandType()->getTypeID() < |
12620 | V2->getPointerOperandType()->getTypeID()) |
12621 | return true; |
12622 | if (V->getPointerOperandType()->getTypeID() > |
12623 | V2->getPointerOperandType()->getTypeID()) |
12624 | return false; |
12625 | // UndefValues are compatible with all other values. |
12626 | if (isa<UndefValue>(V->getValueOperand()) || |
12627 | isa<UndefValue>(V2->getValueOperand())) |
12628 | return false; |
12629 | if (auto *I1 = dyn_cast<Instruction>(V->getValueOperand())) |
12630 | if (auto *I2 = dyn_cast<Instruction>(V2->getValueOperand())) { |
12631 | DomTreeNodeBase<llvm::BasicBlock> *NodeI1 = |
12632 | DT->getNode(I1->getParent()); |
12633 | DomTreeNodeBase<llvm::BasicBlock> *NodeI2 = |
12634 | DT->getNode(I2->getParent()); |
12635 | 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", 12635, __extension__ __PRETTY_FUNCTION__)); |
12636 | 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", 12636, __extension__ __PRETTY_FUNCTION__)); |
12637 | 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", 12639, __extension__ __PRETTY_FUNCTION__)) |
12638 | (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", 12639, __extension__ __PRETTY_FUNCTION__)) |
12639 | "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", 12639, __extension__ __PRETTY_FUNCTION__)); |
12640 | if (NodeI1 != NodeI2) |
12641 | return NodeI1->getDFSNumIn() < NodeI2->getDFSNumIn(); |
12642 | InstructionsState S = getSameOpcode({I1, I2}); |
12643 | if (S.getOpcode()) |
12644 | return false; |
12645 | return I1->getOpcode() < I2->getOpcode(); |
12646 | } |
12647 | if (isa<Constant>(V->getValueOperand()) && |
12648 | isa<Constant>(V2->getValueOperand())) |
12649 | return false; |
12650 | return V->getValueOperand()->getValueID() < |
12651 | V2->getValueOperand()->getValueID(); |
12652 | }; |
12653 | |
12654 | auto &&AreCompatibleStores = [](StoreInst *V1, StoreInst *V2) { |
12655 | if (V1 == V2) |
12656 | return true; |
12657 | if (V1->getPointerOperandType() != V2->getPointerOperandType()) |
12658 | return false; |
12659 | // Undefs are compatible with any other value. |
12660 | if (isa<UndefValue>(V1->getValueOperand()) || |
12661 | isa<UndefValue>(V2->getValueOperand())) |
12662 | return true; |
12663 | if (auto *I1 = dyn_cast<Instruction>(V1->getValueOperand())) |
12664 | if (auto *I2 = dyn_cast<Instruction>(V2->getValueOperand())) { |
12665 | if (I1->getParent() != I2->getParent()) |
12666 | return false; |
12667 | InstructionsState S = getSameOpcode({I1, I2}); |
12668 | return S.getOpcode() > 0; |
12669 | } |
12670 | if (isa<Constant>(V1->getValueOperand()) && |
12671 | isa<Constant>(V2->getValueOperand())) |
12672 | return true; |
12673 | return V1->getValueOperand()->getValueID() == |
12674 | V2->getValueOperand()->getValueID(); |
12675 | }; |
12676 | auto Limit = [&R, this](StoreInst *SI) { |
12677 | unsigned EltSize = DL->getTypeSizeInBits(SI->getValueOperand()->getType()); |
12678 | return R.getMinVF(EltSize); |
12679 | }; |
12680 | |
12681 | // Attempt to sort and vectorize each of the store-groups. |
12682 | for (auto &Pair : Stores) { |
12683 | if (Pair.second.size() < 2) |
12684 | continue; |
12685 | |
12686 | 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 ) |
12687 | << 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 ); |
12688 | |
12689 | if (!isValidElementType(Pair.second.front()->getValueOperand()->getType())) |
12690 | continue; |
12691 | |
12692 | Changed |= tryToVectorizeSequence<StoreInst>( |
12693 | Pair.second, Limit, StoreSorter, AreCompatibleStores, |
12694 | [this, &R](ArrayRef<StoreInst *> Candidates, bool) { |
12695 | return vectorizeStores(Candidates, R); |
12696 | }, |
12697 | /*LimitForRegisterSize=*/false); |
12698 | } |
12699 | return Changed; |
12700 | } |
12701 | |
12702 | char SLPVectorizer::ID = 0; |
12703 | |
12704 | static const char lv_name[] = "SLP Vectorizer"; |
12705 | |
12706 | INITIALIZE_PASS_BEGIN(SLPVectorizer, SV_NAME, lv_name, false, false)static void *initializeSLPVectorizerPassOnce(PassRegistry & Registry) { |
12707 | INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)initializeAAResultsWrapperPassPass(Registry); |
12708 | INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)initializeTargetTransformInfoWrapperPassPass(Registry); |
12709 | INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)initializeAssumptionCacheTrackerPass(Registry); |
12710 | INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)initializeScalarEvolutionWrapperPassPass(Registry); |
12711 | INITIALIZE_PASS_DEPENDENCY(LoopSimplify)initializeLoopSimplifyPass(Registry); |
12712 | INITIALIZE_PASS_DEPENDENCY(DemandedBitsWrapperPass)initializeDemandedBitsWrapperPassPass(Registry); |
12713 | INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass)initializeOptimizationRemarkEmitterWrapperPassPass(Registry); |
12714 | INITIALIZE_PASS_DEPENDENCY(InjectTLIMappingsLegacy)initializeInjectTLIMappingsLegacyPass(Registry); |
12715 | INITIALIZE_PASS_END(SLPVectorizer, SV_NAME, lv_name, false, false)PassInfo *PI = new PassInfo( lv_name, "slp-vectorizer", & SLPVectorizer::ID, PassInfo::NormalCtor_t(callDefaultCtor< SLPVectorizer>), false, false); Registry.registerPass(*PI, true); return PI; } static llvm::once_flag InitializeSLPVectorizerPassFlag ; void llvm::initializeSLPVectorizerPass(PassRegistry &Registry ) { llvm::call_once(InitializeSLPVectorizerPassFlag, initializeSLPVectorizerPassOnce , std::ref(Registry)); } |
12716 | |
12717 | Pass *llvm::createSLPVectorizerPass() { return new SLPVectorizer(); } |