File: | llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp |
Warning: | line 4924, column 22 Called C++ object pointer is null |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | //===- SLPVectorizer.cpp - A bottom up SLP Vectorizer ---------------------===// | ||||
2 | // | ||||
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. | ||||
4 | // See https://llvm.org/LICENSE.txt for license information. | ||||
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception | ||||
6 | // | ||||
7 | //===----------------------------------------------------------------------===// | ||||
8 | // | ||||
9 | // This pass implements the Bottom Up SLP vectorizer. It detects consecutive | ||||
10 | // stores that can be put together into vector-stores. Next, it attempts to | ||||
11 | // construct vectorizable tree using the use-def chains. If a profitable tree | ||||
12 | // was found, the SLP vectorizer performs vectorization on the tree. | ||||
13 | // | ||||
14 | // The pass is inspired by the work described in the paper: | ||||
15 | // "Loop-Aware SLP in GCC" by Ira Rosen, Dorit Nuzman, Ayal Zaks. | ||||
16 | // | ||||
17 | //===----------------------------------------------------------------------===// | ||||
18 | |||||
19 | #include "llvm/Transforms/Vectorize/SLPVectorizer.h" | ||||
20 | #include "llvm/ADT/DenseMap.h" | ||||
21 | #include "llvm/ADT/DenseSet.h" | ||||
22 | #include "llvm/ADT/Optional.h" | ||||
23 | #include "llvm/ADT/PostOrderIterator.h" | ||||
24 | #include "llvm/ADT/STLExtras.h" | ||||
25 | #include "llvm/ADT/SetVector.h" | ||||
26 | #include "llvm/ADT/SmallBitVector.h" | ||||
27 | #include "llvm/ADT/SmallPtrSet.h" | ||||
28 | #include "llvm/ADT/SmallSet.h" | ||||
29 | #include "llvm/ADT/SmallString.h" | ||||
30 | #include "llvm/ADT/Statistic.h" | ||||
31 | #include "llvm/ADT/iterator.h" | ||||
32 | #include "llvm/ADT/iterator_range.h" | ||||
33 | #include "llvm/Analysis/AliasAnalysis.h" | ||||
34 | #include "llvm/Analysis/AssumptionCache.h" | ||||
35 | #include "llvm/Analysis/CodeMetrics.h" | ||||
36 | #include "llvm/Analysis/DemandedBits.h" | ||||
37 | #include "llvm/Analysis/GlobalsModRef.h" | ||||
38 | #include "llvm/Analysis/LoopAccessAnalysis.h" | ||||
39 | #include "llvm/Analysis/LoopInfo.h" | ||||
40 | #include "llvm/Analysis/MemoryLocation.h" | ||||
41 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" | ||||
42 | #include "llvm/Analysis/ScalarEvolution.h" | ||||
43 | #include "llvm/Analysis/ScalarEvolutionExpressions.h" | ||||
44 | #include "llvm/Analysis/TargetLibraryInfo.h" | ||||
45 | #include "llvm/Analysis/TargetTransformInfo.h" | ||||
46 | #include "llvm/Analysis/ValueTracking.h" | ||||
47 | #include "llvm/Analysis/VectorUtils.h" | ||||
48 | #include "llvm/IR/Attributes.h" | ||||
49 | #include "llvm/IR/BasicBlock.h" | ||||
50 | #include "llvm/IR/Constant.h" | ||||
51 | #include "llvm/IR/Constants.h" | ||||
52 | #include "llvm/IR/DataLayout.h" | ||||
53 | #include "llvm/IR/DebugLoc.h" | ||||
54 | #include "llvm/IR/DerivedTypes.h" | ||||
55 | #include "llvm/IR/Dominators.h" | ||||
56 | #include "llvm/IR/Function.h" | ||||
57 | #include "llvm/IR/IRBuilder.h" | ||||
58 | #include "llvm/IR/InstrTypes.h" | ||||
59 | #include "llvm/IR/Instruction.h" | ||||
60 | #include "llvm/IR/Instructions.h" | ||||
61 | #include "llvm/IR/IntrinsicInst.h" | ||||
62 | #include "llvm/IR/Intrinsics.h" | ||||
63 | #include "llvm/IR/Module.h" | ||||
64 | #include "llvm/IR/NoFolder.h" | ||||
65 | #include "llvm/IR/Operator.h" | ||||
66 | #include "llvm/IR/PatternMatch.h" | ||||
67 | #include "llvm/IR/Type.h" | ||||
68 | #include "llvm/IR/Use.h" | ||||
69 | #include "llvm/IR/User.h" | ||||
70 | #include "llvm/IR/Value.h" | ||||
71 | #include "llvm/IR/ValueHandle.h" | ||||
72 | #include "llvm/IR/Verifier.h" | ||||
73 | #include "llvm/InitializePasses.h" | ||||
74 | #include "llvm/Pass.h" | ||||
75 | #include "llvm/Support/Casting.h" | ||||
76 | #include "llvm/Support/CommandLine.h" | ||||
77 | #include "llvm/Support/Compiler.h" | ||||
78 | #include "llvm/Support/DOTGraphTraits.h" | ||||
79 | #include "llvm/Support/Debug.h" | ||||
80 | #include "llvm/Support/ErrorHandling.h" | ||||
81 | #include "llvm/Support/GraphWriter.h" | ||||
82 | #include "llvm/Support/KnownBits.h" | ||||
83 | #include "llvm/Support/MathExtras.h" | ||||
84 | #include "llvm/Support/raw_ostream.h" | ||||
85 | #include "llvm/Transforms/Utils/InjectTLIMappings.h" | ||||
86 | #include "llvm/Transforms/Utils/LoopUtils.h" | ||||
87 | #include "llvm/Transforms/Vectorize.h" | ||||
88 | #include <algorithm> | ||||
89 | #include <cassert> | ||||
90 | #include <cstdint> | ||||
91 | #include <iterator> | ||||
92 | #include <memory> | ||||
93 | #include <set> | ||||
94 | #include <string> | ||||
95 | #include <tuple> | ||||
96 | #include <utility> | ||||
97 | #include <vector> | ||||
98 | |||||
99 | using namespace llvm; | ||||
100 | using namespace llvm::PatternMatch; | ||||
101 | using namespace slpvectorizer; | ||||
102 | |||||
103 | #define SV_NAME"slp-vectorizer" "slp-vectorizer" | ||||
104 | #define DEBUG_TYPE"SLP" "SLP" | ||||
105 | |||||
106 | STATISTIC(NumVectorInstructions, "Number of vector instructions generated")static llvm::Statistic NumVectorInstructions = {"SLP", "NumVectorInstructions" , "Number of vector instructions generated"}; | ||||
107 | |||||
108 | cl::opt<bool> RunSLPVectorization("vectorize-slp", cl::init(true), cl::Hidden, | ||||
109 | cl::desc("Run the SLP vectorization passes")); | ||||
110 | |||||
111 | static cl::opt<int> | ||||
112 | SLPCostThreshold("slp-threshold", cl::init(0), cl::Hidden, | ||||
113 | cl::desc("Only vectorize if you gain more than this " | ||||
114 | "number ")); | ||||
115 | |||||
116 | static cl::opt<bool> | ||||
117 | ShouldVectorizeHor("slp-vectorize-hor", cl::init(true), cl::Hidden, | ||||
118 | cl::desc("Attempt to vectorize horizontal reductions")); | ||||
119 | |||||
120 | static cl::opt<bool> ShouldStartVectorizeHorAtStore( | ||||
121 | "slp-vectorize-hor-store", cl::init(false), cl::Hidden, | ||||
122 | cl::desc( | ||||
123 | "Attempt to vectorize horizontal reductions feeding into a store")); | ||||
124 | |||||
125 | static cl::opt<int> | ||||
126 | MaxVectorRegSizeOption("slp-max-reg-size", cl::init(128), cl::Hidden, | ||||
127 | cl::desc("Attempt to vectorize for this register size in bits")); | ||||
128 | |||||
129 | static cl::opt<int> | ||||
130 | MaxStoreLookup("slp-max-store-lookup", cl::init(32), cl::Hidden, | ||||
131 | cl::desc("Maximum depth of the lookup for consecutive stores.")); | ||||
132 | |||||
133 | /// Limits the size of scheduling regions in a block. | ||||
134 | /// It avoid long compile times for _very_ large blocks where vector | ||||
135 | /// instructions are spread over a wide range. | ||||
136 | /// This limit is way higher than needed by real-world functions. | ||||
137 | static cl::opt<int> | ||||
138 | ScheduleRegionSizeBudget("slp-schedule-budget", cl::init(100000), cl::Hidden, | ||||
139 | cl::desc("Limit the size of the SLP scheduling region per block")); | ||||
140 | |||||
141 | static cl::opt<int> MinVectorRegSizeOption( | ||||
142 | "slp-min-reg-size", cl::init(128), cl::Hidden, | ||||
143 | cl::desc("Attempt to vectorize for this register size in bits")); | ||||
144 | |||||
145 | static cl::opt<unsigned> RecursionMaxDepth( | ||||
146 | "slp-recursion-max-depth", cl::init(12), cl::Hidden, | ||||
147 | cl::desc("Limit the recursion depth when building a vectorizable tree")); | ||||
148 | |||||
149 | static cl::opt<unsigned> MinTreeSize( | ||||
150 | "slp-min-tree-size", cl::init(3), cl::Hidden, | ||||
151 | cl::desc("Only vectorize small trees if they are fully vectorizable")); | ||||
152 | |||||
153 | // The maximum depth that the look-ahead score heuristic will explore. | ||||
154 | // The higher this value, the higher the compilation time overhead. | ||||
155 | static cl::opt<int> LookAheadMaxDepth( | ||||
156 | "slp-max-look-ahead-depth", cl::init(2), cl::Hidden, | ||||
157 | cl::desc("The maximum look-ahead depth for operand reordering scores")); | ||||
158 | |||||
159 | // The Look-ahead heuristic goes through the users of the bundle to calculate | ||||
160 | // the users cost in getExternalUsesCost(). To avoid compilation time increase | ||||
161 | // we limit the number of users visited to this value. | ||||
162 | static cl::opt<unsigned> LookAheadUsersBudget( | ||||
163 | "slp-look-ahead-users-budget", cl::init(2), cl::Hidden, | ||||
164 | cl::desc("The maximum number of users to visit while visiting the " | ||||
165 | "predecessors. This prevents compilation time increase.")); | ||||
166 | |||||
167 | static cl::opt<bool> | ||||
168 | ViewSLPTree("view-slp-tree", cl::Hidden, | ||||
169 | cl::desc("Display the SLP trees with Graphviz")); | ||||
170 | |||||
171 | // Limit the number of alias checks. The limit is chosen so that | ||||
172 | // it has no negative effect on the llvm benchmarks. | ||||
173 | static const unsigned AliasedCheckLimit = 10; | ||||
174 | |||||
175 | // Another limit for the alias checks: The maximum distance between load/store | ||||
176 | // instructions where alias checks are done. | ||||
177 | // This limit is useful for very large basic blocks. | ||||
178 | static const unsigned MaxMemDepDistance = 160; | ||||
179 | |||||
180 | /// If the ScheduleRegionSizeBudget is exhausted, we allow small scheduling | ||||
181 | /// regions to be handled. | ||||
182 | static const int MinScheduleRegionSize = 16; | ||||
183 | |||||
184 | /// Predicate for the element types that the SLP vectorizer supports. | ||||
185 | /// | ||||
186 | /// The most important thing to filter here are types which are invalid in LLVM | ||||
187 | /// vectors. We also filter target specific types which have absolutely no | ||||
188 | /// meaningful vectorization path such as x86_fp80 and ppc_f128. This just | ||||
189 | /// avoids spending time checking the cost model and realizing that they will | ||||
190 | /// be inevitably scalarized. | ||||
191 | static bool isValidElementType(Type *Ty) { | ||||
192 | return VectorType::isValidElementType(Ty) && !Ty->isX86_FP80Ty() && | ||||
193 | !Ty->isPPC_FP128Ty(); | ||||
194 | } | ||||
195 | |||||
196 | /// \returns true if all of the instructions in \p VL are in the same block or | ||||
197 | /// false otherwise. | ||||
198 | static bool allSameBlock(ArrayRef<Value *> VL) { | ||||
199 | Instruction *I0 = dyn_cast<Instruction>(VL[0]); | ||||
200 | if (!I0) | ||||
201 | return false; | ||||
202 | BasicBlock *BB = I0->getParent(); | ||||
203 | for (int I = 1, E = VL.size(); I < E; I++) { | ||||
204 | auto *II = dyn_cast<Instruction>(VL[I]); | ||||
205 | if (!II) | ||||
206 | return false; | ||||
207 | |||||
208 | if (BB != II->getParent()) | ||||
209 | return false; | ||||
210 | } | ||||
211 | return true; | ||||
212 | } | ||||
213 | |||||
214 | /// \returns True if all of the values in \p VL are constants (but not | ||||
215 | /// globals/constant expressions). | ||||
216 | static bool allConstant(ArrayRef<Value *> VL) { | ||||
217 | // Constant expressions and globals can't be vectorized like normal integer/FP | ||||
218 | // constants. | ||||
219 | for (Value *i : VL) | ||||
220 | if (!isa<Constant>(i) || isa<ConstantExpr>(i) || isa<GlobalValue>(i)) | ||||
221 | return false; | ||||
222 | return true; | ||||
223 | } | ||||
224 | |||||
225 | /// \returns True if all of the values in \p VL are identical. | ||||
226 | static bool isSplat(ArrayRef<Value *> VL) { | ||||
227 | for (unsigned i = 1, e = VL.size(); i < e; ++i) | ||||
228 | if (VL[i] != VL[0]) | ||||
229 | return false; | ||||
230 | return true; | ||||
231 | } | ||||
232 | |||||
233 | /// \returns True if \p I is commutative, handles CmpInst and BinaryOperator. | ||||
234 | static bool isCommutative(Instruction *I) { | ||||
235 | if (auto *Cmp = dyn_cast<CmpInst>(I)) | ||||
236 | return Cmp->isCommutative(); | ||||
237 | if (auto *BO = dyn_cast<BinaryOperator>(I)) | ||||
238 | return BO->isCommutative(); | ||||
239 | // TODO: This should check for generic Instruction::isCommutative(), but | ||||
240 | // we need to confirm that the caller code correctly handles Intrinsics | ||||
241 | // for example (does not have 2 operands). | ||||
242 | return false; | ||||
243 | } | ||||
244 | |||||
245 | /// Checks if the vector of instructions can be represented as a shuffle, like: | ||||
246 | /// %x0 = extractelement <4 x i8> %x, i32 0 | ||||
247 | /// %x3 = extractelement <4 x i8> %x, i32 3 | ||||
248 | /// %y1 = extractelement <4 x i8> %y, i32 1 | ||||
249 | /// %y2 = extractelement <4 x i8> %y, i32 2 | ||||
250 | /// %x0x0 = mul i8 %x0, %x0 | ||||
251 | /// %x3x3 = mul i8 %x3, %x3 | ||||
252 | /// %y1y1 = mul i8 %y1, %y1 | ||||
253 | /// %y2y2 = mul i8 %y2, %y2 | ||||
254 | /// %ins1 = insertelement <4 x i8> undef, i8 %x0x0, i32 0 | ||||
255 | /// %ins2 = insertelement <4 x i8> %ins1, i8 %x3x3, i32 1 | ||||
256 | /// %ins3 = insertelement <4 x i8> %ins2, i8 %y1y1, i32 2 | ||||
257 | /// %ins4 = insertelement <4 x i8> %ins3, i8 %y2y2, i32 3 | ||||
258 | /// ret <4 x i8> %ins4 | ||||
259 | /// can be transformed into: | ||||
260 | /// %1 = shufflevector <4 x i8> %x, <4 x i8> %y, <4 x i32> <i32 0, i32 3, i32 5, | ||||
261 | /// i32 6> | ||||
262 | /// %2 = mul <4 x i8> %1, %1 | ||||
263 | /// ret <4 x i8> %2 | ||||
264 | /// We convert this initially to something like: | ||||
265 | /// %x0 = extractelement <4 x i8> %x, i32 0 | ||||
266 | /// %x3 = extractelement <4 x i8> %x, i32 3 | ||||
267 | /// %y1 = extractelement <4 x i8> %y, i32 1 | ||||
268 | /// %y2 = extractelement <4 x i8> %y, i32 2 | ||||
269 | /// %1 = insertelement <4 x i8> undef, i8 %x0, i32 0 | ||||
270 | /// %2 = insertelement <4 x i8> %1, i8 %x3, i32 1 | ||||
271 | /// %3 = insertelement <4 x i8> %2, i8 %y1, i32 2 | ||||
272 | /// %4 = insertelement <4 x i8> %3, i8 %y2, i32 3 | ||||
273 | /// %5 = mul <4 x i8> %4, %4 | ||||
274 | /// %6 = extractelement <4 x i8> %5, i32 0 | ||||
275 | /// %ins1 = insertelement <4 x i8> undef, i8 %6, i32 0 | ||||
276 | /// %7 = extractelement <4 x i8> %5, i32 1 | ||||
277 | /// %ins2 = insertelement <4 x i8> %ins1, i8 %7, i32 1 | ||||
278 | /// %8 = extractelement <4 x i8> %5, i32 2 | ||||
279 | /// %ins3 = insertelement <4 x i8> %ins2, i8 %8, i32 2 | ||||
280 | /// %9 = extractelement <4 x i8> %5, i32 3 | ||||
281 | /// %ins4 = insertelement <4 x i8> %ins3, i8 %9, i32 3 | ||||
282 | /// ret <4 x i8> %ins4 | ||||
283 | /// InstCombiner transforms this into a shuffle and vector mul | ||||
284 | /// TODO: Can we split off and reuse the shuffle mask detection from | ||||
285 | /// TargetTransformInfo::getInstructionThroughput? | ||||
286 | static Optional<TargetTransformInfo::ShuffleKind> | ||||
287 | isShuffle(ArrayRef<Value *> VL) { | ||||
288 | auto *EI0 = cast<ExtractElementInst>(VL[0]); | ||||
289 | unsigned Size = | ||||
290 | cast<FixedVectorType>(EI0->getVectorOperandType())->getNumElements(); | ||||
291 | Value *Vec1 = nullptr; | ||||
292 | Value *Vec2 = nullptr; | ||||
293 | enum ShuffleMode { Unknown, Select, Permute }; | ||||
294 | ShuffleMode CommonShuffleMode = Unknown; | ||||
295 | for (unsigned I = 0, E = VL.size(); I < E; ++I) { | ||||
296 | auto *EI = cast<ExtractElementInst>(VL[I]); | ||||
297 | auto *Vec = EI->getVectorOperand(); | ||||
298 | // All vector operands must have the same number of vector elements. | ||||
299 | if (cast<FixedVectorType>(Vec->getType())->getNumElements() != Size) | ||||
300 | return None; | ||||
301 | auto *Idx = dyn_cast<ConstantInt>(EI->getIndexOperand()); | ||||
302 | if (!Idx) | ||||
303 | return None; | ||||
304 | // Undefined behavior if Idx is negative or >= Size. | ||||
305 | if (Idx->getValue().uge(Size)) | ||||
306 | continue; | ||||
307 | unsigned IntIdx = Idx->getValue().getZExtValue(); | ||||
308 | // We can extractelement from undef vector. | ||||
309 | if (isa<UndefValue>(Vec)) | ||||
310 | continue; | ||||
311 | // For correct shuffling we have to have at most 2 different vector operands | ||||
312 | // in all extractelement instructions. | ||||
313 | if (!Vec1 || Vec1 == Vec) | ||||
314 | Vec1 = Vec; | ||||
315 | else if (!Vec2 || Vec2 == Vec) | ||||
316 | Vec2 = Vec; | ||||
317 | else | ||||
318 | return None; | ||||
319 | if (CommonShuffleMode == Permute) | ||||
320 | continue; | ||||
321 | // If the extract index is not the same as the operation number, it is a | ||||
322 | // permutation. | ||||
323 | if (IntIdx != I) { | ||||
324 | CommonShuffleMode = Permute; | ||||
325 | continue; | ||||
326 | } | ||||
327 | CommonShuffleMode = Select; | ||||
328 | } | ||||
329 | // If we're not crossing lanes in different vectors, consider it as blending. | ||||
330 | if (CommonShuffleMode == Select && Vec2) | ||||
331 | return TargetTransformInfo::SK_Select; | ||||
332 | // If Vec2 was never used, we have a permutation of a single vector, otherwise | ||||
333 | // we have permutation of 2 vectors. | ||||
334 | return Vec2 ? TargetTransformInfo::SK_PermuteTwoSrc | ||||
335 | : TargetTransformInfo::SK_PermuteSingleSrc; | ||||
336 | } | ||||
337 | |||||
338 | namespace { | ||||
339 | |||||
340 | /// Main data required for vectorization of instructions. | ||||
341 | struct InstructionsState { | ||||
342 | /// The very first instruction in the list with the main opcode. | ||||
343 | Value *OpValue = nullptr; | ||||
344 | |||||
345 | /// The main/alternate instruction. | ||||
346 | Instruction *MainOp = nullptr; | ||||
347 | Instruction *AltOp = nullptr; | ||||
348 | |||||
349 | /// The main/alternate opcodes for the list of instructions. | ||||
350 | unsigned getOpcode() const { | ||||
351 | return MainOp ? MainOp->getOpcode() : 0; | ||||
352 | } | ||||
353 | |||||
354 | unsigned getAltOpcode() const { | ||||
355 | return AltOp ? AltOp->getOpcode() : 0; | ||||
356 | } | ||||
357 | |||||
358 | /// Some of the instructions in the list have alternate opcodes. | ||||
359 | bool isAltShuffle() const { return getOpcode() != getAltOpcode(); } | ||||
360 | |||||
361 | bool isOpcodeOrAlt(Instruction *I) const { | ||||
362 | unsigned CheckedOpcode = I->getOpcode(); | ||||
363 | return getOpcode() == CheckedOpcode || getAltOpcode() == CheckedOpcode; | ||||
364 | } | ||||
365 | |||||
366 | InstructionsState() = delete; | ||||
367 | InstructionsState(Value *OpValue, Instruction *MainOp, Instruction *AltOp) | ||||
368 | : OpValue(OpValue), MainOp(MainOp), AltOp(AltOp) {} | ||||
369 | }; | ||||
370 | |||||
371 | } // end anonymous namespace | ||||
372 | |||||
373 | /// Chooses the correct key for scheduling data. If \p Op has the same (or | ||||
374 | /// alternate) opcode as \p OpValue, the key is \p Op. Otherwise the key is \p | ||||
375 | /// OpValue. | ||||
376 | static Value *isOneOf(const InstructionsState &S, Value *Op) { | ||||
377 | auto *I = dyn_cast<Instruction>(Op); | ||||
378 | if (I && S.isOpcodeOrAlt(I)) | ||||
379 | return Op; | ||||
380 | return S.OpValue; | ||||
381 | } | ||||
382 | |||||
383 | /// \returns true if \p Opcode is allowed as part of of the main/alternate | ||||
384 | /// instruction for SLP vectorization. | ||||
385 | /// | ||||
386 | /// Example of unsupported opcode is SDIV that can potentially cause UB if the | ||||
387 | /// "shuffled out" lane would result in division by zero. | ||||
388 | static bool isValidForAlternation(unsigned Opcode) { | ||||
389 | if (Instruction::isIntDivRem(Opcode)) | ||||
390 | return false; | ||||
391 | |||||
392 | return true; | ||||
393 | } | ||||
394 | |||||
395 | /// \returns analysis of the Instructions in \p VL described in | ||||
396 | /// InstructionsState, the Opcode that we suppose the whole list | ||||
397 | /// could be vectorized even if its structure is diverse. | ||||
398 | static InstructionsState getSameOpcode(ArrayRef<Value *> VL, | ||||
399 | unsigned BaseIndex = 0) { | ||||
400 | // Make sure these are all Instructions. | ||||
401 | if (llvm::any_of(VL, [](Value *V) { return !isa<Instruction>(V); })) | ||||
402 | return InstructionsState(VL[BaseIndex], nullptr, nullptr); | ||||
403 | |||||
404 | bool IsCastOp = isa<CastInst>(VL[BaseIndex]); | ||||
405 | bool IsBinOp = isa<BinaryOperator>(VL[BaseIndex]); | ||||
406 | unsigned Opcode = cast<Instruction>(VL[BaseIndex])->getOpcode(); | ||||
407 | unsigned AltOpcode = Opcode; | ||||
408 | unsigned AltIndex = BaseIndex; | ||||
409 | |||||
410 | // Check for one alternate opcode from another BinaryOperator. | ||||
411 | // TODO - generalize to support all operators (types, calls etc.). | ||||
412 | for (int Cnt = 0, E = VL.size(); Cnt < E; Cnt++) { | ||||
413 | unsigned InstOpcode = cast<Instruction>(VL[Cnt])->getOpcode(); | ||||
414 | if (IsBinOp && isa<BinaryOperator>(VL[Cnt])) { | ||||
415 | if (InstOpcode == Opcode || InstOpcode == AltOpcode) | ||||
416 | continue; | ||||
417 | if (Opcode == AltOpcode && isValidForAlternation(InstOpcode) && | ||||
418 | isValidForAlternation(Opcode)) { | ||||
419 | AltOpcode = InstOpcode; | ||||
420 | AltIndex = Cnt; | ||||
421 | continue; | ||||
422 | } | ||||
423 | } else if (IsCastOp && isa<CastInst>(VL[Cnt])) { | ||||
424 | Type *Ty0 = cast<Instruction>(VL[BaseIndex])->getOperand(0)->getType(); | ||||
425 | Type *Ty1 = cast<Instruction>(VL[Cnt])->getOperand(0)->getType(); | ||||
426 | if (Ty0 == Ty1) { | ||||
427 | if (InstOpcode == Opcode || InstOpcode == AltOpcode) | ||||
428 | continue; | ||||
429 | if (Opcode == AltOpcode) { | ||||
430 | assert(isValidForAlternation(Opcode) &&((isValidForAlternation(Opcode) && isValidForAlternation (InstOpcode) && "Cast isn't safe for alternation, logic needs to be updated!" ) ? static_cast<void> (0) : __assert_fail ("isValidForAlternation(Opcode) && isValidForAlternation(InstOpcode) && \"Cast isn't safe for alternation, logic needs to be updated!\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 432, __PRETTY_FUNCTION__)) | ||||
431 | isValidForAlternation(InstOpcode) &&((isValidForAlternation(Opcode) && isValidForAlternation (InstOpcode) && "Cast isn't safe for alternation, logic needs to be updated!" ) ? static_cast<void> (0) : __assert_fail ("isValidForAlternation(Opcode) && isValidForAlternation(InstOpcode) && \"Cast isn't safe for alternation, logic needs to be updated!\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 432, __PRETTY_FUNCTION__)) | ||||
432 | "Cast isn't safe for alternation, logic needs to be updated!")((isValidForAlternation(Opcode) && isValidForAlternation (InstOpcode) && "Cast isn't safe for alternation, logic needs to be updated!" ) ? static_cast<void> (0) : __assert_fail ("isValidForAlternation(Opcode) && isValidForAlternation(InstOpcode) && \"Cast isn't safe for alternation, logic needs to be updated!\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 432, __PRETTY_FUNCTION__)); | ||||
433 | AltOpcode = InstOpcode; | ||||
434 | AltIndex = Cnt; | ||||
435 | continue; | ||||
436 | } | ||||
437 | } | ||||
438 | } else if (InstOpcode == Opcode || InstOpcode == AltOpcode) | ||||
439 | continue; | ||||
440 | return InstructionsState(VL[BaseIndex], nullptr, nullptr); | ||||
441 | } | ||||
442 | |||||
443 | return InstructionsState(VL[BaseIndex], cast<Instruction>(VL[BaseIndex]), | ||||
444 | cast<Instruction>(VL[AltIndex])); | ||||
445 | } | ||||
446 | |||||
447 | /// \returns true if all of the values in \p VL have the same type or false | ||||
448 | /// otherwise. | ||||
449 | static bool allSameType(ArrayRef<Value *> VL) { | ||||
450 | Type *Ty = VL[0]->getType(); | ||||
451 | for (int i = 1, e = VL.size(); i < e; i++) | ||||
452 | if (VL[i]->getType() != Ty) | ||||
453 | return false; | ||||
454 | |||||
455 | return true; | ||||
456 | } | ||||
457 | |||||
458 | /// \returns True if Extract{Value,Element} instruction extracts element Idx. | ||||
459 | static Optional<unsigned> getExtractIndex(Instruction *E) { | ||||
460 | unsigned Opcode = E->getOpcode(); | ||||
461 | assert((Opcode == Instruction::ExtractElement ||(((Opcode == Instruction::ExtractElement || Opcode == Instruction ::ExtractValue) && "Expected extractelement or extractvalue instruction." ) ? static_cast<void> (0) : __assert_fail ("(Opcode == Instruction::ExtractElement || Opcode == Instruction::ExtractValue) && \"Expected extractelement or extractvalue instruction.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 463, __PRETTY_FUNCTION__)) | ||||
462 | Opcode == Instruction::ExtractValue) &&(((Opcode == Instruction::ExtractElement || Opcode == Instruction ::ExtractValue) && "Expected extractelement or extractvalue instruction." ) ? static_cast<void> (0) : __assert_fail ("(Opcode == Instruction::ExtractElement || Opcode == Instruction::ExtractValue) && \"Expected extractelement or extractvalue instruction.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 463, __PRETTY_FUNCTION__)) | ||||
463 | "Expected extractelement or extractvalue instruction.")(((Opcode == Instruction::ExtractElement || Opcode == Instruction ::ExtractValue) && "Expected extractelement or extractvalue instruction." ) ? static_cast<void> (0) : __assert_fail ("(Opcode == Instruction::ExtractElement || Opcode == Instruction::ExtractValue) && \"Expected extractelement or extractvalue instruction.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 463, __PRETTY_FUNCTION__)); | ||||
464 | if (Opcode == Instruction::ExtractElement) { | ||||
465 | auto *CI = dyn_cast<ConstantInt>(E->getOperand(1)); | ||||
466 | if (!CI) | ||||
467 | return None; | ||||
468 | return CI->getZExtValue(); | ||||
469 | } | ||||
470 | ExtractValueInst *EI = cast<ExtractValueInst>(E); | ||||
471 | if (EI->getNumIndices() != 1) | ||||
472 | return None; | ||||
473 | return *EI->idx_begin(); | ||||
474 | } | ||||
475 | |||||
476 | /// \returns True if in-tree use also needs extract. This refers to | ||||
477 | /// possible scalar operand in vectorized instruction. | ||||
478 | static bool InTreeUserNeedToExtract(Value *Scalar, Instruction *UserInst, | ||||
479 | TargetLibraryInfo *TLI) { | ||||
480 | unsigned Opcode = UserInst->getOpcode(); | ||||
481 | switch (Opcode) { | ||||
482 | case Instruction::Load: { | ||||
483 | LoadInst *LI = cast<LoadInst>(UserInst); | ||||
484 | return (LI->getPointerOperand() == Scalar); | ||||
485 | } | ||||
486 | case Instruction::Store: { | ||||
487 | StoreInst *SI = cast<StoreInst>(UserInst); | ||||
488 | return (SI->getPointerOperand() == Scalar); | ||||
489 | } | ||||
490 | case Instruction::Call: { | ||||
491 | CallInst *CI = cast<CallInst>(UserInst); | ||||
492 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); | ||||
493 | for (unsigned i = 0, e = CI->getNumArgOperands(); i != e; ++i) { | ||||
494 | if (hasVectorInstrinsicScalarOpd(ID, i)) | ||||
495 | return (CI->getArgOperand(i) == Scalar); | ||||
496 | } | ||||
497 | LLVM_FALLTHROUGH[[gnu::fallthrough]]; | ||||
498 | } | ||||
499 | default: | ||||
500 | return false; | ||||
501 | } | ||||
502 | } | ||||
503 | |||||
504 | /// \returns the AA location that is being access by the instruction. | ||||
505 | static MemoryLocation getLocation(Instruction *I, AAResults *AA) { | ||||
506 | if (StoreInst *SI = dyn_cast<StoreInst>(I)) | ||||
507 | return MemoryLocation::get(SI); | ||||
508 | if (LoadInst *LI = dyn_cast<LoadInst>(I)) | ||||
509 | return MemoryLocation::get(LI); | ||||
510 | return MemoryLocation(); | ||||
511 | } | ||||
512 | |||||
513 | /// \returns True if the instruction is not a volatile or atomic load/store. | ||||
514 | static bool isSimple(Instruction *I) { | ||||
515 | if (LoadInst *LI = dyn_cast<LoadInst>(I)) | ||||
516 | return LI->isSimple(); | ||||
517 | if (StoreInst *SI = dyn_cast<StoreInst>(I)) | ||||
518 | return SI->isSimple(); | ||||
519 | if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I)) | ||||
520 | return !MI->isVolatile(); | ||||
521 | return true; | ||||
522 | } | ||||
523 | |||||
524 | namespace llvm { | ||||
525 | |||||
526 | static void inversePermutation(ArrayRef<unsigned> Indices, | ||||
527 | SmallVectorImpl<int> &Mask) { | ||||
528 | Mask.clear(); | ||||
529 | const unsigned E = Indices.size(); | ||||
530 | Mask.resize(E, E + 1); | ||||
531 | for (unsigned I = 0; I < E; ++I) | ||||
532 | Mask[Indices[I]] = I; | ||||
533 | } | ||||
534 | |||||
535 | namespace slpvectorizer { | ||||
536 | |||||
537 | /// Bottom Up SLP Vectorizer. | ||||
538 | class BoUpSLP { | ||||
539 | struct TreeEntry; | ||||
540 | struct ScheduleData; | ||||
541 | |||||
542 | public: | ||||
543 | using ValueList = SmallVector<Value *, 8>; | ||||
544 | using InstrList = SmallVector<Instruction *, 16>; | ||||
545 | using ValueSet = SmallPtrSet<Value *, 16>; | ||||
546 | using StoreList = SmallVector<StoreInst *, 8>; | ||||
547 | using ExtraValueToDebugLocsMap = | ||||
548 | MapVector<Value *, SmallVector<Instruction *, 2>>; | ||||
549 | using OrdersType = SmallVector<unsigned, 4>; | ||||
550 | |||||
551 | BoUpSLP(Function *Func, ScalarEvolution *Se, TargetTransformInfo *Tti, | ||||
552 | TargetLibraryInfo *TLi, AAResults *Aa, LoopInfo *Li, | ||||
553 | DominatorTree *Dt, AssumptionCache *AC, DemandedBits *DB, | ||||
554 | const DataLayout *DL, OptimizationRemarkEmitter *ORE) | ||||
555 | : F(Func), SE(Se), TTI(Tti), TLI(TLi), AA(Aa), LI(Li), DT(Dt), AC(AC), | ||||
556 | DB(DB), DL(DL), ORE(ORE), Builder(Se->getContext()) { | ||||
557 | CodeMetrics::collectEphemeralValues(F, AC, EphValues); | ||||
558 | // Use the vector register size specified by the target unless overridden | ||||
559 | // by a command-line option. | ||||
560 | // TODO: It would be better to limit the vectorization factor based on | ||||
561 | // data type rather than just register size. For example, x86 AVX has | ||||
562 | // 256-bit registers, but it does not support integer operations | ||||
563 | // at that width (that requires AVX2). | ||||
564 | if (MaxVectorRegSizeOption.getNumOccurrences()) | ||||
565 | MaxVecRegSize = MaxVectorRegSizeOption; | ||||
566 | else | ||||
567 | MaxVecRegSize = TTI->getRegisterBitWidth(true); | ||||
568 | |||||
569 | if (MinVectorRegSizeOption.getNumOccurrences()) | ||||
570 | MinVecRegSize = MinVectorRegSizeOption; | ||||
571 | else | ||||
572 | MinVecRegSize = TTI->getMinVectorRegisterBitWidth(); | ||||
573 | } | ||||
574 | |||||
575 | /// Vectorize the tree that starts with the elements in \p VL. | ||||
576 | /// Returns the vectorized root. | ||||
577 | Value *vectorizeTree(); | ||||
578 | |||||
579 | /// Vectorize the tree but with the list of externally used values \p | ||||
580 | /// ExternallyUsedValues. Values in this MapVector can be replaced but the | ||||
581 | /// generated extractvalue instructions. | ||||
582 | Value *vectorizeTree(ExtraValueToDebugLocsMap &ExternallyUsedValues); | ||||
583 | |||||
584 | /// \returns the cost incurred by unwanted spills and fills, caused by | ||||
585 | /// holding live values over call sites. | ||||
586 | int getSpillCost() const; | ||||
587 | |||||
588 | /// \returns the vectorization cost of the subtree that starts at \p VL. | ||||
589 | /// A negative number means that this is profitable. | ||||
590 | int getTreeCost(); | ||||
591 | |||||
592 | /// Construct a vectorizable tree that starts at \p Roots, ignoring users for | ||||
593 | /// the purpose of scheduling and extraction in the \p UserIgnoreLst. | ||||
594 | void buildTree(ArrayRef<Value *> Roots, | ||||
595 | ArrayRef<Value *> UserIgnoreLst = None); | ||||
596 | |||||
597 | /// Construct a vectorizable tree that starts at \p Roots, ignoring users for | ||||
598 | /// the purpose of scheduling and extraction in the \p UserIgnoreLst taking | ||||
599 | /// into account (and updating it, if required) list of externally used | ||||
600 | /// values stored in \p ExternallyUsedValues. | ||||
601 | void buildTree(ArrayRef<Value *> Roots, | ||||
602 | ExtraValueToDebugLocsMap &ExternallyUsedValues, | ||||
603 | ArrayRef<Value *> UserIgnoreLst = None); | ||||
604 | |||||
605 | /// Clear the internal data structures that are created by 'buildTree'. | ||||
606 | void deleteTree() { | ||||
607 | VectorizableTree.clear(); | ||||
608 | ScalarToTreeEntry.clear(); | ||||
609 | MustGather.clear(); | ||||
610 | ExternalUses.clear(); | ||||
611 | NumOpsWantToKeepOrder.clear(); | ||||
612 | NumOpsWantToKeepOriginalOrder = 0; | ||||
613 | for (auto &Iter : BlocksSchedules) { | ||||
614 | BlockScheduling *BS = Iter.second.get(); | ||||
615 | BS->clear(); | ||||
616 | } | ||||
617 | MinBWs.clear(); | ||||
618 | } | ||||
619 | |||||
620 | unsigned getTreeSize() const { return VectorizableTree.size(); } | ||||
621 | |||||
622 | /// Perform LICM and CSE on the newly generated gather sequences. | ||||
623 | void optimizeGatherSequence(); | ||||
624 | |||||
625 | /// \returns The best order of instructions for vectorization. | ||||
626 | Optional<ArrayRef<unsigned>> bestOrder() const { | ||||
627 | assert(llvm::all_of(((llvm::all_of( NumOpsWantToKeepOrder, [this](const decltype( NumOpsWantToKeepOrder)::value_type &D) { return D.getFirst ().size() == VectorizableTree[0]->Scalars.size(); }) && "All orders must have the same size as number of instructions in " "tree node.") ? static_cast<void> (0) : __assert_fail ( "llvm::all_of( NumOpsWantToKeepOrder, [this](const decltype(NumOpsWantToKeepOrder)::value_type &D) { return D.getFirst().size() == VectorizableTree[0]->Scalars.size(); }) && \"All orders must have the same size as number of instructions in \" \"tree node.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 634, __PRETTY_FUNCTION__)) | ||||
628 | NumOpsWantToKeepOrder,((llvm::all_of( NumOpsWantToKeepOrder, [this](const decltype( NumOpsWantToKeepOrder)::value_type &D) { return D.getFirst ().size() == VectorizableTree[0]->Scalars.size(); }) && "All orders must have the same size as number of instructions in " "tree node.") ? static_cast<void> (0) : __assert_fail ( "llvm::all_of( NumOpsWantToKeepOrder, [this](const decltype(NumOpsWantToKeepOrder)::value_type &D) { return D.getFirst().size() == VectorizableTree[0]->Scalars.size(); }) && \"All orders must have the same size as number of instructions in \" \"tree node.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 634, __PRETTY_FUNCTION__)) | ||||
629 | [this](const decltype(NumOpsWantToKeepOrder)::value_type &D) {((llvm::all_of( NumOpsWantToKeepOrder, [this](const decltype( NumOpsWantToKeepOrder)::value_type &D) { return D.getFirst ().size() == VectorizableTree[0]->Scalars.size(); }) && "All orders must have the same size as number of instructions in " "tree node.") ? static_cast<void> (0) : __assert_fail ( "llvm::all_of( NumOpsWantToKeepOrder, [this](const decltype(NumOpsWantToKeepOrder)::value_type &D) { return D.getFirst().size() == VectorizableTree[0]->Scalars.size(); }) && \"All orders must have the same size as number of instructions in \" \"tree node.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 634, __PRETTY_FUNCTION__)) | ||||
630 | return D.getFirst().size() ==((llvm::all_of( NumOpsWantToKeepOrder, [this](const decltype( NumOpsWantToKeepOrder)::value_type &D) { return D.getFirst ().size() == VectorizableTree[0]->Scalars.size(); }) && "All orders must have the same size as number of instructions in " "tree node.") ? static_cast<void> (0) : __assert_fail ( "llvm::all_of( NumOpsWantToKeepOrder, [this](const decltype(NumOpsWantToKeepOrder)::value_type &D) { return D.getFirst().size() == VectorizableTree[0]->Scalars.size(); }) && \"All orders must have the same size as number of instructions in \" \"tree node.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 634, __PRETTY_FUNCTION__)) | ||||
631 | VectorizableTree[0]->Scalars.size();((llvm::all_of( NumOpsWantToKeepOrder, [this](const decltype( NumOpsWantToKeepOrder)::value_type &D) { return D.getFirst ().size() == VectorizableTree[0]->Scalars.size(); }) && "All orders must have the same size as number of instructions in " "tree node.") ? static_cast<void> (0) : __assert_fail ( "llvm::all_of( NumOpsWantToKeepOrder, [this](const decltype(NumOpsWantToKeepOrder)::value_type &D) { return D.getFirst().size() == VectorizableTree[0]->Scalars.size(); }) && \"All orders must have the same size as number of instructions in \" \"tree node.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 634, __PRETTY_FUNCTION__)) | ||||
632 | }) &&((llvm::all_of( NumOpsWantToKeepOrder, [this](const decltype( NumOpsWantToKeepOrder)::value_type &D) { return D.getFirst ().size() == VectorizableTree[0]->Scalars.size(); }) && "All orders must have the same size as number of instructions in " "tree node.") ? static_cast<void> (0) : __assert_fail ( "llvm::all_of( NumOpsWantToKeepOrder, [this](const decltype(NumOpsWantToKeepOrder)::value_type &D) { return D.getFirst().size() == VectorizableTree[0]->Scalars.size(); }) && \"All orders must have the same size as number of instructions in \" \"tree node.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 634, __PRETTY_FUNCTION__)) | ||||
633 | "All orders must have the same size as number of instructions in "((llvm::all_of( NumOpsWantToKeepOrder, [this](const decltype( NumOpsWantToKeepOrder)::value_type &D) { return D.getFirst ().size() == VectorizableTree[0]->Scalars.size(); }) && "All orders must have the same size as number of instructions in " "tree node.") ? static_cast<void> (0) : __assert_fail ( "llvm::all_of( NumOpsWantToKeepOrder, [this](const decltype(NumOpsWantToKeepOrder)::value_type &D) { return D.getFirst().size() == VectorizableTree[0]->Scalars.size(); }) && \"All orders must have the same size as number of instructions in \" \"tree node.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 634, __PRETTY_FUNCTION__)) | ||||
634 | "tree node.")((llvm::all_of( NumOpsWantToKeepOrder, [this](const decltype( NumOpsWantToKeepOrder)::value_type &D) { return D.getFirst ().size() == VectorizableTree[0]->Scalars.size(); }) && "All orders must have the same size as number of instructions in " "tree node.") ? static_cast<void> (0) : __assert_fail ( "llvm::all_of( NumOpsWantToKeepOrder, [this](const decltype(NumOpsWantToKeepOrder)::value_type &D) { return D.getFirst().size() == VectorizableTree[0]->Scalars.size(); }) && \"All orders must have the same size as number of instructions in \" \"tree node.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 634, __PRETTY_FUNCTION__)); | ||||
635 | auto I = std::max_element( | ||||
636 | NumOpsWantToKeepOrder.begin(), NumOpsWantToKeepOrder.end(), | ||||
637 | [](const decltype(NumOpsWantToKeepOrder)::value_type &D1, | ||||
638 | const decltype(NumOpsWantToKeepOrder)::value_type &D2) { | ||||
639 | return D1.second < D2.second; | ||||
640 | }); | ||||
641 | if (I == NumOpsWantToKeepOrder.end() || | ||||
642 | I->getSecond() <= NumOpsWantToKeepOriginalOrder) | ||||
643 | return None; | ||||
644 | |||||
645 | return makeArrayRef(I->getFirst()); | ||||
646 | } | ||||
647 | |||||
648 | /// Builds the correct order for root instructions. | ||||
649 | /// If some leaves have the same instructions to be vectorized, we may | ||||
650 | /// incorrectly evaluate the best order for the root node (it is built for the | ||||
651 | /// vector of instructions without repeated instructions and, thus, has less | ||||
652 | /// elements than the root node). This function builds the correct order for | ||||
653 | /// the root node. | ||||
654 | /// For example, if the root node is \<a+b, a+c, a+d, f+e\>, then the leaves | ||||
655 | /// are \<a, a, a, f\> and \<b, c, d, e\>. When we try to vectorize the first | ||||
656 | /// leaf, it will be shrink to \<a, b\>. If instructions in this leaf should | ||||
657 | /// be reordered, the best order will be \<1, 0\>. We need to extend this | ||||
658 | /// order for the root node. For the root node this order should look like | ||||
659 | /// \<3, 0, 1, 2\>. This function extends the order for the reused | ||||
660 | /// instructions. | ||||
661 | void findRootOrder(OrdersType &Order) { | ||||
662 | // If the leaf has the same number of instructions to vectorize as the root | ||||
663 | // - order must be set already. | ||||
664 | unsigned RootSize = VectorizableTree[0]->Scalars.size(); | ||||
665 | if (Order.size() == RootSize) | ||||
666 | return; | ||||
667 | SmallVector<unsigned, 4> RealOrder(Order.size()); | ||||
668 | std::swap(Order, RealOrder); | ||||
669 | SmallVector<int, 4> Mask; | ||||
670 | inversePermutation(RealOrder, Mask); | ||||
671 | Order.assign(Mask.begin(), Mask.end()); | ||||
672 | // The leaf has less number of instructions - need to find the true order of | ||||
673 | // the root. | ||||
674 | // Scan the nodes starting from the leaf back to the root. | ||||
675 | const TreeEntry *PNode = VectorizableTree.back().get(); | ||||
676 | SmallVector<const TreeEntry *, 4> Nodes(1, PNode); | ||||
677 | SmallPtrSet<const TreeEntry *, 4> Visited; | ||||
678 | while (!Nodes.empty() && Order.size() != RootSize) { | ||||
679 | const TreeEntry *PNode = Nodes.pop_back_val(); | ||||
680 | if (!Visited.insert(PNode).second) | ||||
681 | continue; | ||||
682 | const TreeEntry &Node = *PNode; | ||||
683 | for (const EdgeInfo &EI : Node.UserTreeIndices) | ||||
684 | if (EI.UserTE) | ||||
685 | Nodes.push_back(EI.UserTE); | ||||
686 | if (Node.ReuseShuffleIndices.empty()) | ||||
687 | continue; | ||||
688 | // Build the order for the parent node. | ||||
689 | OrdersType NewOrder(Node.ReuseShuffleIndices.size(), RootSize); | ||||
690 | SmallVector<unsigned, 4> OrderCounter(Order.size(), 0); | ||||
691 | // The algorithm of the order extension is: | ||||
692 | // 1. Calculate the number of the same instructions for the order. | ||||
693 | // 2. Calculate the index of the new order: total number of instructions | ||||
694 | // with order less than the order of the current instruction + reuse | ||||
695 | // number of the current instruction. | ||||
696 | // 3. The new order is just the index of the instruction in the original | ||||
697 | // vector of the instructions. | ||||
698 | for (unsigned I : Node.ReuseShuffleIndices) | ||||
699 | ++OrderCounter[Order[I]]; | ||||
700 | SmallVector<unsigned, 4> CurrentCounter(Order.size(), 0); | ||||
701 | for (unsigned I = 0, E = Node.ReuseShuffleIndices.size(); I < E; ++I) { | ||||
702 | unsigned ReusedIdx = Node.ReuseShuffleIndices[I]; | ||||
703 | unsigned OrderIdx = Order[ReusedIdx]; | ||||
704 | unsigned NewIdx = 0; | ||||
705 | for (unsigned J = 0; J < OrderIdx; ++J) | ||||
706 | NewIdx += OrderCounter[J]; | ||||
707 | NewIdx += CurrentCounter[OrderIdx]; | ||||
708 | ++CurrentCounter[OrderIdx]; | ||||
709 | assert(NewOrder[NewIdx] == RootSize &&((NewOrder[NewIdx] == RootSize && "The order index should not be written already." ) ? static_cast<void> (0) : __assert_fail ("NewOrder[NewIdx] == RootSize && \"The order index should not be written already.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 710, __PRETTY_FUNCTION__)) | ||||
710 | "The order index should not be written already.")((NewOrder[NewIdx] == RootSize && "The order index should not be written already." ) ? static_cast<void> (0) : __assert_fail ("NewOrder[NewIdx] == RootSize && \"The order index should not be written already.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 710, __PRETTY_FUNCTION__)); | ||||
711 | NewOrder[NewIdx] = I; | ||||
712 | } | ||||
713 | std::swap(Order, NewOrder); | ||||
714 | } | ||||
715 | assert(Order.size() == RootSize &&((Order.size() == RootSize && "Root node is expected or the size of the order must be the same as " "the number of elements in the root node.") ? static_cast< void> (0) : __assert_fail ("Order.size() == RootSize && \"Root node is expected or the size of the order must be the same as \" \"the number of elements in the root node.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 717, __PRETTY_FUNCTION__)) | ||||
716 | "Root node is expected or the size of the order must be the same as "((Order.size() == RootSize && "Root node is expected or the size of the order must be the same as " "the number of elements in the root node.") ? static_cast< void> (0) : __assert_fail ("Order.size() == RootSize && \"Root node is expected or the size of the order must be the same as \" \"the number of elements in the root node.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 717, __PRETTY_FUNCTION__)) | ||||
717 | "the number of elements in the root node.")((Order.size() == RootSize && "Root node is expected or the size of the order must be the same as " "the number of elements in the root node.") ? static_cast< void> (0) : __assert_fail ("Order.size() == RootSize && \"Root node is expected or the size of the order must be the same as \" \"the number of elements in the root node.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 717, __PRETTY_FUNCTION__)); | ||||
718 | assert(llvm::all_of(Order,((llvm::all_of(Order, [RootSize](unsigned Val) { return Val != RootSize; }) && "All indices must be initialized") ? static_cast<void> (0) : __assert_fail ("llvm::all_of(Order, [RootSize](unsigned Val) { return Val != RootSize; }) && \"All indices must be initialized\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 720, __PRETTY_FUNCTION__)) | ||||
719 | [RootSize](unsigned Val) { return Val != RootSize; }) &&((llvm::all_of(Order, [RootSize](unsigned Val) { return Val != RootSize; }) && "All indices must be initialized") ? static_cast<void> (0) : __assert_fail ("llvm::all_of(Order, [RootSize](unsigned Val) { return Val != RootSize; }) && \"All indices must be initialized\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 720, __PRETTY_FUNCTION__)) | ||||
720 | "All indices must be initialized")((llvm::all_of(Order, [RootSize](unsigned Val) { return Val != RootSize; }) && "All indices must be initialized") ? static_cast<void> (0) : __assert_fail ("llvm::all_of(Order, [RootSize](unsigned Val) { return Val != RootSize; }) && \"All indices must be initialized\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 720, __PRETTY_FUNCTION__)); | ||||
721 | } | ||||
722 | |||||
723 | /// \return The vector element size in bits to use when vectorizing the | ||||
724 | /// expression tree ending at \p V. If V is a store, the size is the width of | ||||
725 | /// the stored value. Otherwise, the size is the width of the largest loaded | ||||
726 | /// value reaching V. This method is used by the vectorizer to calculate | ||||
727 | /// vectorization factors. | ||||
728 | unsigned getVectorElementSize(Value *V); | ||||
729 | |||||
730 | /// Compute the minimum type sizes required to represent the entries in a | ||||
731 | /// vectorizable tree. | ||||
732 | void computeMinimumValueSizes(); | ||||
733 | |||||
734 | // \returns maximum vector register size as set by TTI or overridden by cl::opt. | ||||
735 | unsigned getMaxVecRegSize() const { | ||||
736 | return MaxVecRegSize; | ||||
737 | } | ||||
738 | |||||
739 | // \returns minimum vector register size as set by cl::opt. | ||||
740 | unsigned getMinVecRegSize() const { | ||||
741 | return MinVecRegSize; | ||||
742 | } | ||||
743 | |||||
744 | /// Check if homogeneous aggregate is isomorphic to some VectorType. | ||||
745 | /// Accepts homogeneous multidimensional aggregate of scalars/vectors like | ||||
746 | /// {[4 x i16], [4 x i16]}, { <2 x float>, <2 x float> }, | ||||
747 | /// {{{i16, i16}, {i16, i16}}, {{i16, i16}, {i16, i16}}} and so on. | ||||
748 | /// | ||||
749 | /// \returns number of elements in vector if isomorphism exists, 0 otherwise. | ||||
750 | unsigned canMapToVector(Type *T, const DataLayout &DL) const; | ||||
751 | |||||
752 | /// \returns True if the VectorizableTree is both tiny and not fully | ||||
753 | /// vectorizable. We do not vectorize such trees. | ||||
754 | bool isTreeTinyAndNotFullyVectorizable() const; | ||||
755 | |||||
756 | /// Assume that a legal-sized 'or'-reduction of shifted/zexted loaded values | ||||
757 | /// can be load combined in the backend. Load combining may not be allowed in | ||||
758 | /// the IR optimizer, so we do not want to alter the pattern. For example, | ||||
759 | /// partially transforming a scalar bswap() pattern into vector code is | ||||
760 | /// effectively impossible for the backend to undo. | ||||
761 | /// TODO: If load combining is allowed in the IR optimizer, this analysis | ||||
762 | /// may not be necessary. | ||||
763 | bool isLoadCombineReductionCandidate(unsigned ReductionOpcode) const; | ||||
764 | |||||
765 | /// Assume that a vector of stores of bitwise-or/shifted/zexted loaded values | ||||
766 | /// can be load combined in the backend. Load combining may not be allowed in | ||||
767 | /// the IR optimizer, so we do not want to alter the pattern. For example, | ||||
768 | /// partially transforming a scalar bswap() pattern into vector code is | ||||
769 | /// effectively impossible for the backend to undo. | ||||
770 | /// TODO: If load combining is allowed in the IR optimizer, this analysis | ||||
771 | /// may not be necessary. | ||||
772 | bool isLoadCombineCandidate() const; | ||||
773 | |||||
774 | OptimizationRemarkEmitter *getORE() { return ORE; } | ||||
775 | |||||
776 | /// This structure holds any data we need about the edges being traversed | ||||
777 | /// during buildTree_rec(). We keep track of: | ||||
778 | /// (i) the user TreeEntry index, and | ||||
779 | /// (ii) the index of the edge. | ||||
780 | struct EdgeInfo { | ||||
781 | EdgeInfo() = default; | ||||
782 | EdgeInfo(TreeEntry *UserTE, unsigned EdgeIdx) | ||||
783 | : UserTE(UserTE), EdgeIdx(EdgeIdx) {} | ||||
784 | /// The user TreeEntry. | ||||
785 | TreeEntry *UserTE = nullptr; | ||||
786 | /// The operand index of the use. | ||||
787 | unsigned EdgeIdx = UINT_MAX(2147483647 *2U +1U); | ||||
788 | #ifndef NDEBUG | ||||
789 | friend inline raw_ostream &operator<<(raw_ostream &OS, | ||||
790 | const BoUpSLP::EdgeInfo &EI) { | ||||
791 | EI.dump(OS); | ||||
792 | return OS; | ||||
793 | } | ||||
794 | /// Debug print. | ||||
795 | void dump(raw_ostream &OS) const { | ||||
796 | OS << "{User:" << (UserTE ? std::to_string(UserTE->Idx) : "null") | ||||
797 | << " EdgeIdx:" << EdgeIdx << "}"; | ||||
798 | } | ||||
799 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void dump() const { dump(dbgs()); } | ||||
800 | #endif | ||||
801 | }; | ||||
802 | |||||
803 | /// A helper data structure to hold the operands of a vector of instructions. | ||||
804 | /// This supports a fixed vector length for all operand vectors. | ||||
805 | class VLOperands { | ||||
806 | /// For each operand we need (i) the value, and (ii) the opcode that it | ||||
807 | /// would be attached to if the expression was in a left-linearized form. | ||||
808 | /// This is required to avoid illegal operand reordering. | ||||
809 | /// For example: | ||||
810 | /// \verbatim | ||||
811 | /// 0 Op1 | ||||
812 | /// |/ | ||||
813 | /// Op1 Op2 Linearized + Op2 | ||||
814 | /// \ / ----------> |/ | ||||
815 | /// - - | ||||
816 | /// | ||||
817 | /// Op1 - Op2 (0 + Op1) - Op2 | ||||
818 | /// \endverbatim | ||||
819 | /// | ||||
820 | /// Value Op1 is attached to a '+' operation, and Op2 to a '-'. | ||||
821 | /// | ||||
822 | /// Another way to think of this is to track all the operations across the | ||||
823 | /// path from the operand all the way to the root of the tree and to | ||||
824 | /// calculate the operation that corresponds to this path. For example, the | ||||
825 | /// path from Op2 to the root crosses the RHS of the '-', therefore the | ||||
826 | /// corresponding operation is a '-' (which matches the one in the | ||||
827 | /// linearized tree, as shown above). | ||||
828 | /// | ||||
829 | /// For lack of a better term, we refer to this operation as Accumulated | ||||
830 | /// Path Operation (APO). | ||||
831 | struct OperandData { | ||||
832 | OperandData() = default; | ||||
833 | OperandData(Value *V, bool APO, bool IsUsed) | ||||
834 | : V(V), APO(APO), IsUsed(IsUsed) {} | ||||
835 | /// The operand value. | ||||
836 | Value *V = nullptr; | ||||
837 | /// TreeEntries only allow a single opcode, or an alternate sequence of | ||||
838 | /// them (e.g, +, -). Therefore, we can safely use a boolean value for the | ||||
839 | /// APO. It is set to 'true' if 'V' is attached to an inverse operation | ||||
840 | /// in the left-linearized form (e.g., Sub/Div), and 'false' otherwise | ||||
841 | /// (e.g., Add/Mul) | ||||
842 | bool APO = false; | ||||
843 | /// Helper data for the reordering function. | ||||
844 | bool IsUsed = false; | ||||
845 | }; | ||||
846 | |||||
847 | /// During operand reordering, we are trying to select the operand at lane | ||||
848 | /// that matches best with the operand at the neighboring lane. Our | ||||
849 | /// selection is based on the type of value we are looking for. For example, | ||||
850 | /// if the neighboring lane has a load, we need to look for a load that is | ||||
851 | /// accessing a consecutive address. These strategies are summarized in the | ||||
852 | /// 'ReorderingMode' enumerator. | ||||
853 | enum class ReorderingMode { | ||||
854 | Load, ///< Matching loads to consecutive memory addresses | ||||
855 | Opcode, ///< Matching instructions based on opcode (same or alternate) | ||||
856 | Constant, ///< Matching constants | ||||
857 | Splat, ///< Matching the same instruction multiple times (broadcast) | ||||
858 | Failed, ///< We failed to create a vectorizable group | ||||
859 | }; | ||||
860 | |||||
861 | using OperandDataVec = SmallVector<OperandData, 2>; | ||||
862 | |||||
863 | /// A vector of operand vectors. | ||||
864 | SmallVector<OperandDataVec, 4> OpsVec; | ||||
865 | |||||
866 | const DataLayout &DL; | ||||
867 | ScalarEvolution &SE; | ||||
868 | const BoUpSLP &R; | ||||
869 | |||||
870 | /// \returns the operand data at \p OpIdx and \p Lane. | ||||
871 | OperandData &getData(unsigned OpIdx, unsigned Lane) { | ||||
872 | return OpsVec[OpIdx][Lane]; | ||||
873 | } | ||||
874 | |||||
875 | /// \returns the operand data at \p OpIdx and \p Lane. Const version. | ||||
876 | const OperandData &getData(unsigned OpIdx, unsigned Lane) const { | ||||
877 | return OpsVec[OpIdx][Lane]; | ||||
878 | } | ||||
879 | |||||
880 | /// Clears the used flag for all entries. | ||||
881 | void clearUsed() { | ||||
882 | for (unsigned OpIdx = 0, NumOperands = getNumOperands(); | ||||
883 | OpIdx != NumOperands; ++OpIdx) | ||||
884 | for (unsigned Lane = 0, NumLanes = getNumLanes(); Lane != NumLanes; | ||||
885 | ++Lane) | ||||
886 | OpsVec[OpIdx][Lane].IsUsed = false; | ||||
887 | } | ||||
888 | |||||
889 | /// Swap the operand at \p OpIdx1 with that one at \p OpIdx2. | ||||
890 | void swap(unsigned OpIdx1, unsigned OpIdx2, unsigned Lane) { | ||||
891 | std::swap(OpsVec[OpIdx1][Lane], OpsVec[OpIdx2][Lane]); | ||||
892 | } | ||||
893 | |||||
894 | // The hard-coded scores listed here are not very important. When computing | ||||
895 | // the scores of matching one sub-tree with another, we are basically | ||||
896 | // counting the number of values that are matching. So even if all scores | ||||
897 | // are set to 1, we would still get a decent matching result. | ||||
898 | // However, sometimes we have to break ties. For example we may have to | ||||
899 | // choose between matching loads vs matching opcodes. This is what these | ||||
900 | // scores are helping us with: they provide the order of preference. | ||||
901 | |||||
902 | /// Loads from consecutive memory addresses, e.g. load(A[i]), load(A[i+1]). | ||||
903 | static const int ScoreConsecutiveLoads = 3; | ||||
904 | /// ExtractElementInst from same vector and consecutive indexes. | ||||
905 | static const int ScoreConsecutiveExtracts = 3; | ||||
906 | /// Constants. | ||||
907 | static const int ScoreConstants = 2; | ||||
908 | /// Instructions with the same opcode. | ||||
909 | static const int ScoreSameOpcode = 2; | ||||
910 | /// Instructions with alt opcodes (e.g, add + sub). | ||||
911 | static const int ScoreAltOpcodes = 1; | ||||
912 | /// Identical instructions (a.k.a. splat or broadcast). | ||||
913 | static const int ScoreSplat = 1; | ||||
914 | /// Matching with an undef is preferable to failing. | ||||
915 | static const int ScoreUndef = 1; | ||||
916 | /// Score for failing to find a decent match. | ||||
917 | static const int ScoreFail = 0; | ||||
918 | /// User exteranl to the vectorized code. | ||||
919 | static const int ExternalUseCost = 1; | ||||
920 | /// The user is internal but in a different lane. | ||||
921 | static const int UserInDiffLaneCost = ExternalUseCost; | ||||
922 | |||||
923 | /// \returns the score of placing \p V1 and \p V2 in consecutive lanes. | ||||
924 | static int getShallowScore(Value *V1, Value *V2, const DataLayout &DL, | ||||
925 | ScalarEvolution &SE) { | ||||
926 | auto *LI1 = dyn_cast<LoadInst>(V1); | ||||
927 | auto *LI2 = dyn_cast<LoadInst>(V2); | ||||
928 | if (LI1 && LI2) | ||||
929 | return isConsecutiveAccess(LI1, LI2, DL, SE) | ||||
930 | ? VLOperands::ScoreConsecutiveLoads | ||||
931 | : VLOperands::ScoreFail; | ||||
932 | |||||
933 | auto *C1 = dyn_cast<Constant>(V1); | ||||
934 | auto *C2 = dyn_cast<Constant>(V2); | ||||
935 | if (C1 && C2) | ||||
936 | return VLOperands::ScoreConstants; | ||||
937 | |||||
938 | // Extracts from consecutive indexes of the same vector better score as | ||||
939 | // the extracts could be optimized away. | ||||
940 | Value *EV; | ||||
941 | ConstantInt *Ex1Idx, *Ex2Idx; | ||||
942 | if (match(V1, m_ExtractElt(m_Value(EV), m_ConstantInt(Ex1Idx))) && | ||||
943 | match(V2, m_ExtractElt(m_Deferred(EV), m_ConstantInt(Ex2Idx))) && | ||||
944 | Ex1Idx->getZExtValue() + 1 == Ex2Idx->getZExtValue()) | ||||
945 | return VLOperands::ScoreConsecutiveExtracts; | ||||
946 | |||||
947 | auto *I1 = dyn_cast<Instruction>(V1); | ||||
948 | auto *I2 = dyn_cast<Instruction>(V2); | ||||
949 | if (I1 && I2) { | ||||
950 | if (I1 == I2) | ||||
951 | return VLOperands::ScoreSplat; | ||||
952 | InstructionsState S = getSameOpcode({I1, I2}); | ||||
953 | // Note: Only consider instructions with <= 2 operands to avoid | ||||
954 | // complexity explosion. | ||||
955 | if (S.getOpcode() && S.MainOp->getNumOperands() <= 2) | ||||
956 | return S.isAltShuffle() ? VLOperands::ScoreAltOpcodes | ||||
957 | : VLOperands::ScoreSameOpcode; | ||||
958 | } | ||||
959 | |||||
960 | if (isa<UndefValue>(V2)) | ||||
961 | return VLOperands::ScoreUndef; | ||||
962 | |||||
963 | return VLOperands::ScoreFail; | ||||
964 | } | ||||
965 | |||||
966 | /// Holds the values and their lane that are taking part in the look-ahead | ||||
967 | /// score calculation. This is used in the external uses cost calculation. | ||||
968 | SmallDenseMap<Value *, int> InLookAheadValues; | ||||
969 | |||||
970 | /// \Returns the additinal cost due to uses of \p LHS and \p RHS that are | ||||
971 | /// either external to the vectorized code, or require shuffling. | ||||
972 | int getExternalUsesCost(const std::pair<Value *, int> &LHS, | ||||
973 | const std::pair<Value *, int> &RHS) { | ||||
974 | int Cost = 0; | ||||
975 | std::array<std::pair<Value *, int>, 2> Values = {{LHS, RHS}}; | ||||
976 | for (int Idx = 0, IdxE = Values.size(); Idx != IdxE; ++Idx) { | ||||
977 | Value *V = Values[Idx].first; | ||||
978 | // Calculate the absolute lane, using the minimum relative lane of LHS | ||||
979 | // and RHS as base and Idx as the offset. | ||||
980 | int Ln = std::min(LHS.second, RHS.second) + Idx; | ||||
981 | assert(Ln >= 0 && "Bad lane calculation")((Ln >= 0 && "Bad lane calculation") ? static_cast <void> (0) : __assert_fail ("Ln >= 0 && \"Bad lane calculation\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 981, __PRETTY_FUNCTION__)); | ||||
982 | unsigned UsersBudget = LookAheadUsersBudget; | ||||
983 | for (User *U : V->users()) { | ||||
984 | if (const TreeEntry *UserTE = R.getTreeEntry(U)) { | ||||
985 | // The user is in the VectorizableTree. Check if we need to insert. | ||||
986 | auto It = llvm::find(UserTE->Scalars, U); | ||||
987 | assert(It != UserTE->Scalars.end() && "U is in UserTE")((It != UserTE->Scalars.end() && "U is in UserTE") ? static_cast<void> (0) : __assert_fail ("It != UserTE->Scalars.end() && \"U is in UserTE\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 987, __PRETTY_FUNCTION__)); | ||||
988 | int UserLn = std::distance(UserTE->Scalars.begin(), It); | ||||
989 | assert(UserLn >= 0 && "Bad lane")((UserLn >= 0 && "Bad lane") ? static_cast<void > (0) : __assert_fail ("UserLn >= 0 && \"Bad lane\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 989, __PRETTY_FUNCTION__)); | ||||
990 | if (UserLn != Ln) | ||||
991 | Cost += UserInDiffLaneCost; | ||||
992 | } else { | ||||
993 | // Check if the user is in the look-ahead code. | ||||
994 | auto It2 = InLookAheadValues.find(U); | ||||
995 | if (It2 != InLookAheadValues.end()) { | ||||
996 | // The user is in the look-ahead code. Check the lane. | ||||
997 | if (It2->second != Ln) | ||||
998 | Cost += UserInDiffLaneCost; | ||||
999 | } else { | ||||
1000 | // The user is neither in SLP tree nor in the look-ahead code. | ||||
1001 | Cost += ExternalUseCost; | ||||
1002 | } | ||||
1003 | } | ||||
1004 | // Limit the number of visited uses to cap compilation time. | ||||
1005 | if (--UsersBudget == 0) | ||||
1006 | break; | ||||
1007 | } | ||||
1008 | } | ||||
1009 | return Cost; | ||||
1010 | } | ||||
1011 | |||||
1012 | /// Go through the operands of \p LHS and \p RHS recursively until \p | ||||
1013 | /// MaxLevel, and return the cummulative score. For example: | ||||
1014 | /// \verbatim | ||||
1015 | /// A[0] B[0] A[1] B[1] C[0] D[0] B[1] A[1] | ||||
1016 | /// \ / \ / \ / \ / | ||||
1017 | /// + + + + | ||||
1018 | /// G1 G2 G3 G4 | ||||
1019 | /// \endverbatim | ||||
1020 | /// The getScoreAtLevelRec(G1, G2) function will try to match the nodes at | ||||
1021 | /// each level recursively, accumulating the score. It starts from matching | ||||
1022 | /// the additions at level 0, then moves on to the loads (level 1). The | ||||
1023 | /// score of G1 and G2 is higher than G1 and G3, because {A[0],A[1]} and | ||||
1024 | /// {B[0],B[1]} match with VLOperands::ScoreConsecutiveLoads, while | ||||
1025 | /// {A[0],C[0]} has a score of VLOperands::ScoreFail. | ||||
1026 | /// Please note that the order of the operands does not matter, as we | ||||
1027 | /// evaluate the score of all profitable combinations of operands. In | ||||
1028 | /// other words the score of G1 and G4 is the same as G1 and G2. This | ||||
1029 | /// heuristic is based on ideas described in: | ||||
1030 | /// Look-ahead SLP: Auto-vectorization in the presence of commutative | ||||
1031 | /// operations, CGO 2018 by Vasileios Porpodas, Rodrigo C. O. Rocha, | ||||
1032 | /// LuÃs F. W. Góes | ||||
1033 | int getScoreAtLevelRec(const std::pair<Value *, int> &LHS, | ||||
1034 | const std::pair<Value *, int> &RHS, int CurrLevel, | ||||
1035 | int MaxLevel) { | ||||
1036 | |||||
1037 | Value *V1 = LHS.first; | ||||
1038 | Value *V2 = RHS.first; | ||||
1039 | // Get the shallow score of V1 and V2. | ||||
1040 | int ShallowScoreAtThisLevel = | ||||
1041 | std::max((int)ScoreFail, getShallowScore(V1, V2, DL, SE) - | ||||
1042 | getExternalUsesCost(LHS, RHS)); | ||||
1043 | int Lane1 = LHS.second; | ||||
1044 | int Lane2 = RHS.second; | ||||
1045 | |||||
1046 | // If reached MaxLevel, | ||||
1047 | // or if V1 and V2 are not instructions, | ||||
1048 | // or if they are SPLAT, | ||||
1049 | // or if they are not consecutive, early return the current cost. | ||||
1050 | auto *I1 = dyn_cast<Instruction>(V1); | ||||
1051 | auto *I2 = dyn_cast<Instruction>(V2); | ||||
1052 | if (CurrLevel == MaxLevel || !(I1 && I2) || I1 == I2 || | ||||
1053 | ShallowScoreAtThisLevel == VLOperands::ScoreFail || | ||||
1054 | (isa<LoadInst>(I1) && isa<LoadInst>(I2) && ShallowScoreAtThisLevel)) | ||||
1055 | return ShallowScoreAtThisLevel; | ||||
1056 | assert(I1 && I2 && "Should have early exited.")((I1 && I2 && "Should have early exited.") ? static_cast <void> (0) : __assert_fail ("I1 && I2 && \"Should have early exited.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1056, __PRETTY_FUNCTION__)); | ||||
1057 | |||||
1058 | // Keep track of in-tree values for determining the external-use cost. | ||||
1059 | InLookAheadValues[V1] = Lane1; | ||||
1060 | InLookAheadValues[V2] = Lane2; | ||||
1061 | |||||
1062 | // Contains the I2 operand indexes that got matched with I1 operands. | ||||
1063 | SmallSet<unsigned, 4> Op2Used; | ||||
1064 | |||||
1065 | // Recursion towards the operands of I1 and I2. We are trying all possbile | ||||
1066 | // operand pairs, and keeping track of the best score. | ||||
1067 | for (unsigned OpIdx1 = 0, NumOperands1 = I1->getNumOperands(); | ||||
1068 | OpIdx1 != NumOperands1; ++OpIdx1) { | ||||
1069 | // Try to pair op1I with the best operand of I2. | ||||
1070 | int MaxTmpScore = 0; | ||||
1071 | unsigned MaxOpIdx2 = 0; | ||||
1072 | bool FoundBest = false; | ||||
1073 | // If I2 is commutative try all combinations. | ||||
1074 | unsigned FromIdx = isCommutative(I2) ? 0 : OpIdx1; | ||||
1075 | unsigned ToIdx = isCommutative(I2) | ||||
1076 | ? I2->getNumOperands() | ||||
1077 | : std::min(I2->getNumOperands(), OpIdx1 + 1); | ||||
1078 | assert(FromIdx <= ToIdx && "Bad index")((FromIdx <= ToIdx && "Bad index") ? static_cast< void> (0) : __assert_fail ("FromIdx <= ToIdx && \"Bad index\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1078, __PRETTY_FUNCTION__)); | ||||
1079 | for (unsigned OpIdx2 = FromIdx; OpIdx2 != ToIdx; ++OpIdx2) { | ||||
1080 | // Skip operands already paired with OpIdx1. | ||||
1081 | if (Op2Used.count(OpIdx2)) | ||||
1082 | continue; | ||||
1083 | // Recursively calculate the cost at each level | ||||
1084 | int TmpScore = getScoreAtLevelRec({I1->getOperand(OpIdx1), Lane1}, | ||||
1085 | {I2->getOperand(OpIdx2), Lane2}, | ||||
1086 | CurrLevel + 1, MaxLevel); | ||||
1087 | // Look for the best score. | ||||
1088 | if (TmpScore > VLOperands::ScoreFail && TmpScore > MaxTmpScore) { | ||||
1089 | MaxTmpScore = TmpScore; | ||||
1090 | MaxOpIdx2 = OpIdx2; | ||||
1091 | FoundBest = true; | ||||
1092 | } | ||||
1093 | } | ||||
1094 | if (FoundBest) { | ||||
1095 | // Pair {OpIdx1, MaxOpIdx2} was found to be best. Never revisit it. | ||||
1096 | Op2Used.insert(MaxOpIdx2); | ||||
1097 | ShallowScoreAtThisLevel += MaxTmpScore; | ||||
1098 | } | ||||
1099 | } | ||||
1100 | return ShallowScoreAtThisLevel; | ||||
1101 | } | ||||
1102 | |||||
1103 | /// \Returns the look-ahead score, which tells us how much the sub-trees | ||||
1104 | /// rooted at \p LHS and \p RHS match, the more they match the higher the | ||||
1105 | /// score. This helps break ties in an informed way when we cannot decide on | ||||
1106 | /// the order of the operands by just considering the immediate | ||||
1107 | /// predecessors. | ||||
1108 | int getLookAheadScore(const std::pair<Value *, int> &LHS, | ||||
1109 | const std::pair<Value *, int> &RHS) { | ||||
1110 | InLookAheadValues.clear(); | ||||
1111 | return getScoreAtLevelRec(LHS, RHS, 1, LookAheadMaxDepth); | ||||
1112 | } | ||||
1113 | |||||
1114 | // Search all operands in Ops[*][Lane] for the one that matches best | ||||
1115 | // Ops[OpIdx][LastLane] and return its opreand index. | ||||
1116 | // If no good match can be found, return None. | ||||
1117 | Optional<unsigned> | ||||
1118 | getBestOperand(unsigned OpIdx, int Lane, int LastLane, | ||||
1119 | ArrayRef<ReorderingMode> ReorderingModes) { | ||||
1120 | unsigned NumOperands = getNumOperands(); | ||||
1121 | |||||
1122 | // The operand of the previous lane at OpIdx. | ||||
1123 | Value *OpLastLane = getData(OpIdx, LastLane).V; | ||||
1124 | |||||
1125 | // Our strategy mode for OpIdx. | ||||
1126 | ReorderingMode RMode = ReorderingModes[OpIdx]; | ||||
1127 | |||||
1128 | // The linearized opcode of the operand at OpIdx, Lane. | ||||
1129 | bool OpIdxAPO = getData(OpIdx, Lane).APO; | ||||
1130 | |||||
1131 | // The best operand index and its score. | ||||
1132 | // Sometimes we have more than one option (e.g., Opcode and Undefs), so we | ||||
1133 | // are using the score to differentiate between the two. | ||||
1134 | struct BestOpData { | ||||
1135 | Optional<unsigned> Idx = None; | ||||
1136 | unsigned Score = 0; | ||||
1137 | } BestOp; | ||||
1138 | |||||
1139 | // Iterate through all unused operands and look for the best. | ||||
1140 | for (unsigned Idx = 0; Idx != NumOperands; ++Idx) { | ||||
1141 | // Get the operand at Idx and Lane. | ||||
1142 | OperandData &OpData = getData(Idx, Lane); | ||||
1143 | Value *Op = OpData.V; | ||||
1144 | bool OpAPO = OpData.APO; | ||||
1145 | |||||
1146 | // Skip already selected operands. | ||||
1147 | if (OpData.IsUsed) | ||||
1148 | continue; | ||||
1149 | |||||
1150 | // Skip if we are trying to move the operand to a position with a | ||||
1151 | // different opcode in the linearized tree form. This would break the | ||||
1152 | // semantics. | ||||
1153 | if (OpAPO != OpIdxAPO) | ||||
1154 | continue; | ||||
1155 | |||||
1156 | // Look for an operand that matches the current mode. | ||||
1157 | switch (RMode) { | ||||
1158 | case ReorderingMode::Load: | ||||
1159 | case ReorderingMode::Constant: | ||||
1160 | case ReorderingMode::Opcode: { | ||||
1161 | bool LeftToRight = Lane > LastLane; | ||||
1162 | Value *OpLeft = (LeftToRight) ? OpLastLane : Op; | ||||
1163 | Value *OpRight = (LeftToRight) ? Op : OpLastLane; | ||||
1164 | unsigned Score = | ||||
1165 | getLookAheadScore({OpLeft, LastLane}, {OpRight, Lane}); | ||||
1166 | if (Score > BestOp.Score) { | ||||
1167 | BestOp.Idx = Idx; | ||||
1168 | BestOp.Score = Score; | ||||
1169 | } | ||||
1170 | break; | ||||
1171 | } | ||||
1172 | case ReorderingMode::Splat: | ||||
1173 | if (Op == OpLastLane) | ||||
1174 | BestOp.Idx = Idx; | ||||
1175 | break; | ||||
1176 | case ReorderingMode::Failed: | ||||
1177 | return None; | ||||
1178 | } | ||||
1179 | } | ||||
1180 | |||||
1181 | if (BestOp.Idx) { | ||||
1182 | getData(BestOp.Idx.getValue(), Lane).IsUsed = true; | ||||
1183 | return BestOp.Idx; | ||||
1184 | } | ||||
1185 | // If we could not find a good match return None. | ||||
1186 | return None; | ||||
1187 | } | ||||
1188 | |||||
1189 | /// Helper for reorderOperandVecs. \Returns the lane that we should start | ||||
1190 | /// reordering from. This is the one which has the least number of operands | ||||
1191 | /// that can freely move about. | ||||
1192 | unsigned getBestLaneToStartReordering() const { | ||||
1193 | unsigned BestLane = 0; | ||||
1194 | unsigned Min = UINT_MAX(2147483647 *2U +1U); | ||||
1195 | for (unsigned Lane = 0, NumLanes = getNumLanes(); Lane != NumLanes; | ||||
1196 | ++Lane) { | ||||
1197 | unsigned NumFreeOps = getMaxNumOperandsThatCanBeReordered(Lane); | ||||
1198 | if (NumFreeOps < Min) { | ||||
1199 | Min = NumFreeOps; | ||||
1200 | BestLane = Lane; | ||||
1201 | } | ||||
1202 | } | ||||
1203 | return BestLane; | ||||
1204 | } | ||||
1205 | |||||
1206 | /// \Returns the maximum number of operands that are allowed to be reordered | ||||
1207 | /// for \p Lane. This is used as a heuristic for selecting the first lane to | ||||
1208 | /// start operand reordering. | ||||
1209 | unsigned getMaxNumOperandsThatCanBeReordered(unsigned Lane) const { | ||||
1210 | unsigned CntTrue = 0; | ||||
1211 | unsigned NumOperands = getNumOperands(); | ||||
1212 | // Operands with the same APO can be reordered. We therefore need to count | ||||
1213 | // how many of them we have for each APO, like this: Cnt[APO] = x. | ||||
1214 | // Since we only have two APOs, namely true and false, we can avoid using | ||||
1215 | // a map. Instead we can simply count the number of operands that | ||||
1216 | // correspond to one of them (in this case the 'true' APO), and calculate | ||||
1217 | // the other by subtracting it from the total number of operands. | ||||
1218 | for (unsigned OpIdx = 0; OpIdx != NumOperands; ++OpIdx) | ||||
1219 | if (getData(OpIdx, Lane).APO) | ||||
1220 | ++CntTrue; | ||||
1221 | unsigned CntFalse = NumOperands - CntTrue; | ||||
1222 | return std::max(CntTrue, CntFalse); | ||||
1223 | } | ||||
1224 | |||||
1225 | /// Go through the instructions in VL and append their operands. | ||||
1226 | void appendOperandsOfVL(ArrayRef<Value *> VL) { | ||||
1227 | assert(!VL.empty() && "Bad VL")((!VL.empty() && "Bad VL") ? static_cast<void> ( 0) : __assert_fail ("!VL.empty() && \"Bad VL\"", "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1227, __PRETTY_FUNCTION__)); | ||||
1228 | assert((empty() || VL.size() == getNumLanes()) &&(((empty() || VL.size() == getNumLanes()) && "Expected same number of lanes" ) ? static_cast<void> (0) : __assert_fail ("(empty() || VL.size() == getNumLanes()) && \"Expected same number of lanes\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1229, __PRETTY_FUNCTION__)) | ||||
1229 | "Expected same number of lanes")(((empty() || VL.size() == getNumLanes()) && "Expected same number of lanes" ) ? static_cast<void> (0) : __assert_fail ("(empty() || VL.size() == getNumLanes()) && \"Expected same number of lanes\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1229, __PRETTY_FUNCTION__)); | ||||
1230 | assert(isa<Instruction>(VL[0]) && "Expected instruction")((isa<Instruction>(VL[0]) && "Expected instruction" ) ? static_cast<void> (0) : __assert_fail ("isa<Instruction>(VL[0]) && \"Expected instruction\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1230, __PRETTY_FUNCTION__)); | ||||
1231 | unsigned NumOperands = cast<Instruction>(VL[0])->getNumOperands(); | ||||
1232 | OpsVec.resize(NumOperands); | ||||
1233 | unsigned NumLanes = VL.size(); | ||||
1234 | for (unsigned OpIdx = 0; OpIdx != NumOperands; ++OpIdx) { | ||||
1235 | OpsVec[OpIdx].resize(NumLanes); | ||||
1236 | for (unsigned Lane = 0; Lane != NumLanes; ++Lane) { | ||||
1237 | assert(isa<Instruction>(VL[Lane]) && "Expected instruction")((isa<Instruction>(VL[Lane]) && "Expected instruction" ) ? static_cast<void> (0) : __assert_fail ("isa<Instruction>(VL[Lane]) && \"Expected instruction\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1237, __PRETTY_FUNCTION__)); | ||||
1238 | // Our tree has just 3 nodes: the root and two operands. | ||||
1239 | // It is therefore trivial to get the APO. We only need to check the | ||||
1240 | // opcode of VL[Lane] and whether the operand at OpIdx is the LHS or | ||||
1241 | // RHS operand. The LHS operand of both add and sub is never attached | ||||
1242 | // to an inversese operation in the linearized form, therefore its APO | ||||
1243 | // is false. The RHS is true only if VL[Lane] is an inverse operation. | ||||
1244 | |||||
1245 | // Since operand reordering is performed on groups of commutative | ||||
1246 | // operations or alternating sequences (e.g., +, -), we can safely | ||||
1247 | // tell the inverse operations by checking commutativity. | ||||
1248 | bool IsInverseOperation = !isCommutative(cast<Instruction>(VL[Lane])); | ||||
1249 | bool APO = (OpIdx == 0) ? false : IsInverseOperation; | ||||
1250 | OpsVec[OpIdx][Lane] = {cast<Instruction>(VL[Lane])->getOperand(OpIdx), | ||||
1251 | APO, false}; | ||||
1252 | } | ||||
1253 | } | ||||
1254 | } | ||||
1255 | |||||
1256 | /// \returns the number of operands. | ||||
1257 | unsigned getNumOperands() const { return OpsVec.size(); } | ||||
1258 | |||||
1259 | /// \returns the number of lanes. | ||||
1260 | unsigned getNumLanes() const { return OpsVec[0].size(); } | ||||
1261 | |||||
1262 | /// \returns the operand value at \p OpIdx and \p Lane. | ||||
1263 | Value *getValue(unsigned OpIdx, unsigned Lane) const { | ||||
1264 | return getData(OpIdx, Lane).V; | ||||
1265 | } | ||||
1266 | |||||
1267 | /// \returns true if the data structure is empty. | ||||
1268 | bool empty() const { return OpsVec.empty(); } | ||||
1269 | |||||
1270 | /// Clears the data. | ||||
1271 | void clear() { OpsVec.clear(); } | ||||
1272 | |||||
1273 | /// \Returns true if there are enough operands identical to \p Op to fill | ||||
1274 | /// the whole vector. | ||||
1275 | /// Note: This modifies the 'IsUsed' flag, so a cleanUsed() must follow. | ||||
1276 | bool shouldBroadcast(Value *Op, unsigned OpIdx, unsigned Lane) { | ||||
1277 | bool OpAPO = getData(OpIdx, Lane).APO; | ||||
1278 | for (unsigned Ln = 0, Lns = getNumLanes(); Ln != Lns; ++Ln) { | ||||
1279 | if (Ln == Lane) | ||||
1280 | continue; | ||||
1281 | // This is set to true if we found a candidate for broadcast at Lane. | ||||
1282 | bool FoundCandidate = false; | ||||
1283 | for (unsigned OpI = 0, OpE = getNumOperands(); OpI != OpE; ++OpI) { | ||||
1284 | OperandData &Data = getData(OpI, Ln); | ||||
1285 | if (Data.APO != OpAPO || Data.IsUsed) | ||||
1286 | continue; | ||||
1287 | if (Data.V == Op) { | ||||
1288 | FoundCandidate = true; | ||||
1289 | Data.IsUsed = true; | ||||
1290 | break; | ||||
1291 | } | ||||
1292 | } | ||||
1293 | if (!FoundCandidate) | ||||
1294 | return false; | ||||
1295 | } | ||||
1296 | return true; | ||||
1297 | } | ||||
1298 | |||||
1299 | public: | ||||
1300 | /// Initialize with all the operands of the instruction vector \p RootVL. | ||||
1301 | VLOperands(ArrayRef<Value *> RootVL, const DataLayout &DL, | ||||
1302 | ScalarEvolution &SE, const BoUpSLP &R) | ||||
1303 | : DL(DL), SE(SE), R(R) { | ||||
1304 | // Append all the operands of RootVL. | ||||
1305 | appendOperandsOfVL(RootVL); | ||||
1306 | } | ||||
1307 | |||||
1308 | /// \Returns a value vector with the operands across all lanes for the | ||||
1309 | /// opearnd at \p OpIdx. | ||||
1310 | ValueList getVL(unsigned OpIdx) const { | ||||
1311 | ValueList OpVL(OpsVec[OpIdx].size()); | ||||
1312 | assert(OpsVec[OpIdx].size() == getNumLanes() &&((OpsVec[OpIdx].size() == getNumLanes() && "Expected same num of lanes across all operands" ) ? static_cast<void> (0) : __assert_fail ("OpsVec[OpIdx].size() == getNumLanes() && \"Expected same num of lanes across all operands\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1313, __PRETTY_FUNCTION__)) | ||||
1313 | "Expected same num of lanes across all operands")((OpsVec[OpIdx].size() == getNumLanes() && "Expected same num of lanes across all operands" ) ? static_cast<void> (0) : __assert_fail ("OpsVec[OpIdx].size() == getNumLanes() && \"Expected same num of lanes across all operands\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1313, __PRETTY_FUNCTION__)); | ||||
1314 | for (unsigned Lane = 0, Lanes = getNumLanes(); Lane != Lanes; ++Lane) | ||||
1315 | OpVL[Lane] = OpsVec[OpIdx][Lane].V; | ||||
1316 | return OpVL; | ||||
1317 | } | ||||
1318 | |||||
1319 | // Performs operand reordering for 2 or more operands. | ||||
1320 | // The original operands are in OrigOps[OpIdx][Lane]. | ||||
1321 | // The reordered operands are returned in 'SortedOps[OpIdx][Lane]'. | ||||
1322 | void reorder() { | ||||
1323 | unsigned NumOperands = getNumOperands(); | ||||
1324 | unsigned NumLanes = getNumLanes(); | ||||
1325 | // Each operand has its own mode. We are using this mode to help us select | ||||
1326 | // the instructions for each lane, so that they match best with the ones | ||||
1327 | // we have selected so far. | ||||
1328 | SmallVector<ReorderingMode, 2> ReorderingModes(NumOperands); | ||||
1329 | |||||
1330 | // This is a greedy single-pass algorithm. We are going over each lane | ||||
1331 | // once and deciding on the best order right away with no back-tracking. | ||||
1332 | // However, in order to increase its effectiveness, we start with the lane | ||||
1333 | // that has operands that can move the least. For example, given the | ||||
1334 | // following lanes: | ||||
1335 | // Lane 0 : A[0] = B[0] + C[0] // Visited 3rd | ||||
1336 | // Lane 1 : A[1] = C[1] - B[1] // Visited 1st | ||||
1337 | // Lane 2 : A[2] = B[2] + C[2] // Visited 2nd | ||||
1338 | // Lane 3 : A[3] = C[3] - B[3] // Visited 4th | ||||
1339 | // we will start at Lane 1, since the operands of the subtraction cannot | ||||
1340 | // be reordered. Then we will visit the rest of the lanes in a circular | ||||
1341 | // fashion. That is, Lanes 2, then Lane 0, and finally Lane 3. | ||||
1342 | |||||
1343 | // Find the first lane that we will start our search from. | ||||
1344 | unsigned FirstLane = getBestLaneToStartReordering(); | ||||
1345 | |||||
1346 | // Initialize the modes. | ||||
1347 | for (unsigned OpIdx = 0; OpIdx != NumOperands; ++OpIdx) { | ||||
1348 | Value *OpLane0 = getValue(OpIdx, FirstLane); | ||||
1349 | // Keep track if we have instructions with all the same opcode on one | ||||
1350 | // side. | ||||
1351 | if (isa<LoadInst>(OpLane0)) | ||||
1352 | ReorderingModes[OpIdx] = ReorderingMode::Load; | ||||
1353 | else if (isa<Instruction>(OpLane0)) { | ||||
1354 | // Check if OpLane0 should be broadcast. | ||||
1355 | if (shouldBroadcast(OpLane0, OpIdx, FirstLane)) | ||||
1356 | ReorderingModes[OpIdx] = ReorderingMode::Splat; | ||||
1357 | else | ||||
1358 | ReorderingModes[OpIdx] = ReorderingMode::Opcode; | ||||
1359 | } | ||||
1360 | else if (isa<Constant>(OpLane0)) | ||||
1361 | ReorderingModes[OpIdx] = ReorderingMode::Constant; | ||||
1362 | else if (isa<Argument>(OpLane0)) | ||||
1363 | // Our best hope is a Splat. It may save some cost in some cases. | ||||
1364 | ReorderingModes[OpIdx] = ReorderingMode::Splat; | ||||
1365 | else | ||||
1366 | // NOTE: This should be unreachable. | ||||
1367 | ReorderingModes[OpIdx] = ReorderingMode::Failed; | ||||
1368 | } | ||||
1369 | |||||
1370 | // If the initial strategy fails for any of the operand indexes, then we | ||||
1371 | // perform reordering again in a second pass. This helps avoid assigning | ||||
1372 | // high priority to the failed strategy, and should improve reordering for | ||||
1373 | // the non-failed operand indexes. | ||||
1374 | for (int Pass = 0; Pass != 2; ++Pass) { | ||||
1375 | // Skip the second pass if the first pass did not fail. | ||||
1376 | bool StrategyFailed = false; | ||||
1377 | // Mark all operand data as free to use. | ||||
1378 | clearUsed(); | ||||
1379 | // We keep the original operand order for the FirstLane, so reorder the | ||||
1380 | // rest of the lanes. We are visiting the nodes in a circular fashion, | ||||
1381 | // using FirstLane as the center point and increasing the radius | ||||
1382 | // distance. | ||||
1383 | for (unsigned Distance = 1; Distance != NumLanes; ++Distance) { | ||||
1384 | // Visit the lane on the right and then the lane on the left. | ||||
1385 | for (int Direction : {+1, -1}) { | ||||
1386 | int Lane = FirstLane + Direction * Distance; | ||||
1387 | if (Lane < 0 || Lane >= (int)NumLanes) | ||||
1388 | continue; | ||||
1389 | int LastLane = Lane - Direction; | ||||
1390 | assert(LastLane >= 0 && LastLane < (int)NumLanes &&((LastLane >= 0 && LastLane < (int)NumLanes && "Out of bounds") ? static_cast<void> (0) : __assert_fail ("LastLane >= 0 && LastLane < (int)NumLanes && \"Out of bounds\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1391, __PRETTY_FUNCTION__)) | ||||
1391 | "Out of bounds")((LastLane >= 0 && LastLane < (int)NumLanes && "Out of bounds") ? static_cast<void> (0) : __assert_fail ("LastLane >= 0 && LastLane < (int)NumLanes && \"Out of bounds\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1391, __PRETTY_FUNCTION__)); | ||||
1392 | // Look for a good match for each operand. | ||||
1393 | for (unsigned OpIdx = 0; OpIdx != NumOperands; ++OpIdx) { | ||||
1394 | // Search for the operand that matches SortedOps[OpIdx][Lane-1]. | ||||
1395 | Optional<unsigned> BestIdx = | ||||
1396 | getBestOperand(OpIdx, Lane, LastLane, ReorderingModes); | ||||
1397 | // By not selecting a value, we allow the operands that follow to | ||||
1398 | // select a better matching value. We will get a non-null value in | ||||
1399 | // the next run of getBestOperand(). | ||||
1400 | if (BestIdx) { | ||||
1401 | // Swap the current operand with the one returned by | ||||
1402 | // getBestOperand(). | ||||
1403 | swap(OpIdx, BestIdx.getValue(), Lane); | ||||
1404 | } else { | ||||
1405 | // We failed to find a best operand, set mode to 'Failed'. | ||||
1406 | ReorderingModes[OpIdx] = ReorderingMode::Failed; | ||||
1407 | // Enable the second pass. | ||||
1408 | StrategyFailed = true; | ||||
1409 | } | ||||
1410 | } | ||||
1411 | } | ||||
1412 | } | ||||
1413 | // Skip second pass if the strategy did not fail. | ||||
1414 | if (!StrategyFailed) | ||||
1415 | break; | ||||
1416 | } | ||||
1417 | } | ||||
1418 | |||||
1419 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||
1420 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) static StringRef getModeStr(ReorderingMode RMode) { | ||||
1421 | switch (RMode) { | ||||
1422 | case ReorderingMode::Load: | ||||
1423 | return "Load"; | ||||
1424 | case ReorderingMode::Opcode: | ||||
1425 | return "Opcode"; | ||||
1426 | case ReorderingMode::Constant: | ||||
1427 | return "Constant"; | ||||
1428 | case ReorderingMode::Splat: | ||||
1429 | return "Splat"; | ||||
1430 | case ReorderingMode::Failed: | ||||
1431 | return "Failed"; | ||||
1432 | } | ||||
1433 | llvm_unreachable("Unimplemented Reordering Type")::llvm::llvm_unreachable_internal("Unimplemented Reordering Type" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1433); | ||||
1434 | } | ||||
1435 | |||||
1436 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) static raw_ostream &printMode(ReorderingMode RMode, | ||||
1437 | raw_ostream &OS) { | ||||
1438 | return OS << getModeStr(RMode); | ||||
1439 | } | ||||
1440 | |||||
1441 | /// Debug print. | ||||
1442 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) static void dumpMode(ReorderingMode RMode) { | ||||
1443 | printMode(RMode, dbgs()); | ||||
1444 | } | ||||
1445 | |||||
1446 | friend raw_ostream &operator<<(raw_ostream &OS, ReorderingMode RMode) { | ||||
1447 | return printMode(RMode, OS); | ||||
1448 | } | ||||
1449 | |||||
1450 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) raw_ostream &print(raw_ostream &OS) const { | ||||
1451 | const unsigned Indent = 2; | ||||
1452 | unsigned Cnt = 0; | ||||
1453 | for (const OperandDataVec &OpDataVec : OpsVec) { | ||||
1454 | OS << "Operand " << Cnt++ << "\n"; | ||||
1455 | for (const OperandData &OpData : OpDataVec) { | ||||
1456 | OS.indent(Indent) << "{"; | ||||
1457 | if (Value *V = OpData.V) | ||||
1458 | OS << *V; | ||||
1459 | else | ||||
1460 | OS << "null"; | ||||
1461 | OS << ", APO:" << OpData.APO << "}\n"; | ||||
1462 | } | ||||
1463 | OS << "\n"; | ||||
1464 | } | ||||
1465 | return OS; | ||||
1466 | } | ||||
1467 | |||||
1468 | /// Debug print. | ||||
1469 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void dump() const { print(dbgs()); } | ||||
1470 | #endif | ||||
1471 | }; | ||||
1472 | |||||
1473 | /// Checks if the instruction is marked for deletion. | ||||
1474 | bool isDeleted(Instruction *I) const { return DeletedInstructions.count(I); } | ||||
1475 | |||||
1476 | /// Marks values operands for later deletion by replacing them with Undefs. | ||||
1477 | void eraseInstructions(ArrayRef<Value *> AV); | ||||
1478 | |||||
1479 | ~BoUpSLP(); | ||||
1480 | |||||
1481 | private: | ||||
1482 | /// Checks if all users of \p I are the part of the vectorization tree. | ||||
1483 | bool areAllUsersVectorized(Instruction *I) const; | ||||
1484 | |||||
1485 | /// \returns the cost of the vectorizable entry. | ||||
1486 | int getEntryCost(TreeEntry *E); | ||||
1487 | |||||
1488 | /// This is the recursive part of buildTree. | ||||
1489 | void buildTree_rec(ArrayRef<Value *> Roots, unsigned Depth, | ||||
1490 | const EdgeInfo &EI); | ||||
1491 | |||||
1492 | /// \returns true if the ExtractElement/ExtractValue instructions in \p VL can | ||||
1493 | /// be vectorized to use the original vector (or aggregate "bitcast" to a | ||||
1494 | /// vector) and sets \p CurrentOrder to the identity permutation; otherwise | ||||
1495 | /// returns false, setting \p CurrentOrder to either an empty vector or a | ||||
1496 | /// non-identity permutation that allows to reuse extract instructions. | ||||
1497 | bool canReuseExtract(ArrayRef<Value *> VL, Value *OpValue, | ||||
1498 | SmallVectorImpl<unsigned> &CurrentOrder) const; | ||||
1499 | |||||
1500 | /// Vectorize a single entry in the tree. | ||||
1501 | Value *vectorizeTree(TreeEntry *E); | ||||
1502 | |||||
1503 | /// Vectorize a single entry in the tree, starting in \p VL. | ||||
1504 | Value *vectorizeTree(ArrayRef<Value *> VL); | ||||
1505 | |||||
1506 | /// \returns the scalarization cost for this type. Scalarization in this | ||||
1507 | /// context means the creation of vectors from a group of scalars. | ||||
1508 | int getGatherCost(FixedVectorType *Ty, | ||||
1509 | const DenseSet<unsigned> &ShuffledIndices) const; | ||||
1510 | |||||
1511 | /// \returns the scalarization cost for this list of values. Assuming that | ||||
1512 | /// this subtree gets vectorized, we may need to extract the values from the | ||||
1513 | /// roots. This method calculates the cost of extracting the values. | ||||
1514 | int getGatherCost(ArrayRef<Value *> VL) const; | ||||
1515 | |||||
1516 | /// Set the Builder insert point to one after the last instruction in | ||||
1517 | /// the bundle | ||||
1518 | void setInsertPointAfterBundle(TreeEntry *E); | ||||
1519 | |||||
1520 | /// \returns a vector from a collection of scalars in \p VL. | ||||
1521 | Value *gather(ArrayRef<Value *> VL); | ||||
1522 | |||||
1523 | /// \returns whether the VectorizableTree is fully vectorizable and will | ||||
1524 | /// be beneficial even the tree height is tiny. | ||||
1525 | bool isFullyVectorizableTinyTree() const; | ||||
1526 | |||||
1527 | /// Reorder commutative or alt operands to get better probability of | ||||
1528 | /// generating vectorized code. | ||||
1529 | static void reorderInputsAccordingToOpcode(ArrayRef<Value *> VL, | ||||
1530 | SmallVectorImpl<Value *> &Left, | ||||
1531 | SmallVectorImpl<Value *> &Right, | ||||
1532 | const DataLayout &DL, | ||||
1533 | ScalarEvolution &SE, | ||||
1534 | const BoUpSLP &R); | ||||
1535 | struct TreeEntry { | ||||
1536 | using VecTreeTy = SmallVector<std::unique_ptr<TreeEntry>, 8>; | ||||
1537 | TreeEntry(VecTreeTy &Container) : Container(Container) {} | ||||
1538 | |||||
1539 | /// \returns true if the scalars in VL are equal to this entry. | ||||
1540 | bool isSame(ArrayRef<Value *> VL) const { | ||||
1541 | if (VL.size() == Scalars.size()) | ||||
1542 | return std::equal(VL.begin(), VL.end(), Scalars.begin()); | ||||
1543 | return VL.size() == ReuseShuffleIndices.size() && | ||||
1544 | std::equal( | ||||
1545 | VL.begin(), VL.end(), ReuseShuffleIndices.begin(), | ||||
1546 | [this](Value *V, int Idx) { return V == Scalars[Idx]; }); | ||||
1547 | } | ||||
1548 | |||||
1549 | /// A vector of scalars. | ||||
1550 | ValueList Scalars; | ||||
1551 | |||||
1552 | /// The Scalars are vectorized into this value. It is initialized to Null. | ||||
1553 | Value *VectorizedValue = nullptr; | ||||
1554 | |||||
1555 | /// Do we need to gather this sequence ? | ||||
1556 | enum EntryState { Vectorize, NeedToGather }; | ||||
1557 | EntryState State; | ||||
1558 | |||||
1559 | /// Does this sequence require some shuffling? | ||||
1560 | SmallVector<int, 4> ReuseShuffleIndices; | ||||
1561 | |||||
1562 | /// Does this entry require reordering? | ||||
1563 | SmallVector<unsigned, 4> ReorderIndices; | ||||
1564 | |||||
1565 | /// Points back to the VectorizableTree. | ||||
1566 | /// | ||||
1567 | /// Only used for Graphviz right now. Unfortunately GraphTrait::NodeRef has | ||||
1568 | /// to be a pointer and needs to be able to initialize the child iterator. | ||||
1569 | /// Thus we need a reference back to the container to translate the indices | ||||
1570 | /// to entries. | ||||
1571 | VecTreeTy &Container; | ||||
1572 | |||||
1573 | /// The TreeEntry index containing the user of this entry. We can actually | ||||
1574 | /// have multiple users so the data structure is not truly a tree. | ||||
1575 | SmallVector<EdgeInfo, 1> UserTreeIndices; | ||||
1576 | |||||
1577 | /// The index of this treeEntry in VectorizableTree. | ||||
1578 | int Idx = -1; | ||||
1579 | |||||
1580 | private: | ||||
1581 | /// The operands of each instruction in each lane Operands[op_index][lane]. | ||||
1582 | /// Note: This helps avoid the replication of the code that performs the | ||||
1583 | /// reordering of operands during buildTree_rec() and vectorizeTree(). | ||||
1584 | SmallVector<ValueList, 2> Operands; | ||||
1585 | |||||
1586 | /// The main/alternate instruction. | ||||
1587 | Instruction *MainOp = nullptr; | ||||
1588 | Instruction *AltOp = nullptr; | ||||
1589 | |||||
1590 | public: | ||||
1591 | /// Set this bundle's \p OpIdx'th operand to \p OpVL. | ||||
1592 | void setOperand(unsigned OpIdx, ArrayRef<Value *> OpVL) { | ||||
1593 | if (Operands.size() < OpIdx + 1) | ||||
1594 | Operands.resize(OpIdx + 1); | ||||
1595 | assert(Operands[OpIdx].size() == 0 && "Already resized?")((Operands[OpIdx].size() == 0 && "Already resized?") ? static_cast<void> (0) : __assert_fail ("Operands[OpIdx].size() == 0 && \"Already resized?\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1595, __PRETTY_FUNCTION__)); | ||||
1596 | Operands[OpIdx].resize(Scalars.size()); | ||||
1597 | for (unsigned Lane = 0, E = Scalars.size(); Lane != E; ++Lane) | ||||
1598 | Operands[OpIdx][Lane] = OpVL[Lane]; | ||||
1599 | } | ||||
1600 | |||||
1601 | /// Set the operands of this bundle in their original order. | ||||
1602 | void setOperandsInOrder() { | ||||
1603 | assert(Operands.empty() && "Already initialized?")((Operands.empty() && "Already initialized?") ? static_cast <void> (0) : __assert_fail ("Operands.empty() && \"Already initialized?\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1603, __PRETTY_FUNCTION__)); | ||||
1604 | auto *I0 = cast<Instruction>(Scalars[0]); | ||||
1605 | Operands.resize(I0->getNumOperands()); | ||||
1606 | unsigned NumLanes = Scalars.size(); | ||||
1607 | for (unsigned OpIdx = 0, NumOperands = I0->getNumOperands(); | ||||
1608 | OpIdx != NumOperands; ++OpIdx) { | ||||
1609 | Operands[OpIdx].resize(NumLanes); | ||||
1610 | for (unsigned Lane = 0; Lane != NumLanes; ++Lane) { | ||||
1611 | auto *I = cast<Instruction>(Scalars[Lane]); | ||||
1612 | assert(I->getNumOperands() == NumOperands &&((I->getNumOperands() == NumOperands && "Expected same number of operands" ) ? static_cast<void> (0) : __assert_fail ("I->getNumOperands() == NumOperands && \"Expected same number of operands\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1613, __PRETTY_FUNCTION__)) | ||||
1613 | "Expected same number of operands")((I->getNumOperands() == NumOperands && "Expected same number of operands" ) ? static_cast<void> (0) : __assert_fail ("I->getNumOperands() == NumOperands && \"Expected same number of operands\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1613, __PRETTY_FUNCTION__)); | ||||
1614 | Operands[OpIdx][Lane] = I->getOperand(OpIdx); | ||||
1615 | } | ||||
1616 | } | ||||
1617 | } | ||||
1618 | |||||
1619 | /// \returns the \p OpIdx operand of this TreeEntry. | ||||
1620 | ValueList &getOperand(unsigned OpIdx) { | ||||
1621 | assert(OpIdx < Operands.size() && "Off bounds")((OpIdx < Operands.size() && "Off bounds") ? static_cast <void> (0) : __assert_fail ("OpIdx < Operands.size() && \"Off bounds\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1621, __PRETTY_FUNCTION__)); | ||||
1622 | return Operands[OpIdx]; | ||||
1623 | } | ||||
1624 | |||||
1625 | /// \returns the number of operands. | ||||
1626 | unsigned getNumOperands() const { return Operands.size(); } | ||||
1627 | |||||
1628 | /// \return the single \p OpIdx operand. | ||||
1629 | Value *getSingleOperand(unsigned OpIdx) const { | ||||
1630 | assert(OpIdx < Operands.size() && "Off bounds")((OpIdx < Operands.size() && "Off bounds") ? static_cast <void> (0) : __assert_fail ("OpIdx < Operands.size() && \"Off bounds\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1630, __PRETTY_FUNCTION__)); | ||||
1631 | assert(!Operands[OpIdx].empty() && "No operand available")((!Operands[OpIdx].empty() && "No operand available") ? static_cast<void> (0) : __assert_fail ("!Operands[OpIdx].empty() && \"No operand available\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1631, __PRETTY_FUNCTION__)); | ||||
1632 | return Operands[OpIdx][0]; | ||||
1633 | } | ||||
1634 | |||||
1635 | /// Some of the instructions in the list have alternate opcodes. | ||||
1636 | bool isAltShuffle() const { | ||||
1637 | return getOpcode() != getAltOpcode(); | ||||
1638 | } | ||||
1639 | |||||
1640 | bool isOpcodeOrAlt(Instruction *I) const { | ||||
1641 | unsigned CheckedOpcode = I->getOpcode(); | ||||
1642 | return (getOpcode() == CheckedOpcode || | ||||
1643 | getAltOpcode() == CheckedOpcode); | ||||
1644 | } | ||||
1645 | |||||
1646 | /// Chooses the correct key for scheduling data. If \p Op has the same (or | ||||
1647 | /// alternate) opcode as \p OpValue, the key is \p Op. Otherwise the key is | ||||
1648 | /// \p OpValue. | ||||
1649 | Value *isOneOf(Value *Op) const { | ||||
1650 | auto *I = dyn_cast<Instruction>(Op); | ||||
1651 | if (I && isOpcodeOrAlt(I)) | ||||
1652 | return Op; | ||||
1653 | return MainOp; | ||||
1654 | } | ||||
1655 | |||||
1656 | void setOperations(const InstructionsState &S) { | ||||
1657 | MainOp = S.MainOp; | ||||
1658 | AltOp = S.AltOp; | ||||
1659 | } | ||||
1660 | |||||
1661 | Instruction *getMainOp() const { | ||||
1662 | return MainOp; | ||||
1663 | } | ||||
1664 | |||||
1665 | Instruction *getAltOp() const { | ||||
1666 | return AltOp; | ||||
1667 | } | ||||
1668 | |||||
1669 | /// The main/alternate opcodes for the list of instructions. | ||||
1670 | unsigned getOpcode() const { | ||||
1671 | return MainOp ? MainOp->getOpcode() : 0; | ||||
1672 | } | ||||
1673 | |||||
1674 | unsigned getAltOpcode() const { | ||||
1675 | return AltOp ? AltOp->getOpcode() : 0; | ||||
1676 | } | ||||
1677 | |||||
1678 | /// Update operations state of this entry if reorder occurred. | ||||
1679 | bool updateStateIfReorder() { | ||||
1680 | if (ReorderIndices.empty()) | ||||
1681 | return false; | ||||
1682 | InstructionsState S = getSameOpcode(Scalars, ReorderIndices.front()); | ||||
1683 | setOperations(S); | ||||
1684 | return true; | ||||
1685 | } | ||||
1686 | |||||
1687 | #ifndef NDEBUG | ||||
1688 | /// Debug printer. | ||||
1689 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void dump() const { | ||||
1690 | dbgs() << Idx << ".\n"; | ||||
1691 | for (unsigned OpI = 0, OpE = Operands.size(); OpI != OpE; ++OpI) { | ||||
1692 | dbgs() << "Operand " << OpI << ":\n"; | ||||
1693 | for (const Value *V : Operands[OpI]) | ||||
1694 | dbgs().indent(2) << *V << "\n"; | ||||
1695 | } | ||||
1696 | dbgs() << "Scalars: \n"; | ||||
1697 | for (Value *V : Scalars) | ||||
1698 | dbgs().indent(2) << *V << "\n"; | ||||
1699 | dbgs() << "State: "; | ||||
1700 | switch (State) { | ||||
1701 | case Vectorize: | ||||
1702 | dbgs() << "Vectorize\n"; | ||||
1703 | break; | ||||
1704 | case NeedToGather: | ||||
1705 | dbgs() << "NeedToGather\n"; | ||||
1706 | break; | ||||
1707 | } | ||||
1708 | dbgs() << "MainOp: "; | ||||
1709 | if (MainOp) | ||||
1710 | dbgs() << *MainOp << "\n"; | ||||
1711 | else | ||||
1712 | dbgs() << "NULL\n"; | ||||
1713 | dbgs() << "AltOp: "; | ||||
1714 | if (AltOp) | ||||
1715 | dbgs() << *AltOp << "\n"; | ||||
1716 | else | ||||
1717 | dbgs() << "NULL\n"; | ||||
1718 | dbgs() << "VectorizedValue: "; | ||||
1719 | if (VectorizedValue) | ||||
1720 | dbgs() << *VectorizedValue << "\n"; | ||||
1721 | else | ||||
1722 | dbgs() << "NULL\n"; | ||||
1723 | dbgs() << "ReuseShuffleIndices: "; | ||||
1724 | if (ReuseShuffleIndices.empty()) | ||||
1725 | dbgs() << "Emtpy"; | ||||
1726 | else | ||||
1727 | for (unsigned ReuseIdx : ReuseShuffleIndices) | ||||
1728 | dbgs() << ReuseIdx << ", "; | ||||
1729 | dbgs() << "\n"; | ||||
1730 | dbgs() << "ReorderIndices: "; | ||||
1731 | for (unsigned ReorderIdx : ReorderIndices) | ||||
1732 | dbgs() << ReorderIdx << ", "; | ||||
1733 | dbgs() << "\n"; | ||||
1734 | dbgs() << "UserTreeIndices: "; | ||||
1735 | for (const auto &EInfo : UserTreeIndices) | ||||
1736 | dbgs() << EInfo << ", "; | ||||
1737 | dbgs() << "\n"; | ||||
1738 | } | ||||
1739 | #endif | ||||
1740 | }; | ||||
1741 | |||||
1742 | /// Create a new VectorizableTree entry. | ||||
1743 | TreeEntry *newTreeEntry(ArrayRef<Value *> VL, Optional<ScheduleData *> Bundle, | ||||
1744 | const InstructionsState &S, | ||||
1745 | const EdgeInfo &UserTreeIdx, | ||||
1746 | ArrayRef<unsigned> ReuseShuffleIndices = None, | ||||
1747 | ArrayRef<unsigned> ReorderIndices = None) { | ||||
1748 | bool Vectorized = (bool)Bundle; | ||||
1749 | VectorizableTree.push_back(std::make_unique<TreeEntry>(VectorizableTree)); | ||||
1750 | TreeEntry *Last = VectorizableTree.back().get(); | ||||
1751 | Last->Idx = VectorizableTree.size() - 1; | ||||
1752 | Last->Scalars.insert(Last->Scalars.begin(), VL.begin(), VL.end()); | ||||
1753 | Last->State = Vectorized ? TreeEntry::Vectorize : TreeEntry::NeedToGather; | ||||
1754 | Last->ReuseShuffleIndices.append(ReuseShuffleIndices.begin(), | ||||
1755 | ReuseShuffleIndices.end()); | ||||
1756 | Last->ReorderIndices.append(ReorderIndices.begin(), ReorderIndices.end()); | ||||
1757 | Last->setOperations(S); | ||||
1758 | if (Vectorized) { | ||||
1759 | for (Value *V : VL) { | ||||
1760 | assert(!getTreeEntry(V) && "Scalar already in tree!")((!getTreeEntry(V) && "Scalar already in tree!") ? static_cast <void> (0) : __assert_fail ("!getTreeEntry(V) && \"Scalar already in tree!\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1760, __PRETTY_FUNCTION__)); | ||||
1761 | ScalarToTreeEntry[V] = Last; | ||||
1762 | } | ||||
1763 | // Update the scheduler bundle to point to this TreeEntry. | ||||
1764 | unsigned Lane = 0; | ||||
1765 | for (ScheduleData *BundleMember = Bundle.getValue(); BundleMember; | ||||
1766 | BundleMember = BundleMember->NextInBundle) { | ||||
1767 | BundleMember->TE = Last; | ||||
1768 | BundleMember->Lane = Lane; | ||||
1769 | ++Lane; | ||||
1770 | } | ||||
1771 | assert((!Bundle.getValue() || Lane == VL.size()) &&(((!Bundle.getValue() || Lane == VL.size()) && "Bundle and VL out of sync" ) ? static_cast<void> (0) : __assert_fail ("(!Bundle.getValue() || Lane == VL.size()) && \"Bundle and VL out of sync\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1772, __PRETTY_FUNCTION__)) | ||||
1772 | "Bundle and VL out of sync")(((!Bundle.getValue() || Lane == VL.size()) && "Bundle and VL out of sync" ) ? static_cast<void> (0) : __assert_fail ("(!Bundle.getValue() || Lane == VL.size()) && \"Bundle and VL out of sync\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1772, __PRETTY_FUNCTION__)); | ||||
1773 | } else { | ||||
1774 | MustGather.insert(VL.begin(), VL.end()); | ||||
1775 | } | ||||
1776 | |||||
1777 | if (UserTreeIdx.UserTE) | ||||
1778 | Last->UserTreeIndices.push_back(UserTreeIdx); | ||||
1779 | |||||
1780 | return Last; | ||||
1781 | } | ||||
1782 | |||||
1783 | /// -- Vectorization State -- | ||||
1784 | /// Holds all of the tree entries. | ||||
1785 | TreeEntry::VecTreeTy VectorizableTree; | ||||
1786 | |||||
1787 | #ifndef NDEBUG | ||||
1788 | /// Debug printer. | ||||
1789 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void dumpVectorizableTree() const { | ||||
1790 | for (unsigned Id = 0, IdE = VectorizableTree.size(); Id != IdE; ++Id) { | ||||
1791 | VectorizableTree[Id]->dump(); | ||||
1792 | dbgs() << "\n"; | ||||
1793 | } | ||||
1794 | } | ||||
1795 | #endif | ||||
1796 | |||||
1797 | TreeEntry *getTreeEntry(Value *V) { | ||||
1798 | auto I = ScalarToTreeEntry.find(V); | ||||
1799 | if (I != ScalarToTreeEntry.end()) | ||||
1800 | return I->second; | ||||
1801 | return nullptr; | ||||
1802 | } | ||||
1803 | |||||
1804 | const TreeEntry *getTreeEntry(Value *V) const { | ||||
1805 | auto I = ScalarToTreeEntry.find(V); | ||||
1806 | if (I != ScalarToTreeEntry.end()) | ||||
1807 | return I->second; | ||||
1808 | return nullptr; | ||||
1809 | } | ||||
1810 | |||||
1811 | /// Maps a specific scalar to its tree entry. | ||||
1812 | SmallDenseMap<Value*, TreeEntry *> ScalarToTreeEntry; | ||||
1813 | |||||
1814 | /// Maps a value to the proposed vectorizable size. | ||||
1815 | SmallDenseMap<Value *, unsigned> InstrElementSize; | ||||
1816 | |||||
1817 | /// A list of scalars that we found that we need to keep as scalars. | ||||
1818 | ValueSet MustGather; | ||||
1819 | |||||
1820 | /// This POD struct describes one external user in the vectorized tree. | ||||
1821 | struct ExternalUser { | ||||
1822 | ExternalUser(Value *S, llvm::User *U, int L) | ||||
1823 | : Scalar(S), User(U), Lane(L) {} | ||||
1824 | |||||
1825 | // Which scalar in our function. | ||||
1826 | Value *Scalar; | ||||
1827 | |||||
1828 | // Which user that uses the scalar. | ||||
1829 | llvm::User *User; | ||||
1830 | |||||
1831 | // Which lane does the scalar belong to. | ||||
1832 | int Lane; | ||||
1833 | }; | ||||
1834 | using UserList = SmallVector<ExternalUser, 16>; | ||||
1835 | |||||
1836 | /// Checks if two instructions may access the same memory. | ||||
1837 | /// | ||||
1838 | /// \p Loc1 is the location of \p Inst1. It is passed explicitly because it | ||||
1839 | /// is invariant in the calling loop. | ||||
1840 | bool isAliased(const MemoryLocation &Loc1, Instruction *Inst1, | ||||
1841 | Instruction *Inst2) { | ||||
1842 | // First check if the result is already in the cache. | ||||
1843 | AliasCacheKey key = std::make_pair(Inst1, Inst2); | ||||
1844 | Optional<bool> &result = AliasCache[key]; | ||||
1845 | if (result.hasValue()) { | ||||
1846 | return result.getValue(); | ||||
1847 | } | ||||
1848 | MemoryLocation Loc2 = getLocation(Inst2, AA); | ||||
1849 | bool aliased = true; | ||||
1850 | if (Loc1.Ptr && Loc2.Ptr && isSimple(Inst1) && isSimple(Inst2)) { | ||||
1851 | // Do the alias check. | ||||
1852 | aliased = AA->alias(Loc1, Loc2); | ||||
1853 | } | ||||
1854 | // Store the result in the cache. | ||||
1855 | result = aliased; | ||||
1856 | return aliased; | ||||
1857 | } | ||||
1858 | |||||
1859 | using AliasCacheKey = std::pair<Instruction *, Instruction *>; | ||||
1860 | |||||
1861 | /// Cache for alias results. | ||||
1862 | /// TODO: consider moving this to the AliasAnalysis itself. | ||||
1863 | DenseMap<AliasCacheKey, Optional<bool>> AliasCache; | ||||
1864 | |||||
1865 | /// Removes an instruction from its block and eventually deletes it. | ||||
1866 | /// It's like Instruction::eraseFromParent() except that the actual deletion | ||||
1867 | /// is delayed until BoUpSLP is destructed. | ||||
1868 | /// This is required to ensure that there are no incorrect collisions in the | ||||
1869 | /// AliasCache, which can happen if a new instruction is allocated at the | ||||
1870 | /// same address as a previously deleted instruction. | ||||
1871 | void eraseInstruction(Instruction *I, bool ReplaceOpsWithUndef = false) { | ||||
1872 | auto It = DeletedInstructions.try_emplace(I, ReplaceOpsWithUndef).first; | ||||
1873 | It->getSecond() = It->getSecond() && ReplaceOpsWithUndef; | ||||
1874 | } | ||||
1875 | |||||
1876 | /// Temporary store for deleted instructions. Instructions will be deleted | ||||
1877 | /// eventually when the BoUpSLP is destructed. | ||||
1878 | DenseMap<Instruction *, bool> DeletedInstructions; | ||||
1879 | |||||
1880 | /// A list of values that need to extracted out of the tree. | ||||
1881 | /// This list holds pairs of (Internal Scalar : External User). External User | ||||
1882 | /// can be nullptr, it means that this Internal Scalar will be used later, | ||||
1883 | /// after vectorization. | ||||
1884 | UserList ExternalUses; | ||||
1885 | |||||
1886 | /// Values used only by @llvm.assume calls. | ||||
1887 | SmallPtrSet<const Value *, 32> EphValues; | ||||
1888 | |||||
1889 | /// Holds all of the instructions that we gathered. | ||||
1890 | SetVector<Instruction *> GatherSeq; | ||||
1891 | |||||
1892 | /// A list of blocks that we are going to CSE. | ||||
1893 | SetVector<BasicBlock *> CSEBlocks; | ||||
1894 | |||||
1895 | /// Contains all scheduling relevant data for an instruction. | ||||
1896 | /// A ScheduleData either represents a single instruction or a member of an | ||||
1897 | /// instruction bundle (= a group of instructions which is combined into a | ||||
1898 | /// vector instruction). | ||||
1899 | struct ScheduleData { | ||||
1900 | // The initial value for the dependency counters. It means that the | ||||
1901 | // dependencies are not calculated yet. | ||||
1902 | enum { InvalidDeps = -1 }; | ||||
1903 | |||||
1904 | ScheduleData() = default; | ||||
1905 | |||||
1906 | void init(int BlockSchedulingRegionID, Value *OpVal) { | ||||
1907 | FirstInBundle = this; | ||||
1908 | NextInBundle = nullptr; | ||||
1909 | NextLoadStore = nullptr; | ||||
1910 | IsScheduled = false; | ||||
1911 | SchedulingRegionID = BlockSchedulingRegionID; | ||||
1912 | UnscheduledDepsInBundle = UnscheduledDeps; | ||||
1913 | clearDependencies(); | ||||
1914 | OpValue = OpVal; | ||||
1915 | TE = nullptr; | ||||
1916 | Lane = -1; | ||||
1917 | } | ||||
1918 | |||||
1919 | /// Returns true if the dependency information has been calculated. | ||||
1920 | bool hasValidDependencies() const { return Dependencies != InvalidDeps; } | ||||
1921 | |||||
1922 | /// Returns true for single instructions and for bundle representatives | ||||
1923 | /// (= the head of a bundle). | ||||
1924 | bool isSchedulingEntity() const { return FirstInBundle == this; } | ||||
1925 | |||||
1926 | /// Returns true if it represents an instruction bundle and not only a | ||||
1927 | /// single instruction. | ||||
1928 | bool isPartOfBundle() const { | ||||
1929 | return NextInBundle != nullptr || FirstInBundle != this; | ||||
1930 | } | ||||
1931 | |||||
1932 | /// Returns true if it is ready for scheduling, i.e. it has no more | ||||
1933 | /// unscheduled depending instructions/bundles. | ||||
1934 | bool isReady() const { | ||||
1935 | assert(isSchedulingEntity() &&((isSchedulingEntity() && "can't consider non-scheduling entity for ready list" ) ? static_cast<void> (0) : __assert_fail ("isSchedulingEntity() && \"can't consider non-scheduling entity for ready list\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1936, __PRETTY_FUNCTION__)) | ||||
1936 | "can't consider non-scheduling entity for ready list")((isSchedulingEntity() && "can't consider non-scheduling entity for ready list" ) ? static_cast<void> (0) : __assert_fail ("isSchedulingEntity() && \"can't consider non-scheduling entity for ready list\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1936, __PRETTY_FUNCTION__)); | ||||
1937 | return UnscheduledDepsInBundle == 0 && !IsScheduled; | ||||
1938 | } | ||||
1939 | |||||
1940 | /// Modifies the number of unscheduled dependencies, also updating it for | ||||
1941 | /// the whole bundle. | ||||
1942 | int incrementUnscheduledDeps(int Incr) { | ||||
1943 | UnscheduledDeps += Incr; | ||||
1944 | return FirstInBundle->UnscheduledDepsInBundle += Incr; | ||||
1945 | } | ||||
1946 | |||||
1947 | /// Sets the number of unscheduled dependencies to the number of | ||||
1948 | /// dependencies. | ||||
1949 | void resetUnscheduledDeps() { | ||||
1950 | incrementUnscheduledDeps(Dependencies - UnscheduledDeps); | ||||
1951 | } | ||||
1952 | |||||
1953 | /// Clears all dependency information. | ||||
1954 | void clearDependencies() { | ||||
1955 | Dependencies = InvalidDeps; | ||||
1956 | resetUnscheduledDeps(); | ||||
1957 | MemoryDependencies.clear(); | ||||
1958 | } | ||||
1959 | |||||
1960 | void dump(raw_ostream &os) const { | ||||
1961 | if (!isSchedulingEntity()) { | ||||
1962 | os << "/ " << *Inst; | ||||
1963 | } else if (NextInBundle) { | ||||
1964 | os << '[' << *Inst; | ||||
1965 | ScheduleData *SD = NextInBundle; | ||||
1966 | while (SD) { | ||||
1967 | os << ';' << *SD->Inst; | ||||
1968 | SD = SD->NextInBundle; | ||||
1969 | } | ||||
1970 | os << ']'; | ||||
1971 | } else { | ||||
1972 | os << *Inst; | ||||
1973 | } | ||||
1974 | } | ||||
1975 | |||||
1976 | Instruction *Inst = nullptr; | ||||
1977 | |||||
1978 | /// Points to the head in an instruction bundle (and always to this for | ||||
1979 | /// single instructions). | ||||
1980 | ScheduleData *FirstInBundle = nullptr; | ||||
1981 | |||||
1982 | /// Single linked list of all instructions in a bundle. Null if it is a | ||||
1983 | /// single instruction. | ||||
1984 | ScheduleData *NextInBundle = nullptr; | ||||
1985 | |||||
1986 | /// Single linked list of all memory instructions (e.g. load, store, call) | ||||
1987 | /// in the block - until the end of the scheduling region. | ||||
1988 | ScheduleData *NextLoadStore = nullptr; | ||||
1989 | |||||
1990 | /// The dependent memory instructions. | ||||
1991 | /// This list is derived on demand in calculateDependencies(). | ||||
1992 | SmallVector<ScheduleData *, 4> MemoryDependencies; | ||||
1993 | |||||
1994 | /// This ScheduleData is in the current scheduling region if this matches | ||||
1995 | /// the current SchedulingRegionID of BlockScheduling. | ||||
1996 | int SchedulingRegionID = 0; | ||||
1997 | |||||
1998 | /// Used for getting a "good" final ordering of instructions. | ||||
1999 | int SchedulingPriority = 0; | ||||
2000 | |||||
2001 | /// The number of dependencies. Constitutes of the number of users of the | ||||
2002 | /// instruction plus the number of dependent memory instructions (if any). | ||||
2003 | /// This value is calculated on demand. | ||||
2004 | /// If InvalidDeps, the number of dependencies is not calculated yet. | ||||
2005 | int Dependencies = InvalidDeps; | ||||
2006 | |||||
2007 | /// The number of dependencies minus the number of dependencies of scheduled | ||||
2008 | /// instructions. As soon as this is zero, the instruction/bundle gets ready | ||||
2009 | /// for scheduling. | ||||
2010 | /// Note that this is negative as long as Dependencies is not calculated. | ||||
2011 | int UnscheduledDeps = InvalidDeps; | ||||
2012 | |||||
2013 | /// The sum of UnscheduledDeps in a bundle. Equals to UnscheduledDeps for | ||||
2014 | /// single instructions. | ||||
2015 | int UnscheduledDepsInBundle = InvalidDeps; | ||||
2016 | |||||
2017 | /// True if this instruction is scheduled (or considered as scheduled in the | ||||
2018 | /// dry-run). | ||||
2019 | bool IsScheduled = false; | ||||
2020 | |||||
2021 | /// Opcode of the current instruction in the schedule data. | ||||
2022 | Value *OpValue = nullptr; | ||||
2023 | |||||
2024 | /// The TreeEntry that this instruction corresponds to. | ||||
2025 | TreeEntry *TE = nullptr; | ||||
2026 | |||||
2027 | /// The lane of this node in the TreeEntry. | ||||
2028 | int Lane = -1; | ||||
2029 | }; | ||||
2030 | |||||
2031 | #ifndef NDEBUG | ||||
2032 | friend inline raw_ostream &operator<<(raw_ostream &os, | ||||
2033 | const BoUpSLP::ScheduleData &SD) { | ||||
2034 | SD.dump(os); | ||||
2035 | return os; | ||||
2036 | } | ||||
2037 | #endif | ||||
2038 | |||||
2039 | friend struct GraphTraits<BoUpSLP *>; | ||||
2040 | friend struct DOTGraphTraits<BoUpSLP *>; | ||||
2041 | |||||
2042 | /// Contains all scheduling data for a basic block. | ||||
2043 | struct BlockScheduling { | ||||
2044 | BlockScheduling(BasicBlock *BB) | ||||
2045 | : BB(BB), ChunkSize(BB->size()), ChunkPos(ChunkSize) {} | ||||
2046 | |||||
2047 | void clear() { | ||||
2048 | ReadyInsts.clear(); | ||||
2049 | ScheduleStart = nullptr; | ||||
2050 | ScheduleEnd = nullptr; | ||||
2051 | FirstLoadStoreInRegion = nullptr; | ||||
2052 | LastLoadStoreInRegion = nullptr; | ||||
2053 | |||||
2054 | // Reduce the maximum schedule region size by the size of the | ||||
2055 | // previous scheduling run. | ||||
2056 | ScheduleRegionSizeLimit -= ScheduleRegionSize; | ||||
2057 | if (ScheduleRegionSizeLimit < MinScheduleRegionSize) | ||||
2058 | ScheduleRegionSizeLimit = MinScheduleRegionSize; | ||||
2059 | ScheduleRegionSize = 0; | ||||
2060 | |||||
2061 | // Make a new scheduling region, i.e. all existing ScheduleData is not | ||||
2062 | // in the new region yet. | ||||
2063 | ++SchedulingRegionID; | ||||
2064 | } | ||||
2065 | |||||
2066 | ScheduleData *getScheduleData(Value *V) { | ||||
2067 | ScheduleData *SD = ScheduleDataMap[V]; | ||||
2068 | if (SD && SD->SchedulingRegionID == SchedulingRegionID) | ||||
2069 | return SD; | ||||
2070 | return nullptr; | ||||
2071 | } | ||||
2072 | |||||
2073 | ScheduleData *getScheduleData(Value *V, Value *Key) { | ||||
2074 | if (V == Key) | ||||
2075 | return getScheduleData(V); | ||||
2076 | auto I = ExtraScheduleDataMap.find(V); | ||||
2077 | if (I != ExtraScheduleDataMap.end()) { | ||||
2078 | ScheduleData *SD = I->second[Key]; | ||||
2079 | if (SD && SD->SchedulingRegionID == SchedulingRegionID) | ||||
2080 | return SD; | ||||
2081 | } | ||||
2082 | return nullptr; | ||||
2083 | } | ||||
2084 | |||||
2085 | bool isInSchedulingRegion(ScheduleData *SD) const { | ||||
2086 | return SD->SchedulingRegionID == SchedulingRegionID; | ||||
2087 | } | ||||
2088 | |||||
2089 | /// Marks an instruction as scheduled and puts all dependent ready | ||||
2090 | /// instructions into the ready-list. | ||||
2091 | template <typename ReadyListType> | ||||
2092 | void schedule(ScheduleData *SD, ReadyListType &ReadyList) { | ||||
2093 | SD->IsScheduled = true; | ||||
2094 | LLVM_DEBUG(dbgs() << "SLP: schedule " << *SD << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: schedule " << *SD << "\n"; } } while (false); | ||||
2095 | |||||
2096 | ScheduleData *BundleMember = SD; | ||||
2097 | while (BundleMember) { | ||||
2098 | if (BundleMember->Inst != BundleMember->OpValue) { | ||||
2099 | BundleMember = BundleMember->NextInBundle; | ||||
2100 | continue; | ||||
2101 | } | ||||
2102 | // Handle the def-use chain dependencies. | ||||
2103 | |||||
2104 | // Decrement the unscheduled counter and insert to ready list if ready. | ||||
2105 | auto &&DecrUnsched = [this, &ReadyList](Instruction *I) { | ||||
2106 | doForAllOpcodes(I, [&ReadyList](ScheduleData *OpDef) { | ||||
2107 | if (OpDef && OpDef->hasValidDependencies() && | ||||
2108 | OpDef->incrementUnscheduledDeps(-1) == 0) { | ||||
2109 | // There are no more unscheduled dependencies after | ||||
2110 | // decrementing, so we can put the dependent instruction | ||||
2111 | // into the ready list. | ||||
2112 | ScheduleData *DepBundle = OpDef->FirstInBundle; | ||||
2113 | assert(!DepBundle->IsScheduled &&((!DepBundle->IsScheduled && "already scheduled bundle gets ready" ) ? static_cast<void> (0) : __assert_fail ("!DepBundle->IsScheduled && \"already scheduled bundle gets ready\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2114, __PRETTY_FUNCTION__)) | ||||
2114 | "already scheduled bundle gets ready")((!DepBundle->IsScheduled && "already scheduled bundle gets ready" ) ? static_cast<void> (0) : __assert_fail ("!DepBundle->IsScheduled && \"already scheduled bundle gets ready\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2114, __PRETTY_FUNCTION__)); | ||||
2115 | ReadyList.insert(DepBundle); | ||||
2116 | LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: gets ready (def): " << *DepBundle << "\n"; } } while (false) | ||||
2117 | << "SLP: gets ready (def): " << *DepBundle << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: gets ready (def): " << *DepBundle << "\n"; } } while (false); | ||||
2118 | } | ||||
2119 | }); | ||||
2120 | }; | ||||
2121 | |||||
2122 | // If BundleMember is a vector bundle, its operands may have been | ||||
2123 | // reordered duiring buildTree(). We therefore need to get its operands | ||||
2124 | // through the TreeEntry. | ||||
2125 | if (TreeEntry *TE = BundleMember->TE) { | ||||
2126 | int Lane = BundleMember->Lane; | ||||
2127 | assert(Lane >= 0 && "Lane not set")((Lane >= 0 && "Lane not set") ? static_cast<void > (0) : __assert_fail ("Lane >= 0 && \"Lane not set\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2127, __PRETTY_FUNCTION__)); | ||||
2128 | |||||
2129 | // Since vectorization tree is being built recursively this assertion | ||||
2130 | // ensures that the tree entry has all operands set before reaching | ||||
2131 | // this code. Couple of exceptions known at the moment are extracts | ||||
2132 | // where their second (immediate) operand is not added. Since | ||||
2133 | // immediates do not affect scheduler behavior this is considered | ||||
2134 | // okay. | ||||
2135 | auto *In = TE->getMainOp(); | ||||
2136 | assert(In &&((In && (isa<ExtractValueInst>(In) || isa<ExtractElementInst >(In) || In->getNumOperands() == TE->getNumOperands( )) && "Missed TreeEntry operands?") ? static_cast< void> (0) : __assert_fail ("In && (isa<ExtractValueInst>(In) || isa<ExtractElementInst>(In) || In->getNumOperands() == TE->getNumOperands()) && \"Missed TreeEntry operands?\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2139, __PRETTY_FUNCTION__)) | ||||
2137 | (isa<ExtractValueInst>(In) || isa<ExtractElementInst>(In) ||((In && (isa<ExtractValueInst>(In) || isa<ExtractElementInst >(In) || In->getNumOperands() == TE->getNumOperands( )) && "Missed TreeEntry operands?") ? static_cast< void> (0) : __assert_fail ("In && (isa<ExtractValueInst>(In) || isa<ExtractElementInst>(In) || In->getNumOperands() == TE->getNumOperands()) && \"Missed TreeEntry operands?\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2139, __PRETTY_FUNCTION__)) | ||||
2138 | In->getNumOperands() == TE->getNumOperands()) &&((In && (isa<ExtractValueInst>(In) || isa<ExtractElementInst >(In) || In->getNumOperands() == TE->getNumOperands( )) && "Missed TreeEntry operands?") ? static_cast< void> (0) : __assert_fail ("In && (isa<ExtractValueInst>(In) || isa<ExtractElementInst>(In) || In->getNumOperands() == TE->getNumOperands()) && \"Missed TreeEntry operands?\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2139, __PRETTY_FUNCTION__)) | ||||
2139 | "Missed TreeEntry operands?")((In && (isa<ExtractValueInst>(In) || isa<ExtractElementInst >(In) || In->getNumOperands() == TE->getNumOperands( )) && "Missed TreeEntry operands?") ? static_cast< void> (0) : __assert_fail ("In && (isa<ExtractValueInst>(In) || isa<ExtractElementInst>(In) || In->getNumOperands() == TE->getNumOperands()) && \"Missed TreeEntry operands?\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2139, __PRETTY_FUNCTION__)); | ||||
2140 | (void)In; // fake use to avoid build failure when assertions disabled | ||||
2141 | |||||
2142 | for (unsigned OpIdx = 0, NumOperands = TE->getNumOperands(); | ||||
2143 | OpIdx != NumOperands; ++OpIdx) | ||||
2144 | if (auto *I = dyn_cast<Instruction>(TE->getOperand(OpIdx)[Lane])) | ||||
2145 | DecrUnsched(I); | ||||
2146 | } else { | ||||
2147 | // If BundleMember is a stand-alone instruction, no operand reordering | ||||
2148 | // has taken place, so we directly access its operands. | ||||
2149 | for (Use &U : BundleMember->Inst->operands()) | ||||
2150 | if (auto *I = dyn_cast<Instruction>(U.get())) | ||||
2151 | DecrUnsched(I); | ||||
2152 | } | ||||
2153 | // Handle the memory dependencies. | ||||
2154 | for (ScheduleData *MemoryDepSD : BundleMember->MemoryDependencies) { | ||||
2155 | if (MemoryDepSD->incrementUnscheduledDeps(-1) == 0) { | ||||
2156 | // There are no more unscheduled dependencies after decrementing, | ||||
2157 | // so we can put the dependent instruction into the ready list. | ||||
2158 | ScheduleData *DepBundle = MemoryDepSD->FirstInBundle; | ||||
2159 | assert(!DepBundle->IsScheduled &&((!DepBundle->IsScheduled && "already scheduled bundle gets ready" ) ? static_cast<void> (0) : __assert_fail ("!DepBundle->IsScheduled && \"already scheduled bundle gets ready\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2160, __PRETTY_FUNCTION__)) | ||||
2160 | "already scheduled bundle gets ready")((!DepBundle->IsScheduled && "already scheduled bundle gets ready" ) ? static_cast<void> (0) : __assert_fail ("!DepBundle->IsScheduled && \"already scheduled bundle gets ready\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2160, __PRETTY_FUNCTION__)); | ||||
2161 | ReadyList.insert(DepBundle); | ||||
2162 | LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: gets ready (mem): " << *DepBundle << "\n"; } } while (false) | ||||
2163 | << "SLP: gets ready (mem): " << *DepBundle << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: gets ready (mem): " << *DepBundle << "\n"; } } while (false); | ||||
2164 | } | ||||
2165 | } | ||||
2166 | BundleMember = BundleMember->NextInBundle; | ||||
2167 | } | ||||
2168 | } | ||||
2169 | |||||
2170 | void doForAllOpcodes(Value *V, | ||||
2171 | function_ref<void(ScheduleData *SD)> Action) { | ||||
2172 | if (ScheduleData *SD = getScheduleData(V)) | ||||
2173 | Action(SD); | ||||
2174 | auto I = ExtraScheduleDataMap.find(V); | ||||
2175 | if (I != ExtraScheduleDataMap.end()) | ||||
2176 | for (auto &P : I->second) | ||||
2177 | if (P.second->SchedulingRegionID == SchedulingRegionID) | ||||
2178 | Action(P.second); | ||||
2179 | } | ||||
2180 | |||||
2181 | /// Put all instructions into the ReadyList which are ready for scheduling. | ||||
2182 | template <typename ReadyListType> | ||||
2183 | void initialFillReadyList(ReadyListType &ReadyList) { | ||||
2184 | for (auto *I = ScheduleStart; I != ScheduleEnd; I = I->getNextNode()) { | ||||
2185 | doForAllOpcodes(I, [&](ScheduleData *SD) { | ||||
2186 | if (SD->isSchedulingEntity() && SD->isReady()) { | ||||
2187 | ReadyList.insert(SD); | ||||
2188 | LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: initially in ready list: " << *I << "\n"; } } while (false) | ||||
2189 | << "SLP: initially in ready list: " << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: initially in ready list: " << *I << "\n"; } } while (false); | ||||
2190 | } | ||||
2191 | }); | ||||
2192 | } | ||||
2193 | } | ||||
2194 | |||||
2195 | /// Checks if a bundle of instructions can be scheduled, i.e. has no | ||||
2196 | /// cyclic dependencies. This is only a dry-run, no instructions are | ||||
2197 | /// actually moved at this stage. | ||||
2198 | /// \returns the scheduling bundle. The returned Optional value is non-None | ||||
2199 | /// if \p VL is allowed to be scheduled. | ||||
2200 | Optional<ScheduleData *> | ||||
2201 | tryScheduleBundle(ArrayRef<Value *> VL, BoUpSLP *SLP, | ||||
2202 | const InstructionsState &S); | ||||
2203 | |||||
2204 | /// Un-bundles a group of instructions. | ||||
2205 | void cancelScheduling(ArrayRef<Value *> VL, Value *OpValue); | ||||
2206 | |||||
2207 | /// Allocates schedule data chunk. | ||||
2208 | ScheduleData *allocateScheduleDataChunks(); | ||||
2209 | |||||
2210 | /// Extends the scheduling region so that V is inside the region. | ||||
2211 | /// \returns true if the region size is within the limit. | ||||
2212 | bool extendSchedulingRegion(Value *V, const InstructionsState &S); | ||||
2213 | |||||
2214 | /// Initialize the ScheduleData structures for new instructions in the | ||||
2215 | /// scheduling region. | ||||
2216 | void initScheduleData(Instruction *FromI, Instruction *ToI, | ||||
2217 | ScheduleData *PrevLoadStore, | ||||
2218 | ScheduleData *NextLoadStore); | ||||
2219 | |||||
2220 | /// Updates the dependency information of a bundle and of all instructions/ | ||||
2221 | /// bundles which depend on the original bundle. | ||||
2222 | void calculateDependencies(ScheduleData *SD, bool InsertInReadyList, | ||||
2223 | BoUpSLP *SLP); | ||||
2224 | |||||
2225 | /// Sets all instruction in the scheduling region to un-scheduled. | ||||
2226 | void resetSchedule(); | ||||
2227 | |||||
2228 | BasicBlock *BB; | ||||
2229 | |||||
2230 | /// Simple memory allocation for ScheduleData. | ||||
2231 | std::vector<std::unique_ptr<ScheduleData[]>> ScheduleDataChunks; | ||||
2232 | |||||
2233 | /// The size of a ScheduleData array in ScheduleDataChunks. | ||||
2234 | int ChunkSize; | ||||
2235 | |||||
2236 | /// The allocator position in the current chunk, which is the last entry | ||||
2237 | /// of ScheduleDataChunks. | ||||
2238 | int ChunkPos; | ||||
2239 | |||||
2240 | /// Attaches ScheduleData to Instruction. | ||||
2241 | /// Note that the mapping survives during all vectorization iterations, i.e. | ||||
2242 | /// ScheduleData structures are recycled. | ||||
2243 | DenseMap<Value *, ScheduleData *> ScheduleDataMap; | ||||
2244 | |||||
2245 | /// Attaches ScheduleData to Instruction with the leading key. | ||||
2246 | DenseMap<Value *, SmallDenseMap<Value *, ScheduleData *>> | ||||
2247 | ExtraScheduleDataMap; | ||||
2248 | |||||
2249 | struct ReadyList : SmallVector<ScheduleData *, 8> { | ||||
2250 | void insert(ScheduleData *SD) { push_back(SD); } | ||||
2251 | }; | ||||
2252 | |||||
2253 | /// The ready-list for scheduling (only used for the dry-run). | ||||
2254 | ReadyList ReadyInsts; | ||||
2255 | |||||
2256 | /// The first instruction of the scheduling region. | ||||
2257 | Instruction *ScheduleStart = nullptr; | ||||
2258 | |||||
2259 | /// The first instruction _after_ the scheduling region. | ||||
2260 | Instruction *ScheduleEnd = nullptr; | ||||
2261 | |||||
2262 | /// The first memory accessing instruction in the scheduling region | ||||
2263 | /// (can be null). | ||||
2264 | ScheduleData *FirstLoadStoreInRegion = nullptr; | ||||
2265 | |||||
2266 | /// The last memory accessing instruction in the scheduling region | ||||
2267 | /// (can be null). | ||||
2268 | ScheduleData *LastLoadStoreInRegion = nullptr; | ||||
2269 | |||||
2270 | /// The current size of the scheduling region. | ||||
2271 | int ScheduleRegionSize = 0; | ||||
2272 | |||||
2273 | /// The maximum size allowed for the scheduling region. | ||||
2274 | int ScheduleRegionSizeLimit = ScheduleRegionSizeBudget; | ||||
2275 | |||||
2276 | /// The ID of the scheduling region. For a new vectorization iteration this | ||||
2277 | /// is incremented which "removes" all ScheduleData from the region. | ||||
2278 | // Make sure that the initial SchedulingRegionID is greater than the | ||||
2279 | // initial SchedulingRegionID in ScheduleData (which is 0). | ||||
2280 | int SchedulingRegionID = 1; | ||||
2281 | }; | ||||
2282 | |||||
2283 | /// Attaches the BlockScheduling structures to basic blocks. | ||||
2284 | MapVector<BasicBlock *, std::unique_ptr<BlockScheduling>> BlocksSchedules; | ||||
2285 | |||||
2286 | /// Performs the "real" scheduling. Done before vectorization is actually | ||||
2287 | /// performed in a basic block. | ||||
2288 | void scheduleBlock(BlockScheduling *BS); | ||||
2289 | |||||
2290 | /// List of users to ignore during scheduling and that don't need extracting. | ||||
2291 | ArrayRef<Value *> UserIgnoreList; | ||||
2292 | |||||
2293 | /// A DenseMapInfo implementation for holding DenseMaps and DenseSets of | ||||
2294 | /// sorted SmallVectors of unsigned. | ||||
2295 | struct OrdersTypeDenseMapInfo { | ||||
2296 | static OrdersType getEmptyKey() { | ||||
2297 | OrdersType V; | ||||
2298 | V.push_back(~1U); | ||||
2299 | return V; | ||||
2300 | } | ||||
2301 | |||||
2302 | static OrdersType getTombstoneKey() { | ||||
2303 | OrdersType V; | ||||
2304 | V.push_back(~2U); | ||||
2305 | return V; | ||||
2306 | } | ||||
2307 | |||||
2308 | static unsigned getHashValue(const OrdersType &V) { | ||||
2309 | return static_cast<unsigned>(hash_combine_range(V.begin(), V.end())); | ||||
2310 | } | ||||
2311 | |||||
2312 | static bool isEqual(const OrdersType &LHS, const OrdersType &RHS) { | ||||
2313 | return LHS == RHS; | ||||
2314 | } | ||||
2315 | }; | ||||
2316 | |||||
2317 | /// Contains orders of operations along with the number of bundles that have | ||||
2318 | /// operations in this order. It stores only those orders that require | ||||
2319 | /// reordering, if reordering is not required it is counted using \a | ||||
2320 | /// NumOpsWantToKeepOriginalOrder. | ||||
2321 | DenseMap<OrdersType, unsigned, OrdersTypeDenseMapInfo> NumOpsWantToKeepOrder; | ||||
2322 | /// Number of bundles that do not require reordering. | ||||
2323 | unsigned NumOpsWantToKeepOriginalOrder = 0; | ||||
2324 | |||||
2325 | // Analysis and block reference. | ||||
2326 | Function *F; | ||||
2327 | ScalarEvolution *SE; | ||||
2328 | TargetTransformInfo *TTI; | ||||
2329 | TargetLibraryInfo *TLI; | ||||
2330 | AAResults *AA; | ||||
2331 | LoopInfo *LI; | ||||
2332 | DominatorTree *DT; | ||||
2333 | AssumptionCache *AC; | ||||
2334 | DemandedBits *DB; | ||||
2335 | const DataLayout *DL; | ||||
2336 | OptimizationRemarkEmitter *ORE; | ||||
2337 | |||||
2338 | unsigned MaxVecRegSize; // This is set by TTI or overridden by cl::opt. | ||||
2339 | unsigned MinVecRegSize; // Set by cl::opt (default: 128). | ||||
2340 | |||||
2341 | /// Instruction builder to construct the vectorized tree. | ||||
2342 | IRBuilder<> Builder; | ||||
2343 | |||||
2344 | /// A map of scalar integer values to the smallest bit width with which they | ||||
2345 | /// can legally be represented. The values map to (width, signed) pairs, | ||||
2346 | /// where "width" indicates the minimum bit width and "signed" is True if the | ||||
2347 | /// value must be signed-extended, rather than zero-extended, back to its | ||||
2348 | /// original width. | ||||
2349 | MapVector<Value *, std::pair<uint64_t, bool>> MinBWs; | ||||
2350 | }; | ||||
2351 | |||||
2352 | } // end namespace slpvectorizer | ||||
2353 | |||||
2354 | template <> struct GraphTraits<BoUpSLP *> { | ||||
2355 | using TreeEntry = BoUpSLP::TreeEntry; | ||||
2356 | |||||
2357 | /// NodeRef has to be a pointer per the GraphWriter. | ||||
2358 | using NodeRef = TreeEntry *; | ||||
2359 | |||||
2360 | using ContainerTy = BoUpSLP::TreeEntry::VecTreeTy; | ||||
2361 | |||||
2362 | /// Add the VectorizableTree to the index iterator to be able to return | ||||
2363 | /// TreeEntry pointers. | ||||
2364 | struct ChildIteratorType | ||||
2365 | : public iterator_adaptor_base< | ||||
2366 | ChildIteratorType, SmallVector<BoUpSLP::EdgeInfo, 1>::iterator> { | ||||
2367 | ContainerTy &VectorizableTree; | ||||
2368 | |||||
2369 | ChildIteratorType(SmallVector<BoUpSLP::EdgeInfo, 1>::iterator W, | ||||
2370 | ContainerTy &VT) | ||||
2371 | : ChildIteratorType::iterator_adaptor_base(W), VectorizableTree(VT) {} | ||||
2372 | |||||
2373 | NodeRef operator*() { return I->UserTE; } | ||||
2374 | }; | ||||
2375 | |||||
2376 | static NodeRef getEntryNode(BoUpSLP &R) { | ||||
2377 | return R.VectorizableTree[0].get(); | ||||
2378 | } | ||||
2379 | |||||
2380 | static ChildIteratorType child_begin(NodeRef N) { | ||||
2381 | return {N->UserTreeIndices.begin(), N->Container}; | ||||
2382 | } | ||||
2383 | |||||
2384 | static ChildIteratorType child_end(NodeRef N) { | ||||
2385 | return {N->UserTreeIndices.end(), N->Container}; | ||||
2386 | } | ||||
2387 | |||||
2388 | /// For the node iterator we just need to turn the TreeEntry iterator into a | ||||
2389 | /// TreeEntry* iterator so that it dereferences to NodeRef. | ||||
2390 | class nodes_iterator { | ||||
2391 | using ItTy = ContainerTy::iterator; | ||||
2392 | ItTy It; | ||||
2393 | |||||
2394 | public: | ||||
2395 | nodes_iterator(const ItTy &It2) : It(It2) {} | ||||
2396 | NodeRef operator*() { return It->get(); } | ||||
2397 | nodes_iterator operator++() { | ||||
2398 | ++It; | ||||
2399 | return *this; | ||||
2400 | } | ||||
2401 | bool operator!=(const nodes_iterator &N2) const { return N2.It != It; } | ||||
2402 | }; | ||||
2403 | |||||
2404 | static nodes_iterator nodes_begin(BoUpSLP *R) { | ||||
2405 | return nodes_iterator(R->VectorizableTree.begin()); | ||||
2406 | } | ||||
2407 | |||||
2408 | static nodes_iterator nodes_end(BoUpSLP *R) { | ||||
2409 | return nodes_iterator(R->VectorizableTree.end()); | ||||
2410 | } | ||||
2411 | |||||
2412 | static unsigned size(BoUpSLP *R) { return R->VectorizableTree.size(); } | ||||
2413 | }; | ||||
2414 | |||||
2415 | template <> struct DOTGraphTraits<BoUpSLP *> : public DefaultDOTGraphTraits { | ||||
2416 | using TreeEntry = BoUpSLP::TreeEntry; | ||||
2417 | |||||
2418 | DOTGraphTraits(bool isSimple = false) : DefaultDOTGraphTraits(isSimple) {} | ||||
2419 | |||||
2420 | std::string getNodeLabel(const TreeEntry *Entry, const BoUpSLP *R) { | ||||
2421 | std::string Str; | ||||
2422 | raw_string_ostream OS(Str); | ||||
2423 | if (isSplat(Entry->Scalars)) { | ||||
2424 | OS << "<splat> " << *Entry->Scalars[0]; | ||||
2425 | return Str; | ||||
2426 | } | ||||
2427 | for (auto V : Entry->Scalars) { | ||||
2428 | OS << *V; | ||||
2429 | if (std::any_of( | ||||
2430 | R->ExternalUses.begin(), R->ExternalUses.end(), | ||||
2431 | [&](const BoUpSLP::ExternalUser &EU) { return EU.Scalar == V; })) | ||||
2432 | OS << " <extract>"; | ||||
2433 | OS << "\n"; | ||||
2434 | } | ||||
2435 | return Str; | ||||
2436 | } | ||||
2437 | |||||
2438 | static std::string getNodeAttributes(const TreeEntry *Entry, | ||||
2439 | const BoUpSLP *) { | ||||
2440 | if (Entry->State == TreeEntry::NeedToGather) | ||||
2441 | return "color=red"; | ||||
2442 | return ""; | ||||
2443 | } | ||||
2444 | }; | ||||
2445 | |||||
2446 | } // end namespace llvm | ||||
2447 | |||||
2448 | BoUpSLP::~BoUpSLP() { | ||||
2449 | for (const auto &Pair : DeletedInstructions) { | ||||
2450 | // Replace operands of ignored instructions with Undefs in case if they were | ||||
2451 | // marked for deletion. | ||||
2452 | if (Pair.getSecond()) { | ||||
2453 | Value *Undef = UndefValue::get(Pair.getFirst()->getType()); | ||||
2454 | Pair.getFirst()->replaceAllUsesWith(Undef); | ||||
2455 | } | ||||
2456 | Pair.getFirst()->dropAllReferences(); | ||||
2457 | } | ||||
2458 | for (const auto &Pair : DeletedInstructions) { | ||||
2459 | assert(Pair.getFirst()->use_empty() &&((Pair.getFirst()->use_empty() && "trying to erase instruction with users." ) ? static_cast<void> (0) : __assert_fail ("Pair.getFirst()->use_empty() && \"trying to erase instruction with users.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2460, __PRETTY_FUNCTION__)) | ||||
2460 | "trying to erase instruction with users.")((Pair.getFirst()->use_empty() && "trying to erase instruction with users." ) ? static_cast<void> (0) : __assert_fail ("Pair.getFirst()->use_empty() && \"trying to erase instruction with users.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2460, __PRETTY_FUNCTION__)); | ||||
2461 | Pair.getFirst()->eraseFromParent(); | ||||
2462 | } | ||||
2463 | assert(!verifyFunction(*F, &dbgs()))((!verifyFunction(*F, &dbgs())) ? static_cast<void> (0) : __assert_fail ("!verifyFunction(*F, &dbgs())", "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2463, __PRETTY_FUNCTION__)); | ||||
2464 | } | ||||
2465 | |||||
2466 | void BoUpSLP::eraseInstructions(ArrayRef<Value *> AV) { | ||||
2467 | for (auto *V : AV) { | ||||
2468 | if (auto *I = dyn_cast<Instruction>(V)) | ||||
2469 | eraseInstruction(I, /*ReplaceOpsWithUndef=*/true); | ||||
2470 | }; | ||||
2471 | } | ||||
2472 | |||||
2473 | void BoUpSLP::buildTree(ArrayRef<Value *> Roots, | ||||
2474 | ArrayRef<Value *> UserIgnoreLst) { | ||||
2475 | ExtraValueToDebugLocsMap ExternallyUsedValues; | ||||
2476 | buildTree(Roots, ExternallyUsedValues, UserIgnoreLst); | ||||
2477 | } | ||||
2478 | |||||
2479 | void BoUpSLP::buildTree(ArrayRef<Value *> Roots, | ||||
2480 | ExtraValueToDebugLocsMap &ExternallyUsedValues, | ||||
2481 | ArrayRef<Value *> UserIgnoreLst) { | ||||
2482 | deleteTree(); | ||||
2483 | UserIgnoreList = UserIgnoreLst; | ||||
2484 | if (!allSameType(Roots)) | ||||
2485 | return; | ||||
2486 | buildTree_rec(Roots, 0, EdgeInfo()); | ||||
2487 | |||||
2488 | // Collect the values that we need to extract from the tree. | ||||
2489 | for (auto &TEPtr : VectorizableTree) { | ||||
2490 | TreeEntry *Entry = TEPtr.get(); | ||||
2491 | |||||
2492 | // No need to handle users of gathered values. | ||||
2493 | if (Entry->State == TreeEntry::NeedToGather) | ||||
2494 | continue; | ||||
2495 | |||||
2496 | // For each lane: | ||||
2497 | for (int Lane = 0, LE = Entry->Scalars.size(); Lane != LE; ++Lane) { | ||||
2498 | Value *Scalar = Entry->Scalars[Lane]; | ||||
2499 | int FoundLane = Lane; | ||||
2500 | if (!Entry->ReuseShuffleIndices.empty()) { | ||||
2501 | FoundLane = | ||||
2502 | std::distance(Entry->ReuseShuffleIndices.begin(), | ||||
2503 | llvm::find(Entry->ReuseShuffleIndices, FoundLane)); | ||||
2504 | } | ||||
2505 | |||||
2506 | // Check if the scalar is externally used as an extra arg. | ||||
2507 | auto ExtI = ExternallyUsedValues.find(Scalar); | ||||
2508 | if (ExtI != ExternallyUsedValues.end()) { | ||||
2509 | 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) | ||||
2510 | << 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); | ||||
2511 | ExternalUses.emplace_back(Scalar, nullptr, FoundLane); | ||||
2512 | } | ||||
2513 | for (User *U : Scalar->users()) { | ||||
2514 | LLVM_DEBUG(dbgs() << "SLP: Checking user:" << *U << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Checking user:" << *U << ".\n"; } } while (false); | ||||
2515 | |||||
2516 | Instruction *UserInst = dyn_cast<Instruction>(U); | ||||
2517 | if (!UserInst) | ||||
2518 | continue; | ||||
2519 | |||||
2520 | // Skip in-tree scalars that become vectors | ||||
2521 | if (TreeEntry *UseEntry = getTreeEntry(U)) { | ||||
2522 | Value *UseScalar = UseEntry->Scalars[0]; | ||||
2523 | // Some in-tree scalars will remain as scalar in vectorized | ||||
2524 | // instructions. If that is the case, the one in Lane 0 will | ||||
2525 | // be used. | ||||
2526 | if (UseScalar != U || | ||||
2527 | !InTreeUserNeedToExtract(Scalar, UserInst, TLI)) { | ||||
2528 | 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) | ||||
2529 | << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: \tInternal user will be removed:" << *U << ".\n"; } } while (false); | ||||
2530 | assert(UseEntry->State != TreeEntry::NeedToGather && "Bad state")((UseEntry->State != TreeEntry::NeedToGather && "Bad state" ) ? static_cast<void> (0) : __assert_fail ("UseEntry->State != TreeEntry::NeedToGather && \"Bad state\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2530, __PRETTY_FUNCTION__)); | ||||
2531 | continue; | ||||
2532 | } | ||||
2533 | } | ||||
2534 | |||||
2535 | // Ignore users in the user ignore list. | ||||
2536 | if (is_contained(UserIgnoreList, UserInst)) | ||||
2537 | continue; | ||||
2538 | |||||
2539 | 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) | ||||
2540 | << 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); | ||||
2541 | ExternalUses.push_back(ExternalUser(Scalar, U, FoundLane)); | ||||
2542 | } | ||||
2543 | } | ||||
2544 | } | ||||
2545 | } | ||||
2546 | |||||
2547 | void BoUpSLP::buildTree_rec(ArrayRef<Value *> VL, unsigned Depth, | ||||
2548 | const EdgeInfo &UserTreeIdx) { | ||||
2549 | assert((allConstant(VL) || allSameType(VL)) && "Invalid types!")(((allConstant(VL) || allSameType(VL)) && "Invalid types!" ) ? static_cast<void> (0) : __assert_fail ("(allConstant(VL) || allSameType(VL)) && \"Invalid types!\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2549, __PRETTY_FUNCTION__)); | ||||
2550 | |||||
2551 | InstructionsState S = getSameOpcode(VL); | ||||
2552 | if (Depth == RecursionMaxDepth) { | ||||
2553 | 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); | ||||
2554 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx); | ||||
2555 | return; | ||||
2556 | } | ||||
2557 | |||||
2558 | // Don't handle vectors. | ||||
2559 | if (S.OpValue->getType()->isVectorTy()) { | ||||
2560 | 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); | ||||
2561 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx); | ||||
2562 | return; | ||||
2563 | } | ||||
2564 | |||||
2565 | if (StoreInst *SI = dyn_cast<StoreInst>(S.OpValue)) | ||||
2566 | if (SI->getValueOperand()->getType()->isVectorTy()) { | ||||
2567 | 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); | ||||
2568 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx); | ||||
2569 | return; | ||||
2570 | } | ||||
2571 | |||||
2572 | // If all of the operands are identical or constant we have a simple solution. | ||||
2573 | if (allConstant(VL) || isSplat(VL) || !allSameBlock(VL) || !S.getOpcode()) { | ||||
2574 | LLVM_DEBUG(dbgs() << "SLP: Gathering due to C,S,B,O. \n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering due to C,S,B,O. \n" ; } } while (false); | ||||
2575 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx); | ||||
2576 | return; | ||||
2577 | } | ||||
2578 | |||||
2579 | // We now know that this is a vector of instructions of the same type from | ||||
2580 | // the same block. | ||||
2581 | |||||
2582 | // Don't vectorize ephemeral values. | ||||
2583 | for (Value *V : VL) { | ||||
2584 | if (EphValues.count(V)) { | ||||
2585 | LLVM_DEBUG(dbgs() << "SLP: The instruction (" << *Vdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: The instruction (" << * V << ") is ephemeral.\n"; } } while (false) | ||||
2586 | << ") is ephemeral.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: The instruction (" << * V << ") is ephemeral.\n"; } } while (false); | ||||
2587 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx); | ||||
2588 | return; | ||||
2589 | } | ||||
2590 | } | ||||
2591 | |||||
2592 | // Check if this is a duplicate of another entry. | ||||
2593 | if (TreeEntry *E = getTreeEntry(S.OpValue)) { | ||||
2594 | 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); | ||||
2595 | if (!E->isSame(VL)) { | ||||
2596 | 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); | ||||
2597 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx); | ||||
2598 | return; | ||||
2599 | } | ||||
2600 | // Record the reuse of the tree node. FIXME, currently this is only used to | ||||
2601 | // properly draw the graph rather than for the actual vectorization. | ||||
2602 | E->UserTreeIndices.push_back(UserTreeIdx); | ||||
2603 | 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) | ||||
2604 | << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Perfect diamond merge at " << *S.OpValue << ".\n"; } } while (false); | ||||
2605 | return; | ||||
2606 | } | ||||
2607 | |||||
2608 | // Check that none of the instructions in the bundle are already in the tree. | ||||
2609 | for (Value *V : VL) { | ||||
2610 | auto *I = dyn_cast<Instruction>(V); | ||||
2611 | if (!I) | ||||
2612 | continue; | ||||
2613 | if (getTreeEntry(I)) { | ||||
2614 | 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) | ||||
2615 | << ") is already in tree.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: The instruction (" << * V << ") is already in tree.\n"; } } while (false); | ||||
2616 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx); | ||||
2617 | return; | ||||
2618 | } | ||||
2619 | } | ||||
2620 | |||||
2621 | // If any of the scalars is marked as a value that needs to stay scalar, then | ||||
2622 | // we need to gather the scalars. | ||||
2623 | // The reduction nodes (stored in UserIgnoreList) also should stay scalar. | ||||
2624 | for (Value *V : VL) { | ||||
2625 | if (MustGather.count(V) || is_contained(UserIgnoreList, V)) { | ||||
2626 | 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); | ||||
2627 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx); | ||||
2628 | return; | ||||
2629 | } | ||||
2630 | } | ||||
2631 | |||||
2632 | // Check that all of the users of the scalars that we want to vectorize are | ||||
2633 | // schedulable. | ||||
2634 | auto *VL0 = cast<Instruction>(S.OpValue); | ||||
2635 | BasicBlock *BB = VL0->getParent(); | ||||
2636 | |||||
2637 | if (!DT->isReachableFromEntry(BB)) { | ||||
2638 | // Don't go into unreachable blocks. They may contain instructions with | ||||
2639 | // dependency cycles which confuse the final scheduling. | ||||
2640 | 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); | ||||
2641 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx); | ||||
2642 | return; | ||||
2643 | } | ||||
2644 | |||||
2645 | // Check that every instruction appears once in this bundle. | ||||
2646 | SmallVector<unsigned, 4> ReuseShuffleIndicies; | ||||
2647 | SmallVector<Value *, 4> UniqueValues; | ||||
2648 | DenseMap<Value *, unsigned> UniquePositions; | ||||
2649 | for (Value *V : VL) { | ||||
2650 | auto Res = UniquePositions.try_emplace(V, UniqueValues.size()); | ||||
2651 | ReuseShuffleIndicies.emplace_back(Res.first->second); | ||||
2652 | if (Res.second) | ||||
2653 | UniqueValues.emplace_back(V); | ||||
2654 | } | ||||
2655 | size_t NumUniqueScalarValues = UniqueValues.size(); | ||||
2656 | if (NumUniqueScalarValues == VL.size()) { | ||||
2657 | ReuseShuffleIndicies.clear(); | ||||
2658 | } else { | ||||
2659 | 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); | ||||
2660 | if (NumUniqueScalarValues <= 1 || | ||||
2661 | !llvm::isPowerOf2_32(NumUniqueScalarValues)) { | ||||
2662 | 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); | ||||
2663 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx); | ||||
2664 | return; | ||||
2665 | } | ||||
2666 | VL = UniqueValues; | ||||
2667 | } | ||||
2668 | |||||
2669 | auto &BSRef = BlocksSchedules[BB]; | ||||
2670 | if (!BSRef) | ||||
2671 | BSRef = std::make_unique<BlockScheduling>(BB); | ||||
2672 | |||||
2673 | BlockScheduling &BS = *BSRef.get(); | ||||
2674 | |||||
2675 | Optional<ScheduleData *> Bundle = BS.tryScheduleBundle(VL, this, S); | ||||
2676 | if (!Bundle) { | ||||
2677 | 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); | ||||
2678 | assert((!BS.getScheduleData(VL0) ||(((!BS.getScheduleData(VL0) || !BS.getScheduleData(VL0)->isPartOfBundle ()) && "tryScheduleBundle should cancelScheduling on failure" ) ? static_cast<void> (0) : __assert_fail ("(!BS.getScheduleData(VL0) || !BS.getScheduleData(VL0)->isPartOfBundle()) && \"tryScheduleBundle should cancelScheduling on failure\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2680, __PRETTY_FUNCTION__)) | ||||
2679 | !BS.getScheduleData(VL0)->isPartOfBundle()) &&(((!BS.getScheduleData(VL0) || !BS.getScheduleData(VL0)->isPartOfBundle ()) && "tryScheduleBundle should cancelScheduling on failure" ) ? static_cast<void> (0) : __assert_fail ("(!BS.getScheduleData(VL0) || !BS.getScheduleData(VL0)->isPartOfBundle()) && \"tryScheduleBundle should cancelScheduling on failure\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2680, __PRETTY_FUNCTION__)) | ||||
2680 | "tryScheduleBundle should cancelScheduling on failure")(((!BS.getScheduleData(VL0) || !BS.getScheduleData(VL0)->isPartOfBundle ()) && "tryScheduleBundle should cancelScheduling on failure" ) ? static_cast<void> (0) : __assert_fail ("(!BS.getScheduleData(VL0) || !BS.getScheduleData(VL0)->isPartOfBundle()) && \"tryScheduleBundle should cancelScheduling on failure\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2680, __PRETTY_FUNCTION__)); | ||||
2681 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | ||||
2682 | ReuseShuffleIndicies); | ||||
2683 | return; | ||||
2684 | } | ||||
2685 | 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); | ||||
2686 | |||||
2687 | unsigned ShuffleOrOp = S.isAltShuffle() ? | ||||
2688 | (unsigned) Instruction::ShuffleVector : S.getOpcode(); | ||||
2689 | switch (ShuffleOrOp) { | ||||
2690 | case Instruction::PHI: { | ||||
2691 | auto *PH = cast<PHINode>(VL0); | ||||
2692 | |||||
2693 | // Check for terminator values (e.g. invoke). | ||||
2694 | for (Value *V : VL) | ||||
2695 | for (unsigned I = 0, E = PH->getNumIncomingValues(); I < E; ++I) { | ||||
2696 | Instruction *Term = dyn_cast<Instruction>( | ||||
2697 | cast<PHINode>(V)->getIncomingValueForBlock( | ||||
2698 | PH->getIncomingBlock(I))); | ||||
2699 | if (Term && Term->isTerminator()) { | ||||
2700 | LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Need to swizzle PHINodes (terminator use).\n" ; } } while (false) | ||||
2701 | << "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); | ||||
2702 | BS.cancelScheduling(VL, VL0); | ||||
2703 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | ||||
2704 | ReuseShuffleIndicies); | ||||
2705 | return; | ||||
2706 | } | ||||
2707 | } | ||||
2708 | |||||
2709 | TreeEntry *TE = | ||||
2710 | newTreeEntry(VL, Bundle, S, UserTreeIdx, ReuseShuffleIndicies); | ||||
2711 | 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); | ||||
2712 | |||||
2713 | // Keeps the reordered operands to avoid code duplication. | ||||
2714 | SmallVector<ValueList, 2> OperandsVec; | ||||
2715 | for (unsigned I = 0, E = PH->getNumIncomingValues(); I < E; ++I) { | ||||
2716 | ValueList Operands; | ||||
2717 | // Prepare the operand vector. | ||||
2718 | for (Value *V : VL) | ||||
2719 | Operands.push_back(cast<PHINode>(V)->getIncomingValueForBlock( | ||||
2720 | PH->getIncomingBlock(I))); | ||||
2721 | TE->setOperand(I, Operands); | ||||
2722 | OperandsVec.push_back(Operands); | ||||
2723 | } | ||||
2724 | for (unsigned OpIdx = 0, OpE = OperandsVec.size(); OpIdx != OpE; ++OpIdx) | ||||
2725 | buildTree_rec(OperandsVec[OpIdx], Depth + 1, {TE, OpIdx}); | ||||
2726 | return; | ||||
2727 | } | ||||
2728 | case Instruction::ExtractValue: | ||||
2729 | case Instruction::ExtractElement: { | ||||
2730 | OrdersType CurrentOrder; | ||||
2731 | bool Reuse = canReuseExtract(VL, VL0, CurrentOrder); | ||||
2732 | if (Reuse) { | ||||
2733 | 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); | ||||
2734 | ++NumOpsWantToKeepOriginalOrder; | ||||
2735 | newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, | ||||
2736 | ReuseShuffleIndicies); | ||||
2737 | // This is a special case, as it does not gather, but at the same time | ||||
2738 | // we are not extending buildTree_rec() towards the operands. | ||||
2739 | ValueList Op0; | ||||
2740 | Op0.assign(VL.size(), VL0->getOperand(0)); | ||||
2741 | VectorizableTree.back()->setOperand(0, Op0); | ||||
2742 | return; | ||||
2743 | } | ||||
2744 | if (!CurrentOrder.empty()) { | ||||
2745 | 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) | ||||
2746 | 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) | ||||
2747 | "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) | ||||
2748 | 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) | ||||
2749 | 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) | ||||
2750 | 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) | ||||
2751 | })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); | ||||
2752 | // Insert new order with initial value 0, if it does not exist, | ||||
2753 | // otherwise return the iterator to the existing one. | ||||
2754 | newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, | ||||
2755 | ReuseShuffleIndicies, CurrentOrder); | ||||
2756 | findRootOrder(CurrentOrder); | ||||
2757 | ++NumOpsWantToKeepOrder[CurrentOrder]; | ||||
2758 | // This is a special case, as it does not gather, but at the same time | ||||
2759 | // we are not extending buildTree_rec() towards the operands. | ||||
2760 | ValueList Op0; | ||||
2761 | Op0.assign(VL.size(), VL0->getOperand(0)); | ||||
2762 | VectorizableTree.back()->setOperand(0, Op0); | ||||
2763 | return; | ||||
2764 | } | ||||
2765 | LLVM_DEBUG(dbgs() << "SLP: Gather extract sequence.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gather extract sequence.\n"; } } while (false); | ||||
2766 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | ||||
2767 | ReuseShuffleIndicies); | ||||
2768 | BS.cancelScheduling(VL, VL0); | ||||
2769 | return; | ||||
2770 | } | ||||
2771 | case Instruction::Load: { | ||||
2772 | // Check that a vectorized load would load the same memory as a scalar | ||||
2773 | // load. For example, we don't want to vectorize loads that are smaller | ||||
2774 | // than 8-bit. Even though we have a packed struct {<i2, i2, i2, i2>} LLVM | ||||
2775 | // treats loading/storing it as an i8 struct. If we vectorize loads/stores | ||||
2776 | // from such a struct, we read/write packed bits disagreeing with the | ||||
2777 | // unvectorized version. | ||||
2778 | Type *ScalarTy = VL0->getType(); | ||||
2779 | |||||
2780 | if (DL->getTypeSizeInBits(ScalarTy) != | ||||
2781 | DL->getTypeAllocSizeInBits(ScalarTy)) { | ||||
2782 | BS.cancelScheduling(VL, VL0); | ||||
2783 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | ||||
2784 | ReuseShuffleIndicies); | ||||
2785 | 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); | ||||
2786 | return; | ||||
2787 | } | ||||
2788 | |||||
2789 | // Make sure all loads in the bundle are simple - we can't vectorize | ||||
2790 | // atomic or volatile loads. | ||||
2791 | SmallVector<Value *, 4> PointerOps(VL.size()); | ||||
2792 | auto POIter = PointerOps.begin(); | ||||
2793 | for (Value *V : VL) { | ||||
2794 | auto *L = cast<LoadInst>(V); | ||||
2795 | if (!L->isSimple()) { | ||||
2796 | BS.cancelScheduling(VL, VL0); | ||||
2797 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | ||||
2798 | ReuseShuffleIndicies); | ||||
2799 | 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); | ||||
2800 | return; | ||||
2801 | } | ||||
2802 | *POIter = L->getPointerOperand(); | ||||
2803 | ++POIter; | ||||
2804 | } | ||||
2805 | |||||
2806 | OrdersType CurrentOrder; | ||||
2807 | // Check the order of pointer operands. | ||||
2808 | if (llvm::sortPtrAccesses(PointerOps, *DL, *SE, CurrentOrder)) { | ||||
2809 | Value *Ptr0; | ||||
2810 | Value *PtrN; | ||||
2811 | if (CurrentOrder.empty()) { | ||||
2812 | Ptr0 = PointerOps.front(); | ||||
2813 | PtrN = PointerOps.back(); | ||||
2814 | } else { | ||||
2815 | Ptr0 = PointerOps[CurrentOrder.front()]; | ||||
2816 | PtrN = PointerOps[CurrentOrder.back()]; | ||||
2817 | } | ||||
2818 | const SCEV *Scev0 = SE->getSCEV(Ptr0); | ||||
2819 | const SCEV *ScevN = SE->getSCEV(PtrN); | ||||
2820 | const auto *Diff = | ||||
2821 | dyn_cast<SCEVConstant>(SE->getMinusSCEV(ScevN, Scev0)); | ||||
2822 | uint64_t Size = DL->getTypeAllocSize(ScalarTy); | ||||
2823 | // Check that the sorted loads are consecutive. | ||||
2824 | if (Diff && Diff->getAPInt() == (VL.size() - 1) * Size) { | ||||
2825 | if (CurrentOrder.empty()) { | ||||
2826 | // Original loads are consecutive and does not require reordering. | ||||
2827 | ++NumOpsWantToKeepOriginalOrder; | ||||
2828 | TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, | ||||
2829 | UserTreeIdx, ReuseShuffleIndicies); | ||||
2830 | TE->setOperandsInOrder(); | ||||
2831 | 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); | ||||
2832 | } else { | ||||
2833 | // Need to reorder. | ||||
2834 | TreeEntry *TE = | ||||
2835 | newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, | ||||
2836 | ReuseShuffleIndicies, CurrentOrder); | ||||
2837 | TE->setOperandsInOrder(); | ||||
2838 | 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); | ||||
2839 | findRootOrder(CurrentOrder); | ||||
2840 | ++NumOpsWantToKeepOrder[CurrentOrder]; | ||||
2841 | } | ||||
2842 | return; | ||||
2843 | } | ||||
2844 | } | ||||
2845 | |||||
2846 | 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); | ||||
2847 | BS.cancelScheduling(VL, VL0); | ||||
2848 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | ||||
2849 | ReuseShuffleIndicies); | ||||
2850 | return; | ||||
2851 | } | ||||
2852 | case Instruction::ZExt: | ||||
2853 | case Instruction::SExt: | ||||
2854 | case Instruction::FPToUI: | ||||
2855 | case Instruction::FPToSI: | ||||
2856 | case Instruction::FPExt: | ||||
2857 | case Instruction::PtrToInt: | ||||
2858 | case Instruction::IntToPtr: | ||||
2859 | case Instruction::SIToFP: | ||||
2860 | case Instruction::UIToFP: | ||||
2861 | case Instruction::Trunc: | ||||
2862 | case Instruction::FPTrunc: | ||||
2863 | case Instruction::BitCast: { | ||||
2864 | Type *SrcTy = VL0->getOperand(0)->getType(); | ||||
2865 | for (Value *V : VL) { | ||||
2866 | Type *Ty = cast<Instruction>(V)->getOperand(0)->getType(); | ||||
2867 | if (Ty != SrcTy || !isValidElementType(Ty)) { | ||||
2868 | BS.cancelScheduling(VL, VL0); | ||||
2869 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | ||||
2870 | ReuseShuffleIndicies); | ||||
2871 | LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering casts with different src types.\n" ; } } while (false) | ||||
2872 | << "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); | ||||
2873 | return; | ||||
2874 | } | ||||
2875 | } | ||||
2876 | TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, | ||||
2877 | ReuseShuffleIndicies); | ||||
2878 | 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); | ||||
2879 | |||||
2880 | TE->setOperandsInOrder(); | ||||
2881 | for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) { | ||||
2882 | ValueList Operands; | ||||
2883 | // Prepare the operand vector. | ||||
2884 | for (Value *V : VL) | ||||
2885 | Operands.push_back(cast<Instruction>(V)->getOperand(i)); | ||||
2886 | |||||
2887 | buildTree_rec(Operands, Depth + 1, {TE, i}); | ||||
2888 | } | ||||
2889 | return; | ||||
2890 | } | ||||
2891 | case Instruction::ICmp: | ||||
2892 | case Instruction::FCmp: { | ||||
2893 | // Check that all of the compares have the same predicate. | ||||
2894 | CmpInst::Predicate P0 = cast<CmpInst>(VL0)->getPredicate(); | ||||
2895 | CmpInst::Predicate SwapP0 = CmpInst::getSwappedPredicate(P0); | ||||
2896 | Type *ComparedTy = VL0->getOperand(0)->getType(); | ||||
2897 | for (Value *V : VL) { | ||||
2898 | CmpInst *Cmp = cast<CmpInst>(V); | ||||
2899 | if ((Cmp->getPredicate() != P0 && Cmp->getPredicate() != SwapP0) || | ||||
2900 | Cmp->getOperand(0)->getType() != ComparedTy) { | ||||
2901 | BS.cancelScheduling(VL, VL0); | ||||
2902 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | ||||
2903 | ReuseShuffleIndicies); | ||||
2904 | LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering cmp with different predicate.\n" ; } } while (false) | ||||
2905 | << "SLP: Gathering cmp with different predicate.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering cmp with different predicate.\n" ; } } while (false); | ||||
2906 | return; | ||||
2907 | } | ||||
2908 | } | ||||
2909 | |||||
2910 | TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, | ||||
2911 | ReuseShuffleIndicies); | ||||
2912 | 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); | ||||
2913 | |||||
2914 | ValueList Left, Right; | ||||
2915 | if (cast<CmpInst>(VL0)->isCommutative()) { | ||||
2916 | // Commutative predicate - collect + sort operands of the instructions | ||||
2917 | // so that each side is more likely to have the same opcode. | ||||
2918 | assert(P0 == SwapP0 && "Commutative Predicate mismatch")((P0 == SwapP0 && "Commutative Predicate mismatch") ? static_cast<void> (0) : __assert_fail ("P0 == SwapP0 && \"Commutative Predicate mismatch\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2918, __PRETTY_FUNCTION__)); | ||||
2919 | reorderInputsAccordingToOpcode(VL, Left, Right, *DL, *SE, *this); | ||||
2920 | } else { | ||||
2921 | // Collect operands - commute if it uses the swapped predicate. | ||||
2922 | for (Value *V : VL) { | ||||
2923 | auto *Cmp = cast<CmpInst>(V); | ||||
2924 | Value *LHS = Cmp->getOperand(0); | ||||
2925 | Value *RHS = Cmp->getOperand(1); | ||||
2926 | if (Cmp->getPredicate() != P0) | ||||
2927 | std::swap(LHS, RHS); | ||||
2928 | Left.push_back(LHS); | ||||
2929 | Right.push_back(RHS); | ||||
2930 | } | ||||
2931 | } | ||||
2932 | TE->setOperand(0, Left); | ||||
2933 | TE->setOperand(1, Right); | ||||
2934 | buildTree_rec(Left, Depth + 1, {TE, 0}); | ||||
2935 | buildTree_rec(Right, Depth + 1, {TE, 1}); | ||||
2936 | return; | ||||
2937 | } | ||||
2938 | case Instruction::Select: | ||||
2939 | case Instruction::FNeg: | ||||
2940 | case Instruction::Add: | ||||
2941 | case Instruction::FAdd: | ||||
2942 | case Instruction::Sub: | ||||
2943 | case Instruction::FSub: | ||||
2944 | case Instruction::Mul: | ||||
2945 | case Instruction::FMul: | ||||
2946 | case Instruction::UDiv: | ||||
2947 | case Instruction::SDiv: | ||||
2948 | case Instruction::FDiv: | ||||
2949 | case Instruction::URem: | ||||
2950 | case Instruction::SRem: | ||||
2951 | case Instruction::FRem: | ||||
2952 | case Instruction::Shl: | ||||
2953 | case Instruction::LShr: | ||||
2954 | case Instruction::AShr: | ||||
2955 | case Instruction::And: | ||||
2956 | case Instruction::Or: | ||||
2957 | case Instruction::Xor: { | ||||
2958 | TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, | ||||
2959 | ReuseShuffleIndicies); | ||||
2960 | 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); | ||||
2961 | |||||
2962 | // Sort operands of the instructions so that each side is more likely to | ||||
2963 | // have the same opcode. | ||||
2964 | if (isa<BinaryOperator>(VL0) && VL0->isCommutative()) { | ||||
2965 | ValueList Left, Right; | ||||
2966 | reorderInputsAccordingToOpcode(VL, Left, Right, *DL, *SE, *this); | ||||
2967 | TE->setOperand(0, Left); | ||||
2968 | TE->setOperand(1, Right); | ||||
2969 | buildTree_rec(Left, Depth + 1, {TE, 0}); | ||||
2970 | buildTree_rec(Right, Depth + 1, {TE, 1}); | ||||
2971 | return; | ||||
2972 | } | ||||
2973 | |||||
2974 | TE->setOperandsInOrder(); | ||||
2975 | for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) { | ||||
2976 | ValueList Operands; | ||||
2977 | // Prepare the operand vector. | ||||
2978 | for (Value *V : VL) | ||||
2979 | Operands.push_back(cast<Instruction>(V)->getOperand(i)); | ||||
2980 | |||||
2981 | buildTree_rec(Operands, Depth + 1, {TE, i}); | ||||
2982 | } | ||||
2983 | return; | ||||
2984 | } | ||||
2985 | case Instruction::GetElementPtr: { | ||||
2986 | // We don't combine GEPs with complicated (nested) indexing. | ||||
2987 | for (Value *V : VL) { | ||||
2988 | if (cast<Instruction>(V)->getNumOperands() != 2) { | ||||
2989 | 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); | ||||
2990 | BS.cancelScheduling(VL, VL0); | ||||
2991 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | ||||
2992 | ReuseShuffleIndicies); | ||||
2993 | return; | ||||
2994 | } | ||||
2995 | } | ||||
2996 | |||||
2997 | // We can't combine several GEPs into one vector if they operate on | ||||
2998 | // different types. | ||||
2999 | Type *Ty0 = VL0->getOperand(0)->getType(); | ||||
3000 | for (Value *V : VL) { | ||||
3001 | Type *CurTy = cast<Instruction>(V)->getOperand(0)->getType(); | ||||
3002 | if (Ty0 != CurTy) { | ||||
3003 | LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: not-vectorizable GEP (different types).\n" ; } } while (false) | ||||
3004 | << "SLP: not-vectorizable GEP (different types).\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: not-vectorizable GEP (different types).\n" ; } } while (false); | ||||
3005 | BS.cancelScheduling(VL, VL0); | ||||
3006 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | ||||
3007 | ReuseShuffleIndicies); | ||||
3008 | return; | ||||
3009 | } | ||||
3010 | } | ||||
3011 | |||||
3012 | // We don't combine GEPs with non-constant indexes. | ||||
3013 | Type *Ty1 = VL0->getOperand(1)->getType(); | ||||
3014 | for (Value *V : VL) { | ||||
3015 | auto Op = cast<Instruction>(V)->getOperand(1); | ||||
3016 | if (!isa<ConstantInt>(Op) || | ||||
3017 | (Op->getType() != Ty1 && | ||||
3018 | Op->getType()->getScalarSizeInBits() > | ||||
3019 | DL->getIndexSizeInBits( | ||||
3020 | V->getType()->getPointerAddressSpace()))) { | ||||
3021 | LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: not-vectorizable GEP (non-constant indexes).\n" ; } } while (false) | ||||
3022 | << "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); | ||||
3023 | BS.cancelScheduling(VL, VL0); | ||||
3024 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | ||||
3025 | ReuseShuffleIndicies); | ||||
3026 | return; | ||||
3027 | } | ||||
3028 | } | ||||
3029 | |||||
3030 | TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, | ||||
3031 | ReuseShuffleIndicies); | ||||
3032 | 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); | ||||
3033 | TE->setOperandsInOrder(); | ||||
3034 | for (unsigned i = 0, e = 2; i < e; ++i) { | ||||
3035 | ValueList Operands; | ||||
3036 | // Prepare the operand vector. | ||||
3037 | for (Value *V : VL) | ||||
3038 | Operands.push_back(cast<Instruction>(V)->getOperand(i)); | ||||
3039 | |||||
3040 | buildTree_rec(Operands, Depth + 1, {TE, i}); | ||||
3041 | } | ||||
3042 | return; | ||||
3043 | } | ||||
3044 | case Instruction::Store: { | ||||
3045 | // Check if the stores are consecutive or if we need to swizzle them. | ||||
3046 | llvm::Type *ScalarTy = cast<StoreInst>(VL0)->getValueOperand()->getType(); | ||||
3047 | // Make sure all stores in the bundle are simple - we can't vectorize | ||||
3048 | // atomic or volatile stores. | ||||
3049 | SmallVector<Value *, 4> PointerOps(VL.size()); | ||||
3050 | ValueList Operands(VL.size()); | ||||
3051 | auto POIter = PointerOps.begin(); | ||||
3052 | auto OIter = Operands.begin(); | ||||
3053 | for (Value *V : VL) { | ||||
3054 | auto *SI = cast<StoreInst>(V); | ||||
3055 | if (!SI->isSimple()) { | ||||
3056 | BS.cancelScheduling(VL, VL0); | ||||
3057 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | ||||
3058 | ReuseShuffleIndicies); | ||||
3059 | 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); | ||||
3060 | return; | ||||
3061 | } | ||||
3062 | *POIter = SI->getPointerOperand(); | ||||
3063 | *OIter = SI->getValueOperand(); | ||||
3064 | ++POIter; | ||||
3065 | ++OIter; | ||||
3066 | } | ||||
3067 | |||||
3068 | OrdersType CurrentOrder; | ||||
3069 | // Check the order of pointer operands. | ||||
3070 | if (llvm::sortPtrAccesses(PointerOps, *DL, *SE, CurrentOrder)) { | ||||
3071 | Value *Ptr0; | ||||
3072 | Value *PtrN; | ||||
3073 | if (CurrentOrder.empty()) { | ||||
3074 | Ptr0 = PointerOps.front(); | ||||
3075 | PtrN = PointerOps.back(); | ||||
3076 | } else { | ||||
3077 | Ptr0 = PointerOps[CurrentOrder.front()]; | ||||
3078 | PtrN = PointerOps[CurrentOrder.back()]; | ||||
3079 | } | ||||
3080 | const SCEV *Scev0 = SE->getSCEV(Ptr0); | ||||
3081 | const SCEV *ScevN = SE->getSCEV(PtrN); | ||||
3082 | const auto *Diff = | ||||
3083 | dyn_cast<SCEVConstant>(SE->getMinusSCEV(ScevN, Scev0)); | ||||
3084 | uint64_t Size = DL->getTypeAllocSize(ScalarTy); | ||||
3085 | // Check that the sorted pointer operands are consecutive. | ||||
3086 | if (Diff && Diff->getAPInt() == (VL.size() - 1) * Size) { | ||||
3087 | if (CurrentOrder.empty()) { | ||||
3088 | // Original stores are consecutive and does not require reordering. | ||||
3089 | ++NumOpsWantToKeepOriginalOrder; | ||||
3090 | TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, | ||||
3091 | UserTreeIdx, ReuseShuffleIndicies); | ||||
3092 | TE->setOperandsInOrder(); | ||||
3093 | buildTree_rec(Operands, Depth + 1, {TE, 0}); | ||||
3094 | 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); | ||||
3095 | } else { | ||||
3096 | TreeEntry *TE = | ||||
3097 | newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, | ||||
3098 | ReuseShuffleIndicies, CurrentOrder); | ||||
3099 | TE->setOperandsInOrder(); | ||||
3100 | buildTree_rec(Operands, Depth + 1, {TE, 0}); | ||||
3101 | 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); | ||||
3102 | findRootOrder(CurrentOrder); | ||||
3103 | ++NumOpsWantToKeepOrder[CurrentOrder]; | ||||
3104 | } | ||||
3105 | return; | ||||
3106 | } | ||||
3107 | } | ||||
3108 | |||||
3109 | BS.cancelScheduling(VL, VL0); | ||||
3110 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | ||||
3111 | ReuseShuffleIndicies); | ||||
3112 | LLVM_DEBUG(dbgs() << "SLP: Non-consecutive store.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Non-consecutive store.\n"; } } while (false); | ||||
3113 | return; | ||||
3114 | } | ||||
3115 | case Instruction::Call: { | ||||
3116 | // Check if the calls are all to the same vectorizable intrinsic or | ||||
3117 | // library function. | ||||
3118 | CallInst *CI = cast<CallInst>(VL0); | ||||
3119 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); | ||||
3120 | |||||
3121 | VFShape Shape = VFShape::get( | ||||
3122 | *CI, ElementCount::getFixed(static_cast<unsigned int>(VL.size())), | ||||
3123 | false /*HasGlobalPred*/); | ||||
3124 | Function *VecFunc = VFDatabase(*CI).getVectorizedFunction(Shape); | ||||
3125 | |||||
3126 | if (!VecFunc && !isTriviallyVectorizable(ID)) { | ||||
3127 | BS.cancelScheduling(VL, VL0); | ||||
3128 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | ||||
3129 | ReuseShuffleIndicies); | ||||
3130 | LLVM_DEBUG(dbgs() << "SLP: Non-vectorizable call.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Non-vectorizable call.\n"; } } while (false); | ||||
3131 | return; | ||||
3132 | } | ||||
3133 | Function *F = CI->getCalledFunction(); | ||||
3134 | unsigned NumArgs = CI->getNumArgOperands(); | ||||
3135 | SmallVector<Value*, 4> ScalarArgs(NumArgs, nullptr); | ||||
3136 | for (unsigned j = 0; j != NumArgs; ++j) | ||||
3137 | if (hasVectorInstrinsicScalarOpd(ID, j)) | ||||
3138 | ScalarArgs[j] = CI->getArgOperand(j); | ||||
3139 | for (Value *V : VL) { | ||||
3140 | CallInst *CI2 = dyn_cast<CallInst>(V); | ||||
3141 | if (!CI2 || CI2->getCalledFunction() != F || | ||||
3142 | getVectorIntrinsicIDForCall(CI2, TLI) != ID || | ||||
3143 | (VecFunc && | ||||
3144 | VecFunc != VFDatabase(*CI2).getVectorizedFunction(Shape)) || | ||||
3145 | !CI->hasIdenticalOperandBundleSchema(*CI2)) { | ||||
3146 | BS.cancelScheduling(VL, VL0); | ||||
3147 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | ||||
3148 | ReuseShuffleIndicies); | ||||
3149 | LLVM_DEBUG(dbgs() << "SLP: mismatched calls:" << *CI << "!=" << *Vdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: mismatched calls:" << * CI << "!=" << *V << "\n"; } } while (false) | ||||
3150 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: mismatched calls:" << * CI << "!=" << *V << "\n"; } } while (false); | ||||
3151 | return; | ||||
3152 | } | ||||
3153 | // Some intrinsics have scalar arguments and should be same in order for | ||||
3154 | // them to be vectorized. | ||||
3155 | for (unsigned j = 0; j != NumArgs; ++j) { | ||||
3156 | if (hasVectorInstrinsicScalarOpd(ID, j)) { | ||||
3157 | Value *A1J = CI2->getArgOperand(j); | ||||
3158 | if (ScalarArgs[j] != A1J) { | ||||
3159 | BS.cancelScheduling(VL, VL0); | ||||
3160 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | ||||
3161 | ReuseShuffleIndicies); | ||||
3162 | 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) | ||||
3163 | << " argument " << ScalarArgs[j] << "!=" << A1Jdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: mismatched arguments in call:" << *CI << " argument " << ScalarArgs[j] << "!=" << A1J << "\n"; } } while (false) | ||||
3164 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: mismatched arguments in call:" << *CI << " argument " << ScalarArgs[j] << "!=" << A1J << "\n"; } } while (false); | ||||
3165 | return; | ||||
3166 | } | ||||
3167 | } | ||||
3168 | } | ||||
3169 | // Verify that the bundle operands are identical between the two calls. | ||||
3170 | if (CI->hasOperandBundles() && | ||||
3171 | !std::equal(CI->op_begin() + CI->getBundleOperandsStartIndex(), | ||||
3172 | CI->op_begin() + CI->getBundleOperandsEndIndex(), | ||||
3173 | CI2->op_begin() + CI2->getBundleOperandsStartIndex())) { | ||||
3174 | BS.cancelScheduling(VL, VL0); | ||||
3175 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | ||||
3176 | ReuseShuffleIndicies); | ||||
3177 | 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) | ||||
3178 | << *CI << "!=" << *V << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: mismatched bundle operands in calls:" << *CI << "!=" << *V << '\n'; } } while (false); | ||||
3179 | return; | ||||
3180 | } | ||||
3181 | } | ||||
3182 | |||||
3183 | TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, | ||||
3184 | ReuseShuffleIndicies); | ||||
3185 | TE->setOperandsInOrder(); | ||||
3186 | for (unsigned i = 0, e = CI->getNumArgOperands(); i != e; ++i) { | ||||
3187 | ValueList Operands; | ||||
3188 | // Prepare the operand vector. | ||||
3189 | for (Value *V : VL) { | ||||
3190 | auto *CI2 = cast<CallInst>(V); | ||||
3191 | Operands.push_back(CI2->getArgOperand(i)); | ||||
3192 | } | ||||
3193 | buildTree_rec(Operands, Depth + 1, {TE, i}); | ||||
3194 | } | ||||
3195 | return; | ||||
3196 | } | ||||
3197 | case Instruction::ShuffleVector: { | ||||
3198 | // If this is not an alternate sequence of opcode like add-sub | ||||
3199 | // then do not vectorize this instruction. | ||||
3200 | if (!S.isAltShuffle()) { | ||||
3201 | BS.cancelScheduling(VL, VL0); | ||||
3202 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | ||||
3203 | ReuseShuffleIndicies); | ||||
3204 | 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); | ||||
3205 | return; | ||||
3206 | } | ||||
3207 | TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, | ||||
3208 | ReuseShuffleIndicies); | ||||
3209 | 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); | ||||
3210 | |||||
3211 | // Reorder operands if reordering would enable vectorization. | ||||
3212 | if (isa<BinaryOperator>(VL0)) { | ||||
3213 | ValueList Left, Right; | ||||
3214 | reorderInputsAccordingToOpcode(VL, Left, Right, *DL, *SE, *this); | ||||
3215 | TE->setOperand(0, Left); | ||||
3216 | TE->setOperand(1, Right); | ||||
3217 | buildTree_rec(Left, Depth + 1, {TE, 0}); | ||||
3218 | buildTree_rec(Right, Depth + 1, {TE, 1}); | ||||
3219 | return; | ||||
3220 | } | ||||
3221 | |||||
3222 | TE->setOperandsInOrder(); | ||||
3223 | for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) { | ||||
3224 | ValueList Operands; | ||||
3225 | // Prepare the operand vector. | ||||
3226 | for (Value *V : VL) | ||||
3227 | Operands.push_back(cast<Instruction>(V)->getOperand(i)); | ||||
3228 | |||||
3229 | buildTree_rec(Operands, Depth + 1, {TE, i}); | ||||
3230 | } | ||||
3231 | return; | ||||
3232 | } | ||||
3233 | default: | ||||
3234 | BS.cancelScheduling(VL, VL0); | ||||
3235 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | ||||
3236 | ReuseShuffleIndicies); | ||||
3237 | LLVM_DEBUG(dbgs() << "SLP: Gathering unknown instruction.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering unknown instruction.\n" ; } } while (false); | ||||
3238 | return; | ||||
3239 | } | ||||
3240 | } | ||||
3241 | |||||
3242 | unsigned BoUpSLP::canMapToVector(Type *T, const DataLayout &DL) const { | ||||
3243 | unsigned N = 1; | ||||
3244 | Type *EltTy = T; | ||||
3245 | |||||
3246 | while (isa<StructType>(EltTy) || isa<ArrayType>(EltTy) || | ||||
3247 | isa<VectorType>(EltTy)) { | ||||
3248 | if (auto *ST = dyn_cast<StructType>(EltTy)) { | ||||
3249 | // Check that struct is homogeneous. | ||||
3250 | for (const auto *Ty : ST->elements()) | ||||
3251 | if (Ty != *ST->element_begin()) | ||||
3252 | return 0; | ||||
3253 | N *= ST->getNumElements(); | ||||
3254 | EltTy = *ST->element_begin(); | ||||
3255 | } else if (auto *AT = dyn_cast<ArrayType>(EltTy)) { | ||||
3256 | N *= AT->getNumElements(); | ||||
3257 | EltTy = AT->getElementType(); | ||||
3258 | } else { | ||||
3259 | auto *VT = cast<FixedVectorType>(EltTy); | ||||
3260 | N *= VT->getNumElements(); | ||||
3261 | EltTy = VT->getElementType(); | ||||
3262 | } | ||||
3263 | } | ||||
3264 | |||||
3265 | if (!isValidElementType(EltTy)) | ||||
3266 | return 0; | ||||
3267 | uint64_t VTSize = DL.getTypeStoreSizeInBits(FixedVectorType::get(EltTy, N)); | ||||
3268 | if (VTSize < MinVecRegSize || VTSize > MaxVecRegSize || VTSize != DL.getTypeStoreSizeInBits(T)) | ||||
3269 | return 0; | ||||
3270 | return N; | ||||
3271 | } | ||||
3272 | |||||
3273 | bool BoUpSLP::canReuseExtract(ArrayRef<Value *> VL, Value *OpValue, | ||||
3274 | SmallVectorImpl<unsigned> &CurrentOrder) const { | ||||
3275 | Instruction *E0 = cast<Instruction>(OpValue); | ||||
3276 | assert(E0->getOpcode() == Instruction::ExtractElement ||((E0->getOpcode() == Instruction::ExtractElement || E0-> getOpcode() == Instruction::ExtractValue) ? static_cast<void > (0) : __assert_fail ("E0->getOpcode() == Instruction::ExtractElement || E0->getOpcode() == Instruction::ExtractValue" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3277, __PRETTY_FUNCTION__)) | ||||
3277 | E0->getOpcode() == Instruction::ExtractValue)((E0->getOpcode() == Instruction::ExtractElement || E0-> getOpcode() == Instruction::ExtractValue) ? static_cast<void > (0) : __assert_fail ("E0->getOpcode() == Instruction::ExtractElement || E0->getOpcode() == Instruction::ExtractValue" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3277, __PRETTY_FUNCTION__)); | ||||
3278 | assert(E0->getOpcode() == getSameOpcode(VL).getOpcode() && "Invalid opcode")((E0->getOpcode() == getSameOpcode(VL).getOpcode() && "Invalid opcode") ? static_cast<void> (0) : __assert_fail ("E0->getOpcode() == getSameOpcode(VL).getOpcode() && \"Invalid opcode\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3278, __PRETTY_FUNCTION__)); | ||||
3279 | // Check if all of the extracts come from the same vector and from the | ||||
3280 | // correct offset. | ||||
3281 | Value *Vec = E0->getOperand(0); | ||||
3282 | |||||
3283 | CurrentOrder.clear(); | ||||
3284 | |||||
3285 | // We have to extract from a vector/aggregate with the same number of elements. | ||||
3286 | unsigned NElts; | ||||
3287 | if (E0->getOpcode() == Instruction::ExtractValue) { | ||||
3288 | const DataLayout &DL = E0->getModule()->getDataLayout(); | ||||
3289 | NElts = canMapToVector(Vec->getType(), DL); | ||||
3290 | if (!NElts) | ||||
3291 | return false; | ||||
3292 | // Check if load can be rewritten as load of vector. | ||||
3293 | LoadInst *LI = dyn_cast<LoadInst>(Vec); | ||||
3294 | if (!LI || !LI->isSimple() || !LI->hasNUses(VL.size())) | ||||
3295 | return false; | ||||
3296 | } else { | ||||
3297 | NElts = cast<FixedVectorType>(Vec->getType())->getNumElements(); | ||||
3298 | } | ||||
3299 | |||||
3300 | if (NElts != VL.size()) | ||||
3301 | return false; | ||||
3302 | |||||
3303 | // Check that all of the indices extract from the correct offset. | ||||
3304 | bool ShouldKeepOrder = true; | ||||
3305 | unsigned E = VL.size(); | ||||
3306 | // Assign to all items the initial value E + 1 so we can check if the extract | ||||
3307 | // instruction index was used already. | ||||
3308 | // Also, later we can check that all the indices are used and we have a | ||||
3309 | // consecutive access in the extract instructions, by checking that no | ||||
3310 | // element of CurrentOrder still has value E + 1. | ||||
3311 | CurrentOrder.assign(E, E + 1); | ||||
3312 | unsigned I = 0; | ||||
3313 | for (; I < E; ++I) { | ||||
3314 | auto *Inst = cast<Instruction>(VL[I]); | ||||
3315 | if (Inst->getOperand(0) != Vec) | ||||
3316 | break; | ||||
3317 | Optional<unsigned> Idx = getExtractIndex(Inst); | ||||
3318 | if (!Idx) | ||||
3319 | break; | ||||
3320 | const unsigned ExtIdx = *Idx; | ||||
3321 | if (ExtIdx != I) { | ||||
3322 | if (ExtIdx >= E || CurrentOrder[ExtIdx] != E + 1) | ||||
3323 | break; | ||||
3324 | ShouldKeepOrder = false; | ||||
3325 | CurrentOrder[ExtIdx] = I; | ||||
3326 | } else { | ||||
3327 | if (CurrentOrder[I] != E + 1) | ||||
3328 | break; | ||||
3329 | CurrentOrder[I] = I; | ||||
3330 | } | ||||
3331 | } | ||||
3332 | if (I < E) { | ||||
3333 | CurrentOrder.clear(); | ||||
3334 | return false; | ||||
3335 | } | ||||
3336 | |||||
3337 | return ShouldKeepOrder; | ||||
3338 | } | ||||
3339 | |||||
3340 | bool BoUpSLP::areAllUsersVectorized(Instruction *I) const { | ||||
3341 | return I->hasOneUse() || | ||||
3342 | std::all_of(I->user_begin(), I->user_end(), [this](User *U) { | ||||
3343 | return ScalarToTreeEntry.count(U) > 0; | ||||
3344 | }); | ||||
3345 | } | ||||
3346 | |||||
3347 | static std::pair<unsigned, unsigned> | ||||
3348 | getVectorCallCosts(CallInst *CI, FixedVectorType *VecTy, | ||||
3349 | TargetTransformInfo *TTI, TargetLibraryInfo *TLI) { | ||||
3350 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); | ||||
3351 | |||||
3352 | // Calculate the cost of the scalar and vector calls. | ||||
3353 | IntrinsicCostAttributes CostAttrs(ID, *CI, VecTy->getNumElements()); | ||||
3354 | int IntrinsicCost = | ||||
3355 | TTI->getIntrinsicInstrCost(CostAttrs, TTI::TCK_RecipThroughput); | ||||
3356 | |||||
3357 | auto Shape = VFShape::get(*CI, ElementCount::getFixed(static_cast<unsigned>( | ||||
3358 | VecTy->getNumElements())), | ||||
3359 | false /*HasGlobalPred*/); | ||||
3360 | Function *VecFunc = VFDatabase(*CI).getVectorizedFunction(Shape); | ||||
3361 | int LibCost = IntrinsicCost; | ||||
3362 | if (!CI->isNoBuiltin() && VecFunc) { | ||||
3363 | // Calculate the cost of the vector library call. | ||||
3364 | SmallVector<Type *, 4> VecTys; | ||||
3365 | for (Use &Arg : CI->args()) | ||||
3366 | VecTys.push_back( | ||||
3367 | FixedVectorType::get(Arg->getType(), VecTy->getNumElements())); | ||||
3368 | |||||
3369 | // If the corresponding vector call is cheaper, return its cost. | ||||
3370 | LibCost = TTI->getCallInstrCost(nullptr, VecTy, VecTys, | ||||
3371 | TTI::TCK_RecipThroughput); | ||||
3372 | } | ||||
3373 | return {IntrinsicCost, LibCost}; | ||||
3374 | } | ||||
3375 | |||||
3376 | int BoUpSLP::getEntryCost(TreeEntry *E) { | ||||
3377 | ArrayRef<Value*> VL = E->Scalars; | ||||
3378 | |||||
3379 | Type *ScalarTy = VL[0]->getType(); | ||||
3380 | if (StoreInst *SI = dyn_cast<StoreInst>(VL[0])) | ||||
3381 | ScalarTy = SI->getValueOperand()->getType(); | ||||
3382 | else if (CmpInst *CI = dyn_cast<CmpInst>(VL[0])) | ||||
3383 | ScalarTy = CI->getOperand(0)->getType(); | ||||
3384 | auto *VecTy = FixedVectorType::get(ScalarTy, VL.size()); | ||||
3385 | TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput; | ||||
3386 | |||||
3387 | // If we have computed a smaller type for the expression, update VecTy so | ||||
3388 | // that the costs will be accurate. | ||||
3389 | if (MinBWs.count(VL[0])) | ||||
3390 | VecTy = FixedVectorType::get( | ||||
3391 | IntegerType::get(F->getContext(), MinBWs[VL[0]].first), VL.size()); | ||||
3392 | |||||
3393 | unsigned ReuseShuffleNumbers = E->ReuseShuffleIndices.size(); | ||||
3394 | bool NeedToShuffleReuses = !E->ReuseShuffleIndices.empty(); | ||||
3395 | int ReuseShuffleCost = 0; | ||||
3396 | if (NeedToShuffleReuses) { | ||||
3397 | ReuseShuffleCost = | ||||
3398 | TTI->getShuffleCost(TargetTransformInfo::SK_PermuteSingleSrc, VecTy); | ||||
3399 | } | ||||
3400 | if (E->State == TreeEntry::NeedToGather) { | ||||
3401 | if (allConstant(VL)) | ||||
3402 | return 0; | ||||
3403 | if (isSplat(VL)) { | ||||
3404 | return ReuseShuffleCost + | ||||
3405 | TTI->getShuffleCost(TargetTransformInfo::SK_Broadcast, VecTy, 0); | ||||
3406 | } | ||||
3407 | if (E->getOpcode() == Instruction::ExtractElement && | ||||
3408 | allSameType(VL) && allSameBlock(VL)) { | ||||
3409 | Optional<TargetTransformInfo::ShuffleKind> ShuffleKind = isShuffle(VL); | ||||
3410 | if (ShuffleKind.hasValue()) { | ||||
3411 | int Cost = TTI->getShuffleCost(ShuffleKind.getValue(), VecTy); | ||||
3412 | for (auto *V : VL) { | ||||
3413 | // If all users of instruction are going to be vectorized and this | ||||
3414 | // instruction itself is not going to be vectorized, consider this | ||||
3415 | // instruction as dead and remove its cost from the final cost of the | ||||
3416 | // vectorized tree. | ||||
3417 | if (areAllUsersVectorized(cast<Instruction>(V)) && | ||||
3418 | !ScalarToTreeEntry.count(V)) { | ||||
3419 | auto *IO = cast<ConstantInt>( | ||||
3420 | cast<ExtractElementInst>(V)->getIndexOperand()); | ||||
3421 | Cost -= TTI->getVectorInstrCost(Instruction::ExtractElement, VecTy, | ||||
3422 | IO->getZExtValue()); | ||||
3423 | } | ||||
3424 | } | ||||
3425 | return ReuseShuffleCost + Cost; | ||||
3426 | } | ||||
3427 | } | ||||
3428 | return ReuseShuffleCost + getGatherCost(VL); | ||||
3429 | } | ||||
3430 | assert(E->State == TreeEntry::Vectorize && "Unhandled state")((E->State == TreeEntry::Vectorize && "Unhandled state" ) ? static_cast<void> (0) : __assert_fail ("E->State == TreeEntry::Vectorize && \"Unhandled state\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3430, __PRETTY_FUNCTION__)); | ||||
3431 | assert(E->getOpcode() && allSameType(VL) && allSameBlock(VL) && "Invalid VL")((E->getOpcode() && allSameType(VL) && allSameBlock (VL) && "Invalid VL") ? static_cast<void> (0) : __assert_fail ("E->getOpcode() && allSameType(VL) && allSameBlock(VL) && \"Invalid VL\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3431, __PRETTY_FUNCTION__)); | ||||
3432 | Instruction *VL0 = E->getMainOp(); | ||||
3433 | unsigned ShuffleOrOp = | ||||
3434 | E->isAltShuffle() ? (unsigned)Instruction::ShuffleVector : E->getOpcode(); | ||||
3435 | switch (ShuffleOrOp) { | ||||
3436 | case Instruction::PHI: | ||||
3437 | return 0; | ||||
3438 | |||||
3439 | case Instruction::ExtractValue: | ||||
3440 | case Instruction::ExtractElement: { | ||||
3441 | if (NeedToShuffleReuses) { | ||||
3442 | unsigned Idx = 0; | ||||
3443 | for (unsigned I : E->ReuseShuffleIndices) { | ||||
3444 | if (ShuffleOrOp == Instruction::ExtractElement) { | ||||
3445 | auto *IO = cast<ConstantInt>( | ||||
3446 | cast<ExtractElementInst>(VL[I])->getIndexOperand()); | ||||
3447 | Idx = IO->getZExtValue(); | ||||
3448 | ReuseShuffleCost -= TTI->getVectorInstrCost( | ||||
3449 | Instruction::ExtractElement, VecTy, Idx); | ||||
3450 | } else { | ||||
3451 | ReuseShuffleCost -= TTI->getVectorInstrCost( | ||||
3452 | Instruction::ExtractElement, VecTy, Idx); | ||||
3453 | ++Idx; | ||||
3454 | } | ||||
3455 | } | ||||
3456 | Idx = ReuseShuffleNumbers; | ||||
3457 | for (Value *V : VL) { | ||||
3458 | if (ShuffleOrOp == Instruction::ExtractElement) { | ||||
3459 | auto *IO = cast<ConstantInt>( | ||||
3460 | cast<ExtractElementInst>(V)->getIndexOperand()); | ||||
3461 | Idx = IO->getZExtValue(); | ||||
3462 | } else { | ||||
3463 | --Idx; | ||||
3464 | } | ||||
3465 | ReuseShuffleCost += | ||||
3466 | TTI->getVectorInstrCost(Instruction::ExtractElement, VecTy, Idx); | ||||
3467 | } | ||||
3468 | } | ||||
3469 | int DeadCost = ReuseShuffleCost; | ||||
3470 | if (!E->ReorderIndices.empty()) { | ||||
3471 | // TODO: Merge this shuffle with the ReuseShuffleCost. | ||||
3472 | DeadCost += TTI->getShuffleCost( | ||||
3473 | TargetTransformInfo::SK_PermuteSingleSrc, VecTy); | ||||
3474 | } | ||||
3475 | for (unsigned I = 0, E = VL.size(); I < E; ++I) { | ||||
3476 | Instruction *EI = cast<Instruction>(VL[I]); | ||||
3477 | // If all users are going to be vectorized, instruction can be | ||||
3478 | // considered as dead. | ||||
3479 | // The same, if have only one user, it will be vectorized for sure. | ||||
3480 | if (areAllUsersVectorized(EI)) { | ||||
3481 | // Take credit for instruction that will become dead. | ||||
3482 | if (EI->hasOneUse()) { | ||||
3483 | Instruction *Ext = EI->user_back(); | ||||
3484 | if ((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) && | ||||
3485 | all_of(Ext->users(), | ||||
3486 | [](User *U) { return isa<GetElementPtrInst>(U); })) { | ||||
3487 | // Use getExtractWithExtendCost() to calculate the cost of | ||||
3488 | // extractelement/ext pair. | ||||
3489 | DeadCost -= TTI->getExtractWithExtendCost( | ||||
3490 | Ext->getOpcode(), Ext->getType(), VecTy, I); | ||||
3491 | // Add back the cost of s|zext which is subtracted separately. | ||||
3492 | DeadCost += TTI->getCastInstrCost( | ||||
3493 | Ext->getOpcode(), Ext->getType(), EI->getType(), | ||||
3494 | TTI::getCastContextHint(Ext), CostKind, Ext); | ||||
3495 | continue; | ||||
3496 | } | ||||
3497 | } | ||||
3498 | DeadCost -= | ||||
3499 | TTI->getVectorInstrCost(Instruction::ExtractElement, VecTy, I); | ||||
3500 | } | ||||
3501 | } | ||||
3502 | return DeadCost; | ||||
3503 | } | ||||
3504 | case Instruction::ZExt: | ||||
3505 | case Instruction::SExt: | ||||
3506 | case Instruction::FPToUI: | ||||
3507 | case Instruction::FPToSI: | ||||
3508 | case Instruction::FPExt: | ||||
3509 | case Instruction::PtrToInt: | ||||
3510 | case Instruction::IntToPtr: | ||||
3511 | case Instruction::SIToFP: | ||||
3512 | case Instruction::UIToFP: | ||||
3513 | case Instruction::Trunc: | ||||
3514 | case Instruction::FPTrunc: | ||||
3515 | case Instruction::BitCast: { | ||||
3516 | Type *SrcTy = VL0->getOperand(0)->getType(); | ||||
3517 | int ScalarEltCost = | ||||
3518 | TTI->getCastInstrCost(E->getOpcode(), ScalarTy, SrcTy, | ||||
3519 | TTI::getCastContextHint(VL0), CostKind, VL0); | ||||
3520 | if (NeedToShuffleReuses) { | ||||
3521 | ReuseShuffleCost -= (ReuseShuffleNumbers - VL.size()) * ScalarEltCost; | ||||
3522 | } | ||||
3523 | |||||
3524 | // Calculate the cost of this instruction. | ||||
3525 | int ScalarCost = VL.size() * ScalarEltCost; | ||||
3526 | |||||
3527 | auto *SrcVecTy = FixedVectorType::get(SrcTy, VL.size()); | ||||
3528 | int VecCost = 0; | ||||
3529 | // Check if the values are candidates to demote. | ||||
3530 | if (!MinBWs.count(VL0) || VecTy != SrcVecTy) { | ||||
3531 | VecCost = | ||||
3532 | ReuseShuffleCost + | ||||
3533 | TTI->getCastInstrCost(E->getOpcode(), VecTy, SrcVecTy, | ||||
3534 | TTI::getCastContextHint(VL0), CostKind, VL0); | ||||
3535 | } | ||||
3536 | return VecCost - ScalarCost; | ||||
3537 | } | ||||
3538 | case Instruction::FCmp: | ||||
3539 | case Instruction::ICmp: | ||||
3540 | case Instruction::Select: { | ||||
3541 | // Calculate the cost of this instruction. | ||||
3542 | int ScalarEltCost = TTI->getCmpSelInstrCost(E->getOpcode(), ScalarTy, | ||||
3543 | Builder.getInt1Ty(), | ||||
3544 | CostKind, VL0); | ||||
3545 | if (NeedToShuffleReuses) { | ||||
3546 | ReuseShuffleCost -= (ReuseShuffleNumbers - VL.size()) * ScalarEltCost; | ||||
3547 | } | ||||
3548 | auto *MaskTy = FixedVectorType::get(Builder.getInt1Ty(), VL.size()); | ||||
3549 | int ScalarCost = VecTy->getNumElements() * ScalarEltCost; | ||||
3550 | int VecCost = TTI->getCmpSelInstrCost(E->getOpcode(), VecTy, MaskTy, | ||||
3551 | CostKind, VL0); | ||||
3552 | return ReuseShuffleCost + VecCost - ScalarCost; | ||||
3553 | } | ||||
3554 | case Instruction::FNeg: | ||||
3555 | case Instruction::Add: | ||||
3556 | case Instruction::FAdd: | ||||
3557 | case Instruction::Sub: | ||||
3558 | case Instruction::FSub: | ||||
3559 | case Instruction::Mul: | ||||
3560 | case Instruction::FMul: | ||||
3561 | case Instruction::UDiv: | ||||
3562 | case Instruction::SDiv: | ||||
3563 | case Instruction::FDiv: | ||||
3564 | case Instruction::URem: | ||||
3565 | case Instruction::SRem: | ||||
3566 | case Instruction::FRem: | ||||
3567 | case Instruction::Shl: | ||||
3568 | case Instruction::LShr: | ||||
3569 | case Instruction::AShr: | ||||
3570 | case Instruction::And: | ||||
3571 | case Instruction::Or: | ||||
3572 | case Instruction::Xor: { | ||||
3573 | // Certain instructions can be cheaper to vectorize if they have a | ||||
3574 | // constant second vector operand. | ||||
3575 | TargetTransformInfo::OperandValueKind Op1VK = | ||||
3576 | TargetTransformInfo::OK_AnyValue; | ||||
3577 | TargetTransformInfo::OperandValueKind Op2VK = | ||||
3578 | TargetTransformInfo::OK_UniformConstantValue; | ||||
3579 | TargetTransformInfo::OperandValueProperties Op1VP = | ||||
3580 | TargetTransformInfo::OP_None; | ||||
3581 | TargetTransformInfo::OperandValueProperties Op2VP = | ||||
3582 | TargetTransformInfo::OP_PowerOf2; | ||||
3583 | |||||
3584 | // If all operands are exactly the same ConstantInt then set the | ||||
3585 | // operand kind to OK_UniformConstantValue. | ||||
3586 | // If instead not all operands are constants, then set the operand kind | ||||
3587 | // to OK_AnyValue. If all operands are constants but not the same, | ||||
3588 | // then set the operand kind to OK_NonUniformConstantValue. | ||||
3589 | ConstantInt *CInt0 = nullptr; | ||||
3590 | for (unsigned i = 0, e = VL.size(); i < e; ++i) { | ||||
3591 | const Instruction *I = cast<Instruction>(VL[i]); | ||||
3592 | unsigned OpIdx = isa<BinaryOperator>(I) ? 1 : 0; | ||||
3593 | ConstantInt *CInt = dyn_cast<ConstantInt>(I->getOperand(OpIdx)); | ||||
3594 | if (!CInt) { | ||||
3595 | Op2VK = TargetTransformInfo::OK_AnyValue; | ||||
3596 | Op2VP = TargetTransformInfo::OP_None; | ||||
3597 | break; | ||||
3598 | } | ||||
3599 | if (Op2VP == TargetTransformInfo::OP_PowerOf2 && | ||||
3600 | !CInt->getValue().isPowerOf2()) | ||||
3601 | Op2VP = TargetTransformInfo::OP_None; | ||||
3602 | if (i == 0) { | ||||
3603 | CInt0 = CInt; | ||||
3604 | continue; | ||||
3605 | } | ||||
3606 | if (CInt0 != CInt) | ||||
3607 | Op2VK = TargetTransformInfo::OK_NonUniformConstantValue; | ||||
3608 | } | ||||
3609 | |||||
3610 | SmallVector<const Value *, 4> Operands(VL0->operand_values()); | ||||
3611 | int ScalarEltCost = TTI->getArithmeticInstrCost( | ||||
3612 | E->getOpcode(), ScalarTy, CostKind, Op1VK, Op2VK, Op1VP, Op2VP, | ||||
3613 | Operands, VL0); | ||||
3614 | if (NeedToShuffleReuses) { | ||||
3615 | ReuseShuffleCost -= (ReuseShuffleNumbers - VL.size()) * ScalarEltCost; | ||||
3616 | } | ||||
3617 | int ScalarCost = VecTy->getNumElements() * ScalarEltCost; | ||||
3618 | int VecCost = TTI->getArithmeticInstrCost( | ||||
3619 | E->getOpcode(), VecTy, CostKind, Op1VK, Op2VK, Op1VP, Op2VP, | ||||
3620 | Operands, VL0); | ||||
3621 | return ReuseShuffleCost + VecCost - ScalarCost; | ||||
3622 | } | ||||
3623 | case Instruction::GetElementPtr: { | ||||
3624 | TargetTransformInfo::OperandValueKind Op1VK = | ||||
3625 | TargetTransformInfo::OK_AnyValue; | ||||
3626 | TargetTransformInfo::OperandValueKind Op2VK = | ||||
3627 | TargetTransformInfo::OK_UniformConstantValue; | ||||
3628 | |||||
3629 | int ScalarEltCost = | ||||
3630 | TTI->getArithmeticInstrCost(Instruction::Add, ScalarTy, CostKind, | ||||
3631 | Op1VK, Op2VK); | ||||
3632 | if (NeedToShuffleReuses) { | ||||
3633 | ReuseShuffleCost -= (ReuseShuffleNumbers - VL.size()) * ScalarEltCost; | ||||
3634 | } | ||||
3635 | int ScalarCost = VecTy->getNumElements() * ScalarEltCost; | ||||
3636 | int VecCost = | ||||
3637 | TTI->getArithmeticInstrCost(Instruction::Add, VecTy, CostKind, | ||||
3638 | Op1VK, Op2VK); | ||||
3639 | return ReuseShuffleCost + VecCost - ScalarCost; | ||||
3640 | } | ||||
3641 | case Instruction::Load: { | ||||
3642 | // Cost of wide load - cost of scalar loads. | ||||
3643 | Align alignment = cast<LoadInst>(VL0)->getAlign(); | ||||
3644 | int ScalarEltCost = | ||||
3645 | TTI->getMemoryOpCost(Instruction::Load, ScalarTy, alignment, 0, | ||||
3646 | CostKind, VL0); | ||||
3647 | if (NeedToShuffleReuses) { | ||||
3648 | ReuseShuffleCost -= (ReuseShuffleNumbers - VL.size()) * ScalarEltCost; | ||||
3649 | } | ||||
3650 | int ScalarLdCost = VecTy->getNumElements() * ScalarEltCost; | ||||
3651 | int VecLdCost = | ||||
3652 | TTI->getMemoryOpCost(Instruction::Load, VecTy, alignment, 0, | ||||
3653 | CostKind, VL0); | ||||
3654 | if (!E->ReorderIndices.empty()) { | ||||
3655 | // TODO: Merge this shuffle with the ReuseShuffleCost. | ||||
3656 | VecLdCost += TTI->getShuffleCost( | ||||
3657 | TargetTransformInfo::SK_PermuteSingleSrc, VecTy); | ||||
3658 | } | ||||
3659 | return ReuseShuffleCost + VecLdCost - ScalarLdCost; | ||||
3660 | } | ||||
3661 | case Instruction::Store: { | ||||
3662 | // We know that we can merge the stores. Calculate the cost. | ||||
3663 | bool IsReorder = !E->ReorderIndices.empty(); | ||||
3664 | auto *SI = | ||||
3665 | cast<StoreInst>(IsReorder ? VL[E->ReorderIndices.front()] : VL0); | ||||
3666 | Align Alignment = SI->getAlign(); | ||||
3667 | int ScalarEltCost = | ||||
3668 | TTI->getMemoryOpCost(Instruction::Store, ScalarTy, Alignment, 0, | ||||
3669 | CostKind, VL0); | ||||
3670 | if (NeedToShuffleReuses) | ||||
3671 | ReuseShuffleCost = -(ReuseShuffleNumbers - VL.size()) * ScalarEltCost; | ||||
3672 | int ScalarStCost = VecTy->getNumElements() * ScalarEltCost; | ||||
3673 | int VecStCost = TTI->getMemoryOpCost(Instruction::Store, | ||||
3674 | VecTy, Alignment, 0, CostKind, VL0); | ||||
3675 | if (IsReorder) { | ||||
3676 | // TODO: Merge this shuffle with the ReuseShuffleCost. | ||||
3677 | VecStCost += TTI->getShuffleCost( | ||||
3678 | TargetTransformInfo::SK_PermuteSingleSrc, VecTy); | ||||
3679 | } | ||||
3680 | return ReuseShuffleCost + VecStCost - ScalarStCost; | ||||
3681 | } | ||||
3682 | case Instruction::Call: { | ||||
3683 | CallInst *CI = cast<CallInst>(VL0); | ||||
3684 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); | ||||
3685 | |||||
3686 | // Calculate the cost of the scalar and vector calls. | ||||
3687 | IntrinsicCostAttributes CostAttrs(ID, *CI, 1, 1); | ||||
3688 | int ScalarEltCost = TTI->getIntrinsicInstrCost(CostAttrs, CostKind); | ||||
3689 | if (NeedToShuffleReuses) { | ||||
3690 | ReuseShuffleCost -= (ReuseShuffleNumbers - VL.size()) * ScalarEltCost; | ||||
3691 | } | ||||
3692 | int ScalarCallCost = VecTy->getNumElements() * ScalarEltCost; | ||||
3693 | |||||
3694 | auto VecCallCosts = getVectorCallCosts(CI, VecTy, TTI, TLI); | ||||
3695 | int VecCallCost = std::min(VecCallCosts.first, VecCallCosts.second); | ||||
3696 | |||||
3697 | 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) | ||||
3698 | << " (" << VecCallCost << "-" << ScalarCallCost << ")"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Call cost " << VecCallCost - ScalarCallCost << " (" << VecCallCost << "-" << ScalarCallCost << ")" << " for " << *CI << "\n"; } } while (false) | ||||
3699 | << " for " << *CI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Call cost " << VecCallCost - ScalarCallCost << " (" << VecCallCost << "-" << ScalarCallCost << ")" << " for " << *CI << "\n"; } } while (false); | ||||
3700 | |||||
3701 | return ReuseShuffleCost + VecCallCost - ScalarCallCost; | ||||
3702 | } | ||||
3703 | case Instruction::ShuffleVector: { | ||||
3704 | assert(E->isAltShuffle() &&((E->isAltShuffle() && ((Instruction::isBinaryOp(E ->getOpcode()) && Instruction::isBinaryOp(E->getAltOpcode ())) || (Instruction::isCast(E->getOpcode()) && Instruction ::isCast(E->getAltOpcode()))) && "Invalid Shuffle Vector Operand" ) ? static_cast<void> (0) : __assert_fail ("E->isAltShuffle() && ((Instruction::isBinaryOp(E->getOpcode()) && Instruction::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E->getOpcode()) && Instruction::isCast(E->getAltOpcode()))) && \"Invalid Shuffle Vector Operand\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3709, __PRETTY_FUNCTION__)) | ||||
3705 | ((Instruction::isBinaryOp(E->getOpcode()) &&((E->isAltShuffle() && ((Instruction::isBinaryOp(E ->getOpcode()) && Instruction::isBinaryOp(E->getAltOpcode ())) || (Instruction::isCast(E->getOpcode()) && Instruction ::isCast(E->getAltOpcode()))) && "Invalid Shuffle Vector Operand" ) ? static_cast<void> (0) : __assert_fail ("E->isAltShuffle() && ((Instruction::isBinaryOp(E->getOpcode()) && Instruction::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E->getOpcode()) && Instruction::isCast(E->getAltOpcode()))) && \"Invalid Shuffle Vector Operand\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3709, __PRETTY_FUNCTION__)) | ||||
3706 | Instruction::isBinaryOp(E->getAltOpcode())) ||((E->isAltShuffle() && ((Instruction::isBinaryOp(E ->getOpcode()) && Instruction::isBinaryOp(E->getAltOpcode ())) || (Instruction::isCast(E->getOpcode()) && Instruction ::isCast(E->getAltOpcode()))) && "Invalid Shuffle Vector Operand" ) ? static_cast<void> (0) : __assert_fail ("E->isAltShuffle() && ((Instruction::isBinaryOp(E->getOpcode()) && Instruction::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E->getOpcode()) && Instruction::isCast(E->getAltOpcode()))) && \"Invalid Shuffle Vector Operand\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3709, __PRETTY_FUNCTION__)) | ||||
3707 | (Instruction::isCast(E->getOpcode()) &&((E->isAltShuffle() && ((Instruction::isBinaryOp(E ->getOpcode()) && Instruction::isBinaryOp(E->getAltOpcode ())) || (Instruction::isCast(E->getOpcode()) && Instruction ::isCast(E->getAltOpcode()))) && "Invalid Shuffle Vector Operand" ) ? static_cast<void> (0) : __assert_fail ("E->isAltShuffle() && ((Instruction::isBinaryOp(E->getOpcode()) && Instruction::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E->getOpcode()) && Instruction::isCast(E->getAltOpcode()))) && \"Invalid Shuffle Vector Operand\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3709, __PRETTY_FUNCTION__)) | ||||
3708 | Instruction::isCast(E->getAltOpcode()))) &&((E->isAltShuffle() && ((Instruction::isBinaryOp(E ->getOpcode()) && Instruction::isBinaryOp(E->getAltOpcode ())) || (Instruction::isCast(E->getOpcode()) && Instruction ::isCast(E->getAltOpcode()))) && "Invalid Shuffle Vector Operand" ) ? static_cast<void> (0) : __assert_fail ("E->isAltShuffle() && ((Instruction::isBinaryOp(E->getOpcode()) && Instruction::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E->getOpcode()) && Instruction::isCast(E->getAltOpcode()))) && \"Invalid Shuffle Vector Operand\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3709, __PRETTY_FUNCTION__)) | ||||
3709 | "Invalid Shuffle Vector Operand")((E->isAltShuffle() && ((Instruction::isBinaryOp(E ->getOpcode()) && Instruction::isBinaryOp(E->getAltOpcode ())) || (Instruction::isCast(E->getOpcode()) && Instruction ::isCast(E->getAltOpcode()))) && "Invalid Shuffle Vector Operand" ) ? static_cast<void> (0) : __assert_fail ("E->isAltShuffle() && ((Instruction::isBinaryOp(E->getOpcode()) && Instruction::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E->getOpcode()) && Instruction::isCast(E->getAltOpcode()))) && \"Invalid Shuffle Vector Operand\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3709, __PRETTY_FUNCTION__)); | ||||
3710 | int ScalarCost = 0; | ||||
3711 | if (NeedToShuffleReuses) { | ||||
3712 | for (unsigned Idx : E->ReuseShuffleIndices) { | ||||
3713 | Instruction *I = cast<Instruction>(VL[Idx]); | ||||
3714 | ReuseShuffleCost -= TTI->getInstructionCost(I, CostKind); | ||||
3715 | } | ||||
3716 | for (Value *V : VL) { | ||||
3717 | Instruction *I = cast<Instruction>(V); | ||||
3718 | ReuseShuffleCost += TTI->getInstructionCost(I, CostKind); | ||||
3719 | } | ||||
3720 | } | ||||
3721 | for (Value *V : VL) { | ||||
3722 | Instruction *I = cast<Instruction>(V); | ||||
3723 | assert(E->isOpcodeOrAlt(I) && "Unexpected main/alternate opcode")((E->isOpcodeOrAlt(I) && "Unexpected main/alternate opcode" ) ? static_cast<void> (0) : __assert_fail ("E->isOpcodeOrAlt(I) && \"Unexpected main/alternate opcode\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3723, __PRETTY_FUNCTION__)); | ||||
3724 | ScalarCost += TTI->getInstructionCost(I, CostKind); | ||||
3725 | } | ||||
3726 | // VecCost is equal to sum of the cost of creating 2 vectors | ||||
3727 | // and the cost of creating shuffle. | ||||
3728 | int VecCost = 0; | ||||
3729 | if (Instruction::isBinaryOp(E->getOpcode())) { | ||||
3730 | VecCost = TTI->getArithmeticInstrCost(E->getOpcode(), VecTy, CostKind); | ||||
3731 | VecCost += TTI->getArithmeticInstrCost(E->getAltOpcode(), VecTy, | ||||
3732 | CostKind); | ||||
3733 | } else { | ||||
3734 | Type *Src0SclTy = E->getMainOp()->getOperand(0)->getType(); | ||||
3735 | Type *Src1SclTy = E->getAltOp()->getOperand(0)->getType(); | ||||
3736 | auto *Src0Ty = FixedVectorType::get(Src0SclTy, VL.size()); | ||||
3737 | auto *Src1Ty = FixedVectorType::get(Src1SclTy, VL.size()); | ||||
3738 | VecCost = TTI->getCastInstrCost(E->getOpcode(), VecTy, Src0Ty, | ||||
3739 | TTI::CastContextHint::None, CostKind); | ||||
3740 | VecCost += TTI->getCastInstrCost(E->getAltOpcode(), VecTy, Src1Ty, | ||||
3741 | TTI::CastContextHint::None, CostKind); | ||||
3742 | } | ||||
3743 | VecCost += TTI->getShuffleCost(TargetTransformInfo::SK_Select, VecTy, 0); | ||||
3744 | return ReuseShuffleCost + VecCost - ScalarCost; | ||||
3745 | } | ||||
3746 | default: | ||||
3747 | llvm_unreachable("Unknown instruction")::llvm::llvm_unreachable_internal("Unknown instruction", "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3747); | ||||
3748 | } | ||||
3749 | } | ||||
3750 | |||||
3751 | bool BoUpSLP::isFullyVectorizableTinyTree() const { | ||||
3752 | 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) | ||||
3753 | << 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); | ||||
3754 | |||||
3755 | // We only handle trees of heights 1 and 2. | ||||
3756 | if (VectorizableTree.size() == 1 && | ||||
3757 | VectorizableTree[0]->State == TreeEntry::Vectorize) | ||||
3758 | return true; | ||||
3759 | |||||
3760 | if (VectorizableTree.size() != 2) | ||||
3761 | return false; | ||||
3762 | |||||
3763 | // Handle splat and all-constants stores. | ||||
3764 | if (VectorizableTree[0]->State == TreeEntry::Vectorize && | ||||
3765 | (allConstant(VectorizableTree[1]->Scalars) || | ||||
3766 | isSplat(VectorizableTree[1]->Scalars))) | ||||
3767 | return true; | ||||
3768 | |||||
3769 | // Gathering cost would be too much for tiny trees. | ||||
3770 | if (VectorizableTree[0]->State == TreeEntry::NeedToGather || | ||||
3771 | VectorizableTree[1]->State == TreeEntry::NeedToGather) | ||||
3772 | return false; | ||||
3773 | |||||
3774 | return true; | ||||
3775 | } | ||||
3776 | |||||
3777 | static bool isLoadCombineCandidateImpl(Value *Root, unsigned NumElts, | ||||
3778 | TargetTransformInfo *TTI) { | ||||
3779 | // Look past the root to find a source value. Arbitrarily follow the | ||||
3780 | // path through operand 0 of any 'or'. Also, peek through optional | ||||
3781 | // shift-left-by-multiple-of-8-bits. | ||||
3782 | Value *ZextLoad = Root; | ||||
3783 | const APInt *ShAmtC; | ||||
3784 | while (!isa<ConstantExpr>(ZextLoad) && | ||||
3785 | (match(ZextLoad, m_Or(m_Value(), m_Value())) || | ||||
3786 | (match(ZextLoad, m_Shl(m_Value(), m_APInt(ShAmtC))) && | ||||
3787 | ShAmtC->urem(8) == 0))) | ||||
3788 | ZextLoad = cast<BinaryOperator>(ZextLoad)->getOperand(0); | ||||
3789 | |||||
3790 | // Check if the input is an extended load of the required or/shift expression. | ||||
3791 | Value *LoadPtr; | ||||
3792 | if (ZextLoad == Root || !match(ZextLoad, m_ZExt(m_Load(m_Value(LoadPtr))))) | ||||
3793 | return false; | ||||
3794 | |||||
3795 | // Require that the total load bit width is a legal integer type. | ||||
3796 | // For example, <8 x i8> --> i64 is a legal integer on a 64-bit target. | ||||
3797 | // But <16 x i8> --> i128 is not, so the backend probably can't reduce it. | ||||
3798 | Type *SrcTy = LoadPtr->getType()->getPointerElementType(); | ||||
3799 | unsigned LoadBitWidth = SrcTy->getIntegerBitWidth() * NumElts; | ||||
3800 | if (!TTI->isTypeLegal(IntegerType::get(Root->getContext(), LoadBitWidth))) | ||||
3801 | return false; | ||||
3802 | |||||
3803 | // Everything matched - assume that we can fold the whole sequence using | ||||
3804 | // load combining. | ||||
3805 | 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) | ||||
3806 | << *(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); | ||||
3807 | |||||
3808 | return true; | ||||
3809 | } | ||||
3810 | |||||
3811 | bool BoUpSLP::isLoadCombineReductionCandidate(unsigned RdxOpcode) const { | ||||
3812 | if (RdxOpcode != Instruction::Or) | ||||
3813 | return false; | ||||
3814 | |||||
3815 | unsigned NumElts = VectorizableTree[0]->Scalars.size(); | ||||
3816 | Value *FirstReduced = VectorizableTree[0]->Scalars[0]; | ||||
3817 | return isLoadCombineCandidateImpl(FirstReduced, NumElts, TTI); | ||||
3818 | } | ||||
3819 | |||||
3820 | bool BoUpSLP::isLoadCombineCandidate() const { | ||||
3821 | // Peek through a final sequence of stores and check if all operations are | ||||
3822 | // likely to be load-combined. | ||||
3823 | unsigned NumElts = VectorizableTree[0]->Scalars.size(); | ||||
3824 | for (Value *Scalar : VectorizableTree[0]->Scalars) { | ||||
3825 | Value *X; | ||||
3826 | if (!match(Scalar, m_Store(m_Value(X), m_Value())) || | ||||
3827 | !isLoadCombineCandidateImpl(X, NumElts, TTI)) | ||||
3828 | return false; | ||||
3829 | } | ||||
3830 | return true; | ||||
3831 | } | ||||
3832 | |||||
3833 | bool BoUpSLP::isTreeTinyAndNotFullyVectorizable() const { | ||||
3834 | // We can vectorize the tree if its size is greater than or equal to the | ||||
3835 | // minimum size specified by the MinTreeSize command line option. | ||||
3836 | if (VectorizableTree.size() >= MinTreeSize) | ||||
3837 | return false; | ||||
3838 | |||||
3839 | // If we have a tiny tree (a tree whose size is less than MinTreeSize), we | ||||
3840 | // can vectorize it if we can prove it fully vectorizable. | ||||
3841 | if (isFullyVectorizableTinyTree()) | ||||
3842 | return false; | ||||
3843 | |||||
3844 | assert(VectorizableTree.empty()((VectorizableTree.empty() ? ExternalUses.empty() : true && "We shouldn't have any external users") ? static_cast<void > (0) : __assert_fail ("VectorizableTree.empty() ? ExternalUses.empty() : true && \"We shouldn't have any external users\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3846, __PRETTY_FUNCTION__)) | ||||
3845 | ? ExternalUses.empty()((VectorizableTree.empty() ? ExternalUses.empty() : true && "We shouldn't have any external users") ? static_cast<void > (0) : __assert_fail ("VectorizableTree.empty() ? ExternalUses.empty() : true && \"We shouldn't have any external users\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3846, __PRETTY_FUNCTION__)) | ||||
3846 | : true && "We shouldn't have any external users")((VectorizableTree.empty() ? ExternalUses.empty() : true && "We shouldn't have any external users") ? static_cast<void > (0) : __assert_fail ("VectorizableTree.empty() ? ExternalUses.empty() : true && \"We shouldn't have any external users\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3846, __PRETTY_FUNCTION__)); | ||||
3847 | |||||
3848 | // Otherwise, we can't vectorize the tree. It is both tiny and not fully | ||||
3849 | // vectorizable. | ||||
3850 | return true; | ||||
3851 | } | ||||
3852 | |||||
3853 | int BoUpSLP::getSpillCost() const { | ||||
3854 | // Walk from the bottom of the tree to the top, tracking which values are | ||||
3855 | // live. When we see a call instruction that is not part of our tree, | ||||
3856 | // query TTI to see if there is a cost to keeping values live over it | ||||
3857 | // (for example, if spills and fills are required). | ||||
3858 | unsigned BundleWidth = VectorizableTree.front()->Scalars.size(); | ||||
3859 | int Cost = 0; | ||||
3860 | |||||
3861 | SmallPtrSet<Instruction*, 4> LiveValues; | ||||
3862 | Instruction *PrevInst = nullptr; | ||||
3863 | |||||
3864 | // The entries in VectorizableTree are not necessarily ordered by their | ||||
3865 | // position in basic blocks. Collect them and order them by dominance so later | ||||
3866 | // instructions are guaranteed to be visited first. For instructions in | ||||
3867 | // different basic blocks, we only scan to the beginning of the block, so | ||||
3868 | // their order does not matter, as long as all instructions in a basic block | ||||
3869 | // are grouped together. Using dominance ensures a deterministic order. | ||||
3870 | SmallVector<Instruction *, 16> OrderedScalars; | ||||
3871 | for (const auto &TEPtr : VectorizableTree) { | ||||
3872 | Instruction *Inst = dyn_cast<Instruction>(TEPtr->Scalars[0]); | ||||
3873 | if (!Inst) | ||||
3874 | continue; | ||||
3875 | OrderedScalars.push_back(Inst); | ||||
3876 | } | ||||
3877 | llvm::stable_sort(OrderedScalars, [this](Instruction *A, Instruction *B) { | ||||
3878 | return DT->dominates(B, A); | ||||
3879 | }); | ||||
3880 | |||||
3881 | for (Instruction *Inst : OrderedScalars) { | ||||
3882 | if (!PrevInst) { | ||||
3883 | PrevInst = Inst; | ||||
3884 | continue; | ||||
3885 | } | ||||
3886 | |||||
3887 | // Update LiveValues. | ||||
3888 | LiveValues.erase(PrevInst); | ||||
3889 | for (auto &J : PrevInst->operands()) { | ||||
3890 | if (isa<Instruction>(&*J) && getTreeEntry(&*J)) | ||||
3891 | LiveValues.insert(cast<Instruction>(&*J)); | ||||
3892 | } | ||||
3893 | |||||
3894 | 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) | ||||
3895 | 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) | ||||
3896 | 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) | ||||
3897 | 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) | ||||
3898 | 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) | ||||
3899 | 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) | ||||
3900 | })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); | ||||
3901 | |||||
3902 | // Now find the sequence of instructions between PrevInst and Inst. | ||||
3903 | unsigned NumCalls = 0; | ||||
3904 | BasicBlock::reverse_iterator InstIt = ++Inst->getIterator().getReverse(), | ||||
3905 | PrevInstIt = | ||||
3906 | PrevInst->getIterator().getReverse(); | ||||
3907 | while (InstIt != PrevInstIt) { | ||||
3908 | if (PrevInstIt == PrevInst->getParent()->rend()) { | ||||
3909 | PrevInstIt = Inst->getParent()->rbegin(); | ||||
3910 | continue; | ||||
3911 | } | ||||
3912 | |||||
3913 | // Debug information does not impact spill cost. | ||||
3914 | if ((isa<CallInst>(&*PrevInstIt) && | ||||
3915 | !isa<DbgInfoIntrinsic>(&*PrevInstIt)) && | ||||
3916 | &*PrevInstIt != PrevInst) | ||||
3917 | NumCalls++; | ||||
3918 | |||||
3919 | ++PrevInstIt; | ||||
3920 | } | ||||
3921 | |||||
3922 | if (NumCalls) { | ||||
3923 | SmallVector<Type*, 4> V; | ||||
3924 | for (auto *II : LiveValues) | ||||
3925 | V.push_back(FixedVectorType::get(II->getType(), BundleWidth)); | ||||
3926 | Cost += NumCalls * TTI->getCostOfKeepingLiveOverCall(V); | ||||
3927 | } | ||||
3928 | |||||
3929 | PrevInst = Inst; | ||||
3930 | } | ||||
3931 | |||||
3932 | return Cost; | ||||
3933 | } | ||||
3934 | |||||
3935 | int BoUpSLP::getTreeCost() { | ||||
3936 | int Cost = 0; | ||||
3937 | 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) | ||||
3938 | << VectorizableTree.size() << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Calculating cost for tree of size " << VectorizableTree.size() << ".\n"; } } while ( false); | ||||
3939 | |||||
3940 | unsigned BundleWidth = VectorizableTree[0]->Scalars.size(); | ||||
3941 | |||||
3942 | for (unsigned I = 0, E = VectorizableTree.size(); I < E; ++I) { | ||||
3943 | TreeEntry &TE = *VectorizableTree[I].get(); | ||||
3944 | |||||
3945 | // We create duplicate tree entries for gather sequences that have multiple | ||||
3946 | // uses. However, we should not compute the cost of duplicate sequences. | ||||
3947 | // For example, if we have a build vector (i.e., insertelement sequence) | ||||
3948 | // that is used by more than one vector instruction, we only need to | ||||
3949 | // compute the cost of the insertelement instructions once. The redundant | ||||
3950 | // instructions will be eliminated by CSE. | ||||
3951 | // | ||||
3952 | // We should consider not creating duplicate tree entries for gather | ||||
3953 | // sequences, and instead add additional edges to the tree representing | ||||
3954 | // their uses. Since such an approach results in fewer total entries, | ||||
3955 | // existing heuristics based on tree size may yield different results. | ||||
3956 | // | ||||
3957 | if (TE.State == TreeEntry::NeedToGather && | ||||
3958 | std::any_of(std::next(VectorizableTree.begin(), I + 1), | ||||
3959 | VectorizableTree.end(), | ||||
3960 | [TE](const std::unique_ptr<TreeEntry> &EntryPtr) { | ||||
3961 | return EntryPtr->State == TreeEntry::NeedToGather && | ||||
3962 | EntryPtr->isSame(TE.Scalars); | ||||
3963 | })) | ||||
3964 | continue; | ||||
3965 | |||||
3966 | int C = getEntryCost(&TE); | ||||
3967 | 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"; } } while (false) | ||||
3968 | << " 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"; } } while (false) | ||||
3969 | << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Adding cost " << C << " for bundle that starts with " << *TE.Scalars[0] << ".\n"; } } while (false); | ||||
3970 | Cost += C; | ||||
3971 | } | ||||
3972 | |||||
3973 | SmallPtrSet<Value *, 16> ExtractCostCalculated; | ||||
3974 | int ExtractCost = 0; | ||||
3975 | for (ExternalUser &EU : ExternalUses) { | ||||
3976 | // We only add extract cost once for the same scalar. | ||||
3977 | if (!ExtractCostCalculated.insert(EU.Scalar).second) | ||||
3978 | continue; | ||||
3979 | |||||
3980 | // Uses by ephemeral values are free (because the ephemeral value will be | ||||
3981 | // removed prior to code generation, and so the extraction will be | ||||
3982 | // removed as well). | ||||
3983 | if (EphValues.count(EU.User)) | ||||
3984 | continue; | ||||
3985 | |||||
3986 | // If we plan to rewrite the tree in a smaller type, we will need to sign | ||||
3987 | // extend the extracted value back to the original type. Here, we account | ||||
3988 | // for the extract and the added cost of the sign extend if needed. | ||||
3989 | auto *VecTy = FixedVectorType::get(EU.Scalar->getType(), BundleWidth); | ||||
3990 | auto *ScalarRoot = VectorizableTree[0]->Scalars[0]; | ||||
3991 | if (MinBWs.count(ScalarRoot)) { | ||||
3992 | auto *MinTy = IntegerType::get(F->getContext(), MinBWs[ScalarRoot].first); | ||||
3993 | auto Extend = | ||||
3994 | MinBWs[ScalarRoot].second ? Instruction::SExt : Instruction::ZExt; | ||||
3995 | VecTy = FixedVectorType::get(MinTy, BundleWidth); | ||||
3996 | ExtractCost += TTI->getExtractWithExtendCost(Extend, EU.Scalar->getType(), | ||||
3997 | VecTy, EU.Lane); | ||||
3998 | } else { | ||||
3999 | ExtractCost += | ||||
4000 | TTI->getVectorInstrCost(Instruction::ExtractElement, VecTy, EU.Lane); | ||||
4001 | } | ||||
4002 | } | ||||
4003 | |||||
4004 | int SpillCost = getSpillCost(); | ||||
4005 | Cost += SpillCost + ExtractCost; | ||||
4006 | |||||
4007 | #ifndef NDEBUG | ||||
4008 | SmallString<256> Str; | ||||
4009 | { | ||||
4010 | raw_svector_ostream OS(Str); | ||||
4011 | OS << "SLP: Spill Cost = " << SpillCost << ".\n" | ||||
4012 | << "SLP: Extract Cost = " << ExtractCost << ".\n" | ||||
4013 | << "SLP: Total Cost = " << Cost << ".\n"; | ||||
4014 | } | ||||
4015 | LLVM_DEBUG(dbgs() << Str)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << Str; } } while (false); | ||||
4016 | if (ViewSLPTree) | ||||
4017 | ViewGraph(this, "SLP" + F->getName(), false, Str); | ||||
4018 | #endif | ||||
4019 | |||||
4020 | return Cost; | ||||
4021 | } | ||||
4022 | |||||
4023 | int BoUpSLP::getGatherCost(FixedVectorType *Ty, | ||||
4024 | const DenseSet<unsigned> &ShuffledIndices) const { | ||||
4025 | unsigned NumElts = Ty->getNumElements(); | ||||
4026 | APInt DemandedElts = APInt::getNullValue(NumElts); | ||||
4027 | for (unsigned I = 0; I < NumElts; ++I) | ||||
4028 | if (!ShuffledIndices.count(I)) | ||||
4029 | DemandedElts.setBit(I); | ||||
4030 | int Cost = TTI->getScalarizationOverhead(Ty, DemandedElts, /*Insert*/ true, | ||||
4031 | /*Extract*/ false); | ||||
4032 | if (!ShuffledIndices.empty()) | ||||
4033 | Cost += TTI->getShuffleCost(TargetTransformInfo::SK_PermuteSingleSrc, Ty); | ||||
4034 | return Cost; | ||||
4035 | } | ||||
4036 | |||||
4037 | int BoUpSLP::getGatherCost(ArrayRef<Value *> VL) const { | ||||
4038 | // Find the type of the operands in VL. | ||||
4039 | Type *ScalarTy = VL[0]->getType(); | ||||
4040 | if (StoreInst *SI = dyn_cast<StoreInst>(VL[0])) | ||||
4041 | ScalarTy = SI->getValueOperand()->getType(); | ||||
4042 | auto *VecTy = FixedVectorType::get(ScalarTy, VL.size()); | ||||
4043 | // Find the cost of inserting/extracting values from the vector. | ||||
4044 | // Check if the same elements are inserted several times and count them as | ||||
4045 | // shuffle candidates. | ||||
4046 | DenseSet<unsigned> ShuffledElements; | ||||
4047 | DenseSet<Value *> UniqueElements; | ||||
4048 | // Iterate in reverse order to consider insert elements with the high cost. | ||||
4049 | for (unsigned I = VL.size(); I > 0; --I) { | ||||
4050 | unsigned Idx = I - 1; | ||||
4051 | if (!UniqueElements.insert(VL[Idx]).second) | ||||
4052 | ShuffledElements.insert(Idx); | ||||
4053 | } | ||||
4054 | return getGatherCost(VecTy, ShuffledElements); | ||||
4055 | } | ||||
4056 | |||||
4057 | // Perform operand reordering on the instructions in VL and return the reordered | ||||
4058 | // operands in Left and Right. | ||||
4059 | void BoUpSLP::reorderInputsAccordingToOpcode(ArrayRef<Value *> VL, | ||||
4060 | SmallVectorImpl<Value *> &Left, | ||||
4061 | SmallVectorImpl<Value *> &Right, | ||||
4062 | const DataLayout &DL, | ||||
4063 | ScalarEvolution &SE, | ||||
4064 | const BoUpSLP &R) { | ||||
4065 | if (VL.empty()) | ||||
4066 | return; | ||||
4067 | VLOperands Ops(VL, DL, SE, R); | ||||
4068 | // Reorder the operands in place. | ||||
4069 | Ops.reorder(); | ||||
4070 | Left = Ops.getVL(0); | ||||
4071 | Right = Ops.getVL(1); | ||||
4072 | } | ||||
4073 | |||||
4074 | void BoUpSLP::setInsertPointAfterBundle(TreeEntry *E) { | ||||
4075 | // Get the basic block this bundle is in. All instructions in the bundle | ||||
4076 | // should be in this block. | ||||
4077 | auto *Front = E->getMainOp(); | ||||
4078 | auto *BB = Front->getParent(); | ||||
4079 | assert(llvm::all_of(make_range(E->Scalars.begin(), E->Scalars.end()),((llvm::all_of(make_range(E->Scalars.begin(), E->Scalars .end()), [=](Value *V) -> bool { auto *I = cast<Instruction >(V); return !E->isOpcodeOrAlt(I) || I->getParent() == BB; })) ? static_cast<void> (0) : __assert_fail ("llvm::all_of(make_range(E->Scalars.begin(), E->Scalars.end()), [=](Value *V) -> bool { auto *I = cast<Instruction>(V); return !E->isOpcodeOrAlt(I) || I->getParent() == BB; })" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4083, __PRETTY_FUNCTION__)) | ||||
4080 | [=](Value *V) -> bool {((llvm::all_of(make_range(E->Scalars.begin(), E->Scalars .end()), [=](Value *V) -> bool { auto *I = cast<Instruction >(V); return !E->isOpcodeOrAlt(I) || I->getParent() == BB; })) ? static_cast<void> (0) : __assert_fail ("llvm::all_of(make_range(E->Scalars.begin(), E->Scalars.end()), [=](Value *V) -> bool { auto *I = cast<Instruction>(V); return !E->isOpcodeOrAlt(I) || I->getParent() == BB; })" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4083, __PRETTY_FUNCTION__)) | ||||
4081 | auto *I = cast<Instruction>(V);((llvm::all_of(make_range(E->Scalars.begin(), E->Scalars .end()), [=](Value *V) -> bool { auto *I = cast<Instruction >(V); return !E->isOpcodeOrAlt(I) || I->getParent() == BB; })) ? static_cast<void> (0) : __assert_fail ("llvm::all_of(make_range(E->Scalars.begin(), E->Scalars.end()), [=](Value *V) -> bool { auto *I = cast<Instruction>(V); return !E->isOpcodeOrAlt(I) || I->getParent() == BB; })" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4083, __PRETTY_FUNCTION__)) | ||||
4082 | return !E->isOpcodeOrAlt(I) || I->getParent() == BB;((llvm::all_of(make_range(E->Scalars.begin(), E->Scalars .end()), [=](Value *V) -> bool { auto *I = cast<Instruction >(V); return !E->isOpcodeOrAlt(I) || I->getParent() == BB; })) ? static_cast<void> (0) : __assert_fail ("llvm::all_of(make_range(E->Scalars.begin(), E->Scalars.end()), [=](Value *V) -> bool { auto *I = cast<Instruction>(V); return !E->isOpcodeOrAlt(I) || I->getParent() == BB; })" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4083, __PRETTY_FUNCTION__)) | ||||
4083 | }))((llvm::all_of(make_range(E->Scalars.begin(), E->Scalars .end()), [=](Value *V) -> bool { auto *I = cast<Instruction >(V); return !E->isOpcodeOrAlt(I) || I->getParent() == BB; })) ? static_cast<void> (0) : __assert_fail ("llvm::all_of(make_range(E->Scalars.begin(), E->Scalars.end()), [=](Value *V) -> bool { auto *I = cast<Instruction>(V); return !E->isOpcodeOrAlt(I) || I->getParent() == BB; })" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4083, __PRETTY_FUNCTION__)); | ||||
4084 | |||||
4085 | // The last instruction in the bundle in program order. | ||||
4086 | Instruction *LastInst = nullptr; | ||||
4087 | |||||
4088 | // Find the last instruction. The common case should be that BB has been | ||||
4089 | // scheduled, and the last instruction is VL.back(). So we start with | ||||
4090 | // VL.back() and iterate over schedule data until we reach the end of the | ||||
4091 | // bundle. The end of the bundle is marked by null ScheduleData. | ||||
4092 | if (BlocksSchedules.count(BB)) { | ||||
4093 | auto *Bundle = | ||||
4094 | BlocksSchedules[BB]->getScheduleData(E->isOneOf(E->Scalars.back())); | ||||
4095 | if (Bundle && Bundle->isPartOfBundle()) | ||||
4096 | for (; Bundle; Bundle = Bundle->NextInBundle) | ||||
4097 | if (Bundle->OpValue == Bundle->Inst) | ||||
4098 | LastInst = Bundle->Inst; | ||||
4099 | } | ||||
4100 | |||||
4101 | // LastInst can still be null at this point if there's either not an entry | ||||
4102 | // for BB in BlocksSchedules or there's no ScheduleData available for | ||||
4103 | // VL.back(). This can be the case if buildTree_rec aborts for various | ||||
4104 | // reasons (e.g., the maximum recursion depth is reached, the maximum region | ||||
4105 | // size is reached, etc.). ScheduleData is initialized in the scheduling | ||||
4106 | // "dry-run". | ||||
4107 | // | ||||
4108 | // If this happens, we can still find the last instruction by brute force. We | ||||
4109 | // iterate forwards from Front (inclusive) until we either see all | ||||
4110 | // instructions in the bundle or reach the end of the block. If Front is the | ||||
4111 | // last instruction in program order, LastInst will be set to Front, and we | ||||
4112 | // will visit all the remaining instructions in the block. | ||||
4113 | // | ||||
4114 | // One of the reasons we exit early from buildTree_rec is to place an upper | ||||
4115 | // bound on compile-time. Thus, taking an additional compile-time hit here is | ||||
4116 | // not ideal. However, this should be exceedingly rare since it requires that | ||||
4117 | // we both exit early from buildTree_rec and that the bundle be out-of-order | ||||
4118 | // (causing us to iterate all the way to the end of the block). | ||||
4119 | if (!LastInst) { | ||||
4120 | SmallPtrSet<Value *, 16> Bundle(E->Scalars.begin(), E->Scalars.end()); | ||||
4121 | for (auto &I : make_range(BasicBlock::iterator(Front), BB->end())) { | ||||
4122 | if (Bundle.erase(&I) && E->isOpcodeOrAlt(&I)) | ||||
4123 | LastInst = &I; | ||||
4124 | if (Bundle.empty()) | ||||
4125 | break; | ||||
4126 | } | ||||
4127 | } | ||||
4128 | assert(LastInst && "Failed to find last instruction in bundle")((LastInst && "Failed to find last instruction in bundle" ) ? static_cast<void> (0) : __assert_fail ("LastInst && \"Failed to find last instruction in bundle\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4128, __PRETTY_FUNCTION__)); | ||||
4129 | |||||
4130 | // Set the insertion point after the last instruction in the bundle. Set the | ||||
4131 | // debug location to Front. | ||||
4132 | Builder.SetInsertPoint(BB, ++LastInst->getIterator()); | ||||
4133 | Builder.SetCurrentDebugLocation(Front->getDebugLoc()); | ||||
4134 | } | ||||
4135 | |||||
4136 | Value *BoUpSLP::gather(ArrayRef<Value *> VL) { | ||||
4137 | Value *Val0 = | ||||
4138 | isa<StoreInst>(VL[0]) ? cast<StoreInst>(VL[0])->getValueOperand() : VL[0]; | ||||
4139 | FixedVectorType *VecTy = FixedVectorType::get(Val0->getType(), VL.size()); | ||||
4140 | Value *Vec = UndefValue::get(VecTy); | ||||
4141 | unsigned InsIndex = 0; | ||||
4142 | for (Value *Val : VL) { | ||||
4143 | Vec = Builder.CreateInsertElement(Vec, Val, Builder.getInt32(InsIndex++)); | ||||
4144 | auto *InsElt = dyn_cast<InsertElementInst>(Vec); | ||||
4145 | if (!InsElt) | ||||
4146 | continue; | ||||
4147 | GatherSeq.insert(InsElt); | ||||
4148 | CSEBlocks.insert(InsElt->getParent()); | ||||
4149 | // Add to our 'need-to-extract' list. | ||||
4150 | if (TreeEntry *Entry = getTreeEntry(Val)) { | ||||
4151 | // Find which lane we need to extract. | ||||
4152 | unsigned FoundLane = std::distance(Entry->Scalars.begin(), | ||||
4153 | find(Entry->Scalars, Val)); | ||||
4154 | assert(FoundLane < Entry->Scalars.size() && "Couldn't find extract lane")((FoundLane < Entry->Scalars.size() && "Couldn't find extract lane" ) ? static_cast<void> (0) : __assert_fail ("FoundLane < Entry->Scalars.size() && \"Couldn't find extract lane\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4154, __PRETTY_FUNCTION__)); | ||||
4155 | if (!Entry->ReuseShuffleIndices.empty()) { | ||||
4156 | FoundLane = std::distance(Entry->ReuseShuffleIndices.begin(), | ||||
4157 | find(Entry->ReuseShuffleIndices, FoundLane)); | ||||
4158 | } | ||||
4159 | ExternalUses.push_back(ExternalUser(Val, InsElt, FoundLane)); | ||||
4160 | } | ||||
4161 | } | ||||
4162 | |||||
4163 | return Vec; | ||||
4164 | } | ||||
4165 | |||||
4166 | Value *BoUpSLP::vectorizeTree(ArrayRef<Value *> VL) { | ||||
4167 | InstructionsState S = getSameOpcode(VL); | ||||
4168 | if (S.getOpcode()) { | ||||
4169 | if (TreeEntry *E = getTreeEntry(S.OpValue)) { | ||||
4170 | if (E->isSame(VL)) { | ||||
4171 | Value *V = vectorizeTree(E); | ||||
4172 | if (VL.size() == E->Scalars.size() && !E->ReuseShuffleIndices.empty()) { | ||||
4173 | // We need to get the vectorized value but without shuffle. | ||||
4174 | if (auto *SV = dyn_cast<ShuffleVectorInst>(V)) { | ||||
4175 | V = SV->getOperand(0); | ||||
4176 | } else { | ||||
4177 | // Reshuffle to get only unique values. | ||||
4178 | SmallVector<int, 4> UniqueIdxs; | ||||
4179 | SmallSet<int, 4> UsedIdxs; | ||||
4180 | for (int Idx : E->ReuseShuffleIndices) | ||||
4181 | if (UsedIdxs.insert(Idx).second) | ||||
4182 | UniqueIdxs.emplace_back(Idx); | ||||
4183 | V = Builder.CreateShuffleVector(V, UniqueIdxs); | ||||
4184 | } | ||||
4185 | } | ||||
4186 | return V; | ||||
4187 | } | ||||
4188 | } | ||||
4189 | } | ||||
4190 | |||||
4191 | // Check that every instruction appears once in this bundle. | ||||
4192 | SmallVector<int, 4> ReuseShuffleIndicies; | ||||
4193 | SmallVector<Value *, 4> UniqueValues; | ||||
4194 | if (VL.size() > 2) { | ||||
4195 | DenseMap<Value *, unsigned> UniquePositions; | ||||
4196 | for (Value *V : VL) { | ||||
4197 | auto Res = UniquePositions.try_emplace(V, UniqueValues.size()); | ||||
4198 | ReuseShuffleIndicies.emplace_back(Res.first->second); | ||||
4199 | if (Res.second || isa<Constant>(V)) | ||||
4200 | UniqueValues.emplace_back(V); | ||||
4201 | } | ||||
4202 | // Do not shuffle single element or if number of unique values is not power | ||||
4203 | // of 2. | ||||
4204 | if (UniqueValues.size() == VL.size() || UniqueValues.size() <= 1 || | ||||
4205 | !llvm::isPowerOf2_32(UniqueValues.size())) | ||||
4206 | ReuseShuffleIndicies.clear(); | ||||
4207 | else | ||||
4208 | VL = UniqueValues; | ||||
4209 | } | ||||
4210 | |||||
4211 | Value *Vec = gather(VL); | ||||
4212 | if (!ReuseShuffleIndicies.empty()) { | ||||
4213 | Vec = Builder.CreateShuffleVector(Vec, ReuseShuffleIndicies, "shuffle"); | ||||
4214 | if (auto *I = dyn_cast<Instruction>(Vec)) { | ||||
4215 | GatherSeq.insert(I); | ||||
4216 | CSEBlocks.insert(I->getParent()); | ||||
4217 | } | ||||
4218 | } | ||||
4219 | return Vec; | ||||
4220 | } | ||||
4221 | |||||
4222 | Value *BoUpSLP::vectorizeTree(TreeEntry *E) { | ||||
4223 | IRBuilder<>::InsertPointGuard Guard(Builder); | ||||
4224 | |||||
4225 | if (E->VectorizedValue) { | ||||
4226 | 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); | ||||
4227 | return E->VectorizedValue; | ||||
4228 | } | ||||
4229 | |||||
4230 | bool NeedToShuffleReuses = !E->ReuseShuffleIndices.empty(); | ||||
4231 | if (E->State == TreeEntry::NeedToGather) { | ||||
4232 | setInsertPointAfterBundle(E); | ||||
4233 | Value *Vec = gather(E->Scalars); | ||||
4234 | if (NeedToShuffleReuses) { | ||||
4235 | Vec = Builder.CreateShuffleVector(Vec, E->ReuseShuffleIndices, "shuffle"); | ||||
4236 | if (auto *I = dyn_cast<Instruction>(Vec)) { | ||||
4237 | GatherSeq.insert(I); | ||||
4238 | CSEBlocks.insert(I->getParent()); | ||||
4239 | } | ||||
4240 | } | ||||
4241 | E->VectorizedValue = Vec; | ||||
4242 | return Vec; | ||||
4243 | } | ||||
4244 | |||||
4245 | assert(E->State == TreeEntry::Vectorize && "Unhandled state")((E->State == TreeEntry::Vectorize && "Unhandled state" ) ? static_cast<void> (0) : __assert_fail ("E->State == TreeEntry::Vectorize && \"Unhandled state\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4245, __PRETTY_FUNCTION__)); | ||||
4246 | unsigned ShuffleOrOp = | ||||
4247 | E->isAltShuffle() ? (unsigned)Instruction::ShuffleVector : E->getOpcode(); | ||||
4248 | Instruction *VL0 = E->getMainOp(); | ||||
4249 | Type *ScalarTy = VL0->getType(); | ||||
4250 | if (auto *Store = dyn_cast<StoreInst>(VL0)) | ||||
4251 | ScalarTy = Store->getValueOperand()->getType(); | ||||
4252 | auto *VecTy = FixedVectorType::get(ScalarTy, E->Scalars.size()); | ||||
4253 | switch (ShuffleOrOp) { | ||||
4254 | case Instruction::PHI: { | ||||
4255 | auto *PH = cast<PHINode>(VL0); | ||||
4256 | Builder.SetInsertPoint(PH->getParent()->getFirstNonPHI()); | ||||
4257 | Builder.SetCurrentDebugLocation(PH->getDebugLoc()); | ||||
4258 | PHINode *NewPhi = Builder.CreatePHI(VecTy, PH->getNumIncomingValues()); | ||||
4259 | Value *V = NewPhi; | ||||
4260 | if (NeedToShuffleReuses) | ||||
4261 | V = Builder.CreateShuffleVector(V, E->ReuseShuffleIndices, "shuffle"); | ||||
4262 | |||||
4263 | E->VectorizedValue = V; | ||||
4264 | |||||
4265 | // PHINodes may have multiple entries from the same block. We want to | ||||
4266 | // visit every block once. | ||||
4267 | SmallPtrSet<BasicBlock*, 4> VisitedBBs; | ||||
4268 | |||||
4269 | for (unsigned i = 0, e = PH->getNumIncomingValues(); i < e; ++i) { | ||||
4270 | ValueList Operands; | ||||
4271 | BasicBlock *IBB = PH->getIncomingBlock(i); | ||||
4272 | |||||
4273 | if (!VisitedBBs.insert(IBB).second) { | ||||
4274 | NewPhi->addIncoming(NewPhi->getIncomingValueForBlock(IBB), IBB); | ||||
4275 | continue; | ||||
4276 | } | ||||
4277 | |||||
4278 | Builder.SetInsertPoint(IBB->getTerminator()); | ||||
4279 | Builder.SetCurrentDebugLocation(PH->getDebugLoc()); | ||||
4280 | Value *Vec = vectorizeTree(E->getOperand(i)); | ||||
4281 | NewPhi->addIncoming(Vec, IBB); | ||||
4282 | } | ||||
4283 | |||||
4284 | assert(NewPhi->getNumIncomingValues() == PH->getNumIncomingValues() &&((NewPhi->getNumIncomingValues() == PH->getNumIncomingValues () && "Invalid number of incoming values") ? static_cast <void> (0) : __assert_fail ("NewPhi->getNumIncomingValues() == PH->getNumIncomingValues() && \"Invalid number of incoming values\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4285, __PRETTY_FUNCTION__)) | ||||
4285 | "Invalid number of incoming values")((NewPhi->getNumIncomingValues() == PH->getNumIncomingValues () && "Invalid number of incoming values") ? static_cast <void> (0) : __assert_fail ("NewPhi->getNumIncomingValues() == PH->getNumIncomingValues() && \"Invalid number of incoming values\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4285, __PRETTY_FUNCTION__)); | ||||
4286 | return V; | ||||
4287 | } | ||||
4288 | |||||
4289 | case Instruction::ExtractElement: { | ||||
4290 | Value *V = E->getSingleOperand(0); | ||||
4291 | if (!E->ReorderIndices.empty()) { | ||||
4292 | SmallVector<int, 4> Mask; | ||||
4293 | inversePermutation(E->ReorderIndices, Mask); | ||||
4294 | Builder.SetInsertPoint(VL0); | ||||
4295 | V = Builder.CreateShuffleVector(V, Mask, "reorder_shuffle"); | ||||
4296 | } | ||||
4297 | if (NeedToShuffleReuses) { | ||||
4298 | // TODO: Merge this shuffle with the ReorderShuffleMask. | ||||
4299 | if (E->ReorderIndices.empty()) | ||||
4300 | Builder.SetInsertPoint(VL0); | ||||
4301 | V = Builder.CreateShuffleVector(V, E->ReuseShuffleIndices, "shuffle"); | ||||
4302 | } | ||||
4303 | E->VectorizedValue = V; | ||||
4304 | return V; | ||||
4305 | } | ||||
4306 | case Instruction::ExtractValue: { | ||||
4307 | auto *LI = cast<LoadInst>(E->getSingleOperand(0)); | ||||
4308 | Builder.SetInsertPoint(LI); | ||||
4309 | auto *PtrTy = PointerType::get(VecTy, LI->getPointerAddressSpace()); | ||||
4310 | Value *Ptr = Builder.CreateBitCast(LI->getOperand(0), PtrTy); | ||||
4311 | LoadInst *V = Builder.CreateAlignedLoad(VecTy, Ptr, LI->getAlign()); | ||||
4312 | Value *NewV = propagateMetadata(V, E->Scalars); | ||||
4313 | if (!E->ReorderIndices.empty()) { | ||||
4314 | SmallVector<int, 4> Mask; | ||||
4315 | inversePermutation(E->ReorderIndices, Mask); | ||||
4316 | NewV = Builder.CreateShuffleVector(NewV, Mask, "reorder_shuffle"); | ||||
4317 | } | ||||
4318 | if (NeedToShuffleReuses) { | ||||
4319 | // TODO: Merge this shuffle with the ReorderShuffleMask. | ||||
4320 | NewV = Builder.CreateShuffleVector(NewV, E->ReuseShuffleIndices, | ||||
4321 | "shuffle"); | ||||
4322 | } | ||||
4323 | E->VectorizedValue = NewV; | ||||
4324 | return NewV; | ||||
4325 | } | ||||
4326 | case Instruction::ZExt: | ||||
4327 | case Instruction::SExt: | ||||
4328 | case Instruction::FPToUI: | ||||
4329 | case Instruction::FPToSI: | ||||
4330 | case Instruction::FPExt: | ||||
4331 | case Instruction::PtrToInt: | ||||
4332 | case Instruction::IntToPtr: | ||||
4333 | case Instruction::SIToFP: | ||||
4334 | case Instruction::UIToFP: | ||||
4335 | case Instruction::Trunc: | ||||
4336 | case Instruction::FPTrunc: | ||||
4337 | case Instruction::BitCast: { | ||||
4338 | setInsertPointAfterBundle(E); | ||||
4339 | |||||
4340 | Value *InVec = vectorizeTree(E->getOperand(0)); | ||||
4341 | |||||
4342 | if (E->VectorizedValue) { | ||||
4343 | 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); | ||||
4344 | return E->VectorizedValue; | ||||
4345 | } | ||||
4346 | |||||
4347 | auto *CI = cast<CastInst>(VL0); | ||||
4348 | Value *V = Builder.CreateCast(CI->getOpcode(), InVec, VecTy); | ||||
4349 | if (NeedToShuffleReuses) | ||||
4350 | V = Builder.CreateShuffleVector(V, E->ReuseShuffleIndices, "shuffle"); | ||||
4351 | |||||
4352 | E->VectorizedValue = V; | ||||
4353 | ++NumVectorInstructions; | ||||
4354 | return V; | ||||
4355 | } | ||||
4356 | case Instruction::FCmp: | ||||
4357 | case Instruction::ICmp: { | ||||
4358 | setInsertPointAfterBundle(E); | ||||
4359 | |||||
4360 | Value *L = vectorizeTree(E->getOperand(0)); | ||||
4361 | Value *R = vectorizeTree(E->getOperand(1)); | ||||
4362 | |||||
4363 | if (E->VectorizedValue) { | ||||
4364 | 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); | ||||
4365 | return E->VectorizedValue; | ||||
4366 | } | ||||
4367 | |||||
4368 | CmpInst::Predicate P0 = cast<CmpInst>(VL0)->getPredicate(); | ||||
4369 | Value *V = Builder.CreateCmp(P0, L, R); | ||||
4370 | propagateIRFlags(V, E->Scalars, VL0); | ||||
4371 | if (NeedToShuffleReuses) | ||||
4372 | V = Builder.CreateShuffleVector(V, E->ReuseShuffleIndices, "shuffle"); | ||||
4373 | |||||
4374 | E->VectorizedValue = V; | ||||
4375 | ++NumVectorInstructions; | ||||
4376 | return V; | ||||
4377 | } | ||||
4378 | case Instruction::Select: { | ||||
4379 | setInsertPointAfterBundle(E); | ||||
4380 | |||||
4381 | Value *Cond = vectorizeTree(E->getOperand(0)); | ||||
4382 | Value *True = vectorizeTree(E->getOperand(1)); | ||||
4383 | Value *False = vectorizeTree(E->getOperand(2)); | ||||
4384 | |||||
4385 | if (E->VectorizedValue) { | ||||
4386 | 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); | ||||
4387 | return E->VectorizedValue; | ||||
4388 | } | ||||
4389 | |||||
4390 | Value *V = Builder.CreateSelect(Cond, True, False); | ||||
4391 | if (NeedToShuffleReuses) | ||||
4392 | V = Builder.CreateShuffleVector(V, E->ReuseShuffleIndices, "shuffle"); | ||||
4393 | |||||
4394 | E->VectorizedValue = V; | ||||
4395 | ++NumVectorInstructions; | ||||
4396 | return V; | ||||
4397 | } | ||||
4398 | case Instruction::FNeg: { | ||||
4399 | setInsertPointAfterBundle(E); | ||||
4400 | |||||
4401 | Value *Op = vectorizeTree(E->getOperand(0)); | ||||
4402 | |||||
4403 | if (E->VectorizedValue) { | ||||
4404 | 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); | ||||
4405 | return E->VectorizedValue; | ||||
4406 | } | ||||
4407 | |||||
4408 | Value *V = Builder.CreateUnOp( | ||||
4409 | static_cast<Instruction::UnaryOps>(E->getOpcode()), Op); | ||||
4410 | propagateIRFlags(V, E->Scalars, VL0); | ||||
4411 | if (auto *I = dyn_cast<Instruction>(V)) | ||||
4412 | V = propagateMetadata(I, E->Scalars); | ||||
4413 | |||||
4414 | if (NeedToShuffleReuses) | ||||
4415 | V = Builder.CreateShuffleVector(V, E->ReuseShuffleIndices, "shuffle"); | ||||
4416 | |||||
4417 | E->VectorizedValue = V; | ||||
4418 | ++NumVectorInstructions; | ||||
4419 | |||||
4420 | return V; | ||||
4421 | } | ||||
4422 | case Instruction::Add: | ||||
4423 | case Instruction::FAdd: | ||||
4424 | case Instruction::Sub: | ||||
4425 | case Instruction::FSub: | ||||
4426 | case Instruction::Mul: | ||||
4427 | case Instruction::FMul: | ||||
4428 | case Instruction::UDiv: | ||||
4429 | case Instruction::SDiv: | ||||
4430 | case Instruction::FDiv: | ||||
4431 | case Instruction::URem: | ||||
4432 | case Instruction::SRem: | ||||
4433 | case Instruction::FRem: | ||||
4434 | case Instruction::Shl: | ||||
4435 | case Instruction::LShr: | ||||
4436 | case Instruction::AShr: | ||||
4437 | case Instruction::And: | ||||
4438 | case Instruction::Or: | ||||
4439 | case Instruction::Xor: { | ||||
4440 | setInsertPointAfterBundle(E); | ||||
4441 | |||||
4442 | Value *LHS = vectorizeTree(E->getOperand(0)); | ||||
4443 | Value *RHS = vectorizeTree(E->getOperand(1)); | ||||
4444 | |||||
4445 | if (E->VectorizedValue) { | ||||
4446 | 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); | ||||
4447 | return E->VectorizedValue; | ||||
4448 | } | ||||
4449 | |||||
4450 | Value *V = Builder.CreateBinOp( | ||||
4451 | static_cast<Instruction::BinaryOps>(E->getOpcode()), LHS, | ||||
4452 | RHS); | ||||
4453 | propagateIRFlags(V, E->Scalars, VL0); | ||||
4454 | if (auto *I = dyn_cast<Instruction>(V)) | ||||
4455 | V = propagateMetadata(I, E->Scalars); | ||||
4456 | |||||
4457 | if (NeedToShuffleReuses) | ||||
4458 | V = Builder.CreateShuffleVector(V, E->ReuseShuffleIndices, "shuffle"); | ||||
4459 | |||||
4460 | E->VectorizedValue = V; | ||||
4461 | ++NumVectorInstructions; | ||||
4462 | |||||
4463 | return V; | ||||
4464 | } | ||||
4465 | case Instruction::Load: { | ||||
4466 | // Loads are inserted at the head of the tree because we don't want to | ||||
4467 | // sink them all the way down past store instructions. | ||||
4468 | bool IsReorder = E->updateStateIfReorder(); | ||||
4469 | if (IsReorder) | ||||
4470 | VL0 = E->getMainOp(); | ||||
4471 | setInsertPointAfterBundle(E); | ||||
4472 | |||||
4473 | LoadInst *LI = cast<LoadInst>(VL0); | ||||
4474 | unsigned AS = LI->getPointerAddressSpace(); | ||||
4475 | |||||
4476 | Value *VecPtr = Builder.CreateBitCast(LI->getPointerOperand(), | ||||
4477 | VecTy->getPointerTo(AS)); | ||||
4478 | |||||
4479 | // The pointer operand uses an in-tree scalar so we add the new BitCast to | ||||
4480 | // ExternalUses list to make sure that an extract will be generated in the | ||||
4481 | // future. | ||||
4482 | Value *PO = LI->getPointerOperand(); | ||||
4483 | if (getTreeEntry(PO)) | ||||
4484 | ExternalUses.push_back(ExternalUser(PO, cast<User>(VecPtr), 0)); | ||||
4485 | |||||
4486 | LI = Builder.CreateAlignedLoad(VecTy, VecPtr, LI->getAlign()); | ||||
4487 | Value *V = propagateMetadata(LI, E->Scalars); | ||||
4488 | if (IsReorder) { | ||||
4489 | SmallVector<int, 4> Mask; | ||||
4490 | inversePermutation(E->ReorderIndices, Mask); | ||||
4491 | V = Builder.CreateShuffleVector(V, Mask, "reorder_shuffle"); | ||||
4492 | } | ||||
4493 | if (NeedToShuffleReuses) { | ||||
4494 | // TODO: Merge this shuffle with the ReorderShuffleMask. | ||||
4495 | V = Builder.CreateShuffleVector(V, E->ReuseShuffleIndices, "shuffle"); | ||||
4496 | } | ||||
4497 | E->VectorizedValue = V; | ||||
4498 | ++NumVectorInstructions; | ||||
4499 | return V; | ||||
4500 | } | ||||
4501 | case Instruction::Store: { | ||||
4502 | bool IsReorder = !E->ReorderIndices.empty(); | ||||
4503 | auto *SI = cast<StoreInst>( | ||||
4504 | IsReorder ? E->Scalars[E->ReorderIndices.front()] : VL0); | ||||
4505 | unsigned AS = SI->getPointerAddressSpace(); | ||||
4506 | |||||
4507 | setInsertPointAfterBundle(E); | ||||
4508 | |||||
4509 | Value *VecValue = vectorizeTree(E->getOperand(0)); | ||||
4510 | if (IsReorder) { | ||||
4511 | SmallVector<int, 4> Mask(E->ReorderIndices.begin(), | ||||
4512 | E->ReorderIndices.end()); | ||||
4513 | VecValue = Builder.CreateShuffleVector(VecValue, Mask, "reorder_shuf"); | ||||
4514 | } | ||||
4515 | Value *ScalarPtr = SI->getPointerOperand(); | ||||
4516 | Value *VecPtr = Builder.CreateBitCast( | ||||
4517 | ScalarPtr, VecValue->getType()->getPointerTo(AS)); | ||||
4518 | StoreInst *ST = Builder.CreateAlignedStore(VecValue, VecPtr, | ||||
4519 | SI->getAlign()); | ||||
4520 | |||||
4521 | // The pointer operand uses an in-tree scalar, so add the new BitCast to | ||||
4522 | // ExternalUses to make sure that an extract will be generated in the | ||||
4523 | // future. | ||||
4524 | if (getTreeEntry(ScalarPtr)) | ||||
4525 | ExternalUses.push_back(ExternalUser(ScalarPtr, cast<User>(VecPtr), 0)); | ||||
4526 | |||||
4527 | Value *V = propagateMetadata(ST, E->Scalars); | ||||
4528 | if (NeedToShuffleReuses) | ||||
4529 | V = Builder.CreateShuffleVector(V, E->ReuseShuffleIndices, "shuffle"); | ||||
4530 | |||||
4531 | E->VectorizedValue = V; | ||||
4532 | ++NumVectorInstructions; | ||||
4533 | return V; | ||||
4534 | } | ||||
4535 | case Instruction::GetElementPtr: { | ||||
4536 | setInsertPointAfterBundle(E); | ||||
4537 | |||||
4538 | Value *Op0 = vectorizeTree(E->getOperand(0)); | ||||
4539 | |||||
4540 | std::vector<Value *> OpVecs; | ||||
4541 | for (int j = 1, e = cast<GetElementPtrInst>(VL0)->getNumOperands(); j < e; | ||||
4542 | ++j) { | ||||
4543 | ValueList &VL = E->getOperand(j); | ||||
4544 | // Need to cast all elements to the same type before vectorization to | ||||
4545 | // avoid crash. | ||||
4546 | Type *VL0Ty = VL0->getOperand(j)->getType(); | ||||
4547 | Type *Ty = llvm::all_of( | ||||
4548 | VL, [VL0Ty](Value *V) { return VL0Ty == V->getType(); }) | ||||
4549 | ? VL0Ty | ||||
4550 | : DL->getIndexType(cast<GetElementPtrInst>(VL0) | ||||
4551 | ->getPointerOperandType() | ||||
4552 | ->getScalarType()); | ||||
4553 | for (Value *&V : VL) { | ||||
4554 | auto *CI = cast<ConstantInt>(V); | ||||
4555 | V = ConstantExpr::getIntegerCast(CI, Ty, | ||||
4556 | CI->getValue().isSignBitSet()); | ||||
4557 | } | ||||
4558 | Value *OpVec = vectorizeTree(VL); | ||||
4559 | OpVecs.push_back(OpVec); | ||||
4560 | } | ||||
4561 | |||||
4562 | Value *V = Builder.CreateGEP( | ||||
4563 | cast<GetElementPtrInst>(VL0)->getSourceElementType(), Op0, OpVecs); | ||||
4564 | if (Instruction *I = dyn_cast<Instruction>(V)) | ||||
4565 | V = propagateMetadata(I, E->Scalars); | ||||
4566 | |||||
4567 | if (NeedToShuffleReuses) | ||||
4568 | V = Builder.CreateShuffleVector(V, E->ReuseShuffleIndices, "shuffle"); | ||||
4569 | |||||
4570 | E->VectorizedValue = V; | ||||
4571 | ++NumVectorInstructions; | ||||
4572 | |||||
4573 | return V; | ||||
4574 | } | ||||
4575 | case Instruction::Call: { | ||||
4576 | CallInst *CI = cast<CallInst>(VL0); | ||||
4577 | setInsertPointAfterBundle(E); | ||||
4578 | |||||
4579 | Intrinsic::ID IID = Intrinsic::not_intrinsic; | ||||
4580 | if (Function *FI = CI->getCalledFunction()) | ||||
4581 | IID = FI->getIntrinsicID(); | ||||
4582 | |||||
4583 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); | ||||
4584 | |||||
4585 | auto VecCallCosts = getVectorCallCosts(CI, VecTy, TTI, TLI); | ||||
4586 | bool UseIntrinsic = ID != Intrinsic::not_intrinsic && | ||||
4587 | VecCallCosts.first <= VecCallCosts.second; | ||||
4588 | |||||
4589 | Value *ScalarArg = nullptr; | ||||
4590 | std::vector<Value *> OpVecs; | ||||
4591 | for (int j = 0, e = CI->getNumArgOperands(); j < e; ++j) { | ||||
4592 | ValueList OpVL; | ||||
4593 | // Some intrinsics have scalar arguments. This argument should not be | ||||
4594 | // vectorized. | ||||
4595 | if (UseIntrinsic && hasVectorInstrinsicScalarOpd(IID, j)) { | ||||
4596 | CallInst *CEI = cast<CallInst>(VL0); | ||||
4597 | ScalarArg = CEI->getArgOperand(j); | ||||
4598 | OpVecs.push_back(CEI->getArgOperand(j)); | ||||
4599 | continue; | ||||
4600 | } | ||||
4601 | |||||
4602 | Value *OpVec = vectorizeTree(E->getOperand(j)); | ||||
4603 | LLVM_DEBUG(dbgs() << "SLP: OpVec[" << j << "]: " << *OpVec << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: OpVec[" << j << "]: " << *OpVec << "\n"; } } while (false); | ||||
4604 | OpVecs.push_back(OpVec); | ||||
4605 | } | ||||
4606 | |||||
4607 | Function *CF; | ||||
4608 | if (!UseIntrinsic) { | ||||
4609 | VFShape Shape = | ||||
4610 | VFShape::get(*CI, ElementCount::getFixed(static_cast<unsigned>( | ||||
4611 | VecTy->getNumElements())), | ||||
4612 | false /*HasGlobalPred*/); | ||||
4613 | CF = VFDatabase(*CI).getVectorizedFunction(Shape); | ||||
4614 | } else { | ||||
4615 | Type *Tys[] = {FixedVectorType::get(CI->getType(), E->Scalars.size())}; | ||||
4616 | CF = Intrinsic::getDeclaration(F->getParent(), ID, Tys); | ||||
4617 | } | ||||
4618 | |||||
4619 | SmallVector<OperandBundleDef, 1> OpBundles; | ||||
4620 | CI->getOperandBundlesAsDefs(OpBundles); | ||||
4621 | Value *V = Builder.CreateCall(CF, OpVecs, OpBundles); | ||||
4622 | |||||
4623 | // The scalar argument uses an in-tree scalar so we add the new vectorized | ||||
4624 | // call to ExternalUses list to make sure that an extract will be | ||||
4625 | // generated in the future. | ||||
4626 | if (ScalarArg && getTreeEntry(ScalarArg)) | ||||
4627 | ExternalUses.push_back(ExternalUser(ScalarArg, cast<User>(V), 0)); | ||||
4628 | |||||
4629 | propagateIRFlags(V, E->Scalars, VL0); | ||||
4630 | if (NeedToShuffleReuses) | ||||
4631 | V = Builder.CreateShuffleVector(V, E->ReuseShuffleIndices, "shuffle"); | ||||
4632 | |||||
4633 | E->VectorizedValue = V; | ||||
4634 | ++NumVectorInstructions; | ||||
4635 | return V; | ||||
4636 | } | ||||
4637 | case Instruction::ShuffleVector: { | ||||
4638 | assert(E->isAltShuffle() &&((E->isAltShuffle() && ((Instruction::isBinaryOp(E ->getOpcode()) && Instruction::isBinaryOp(E->getAltOpcode ())) || (Instruction::isCast(E->getOpcode()) && Instruction ::isCast(E->getAltOpcode()))) && "Invalid Shuffle Vector Operand" ) ? static_cast<void> (0) : __assert_fail ("E->isAltShuffle() && ((Instruction::isBinaryOp(E->getOpcode()) && Instruction::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E->getOpcode()) && Instruction::isCast(E->getAltOpcode()))) && \"Invalid Shuffle Vector Operand\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4643, __PRETTY_FUNCTION__)) | ||||
4639 | ((Instruction::isBinaryOp(E->getOpcode()) &&((E->isAltShuffle() && ((Instruction::isBinaryOp(E ->getOpcode()) && Instruction::isBinaryOp(E->getAltOpcode ())) || (Instruction::isCast(E->getOpcode()) && Instruction ::isCast(E->getAltOpcode()))) && "Invalid Shuffle Vector Operand" ) ? static_cast<void> (0) : __assert_fail ("E->isAltShuffle() && ((Instruction::isBinaryOp(E->getOpcode()) && Instruction::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E->getOpcode()) && Instruction::isCast(E->getAltOpcode()))) && \"Invalid Shuffle Vector Operand\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4643, __PRETTY_FUNCTION__)) | ||||
4640 | Instruction::isBinaryOp(E->getAltOpcode())) ||((E->isAltShuffle() && ((Instruction::isBinaryOp(E ->getOpcode()) && Instruction::isBinaryOp(E->getAltOpcode ())) || (Instruction::isCast(E->getOpcode()) && Instruction ::isCast(E->getAltOpcode()))) && "Invalid Shuffle Vector Operand" ) ? static_cast<void> (0) : __assert_fail ("E->isAltShuffle() && ((Instruction::isBinaryOp(E->getOpcode()) && Instruction::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E->getOpcode()) && Instruction::isCast(E->getAltOpcode()))) && \"Invalid Shuffle Vector Operand\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4643, __PRETTY_FUNCTION__)) | ||||
4641 | (Instruction::isCast(E->getOpcode()) &&((E->isAltShuffle() && ((Instruction::isBinaryOp(E ->getOpcode()) && Instruction::isBinaryOp(E->getAltOpcode ())) || (Instruction::isCast(E->getOpcode()) && Instruction ::isCast(E->getAltOpcode()))) && "Invalid Shuffle Vector Operand" ) ? static_cast<void> (0) : __assert_fail ("E->isAltShuffle() && ((Instruction::isBinaryOp(E->getOpcode()) && Instruction::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E->getOpcode()) && Instruction::isCast(E->getAltOpcode()))) && \"Invalid Shuffle Vector Operand\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4643, __PRETTY_FUNCTION__)) | ||||
4642 | Instruction::isCast(E->getAltOpcode()))) &&((E->isAltShuffle() && ((Instruction::isBinaryOp(E ->getOpcode()) && Instruction::isBinaryOp(E->getAltOpcode ())) || (Instruction::isCast(E->getOpcode()) && Instruction ::isCast(E->getAltOpcode()))) && "Invalid Shuffle Vector Operand" ) ? static_cast<void> (0) : __assert_fail ("E->isAltShuffle() && ((Instruction::isBinaryOp(E->getOpcode()) && Instruction::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E->getOpcode()) && Instruction::isCast(E->getAltOpcode()))) && \"Invalid Shuffle Vector Operand\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4643, __PRETTY_FUNCTION__)) | ||||
4643 | "Invalid Shuffle Vector Operand")((E->isAltShuffle() && ((Instruction::isBinaryOp(E ->getOpcode()) && Instruction::isBinaryOp(E->getAltOpcode ())) || (Instruction::isCast(E->getOpcode()) && Instruction ::isCast(E->getAltOpcode()))) && "Invalid Shuffle Vector Operand" ) ? static_cast<void> (0) : __assert_fail ("E->isAltShuffle() && ((Instruction::isBinaryOp(E->getOpcode()) && Instruction::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E->getOpcode()) && Instruction::isCast(E->getAltOpcode()))) && \"Invalid Shuffle Vector Operand\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4643, __PRETTY_FUNCTION__)); | ||||
4644 | |||||
4645 | Value *LHS = nullptr, *RHS = nullptr; | ||||
4646 | if (Instruction::isBinaryOp(E->getOpcode())) { | ||||
4647 | setInsertPointAfterBundle(E); | ||||
4648 | LHS = vectorizeTree(E->getOperand(0)); | ||||
4649 | RHS = vectorizeTree(E->getOperand(1)); | ||||
4650 | } else { | ||||
4651 | setInsertPointAfterBundle(E); | ||||
4652 | LHS = vectorizeTree(E->getOperand(0)); | ||||
4653 | } | ||||
4654 | |||||
4655 | if (E->VectorizedValue) { | ||||
4656 | 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); | ||||
4657 | return E->VectorizedValue; | ||||
4658 | } | ||||
4659 | |||||
4660 | Value *V0, *V1; | ||||
4661 | if (Instruction::isBinaryOp(E->getOpcode())) { | ||||
4662 | V0 = Builder.CreateBinOp( | ||||
4663 | static_cast<Instruction::BinaryOps>(E->getOpcode()), LHS, RHS); | ||||
4664 | V1 = Builder.CreateBinOp( | ||||
4665 | static_cast<Instruction::BinaryOps>(E->getAltOpcode()), LHS, RHS); | ||||
4666 | } else { | ||||
4667 | V0 = Builder.CreateCast( | ||||
4668 | static_cast<Instruction::CastOps>(E->getOpcode()), LHS, VecTy); | ||||
4669 | V1 = Builder.CreateCast( | ||||
4670 | static_cast<Instruction::CastOps>(E->getAltOpcode()), LHS, VecTy); | ||||
4671 | } | ||||
4672 | |||||
4673 | // Create shuffle to take alternate operations from the vector. | ||||
4674 | // Also, gather up main and alt scalar ops to propagate IR flags to | ||||
4675 | // each vector operation. | ||||
4676 | ValueList OpScalars, AltScalars; | ||||
4677 | unsigned e = E->Scalars.size(); | ||||
4678 | SmallVector<int, 8> Mask(e); | ||||
4679 | for (unsigned i = 0; i < e; ++i) { | ||||
4680 | auto *OpInst = cast<Instruction>(E->Scalars[i]); | ||||
4681 | assert(E->isOpcodeOrAlt(OpInst) && "Unexpected main/alternate opcode")((E->isOpcodeOrAlt(OpInst) && "Unexpected main/alternate opcode" ) ? static_cast<void> (0) : __assert_fail ("E->isOpcodeOrAlt(OpInst) && \"Unexpected main/alternate opcode\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4681, __PRETTY_FUNCTION__)); | ||||
4682 | if (OpInst->getOpcode() == E->getAltOpcode()) { | ||||
4683 | Mask[i] = e + i; | ||||
4684 | AltScalars.push_back(E->Scalars[i]); | ||||
4685 | } else { | ||||
4686 | Mask[i] = i; | ||||
4687 | OpScalars.push_back(E->Scalars[i]); | ||||
4688 | } | ||||
4689 | } | ||||
4690 | |||||
4691 | propagateIRFlags(V0, OpScalars); | ||||
4692 | propagateIRFlags(V1, AltScalars); | ||||
4693 | |||||
4694 | Value *V = Builder.CreateShuffleVector(V0, V1, Mask); | ||||
4695 | if (Instruction *I = dyn_cast<Instruction>(V)) | ||||
4696 | V = propagateMetadata(I, E->Scalars); | ||||
4697 | if (NeedToShuffleReuses) | ||||
4698 | V = Builder.CreateShuffleVector(V, E->ReuseShuffleIndices, "shuffle"); | ||||
4699 | |||||
4700 | E->VectorizedValue = V; | ||||
4701 | ++NumVectorInstructions; | ||||
4702 | |||||
4703 | return V; | ||||
4704 | } | ||||
4705 | default: | ||||
4706 | llvm_unreachable("unknown inst")::llvm::llvm_unreachable_internal("unknown inst", "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4706); | ||||
4707 | } | ||||
4708 | return nullptr; | ||||
4709 | } | ||||
4710 | |||||
4711 | Value *BoUpSLP::vectorizeTree() { | ||||
4712 | ExtraValueToDebugLocsMap ExternallyUsedValues; | ||||
4713 | return vectorizeTree(ExternallyUsedValues); | ||||
4714 | } | ||||
4715 | |||||
4716 | Value * | ||||
4717 | BoUpSLP::vectorizeTree(ExtraValueToDebugLocsMap &ExternallyUsedValues) { | ||||
4718 | // All blocks must be scheduled before any instructions are inserted. | ||||
4719 | for (auto &BSIter : BlocksSchedules) { | ||||
4720 | scheduleBlock(BSIter.second.get()); | ||||
4721 | } | ||||
4722 | |||||
4723 | Builder.SetInsertPoint(&F->getEntryBlock().front()); | ||||
4724 | auto *VectorRoot = vectorizeTree(VectorizableTree[0].get()); | ||||
4725 | |||||
4726 | // If the vectorized tree can be rewritten in a smaller type, we truncate the | ||||
4727 | // vectorized root. InstCombine will then rewrite the entire expression. We | ||||
4728 | // sign extend the extracted values below. | ||||
4729 | auto *ScalarRoot = VectorizableTree[0]->Scalars[0]; | ||||
4730 | if (MinBWs.count(ScalarRoot)) { | ||||
4731 | if (auto *I = dyn_cast<Instruction>(VectorRoot)) | ||||
4732 | Builder.SetInsertPoint(&*++BasicBlock::iterator(I)); | ||||
4733 | auto BundleWidth = VectorizableTree[0]->Scalars.size(); | ||||
4734 | auto *MinTy = IntegerType::get(F->getContext(), MinBWs[ScalarRoot].first); | ||||
4735 | auto *VecTy = FixedVectorType::get(MinTy, BundleWidth); | ||||
4736 | auto *Trunc = Builder.CreateTrunc(VectorRoot, VecTy); | ||||
4737 | VectorizableTree[0]->VectorizedValue = Trunc; | ||||
4738 | } | ||||
4739 | |||||
4740 | LLVM_DEBUG(dbgs() << "SLP: Extracting " << ExternalUses.size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Extracting " << ExternalUses .size() << " values .\n"; } } while (false) | ||||
4741 | << " values .\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Extracting " << ExternalUses .size() << " values .\n"; } } while (false); | ||||
4742 | |||||
4743 | // If necessary, sign-extend or zero-extend ScalarRoot to the larger type | ||||
4744 | // specified by ScalarType. | ||||
4745 | auto extend = [&](Value *ScalarRoot, Value *Ex, Type *ScalarType) { | ||||
4746 | if (!MinBWs.count(ScalarRoot)) | ||||
4747 | return Ex; | ||||
4748 | if (MinBWs[ScalarRoot].second) | ||||
4749 | return Builder.CreateSExt(Ex, ScalarType); | ||||
4750 | return Builder.CreateZExt(Ex, ScalarType); | ||||
4751 | }; | ||||
4752 | |||||
4753 | // Extract all of the elements with the external uses. | ||||
4754 | for (const auto &ExternalUse : ExternalUses) { | ||||
4755 | Value *Scalar = ExternalUse.Scalar; | ||||
4756 | llvm::User *User = ExternalUse.User; | ||||
4757 | |||||
4758 | // Skip users that we already RAUW. This happens when one instruction | ||||
4759 | // has multiple uses of the same value. | ||||
4760 | if (User && !is_contained(Scalar->users(), User)) | ||||
4761 | continue; | ||||
4762 | TreeEntry *E = getTreeEntry(Scalar); | ||||
4763 | assert(E && "Invalid scalar")((E && "Invalid scalar") ? static_cast<void> (0 ) : __assert_fail ("E && \"Invalid scalar\"", "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4763, __PRETTY_FUNCTION__)); | ||||
4764 | assert(E->State == TreeEntry::Vectorize && "Extracting from a gather list")((E->State == TreeEntry::Vectorize && "Extracting from a gather list" ) ? static_cast<void> (0) : __assert_fail ("E->State == TreeEntry::Vectorize && \"Extracting from a gather list\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4764, __PRETTY_FUNCTION__)); | ||||
4765 | |||||
4766 | Value *Vec = E->VectorizedValue; | ||||
4767 | assert(Vec && "Can't find vectorizable value")((Vec && "Can't find vectorizable value") ? static_cast <void> (0) : __assert_fail ("Vec && \"Can't find vectorizable value\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4767, __PRETTY_FUNCTION__)); | ||||
4768 | |||||
4769 | Value *Lane = Builder.getInt32(ExternalUse.Lane); | ||||
4770 | // If User == nullptr, the Scalar is used as extra arg. Generate | ||||
4771 | // ExtractElement instruction and update the record for this scalar in | ||||
4772 | // ExternallyUsedValues. | ||||
4773 | if (!User) { | ||||
4774 | assert(ExternallyUsedValues.count(Scalar) &&((ExternallyUsedValues.count(Scalar) && "Scalar with nullptr as an external user must be registered in " "ExternallyUsedValues map") ? static_cast<void> (0) : __assert_fail ("ExternallyUsedValues.count(Scalar) && \"Scalar with nullptr as an external user must be registered in \" \"ExternallyUsedValues map\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4776, __PRETTY_FUNCTION__)) | ||||
4775 | "Scalar with nullptr as an external user must be registered in "((ExternallyUsedValues.count(Scalar) && "Scalar with nullptr as an external user must be registered in " "ExternallyUsedValues map") ? static_cast<void> (0) : __assert_fail ("ExternallyUsedValues.count(Scalar) && \"Scalar with nullptr as an external user must be registered in \" \"ExternallyUsedValues map\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4776, __PRETTY_FUNCTION__)) | ||||
4776 | "ExternallyUsedValues map")((ExternallyUsedValues.count(Scalar) && "Scalar with nullptr as an external user must be registered in " "ExternallyUsedValues map") ? static_cast<void> (0) : __assert_fail ("ExternallyUsedValues.count(Scalar) && \"Scalar with nullptr as an external user must be registered in \" \"ExternallyUsedValues map\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4776, __PRETTY_FUNCTION__)); | ||||
4777 | if (auto *VecI = dyn_cast<Instruction>(Vec)) { | ||||
4778 | Builder.SetInsertPoint(VecI->getParent(), | ||||
4779 | std::next(VecI->getIterator())); | ||||
4780 | } else { | ||||
4781 | Builder.SetInsertPoint(&F->getEntryBlock().front()); | ||||
4782 | } | ||||
4783 | Value *Ex = Builder.CreateExtractElement(Vec, Lane); | ||||
4784 | Ex = extend(ScalarRoot, Ex, Scalar->getType()); | ||||
4785 | CSEBlocks.insert(cast<Instruction>(Scalar)->getParent()); | ||||
4786 | auto &Locs = ExternallyUsedValues[Scalar]; | ||||
4787 | ExternallyUsedValues.insert({Ex, Locs}); | ||||
4788 | ExternallyUsedValues.erase(Scalar); | ||||
4789 | // Required to update internally referenced instructions. | ||||
4790 | Scalar->replaceAllUsesWith(Ex); | ||||
4791 | continue; | ||||
4792 | } | ||||
4793 | |||||
4794 | // Generate extracts for out-of-tree users. | ||||
4795 | // Find the insertion point for the extractelement lane. | ||||
4796 | if (auto *VecI = dyn_cast<Instruction>(Vec)) { | ||||
4797 | if (PHINode *PH = dyn_cast<PHINode>(User)) { | ||||
4798 | for (int i = 0, e = PH->getNumIncomingValues(); i != e; ++i) { | ||||
4799 | if (PH->getIncomingValue(i) == Scalar) { | ||||
4800 | Instruction *IncomingTerminator = | ||||
4801 | PH->getIncomingBlock(i)->getTerminator(); | ||||
4802 | if (isa<CatchSwitchInst>(IncomingTerminator)) { | ||||
4803 | Builder.SetInsertPoint(VecI->getParent(), | ||||
4804 | std::next(VecI->getIterator())); | ||||
4805 | } else { | ||||
4806 | Builder.SetInsertPoint(PH->getIncomingBlock(i)->getTerminator()); | ||||
4807 | } | ||||
4808 | Value *Ex = Builder.CreateExtractElement(Vec, Lane); | ||||
4809 | Ex = extend(ScalarRoot, Ex, Scalar->getType()); | ||||
4810 | CSEBlocks.insert(PH->getIncomingBlock(i)); | ||||
4811 | PH->setOperand(i, Ex); | ||||
4812 | } | ||||
4813 | } | ||||
4814 | } else { | ||||
4815 | Builder.SetInsertPoint(cast<Instruction>(User)); | ||||
4816 | Value *Ex = Builder.CreateExtractElement(Vec, Lane); | ||||
4817 | Ex = extend(ScalarRoot, Ex, Scalar->getType()); | ||||
4818 | CSEBlocks.insert(cast<Instruction>(User)->getParent()); | ||||
4819 | User->replaceUsesOfWith(Scalar, Ex); | ||||
4820 | } | ||||
4821 | } else { | ||||
4822 | Builder.SetInsertPoint(&F->getEntryBlock().front()); | ||||
4823 | Value *Ex = Builder.CreateExtractElement(Vec, Lane); | ||||
4824 | Ex = extend(ScalarRoot, Ex, Scalar->getType()); | ||||
4825 | CSEBlocks.insert(&F->getEntryBlock()); | ||||
4826 | User->replaceUsesOfWith(Scalar, Ex); | ||||
4827 | } | ||||
4828 | |||||
4829 | LLVM_DEBUG(dbgs() << "SLP: Replaced:" << *User << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Replaced:" << *User << ".\n"; } } while (false); | ||||
4830 | } | ||||
4831 | |||||
4832 | // For each vectorized value: | ||||
4833 | for (auto &TEPtr : VectorizableTree) { | ||||
4834 | TreeEntry *Entry = TEPtr.get(); | ||||
4835 | |||||
4836 | // No need to handle users of gathered values. | ||||
4837 | if (Entry->State == TreeEntry::NeedToGather) | ||||
4838 | continue; | ||||
4839 | |||||
4840 | assert(Entry->VectorizedValue && "Can't find vectorizable value")((Entry->VectorizedValue && "Can't find vectorizable value" ) ? static_cast<void> (0) : __assert_fail ("Entry->VectorizedValue && \"Can't find vectorizable value\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4840, __PRETTY_FUNCTION__)); | ||||
4841 | |||||
4842 | // For each lane: | ||||
4843 | for (int Lane = 0, LE = Entry->Scalars.size(); Lane != LE; ++Lane) { | ||||
4844 | Value *Scalar = Entry->Scalars[Lane]; | ||||
4845 | |||||
4846 | #ifndef NDEBUG | ||||
4847 | Type *Ty = Scalar->getType(); | ||||
4848 | if (!Ty->isVoidTy()) { | ||||
4849 | for (User *U : Scalar->users()) { | ||||
4850 | LLVM_DEBUG(dbgs() << "SLP: \tvalidating user:" << *U << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: \tvalidating user:" << *U << ".\n"; } } while (false); | ||||
4851 | |||||
4852 | // It is legal to delete users in the ignorelist. | ||||
4853 | assert((getTreeEntry(U) || is_contained(UserIgnoreList, U)) &&(((getTreeEntry(U) || is_contained(UserIgnoreList, U)) && "Deleting out-of-tree value") ? static_cast<void> (0) : __assert_fail ("(getTreeEntry(U) || is_contained(UserIgnoreList, U)) && \"Deleting out-of-tree value\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4854, __PRETTY_FUNCTION__)) | ||||
4854 | "Deleting out-of-tree value")(((getTreeEntry(U) || is_contained(UserIgnoreList, U)) && "Deleting out-of-tree value") ? static_cast<void> (0) : __assert_fail ("(getTreeEntry(U) || is_contained(UserIgnoreList, U)) && \"Deleting out-of-tree value\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4854, __PRETTY_FUNCTION__)); | ||||
4855 | } | ||||
4856 | } | ||||
4857 | #endif | ||||
4858 | LLVM_DEBUG(dbgs() << "SLP: \tErasing scalar:" << *Scalar << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: \tErasing scalar:" << * Scalar << ".\n"; } } while (false); | ||||
4859 | eraseInstruction(cast<Instruction>(Scalar)); | ||||
4860 | } | ||||
4861 | } | ||||
4862 | |||||
4863 | Builder.ClearInsertionPoint(); | ||||
4864 | InstrElementSize.clear(); | ||||
4865 | |||||
4866 | return VectorizableTree[0]->VectorizedValue; | ||||
4867 | } | ||||
4868 | |||||
4869 | void BoUpSLP::optimizeGatherSequence() { | ||||
4870 | LLVM_DEBUG(dbgs() << "SLP: Optimizing " << GatherSeq.size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Optimizing " << GatherSeq .size() << " gather sequences instructions.\n"; } } while (false) | ||||
4871 | << " gather sequences instructions.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Optimizing " << GatherSeq .size() << " gather sequences instructions.\n"; } } while (false); | ||||
4872 | // LICM InsertElementInst sequences. | ||||
4873 | for (Instruction *I : GatherSeq) { | ||||
4874 | if (isDeleted(I)) | ||||
4875 | continue; | ||||
4876 | |||||
4877 | // Check if this block is inside a loop. | ||||
4878 | Loop *L = LI->getLoopFor(I->getParent()); | ||||
4879 | if (!L) | ||||
4880 | continue; | ||||
4881 | |||||
4882 | // Check if it has a preheader. | ||||
4883 | BasicBlock *PreHeader = L->getLoopPreheader(); | ||||
4884 | if (!PreHeader) | ||||
4885 | continue; | ||||
4886 | |||||
4887 | // If the vector or the element that we insert into it are | ||||
4888 | // instructions that are defined in this basic block then we can't | ||||
4889 | // hoist this instruction. | ||||
4890 | auto *Op0 = dyn_cast<Instruction>(I->getOperand(0)); | ||||
4891 | auto *Op1 = dyn_cast<Instruction>(I->getOperand(1)); | ||||
4892 | if (Op0 && L->contains(Op0)) | ||||
4893 | continue; | ||||
4894 | if (Op1 && L->contains(Op1)) | ||||
4895 | continue; | ||||
4896 | |||||
4897 | // We can hoist this instruction. Move it to the pre-header. | ||||
4898 | I->moveBefore(PreHeader->getTerminator()); | ||||
4899 | } | ||||
4900 | |||||
4901 | // Make a list of all reachable blocks in our CSE queue. | ||||
4902 | SmallVector<const DomTreeNode *, 8> CSEWorkList; | ||||
4903 | CSEWorkList.reserve(CSEBlocks.size()); | ||||
4904 | for (BasicBlock *BB : CSEBlocks) | ||||
4905 | if (DomTreeNode *N = DT->getNode(BB)) { | ||||
4906 | assert(DT->isReachableFromEntry(N))((DT->isReachableFromEntry(N)) ? static_cast<void> ( 0) : __assert_fail ("DT->isReachableFromEntry(N)", "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4906, __PRETTY_FUNCTION__)); | ||||
4907 | CSEWorkList.push_back(N); | ||||
4908 | } | ||||
4909 | |||||
4910 | // Sort blocks by domination. This ensures we visit a block after all blocks | ||||
4911 | // dominating it are visited. | ||||
4912 | llvm::stable_sort(CSEWorkList, | ||||
4913 | [this](const DomTreeNode *A, const DomTreeNode *B) { | ||||
4914 | return DT->properlyDominates(A, B); | ||||
4915 | }); | ||||
4916 | |||||
4917 | // Perform O(N^2) search over the gather sequences and merge identical | ||||
4918 | // instructions. TODO: We can further optimize this scan if we split the | ||||
4919 | // instructions into different buckets based on the insert lane. | ||||
4920 | SmallVector<Instruction *, 16> Visited; | ||||
4921 | for (auto I = CSEWorkList.begin(), E = CSEWorkList.end(); I != E; ++I) { | ||||
4922 | assert((I == CSEWorkList.begin() || !DT->dominates(*I, *std::prev(I))) &&(((I == CSEWorkList.begin() || !DT->dominates(*I, *std::prev (I))) && "Worklist not sorted properly!") ? static_cast <void> (0) : __assert_fail ("(I == CSEWorkList.begin() || !DT->dominates(*I, *std::prev(I))) && \"Worklist not sorted properly!\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4923, __PRETTY_FUNCTION__)) | ||||
4923 | "Worklist not sorted properly!")(((I == CSEWorkList.begin() || !DT->dominates(*I, *std::prev (I))) && "Worklist not sorted properly!") ? static_cast <void> (0) : __assert_fail ("(I == CSEWorkList.begin() || !DT->dominates(*I, *std::prev(I))) && \"Worklist not sorted properly!\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4923, __PRETTY_FUNCTION__)); | ||||
4924 | BasicBlock *BB = (*I)->getBlock(); | ||||
| |||||
4925 | // For all instructions in blocks containing gather sequences: | ||||
4926 | for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e;) { | ||||
4927 | Instruction *In = &*it++; | ||||
4928 | if (isDeleted(In)) | ||||
4929 | continue; | ||||
4930 | if (!isa<InsertElementInst>(In) && !isa<ExtractElementInst>(In)) | ||||
4931 | continue; | ||||
4932 | |||||
4933 | // Check if we can replace this instruction with any of the | ||||
4934 | // visited instructions. | ||||
4935 | for (Instruction *v : Visited) { | ||||
4936 | if (In->isIdenticalTo(v) && | ||||
4937 | DT->dominates(v->getParent(), In->getParent())) { | ||||
4938 | In->replaceAllUsesWith(v); | ||||
4939 | eraseInstruction(In); | ||||
4940 | In = nullptr; | ||||
4941 | break; | ||||
4942 | } | ||||
4943 | } | ||||
4944 | if (In) { | ||||
4945 | assert(!is_contained(Visited, In))((!is_contained(Visited, In)) ? static_cast<void> (0) : __assert_fail ("!is_contained(Visited, In)", "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4945, __PRETTY_FUNCTION__)); | ||||
4946 | Visited.push_back(In); | ||||
4947 | } | ||||
4948 | } | ||||
4949 | } | ||||
4950 | CSEBlocks.clear(); | ||||
4951 | GatherSeq.clear(); | ||||
4952 | } | ||||
4953 | |||||
4954 | // Groups the instructions to a bundle (which is then a single scheduling entity) | ||||
4955 | // and schedules instructions until the bundle gets ready. | ||||
4956 | Optional<BoUpSLP::ScheduleData *> | ||||
4957 | BoUpSLP::BlockScheduling::tryScheduleBundle(ArrayRef<Value *> VL, BoUpSLP *SLP, | ||||
4958 | const InstructionsState &S) { | ||||
4959 | if (isa<PHINode>(S.OpValue)) | ||||
4960 | return nullptr; | ||||
4961 | |||||
4962 | // Initialize the instruction bundle. | ||||
4963 | Instruction *OldScheduleEnd = ScheduleEnd; | ||||
4964 | ScheduleData *PrevInBundle = nullptr; | ||||
4965 | ScheduleData *Bundle = nullptr; | ||||
4966 | bool ReSchedule = false; | ||||
4967 | LLVM_DEBUG(dbgs() << "SLP: bundle: " << *S.OpValue << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: bundle: " << *S.OpValue << "\n"; } } while (false); | ||||
4968 | |||||
4969 | // Make sure that the scheduling region contains all | ||||
4970 | // instructions of the bundle. | ||||
4971 | for (Value *V : VL) { | ||||
4972 | if (!extendSchedulingRegion(V, S)) | ||||
4973 | return None; | ||||
4974 | } | ||||
4975 | |||||
4976 | for (Value *V : VL) { | ||||
4977 | ScheduleData *BundleMember = getScheduleData(V); | ||||
4978 | assert(BundleMember &&((BundleMember && "no ScheduleData for bundle member (maybe not in same basic block)" ) ? static_cast<void> (0) : __assert_fail ("BundleMember && \"no ScheduleData for bundle member (maybe not in same basic block)\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4979, __PRETTY_FUNCTION__)) | ||||
4979 | "no ScheduleData for bundle member (maybe not in same basic block)")((BundleMember && "no ScheduleData for bundle member (maybe not in same basic block)" ) ? static_cast<void> (0) : __assert_fail ("BundleMember && \"no ScheduleData for bundle member (maybe not in same basic block)\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4979, __PRETTY_FUNCTION__)); | ||||
4980 | if (BundleMember->IsScheduled) { | ||||
4981 | // A bundle member was scheduled as single instruction before and now | ||||
4982 | // needs to be scheduled as part of the bundle. We just get rid of the | ||||
4983 | // existing schedule. | ||||
4984 | 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) | ||||
4985 | << " was already scheduled\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: reset schedule because " << *BundleMember << " was already scheduled\n"; } } while (false); | ||||
4986 | ReSchedule = true; | ||||
4987 | } | ||||
4988 | assert(BundleMember->isSchedulingEntity() &&((BundleMember->isSchedulingEntity() && "bundle member already part of other bundle" ) ? static_cast<void> (0) : __assert_fail ("BundleMember->isSchedulingEntity() && \"bundle member already part of other bundle\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4989, __PRETTY_FUNCTION__)) | ||||
4989 | "bundle member already part of other bundle")((BundleMember->isSchedulingEntity() && "bundle member already part of other bundle" ) ? static_cast<void> (0) : __assert_fail ("BundleMember->isSchedulingEntity() && \"bundle member already part of other bundle\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4989, __PRETTY_FUNCTION__)); | ||||
4990 | if (PrevInBundle) { | ||||
4991 | PrevInBundle->NextInBundle = BundleMember; | ||||
4992 | } else { | ||||
4993 | Bundle = BundleMember; | ||||
4994 | } | ||||
4995 | BundleMember->UnscheduledDepsInBundle = 0; | ||||
4996 | Bundle->UnscheduledDepsInBundle += BundleMember->UnscheduledDeps; | ||||
4997 | |||||
4998 | // Group the instructions to a bundle. | ||||
4999 | BundleMember->FirstInBundle = Bundle; | ||||
5000 | PrevInBundle = BundleMember; | ||||
5001 | } | ||||
5002 | if (ScheduleEnd != OldScheduleEnd) { | ||||
5003 | // The scheduling region got new instructions at the lower end (or it is a | ||||
5004 | // new region for the first bundle). This makes it necessary to | ||||
5005 | // recalculate all dependencies. | ||||
5006 | // It is seldom that this needs to be done a second time after adding the | ||||
5007 | // initial bundle to the region. | ||||
5008 | for (auto *I = ScheduleStart; I != ScheduleEnd; I = I->getNextNode()) { | ||||
5009 | doForAllOpcodes(I, [](ScheduleData *SD) { | ||||
5010 | SD->clearDependencies(); | ||||
5011 | }); | ||||
5012 | } | ||||
5013 | ReSchedule = true; | ||||
5014 | } | ||||
5015 | if (ReSchedule) { | ||||
5016 | resetSchedule(); | ||||
5017 | initialFillReadyList(ReadyInsts); | ||||
5018 | } | ||||
5019 | assert(Bundle && "Failed to find schedule bundle")((Bundle && "Failed to find schedule bundle") ? static_cast <void> (0) : __assert_fail ("Bundle && \"Failed to find schedule bundle\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5019, __PRETTY_FUNCTION__)); | ||||
5020 | |||||
5021 | LLVM_DEBUG(dbgs() << "SLP: try schedule bundle " << *Bundle << " in block "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: try schedule bundle " << *Bundle << " in block " << BB->getName() << "\n"; } } while (false) | ||||
5022 | << BB->getName() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: try schedule bundle " << *Bundle << " in block " << BB->getName() << "\n"; } } while (false); | ||||
5023 | |||||
5024 | calculateDependencies(Bundle, true, SLP); | ||||
5025 | |||||
5026 | // Now try to schedule the new bundle. As soon as the bundle is "ready" it | ||||
5027 | // means that there are no cyclic dependencies and we can schedule it. | ||||
5028 | // Note that's important that we don't "schedule" the bundle yet (see | ||||
5029 | // cancelScheduling). | ||||
5030 | while (!Bundle->isReady() && !ReadyInsts.empty()) { | ||||
5031 | |||||
5032 | ScheduleData *pickedSD = ReadyInsts.back(); | ||||
5033 | ReadyInsts.pop_back(); | ||||
5034 | |||||
5035 | if (pickedSD->isSchedulingEntity() && pickedSD->isReady()) { | ||||
5036 | schedule(pickedSD, ReadyInsts); | ||||
5037 | } | ||||
5038 | } | ||||
5039 | if (!Bundle->isReady()) { | ||||
5040 | cancelScheduling(VL, S.OpValue); | ||||
5041 | return None; | ||||
5042 | } | ||||
5043 | return Bundle; | ||||
5044 | } | ||||
5045 | |||||
5046 | void BoUpSLP::BlockScheduling::cancelScheduling(ArrayRef<Value *> VL, | ||||
5047 | Value *OpValue) { | ||||
5048 | if (isa<PHINode>(OpValue)) | ||||
5049 | return; | ||||
5050 | |||||
5051 | ScheduleData *Bundle = getScheduleData(OpValue); | ||||
5052 | 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); | ||||
5053 | assert(!Bundle->IsScheduled &&((!Bundle->IsScheduled && "Can't cancel bundle which is already scheduled" ) ? static_cast<void> (0) : __assert_fail ("!Bundle->IsScheduled && \"Can't cancel bundle which is already scheduled\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5054, __PRETTY_FUNCTION__)) | ||||
5054 | "Can't cancel bundle which is already scheduled")((!Bundle->IsScheduled && "Can't cancel bundle which is already scheduled" ) ? static_cast<void> (0) : __assert_fail ("!Bundle->IsScheduled && \"Can't cancel bundle which is already scheduled\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5054, __PRETTY_FUNCTION__)); | ||||
5055 | assert(Bundle->isSchedulingEntity() && Bundle->isPartOfBundle() &&((Bundle->isSchedulingEntity() && Bundle->isPartOfBundle () && "tried to unbundle something which is not a bundle" ) ? static_cast<void> (0) : __assert_fail ("Bundle->isSchedulingEntity() && Bundle->isPartOfBundle() && \"tried to unbundle something which is not a bundle\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5056, __PRETTY_FUNCTION__)) | ||||
5056 | "tried to unbundle something which is not a bundle")((Bundle->isSchedulingEntity() && Bundle->isPartOfBundle () && "tried to unbundle something which is not a bundle" ) ? static_cast<void> (0) : __assert_fail ("Bundle->isSchedulingEntity() && Bundle->isPartOfBundle() && \"tried to unbundle something which is not a bundle\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5056, __PRETTY_FUNCTION__)); | ||||
5057 | |||||
5058 | // Un-bundle: make single instructions out of the bundle. | ||||
5059 | ScheduleData *BundleMember = Bundle; | ||||
5060 | while (BundleMember) { | ||||
5061 | assert(BundleMember->FirstInBundle == Bundle && "corrupt bundle links")((BundleMember->FirstInBundle == Bundle && "corrupt bundle links" ) ? static_cast<void> (0) : __assert_fail ("BundleMember->FirstInBundle == Bundle && \"corrupt bundle links\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5061, __PRETTY_FUNCTION__)); | ||||
5062 | BundleMember->FirstInBundle = BundleMember; | ||||
5063 | ScheduleData *Next = BundleMember->NextInBundle; | ||||
5064 | BundleMember->NextInBundle = nullptr; | ||||
5065 | BundleMember->UnscheduledDepsInBundle = BundleMember->UnscheduledDeps; | ||||
5066 | if (BundleMember->UnscheduledDepsInBundle == 0) { | ||||
5067 | ReadyInsts.insert(BundleMember); | ||||
5068 | } | ||||
5069 | BundleMember = Next; | ||||
5070 | } | ||||
5071 | } | ||||
5072 | |||||
5073 | BoUpSLP::ScheduleData *BoUpSLP::BlockScheduling::allocateScheduleDataChunks() { | ||||
5074 | // Allocate a new ScheduleData for the instruction. | ||||
5075 | if (ChunkPos >= ChunkSize) { | ||||
5076 | ScheduleDataChunks.push_back(std::make_unique<ScheduleData[]>(ChunkSize)); | ||||
5077 | ChunkPos = 0; | ||||
5078 | } | ||||
5079 | return &(ScheduleDataChunks.back()[ChunkPos++]); | ||||
5080 | } | ||||
5081 | |||||
5082 | bool BoUpSLP::BlockScheduling::extendSchedulingRegion(Value *V, | ||||
5083 | const InstructionsState &S) { | ||||
5084 | if (getScheduleData(V, isOneOf(S, V))) | ||||
5085 | return true; | ||||
5086 | Instruction *I = dyn_cast<Instruction>(V); | ||||
5087 | assert(I && "bundle member must be an instruction")((I && "bundle member must be an instruction") ? static_cast <void> (0) : __assert_fail ("I && \"bundle member must be an instruction\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5087, __PRETTY_FUNCTION__)); | ||||
5088 | assert(!isa<PHINode>(I) && "phi nodes don't need to be scheduled")((!isa<PHINode>(I) && "phi nodes don't need to be scheduled" ) ? static_cast<void> (0) : __assert_fail ("!isa<PHINode>(I) && \"phi nodes don't need to be scheduled\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5088, __PRETTY_FUNCTION__)); | ||||
5089 | auto &&CheckSheduleForI = [this, &S](Instruction *I) -> bool { | ||||
5090 | ScheduleData *ISD = getScheduleData(I); | ||||
5091 | if (!ISD) | ||||
5092 | return false; | ||||
5093 | assert(isInSchedulingRegion(ISD) &&((isInSchedulingRegion(ISD) && "ScheduleData not in scheduling region" ) ? static_cast<void> (0) : __assert_fail ("isInSchedulingRegion(ISD) && \"ScheduleData not in scheduling region\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5094, __PRETTY_FUNCTION__)) | ||||
5094 | "ScheduleData not in scheduling region")((isInSchedulingRegion(ISD) && "ScheduleData not in scheduling region" ) ? static_cast<void> (0) : __assert_fail ("isInSchedulingRegion(ISD) && \"ScheduleData not in scheduling region\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5094, __PRETTY_FUNCTION__)); | ||||
5095 | ScheduleData *SD = allocateScheduleDataChunks(); | ||||
5096 | SD->Inst = I; | ||||
5097 | SD->init(SchedulingRegionID, S.OpValue); | ||||
5098 | ExtraScheduleDataMap[I][S.OpValue] = SD; | ||||
5099 | return true; | ||||
5100 | }; | ||||
5101 | if (CheckSheduleForI(I)) | ||||
5102 | return true; | ||||
5103 | if (!ScheduleStart) { | ||||
5104 | // It's the first instruction in the new region. | ||||
5105 | initScheduleData(I, I->getNextNode(), nullptr, nullptr); | ||||
5106 | ScheduleStart = I; | ||||
5107 | ScheduleEnd = I->getNextNode(); | ||||
5108 | if (isOneOf(S, I) != I) | ||||
5109 | CheckSheduleForI(I); | ||||
5110 | assert(ScheduleEnd && "tried to vectorize a terminator?")((ScheduleEnd && "tried to vectorize a terminator?") ? static_cast<void> (0) : __assert_fail ("ScheduleEnd && \"tried to vectorize a terminator?\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5110, __PRETTY_FUNCTION__)); | ||||
5111 | 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); | ||||
5112 | return true; | ||||
5113 | } | ||||
5114 | // Search up and down at the same time, because we don't know if the new | ||||
5115 | // instruction is above or below the existing scheduling region. | ||||
5116 | BasicBlock::reverse_iterator UpIter = | ||||
5117 | ++ScheduleStart->getIterator().getReverse(); | ||||
5118 | BasicBlock::reverse_iterator UpperEnd = BB->rend(); | ||||
5119 | BasicBlock::iterator DownIter = ScheduleEnd->getIterator(); | ||||
5120 | BasicBlock::iterator LowerEnd = BB->end(); | ||||
5121 | while (true) { | ||||
5122 | if (++ScheduleRegionSize > ScheduleRegionSizeLimit) { | ||||
5123 | 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); | ||||
5124 | return false; | ||||
5125 | } | ||||
5126 | |||||
5127 | if (UpIter != UpperEnd) { | ||||
5128 | if (&*UpIter == I) { | ||||
5129 | initScheduleData(I, ScheduleStart, nullptr, FirstLoadStoreInRegion); | ||||
5130 | ScheduleStart = I; | ||||
5131 | if (isOneOf(S, I) != I) | ||||
5132 | CheckSheduleForI(I); | ||||
5133 | 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) | ||||
5134 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: extend schedule region start to " << *I << "\n"; } } while (false); | ||||
5135 | return true; | ||||
5136 | } | ||||
5137 | ++UpIter; | ||||
5138 | } | ||||
5139 | if (DownIter != LowerEnd) { | ||||
5140 | if (&*DownIter == I) { | ||||
5141 | initScheduleData(ScheduleEnd, I->getNextNode(), LastLoadStoreInRegion, | ||||
5142 | nullptr); | ||||
5143 | ScheduleEnd = I->getNextNode(); | ||||
5144 | if (isOneOf(S, I) != I) | ||||
5145 | CheckSheduleForI(I); | ||||
5146 | assert(ScheduleEnd && "tried to vectorize a terminator?")((ScheduleEnd && "tried to vectorize a terminator?") ? static_cast<void> (0) : __assert_fail ("ScheduleEnd && \"tried to vectorize a terminator?\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5146, __PRETTY_FUNCTION__)); | ||||
5147 | LLVM_DEBUG(dbgs() << "SLP: extend schedule region end to " << *Ido { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: extend schedule region end to " << *I << "\n"; } } while (false) | ||||
5148 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: extend schedule region end to " << *I << "\n"; } } while (false); | ||||
5149 | return true; | ||||
5150 | } | ||||
5151 | ++DownIter; | ||||
5152 | } | ||||
5153 | assert((UpIter != UpperEnd || DownIter != LowerEnd) &&(((UpIter != UpperEnd || DownIter != LowerEnd) && "instruction not found in block" ) ? static_cast<void> (0) : __assert_fail ("(UpIter != UpperEnd || DownIter != LowerEnd) && \"instruction not found in block\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5154, __PRETTY_FUNCTION__)) | ||||
5154 | "instruction not found in block")(((UpIter != UpperEnd || DownIter != LowerEnd) && "instruction not found in block" ) ? static_cast<void> (0) : __assert_fail ("(UpIter != UpperEnd || DownIter != LowerEnd) && \"instruction not found in block\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5154, __PRETTY_FUNCTION__)); | ||||
5155 | } | ||||
5156 | return true; | ||||
5157 | } | ||||
5158 | |||||
5159 | void BoUpSLP::BlockScheduling::initScheduleData(Instruction *FromI, | ||||
5160 | Instruction *ToI, | ||||
5161 | ScheduleData *PrevLoadStore, | ||||
5162 | ScheduleData *NextLoadStore) { | ||||
5163 | ScheduleData *CurrentLoadStore = PrevLoadStore; | ||||
5164 | for (Instruction *I = FromI; I != ToI; I = I->getNextNode()) { | ||||
5165 | ScheduleData *SD = ScheduleDataMap[I]; | ||||
5166 | if (!SD) { | ||||
5167 | SD = allocateScheduleDataChunks(); | ||||
5168 | ScheduleDataMap[I] = SD; | ||||
5169 | SD->Inst = I; | ||||
5170 | } | ||||
5171 | assert(!isInSchedulingRegion(SD) &&((!isInSchedulingRegion(SD) && "new ScheduleData already in scheduling region" ) ? static_cast<void> (0) : __assert_fail ("!isInSchedulingRegion(SD) && \"new ScheduleData already in scheduling region\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5172, __PRETTY_FUNCTION__)) | ||||
5172 | "new ScheduleData already in scheduling region")((!isInSchedulingRegion(SD) && "new ScheduleData already in scheduling region" ) ? static_cast<void> (0) : __assert_fail ("!isInSchedulingRegion(SD) && \"new ScheduleData already in scheduling region\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5172, __PRETTY_FUNCTION__)); | ||||
5173 | SD->init(SchedulingRegionID, I); | ||||
5174 | |||||
5175 | if (I->mayReadOrWriteMemory() && | ||||
5176 | (!isa<IntrinsicInst>(I) || | ||||
5177 | cast<IntrinsicInst>(I)->getIntrinsicID() != Intrinsic::sideeffect)) { | ||||
5178 | // Update the linked list of memory accessing instructions. | ||||
5179 | if (CurrentLoadStore) { | ||||
5180 | CurrentLoadStore->NextLoadStore = SD; | ||||
5181 | } else { | ||||
5182 | FirstLoadStoreInRegion = SD; | ||||
5183 | } | ||||
5184 | CurrentLoadStore = SD; | ||||
5185 | } | ||||
5186 | } | ||||
5187 | if (NextLoadStore) { | ||||
5188 | if (CurrentLoadStore) | ||||
5189 | CurrentLoadStore->NextLoadStore = NextLoadStore; | ||||
5190 | } else { | ||||
5191 | LastLoadStoreInRegion = CurrentLoadStore; | ||||
5192 | } | ||||
5193 | } | ||||
5194 | |||||
5195 | void BoUpSLP::BlockScheduling::calculateDependencies(ScheduleData *SD, | ||||
5196 | bool InsertInReadyList, | ||||
5197 | BoUpSLP *SLP) { | ||||
5198 | assert(SD->isSchedulingEntity())((SD->isSchedulingEntity()) ? static_cast<void> (0) : __assert_fail ("SD->isSchedulingEntity()", "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5198, __PRETTY_FUNCTION__)); | ||||
5199 | |||||
5200 | SmallVector<ScheduleData *, 10> WorkList; | ||||
5201 | WorkList.push_back(SD); | ||||
5202 | |||||
5203 | while (!WorkList.empty()) { | ||||
5204 | ScheduleData *SD = WorkList.back(); | ||||
5205 | WorkList.pop_back(); | ||||
5206 | |||||
5207 | ScheduleData *BundleMember = SD; | ||||
5208 | while (BundleMember) { | ||||
5209 | assert(isInSchedulingRegion(BundleMember))((isInSchedulingRegion(BundleMember)) ? static_cast<void> (0) : __assert_fail ("isInSchedulingRegion(BundleMember)", "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5209, __PRETTY_FUNCTION__)); | ||||
5210 | if (!BundleMember->hasValidDependencies()) { | ||||
5211 | |||||
5212 | LLVM_DEBUG(dbgs() << "SLP: update deps of " << *BundleMemberdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: update deps of " << *BundleMember << "\n"; } } while (false) | ||||
5213 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: update deps of " << *BundleMember << "\n"; } } while (false); | ||||
5214 | BundleMember->Dependencies = 0; | ||||
5215 | BundleMember->resetUnscheduledDeps(); | ||||
5216 | |||||
5217 | // Handle def-use chain dependencies. | ||||
5218 | if (BundleMember->OpValue != BundleMember->Inst) { | ||||
5219 | ScheduleData *UseSD = getScheduleData(BundleMember->Inst); | ||||
5220 | if (UseSD && isInSchedulingRegion(UseSD->FirstInBundle)) { | ||||
5221 | BundleMember->Dependencies++; | ||||
5222 | ScheduleData *DestBundle = UseSD->FirstInBundle; | ||||
5223 | if (!DestBundle->IsScheduled) | ||||
5224 | BundleMember->incrementUnscheduledDeps(1); | ||||
5225 | if (!DestBundle->hasValidDependencies()) | ||||
5226 | WorkList.push_back(DestBundle); | ||||
5227 | } | ||||
5228 | } else { | ||||
5229 | for (User *U : BundleMember->Inst->users()) { | ||||
5230 | if (isa<Instruction>(U)) { | ||||
5231 | ScheduleData *UseSD = getScheduleData(U); | ||||
5232 | if (UseSD && isInSchedulingRegion(UseSD->FirstInBundle)) { | ||||
5233 | BundleMember->Dependencies++; | ||||
5234 | ScheduleData *DestBundle = UseSD->FirstInBundle; | ||||
5235 | if (!DestBundle->IsScheduled) | ||||
5236 | BundleMember->incrementUnscheduledDeps(1); | ||||
5237 | if (!DestBundle->hasValidDependencies()) | ||||
5238 | WorkList.push_back(DestBundle); | ||||
5239 | } | ||||
5240 | } else { | ||||
5241 | // I'm not sure if this can ever happen. But we need to be safe. | ||||
5242 | // This lets the instruction/bundle never be scheduled and | ||||
5243 | // eventually disable vectorization. | ||||
5244 | BundleMember->Dependencies++; | ||||
5245 | BundleMember->incrementUnscheduledDeps(1); | ||||
5246 | } | ||||
5247 | } | ||||
5248 | } | ||||
5249 | |||||
5250 | // Handle the memory dependencies. | ||||
5251 | ScheduleData *DepDest = BundleMember->NextLoadStore; | ||||
5252 | if (DepDest) { | ||||
5253 | Instruction *SrcInst = BundleMember->Inst; | ||||
5254 | MemoryLocation SrcLoc = getLocation(SrcInst, SLP->AA); | ||||
5255 | bool SrcMayWrite = BundleMember->Inst->mayWriteToMemory(); | ||||
5256 | unsigned numAliased = 0; | ||||
5257 | unsigned DistToSrc = 1; | ||||
5258 | |||||
5259 | while (DepDest) { | ||||
5260 | assert(isInSchedulingRegion(DepDest))((isInSchedulingRegion(DepDest)) ? static_cast<void> (0 ) : __assert_fail ("isInSchedulingRegion(DepDest)", "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5260, __PRETTY_FUNCTION__)); | ||||
5261 | |||||
5262 | // We have two limits to reduce the complexity: | ||||
5263 | // 1) AliasedCheckLimit: It's a small limit to reduce calls to | ||||
5264 | // SLP->isAliased (which is the expensive part in this loop). | ||||
5265 | // 2) MaxMemDepDistance: It's for very large blocks and it aborts | ||||
5266 | // the whole loop (even if the loop is fast, it's quadratic). | ||||
5267 | // It's important for the loop break condition (see below) to | ||||
5268 | // check this limit even between two read-only instructions. | ||||
5269 | if (DistToSrc >= MaxMemDepDistance || | ||||
5270 | ((SrcMayWrite || DepDest->Inst->mayWriteToMemory()) && | ||||
5271 | (numAliased >= AliasedCheckLimit || | ||||
5272 | SLP->isAliased(SrcLoc, SrcInst, DepDest->Inst)))) { | ||||
5273 | |||||
5274 | // We increment the counter only if the locations are aliased | ||||
5275 | // (instead of counting all alias checks). This gives a better | ||||
5276 | // balance between reduced runtime and accurate dependencies. | ||||
5277 | numAliased++; | ||||
5278 | |||||
5279 | DepDest->MemoryDependencies.push_back(BundleMember); | ||||
5280 | BundleMember->Dependencies++; | ||||
5281 | ScheduleData *DestBundle = DepDest->FirstInBundle; | ||||
5282 | if (!DestBundle->IsScheduled) { | ||||
5283 | BundleMember->incrementUnscheduledDeps(1); | ||||
5284 | } | ||||
5285 | if (!DestBundle->hasValidDependencies()) { | ||||
5286 | WorkList.push_back(DestBundle); | ||||
5287 | } | ||||
5288 | } | ||||
5289 | DepDest = DepDest->NextLoadStore; | ||||
5290 | |||||
5291 | // Example, explaining the loop break condition: Let's assume our | ||||
5292 | // starting instruction is i0 and MaxMemDepDistance = 3. | ||||
5293 | // | ||||
5294 | // +--------v--v--v | ||||
5295 | // i0,i1,i2,i3,i4,i5,i6,i7,i8 | ||||
5296 | // +--------^--^--^ | ||||
5297 | // | ||||
5298 | // MaxMemDepDistance let us stop alias-checking at i3 and we add | ||||
5299 | // dependencies from i0 to i3,i4,.. (even if they are not aliased). | ||||
5300 | // Previously we already added dependencies from i3 to i6,i7,i8 | ||||
5301 | // (because of MaxMemDepDistance). As we added a dependency from | ||||
5302 | // i0 to i3, we have transitive dependencies from i0 to i6,i7,i8 | ||||
5303 | // and we can abort this loop at i6. | ||||
5304 | if (DistToSrc >= 2 * MaxMemDepDistance) | ||||
5305 | break; | ||||
5306 | DistToSrc++; | ||||
5307 | } | ||||
5308 | } | ||||
5309 | } | ||||
5310 | BundleMember = BundleMember->NextInBundle; | ||||
5311 | } | ||||
5312 | if (InsertInReadyList && SD->isReady()) { | ||||
5313 | ReadyInsts.push_back(SD); | ||||
5314 | 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) | ||||
5315 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: gets ready on update: " << *SD->Inst << "\n"; } } while (false); | ||||
5316 | } | ||||
5317 | } | ||||
5318 | } | ||||
5319 | |||||
5320 | void BoUpSLP::BlockScheduling::resetSchedule() { | ||||
5321 | assert(ScheduleStart &&((ScheduleStart && "tried to reset schedule on block which has not been scheduled" ) ? static_cast<void> (0) : __assert_fail ("ScheduleStart && \"tried to reset schedule on block which has not been scheduled\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5322, __PRETTY_FUNCTION__)) | ||||
5322 | "tried to reset schedule on block which has not been scheduled")((ScheduleStart && "tried to reset schedule on block which has not been scheduled" ) ? static_cast<void> (0) : __assert_fail ("ScheduleStart && \"tried to reset schedule on block which has not been scheduled\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5322, __PRETTY_FUNCTION__)); | ||||
5323 | for (Instruction *I = ScheduleStart; I != ScheduleEnd; I = I->getNextNode()) { | ||||
5324 | doForAllOpcodes(I, [&](ScheduleData *SD) { | ||||
5325 | assert(isInSchedulingRegion(SD) &&((isInSchedulingRegion(SD) && "ScheduleData not in scheduling region" ) ? static_cast<void> (0) : __assert_fail ("isInSchedulingRegion(SD) && \"ScheduleData not in scheduling region\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5326, __PRETTY_FUNCTION__)) | ||||
5326 | "ScheduleData not in scheduling region")((isInSchedulingRegion(SD) && "ScheduleData not in scheduling region" ) ? static_cast<void> (0) : __assert_fail ("isInSchedulingRegion(SD) && \"ScheduleData not in scheduling region\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5326, __PRETTY_FUNCTION__)); | ||||
5327 | SD->IsScheduled = false; | ||||
5328 | SD->resetUnscheduledDeps(); | ||||
5329 | }); | ||||
5330 | } | ||||
5331 | ReadyInsts.clear(); | ||||
5332 | } | ||||
5333 | |||||
5334 | void BoUpSLP::scheduleBlock(BlockScheduling *BS) { | ||||
5335 | if (!BS->ScheduleStart) | ||||
5336 | return; | ||||
5337 | |||||
5338 | 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); | ||||
5339 | |||||
5340 | BS->resetSchedule(); | ||||
5341 | |||||
5342 | // For the real scheduling we use a more sophisticated ready-list: it is | ||||
5343 | // sorted by the original instruction location. This lets the final schedule | ||||
5344 | // be as close as possible to the original instruction order. | ||||
5345 | struct ScheduleDataCompare { | ||||
5346 | bool operator()(ScheduleData *SD1, ScheduleData *SD2) const { | ||||
5347 | return SD2->SchedulingPriority < SD1->SchedulingPriority; | ||||
5348 | } | ||||
5349 | }; | ||||
5350 | std::set<ScheduleData *, ScheduleDataCompare> ReadyInsts; | ||||
5351 | |||||
5352 | // Ensure that all dependency data is updated and fill the ready-list with | ||||
5353 | // initial instructions. | ||||
5354 | int Idx = 0; | ||||
5355 | int NumToSchedule = 0; | ||||
5356 | for (auto *I = BS->ScheduleStart; I != BS->ScheduleEnd; | ||||
5357 | I = I->getNextNode()) { | ||||
5358 | BS->doForAllOpcodes(I, [this, &Idx, &NumToSchedule, BS](ScheduleData *SD) { | ||||
5359 | assert(SD->isPartOfBundle() ==((SD->isPartOfBundle() == (getTreeEntry(SD->Inst) != nullptr ) && "scheduler and vectorizer bundle mismatch") ? static_cast <void> (0) : __assert_fail ("SD->isPartOfBundle() == (getTreeEntry(SD->Inst) != nullptr) && \"scheduler and vectorizer bundle mismatch\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5361, __PRETTY_FUNCTION__)) | ||||
5360 | (getTreeEntry(SD->Inst) != nullptr) &&((SD->isPartOfBundle() == (getTreeEntry(SD->Inst) != nullptr ) && "scheduler and vectorizer bundle mismatch") ? static_cast <void> (0) : __assert_fail ("SD->isPartOfBundle() == (getTreeEntry(SD->Inst) != nullptr) && \"scheduler and vectorizer bundle mismatch\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5361, __PRETTY_FUNCTION__)) | ||||
5361 | "scheduler and vectorizer bundle mismatch")((SD->isPartOfBundle() == (getTreeEntry(SD->Inst) != nullptr ) && "scheduler and vectorizer bundle mismatch") ? static_cast <void> (0) : __assert_fail ("SD->isPartOfBundle() == (getTreeEntry(SD->Inst) != nullptr) && \"scheduler and vectorizer bundle mismatch\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5361, __PRETTY_FUNCTION__)); | ||||
5362 | SD->FirstInBundle->SchedulingPriority = Idx++; | ||||
5363 | if (SD->isSchedulingEntity()) { | ||||
5364 | BS->calculateDependencies(SD, false, this); | ||||
5365 | NumToSchedule++; | ||||
5366 | } | ||||
5367 | }); | ||||
5368 | } | ||||
5369 | BS->initialFillReadyList(ReadyInsts); | ||||
5370 | |||||
5371 | Instruction *LastScheduledInst = BS->ScheduleEnd; | ||||
5372 | |||||
5373 | // Do the "real" scheduling. | ||||
5374 | while (!ReadyInsts.empty()) { | ||||
5375 | ScheduleData *picked = *ReadyInsts.begin(); | ||||
5376 | ReadyInsts.erase(ReadyInsts.begin()); | ||||
5377 | |||||
5378 | // Move the scheduled instruction(s) to their dedicated places, if not | ||||
5379 | // there yet. | ||||
5380 | ScheduleData *BundleMember = picked; | ||||
5381 | while (BundleMember) { | ||||
5382 | Instruction *pickedInst = BundleMember->Inst; | ||||
5383 | if (LastScheduledInst->getNextNode() != pickedInst) { | ||||
5384 | BS->BB->getInstList().remove(pickedInst); | ||||
5385 | BS->BB->getInstList().insert(LastScheduledInst->getIterator(), | ||||
5386 | pickedInst); | ||||
5387 | } | ||||
5388 | LastScheduledInst = pickedInst; | ||||
5389 | BundleMember = BundleMember->NextInBundle; | ||||
5390 | } | ||||
5391 | |||||
5392 | BS->schedule(picked, ReadyInsts); | ||||
5393 | NumToSchedule--; | ||||
5394 | } | ||||
5395 | assert(NumToSchedule == 0 && "could not schedule all instructions")((NumToSchedule == 0 && "could not schedule all instructions" ) ? static_cast<void> (0) : __assert_fail ("NumToSchedule == 0 && \"could not schedule all instructions\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5395, __PRETTY_FUNCTION__)); | ||||
5396 | |||||
5397 | // Avoid duplicate scheduling of the block. | ||||
5398 | BS->ScheduleStart = nullptr; | ||||
5399 | } | ||||
5400 | |||||
5401 | unsigned BoUpSLP::getVectorElementSize(Value *V) { | ||||
5402 | // If V is a store, just return the width of the stored value without | ||||
5403 | // traversing the expression tree. This is the common case. | ||||
5404 | if (auto *Store = dyn_cast<StoreInst>(V)) | ||||
5405 | return DL->getTypeSizeInBits(Store->getValueOperand()->getType()); | ||||
5406 | |||||
5407 | auto E = InstrElementSize.find(V); | ||||
5408 | if (E != InstrElementSize.end()) | ||||
5409 | return E->second; | ||||
5410 | |||||
5411 | // If V is not a store, we can traverse the expression tree to find loads | ||||
5412 | // that feed it. The type of the loaded value may indicate a more suitable | ||||
5413 | // width than V's type. We want to base the vector element size on the width | ||||
5414 | // of memory operations where possible. | ||||
5415 | SmallVector<Instruction *, 16> Worklist; | ||||
5416 | SmallPtrSet<Instruction *, 16> Visited; | ||||
5417 | if (auto *I = dyn_cast<Instruction>(V)) { | ||||
5418 | Worklist.push_back(I); | ||||
5419 | Visited.insert(I); | ||||
5420 | } | ||||
5421 | |||||
5422 | // Traverse the expression tree in bottom-up order looking for loads. If we | ||||
5423 | // encounter an instruction we don't yet handle, we give up. | ||||
5424 | auto MaxWidth = 0u; | ||||
5425 | auto FoundUnknownInst = false; | ||||
5426 | while (!Worklist.empty() && !FoundUnknownInst) { | ||||
5427 | auto *I = Worklist.pop_back_val(); | ||||
5428 | |||||
5429 | // We should only be looking at scalar instructions here. If the current | ||||
5430 | // instruction has a vector type, give up. | ||||
5431 | auto *Ty = I->getType(); | ||||
5432 | if (isa<VectorType>(Ty)) | ||||
5433 | FoundUnknownInst = true; | ||||
5434 | |||||
5435 | // If the current instruction is a load, update MaxWidth to reflect the | ||||
5436 | // width of the loaded value. | ||||
5437 | else if (isa<LoadInst>(I)) | ||||
5438 | MaxWidth = std::max<unsigned>(MaxWidth, DL->getTypeSizeInBits(Ty)); | ||||
5439 | |||||
5440 | // Otherwise, we need to visit the operands of the instruction. We only | ||||
5441 | // handle the interesting cases from buildTree here. If an operand is an | ||||
5442 | // instruction we haven't yet visited, we add it to the worklist. | ||||
5443 | else if (isa<PHINode>(I) || isa<CastInst>(I) || isa<GetElementPtrInst>(I) || | ||||
5444 | isa<CmpInst>(I) || isa<SelectInst>(I) || isa<BinaryOperator>(I)) { | ||||
5445 | for (Use &U : I->operands()) | ||||
5446 | if (auto *J = dyn_cast<Instruction>(U.get())) | ||||
5447 | if (Visited.insert(J).second) | ||||
5448 | Worklist.push_back(J); | ||||
5449 | } | ||||
5450 | |||||
5451 | // If we don't yet handle the instruction, give up. | ||||
5452 | else | ||||
5453 | FoundUnknownInst = true; | ||||
5454 | } | ||||
5455 | |||||
5456 | int Width = MaxWidth; | ||||
5457 | // If we didn't encounter a memory access in the expression tree, or if we | ||||
5458 | // gave up for some reason, just return the width of V. Otherwise, return the | ||||
5459 | // maximum width we found. | ||||
5460 | if (!MaxWidth || FoundUnknownInst) | ||||
5461 | Width = DL->getTypeSizeInBits(V->getType()); | ||||
5462 | |||||
5463 | for (Instruction *I : Visited) | ||||
5464 | InstrElementSize[I] = Width; | ||||
5465 | |||||
5466 | return Width; | ||||
5467 | } | ||||
5468 | |||||
5469 | // Determine if a value V in a vectorizable expression Expr can be demoted to a | ||||
5470 | // smaller type with a truncation. We collect the values that will be demoted | ||||
5471 | // in ToDemote and additional roots that require investigating in Roots. | ||||
5472 | static bool collectValuesToDemote(Value *V, SmallPtrSetImpl<Value *> &Expr, | ||||
5473 | SmallVectorImpl<Value *> &ToDemote, | ||||
5474 | SmallVectorImpl<Value *> &Roots) { | ||||
5475 | // We can always demote constants. | ||||
5476 | if (isa<Constant>(V)) { | ||||
5477 | ToDemote.push_back(V); | ||||
5478 | return true; | ||||
5479 | } | ||||
5480 | |||||
5481 | // If the value is not an instruction in the expression with only one use, it | ||||
5482 | // cannot be demoted. | ||||
5483 | auto *I = dyn_cast<Instruction>(V); | ||||
5484 | if (!I || !I->hasOneUse() || !Expr.count(I)) | ||||
5485 | return false; | ||||
5486 | |||||
5487 | switch (I->getOpcode()) { | ||||
5488 | |||||
5489 | // We can always demote truncations and extensions. Since truncations can | ||||
5490 | // seed additional demotion, we save the truncated value. | ||||
5491 | case Instruction::Trunc: | ||||
5492 | Roots.push_back(I->getOperand(0)); | ||||
5493 | break; | ||||
5494 | case Instruction::ZExt: | ||||
5495 | case Instruction::SExt: | ||||
5496 | break; | ||||
5497 | |||||
5498 | // We can demote certain binary operations if we can demote both of their | ||||
5499 | // operands. | ||||
5500 | case Instruction::Add: | ||||
5501 | case Instruction::Sub: | ||||
5502 | case Instruction::Mul: | ||||
5503 | case Instruction::And: | ||||
5504 | case Instruction::Or: | ||||
5505 | case Instruction::Xor: | ||||
5506 | if (!collectValuesToDemote(I->getOperand(0), Expr, ToDemote, Roots) || | ||||
5507 | !collectValuesToDemote(I->getOperand(1), Expr, ToDemote, Roots)) | ||||
5508 | return false; | ||||
5509 | break; | ||||
5510 | |||||
5511 | // We can demote selects if we can demote their true and false values. | ||||
5512 | case Instruction::Select: { | ||||
5513 | SelectInst *SI = cast<SelectInst>(I); | ||||
5514 | if (!collectValuesToDemote(SI->getTrueValue(), Expr, ToDemote, Roots) || | ||||
5515 | !collectValuesToDemote(SI->getFalseValue(), Expr, ToDemote, Roots)) | ||||
5516 | return false; | ||||
5517 | break; | ||||
5518 | } | ||||
5519 | |||||
5520 | // We can demote phis if we can demote all their incoming operands. Note that | ||||
5521 | // we don't need to worry about cycles since we ensure single use above. | ||||
5522 | case Instruction::PHI: { | ||||
5523 | PHINode *PN = cast<PHINode>(I); | ||||
5524 | for (Value *IncValue : PN->incoming_values()) | ||||
5525 | if (!collectValuesToDemote(IncValue, Expr, ToDemote, Roots)) | ||||
5526 | return false; | ||||
5527 | break; | ||||
5528 | } | ||||
5529 | |||||
5530 | // Otherwise, conservatively give up. | ||||
5531 | default: | ||||
5532 | return false; | ||||
5533 | } | ||||
5534 | |||||
5535 | // Record the value that we can demote. | ||||
5536 | ToDemote.push_back(V); | ||||
5537 | return true; | ||||
5538 | } | ||||
5539 | |||||
5540 | void BoUpSLP::computeMinimumValueSizes() { | ||||
5541 | // If there are no external uses, the expression tree must be rooted by a | ||||
5542 | // store. We can't demote in-memory values, so there is nothing to do here. | ||||
5543 | if (ExternalUses.empty()) | ||||
5544 | return; | ||||
5545 | |||||
5546 | // We only attempt to truncate integer expressions. | ||||
5547 | auto &TreeRoot = VectorizableTree[0]->Scalars; | ||||
5548 | auto *TreeRootIT = dyn_cast<IntegerType>(TreeRoot[0]->getType()); | ||||
5549 | if (!TreeRootIT) | ||||
5550 | return; | ||||
5551 | |||||
5552 | // If the expression is not rooted by a store, these roots should have | ||||
5553 | // external uses. We will rely on InstCombine to rewrite the expression in | ||||
5554 | // the narrower type. However, InstCombine only rewrites single-use values. | ||||
5555 | // This means that if a tree entry other than a root is used externally, it | ||||
5556 | // must have multiple uses and InstCombine will not rewrite it. The code | ||||
5557 | // below ensures that only the roots are used externally. | ||||
5558 | SmallPtrSet<Value *, 32> Expr(TreeRoot.begin(), TreeRoot.end()); | ||||
5559 | for (auto &EU : ExternalUses) | ||||
5560 | if (!Expr.erase(EU.Scalar)) | ||||
5561 | return; | ||||
5562 | if (!Expr.empty()) | ||||
5563 | return; | ||||
5564 | |||||
5565 | // Collect the scalar values of the vectorizable expression. We will use this | ||||
5566 | // context to determine which values can be demoted. If we see a truncation, | ||||
5567 | // we mark it as seeding another demotion. | ||||
5568 | for (auto &EntryPtr : VectorizableTree) | ||||
5569 | Expr.insert(EntryPtr->Scalars.begin(), EntryPtr->Scalars.end()); | ||||
5570 | |||||
5571 | // Ensure the roots of the vectorizable tree don't form a cycle. They must | ||||
5572 | // have a single external user that is not in the vectorizable tree. | ||||
5573 | for (auto *Root : TreeRoot) | ||||
5574 | if (!Root->hasOneUse() || Expr.count(*Root->user_begin())) | ||||
5575 | return; | ||||
5576 | |||||
5577 | // Conservatively determine if we can actually truncate the roots of the | ||||
5578 | // expression. Collect the values that can be demoted in ToDemote and | ||||
5579 | // additional roots that require investigating in Roots. | ||||
5580 | SmallVector<Value *, 32> ToDemote; | ||||
5581 | SmallVector<Value *, 4> Roots; | ||||
5582 | for (auto *Root : TreeRoot) | ||||
5583 | if (!collectValuesToDemote(Root, Expr, ToDemote, Roots)) | ||||
5584 | return; | ||||
5585 | |||||
5586 | // The maximum bit width required to represent all the values that can be | ||||
5587 | // demoted without loss of precision. It would be safe to truncate the roots | ||||
5588 | // of the expression to this width. | ||||
5589 | auto MaxBitWidth = 8u; | ||||
5590 | |||||
5591 | // We first check if all the bits of the roots are demanded. If they're not, | ||||
5592 | // we can truncate the roots to this narrower type. | ||||
5593 | for (auto *Root : TreeRoot) { | ||||
5594 | auto Mask = DB->getDemandedBits(cast<Instruction>(Root)); | ||||
5595 | MaxBitWidth = std::max<unsigned>( | ||||
5596 | Mask.getBitWidth() - Mask.countLeadingZeros(), MaxBitWidth); | ||||
5597 | } | ||||
5598 | |||||
5599 | // True if the roots can be zero-extended back to their original type, rather | ||||
5600 | // than sign-extended. We know that if the leading bits are not demanded, we | ||||
5601 | // can safely zero-extend. So we initialize IsKnownPositive to True. | ||||
5602 | bool IsKnownPositive = true; | ||||
5603 | |||||
5604 | // If all the bits of the roots are demanded, we can try a little harder to | ||||
5605 | // compute a narrower type. This can happen, for example, if the roots are | ||||
5606 | // getelementptr indices. InstCombine promotes these indices to the pointer | ||||
5607 | // width. Thus, all their bits are technically demanded even though the | ||||
5608 | // address computation might be vectorized in a smaller type. | ||||
5609 | // | ||||
5610 | // We start by looking at each entry that can be demoted. We compute the | ||||
5611 | // maximum bit width required to store the scalar by using ValueTracking to | ||||
5612 | // compute the number of high-order bits we can truncate. | ||||
5613 | if (MaxBitWidth == DL->getTypeSizeInBits(TreeRoot[0]->getType()) && | ||||
5614 | llvm::all_of(TreeRoot, [](Value *R) { | ||||
5615 | assert(R->hasOneUse() && "Root should have only one use!")((R->hasOneUse() && "Root should have only one use!" ) ? static_cast<void> (0) : __assert_fail ("R->hasOneUse() && \"Root should have only one use!\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5615, __PRETTY_FUNCTION__)); | ||||
5616 | return isa<GetElementPtrInst>(R->user_back()); | ||||
5617 | })) { | ||||
5618 | MaxBitWidth = 8u; | ||||
5619 | |||||
5620 | // Determine if the sign bit of all the roots is known to be zero. If not, | ||||
5621 | // IsKnownPositive is set to False. | ||||
5622 | IsKnownPositive = llvm::all_of(TreeRoot, [&](Value *R) { | ||||
5623 | KnownBits Known = computeKnownBits(R, *DL); | ||||
5624 | return Known.isNonNegative(); | ||||
5625 | }); | ||||
5626 | |||||
5627 | // Determine the maximum number of bits required to store the scalar | ||||
5628 | // values. | ||||
5629 | for (auto *Scalar : ToDemote) { | ||||
5630 | auto NumSignBits = ComputeNumSignBits(Scalar, *DL, 0, AC, nullptr, DT); | ||||
5631 | auto NumTypeBits = DL->getTypeSizeInBits(Scalar->getType()); | ||||
5632 | MaxBitWidth = std::max<unsigned>(NumTypeBits - NumSignBits, MaxBitWidth); | ||||
5633 | } | ||||
5634 | |||||
5635 | // If we can't prove that the sign bit is zero, we must add one to the | ||||
5636 | // maximum bit width to account for the unknown sign bit. This preserves | ||||
5637 | // the existing sign bit so we can safely sign-extend the root back to the | ||||
5638 | // original type. Otherwise, if we know the sign bit is zero, we will | ||||
5639 | // zero-extend the root instead. | ||||
5640 | // | ||||
5641 | // FIXME: This is somewhat suboptimal, as there will be cases where adding | ||||
5642 | // one to the maximum bit width will yield a larger-than-necessary | ||||
5643 | // type. In general, we need to add an extra bit only if we can't | ||||
5644 | // prove that the upper bit of the original type is equal to the | ||||
5645 | // upper bit of the proposed smaller type. If these two bits are the | ||||
5646 | // same (either zero or one) we know that sign-extending from the | ||||
5647 | // smaller type will result in the same value. Here, since we can't | ||||
5648 | // yet prove this, we are just making the proposed smaller type | ||||
5649 | // larger to ensure correctness. | ||||
5650 | if (!IsKnownPositive) | ||||
5651 | ++MaxBitWidth; | ||||
5652 | } | ||||
5653 | |||||
5654 | // Round MaxBitWidth up to the next power-of-two. | ||||
5655 | if (!isPowerOf2_64(MaxBitWidth)) | ||||
5656 | MaxBitWidth = NextPowerOf2(MaxBitWidth); | ||||
5657 | |||||
5658 | // If the maximum bit width we compute is less than the with of the roots' | ||||
5659 | // type, we can proceed with the narrowing. Otherwise, do nothing. | ||||
5660 | if (MaxBitWidth >= TreeRootIT->getBitWidth()) | ||||
5661 | return; | ||||
5662 | |||||
5663 | // If we can truncate the root, we must collect additional values that might | ||||
5664 | // be demoted as a result. That is, those seeded by truncations we will | ||||
5665 | // modify. | ||||
5666 | while (!Roots.empty()) | ||||
5667 | collectValuesToDemote(Roots.pop_back_val(), Expr, ToDemote, Roots); | ||||
5668 | |||||
5669 | // Finally, map the values we can demote to the maximum bit with we computed. | ||||
5670 | for (auto *Scalar : ToDemote) | ||||
5671 | MinBWs[Scalar] = std::make_pair(MaxBitWidth, !IsKnownPositive); | ||||
5672 | } | ||||
5673 | |||||
5674 | namespace { | ||||
5675 | |||||
5676 | /// The SLPVectorizer Pass. | ||||
5677 | struct SLPVectorizer : public FunctionPass { | ||||
5678 | SLPVectorizerPass Impl; | ||||
5679 | |||||
5680 | /// Pass identification, replacement for typeid | ||||
5681 | static char ID; | ||||
5682 | |||||
5683 | explicit SLPVectorizer() : FunctionPass(ID) { | ||||
5684 | initializeSLPVectorizerPass(*PassRegistry::getPassRegistry()); | ||||
5685 | } | ||||
5686 | |||||
5687 | bool doInitialization(Module &M) override { | ||||
5688 | return false; | ||||
5689 | } | ||||
5690 | |||||
5691 | bool runOnFunction(Function &F) override { | ||||
5692 | if (skipFunction(F)) | ||||
| |||||
5693 | return false; | ||||
5694 | |||||
5695 | auto *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE(); | ||||
5696 | auto *TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F); | ||||
5697 | auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>(); | ||||
5698 | auto *TLI = TLIP
| ||||
5699 | auto *AA = &getAnalysis<AAResultsWrapperPass>().getAAResults(); | ||||
5700 | auto *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); | ||||
5701 | auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); | ||||
5702 | auto *AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F); | ||||
5703 | auto *DB = &getAnalysis<DemandedBitsWrapperPass>().getDemandedBits(); | ||||
5704 | auto *ORE = &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE(); | ||||
5705 | |||||
5706 | return Impl.runImpl(F, SE, TTI, TLI, AA, LI, DT, AC, DB, ORE); | ||||
5707 | } | ||||
5708 | |||||
5709 | void getAnalysisUsage(AnalysisUsage &AU) const override { | ||||
5710 | FunctionPass::getAnalysisUsage(AU); | ||||
5711 | AU.addRequired<AssumptionCacheTracker>(); | ||||
5712 | AU.addRequired<ScalarEvolutionWrapperPass>(); | ||||
5713 | AU.addRequired<AAResultsWrapperPass>(); | ||||
5714 | AU.addRequired<TargetTransformInfoWrapperPass>(); | ||||
5715 | AU.addRequired<LoopInfoWrapperPass>(); | ||||
5716 | AU.addRequired<DominatorTreeWrapperPass>(); | ||||
5717 | AU.addRequired<DemandedBitsWrapperPass>(); | ||||
5718 | AU.addRequired<OptimizationRemarkEmitterWrapperPass>(); | ||||
5719 | AU.addRequired<InjectTLIMappingsLegacy>(); | ||||
5720 | AU.addPreserved<LoopInfoWrapperPass>(); | ||||
5721 | AU.addPreserved<DominatorTreeWrapperPass>(); | ||||
5722 | AU.addPreserved<AAResultsWrapperPass>(); | ||||
5723 | AU.addPreserved<GlobalsAAWrapperPass>(); | ||||
5724 | AU.setPreservesCFG(); | ||||
5725 | } | ||||
5726 | }; | ||||
5727 | |||||
5728 | } // end anonymous namespace | ||||
5729 | |||||
5730 | PreservedAnalyses SLPVectorizerPass::run(Function &F, FunctionAnalysisManager &AM) { | ||||
5731 | auto *SE = &AM.getResult<ScalarEvolutionAnalysis>(F); | ||||
5732 | auto *TTI = &AM.getResult<TargetIRAnalysis>(F); | ||||
5733 | auto *TLI = AM.getCachedResult<TargetLibraryAnalysis>(F); | ||||
5734 | auto *AA = &AM.getResult<AAManager>(F); | ||||
5735 | auto *LI = &AM.getResult<LoopAnalysis>(F); | ||||
5736 | auto *DT = &AM.getResult<DominatorTreeAnalysis>(F); | ||||
5737 | auto *AC = &AM.getResult<AssumptionAnalysis>(F); | ||||
5738 | auto *DB = &AM.getResult<DemandedBitsAnalysis>(F); | ||||
5739 | auto *ORE = &AM.getResult<OptimizationRemarkEmitterAnalysis>(F); | ||||
5740 | |||||
5741 | bool Changed = runImpl(F, SE, TTI, TLI, AA, LI, DT, AC, DB, ORE); | ||||
5742 | if (!Changed) | ||||
5743 | return PreservedAnalyses::all(); | ||||
5744 | |||||
5745 | PreservedAnalyses PA; | ||||
5746 | PA.preserveSet<CFGAnalyses>(); | ||||
5747 | PA.preserve<AAManager>(); | ||||
5748 | PA.preserve<GlobalsAA>(); | ||||
5749 | return PA; | ||||
5750 | } | ||||
5751 | |||||
5752 | bool SLPVectorizerPass::runImpl(Function &F, ScalarEvolution *SE_, | ||||
5753 | TargetTransformInfo *TTI_, | ||||
5754 | TargetLibraryInfo *TLI_, AAResults *AA_, | ||||
5755 | LoopInfo *LI_, DominatorTree *DT_, | ||||
5756 | AssumptionCache *AC_, DemandedBits *DB_, | ||||
5757 | OptimizationRemarkEmitter *ORE_) { | ||||
5758 | if (!RunSLPVectorization) | ||||
5759 | return false; | ||||
5760 | SE = SE_; | ||||
5761 | TTI = TTI_; | ||||
5762 | TLI = TLI_; | ||||
5763 | AA = AA_; | ||||
5764 | LI = LI_; | ||||
5765 | DT = DT_; | ||||
5766 | AC = AC_; | ||||
5767 | DB = DB_; | ||||
5768 | DL = &F.getParent()->getDataLayout(); | ||||
5769 | |||||
5770 | Stores.clear(); | ||||
5771 | GEPs.clear(); | ||||
5772 | bool Changed = false; | ||||
5773 | |||||
5774 | // If the target claims to have no vector registers don't attempt | ||||
5775 | // vectorization. | ||||
5776 | if (!TTI->getNumberOfRegisters(TTI->getRegisterClassForType(true))) | ||||
5777 | return false; | ||||
5778 | |||||
5779 | // Don't vectorize when the attribute NoImplicitFloat is used. | ||||
5780 | if (F.hasFnAttribute(Attribute::NoImplicitFloat)) | ||||
5781 | return false; | ||||
5782 | |||||
5783 | 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); | ||||
5784 | |||||
5785 | // Use the bottom up slp vectorizer to construct chains that start with | ||||
5786 | // store instructions. | ||||
5787 | BoUpSLP R(&F, SE, TTI, TLI, AA, LI, DT, AC, DB, DL, ORE_); | ||||
5788 | |||||
5789 | // A general note: the vectorizer must use BoUpSLP::eraseInstruction() to | ||||
5790 | // delete instructions. | ||||
5791 | |||||
5792 | // Scan the blocks in the function in post order. | ||||
5793 | for (auto BB : post_order(&F.getEntryBlock())) { | ||||
5794 | collectSeedInstructions(BB); | ||||
5795 | |||||
5796 | // Vectorize trees that end at stores. | ||||
5797 | if (!Stores.empty()) { | ||||
5798 | 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) | ||||
5799 | << " underlying objects.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Found stores for " << Stores .size() << " underlying objects.\n"; } } while (false); | ||||
5800 | Changed |= vectorizeStoreChains(R); | ||||
5801 | } | ||||
5802 | |||||
5803 | // Vectorize trees that end at reductions. | ||||
5804 | Changed |= vectorizeChainsInBlock(BB, R); | ||||
5805 | |||||
5806 | // Vectorize the index computations of getelementptr instructions. This | ||||
5807 | // is primarily intended to catch gather-like idioms ending at | ||||
5808 | // non-consecutive loads. | ||||
5809 | if (!GEPs.empty()) { | ||||
5810 | 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) | ||||
5811 | << " underlying objects.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Found GEPs for " << GEPs .size() << " underlying objects.\n"; } } while (false); | ||||
5812 | Changed |= vectorizeGEPIndices(BB, R); | ||||
5813 | } | ||||
5814 | } | ||||
5815 | |||||
5816 | if (Changed) { | ||||
5817 | R.optimizeGatherSequence(); | ||||
5818 | LLVM_DEBUG(dbgs() << "SLP: vectorized \"" << F.getName() << "\"\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: vectorized \"" << F.getName () << "\"\n"; } } while (false); | ||||
5819 | } | ||||
5820 | return Changed; | ||||
5821 | } | ||||
5822 | |||||
5823 | bool SLPVectorizerPass::vectorizeStoreChain(ArrayRef<Value *> Chain, BoUpSLP &R, | ||||
5824 | unsigned Idx) { | ||||
5825 | 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) | ||||
5826 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing a store chain of length " << Chain.size() << "\n"; } } while (false); | ||||
5827 | const unsigned Sz = R.getVectorElementSize(Chain[0]); | ||||
5828 | const unsigned MinVF = R.getMinVecRegSize() / Sz; | ||||
5829 | unsigned VF = Chain.size(); | ||||
5830 | |||||
5831 | if (!isPowerOf2_32(Sz) || !isPowerOf2_32(VF) || VF < 2 || VF < MinVF) | ||||
5832 | return false; | ||||
5833 | |||||
5834 | 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) | ||||
5835 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing " << VF << " stores at offset " << Idx << "\n"; } } while ( false); | ||||
5836 | |||||
5837 | R.buildTree(Chain); | ||||
5838 | Optional<ArrayRef<unsigned>> Order = R.bestOrder(); | ||||
5839 | // TODO: Handle orders of size less than number of elements in the vector. | ||||
5840 | if (Order && Order->size() == Chain.size()) { | ||||
5841 | // TODO: reorder tree nodes without tree rebuilding. | ||||
5842 | SmallVector<Value *, 4> ReorderedOps(Chain.rbegin(), Chain.rend()); | ||||
5843 | llvm::transform(*Order, ReorderedOps.begin(), | ||||
5844 | [Chain](const unsigned Idx) { return Chain[Idx]; }); | ||||
5845 | R.buildTree(ReorderedOps); | ||||
5846 | } | ||||
5847 | if (R.isTreeTinyAndNotFullyVectorizable()) | ||||
5848 | return false; | ||||
5849 | if (R.isLoadCombineCandidate()) | ||||
5850 | return false; | ||||
5851 | |||||
5852 | R.computeMinimumValueSizes(); | ||||
5853 | |||||
5854 | int Cost = R.getTreeCost(); | ||||
5855 | |||||
5856 | 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); | ||||
5857 | if (Cost < -SLPCostThreshold) { | ||||
5858 | 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); | ||||
5859 | |||||
5860 | using namespace ore; | ||||
5861 | |||||
5862 | R.getORE()->emit(OptimizationRemark(SV_NAME"slp-vectorizer", "StoresVectorized", | ||||
5863 | cast<StoreInst>(Chain[0])) | ||||
5864 | << "Stores SLP vectorized with cost " << NV("Cost", Cost) | ||||
5865 | << " and with tree size " | ||||
5866 | << NV("TreeSize", R.getTreeSize())); | ||||
5867 | |||||
5868 | R.vectorizeTree(); | ||||
5869 | return true; | ||||
5870 | } | ||||
5871 | |||||
5872 | return false; | ||||
5873 | } | ||||
5874 | |||||
5875 | bool SLPVectorizerPass::vectorizeStores(ArrayRef<StoreInst *> Stores, | ||||
5876 | BoUpSLP &R) { | ||||
5877 | // We may run into multiple chains that merge into a single chain. We mark the | ||||
5878 | // stores that we vectorized so that we don't visit the same store twice. | ||||
5879 | BoUpSLP::ValueSet VectorizedStores; | ||||
5880 | bool Changed = false; | ||||
5881 | |||||
5882 | int E = Stores.size(); | ||||
5883 | SmallBitVector Tails(E, false); | ||||
5884 | SmallVector<int, 16> ConsecutiveChain(E, E + 1); | ||||
5885 | int MaxIter = MaxStoreLookup.getValue(); | ||||
5886 | int IterCnt; | ||||
5887 | auto &&FindConsecutiveAccess = [this, &Stores, &Tails, &IterCnt, MaxIter, | ||||
5888 | &ConsecutiveChain](int K, int Idx) { | ||||
5889 | if (IterCnt >= MaxIter) | ||||
5890 | return true; | ||||
5891 | ++IterCnt; | ||||
5892 | if (!isConsecutiveAccess(Stores[K], Stores[Idx], *DL, *SE)) | ||||
5893 | return false; | ||||
5894 | |||||
5895 | Tails.set(Idx); | ||||
5896 | ConsecutiveChain[K] = Idx; | ||||
5897 | return true; | ||||
5898 | }; | ||||
5899 | // Do a quadratic search on all of the given stores in reverse order and find | ||||
5900 | // all of the pairs of stores that follow each other. | ||||
5901 | for (int Idx = E - 1; Idx >= 0; --Idx) { | ||||
5902 | // If a store has multiple consecutive store candidates, search according | ||||
5903 | // to the sequence: Idx-1, Idx+1, Idx-2, Idx+2, ... | ||||
5904 | // This is because usually pairing with immediate succeeding or preceding | ||||
5905 | // candidate create the best chance to find slp vectorization opportunity. | ||||
5906 | const int MaxLookDepth = std::max(E - Idx, Idx + 1); | ||||
5907 | IterCnt = 0; | ||||
5908 | for (int Offset = 1, F = MaxLookDepth; Offset < F; ++Offset) | ||||
5909 | if ((Idx >= Offset && FindConsecutiveAccess(Idx - Offset, Idx)) || | ||||
5910 | (Idx + Offset < E && FindConsecutiveAccess(Idx + Offset, Idx))) | ||||
5911 | break; | ||||
5912 | } | ||||
5913 | |||||
5914 | // For stores that start but don't end a link in the chain: | ||||
5915 | for (int Cnt = E; Cnt > 0; --Cnt) { | ||||
5916 | int I = Cnt - 1; | ||||
5917 | if (ConsecutiveChain[I] == E + 1 || Tails.test(I)) | ||||
5918 | continue; | ||||
5919 | // We found a store instr that starts a chain. Now follow the chain and try | ||||
5920 | // to vectorize it. | ||||
5921 | BoUpSLP::ValueList Operands; | ||||
5922 | // Collect the chain into a list. | ||||
5923 | while (I != E + 1 && !VectorizedStores.count(Stores[I])) { | ||||
5924 | Operands.push_back(Stores[I]); | ||||
5925 | // Move to the next value in the chain. | ||||
5926 | I = ConsecutiveChain[I]; | ||||
5927 | } | ||||
5928 | |||||
5929 | // If a vector register can't hold 1 element, we are done. | ||||
5930 | unsigned MaxVecRegSize = R.getMaxVecRegSize(); | ||||
5931 | unsigned EltSize = R.getVectorElementSize(Stores[0]); | ||||
5932 | if (MaxVecRegSize % EltSize != 0) | ||||
5933 | continue; | ||||
5934 | |||||
5935 | unsigned MaxElts = MaxVecRegSize / EltSize; | ||||
5936 | // FIXME: Is division-by-2 the correct step? Should we assert that the | ||||
5937 | // register size is a power-of-2? | ||||
5938 | unsigned StartIdx = 0; | ||||
5939 | for (unsigned Size = llvm::PowerOf2Ceil(MaxElts); Size >= 2; Size /= 2) { | ||||
5940 | for (unsigned Cnt = StartIdx, E = Operands.size(); Cnt + Size <= E;) { | ||||
5941 | ArrayRef<Value *> Slice = makeArrayRef(Operands).slice(Cnt, Size); | ||||
5942 | if (!VectorizedStores.count(Slice.front()) && | ||||
5943 | !VectorizedStores.count(Slice.back()) && | ||||
5944 | vectorizeStoreChain(Slice, R, Cnt)) { | ||||
5945 | // Mark the vectorized stores so that we don't vectorize them again. | ||||
5946 | VectorizedStores.insert(Slice.begin(), Slice.end()); | ||||
5947 | Changed = true; | ||||
5948 | // If we vectorized initial block, no need to try to vectorize it | ||||
5949 | // again. | ||||
5950 | if (Cnt == StartIdx) | ||||
5951 | StartIdx += Size; | ||||
5952 | Cnt += Size; | ||||
5953 | continue; | ||||
5954 | } | ||||
5955 | ++Cnt; | ||||
5956 | } | ||||
5957 | // Check if the whole array was vectorized already - exit. | ||||
5958 | if (StartIdx >= Operands.size()) | ||||
5959 | break; | ||||
5960 | } | ||||
5961 | } | ||||
5962 | |||||
5963 | return Changed; | ||||
5964 | } | ||||
5965 | |||||
5966 | void SLPVectorizerPass::collectSeedInstructions(BasicBlock *BB) { | ||||
5967 | // Initialize the collections. We will make a single pass over the block. | ||||
5968 | Stores.clear(); | ||||
5969 | GEPs.clear(); | ||||
5970 | |||||
5971 | // Visit the store and getelementptr instructions in BB and organize them in | ||||
5972 | // Stores and GEPs according to the underlying objects of their pointer | ||||
5973 | // operands. | ||||
5974 | for (Instruction &I : *BB) { | ||||
5975 | // Ignore store instructions that are volatile or have a pointer operand | ||||
5976 | // that doesn't point to a scalar type. | ||||
5977 | if (auto *SI = dyn_cast<StoreInst>(&I)) { | ||||
5978 | if (!SI->isSimple()) | ||||
5979 | continue; | ||||
5980 | if (!isValidElementType(SI->getValueOperand()->getType())) | ||||
5981 | continue; | ||||
5982 | Stores[getUnderlyingObject(SI->getPointerOperand())].push_back(SI); | ||||
5983 | } | ||||
5984 | |||||
5985 | // Ignore getelementptr instructions that have more than one index, a | ||||
5986 | // constant index, or a pointer operand that doesn't point to a scalar | ||||
5987 | // type. | ||||
5988 | else if (auto *GEP = dyn_cast<GetElementPtrInst>(&I)) { | ||||
5989 | auto Idx = GEP->idx_begin()->get(); | ||||
5990 | if (GEP->getNumIndices() > 1 || isa<Constant>(Idx)) | ||||
5991 | continue; | ||||
5992 | if (!isValidElementType(Idx->getType())) | ||||
5993 | continue; | ||||
5994 | if (GEP->getType()->isVectorTy()) | ||||
5995 | continue; | ||||
5996 | GEPs[GEP->getPointerOperand()].push_back(GEP); | ||||
5997 | } | ||||
5998 | } | ||||
5999 | } | ||||
6000 | |||||
6001 | bool SLPVectorizerPass::tryToVectorizePair(Value *A, Value *B, BoUpSLP &R) { | ||||
6002 | if (!A || !B) | ||||
6003 | return false; | ||||
6004 | Value *VL[] = {A, B}; | ||||
6005 | return tryToVectorizeList(VL, R, /*AllowReorder=*/true); | ||||
6006 | } | ||||
6007 | |||||
6008 | bool SLPVectorizerPass::tryToVectorizeList(ArrayRef<Value *> VL, BoUpSLP &R, | ||||
6009 | bool AllowReorder, | ||||
6010 | ArrayRef<Value *> InsertUses) { | ||||
6011 | if (VL.size() < 2) | ||||
6012 | return false; | ||||
6013 | |||||
6014 | 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) | ||||
6015 | << VL.size() << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Trying to vectorize a list of length = " << VL.size() << ".\n"; } } while (false); | ||||
6016 | |||||
6017 | // Check that all of the parts are instructions of the same type, | ||||
6018 | // we permit an alternate opcode via InstructionsState. | ||||
6019 | InstructionsState S = getSameOpcode(VL); | ||||
6020 | if (!S.getOpcode()) | ||||
6021 | return false; | ||||
6022 | |||||
6023 | Instruction *I0 = cast<Instruction>(S.OpValue); | ||||
6024 | // Make sure invalid types (including vector type) are rejected before | ||||
6025 | // determining vectorization factor for scalar instructions. | ||||
6026 | for (Value *V : VL) { | ||||
6027 | Type *Ty = V->getType(); | ||||
6028 | if (!isValidElementType(Ty)) { | ||||
6029 | // NOTE: the following will give user internal llvm type name, which may | ||||
6030 | // not be useful. | ||||
6031 | R.getORE()->emit([&]() { | ||||
6032 | std::string type_str; | ||||
6033 | llvm::raw_string_ostream rso(type_str); | ||||
6034 | Ty->print(rso); | ||||
6035 | return OptimizationRemarkMissed(SV_NAME"slp-vectorizer", "UnsupportedType", I0) | ||||
6036 | << "Cannot SLP vectorize list: type " | ||||
6037 | << rso.str() + " is unsupported by vectorizer"; | ||||
6038 | }); | ||||
6039 | return false; | ||||
6040 | } | ||||
6041 | } | ||||
6042 | |||||
6043 | unsigned Sz = R.getVectorElementSize(I0); | ||||
6044 | unsigned MinVF = std::max(2U, R.getMinVecRegSize() / Sz); | ||||
6045 | unsigned MaxVF = std::max<unsigned>(PowerOf2Floor(VL.size()), MinVF); | ||||
6046 | if (MaxVF < 2) { | ||||
6047 | R.getORE()->emit([&]() { | ||||
6048 | return OptimizationRemarkMissed(SV_NAME"slp-vectorizer", "SmallVF", I0) | ||||
6049 | << "Cannot SLP vectorize list: vectorization factor " | ||||
6050 | << "less than 2 is not supported"; | ||||
6051 | }); | ||||
6052 | return false; | ||||
6053 | } | ||||
6054 | |||||
6055 | bool Changed = false; | ||||
6056 | bool CandidateFound = false; | ||||
6057 | int MinCost = SLPCostThreshold; | ||||
6058 | |||||
6059 | bool CompensateUseCost = | ||||
6060 | !InsertUses.empty() && llvm::all_of(InsertUses, [](const Value *V) { | ||||
6061 | return V && isa<InsertElementInst>(V); | ||||
6062 | }); | ||||
6063 | assert((!CompensateUseCost || InsertUses.size() == VL.size()) &&(((!CompensateUseCost || InsertUses.size() == VL.size()) && "Each scalar expected to have an associated InsertElement user." ) ? static_cast<void> (0) : __assert_fail ("(!CompensateUseCost || InsertUses.size() == VL.size()) && \"Each scalar expected to have an associated InsertElement user.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 6064, __PRETTY_FUNCTION__)) | ||||
6064 | "Each scalar expected to have an associated InsertElement user.")(((!CompensateUseCost || InsertUses.size() == VL.size()) && "Each scalar expected to have an associated InsertElement user." ) ? static_cast<void> (0) : __assert_fail ("(!CompensateUseCost || InsertUses.size() == VL.size()) && \"Each scalar expected to have an associated InsertElement user.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 6064, __PRETTY_FUNCTION__)); | ||||
6065 | |||||
6066 | unsigned NextInst = 0, MaxInst = VL.size(); | ||||
6067 | for (unsigned VF = MaxVF; NextInst + 1 < MaxInst && VF >= MinVF; VF /= 2) { | ||||
6068 | // No actual vectorization should happen, if number of parts is the same as | ||||
6069 | // provided vectorization factor (i.e. the scalar type is used for vector | ||||
6070 | // code during codegen). | ||||
6071 | auto *VecTy = FixedVectorType::get(VL[0]->getType(), VF); | ||||
6072 | if (TTI->getNumberOfParts(VecTy) == VF) | ||||
6073 | continue; | ||||
6074 | for (unsigned I = NextInst; I < MaxInst; ++I) { | ||||
6075 | unsigned OpsWidth = 0; | ||||
6076 | |||||
6077 | if (I + VF > MaxInst) | ||||
6078 | OpsWidth = MaxInst - I; | ||||
6079 | else | ||||
6080 | OpsWidth = VF; | ||||
6081 | |||||
6082 | if (!isPowerOf2_32(OpsWidth) || OpsWidth < 2) | ||||
6083 | break; | ||||
6084 | |||||
6085 | ArrayRef<Value *> Ops = VL.slice(I, OpsWidth); | ||||
6086 | // Check that a previous iteration of this loop did not delete the Value. | ||||
6087 | if (llvm::any_of(Ops, [&R](Value *V) { | ||||
6088 | auto *I = dyn_cast<Instruction>(V); | ||||
6089 | return I && R.isDeleted(I); | ||||
6090 | })) | ||||
6091 | continue; | ||||
6092 | |||||
6093 | LLVM_DEBUG(dbgs() << "SLP: Analyzing " << OpsWidth << " operations "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing " << OpsWidth << " operations " << "\n"; } } while (false) | ||||
6094 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing " << OpsWidth << " operations " << "\n"; } } while (false); | ||||
6095 | |||||
6096 | R.buildTree(Ops); | ||||
6097 | Optional<ArrayRef<unsigned>> Order = R.bestOrder(); | ||||
6098 | // TODO: check if we can allow reordering for more cases. | ||||
6099 | if (AllowReorder && Order) { | ||||
6100 | // TODO: reorder tree nodes without tree rebuilding. | ||||
6101 | // Conceptually, there is nothing actually preventing us from trying to | ||||
6102 | // reorder a larger list. In fact, we do exactly this when vectorizing | ||||
6103 | // reductions. However, at this point, we only expect to get here when | ||||
6104 | // there are exactly two operations. | ||||
6105 | assert(Ops.size() == 2)((Ops.size() == 2) ? static_cast<void> (0) : __assert_fail ("Ops.size() == 2", "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 6105, __PRETTY_FUNCTION__)); | ||||
6106 | Value *ReorderedOps[] = {Ops[1], Ops[0]}; | ||||
6107 | R.buildTree(ReorderedOps, None); | ||||
6108 | } | ||||
6109 | if (R.isTreeTinyAndNotFullyVectorizable()) | ||||
6110 | continue; | ||||
6111 | |||||
6112 | R.computeMinimumValueSizes(); | ||||
6113 | int Cost = R.getTreeCost(); | ||||
6114 | CandidateFound = true; | ||||
6115 | if (CompensateUseCost) { | ||||
6116 | // TODO: Use TTI's getScalarizationOverhead for sequence of inserts | ||||
6117 | // rather than sum of single inserts as the latter may overestimate | ||||
6118 | // cost. This work should imply improving cost estimation for extracts | ||||
6119 | // that added in for external (for vectorization tree) users,i.e. that | ||||
6120 | // part should also switch to same interface. | ||||
6121 | // For example, the following case is projected code after SLP: | ||||
6122 | // %4 = extractelement <4 x i64> %3, i32 0 | ||||
6123 | // %v0 = insertelement <4 x i64> undef, i64 %4, i32 0 | ||||
6124 | // %5 = extractelement <4 x i64> %3, i32 1 | ||||
6125 | // %v1 = insertelement <4 x i64> %v0, i64 %5, i32 1 | ||||
6126 | // %6 = extractelement <4 x i64> %3, i32 2 | ||||
6127 | // %v2 = insertelement <4 x i64> %v1, i64 %6, i32 2 | ||||
6128 | // %7 = extractelement <4 x i64> %3, i32 3 | ||||
6129 | // %v3 = insertelement <4 x i64> %v2, i64 %7, i32 3 | ||||
6130 | // | ||||
6131 | // Extracts here added by SLP in order to feed users (the inserts) of | ||||
6132 | // original scalars and contribute to "ExtractCost" at cost evaluation. | ||||
6133 | // The inserts in turn form sequence to build an aggregate that | ||||
6134 | // detected by findBuildAggregate routine. | ||||
6135 | // SLP makes an assumption that such sequence will be optimized away | ||||
6136 | // later (instcombine) so it tries to compensate ExctractCost with | ||||
6137 | // cost of insert sequence. | ||||
6138 | // Current per element cost calculation approach is not quite accurate | ||||
6139 | // and tends to create bias toward favoring vectorization. | ||||
6140 | // Switching to the TTI interface might help a bit. | ||||
6141 | // Alternative solution could be pattern-match to detect a no-op or | ||||
6142 | // shuffle. | ||||
6143 | unsigned UserCost = 0; | ||||
6144 | for (unsigned Lane = 0; Lane < OpsWidth; Lane++) { | ||||
6145 | auto *IE = cast<InsertElementInst>(InsertUses[I + Lane]); | ||||
6146 | if (auto *CI = dyn_cast<ConstantInt>(IE->getOperand(2))) | ||||
6147 | UserCost += TTI->getVectorInstrCost( | ||||
6148 | Instruction::InsertElement, IE->getType(), CI->getZExtValue()); | ||||
6149 | } | ||||
6150 | LLVM_DEBUG(dbgs() << "SLP: Compensate cost of users by: " << UserCostdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Compensate cost of users by: " << UserCost << ".\n"; } } while (false) | ||||
6151 | << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Compensate cost of users by: " << UserCost << ".\n"; } } while (false); | ||||
6152 | Cost -= UserCost; | ||||
6153 | } | ||||
6154 | |||||
6155 | MinCost = std::min(MinCost, Cost); | ||||
6156 | |||||
6157 | if (Cost < -SLPCostThreshold) { | ||||
6158 | 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); | ||||
6159 | R.getORE()->emit(OptimizationRemark(SV_NAME"slp-vectorizer", "VectorizedList", | ||||
6160 | cast<Instruction>(Ops[0])) | ||||
6161 | << "SLP vectorized with cost " << ore::NV("Cost", Cost) | ||||
6162 | << " and with tree size " | ||||
6163 | << ore::NV("TreeSize", R.getTreeSize())); | ||||
6164 | |||||
6165 | R.vectorizeTree(); | ||||
6166 | // Move to the next bundle. | ||||
6167 | I += VF - 1; | ||||
6168 | NextInst = I + 1; | ||||
6169 | Changed = true; | ||||
6170 | } | ||||
6171 | } | ||||
6172 | } | ||||
6173 | |||||
6174 | if (!Changed && CandidateFound) { | ||||
6175 | R.getORE()->emit([&]() { | ||||
6176 | return OptimizationRemarkMissed(SV_NAME"slp-vectorizer", "NotBeneficial", I0) | ||||
6177 | << "List vectorization was possible but not beneficial with cost " | ||||
6178 | << ore::NV("Cost", MinCost) << " >= " | ||||
6179 | << ore::NV("Treshold", -SLPCostThreshold); | ||||
6180 | }); | ||||
6181 | } else if (!Changed) { | ||||
6182 | R.getORE()->emit([&]() { | ||||
6183 | return OptimizationRemarkMissed(SV_NAME"slp-vectorizer", "NotPossible", I0) | ||||
6184 | << "Cannot SLP vectorize list: vectorization was impossible" | ||||
6185 | << " with available vectorization factors"; | ||||
6186 | }); | ||||
6187 | } | ||||
6188 | return Changed; | ||||
6189 | } | ||||
6190 | |||||
6191 | bool SLPVectorizerPass::tryToVectorize(Instruction *I, BoUpSLP &R) { | ||||
6192 | if (!I) | ||||
6193 | return false; | ||||
6194 | |||||
6195 | if (!isa<BinaryOperator>(I) && !isa<CmpInst>(I)) | ||||
6196 | return false; | ||||
6197 | |||||
6198 | Value *P = I->getParent(); | ||||
6199 | |||||
6200 | // Vectorize in current basic block only. | ||||
6201 | auto *Op0 = dyn_cast<Instruction>(I->getOperand(0)); | ||||
6202 | auto *Op1 = dyn_cast<Instruction>(I->getOperand(1)); | ||||
6203 | if (!Op0 || !Op1 || Op0->getParent() != P || Op1->getParent() != P) | ||||
6204 | return false; | ||||
6205 | |||||
6206 | // Try to vectorize V. | ||||
6207 | if (tryToVectorizePair(Op0, Op1, R)) | ||||
6208 | return true; | ||||
6209 | |||||
6210 | auto *A = dyn_cast<BinaryOperator>(Op0); | ||||
6211 | auto *B = dyn_cast<BinaryOperator>(Op1); | ||||
6212 | // Try to skip B. | ||||
6213 | if (B && B->hasOneUse()) { | ||||
6214 | auto *B0 = dyn_cast<BinaryOperator>(B->getOperand(0)); | ||||
6215 | auto *B1 = dyn_cast<BinaryOperator>(B->getOperand(1)); | ||||
6216 | if (B0 && B0->getParent() == P && tryToVectorizePair(A, B0, R)) | ||||
6217 | return true; | ||||
6218 | if (B1 && B1->getParent() == P && tryToVectorizePair(A, B1, R)) | ||||
6219 | return true; | ||||
6220 | } | ||||
6221 | |||||
6222 | // Try to skip A. | ||||
6223 | if (A && A->hasOneUse()) { | ||||
6224 | auto *A0 = dyn_cast<BinaryOperator>(A->getOperand(0)); | ||||
6225 | auto *A1 = dyn_cast<BinaryOperator>(A->getOperand(1)); | ||||
6226 | if (A0 && A0->getParent() == P && tryToVectorizePair(A0, B, R)) | ||||
6227 | return true; | ||||
6228 | if (A1 && A1->getParent() == P && tryToVectorizePair(A1, B, R)) | ||||
6229 | return true; | ||||
6230 | } | ||||
6231 | return false; | ||||
6232 | } | ||||
6233 | |||||
6234 | /// Generate a shuffle mask to be used in a reduction tree. | ||||
6235 | /// | ||||
6236 | /// \param VecLen The length of the vector to be reduced. | ||||
6237 | /// \param NumEltsToRdx The number of elements that should be reduced in the | ||||
6238 | /// vector. | ||||
6239 | /// \param IsPairwise Whether the reduction is a pairwise or splitting | ||||
6240 | /// reduction. A pairwise reduction will generate a mask of | ||||
6241 | /// <0,2,...> or <1,3,..> while a splitting reduction will generate | ||||
6242 | /// <2,3, undef,undef> for a vector of 4 and NumElts = 2. | ||||
6243 | /// \param IsLeft True will generate a mask of even elements, odd otherwise. | ||||
6244 | static SmallVector<int, 32> createRdxShuffleMask(unsigned VecLen, | ||||
6245 | unsigned NumEltsToRdx, | ||||
6246 | bool IsPairwise, bool IsLeft) { | ||||
6247 | assert((IsPairwise || !IsLeft) && "Don't support a <0,1,undef,...> mask")(((IsPairwise || !IsLeft) && "Don't support a <0,1,undef,...> mask" ) ? static_cast<void> (0) : __assert_fail ("(IsPairwise || !IsLeft) && \"Don't support a <0,1,undef,...> mask\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 6247, __PRETTY_FUNCTION__)); | ||||
6248 | |||||
6249 | SmallVector<int, 32> ShuffleMask(VecLen, -1); | ||||
6250 | |||||
6251 | if (IsPairwise) | ||||
6252 | // Build a mask of 0, 2, ... (left) or 1, 3, ... (right). | ||||
6253 | for (unsigned i = 0; i != NumEltsToRdx; ++i) | ||||
6254 | ShuffleMask[i] = 2 * i + !IsLeft; | ||||
6255 | else | ||||
6256 | // Move the upper half of the vector to the lower half. | ||||
6257 | for (unsigned i = 0; i != NumEltsToRdx; ++i) | ||||
6258 | ShuffleMask[i] = NumEltsToRdx + i; | ||||
6259 | |||||
6260 | return ShuffleMask; | ||||
6261 | } | ||||
6262 | |||||
6263 | namespace { | ||||
6264 | |||||
6265 | /// Model horizontal reductions. | ||||
6266 | /// | ||||
6267 | /// A horizontal reduction is a tree of reduction operations (currently add and | ||||
6268 | /// fadd) that has operations that can be put into a vector as its leaf. | ||||
6269 | /// For example, this tree: | ||||
6270 | /// | ||||
6271 | /// mul mul mul mul | ||||
6272 | /// \ / \ / | ||||
6273 | /// + + | ||||
6274 | /// \ / | ||||
6275 | /// + | ||||
6276 | /// This tree has "mul" as its reduced values and "+" as its reduction | ||||
6277 | /// operations. A reduction might be feeding into a store or a binary operation | ||||
6278 | /// feeding a phi. | ||||
6279 | /// ... | ||||
6280 | /// \ / | ||||
6281 | /// + | ||||
6282 | /// | | ||||
6283 | /// phi += | ||||
6284 | /// | ||||
6285 | /// Or: | ||||
6286 | /// ... | ||||
6287 | /// \ / | ||||
6288 | /// + | ||||
6289 | /// | | ||||
6290 | /// *p = | ||||
6291 | /// | ||||
6292 | class HorizontalReduction { | ||||
6293 | using ReductionOpsType = SmallVector<Value *, 16>; | ||||
6294 | using ReductionOpsListType = SmallVector<ReductionOpsType, 2>; | ||||
6295 | ReductionOpsListType ReductionOps; | ||||
6296 | SmallVector<Value *, 32> ReducedVals; | ||||
6297 | // Use map vector to make stable output. | ||||
6298 | MapVector<Instruction *, Value *> ExtraArgs; | ||||
6299 | |||||
6300 | /// Kind of the reduction data. | ||||
6301 | enum ReductionKind { | ||||
6302 | RK_None, /// Not a reduction. | ||||
6303 | RK_Arithmetic, /// Binary reduction data. | ||||
6304 | RK_SMin, /// Signed minimum reduction data. | ||||
6305 | RK_UMin, /// Unsigned minimum reduction data. | ||||
6306 | RK_SMax, /// Signed maximum reduction data. | ||||
6307 | RK_UMax, /// Unsigned maximum reduction data. | ||||
6308 | }; | ||||
6309 | |||||
6310 | /// Contains info about operation, like its opcode, left and right operands. | ||||
6311 | class OperationData { | ||||
6312 | /// Opcode of the instruction. | ||||
6313 | unsigned Opcode = 0; | ||||
6314 | |||||
6315 | /// Kind of the reduction operation. | ||||
6316 | ReductionKind Kind = RK_None; | ||||
6317 | |||||
6318 | /// Checks if the reduction operation can be vectorized. | ||||
6319 | bool isVectorizable() const { | ||||
6320 | // We currently only support add/mul/logical && min/max reductions. | ||||
6321 | return ((Kind == RK_Arithmetic && | ||||
6322 | (Opcode == Instruction::Add || Opcode == Instruction::FAdd || | ||||
6323 | Opcode == Instruction::Mul || Opcode == Instruction::FMul || | ||||
6324 | Opcode == Instruction::And || Opcode == Instruction::Or || | ||||
6325 | Opcode == Instruction::Xor)) || | ||||
6326 | (Opcode == Instruction::ICmp && | ||||
6327 | (Kind == RK_SMin || Kind == RK_SMax || | ||||
6328 | Kind == RK_UMin || Kind == RK_UMax))); | ||||
6329 | } | ||||
6330 | |||||
6331 | /// Creates reduction operation with the current opcode. | ||||
6332 | Value *createOp(IRBuilder<> &Builder, Value *LHS, Value *RHS, | ||||
6333 | const Twine &Name) const { | ||||
6334 | assert(isVectorizable() &&((isVectorizable() && "Expected add|fadd or min/max reduction operation." ) ? static_cast<void> (0) : __assert_fail ("isVectorizable() && \"Expected add|fadd or min/max reduction operation.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 6335, __PRETTY_FUNCTION__)) | ||||
6335 | "Expected add|fadd or min/max reduction operation.")((isVectorizable() && "Expected add|fadd or min/max reduction operation." ) ? static_cast<void> (0) : __assert_fail ("isVectorizable() && \"Expected add|fadd or min/max reduction operation.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 6335, __PRETTY_FUNCTION__)); | ||||
6336 | Value *Cmp = nullptr; | ||||
6337 | switch (Kind) { | ||||
6338 | case RK_Arithmetic: | ||||
6339 | return Builder.CreateBinOp((Instruction::BinaryOps)Opcode, LHS, RHS, | ||||
6340 | Name); | ||||
6341 | case RK_SMin: | ||||
6342 | assert(Opcode == Instruction::ICmp && "Expected integer types.")((Opcode == Instruction::ICmp && "Expected integer types." ) ? static_cast<void> (0) : __assert_fail ("Opcode == Instruction::ICmp && \"Expected integer types.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 6342, __PRETTY_FUNCTION__)); | ||||
6343 | Cmp = Builder.CreateICmpSLT(LHS, RHS); | ||||
6344 | return Builder.CreateSelect(Cmp, LHS, RHS, Name); | ||||
6345 | case RK_SMax: | ||||
6346 | assert(Opcode == Instruction::ICmp && "Expected integer types.")((Opcode == Instruction::ICmp && "Expected integer types." ) ? static_cast<void> (0) : __assert_fail ("Opcode == Instruction::ICmp && \"Expected integer types.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 6346, __PRETTY_FUNCTION__)); | ||||
6347 | Cmp = Builder.CreateICmpSGT(LHS, RHS); | ||||
6348 | return Builder.CreateSelect(Cmp, LHS, RHS, Name); | ||||
6349 | case RK_UMin: | ||||
6350 | assert(Opcode == Instruction::ICmp && "Expected integer types.")((Opcode == Instruction::ICmp && "Expected integer types." ) ? static_cast<void> (0) : __assert_fail ("Opcode == Instruction::ICmp && \"Expected integer types.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 6350, __PRETTY_FUNCTION__)); | ||||
6351 | Cmp = Builder.CreateICmpULT(LHS, RHS); | ||||
6352 | return Builder.CreateSelect(Cmp, LHS, RHS, Name); | ||||
6353 | case RK_UMax: | ||||
6354 | assert(Opcode == Instruction::ICmp && "Expected integer types.")((Opcode == Instruction::ICmp && "Expected integer types." ) ? static_cast<void> (0) : __assert_fail ("Opcode == Instruction::ICmp && \"Expected integer types.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 6354, __PRETTY_FUNCTION__)); | ||||
6355 | Cmp = Builder.CreateICmpUGT(LHS, RHS); | ||||
6356 | return Builder.CreateSelect(Cmp, LHS, RHS, Name); | ||||
6357 | case RK_None: | ||||
6358 | break; | ||||
6359 | } | ||||
6360 | llvm_unreachable("Unknown reduction operation.")::llvm::llvm_unreachable_internal("Unknown reduction operation." , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 6360); | ||||
6361 | } | ||||
6362 | |||||
6363 | public: | ||||
6364 | explicit OperationData() = default; | ||||
6365 | |||||
6366 | /// Construction for reduced values. They are identified by opcode only and | ||||
6367 | /// don't have associated LHS/RHS values. | ||||
6368 | explicit OperationData(Instruction &I) { | ||||
6369 | Opcode = I.getOpcode(); | ||||
6370 | } | ||||
6371 | |||||
6372 | /// Constructor for reduction operations with opcode and its left and | ||||
6373 | /// right operands. | ||||
6374 | OperationData(unsigned Opcode, ReductionKind Kind) | ||||
6375 | : Opcode(Opcode), Kind(Kind) { | ||||
6376 | assert(Kind != RK_None && "One of the reduction operations is expected.")((Kind != RK_None && "One of the reduction operations is expected." ) ? static_cast<void> (0) : __assert_fail ("Kind != RK_None && \"One of the reduction operations is expected.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 6376, __PRETTY_FUNCTION__)); | ||||
6377 | } | ||||
6378 | |||||
6379 | explicit operator bool() const { return Opcode; } | ||||
6380 | |||||
6381 | /// Return true if this operation is any kind of minimum or maximum. | ||||
6382 | bool isMinMax() const { | ||||
6383 | switch (Kind) { | ||||
6384 | case RK_Arithmetic: | ||||
6385 | return false; | ||||
6386 | case RK_SMin: | ||||
6387 | case RK_SMax: | ||||
6388 | case RK_UMin: | ||||
6389 | case RK_UMax: | ||||
6390 | return true; | ||||
6391 | case RK_None: | ||||
6392 | break; | ||||
6393 | } | ||||
6394 | llvm_unreachable("Reduction kind is not set")::llvm::llvm_unreachable_internal("Reduction kind is not set" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 6394); | ||||
6395 | } | ||||
6396 | |||||
6397 | /// Get the index of the first operand. | ||||
6398 | unsigned getFirstOperandIndex() const { | ||||
6399 | assert(!!*this && "The opcode is not set.")((!!*this && "The opcode is not set.") ? static_cast< void> (0) : __assert_fail ("!!*this && \"The opcode is not set.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 6399, __PRETTY_FUNCTION__)); | ||||
6400 | // We allow calling this before 'Kind' is set, so handle that specially. | ||||
6401 | if (Kind == RK_None) | ||||
6402 | return 0; | ||||
6403 | return isMinMax() ? 1 : 0; | ||||
6404 | } | ||||
6405 | |||||
6406 | /// Total number of operands in the reduction operation. | ||||
6407 | unsigned getNumberOfOperands() const { | ||||
6408 | assert(Kind != RK_None && !!*this && "Expected reduction operation.")((Kind != RK_None && !!*this && "Expected reduction operation." ) ? static_cast<void> (0) : __assert_fail ("Kind != RK_None && !!*this && \"Expected reduction operation.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 6408, __PRETTY_FUNCTION__)); | ||||
6409 | return isMinMax() ? 3 : 2; | ||||
6410 | } | ||||
6411 | |||||
6412 | /// Checks if the instruction is in basic block \p BB. | ||||
6413 | /// For a min/max reduction check that both compare and select are in \p BB. | ||||
6414 | bool hasSameParent(Instruction *I, BasicBlock *BB, bool IsRedOp) const { | ||||
6415 | assert(Kind != RK_None && !!*this && "Expected reduction operation.")((Kind != RK_None && !!*this && "Expected reduction operation." ) ? static_cast<void> (0) : __assert_fail ("Kind != RK_None && !!*this && \"Expected reduction operation.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 6415, __PRETTY_FUNCTION__)); | ||||
6416 | if (IsRedOp && isMinMax()) { | ||||
6417 | auto *Cmp = cast<Instruction>(cast<SelectInst>(I)->getCondition()); | ||||
6418 | return I->getParent() == BB && Cmp && Cmp->getParent() == BB; | ||||
6419 | } | ||||
6420 | return I->getParent() == BB; | ||||
6421 | } | ||||
6422 | |||||
6423 | /// Expected number of uses for reduction operations/reduced values. | ||||
6424 | bool hasRequiredNumberOfUses(Instruction *I, bool IsReductionOp) const { | ||||
6425 | assert(Kind != RK_None && !!*this && "Expected reduction operation.")((Kind != RK_None && !!*this && "Expected reduction operation." ) ? static_cast<void> (0) : __assert_fail ("Kind != RK_None && !!*this && \"Expected reduction operation.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 6425, __PRETTY_FUNCTION__)); | ||||
6426 | // SelectInst must be used twice while the condition op must have single | ||||
6427 | // use only. | ||||
6428 | if (isMinMax()) | ||||
6429 | return I->hasNUses(2) && | ||||
6430 | (!IsReductionOp || | ||||
6431 | cast<SelectInst>(I)->getCondition()->hasOneUse()); | ||||
6432 | |||||
6433 | // Arithmetic reduction operation must be used once only. | ||||
6434 | return I->hasOneUse(); | ||||
6435 | } | ||||
6436 | |||||
6437 | /// Initializes the list of reduction operations. | ||||
6438 | void initReductionOps(ReductionOpsListType &ReductionOps) { | ||||
6439 | assert(Kind != RK_None && !!*this && "Expected reduction operation.")((Kind != RK_None && !!*this && "Expected reduction operation." ) ? static_cast<void> (0) : __assert_fail ("Kind != RK_None && !!*this && \"Expected reduction operation.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 6439, __PRETTY_FUNCTION__)); | ||||
6440 | if (isMinMax()) | ||||
6441 | ReductionOps.assign(2, ReductionOpsType()); | ||||
6442 | else | ||||
6443 | ReductionOps.assign(1, ReductionOpsType()); | ||||
6444 | } | ||||
6445 | |||||
6446 | /// Add all reduction operations for the reduction instruction \p I. | ||||
6447 | void addReductionOps(Instruction *I, ReductionOpsListType &ReductionOps) { | ||||
6448 | assert(Kind != RK_None && !!*this && "Expected reduction operation.")((Kind != RK_None && !!*this && "Expected reduction operation." ) ? static_cast<void> (0) : __assert_fail ("Kind != RK_None && !!*this && \"Expected reduction operation.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 6448, __PRETTY_FUNCTION__)); | ||||
6449 | if (isMinMax()) { | ||||
6450 | ReductionOps[0].emplace_back(cast<SelectInst>(I)->getCondition()); | ||||
6451 | ReductionOps[1].emplace_back(I); | ||||
6452 | } else { | ||||
6453 | ReductionOps[0].emplace_back(I); | ||||
6454 | } | ||||
6455 | } | ||||
6456 | |||||
6457 | /// Checks if instruction is associative and can be vectorized. | ||||
6458 | bool isAssociative(Instruction *I) const { | ||||
6459 | assert(Kind != RK_None && *this && "Expected reduction operation.")((Kind != RK_None && *this && "Expected reduction operation." ) ? static_cast<void> (0) : __assert_fail ("Kind != RK_None && *this && \"Expected reduction operation.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 6459, __PRETTY_FUNCTION__)); | ||||
6460 | switch (Kind) { | ||||
6461 | case RK_Arithmetic: | ||||
6462 | return I->isAssociative(); | ||||
6463 | case RK_SMin: | ||||
6464 | case RK_SMax: | ||||
6465 | case RK_UMin: | ||||
6466 | case RK_UMax: | ||||
6467 | assert(Opcode == Instruction::ICmp &&((Opcode == Instruction::ICmp && "Only integer compare operation is expected." ) ? static_cast<void> (0) : __assert_fail ("Opcode == Instruction::ICmp && \"Only integer compare operation is expected.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 6468, __PRETTY_FUNCTION__)) | ||||
6468 | "Only integer compare operation is expected.")((Opcode == Instruction::ICmp && "Only integer compare operation is expected." ) ? static_cast<void> (0) : __assert_fail ("Opcode == Instruction::ICmp && \"Only integer compare operation is expected.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 6468, __PRETTY_FUNCTION__)); | ||||
6469 | return true; | ||||
6470 | case RK_None: | ||||
6471 | break; | ||||
6472 | } | ||||
6473 | llvm_unreachable("Reduction kind is not set")::llvm::llvm_unreachable_internal("Reduction kind is not set" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 6473); | ||||
6474 | } | ||||
6475 | |||||
6476 | /// Checks if the reduction operation can be vectorized. | ||||
6477 | bool isVectorizable(Instruction *I) const { | ||||
6478 | return isVectorizable() && isAssociative(I); | ||||
6479 | } | ||||
6480 | |||||
6481 | /// Checks if two operation data are both a reduction op or both a reduced | ||||
6482 | /// value. | ||||
6483 | bool operator==(const OperationData &OD) const { | ||||
6484 | assert(((Kind != OD.Kind) || (Opcode != 0 && OD.Opcode != 0)) &&((((Kind != OD.Kind) || (Opcode != 0 && OD.Opcode != 0 )) && "One of the comparing operations is incorrect." ) ? static_cast<void> (0) : __assert_fail ("((Kind != OD.Kind) || (Opcode != 0 && OD.Opcode != 0)) && \"One of the comparing operations is incorrect.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 6485, __PRETTY_FUNCTION__)) | ||||
6485 | "One of the comparing operations is incorrect.")((((Kind != OD.Kind) || (Opcode != 0 && OD.Opcode != 0 )) && "One of the comparing operations is incorrect." ) ? static_cast<void> (0) : __assert_fail ("((Kind != OD.Kind) || (Opcode != 0 && OD.Opcode != 0)) && \"One of the comparing operations is incorrect.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 6485, __PRETTY_FUNCTION__)); | ||||
6486 | return Kind == OD.Kind && Opcode == OD.Opcode; | ||||
6487 | } | ||||
6488 | bool operator!=(const OperationData &OD) const { return !(*this == OD); } | ||||
6489 | void clear() { | ||||
6490 | Opcode = 0; | ||||
6491 | Kind = RK_None; | ||||
6492 | } | ||||
6493 | |||||
6494 | /// Get the opcode of the reduction operation. | ||||
6495 | unsigned getOpcode() const { | ||||
6496 | assert(isVectorizable() && "Expected vectorizable operation.")((isVectorizable() && "Expected vectorizable operation." ) ? static_cast<void> (0) : __assert_fail ("isVectorizable() && \"Expected vectorizable operation.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 6496, __PRETTY_FUNCTION__)); | ||||
6497 | return Opcode; | ||||
6498 | } | ||||
6499 | |||||
6500 | /// Get kind of reduction data. | ||||
6501 | ReductionKind getKind() const { return Kind; } | ||||
6502 | Value *getLHS(Instruction *I) const { | ||||
6503 | if (Kind == RK_None) | ||||
6504 | return nullptr; | ||||
6505 | return I->getOperand(getFirstOperandIndex()); | ||||
6506 | } | ||||
6507 | Value *getRHS(Instruction *I) const { | ||||
6508 | if (Kind == RK_None) | ||||
6509 | return nullptr; | ||||
6510 | return I->getOperand(getFirstOperandIndex() + 1); | ||||
6511 | } | ||||
6512 | |||||
6513 | /// Creates reduction operation with the current opcode with the IR flags | ||||
6514 | /// from \p ReductionOps. | ||||
6515 | Value *createOp(IRBuilder<> &Builder, Value *LHS, Value *RHS, | ||||
6516 | const Twine &Name, | ||||
6517 | const ReductionOpsListType &ReductionOps) const { | ||||
6518 | assert(isVectorizable() &&((isVectorizable() && "Expected add|fadd or min/max reduction operation." ) ? static_cast<void> (0) : __assert_fail ("isVectorizable() && \"Expected add|fadd or min/max reduction operation.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 6519, __PRETTY_FUNCTION__)) | ||||
6519 | "Expected add|fadd or min/max reduction operation.")((isVectorizable() && "Expected add|fadd or min/max reduction operation." ) ? static_cast<void> (0) : __assert_fail ("isVectorizable() && \"Expected add|fadd or min/max reduction operation.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 6519, __PRETTY_FUNCTION__)); | ||||
6520 | auto *Op = createOp(Builder, LHS, RHS, Name); | ||||
6521 | switch (Kind) { | ||||
6522 | case RK_Arithmetic: | ||||
6523 | propagateIRFlags(Op, ReductionOps[0]); | ||||
6524 | return Op; | ||||
6525 | case RK_SMin: | ||||
6526 | case RK_SMax: | ||||
6527 | case RK_UMin: | ||||
6528 | case RK_UMax: | ||||
6529 | if (auto *SI = dyn_cast<SelectInst>(Op)) | ||||
6530 | propagateIRFlags(SI->getCondition(), ReductionOps[0]); | ||||
6531 | propagateIRFlags(Op, ReductionOps[1]); | ||||
6532 | return Op; | ||||
6533 | case RK_None: | ||||
6534 | break; | ||||
6535 | } | ||||
6536 | llvm_unreachable("Unknown reduction operation.")::llvm::llvm_unreachable_internal("Unknown reduction operation." , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 6536); | ||||
6537 | } | ||||
6538 | /// Creates reduction operation with the current opcode with the IR flags | ||||
6539 | /// from \p I. | ||||
6540 | Value *createOp(IRBuilder<> &Builder, Value *LHS, Value *RHS, | ||||
6541 | const Twine &Name, Instruction *I) const { | ||||
6542 | assert(isVectorizable() &&((isVectorizable() && "Expected add|fadd or min/max reduction operation." ) ? static_cast<void> (0) : __assert_fail ("isVectorizable() && \"Expected add|fadd or min/max reduction operation.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 6543, __PRETTY_FUNCTION__)) | ||||
6543 | "Expected add|fadd or min/max reduction operation.")((isVectorizable() && "Expected add|fadd or min/max reduction operation." ) ? static_cast<void> (0) : __assert_fail ("isVectorizable() && \"Expected add|fadd or min/max reduction operation.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 6543, __PRETTY_FUNCTION__)); | ||||
6544 | auto *Op = createOp(Builder, LHS, RHS, Name); | ||||
6545 | switch (Kind) { | ||||
6546 | case RK_Arithmetic: | ||||
6547 | propagateIRFlags(Op, I); | ||||
6548 | return Op; | ||||
6549 | case RK_SMin: | ||||
6550 | case RK_SMax: | ||||
6551 | case RK_UMin: | ||||
6552 | case RK_UMax: | ||||
6553 | if (auto *SI = dyn_cast<SelectInst>(Op)) { | ||||
6554 | propagateIRFlags(SI->getCondition(), | ||||
6555 | cast<SelectInst>(I)->getCondition()); | ||||
6556 | } | ||||
6557 | propagateIRFlags(Op, I); | ||||
6558 | return Op; | ||||
6559 | case RK_None: | ||||
6560 | break; | ||||
6561 | } | ||||
6562 | llvm_unreachable("Unknown reduction operation.")::llvm::llvm_unreachable_internal("Unknown reduction operation." , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 6562); | ||||
6563 | } | ||||
6564 | |||||
6565 | TargetTransformInfo::ReductionFlags getFlags() const { | ||||
6566 | TargetTransformInfo::ReductionFlags Flags; | ||||
6567 | switch (Kind) { | ||||
6568 | case RK_Arithmetic: | ||||
6569 | break; | ||||
6570 | case RK_SMin: | ||||
6571 | Flags.IsSigned = true; | ||||
6572 | Flags.IsMaxOp = false; | ||||
6573 | break; | ||||
6574 | case RK_SMax: | ||||
6575 | Flags.IsSigned = true; | ||||
6576 | Flags.IsMaxOp = true; | ||||
6577 | break; | ||||
6578 | case RK_UMin: | ||||
6579 | Flags.IsSigned = false; | ||||
6580 | Flags.IsMaxOp = false; | ||||
6581 | break; | ||||
6582 | case RK_UMax: | ||||
6583 | Flags.IsSigned = false; | ||||
6584 | Flags.IsMaxOp = true; | ||||
6585 | break; | ||||
6586 | case RK_None: | ||||
6587 | llvm_unreachable("Reduction kind is not set")::llvm::llvm_unreachable_internal("Reduction kind is not set" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 6587); | ||||
6588 | } | ||||
6589 | return Flags; | ||||
6590 | } | ||||
6591 | }; | ||||
6592 | |||||
6593 | WeakTrackingVH ReductionRoot; | ||||
6594 | |||||
6595 | /// The operation data of the reduction operation. | ||||
6596 | OperationData ReductionData; | ||||
6597 | |||||
6598 | /// The operation data of the values we perform a reduction on. | ||||
6599 | OperationData ReducedValueData; | ||||
6600 | |||||
6601 | /// Should we model this reduction as a pairwise reduction tree or a tree that | ||||
6602 | /// splits the vector in halves and adds those halves. | ||||
6603 | bool IsPairwiseReduction = false; | ||||
6604 | |||||
6605 | /// Checks if the ParentStackElem.first should be marked as a reduction | ||||
6606 | /// operation with an extra argument or as extra argument itself. | ||||
6607 | void markExtraArg(std::pair<Instruction *, unsigned> &ParentStackElem, | ||||
6608 | Value *ExtraArg) { | ||||
6609 | if (ExtraArgs.count(ParentStackElem.first)) { | ||||
6610 | ExtraArgs[ParentStackElem.first] = nullptr; | ||||
6611 | // We ran into something like: | ||||
6612 | // ParentStackElem.first = ExtraArgs[ParentStackElem.first] + ExtraArg. | ||||
6613 | // The whole ParentStackElem.first should be considered as an extra value | ||||
6614 | // in this case. | ||||
6615 | // Do not perform analysis of remaining operands of ParentStackElem.first | ||||
6616 | // instruction, this whole instruction is an extra argument. | ||||
6617 | ParentStackElem.second = ParentStackElem.first->getNumOperands(); | ||||
6618 | } else { | ||||
6619 | // We ran into something like: | ||||
6620 | // ParentStackElem.first += ... + ExtraArg + ... | ||||
6621 | ExtraArgs[ParentStackElem.first] = ExtraArg; | ||||
6622 | } | ||||
6623 | } | ||||
6624 | |||||
6625 | static OperationData getOperationData(Instruction *I) { | ||||
6626 | if (!I) | ||||
6627 | return OperationData(); | ||||
6628 | |||||
6629 | Value *LHS; | ||||
6630 | Value *RHS; | ||||
6631 | if (m_BinOp(m_Value(LHS), m_Value(RHS)).match(I)) { | ||||
6632 | return OperationData(cast<BinaryOperator>(I)->getOpcode(), RK_Arithmetic); | ||||
6633 | } | ||||
6634 | if (auto *Select = dyn_cast<SelectInst>(I)) { | ||||
6635 | // Look for a min/max pattern. | ||||
6636 | if (m_UMin(m_Value(LHS), m_Value(RHS)).match(Select)) { | ||||
6637 | return OperationData(Instruction::ICmp, RK_UMin); | ||||
6638 | } else if (m_SMin(m_Value(LHS), m_Value(RHS)).match(Select)) { | ||||
6639 | return OperationData(Instruction::ICmp, RK_SMin); | ||||
6640 | } else if (m_UMax(m_Value(LHS), m_Value(RHS)).match(Select)) { | ||||
6641 | return OperationData(Instruction::ICmp, RK_UMax); | ||||
6642 | } else if (m_SMax(m_Value(LHS), m_Value(RHS)).match(Select)) { | ||||
6643 | return OperationData(Instruction::ICmp, RK_SMax); | ||||
6644 | } else { | ||||
6645 | // Try harder: look for min/max pattern based on instructions producing | ||||
6646 | // same values such as: select ((cmp Inst1, Inst2), Inst1, Inst2). | ||||
6647 | // During the intermediate stages of SLP, it's very common to have | ||||
6648 | // pattern like this (since optimizeGatherSequence is run only once | ||||
6649 | // at the end): | ||||
6650 | // %1 = extractelement <2 x i32> %a, i32 0 | ||||
6651 | // %2 = extractelement <2 x i32> %a, i32 1 | ||||
6652 | // %cond = icmp sgt i32 %1, %2 | ||||
6653 | // %3 = extractelement <2 x i32> %a, i32 0 | ||||
6654 | // %4 = extractelement <2 x i32> %a, i32 1 | ||||
6655 | // %select = select i1 %cond, i32 %3, i32 %4 | ||||
6656 | CmpInst::Predicate Pred; | ||||
6657 | Instruction *L1; | ||||
6658 | Instruction *L2; | ||||
6659 | |||||
6660 | LHS = Select->getTrueValue(); | ||||
6661 | RHS = Select->getFalseValue(); | ||||
6662 | Value *Cond = Select->getCondition(); | ||||
6663 | |||||
6664 | // TODO: Support inverse predicates. | ||||
6665 | if (match(Cond, m_Cmp(Pred, m_Specific(LHS), m_Instruction(L2)))) { | ||||
6666 | if (!isa<ExtractElementInst>(RHS) || | ||||
6667 | !L2->isIdenticalTo(cast<Instruction>(RHS))) | ||||
6668 | return OperationData(*I); | ||||
6669 | } else if (match(Cond, m_Cmp(Pred, m_Instruction(L1), m_Specific(RHS)))) { | ||||
6670 | if (!isa<ExtractElementInst>(LHS) || | ||||
6671 | !L1->isIdenticalTo(cast<Instruction>(LHS))) | ||||
6672 | return OperationData(*I); | ||||
6673 | } else { | ||||
6674 | if (!isa<ExtractElementInst>(LHS) || !isa<ExtractElementInst>(RHS)) | ||||
6675 | return OperationData(*I); | ||||
6676 | if (!match(Cond, m_Cmp(Pred, m_Instruction(L1), m_Instruction(L2))) || | ||||
6677 | !L1->isIdenticalTo(cast<Instruction>(LHS)) || | ||||
6678 | !L2->isIdenticalTo(cast<Instruction>(RHS))) | ||||
6679 | return OperationData(*I); | ||||
6680 | } | ||||
6681 | switch (Pred) { | ||||
6682 | default: | ||||
6683 | return OperationData(*I); | ||||
6684 | |||||
6685 | case CmpInst::ICMP_ULT: | ||||
6686 | case CmpInst::ICMP_ULE: | ||||
6687 | return OperationData(Instruction::ICmp, RK_UMin); | ||||
6688 | |||||
6689 | case CmpInst::ICMP_SLT: | ||||
6690 | case CmpInst::ICMP_SLE: | ||||
6691 | return OperationData(Instruction::ICmp, RK_SMin); | ||||
6692 | |||||
6693 | case CmpInst::ICMP_UGT: | ||||
6694 | case CmpInst::ICMP_UGE: | ||||
6695 | return OperationData(Instruction::ICmp, RK_UMax); | ||||
6696 | |||||
6697 | case CmpInst::ICMP_SGT: | ||||
6698 | case CmpInst::ICMP_SGE: | ||||
6699 | return OperationData(Instruction::ICmp, RK_SMax); | ||||
6700 | } | ||||
6701 | } | ||||
6702 | } | ||||
6703 | return OperationData(*I); | ||||
6704 | } | ||||
6705 | |||||
6706 | public: | ||||
6707 | HorizontalReduction() = default; | ||||
6708 | |||||
6709 | /// Try to find a reduction tree. | ||||
6710 | bool matchAssociativeReduction(PHINode *Phi, Instruction *B) { | ||||
6711 | assert((!Phi || is_contained(Phi->operands(), B)) &&(((!Phi || is_contained(Phi->operands(), B)) && "Thi phi needs to use the binary operator" ) ? static_cast<void> (0) : __assert_fail ("(!Phi || is_contained(Phi->operands(), B)) && \"Thi phi needs to use the binary operator\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 6712, __PRETTY_FUNCTION__)) | ||||
6712 | "Thi phi needs to use the binary operator")(((!Phi || is_contained(Phi->operands(), B)) && "Thi phi needs to use the binary operator" ) ? static_cast<void> (0) : __assert_fail ("(!Phi || is_contained(Phi->operands(), B)) && \"Thi phi needs to use the binary operator\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 6712, __PRETTY_FUNCTION__)); | ||||
6713 | |||||
6714 | ReductionData = getOperationData(B); | ||||
6715 | |||||
6716 | // We could have a initial reductions that is not an add. | ||||
6717 | // r *= v1 + v2 + v3 + v4 | ||||
6718 | // In such a case start looking for a tree rooted in the first '+'. | ||||
6719 | if (Phi) { | ||||
6720 | if (ReductionData.getLHS(B) == Phi) { | ||||
6721 | Phi = nullptr; | ||||
6722 | B = dyn_cast<Instruction>(ReductionData.getRHS(B)); | ||||
6723 | ReductionData = getOperationData(B); | ||||
6724 | } else if (ReductionData.getRHS(B) == Phi) { | ||||
6725 | Phi = nullptr; | ||||
6726 | B = dyn_cast<Instruction>(ReductionData.getLHS(B)); | ||||
6727 | ReductionData = getOperationData(B); | ||||
6728 | } | ||||
6729 | } | ||||
6730 | |||||
6731 | if (!ReductionData.isVectorizable(B)) | ||||
6732 | return false; | ||||
6733 | |||||
6734 | Type *Ty = B->getType(); | ||||
6735 | if (!isValidElementType(Ty)) | ||||
6736 | return false; | ||||
6737 | if (!Ty->isIntOrIntVectorTy() && !Ty->isFPOrFPVectorTy()) | ||||
6738 | return false; | ||||
6739 | |||||
6740 | ReducedValueData.clear(); | ||||
6741 | ReductionRoot = B; | ||||
6742 | |||||
6743 | // Post order traverse the reduction tree starting at B. We only handle true | ||||
6744 | // trees containing only binary operators. | ||||
6745 | SmallVector<std::pair<Instruction *, unsigned>, 32> Stack; | ||||
6746 | Stack.push_back(std::make_pair(B, ReductionData.getFirstOperandIndex())); | ||||
6747 | ReductionData.initReductionOps(ReductionOps); | ||||
6748 | while (!Stack.empty()) { | ||||
6749 | Instruction *TreeN = Stack.back().first; | ||||
6750 | unsigned EdgeToVist = Stack.back().second++; | ||||
6751 | OperationData OpData = getOperationData(TreeN); | ||||
6752 | bool IsReducedValue = OpData != ReductionData; | ||||
6753 | |||||
6754 | // Postorder vist. | ||||
6755 | if (IsReducedValue || EdgeToVist == OpData.getNumberOfOperands()) { | ||||
6756 | if (IsReducedValue) | ||||
6757 | ReducedVals.push_back(TreeN); | ||||
6758 | else { | ||||
6759 | auto I = ExtraArgs.find(TreeN); | ||||
6760 | if (I != ExtraArgs.end() && !I->second) { | ||||
6761 | // Check if TreeN is an extra argument of its parent operation. | ||||
6762 | if (Stack.size() <= 1) { | ||||
6763 | // TreeN can't be an extra argument as it is a root reduction | ||||
6764 | // operation. | ||||
6765 | return false; | ||||
6766 | } | ||||
6767 | // Yes, TreeN is an extra argument, do not add it to a list of | ||||
6768 | // reduction operations. | ||||
6769 | // Stack[Stack.size() - 2] always points to the parent operation. | ||||
6770 | markExtraArg(Stack[Stack.size() - 2], TreeN); | ||||
6771 | ExtraArgs.erase(TreeN); | ||||
6772 | } else | ||||
6773 | ReductionData.addReductionOps(TreeN, ReductionOps); | ||||
6774 | } | ||||
6775 | // Retract. | ||||
6776 | Stack.pop_back(); | ||||
6777 | continue; | ||||
6778 | } | ||||
6779 | |||||
6780 | // Visit left or right. | ||||
6781 | Value *NextV = TreeN->getOperand(EdgeToVist); | ||||
6782 | if (NextV != Phi) { | ||||
6783 | auto *I = dyn_cast<Instruction>(NextV); | ||||
6784 | OpData = getOperationData(I); | ||||
6785 | // Continue analysis if the next operand is a reduction operation or | ||||
6786 | // (possibly) a reduced value. If the reduced value opcode is not set, | ||||
6787 | // the first met operation != reduction operation is considered as the | ||||
6788 | // reduced value class. | ||||
6789 | if (I && (!ReducedValueData || OpData == ReducedValueData || | ||||
6790 | OpData == ReductionData)) { | ||||
6791 | const bool IsReductionOperation = OpData == ReductionData; | ||||
6792 | // Only handle trees in the current basic block. | ||||
6793 | if (!ReductionData.hasSameParent(I, B->getParent(), | ||||
6794 | IsReductionOperation)) { | ||||
6795 | // I is an extra argument for TreeN (its parent operation). | ||||
6796 | markExtraArg(Stack.back(), I); | ||||
6797 | continue; | ||||
6798 | } | ||||
6799 | |||||
6800 | // Each tree node needs to have minimal number of users except for the | ||||
6801 | // ultimate reduction. | ||||
6802 | if (!ReductionData.hasRequiredNumberOfUses(I, | ||||
6803 | OpData == ReductionData) && | ||||
6804 | I != B) { | ||||
6805 | // I is an extra argument for TreeN (its parent operation). | ||||
6806 | markExtraArg(Stack.back(), I); | ||||
6807 | continue; | ||||
6808 | } | ||||
6809 | |||||
6810 | if (IsReductionOperation) { | ||||
6811 | // We need to be able to reassociate the reduction operations. | ||||
6812 | if (!OpData.isAssociative(I)) { | ||||
6813 | // I is an extra argument for TreeN (its parent operation). | ||||
6814 | markExtraArg(Stack.back(), I); | ||||
6815 | continue; | ||||
6816 | } | ||||
6817 | } else if (ReducedValueData && | ||||
6818 | ReducedValueData != OpData) { | ||||
6819 | // Make sure that the opcodes of the operations that we are going to | ||||
6820 | // reduce match. | ||||
6821 | // I is an extra argument for TreeN (its parent operation). | ||||
6822 | markExtraArg(Stack.back(), I); | ||||
6823 | continue; | ||||
6824 | } else if (!ReducedValueData) | ||||
6825 | ReducedValueData = OpData; | ||||
6826 | |||||
6827 | Stack.push_back(std::make_pair(I, OpData.getFirstOperandIndex())); | ||||
6828 | continue; | ||||
6829 | } | ||||
6830 | } | ||||
6831 | // NextV is an extra argument for TreeN (its parent operation). | ||||
6832 | markExtraArg(Stack.back(), NextV); | ||||
6833 | } | ||||
6834 | return true; | ||||
6835 | } | ||||
6836 | |||||
6837 | /// Attempt to vectorize the tree found by matchAssociativeReduction. | ||||
6838 | bool tryToReduce(BoUpSLP &V, TargetTransformInfo *TTI) { | ||||
6839 | // If there are a sufficient number of reduction values, reduce | ||||
6840 | // to a nearby power-of-2. We can safely generate oversized | ||||
6841 | // vectors and rely on the backend to split them to legal sizes. | ||||
6842 | unsigned NumReducedVals = ReducedVals.size(); | ||||
6843 | if (NumReducedVals < 4) | ||||
6844 | return false; | ||||
6845 | |||||
6846 | // FIXME: Fast-math-flags should be set based on the instructions in the | ||||
6847 | // reduction (not all of 'fast' are required). | ||||
6848 | IRBuilder<> Builder(cast<Instruction>(ReductionRoot)); | ||||
6849 | FastMathFlags Unsafe; | ||||
6850 | Unsafe.setFast(); | ||||
6851 | Builder.setFastMathFlags(Unsafe); | ||||
6852 | |||||
6853 | BoUpSLP::ExtraValueToDebugLocsMap ExternallyUsedValues; | ||||
6854 | // The same extra argument may be used several times, so log each attempt | ||||
6855 | // to use it. | ||||
6856 | for (const std::pair<Instruction *, Value *> &Pair : ExtraArgs) { | ||||
6857 | assert(Pair.first && "DebugLoc must be set.")((Pair.first && "DebugLoc must be set.") ? static_cast <void> (0) : __assert_fail ("Pair.first && \"DebugLoc must be set.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 6857, __PRETTY_FUNCTION__)); | ||||
6858 | ExternallyUsedValues[Pair.second].push_back(Pair.first); | ||||
6859 | } | ||||
6860 | |||||
6861 | // The compare instruction of a min/max is the insertion point for new | ||||
6862 | // instructions and may be replaced with a new compare instruction. | ||||
6863 | auto getCmpForMinMaxReduction = [](Instruction *RdxRootInst) { | ||||
6864 | assert(isa<SelectInst>(RdxRootInst) &&((isa<SelectInst>(RdxRootInst) && "Expected min/max reduction to have select root instruction" ) ? static_cast<void> (0) : __assert_fail ("isa<SelectInst>(RdxRootInst) && \"Expected min/max reduction to have select root instruction\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 6865, __PRETTY_FUNCTION__)) | ||||
6865 | "Expected min/max reduction to have select root instruction")((isa<SelectInst>(RdxRootInst) && "Expected min/max reduction to have select root instruction" ) ? static_cast<void> (0) : __assert_fail ("isa<SelectInst>(RdxRootInst) && \"Expected min/max reduction to have select root instruction\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 6865, __PRETTY_FUNCTION__)); | ||||
6866 | Value *ScalarCond = cast<SelectInst>(RdxRootInst)->getCondition(); | ||||
6867 | assert(isa<Instruction>(ScalarCond) &&((isa<Instruction>(ScalarCond) && "Expected min/max reduction to have compare condition" ) ? static_cast<void> (0) : __assert_fail ("isa<Instruction>(ScalarCond) && \"Expected min/max reduction to have compare condition\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 6868, __PRETTY_FUNCTION__)) | ||||
6868 | "Expected min/max reduction to have compare condition")((isa<Instruction>(ScalarCond) && "Expected min/max reduction to have compare condition" ) ? static_cast<void> (0) : __assert_fail ("isa<Instruction>(ScalarCond) && \"Expected min/max reduction to have compare condition\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 6868, __PRETTY_FUNCTION__)); | ||||
6869 | return cast<Instruction>(ScalarCond); | ||||
6870 | }; | ||||
6871 | |||||
6872 | // The reduction root is used as the insertion point for new instructions, | ||||
6873 | // so set it as externally used to prevent it from being deleted. | ||||
6874 | ExternallyUsedValues[ReductionRoot]; | ||||
6875 | SmallVector<Value *, 16> IgnoreList; | ||||
6876 | for (ReductionOpsType &RdxOp : ReductionOps) | ||||
6877 | IgnoreList.append(RdxOp.begin(), RdxOp.end()); | ||||
6878 | |||||
6879 | unsigned ReduxWidth = PowerOf2Floor(NumReducedVals); | ||||
6880 | if (NumReducedVals > ReduxWidth) { | ||||
6881 | // In the loop below, we are building a tree based on a window of | ||||
6882 | // 'ReduxWidth' values. | ||||
6883 | // If the operands of those values have common traits (compare predicate, | ||||
6884 | // constant operand, etc), then we want to group those together to | ||||
6885 | // minimize the cost of the reduction. | ||||
6886 | |||||
6887 | // TODO: This should be extended to count common operands for | ||||
6888 | // compares and binops. | ||||
6889 | |||||
6890 | // Step 1: Count the number of times each compare predicate occurs. | ||||
6891 | SmallDenseMap<unsigned, unsigned> PredCountMap; | ||||
6892 | for (Value *RdxVal : ReducedVals) { | ||||
6893 | CmpInst::Predicate Pred; | ||||
6894 | if (match(RdxVal, m_Cmp(Pred, m_Value(), m_Value()))) | ||||
6895 | ++PredCountMap[Pred]; | ||||
6896 | } | ||||
6897 | // Step 2: Sort the values so the most common predicates come first. | ||||
6898 | stable_sort(ReducedVals, [&PredCountMap](Value *A, Value *B) { | ||||
6899 | CmpInst::Predicate PredA, PredB; | ||||
6900 | if (match(A, m_Cmp(PredA, m_Value(), m_Value())) && | ||||
6901 | match(B, m_Cmp(PredB, m_Value(), m_Value()))) { | ||||
6902 | return PredCountMap[PredA] > PredCountMap[PredB]; | ||||
6903 | } | ||||
6904 | return false; | ||||
6905 | }); | ||||
6906 | } | ||||
6907 | |||||
6908 | Value *VectorizedTree = nullptr; | ||||
6909 | unsigned i = 0; | ||||
6910 | while (i < NumReducedVals - ReduxWidth + 1 && ReduxWidth > 2) { | ||||
6911 | ArrayRef<Value *> VL(&ReducedVals[i], ReduxWidth); | ||||
6912 | V.buildTree(VL, ExternallyUsedValues, IgnoreList); | ||||
6913 | Optional<ArrayRef<unsigned>> Order = V.bestOrder(); | ||||
6914 | if (Order) { | ||||
6915 | assert(Order->size() == VL.size() &&((Order->size() == VL.size() && "Order size must be the same as number of vectorized " "instructions.") ? static_cast<void> (0) : __assert_fail ("Order->size() == VL.size() && \"Order size must be the same as number of vectorized \" \"instructions.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 6917, __PRETTY_FUNCTION__)) | ||||
6916 | "Order size must be the same as number of vectorized "((Order->size() == VL.size() && "Order size must be the same as number of vectorized " "instructions.") ? static_cast<void> (0) : __assert_fail ("Order->size() == VL.size() && \"Order size must be the same as number of vectorized \" \"instructions.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 6917, __PRETTY_FUNCTION__)) | ||||
6917 | "instructions.")((Order->size() == VL.size() && "Order size must be the same as number of vectorized " "instructions.") ? static_cast<void> (0) : __assert_fail ("Order->size() == VL.size() && \"Order size must be the same as number of vectorized \" \"instructions.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 6917, __PRETTY_FUNCTION__)); | ||||
6918 | // TODO: reorder tree nodes without tree rebuilding. | ||||
6919 | SmallVector<Value *, 4> ReorderedOps(VL.size()); | ||||
6920 | llvm::transform(*Order, ReorderedOps.begin(), | ||||
6921 | [VL](const unsigned Idx) { return VL[Idx]; }); | ||||
6922 | V.buildTree(ReorderedOps, ExternallyUsedValues, IgnoreList); | ||||
6923 | } | ||||
6924 | if (V.isTreeTinyAndNotFullyVectorizable()) | ||||
6925 | break; | ||||
6926 | if (V.isLoadCombineReductionCandidate(ReductionData.getOpcode())) | ||||
6927 | break; | ||||
6928 | |||||
6929 | V.computeMinimumValueSizes(); | ||||
6930 | |||||
6931 | // Estimate cost. | ||||
6932 | int TreeCost = V.getTreeCost(); | ||||
6933 | int ReductionCost = getReductionCost(TTI, ReducedVals[i], ReduxWidth); | ||||
6934 | int Cost = TreeCost + ReductionCost; | ||||
6935 | if (Cost >= -SLPCostThreshold) { | ||||
6936 | V.getORE()->emit([&]() { | ||||
6937 | return OptimizationRemarkMissed(SV_NAME"slp-vectorizer", "HorSLPNotBeneficial", | ||||
6938 | cast<Instruction>(VL[0])) | ||||
6939 | << "Vectorizing horizontal reduction is possible" | ||||
6940 | << "but not beneficial with cost " << ore::NV("Cost", Cost) | ||||
6941 | << " and threshold " | ||||
6942 | << ore::NV("Threshold", -SLPCostThreshold); | ||||
6943 | }); | ||||
6944 | break; | ||||
6945 | } | ||||
6946 | |||||
6947 | 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) | ||||
6948 | << Cost << ". (HorRdx)\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Vectorizing horizontal reduction at cost:" << Cost << ". (HorRdx)\n"; } } while (false); | ||||
6949 | V.getORE()->emit([&]() { | ||||
6950 | return OptimizationRemark(SV_NAME"slp-vectorizer", "VectorizedHorizontalReduction", | ||||
6951 | cast<Instruction>(VL[0])) | ||||
6952 | << "Vectorized horizontal reduction with cost " | ||||
6953 | << ore::NV("Cost", Cost) << " and with tree size " | ||||
6954 | << ore::NV("TreeSize", V.getTreeSize()); | ||||
6955 | }); | ||||
6956 | |||||
6957 | // Vectorize a tree. | ||||
6958 | DebugLoc Loc = cast<Instruction>(ReducedVals[i])->getDebugLoc(); | ||||
6959 | Value *VectorizedRoot = V.vectorizeTree(ExternallyUsedValues); | ||||
6960 | |||||
6961 | // Emit a reduction. For min/max, the root is a select, but the insertion | ||||
6962 | // point is the compare condition of that select. | ||||
6963 | Instruction *RdxRootInst = cast<Instruction>(ReductionRoot); | ||||
6964 | if (ReductionData.isMinMax()) | ||||
6965 | Builder.SetInsertPoint(getCmpForMinMaxReduction(RdxRootInst)); | ||||
6966 | else | ||||
6967 | Builder.SetInsertPoint(RdxRootInst); | ||||
6968 | |||||
6969 | Value *ReducedSubTree = | ||||
6970 | emitReduction(VectorizedRoot, Builder, ReduxWidth, TTI); | ||||
6971 | |||||
6972 | if (!VectorizedTree) { | ||||
6973 | // Initialize the final value in the reduction. | ||||
6974 | VectorizedTree = ReducedSubTree; | ||||
6975 | } else { | ||||
6976 | // Update the final value in the reduction. | ||||
6977 | Builder.SetCurrentDebugLocation(Loc); | ||||
6978 | VectorizedTree = ReductionData.createOp( | ||||
6979 | Builder, VectorizedTree, ReducedSubTree, "op.rdx", ReductionOps); | ||||
6980 | } | ||||
6981 | i += ReduxWidth; | ||||
6982 | ReduxWidth = PowerOf2Floor(NumReducedVals - i); | ||||
6983 | } | ||||
6984 | |||||
6985 | if (VectorizedTree) { | ||||
6986 | // Finish the reduction. | ||||
6987 | for (; i < NumReducedVals; ++i) { | ||||
6988 | auto *I = cast<Instruction>(ReducedVals[i]); | ||||
6989 | Builder.SetCurrentDebugLocation(I->getDebugLoc()); | ||||
6990 | VectorizedTree = ReductionData.createOp(Builder, VectorizedTree, I, "", | ||||
6991 | ReductionOps); | ||||
6992 | } | ||||
6993 | for (auto &Pair : ExternallyUsedValues) { | ||||
6994 | // Add each externally used value to the final reduction. | ||||
6995 | for (auto *I : Pair.second) { | ||||
6996 | Builder.SetCurrentDebugLocation(I->getDebugLoc()); | ||||
6997 | VectorizedTree = ReductionData.createOp(Builder, VectorizedTree, | ||||
6998 | Pair.first, "op.extra", I); | ||||
6999 | } | ||||
7000 | } | ||||
7001 | |||||
7002 | // Update users. For a min/max reduction that ends with a compare and | ||||
7003 | // select, we also have to RAUW for the compare instruction feeding the | ||||
7004 | // reduction root. That's because the original compare may have extra uses | ||||
7005 | // besides the final select of the reduction. | ||||
7006 | if (ReductionData.isMinMax()) { | ||||
7007 | if (auto *VecSelect = dyn_cast<SelectInst>(VectorizedTree)) { | ||||
7008 | Instruction *ScalarCmp = | ||||
7009 | getCmpForMinMaxReduction(cast<Instruction>(ReductionRoot)); | ||||
7010 | ScalarCmp->replaceAllUsesWith(VecSelect->getCondition()); | ||||
7011 | } | ||||
7012 | } | ||||
7013 | ReductionRoot->replaceAllUsesWith(VectorizedTree); | ||||
7014 | |||||
7015 | // Mark all scalar reduction ops for deletion, they are replaced by the | ||||
7016 | // vector reductions. | ||||
7017 | V.eraseInstructions(IgnoreList); | ||||
7018 | } | ||||
7019 | return VectorizedTree != nullptr; | ||||
7020 | } | ||||
7021 | |||||
7022 | unsigned numReductionValues() const { | ||||
7023 | return ReducedVals.size(); | ||||
7024 | } | ||||
7025 | |||||
7026 | private: | ||||
7027 | /// Calculate the cost of a reduction. | ||||
7028 | int getReductionCost(TargetTransformInfo *TTI, Value *FirstReducedVal, | ||||
7029 | unsigned ReduxWidth) { | ||||
7030 | Type *ScalarTy = FirstReducedVal->getType(); | ||||
7031 | auto *VecTy = FixedVectorType::get(ScalarTy, ReduxWidth); | ||||
7032 | |||||
7033 | int PairwiseRdxCost; | ||||
7034 | int SplittingRdxCost; | ||||
7035 | switch (ReductionData.getKind()) { | ||||
7036 | case RK_Arithmetic: | ||||
7037 | PairwiseRdxCost = | ||||
7038 | TTI->getArithmeticReductionCost(ReductionData.getOpcode(), VecTy, | ||||
7039 | /*IsPairwiseForm=*/true); | ||||
7040 | SplittingRdxCost = | ||||
7041 | TTI->getArithmeticReductionCost(ReductionData.getOpcode(), VecTy, | ||||
7042 | /*IsPairwiseForm=*/false); | ||||
7043 | break; | ||||
7044 | case RK_SMin: | ||||
7045 | case RK_SMax: | ||||
7046 | case RK_UMin: | ||||
7047 | case RK_UMax: { | ||||
7048 | auto *VecCondTy = cast<VectorType>(CmpInst::makeCmpResultType(VecTy)); | ||||
7049 | bool IsUnsigned = ReductionData.getKind() == RK_UMin || | ||||
7050 | ReductionData.getKind() == RK_UMax; | ||||
7051 | PairwiseRdxCost = | ||||
7052 | TTI->getMinMaxReductionCost(VecTy, VecCondTy, | ||||
7053 | /*IsPairwiseForm=*/true, IsUnsigned); | ||||
7054 | SplittingRdxCost = | ||||
7055 | TTI->getMinMaxReductionCost(VecTy, VecCondTy, | ||||
7056 | /*IsPairwiseForm=*/false, IsUnsigned); | ||||
7057 | break; | ||||
7058 | } | ||||
7059 | case RK_None: | ||||
7060 | llvm_unreachable("Expected arithmetic or min/max reduction operation")::llvm::llvm_unreachable_internal("Expected arithmetic or min/max reduction operation" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 7060); | ||||
7061 | } | ||||
7062 | |||||
7063 | IsPairwiseReduction = PairwiseRdxCost < SplittingRdxCost; | ||||
7064 | int VecReduxCost = IsPairwiseReduction ? PairwiseRdxCost : SplittingRdxCost; | ||||
7065 | |||||
7066 | int ScalarReduxCost = 0; | ||||
7067 | switch (ReductionData.getKind()) { | ||||
7068 | case RK_Arithmetic: | ||||
7069 | ScalarReduxCost = | ||||
7070 | TTI->getArithmeticInstrCost(ReductionData.getOpcode(), ScalarTy); | ||||
7071 | break; | ||||
7072 | case RK_SMin: | ||||
7073 | case RK_SMax: | ||||
7074 | case RK_UMin: | ||||
7075 | case RK_UMax: | ||||
7076 | ScalarReduxCost = | ||||
7077 | TTI->getCmpSelInstrCost(ReductionData.getOpcode(), ScalarTy) + | ||||
7078 | TTI->getCmpSelInstrCost(Instruction::Select, ScalarTy, | ||||
7079 | CmpInst::makeCmpResultType(ScalarTy)); | ||||
7080 | break; | ||||
7081 | case RK_None: | ||||
7082 | llvm_unreachable("Expected arithmetic or min/max reduction operation")::llvm::llvm_unreachable_internal("Expected arithmetic or min/max reduction operation" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 7082); | ||||
7083 | } | ||||
7084 | ScalarReduxCost *= (ReduxWidth - 1); | ||||
7085 | |||||
7086 | LLVM_DEBUG(dbgs() << "SLP: Adding cost " << VecReduxCost - ScalarReduxCostdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Adding cost " << VecReduxCost - ScalarReduxCost << " for reduction that starts with " << *FirstReducedVal << " (It is a " << (IsPairwiseReduction ? "pairwise" : "splitting") << " reduction)\n"; } } while (false) | ||||
7087 | << " for reduction that starts with " << *FirstReducedValdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Adding cost " << VecReduxCost - ScalarReduxCost << " for reduction that starts with " << *FirstReducedVal << " (It is a " << (IsPairwiseReduction ? "pairwise" : "splitting") << " reduction)\n"; } } while (false) | ||||
7088 | << " (It is a "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Adding cost " << VecReduxCost - ScalarReduxCost << " for reduction that starts with " << *FirstReducedVal << " (It is a " << (IsPairwiseReduction ? "pairwise" : "splitting") << " reduction)\n"; } } while (false) | ||||
7089 | << (IsPairwiseReduction ? "pairwise" : "splitting")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Adding cost " << VecReduxCost - ScalarReduxCost << " for reduction that starts with " << *FirstReducedVal << " (It is a " << (IsPairwiseReduction ? "pairwise" : "splitting") << " reduction)\n"; } } while (false) | ||||
7090 | << " reduction)\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Adding cost " << VecReduxCost - ScalarReduxCost << " for reduction that starts with " << *FirstReducedVal << " (It is a " << (IsPairwiseReduction ? "pairwise" : "splitting") << " reduction)\n"; } } while (false); | ||||
7091 | |||||
7092 | return VecReduxCost - ScalarReduxCost; | ||||
7093 | } | ||||
7094 | |||||
7095 | /// Emit a horizontal reduction of the vectorized value. | ||||
7096 | Value *emitReduction(Value *VectorizedValue, IRBuilder<> &Builder, | ||||
7097 | unsigned ReduxWidth, const TargetTransformInfo *TTI) { | ||||
7098 | assert(VectorizedValue && "Need to have a vectorized tree node")((VectorizedValue && "Need to have a vectorized tree node" ) ? static_cast<void> (0) : __assert_fail ("VectorizedValue && \"Need to have a vectorized tree node\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 7098, __PRETTY_FUNCTION__)); | ||||
7099 | assert(isPowerOf2_32(ReduxWidth) &&((isPowerOf2_32(ReduxWidth) && "We only handle power-of-two reductions for now" ) ? static_cast<void> (0) : __assert_fail ("isPowerOf2_32(ReduxWidth) && \"We only handle power-of-two reductions for now\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 7100, __PRETTY_FUNCTION__)) | ||||
7100 | "We only handle power-of-two reductions for now")((isPowerOf2_32(ReduxWidth) && "We only handle power-of-two reductions for now" ) ? static_cast<void> (0) : __assert_fail ("isPowerOf2_32(ReduxWidth) && \"We only handle power-of-two reductions for now\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 7100, __PRETTY_FUNCTION__)); | ||||
7101 | |||||
7102 | if (!IsPairwiseReduction) { | ||||
7103 | // FIXME: The builder should use an FMF guard. It should not be hard-coded | ||||
7104 | // to 'fast'. | ||||
7105 | assert(Builder.getFastMathFlags().isFast() && "Expected 'fast' FMF")((Builder.getFastMathFlags().isFast() && "Expected 'fast' FMF" ) ? static_cast<void> (0) : __assert_fail ("Builder.getFastMathFlags().isFast() && \"Expected 'fast' FMF\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 7105, __PRETTY_FUNCTION__)); | ||||
7106 | return createSimpleTargetReduction( | ||||
7107 | Builder, TTI, ReductionData.getOpcode(), VectorizedValue, | ||||
7108 | ReductionData.getFlags(), ReductionOps.back()); | ||||
7109 | } | ||||
7110 | |||||
7111 | Value *TmpVec = VectorizedValue; | ||||
7112 | for (unsigned i = ReduxWidth / 2; i != 0; i >>= 1) { | ||||
7113 | auto LeftMask = createRdxShuffleMask(ReduxWidth, i, true, true); | ||||
7114 | auto RightMask = createRdxShuffleMask(ReduxWidth, i, true, false); | ||||
7115 | |||||
7116 | Value *LeftShuf = | ||||
7117 | Builder.CreateShuffleVector(TmpVec, LeftMask, "rdx.shuf.l"); | ||||
7118 | Value *RightShuf = | ||||
7119 | Builder.CreateShuffleVector(TmpVec, RightMask, "rdx.shuf.r"); | ||||
7120 | TmpVec = ReductionData.createOp(Builder, LeftShuf, RightShuf, "op.rdx", | ||||
7121 | ReductionOps); | ||||
7122 | } | ||||
7123 | |||||
7124 | // The result is in the first element of the vector. | ||||
7125 | return Builder.CreateExtractElement(TmpVec, Builder.getInt32(0)); | ||||
7126 | } | ||||
7127 | }; | ||||
7128 | |||||
7129 | } // end anonymous namespace | ||||
7130 | |||||
7131 | static Optional<unsigned> getAggregateSize(Instruction *InsertInst) { | ||||
7132 | if (auto *IE = dyn_cast<InsertElementInst>(InsertInst)) | ||||
7133 | return cast<FixedVectorType>(IE->getType())->getNumElements(); | ||||
7134 | |||||
7135 | unsigned AggregateSize = 1; | ||||
7136 | auto *IV = cast<InsertValueInst>(InsertInst); | ||||
7137 | Type *CurrentType = IV->getType(); | ||||
7138 | do { | ||||
7139 | if (auto *ST = dyn_cast<StructType>(CurrentType)) { | ||||
7140 | for (auto *Elt : ST->elements()) | ||||
7141 | if (Elt != ST->getElementType(0)) // check homogeneity | ||||
7142 | return None; | ||||
7143 | AggregateSize *= ST->getNumElements(); | ||||
7144 | CurrentType = ST->getElementType(0); | ||||
7145 | } else if (auto *AT = dyn_cast<ArrayType>(CurrentType)) { | ||||
7146 | AggregateSize *= AT->getNumElements(); | ||||
7147 | CurrentType = AT->getElementType(); | ||||
7148 | } else if (auto *VT = dyn_cast<FixedVectorType>(CurrentType)) { | ||||
7149 | AggregateSize *= VT->getNumElements(); | ||||
7150 | return AggregateSize; | ||||
7151 | } else if (CurrentType->isSingleValueType()) { | ||||
7152 | return AggregateSize; | ||||
7153 | } else { | ||||
7154 | return None; | ||||
7155 | } | ||||
7156 | } while (true); | ||||
7157 | } | ||||
7158 | |||||
7159 | static Optional<unsigned> getOperandIndex(Instruction *InsertInst, | ||||
7160 | unsigned OperandOffset) { | ||||
7161 | unsigned OperandIndex = OperandOffset; | ||||
7162 | if (auto *IE = dyn_cast<InsertElementInst>(InsertInst)) { | ||||
7163 | if (auto *CI = dyn_cast<ConstantInt>(IE->getOperand(2))) { | ||||
7164 | auto *VT = cast<FixedVectorType>(IE->getType()); | ||||
7165 | OperandIndex *= VT->getNumElements(); | ||||
7166 | OperandIndex += CI->getZExtValue(); | ||||
7167 | return OperandIndex; | ||||
7168 | } | ||||
7169 | return None; | ||||
7170 | } | ||||
7171 | |||||
7172 | auto *IV = cast<InsertValueInst>(InsertInst); | ||||
7173 | Type *CurrentType = IV->getType(); | ||||
7174 | for (unsigned int Index : IV->indices()) { | ||||
7175 | if (auto *ST = dyn_cast<StructType>(CurrentType)) { | ||||
7176 | OperandIndex *= ST->getNumElements(); | ||||
7177 | CurrentType = ST->getElementType(Index); | ||||
7178 | } else if (auto *AT = dyn_cast<ArrayType>(CurrentType)) { | ||||
7179 | OperandIndex *= AT->getNumElements(); | ||||
7180 | CurrentType = AT->getElementType(); | ||||
7181 | } else { | ||||
7182 | return None; | ||||
7183 | } | ||||
7184 | OperandIndex += Index; | ||||
7185 | } | ||||
7186 | return OperandIndex; | ||||
7187 | } | ||||
7188 | |||||
7189 | static bool findBuildAggregate_rec(Instruction *LastInsertInst, | ||||
7190 | TargetTransformInfo *TTI, | ||||
7191 | SmallVectorImpl<Value *> &BuildVectorOpds, | ||||
7192 | SmallVectorImpl<Value *> &InsertElts, | ||||
7193 | unsigned OperandOffset) { | ||||
7194 | do { | ||||
7195 | Value *InsertedOperand = LastInsertInst->getOperand(1); | ||||
7196 | Optional<unsigned> OperandIndex = | ||||
7197 | getOperandIndex(LastInsertInst, OperandOffset); | ||||
7198 | if (!OperandIndex) | ||||
7199 | return false; | ||||
7200 | if (isa<InsertElementInst>(InsertedOperand) || | ||||
7201 | isa<InsertValueInst>(InsertedOperand)) { | ||||
7202 | if (!findBuildAggregate_rec(cast<Instruction>(InsertedOperand), TTI, | ||||
7203 | BuildVectorOpds, InsertElts, *OperandIndex)) | ||||
7204 | return false; | ||||
7205 | } else { | ||||
7206 | BuildVectorOpds[*OperandIndex] = InsertedOperand; | ||||
7207 | InsertElts[*OperandIndex] = LastInsertInst; | ||||
7208 | } | ||||
7209 | if (isa<UndefValue>(LastInsertInst->getOperand(0))) | ||||
7210 | return true; | ||||
7211 | LastInsertInst = dyn_cast<Instruction>(LastInsertInst->getOperand(0)); | ||||
7212 | } while (LastInsertInst != nullptr && | ||||
7213 | (isa<InsertValueInst>(LastInsertInst) || | ||||
7214 | isa<InsertElementInst>(LastInsertInst)) && | ||||
7215 | LastInsertInst->hasOneUse()); | ||||
7216 | return false; | ||||
7217 | } | ||||
7218 | |||||
7219 | /// Recognize construction of vectors like | ||||
7220 | /// %ra = insertelement <4 x float> undef, float %s0, i32 0 | ||||
7221 | /// %rb = insertelement <4 x float> %ra, float %s1, i32 1 | ||||
7222 | /// %rc = insertelement <4 x float> %rb, float %s2, i32 2 | ||||
7223 | /// %rd = insertelement <4 x float> %rc, float %s3, i32 3 | ||||
7224 | /// starting from the last insertelement or insertvalue instruction. | ||||
7225 | /// | ||||
7226 | /// Also recognize homogeneous aggregates like {<2 x float>, <2 x float>}, | ||||
7227 | /// {{float, float}, {float, float}}, [2 x {float, float}] and so on. | ||||
7228 | /// See llvm/test/Transforms/SLPVectorizer/X86/pr42022.ll for examples. | ||||
7229 | /// | ||||
7230 | /// Assume LastInsertInst is of InsertElementInst or InsertValueInst type. | ||||
7231 | /// | ||||
7232 | /// \return true if it matches. | ||||
7233 | static bool findBuildAggregate(Instruction *LastInsertInst, | ||||
7234 | TargetTransformInfo *TTI, | ||||
7235 | SmallVectorImpl<Value *> &BuildVectorOpds, | ||||
7236 | SmallVectorImpl<Value *> &InsertElts) { | ||||
7237 | |||||
7238 | assert((isa<InsertElementInst>(LastInsertInst) ||(((isa<InsertElementInst>(LastInsertInst) || isa<InsertValueInst >(LastInsertInst)) && "Expected insertelement or insertvalue instruction!" ) ? static_cast<void> (0) : __assert_fail ("(isa<InsertElementInst>(LastInsertInst) || isa<InsertValueInst>(LastInsertInst)) && \"Expected insertelement or insertvalue instruction!\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 7240, __PRETTY_FUNCTION__)) | ||||
7239 | isa<InsertValueInst>(LastInsertInst)) &&(((isa<InsertElementInst>(LastInsertInst) || isa<InsertValueInst >(LastInsertInst)) && "Expected insertelement or insertvalue instruction!" ) ? static_cast<void> (0) : __assert_fail ("(isa<InsertElementInst>(LastInsertInst) || isa<InsertValueInst>(LastInsertInst)) && \"Expected insertelement or insertvalue instruction!\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 7240, __PRETTY_FUNCTION__)) | ||||
7240 | "Expected insertelement or insertvalue instruction!")(((isa<InsertElementInst>(LastInsertInst) || isa<InsertValueInst >(LastInsertInst)) && "Expected insertelement or insertvalue instruction!" ) ? static_cast<void> (0) : __assert_fail ("(isa<InsertElementInst>(LastInsertInst) || isa<InsertValueInst>(LastInsertInst)) && \"Expected insertelement or insertvalue instruction!\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 7240, __PRETTY_FUNCTION__)); | ||||
7241 | |||||
7242 | assert((BuildVectorOpds.empty() && InsertElts.empty()) &&(((BuildVectorOpds.empty() && InsertElts.empty()) && "Expected empty result vectors!") ? static_cast<void> ( 0) : __assert_fail ("(BuildVectorOpds.empty() && InsertElts.empty()) && \"Expected empty result vectors!\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 7243, __PRETTY_FUNCTION__)) | ||||
7243 | "Expected empty result vectors!")(((BuildVectorOpds.empty() && InsertElts.empty()) && "Expected empty result vectors!") ? static_cast<void> ( 0) : __assert_fail ("(BuildVectorOpds.empty() && InsertElts.empty()) && \"Expected empty result vectors!\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 7243, __PRETTY_FUNCTION__)); | ||||
7244 | |||||
7245 | Optional<unsigned> AggregateSize = getAggregateSize(LastInsertInst); | ||||
7246 | if (!AggregateSize) | ||||
7247 | return false; | ||||
7248 | BuildVectorOpds.resize(*AggregateSize); | ||||
7249 | InsertElts.resize(*AggregateSize); | ||||
7250 | |||||
7251 | if (findBuildAggregate_rec(LastInsertInst, TTI, BuildVectorOpds, InsertElts, | ||||
7252 | 0)) { | ||||
7253 | llvm::erase_if(BuildVectorOpds, | ||||
7254 | [](const Value *V) { return V == nullptr; }); | ||||
7255 | llvm::erase_if(InsertElts, [](const Value *V) { return V == nullptr; }); | ||||
7256 | if (BuildVectorOpds.size() >= 2) | ||||
7257 | return true; | ||||
7258 | } | ||||
7259 | |||||
7260 | return false; | ||||
7261 | } | ||||
7262 | |||||
7263 | static bool PhiTypeSorterFunc(Value *V, Value *V2) { | ||||
7264 | return V->getType() < V2->getType(); | ||||
7265 | } | ||||
7266 | |||||
7267 | /// Try and get a reduction value from a phi node. | ||||
7268 | /// | ||||
7269 | /// Given a phi node \p P in a block \p ParentBB, consider possible reductions | ||||
7270 | /// if they come from either \p ParentBB or a containing loop latch. | ||||
7271 | /// | ||||
7272 | /// \returns A candidate reduction value if possible, or \code nullptr \endcode | ||||
7273 | /// if not possible. | ||||
7274 | static Value *getReductionValue(const DominatorTree *DT, PHINode *P, | ||||
7275 | BasicBlock *ParentBB, LoopInfo *LI) { | ||||
7276 | // There are situations where the reduction value is not dominated by the | ||||
7277 | // reduction phi. Vectorizing such cases has been reported to cause | ||||
7278 | // miscompiles. See PR25787. | ||||
7279 | auto DominatedReduxValue = [&](Value *R) { | ||||
7280 | return isa<Instruction>(R) && | ||||
7281 | DT->dominates(P->getParent(), cast<Instruction>(R)->getParent()); | ||||
7282 | }; | ||||
7283 | |||||
7284 | Value *Rdx = nullptr; | ||||
7285 | |||||
7286 | // Return the incoming value if it comes from the same BB as the phi node. | ||||
7287 | if (P->getIncomingBlock(0) == ParentBB) { | ||||
7288 | Rdx = P->getIncomingValue(0); | ||||
7289 | } else if (P->getIncomingBlock(1) == ParentBB) { | ||||
7290 | Rdx = P->getIncomingValue(1); | ||||
7291 | } | ||||
7292 | |||||
7293 | if (Rdx && DominatedReduxValue(Rdx)) | ||||
7294 | return Rdx; | ||||
7295 | |||||
7296 | // Otherwise, check whether we have a loop latch to look at. | ||||
7297 | Loop *BBL = LI->getLoopFor(ParentBB); | ||||
7298 | if (!BBL) | ||||
7299 | return nullptr; | ||||
7300 | BasicBlock *BBLatch = BBL->getLoopLatch(); | ||||
7301 | if (!BBLatch) | ||||
7302 | return nullptr; | ||||
7303 | |||||
7304 | // There is a loop latch, return the incoming value if it comes from | ||||
7305 | // that. This reduction pattern occasionally turns up. | ||||
7306 | if (P->getIncomingBlock(0) == BBLatch) { | ||||
7307 | Rdx = P->getIncomingValue(0); | ||||
7308 | } else if (P->getIncomingBlock(1) == BBLatch) { | ||||
7309 | Rdx = P->getIncomingValue(1); | ||||
7310 | } | ||||
7311 | |||||
7312 | if (Rdx && DominatedReduxValue(Rdx)) | ||||
7313 | return Rdx; | ||||
7314 | |||||
7315 | return nullptr; | ||||
7316 | } | ||||
7317 | |||||
7318 | /// Attempt to reduce a horizontal reduction. | ||||
7319 | /// If it is legal to match a horizontal reduction feeding the phi node \a P | ||||
7320 | /// with reduction operators \a Root (or one of its operands) in a basic block | ||||
7321 | /// \a BB, then check if it can be done. If horizontal reduction is not found | ||||
7322 | /// and root instruction is a binary operation, vectorization of the operands is | ||||
7323 | /// attempted. | ||||
7324 | /// \returns true if a horizontal reduction was matched and reduced or operands | ||||
7325 | /// of one of the binary instruction were vectorized. | ||||
7326 | /// \returns false if a horizontal reduction was not matched (or not possible) | ||||
7327 | /// or no vectorization of any binary operation feeding \a Root instruction was | ||||
7328 | /// performed. | ||||
7329 | static bool tryToVectorizeHorReductionOrInstOperands( | ||||
7330 | PHINode *P, Instruction *Root, BasicBlock *BB, BoUpSLP &R, | ||||
7331 | TargetTransformInfo *TTI, | ||||
7332 | const function_ref<bool(Instruction *, BoUpSLP &)> Vectorize) { | ||||
7333 | if (!ShouldVectorizeHor) | ||||
7334 | return false; | ||||
7335 | |||||
7336 | if (!Root) | ||||
7337 | return false; | ||||
7338 | |||||
7339 | if (Root->getParent() != BB || isa<PHINode>(Root)) | ||||
7340 | return false; | ||||
7341 | // Start analysis starting from Root instruction. If horizontal reduction is | ||||
7342 | // found, try to vectorize it. If it is not a horizontal reduction or | ||||
7343 | // vectorization is not possible or not effective, and currently analyzed | ||||
7344 | // instruction is a binary operation, try to vectorize the operands, using | ||||
7345 | // pre-order DFS traversal order. If the operands were not vectorized, repeat | ||||
7346 | // the same procedure considering each operand as a possible root of the | ||||
7347 | // horizontal reduction. | ||||
7348 | // Interrupt the process if the Root instruction itself was vectorized or all | ||||
7349 | // sub-trees not higher that RecursionMaxDepth were analyzed/vectorized. | ||||
7350 | SmallVector<std::pair<Instruction *, unsigned>, 8> Stack(1, {Root, 0}); | ||||
7351 | SmallPtrSet<Value *, 8> VisitedInstrs; | ||||
7352 | bool Res = false; | ||||
7353 | while (!Stack.empty()) { | ||||
7354 | Instruction *Inst; | ||||
7355 | unsigned Level; | ||||
7356 | std::tie(Inst, Level) = Stack.pop_back_val(); | ||||
7357 | auto *BI = dyn_cast<BinaryOperator>(Inst); | ||||
7358 | auto *SI = dyn_cast<SelectInst>(Inst); | ||||
7359 | if (BI || SI) { | ||||
7360 | HorizontalReduction HorRdx; | ||||
7361 | if (HorRdx.matchAssociativeReduction(P, Inst)) { | ||||
7362 | if (HorRdx.tryToReduce(R, TTI)) { | ||||
7363 | Res = true; | ||||
7364 | // Set P to nullptr to avoid re-analysis of phi node in | ||||
7365 | // matchAssociativeReduction function unless this is the root node. | ||||
7366 | P = nullptr; | ||||
7367 | continue; | ||||
7368 | } | ||||
7369 | } | ||||
7370 | if (P && BI) { | ||||
7371 | Inst = dyn_cast<Instruction>(BI->getOperand(0)); | ||||
7372 | if (Inst == P) | ||||
7373 | Inst = dyn_cast<Instruction>(BI->getOperand(1)); | ||||
7374 | if (!Inst) { | ||||
7375 | // Set P to nullptr to avoid re-analysis of phi node in | ||||
7376 | // matchAssociativeReduction function unless this is the root node. | ||||
7377 | P = nullptr; | ||||
7378 | continue; | ||||
7379 | } | ||||
7380 | } | ||||
7381 | } | ||||
7382 | // Set P to nullptr to avoid re-analysis of phi node in | ||||
7383 | // matchAssociativeReduction function unless this is the root node. | ||||
7384 | P = nullptr; | ||||
7385 | if (Vectorize(Inst, R)) { | ||||
7386 | Res = true; | ||||
7387 | continue; | ||||
7388 | } | ||||
7389 | |||||
7390 | // Try to vectorize operands. | ||||
7391 | // Continue analysis for the instruction from the same basic block only to | ||||
7392 | // save compile time. | ||||
7393 | if (++Level < RecursionMaxDepth) | ||||
7394 | for (auto *Op : Inst->operand_values()) | ||||
7395 | if (VisitedInstrs.insert(Op).second) | ||||
7396 | if (auto *I = dyn_cast<Instruction>(Op)) | ||||
7397 | if (!isa<PHINode>(I) && !R.isDeleted(I) && I->getParent() == BB) | ||||
7398 | Stack.emplace_back(I, Level); | ||||
7399 | } | ||||
7400 | return Res; | ||||
7401 | } | ||||
7402 | |||||
7403 | bool SLPVectorizerPass::vectorizeRootInstruction(PHINode *P, Value *V, | ||||
7404 | BasicBlock *BB, BoUpSLP &R, | ||||
7405 | TargetTransformInfo *TTI) { | ||||
7406 | auto *I = dyn_cast_or_null<Instruction>(V); | ||||
7407 | if (!I) | ||||
7408 | return false; | ||||
7409 | |||||
7410 | if (!isa<BinaryOperator>(I)) | ||||
7411 | P = nullptr; | ||||
7412 | // Try to match and vectorize a horizontal reduction. | ||||
7413 | auto &&ExtraVectorization = [this](Instruction *I, BoUpSLP &R) -> bool { | ||||
7414 | return tryToVectorize(I, R); | ||||
7415 | }; | ||||
7416 | return tryToVectorizeHorReductionOrInstOperands(P, I, BB, R, TTI, | ||||
7417 | ExtraVectorization); | ||||
7418 | } | ||||
7419 | |||||
7420 | bool SLPVectorizerPass::vectorizeInsertValueInst(InsertValueInst *IVI, | ||||
7421 | BasicBlock *BB, BoUpSLP &R) { | ||||
7422 | const DataLayout &DL = BB->getModule()->getDataLayout(); | ||||
7423 | if (!R.canMapToVector(IVI->getType(), DL)) | ||||
7424 | return false; | ||||
7425 | |||||
7426 | SmallVector<Value *, 16> BuildVectorOpds; | ||||
7427 | SmallVector<Value *, 16> BuildVectorInsts; | ||||
7428 | if (!findBuildAggregate(IVI, TTI, BuildVectorOpds, BuildVectorInsts)) | ||||
7429 | return false; | ||||
7430 | |||||
7431 | 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); | ||||
7432 | // Aggregate value is unlikely to be processed in vector register, we need to | ||||
7433 | // extract scalars into scalar registers, so NeedExtraction is set true. | ||||
7434 | return tryToVectorizeList(BuildVectorOpds, R, /*AllowReorder=*/false, | ||||
7435 | BuildVectorInsts); | ||||
7436 | } | ||||
7437 | |||||
7438 | bool SLPVectorizerPass::vectorizeInsertElementInst(InsertElementInst *IEI, | ||||
7439 | BasicBlock *BB, BoUpSLP &R) { | ||||
7440 | SmallVector<Value *, 16> BuildVectorInsts; | ||||
7441 | SmallVector<Value *, 16> BuildVectorOpds; | ||||
7442 | if (!findBuildAggregate(IEI, TTI, BuildVectorOpds, BuildVectorInsts) || | ||||
7443 | (llvm::all_of(BuildVectorOpds, | ||||
7444 | [](Value *V) { return isa<ExtractElementInst>(V); }) && | ||||
7445 | isShuffle(BuildVectorOpds))) | ||||
7446 | return false; | ||||
7447 | |||||
7448 | // Vectorize starting with the build vector operands ignoring the BuildVector | ||||
7449 | // instructions for the purpose of scheduling and user extraction. | ||||
7450 | return tryToVectorizeList(BuildVectorOpds, R, /*AllowReorder=*/false, | ||||
7451 | BuildVectorInsts); | ||||
7452 | } | ||||
7453 | |||||
7454 | bool SLPVectorizerPass::vectorizeCmpInst(CmpInst *CI, BasicBlock *BB, | ||||
7455 | BoUpSLP &R) { | ||||
7456 | if (tryToVectorizePair(CI->getOperand(0), CI->getOperand(1), R)) | ||||
7457 | return true; | ||||
7458 | |||||
7459 | bool OpsChanged = false; | ||||
7460 | for (int Idx = 0; Idx < 2; ++Idx) { | ||||
7461 | OpsChanged |= | ||||
7462 | vectorizeRootInstruction(nullptr, CI->getOperand(Idx), BB, R, TTI); | ||||
7463 | } | ||||
7464 | return OpsChanged; | ||||
7465 | } | ||||
7466 | |||||
7467 | bool SLPVectorizerPass::vectorizeSimpleInstructions( | ||||
7468 | SmallVectorImpl<Instruction *> &Instructions, BasicBlock *BB, BoUpSLP &R) { | ||||
7469 | bool OpsChanged = false; | ||||
7470 | for (auto *I : reverse(Instructions)) { | ||||
7471 | if (R.isDeleted(I)) | ||||
7472 | continue; | ||||
7473 | if (auto *LastInsertValue = dyn_cast<InsertValueInst>(I)) | ||||
7474 | OpsChanged |= vectorizeInsertValueInst(LastInsertValue, BB, R); | ||||
7475 | else if (auto *LastInsertElem = dyn_cast<InsertElementInst>(I)) | ||||
7476 | OpsChanged |= vectorizeInsertElementInst(LastInsertElem, BB, R); | ||||
7477 | else if (auto *CI = dyn_cast<CmpInst>(I)) | ||||
7478 | OpsChanged |= vectorizeCmpInst(CI, BB, R); | ||||
7479 | } | ||||
7480 | Instructions.clear(); | ||||
7481 | return OpsChanged; | ||||
7482 | } | ||||
7483 | |||||
7484 | bool SLPVectorizerPass::vectorizeChainsInBlock(BasicBlock *BB, BoUpSLP &R) { | ||||
7485 | bool Changed = false; | ||||
7486 | SmallVector<Value *, 4> Incoming; | ||||
7487 | SmallPtrSet<Value *, 16> VisitedInstrs; | ||||
7488 | unsigned MaxVecRegSize = R.getMaxVecRegSize(); | ||||
7489 | |||||
7490 | bool HaveVectorizedPhiNodes = true; | ||||
7491 | while (HaveVectorizedPhiNodes) { | ||||
7492 | HaveVectorizedPhiNodes = false; | ||||
7493 | |||||
7494 | // Collect the incoming values from the PHIs. | ||||
7495 | Incoming.clear(); | ||||
7496 | for (Instruction &I : *BB) { | ||||
7497 | PHINode *P = dyn_cast<PHINode>(&I); | ||||
7498 | if (!P) | ||||
7499 | break; | ||||
7500 | |||||
7501 | if (!VisitedInstrs.count(P) && !R.isDeleted(P)) | ||||
7502 | Incoming.push_back(P); | ||||
7503 | } | ||||
7504 | |||||
7505 | // Sort by type. | ||||
7506 | llvm::stable_sort(Incoming, PhiTypeSorterFunc); | ||||
7507 | |||||
7508 | // Try to vectorize elements base on their type. | ||||
7509 | for (SmallVector<Value *, 4>::iterator IncIt = Incoming.begin(), | ||||
7510 | E = Incoming.end(); | ||||
7511 | IncIt != E;) { | ||||
7512 | |||||
7513 | // Look for the next elements with the same type. | ||||
7514 | SmallVector<Value *, 4>::iterator SameTypeIt = IncIt; | ||||
7515 | Type *EltTy = (*IncIt)->getType(); | ||||
7516 | |||||
7517 | assert(EltTy->isSized() &&((EltTy->isSized() && "Instructions should all be sized at this point" ) ? static_cast<void> (0) : __assert_fail ("EltTy->isSized() && \"Instructions should all be sized at this point\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 7518, __PRETTY_FUNCTION__)) | ||||
7518 | "Instructions should all be sized at this point")((EltTy->isSized() && "Instructions should all be sized at this point" ) ? static_cast<void> (0) : __assert_fail ("EltTy->isSized() && \"Instructions should all be sized at this point\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 7518, __PRETTY_FUNCTION__)); | ||||
7519 | TypeSize EltTS = DL->getTypeSizeInBits(EltTy); | ||||
7520 | if (EltTS.isScalable()) { | ||||
7521 | // For now, just ignore vectorizing scalable types. | ||||
7522 | ++IncIt; | ||||
7523 | continue; | ||||
7524 | } | ||||
7525 | |||||
7526 | unsigned EltSize = EltTS.getFixedSize(); | ||||
7527 | unsigned MaxNumElts = MaxVecRegSize / EltSize; | ||||
7528 | if (MaxNumElts < 2) { | ||||
7529 | ++IncIt; | ||||
7530 | continue; | ||||
7531 | } | ||||
7532 | |||||
7533 | while (SameTypeIt != E && | ||||
7534 | (*SameTypeIt)->getType() == EltTy && | ||||
7535 | static_cast<unsigned>(SameTypeIt - IncIt) < MaxNumElts) { | ||||
7536 | VisitedInstrs.insert(*SameTypeIt); | ||||
7537 | ++SameTypeIt; | ||||
7538 | } | ||||
7539 | |||||
7540 | // Try to vectorize them. | ||||
7541 | unsigned NumElts = (SameTypeIt - IncIt); | ||||
7542 | LLVM_DEBUG(dbgs() << "SLP: Trying to vectorize starting at PHIs ("do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Trying to vectorize starting at PHIs (" << NumElts << ")\n"; } } while (false) | ||||
7543 | << NumElts << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Trying to vectorize starting at PHIs (" << NumElts << ")\n"; } } while (false); | ||||
7544 | // The order in which the phi nodes appear in the program does not matter. | ||||
7545 | // So allow tryToVectorizeList to reorder them if it is beneficial. This | ||||
7546 | // is done when there are exactly two elements since tryToVectorizeList | ||||
7547 | // asserts that there are only two values when AllowReorder is true. | ||||
7548 | bool AllowReorder = NumElts == 2; | ||||
7549 | if (NumElts > 1 && | ||||
7550 | tryToVectorizeList(makeArrayRef(IncIt, NumElts), R, AllowReorder)) { | ||||
7551 | // Success start over because instructions might have been changed. | ||||
7552 | HaveVectorizedPhiNodes = true; | ||||
7553 | Changed = true; | ||||
7554 | break; | ||||
7555 | } | ||||
7556 | |||||
7557 | // Start over at the next instruction of a different type (or the end). | ||||
7558 | IncIt = SameTypeIt; | ||||
7559 | } | ||||
7560 | } | ||||
7561 | |||||
7562 | VisitedInstrs.clear(); | ||||
7563 | |||||
7564 | SmallVector<Instruction *, 8> PostProcessInstructions; | ||||
7565 | SmallDenseSet<Instruction *, 4> KeyNodes; | ||||
7566 | for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) { | ||||
7567 | // Skip instructions with scalable type. The num of elements is unknown at | ||||
7568 | // compile-time for scalable type. | ||||
7569 | if (isa<ScalableVectorType>(it->getType())) | ||||
7570 | continue; | ||||
7571 | |||||
7572 | // Skip instructions marked for the deletion. | ||||
7573 | if (R.isDeleted(&*it)) | ||||
7574 | continue; | ||||
7575 | // We may go through BB multiple times so skip the one we have checked. | ||||
7576 | if (!VisitedInstrs.insert(&*it).second) { | ||||
7577 | if (it->use_empty() && KeyNodes.count(&*it) > 0 && | ||||
7578 | vectorizeSimpleInstructions(PostProcessInstructions, BB, R)) { | ||||
7579 | // We would like to start over since some instructions are deleted | ||||
7580 | // and the iterator may become invalid value. | ||||
7581 | Changed = true; | ||||
7582 | it = BB->begin(); | ||||
7583 | e = BB->end(); | ||||
7584 | } | ||||
7585 | continue; | ||||
7586 | } | ||||
7587 | |||||
7588 | if (isa<DbgInfoIntrinsic>(it)) | ||||
7589 | continue; | ||||
7590 | |||||
7591 | // Try to vectorize reductions that use PHINodes. | ||||
7592 | if (PHINode *P = dyn_cast<PHINode>(it)) { | ||||
7593 | // Check that the PHI is a reduction PHI. | ||||
7594 | if (P->getNumIncomingValues() != 2) | ||||
7595 | return Changed; | ||||
7596 | |||||
7597 | // Try to match and vectorize a horizontal reduction. | ||||
7598 | if (vectorizeRootInstruction(P, getReductionValue(DT, P, BB, LI), BB, R, | ||||
7599 | TTI)) { | ||||
7600 | Changed = true; | ||||
7601 | it = BB->begin(); | ||||
7602 | e = BB->end(); | ||||
7603 | continue; | ||||
7604 | } | ||||
7605 | continue; | ||||
7606 | } | ||||
7607 | |||||
7608 | // Ran into an instruction without users, like terminator, or function call | ||||
7609 | // with ignored return value, store. Ignore unused instructions (basing on | ||||
7610 | // instruction type, except for CallInst and InvokeInst). | ||||
7611 | if (it->use_empty() && (it->getType()->isVoidTy() || isa<CallInst>(it) || | ||||
7612 | isa<InvokeInst>(it))) { | ||||
7613 | KeyNodes.insert(&*it); | ||||
7614 | bool OpsChanged = false; | ||||
7615 | if (ShouldStartVectorizeHorAtStore || !isa<StoreInst>(it)) { | ||||
7616 | for (auto *V : it->operand_values()) { | ||||
7617 | // Try to match and vectorize a horizontal reduction. | ||||
7618 | OpsChanged |= vectorizeRootInstruction(nullptr, V, BB, R, TTI); | ||||
7619 | } | ||||
7620 | } | ||||
7621 | // Start vectorization of post-process list of instructions from the | ||||
7622 | // top-tree instructions to try to vectorize as many instructions as | ||||
7623 | // possible. | ||||
7624 | OpsChanged |= vectorizeSimpleInstructions(PostProcessInstructions, BB, R); | ||||
7625 | if (OpsChanged) { | ||||
7626 | // We would like to start over since some instructions are deleted | ||||
7627 | // and the iterator may become invalid value. | ||||
7628 | Changed = true; | ||||
7629 | it = BB->begin(); | ||||
7630 | e = BB->end(); | ||||
7631 | continue; | ||||
7632 | } | ||||
7633 | } | ||||
7634 | |||||
7635 | if (isa<InsertElementInst>(it) || isa<CmpInst>(it) || | ||||
7636 | isa<InsertValueInst>(it)) | ||||
7637 | PostProcessInstructions.push_back(&*it); | ||||
7638 | } | ||||
7639 | |||||
7640 | return Changed; | ||||
7641 | } | ||||
7642 | |||||
7643 | bool SLPVectorizerPass::vectorizeGEPIndices(BasicBlock *BB, BoUpSLP &R) { | ||||
7644 | auto Changed = false; | ||||
7645 | for (auto &Entry : GEPs) { | ||||
7646 | // If the getelementptr list has fewer than two elements, there's nothing | ||||
7647 | // to do. | ||||
7648 | if (Entry.second.size() < 2) | ||||
7649 | continue; | ||||
7650 | |||||
7651 | 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 ) | ||||
7652 | << 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 ); | ||||
7653 | |||||
7654 | // Process the GEP list in chunks suitable for the target's supported | ||||
7655 | // vector size. If a vector register can't hold 1 element, we are done. We | ||||
7656 | // are trying to vectorize the index computations, so the maximum number of | ||||
7657 | // elements is based on the size of the index expression, rather than the | ||||
7658 | // size of the GEP itself (the target's pointer size). | ||||
7659 | unsigned MaxVecRegSize = R.getMaxVecRegSize(); | ||||
7660 | unsigned EltSize = R.getVectorElementSize(*Entry.second[0]->idx_begin()); | ||||
7661 | if (MaxVecRegSize < EltSize) | ||||
7662 | continue; | ||||
7663 | |||||
7664 | unsigned MaxElts = MaxVecRegSize / EltSize; | ||||
7665 | for (unsigned BI = 0, BE = Entry.second.size(); BI < BE; BI += MaxElts) { | ||||
7666 | auto Len = std::min<unsigned>(BE - BI, MaxElts); | ||||
7667 | ArrayRef<GetElementPtrInst *> GEPList(&Entry.second[BI], Len); | ||||
7668 | |||||
7669 | // Initialize a set a candidate getelementptrs. Note that we use a | ||||
7670 | // SetVector here to preserve program order. If the index computations | ||||
7671 | // are vectorizable and begin with loads, we want to minimize the chance | ||||
7672 | // of having to reorder them later. | ||||
7673 | SetVector<Value *> Candidates(GEPList.begin(), GEPList.end()); | ||||
7674 | |||||
7675 | // Some of the candidates may have already been vectorized after we | ||||
7676 | // initially collected them. If so, they are marked as deleted, so remove | ||||
7677 | // them from the set of candidates. | ||||
7678 | Candidates.remove_if( | ||||
7679 | [&R](Value *I) { return R.isDeleted(cast<Instruction>(I)); }); | ||||
7680 | |||||
7681 | // Remove from the set of candidates all pairs of getelementptrs with | ||||
7682 | // constant differences. Such getelementptrs are likely not good | ||||
7683 | // candidates for vectorization in a bottom-up phase since one can be | ||||
7684 | // computed from the other. We also ensure all candidate getelementptr | ||||
7685 | // indices are unique. | ||||
7686 | for (int I = 0, E = GEPList.size(); I < E && Candidates.size() > 1; ++I) { | ||||
7687 | auto *GEPI = GEPList[I]; | ||||
7688 | if (!Candidates.count(GEPI)) | ||||
7689 | continue; | ||||
7690 | auto *SCEVI = SE->getSCEV(GEPList[I]); | ||||
7691 | for (int J = I + 1; J < E && Candidates.size() > 1; ++J) { | ||||
7692 | auto *GEPJ = GEPList[J]; | ||||
7693 | auto *SCEVJ = SE->getSCEV(GEPList[J]); | ||||
7694 | if (isa<SCEVConstant>(SE->getMinusSCEV(SCEVI, SCEVJ))) { | ||||
7695 | Candidates.remove(GEPI); | ||||
7696 | Candidates.remove(GEPJ); | ||||
7697 | } else if (GEPI->idx_begin()->get() == GEPJ->idx_begin()->get()) { | ||||
7698 | Candidates.remove(GEPJ); | ||||
7699 | } | ||||
7700 | } | ||||
7701 | } | ||||
7702 | |||||
7703 | // We break out of the above computation as soon as we know there are | ||||
7704 | // fewer than two candidates remaining. | ||||
7705 | if (Candidates.size() < 2) | ||||
7706 | continue; | ||||
7707 | |||||
7708 | // Add the single, non-constant index of each candidate to the bundle. We | ||||
7709 | // ensured the indices met these constraints when we originally collected | ||||
7710 | // the getelementptrs. | ||||
7711 | SmallVector<Value *, 16> Bundle(Candidates.size()); | ||||
7712 | auto BundleIndex = 0u; | ||||
7713 | for (auto *V : Candidates) { | ||||
7714 | auto *GEP = cast<GetElementPtrInst>(V); | ||||
7715 | auto *GEPIdx = GEP->idx_begin()->get(); | ||||
7716 | assert(GEP->getNumIndices() == 1 || !isa<Constant>(GEPIdx))((GEP->getNumIndices() == 1 || !isa<Constant>(GEPIdx )) ? static_cast<void> (0) : __assert_fail ("GEP->getNumIndices() == 1 || !isa<Constant>(GEPIdx)" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 7716, __PRETTY_FUNCTION__)); | ||||
7717 | Bundle[BundleIndex++] = GEPIdx; | ||||
7718 | } | ||||
7719 | |||||
7720 | // Try and vectorize the indices. We are currently only interested in | ||||
7721 | // gather-like cases of the form: | ||||
7722 | // | ||||
7723 | // ... = g[a[0] - b[0]] + g[a[1] - b[1]] + ... | ||||
7724 | // | ||||
7725 | // where the loads of "a", the loads of "b", and the subtractions can be | ||||
7726 | // performed in parallel. It's likely that detecting this pattern in a | ||||
7727 | // bottom-up phase will be simpler and less costly than building a | ||||
7728 | // full-blown top-down phase beginning at the consecutive loads. | ||||
7729 | Changed |= tryToVectorizeList(Bundle, R); | ||||
7730 | } | ||||
7731 | } | ||||
7732 | return Changed; | ||||
7733 | } | ||||
7734 | |||||
7735 | bool SLPVectorizerPass::vectorizeStoreChains(BoUpSLP &R) { | ||||
7736 | bool Changed = false; | ||||
7737 | // Attempt to sort and vectorize each of the store-groups. | ||||
7738 | for (StoreListMap::iterator it = Stores.begin(), e = Stores.end(); it != e; | ||||
7739 | ++it) { | ||||
7740 | if (it->second.size() < 2) | ||||
7741 | continue; | ||||
7742 | |||||
7743 | 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 " << it->second.size() << ".\n"; } } while (false ) | ||||
7744 | << it->second.size() << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing a store chain of length " << it->second.size() << ".\n"; } } while (false ); | ||||
7745 | |||||
7746 | Changed |= vectorizeStores(it->second, R); | ||||
7747 | } | ||||
7748 | return Changed; | ||||
7749 | } | ||||
7750 | |||||
7751 | char SLPVectorizer::ID = 0; | ||||
7752 | |||||
7753 | static const char lv_name[] = "SLP Vectorizer"; | ||||
7754 | |||||
7755 | INITIALIZE_PASS_BEGIN(SLPVectorizer, SV_NAME, lv_name, false, false)static void *initializeSLPVectorizerPassOnce(PassRegistry & Registry) { | ||||
7756 | INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)initializeAAResultsWrapperPassPass(Registry); | ||||
7757 | INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)initializeTargetTransformInfoWrapperPassPass(Registry); | ||||
7758 | INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)initializeAssumptionCacheTrackerPass(Registry); | ||||
7759 | INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)initializeScalarEvolutionWrapperPassPass(Registry); | ||||
7760 | INITIALIZE_PASS_DEPENDENCY(LoopSimplify)initializeLoopSimplifyPass(Registry); | ||||
7761 | INITIALIZE_PASS_DEPENDENCY(DemandedBitsWrapperPass)initializeDemandedBitsWrapperPassPass(Registry); | ||||
7762 | INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass)initializeOptimizationRemarkEmitterWrapperPassPass(Registry); | ||||
7763 | INITIALIZE_PASS_DEPENDENCY(InjectTLIMappingsLegacy)initializeInjectTLIMappingsLegacyPass(Registry); | ||||
7764 | 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)); } | ||||
7765 | |||||
7766 | Pass *llvm::createSLPVectorizerPass() { return new SLPVectorizer(); } |
1 | //===- GenericDomTree.h - Generic dominator trees for graphs ----*- C++ -*-===// |
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 | /// \file |
9 | /// |
10 | /// This file defines a set of templates that efficiently compute a dominator |
11 | /// tree over a generic graph. This is used typically in LLVM for fast |
12 | /// dominance queries on the CFG, but is fully generic w.r.t. the underlying |
13 | /// graph types. |
14 | /// |
15 | /// Unlike ADT/* graph algorithms, generic dominator tree has more requirements |
16 | /// on the graph's NodeRef. The NodeRef should be a pointer and, |
17 | /// NodeRef->getParent() must return the parent node that is also a pointer. |
18 | /// |
19 | /// FIXME: Maybe GenericDomTree needs a TreeTraits, instead of GraphTraits. |
20 | /// |
21 | //===----------------------------------------------------------------------===// |
22 | |
23 | #ifndef LLVM_SUPPORT_GENERICDOMTREE_H |
24 | #define LLVM_SUPPORT_GENERICDOMTREE_H |
25 | |
26 | #include "llvm/ADT/DenseMap.h" |
27 | #include "llvm/ADT/GraphTraits.h" |
28 | #include "llvm/ADT/STLExtras.h" |
29 | #include "llvm/ADT/SmallPtrSet.h" |
30 | #include "llvm/ADT/SmallVector.h" |
31 | #include "llvm/Support/CFGDiff.h" |
32 | #include "llvm/Support/CFGUpdate.h" |
33 | #include "llvm/Support/raw_ostream.h" |
34 | #include <algorithm> |
35 | #include <cassert> |
36 | #include <cstddef> |
37 | #include <iterator> |
38 | #include <memory> |
39 | #include <type_traits> |
40 | #include <utility> |
41 | |
42 | namespace llvm { |
43 | |
44 | template <typename NodeT, bool IsPostDom> |
45 | class DominatorTreeBase; |
46 | |
47 | namespace DomTreeBuilder { |
48 | template <typename DomTreeT> |
49 | struct SemiNCAInfo; |
50 | } // namespace DomTreeBuilder |
51 | |
52 | /// Base class for the actual dominator tree node. |
53 | template <class NodeT> class DomTreeNodeBase { |
54 | friend class PostDominatorTree; |
55 | friend class DominatorTreeBase<NodeT, false>; |
56 | friend class DominatorTreeBase<NodeT, true>; |
57 | friend struct DomTreeBuilder::SemiNCAInfo<DominatorTreeBase<NodeT, false>>; |
58 | friend struct DomTreeBuilder::SemiNCAInfo<DominatorTreeBase<NodeT, true>>; |
59 | |
60 | NodeT *TheBB; |
61 | DomTreeNodeBase *IDom; |
62 | unsigned Level; |
63 | SmallVector<DomTreeNodeBase *, 4> Children; |
64 | mutable unsigned DFSNumIn = ~0; |
65 | mutable unsigned DFSNumOut = ~0; |
66 | |
67 | public: |
68 | DomTreeNodeBase(NodeT *BB, DomTreeNodeBase *iDom) |
69 | : TheBB(BB), IDom(iDom), Level(IDom ? IDom->Level + 1 : 0) {} |
70 | |
71 | using iterator = typename SmallVector<DomTreeNodeBase *, 4>::iterator; |
72 | using const_iterator = |
73 | typename SmallVector<DomTreeNodeBase *, 4>::const_iterator; |
74 | |
75 | iterator begin() { return Children.begin(); } |
76 | iterator end() { return Children.end(); } |
77 | const_iterator begin() const { return Children.begin(); } |
78 | const_iterator end() const { return Children.end(); } |
79 | |
80 | DomTreeNodeBase *const &back() const { return Children.back(); } |
81 | DomTreeNodeBase *&back() { return Children.back(); } |
82 | |
83 | iterator_range<iterator> children() { return make_range(begin(), end()); } |
84 | iterator_range<const_iterator> children() const { |
85 | return make_range(begin(), end()); |
86 | } |
87 | |
88 | NodeT *getBlock() const { return TheBB; } |
89 | DomTreeNodeBase *getIDom() const { return IDom; } |
90 | unsigned getLevel() const { return Level; } |
91 | |
92 | std::unique_ptr<DomTreeNodeBase> addChild( |
93 | std::unique_ptr<DomTreeNodeBase> C) { |
94 | Children.push_back(C.get()); |
95 | return C; |
96 | } |
97 | |
98 | bool isLeaf() const { return Children.empty(); } |
99 | size_t getNumChildren() const { return Children.size(); } |
100 | |
101 | void clearAllChildren() { Children.clear(); } |
102 | |
103 | bool compare(const DomTreeNodeBase *Other) const { |
104 | if (getNumChildren() != Other->getNumChildren()) |
105 | return true; |
106 | |
107 | if (Level != Other->Level) return true; |
108 | |
109 | SmallPtrSet<const NodeT *, 4> OtherChildren; |
110 | for (const DomTreeNodeBase *I : *Other) { |
111 | const NodeT *Nd = I->getBlock(); |
112 | OtherChildren.insert(Nd); |
113 | } |
114 | |
115 | for (const DomTreeNodeBase *I : *this) { |
116 | const NodeT *N = I->getBlock(); |
117 | if (OtherChildren.count(N) == 0) |
118 | return true; |
119 | } |
120 | return false; |
121 | } |
122 | |
123 | void setIDom(DomTreeNodeBase *NewIDom) { |
124 | assert(IDom && "No immediate dominator?")((IDom && "No immediate dominator?") ? static_cast< void> (0) : __assert_fail ("IDom && \"No immediate dominator?\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/include/llvm/Support/GenericDomTree.h" , 124, __PRETTY_FUNCTION__)); |
125 | if (IDom == NewIDom) return; |
126 | |
127 | auto I = find(IDom->Children, this); |
128 | assert(I != IDom->Children.end() &&((I != IDom->Children.end() && "Not in immediate dominator children set!" ) ? static_cast<void> (0) : __assert_fail ("I != IDom->Children.end() && \"Not in immediate dominator children set!\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/include/llvm/Support/GenericDomTree.h" , 129, __PRETTY_FUNCTION__)) |
129 | "Not in immediate dominator children set!")((I != IDom->Children.end() && "Not in immediate dominator children set!" ) ? static_cast<void> (0) : __assert_fail ("I != IDom->Children.end() && \"Not in immediate dominator children set!\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/include/llvm/Support/GenericDomTree.h" , 129, __PRETTY_FUNCTION__)); |
130 | // I am no longer your child... |
131 | IDom->Children.erase(I); |
132 | |
133 | // Switch to new dominator |
134 | IDom = NewIDom; |
135 | IDom->Children.push_back(this); |
136 | |
137 | UpdateLevel(); |
138 | } |
139 | |
140 | /// getDFSNumIn/getDFSNumOut - These return the DFS visitation order for nodes |
141 | /// in the dominator tree. They are only guaranteed valid if |
142 | /// updateDFSNumbers() has been called. |
143 | unsigned getDFSNumIn() const { return DFSNumIn; } |
144 | unsigned getDFSNumOut() const { return DFSNumOut; } |
145 | |
146 | private: |
147 | // Return true if this node is dominated by other. Use this only if DFS info |
148 | // is valid. |
149 | bool DominatedBy(const DomTreeNodeBase *other) const { |
150 | return this->DFSNumIn >= other->DFSNumIn && |
151 | this->DFSNumOut <= other->DFSNumOut; |
152 | } |
153 | |
154 | void UpdateLevel() { |
155 | assert(IDom)((IDom) ? static_cast<void> (0) : __assert_fail ("IDom" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/include/llvm/Support/GenericDomTree.h" , 155, __PRETTY_FUNCTION__)); |
156 | if (Level == IDom->Level + 1) return; |
157 | |
158 | SmallVector<DomTreeNodeBase *, 64> WorkStack = {this}; |
159 | |
160 | while (!WorkStack.empty()) { |
161 | DomTreeNodeBase *Current = WorkStack.pop_back_val(); |
162 | Current->Level = Current->IDom->Level + 1; |
163 | |
164 | for (DomTreeNodeBase *C : *Current) { |
165 | assert(C->IDom)((C->IDom) ? static_cast<void> (0) : __assert_fail ( "C->IDom", "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/include/llvm/Support/GenericDomTree.h" , 165, __PRETTY_FUNCTION__)); |
166 | if (C->Level != C->IDom->Level + 1) WorkStack.push_back(C); |
167 | } |
168 | } |
169 | } |
170 | }; |
171 | |
172 | template <class NodeT> |
173 | raw_ostream &operator<<(raw_ostream &O, const DomTreeNodeBase<NodeT> *Node) { |
174 | if (Node->getBlock()) |
175 | Node->getBlock()->printAsOperand(O, false); |
176 | else |
177 | O << " <<exit node>>"; |
178 | |
179 | O << " {" << Node->getDFSNumIn() << "," << Node->getDFSNumOut() << "} [" |
180 | << Node->getLevel() << "]\n"; |
181 | |
182 | return O; |
183 | } |
184 | |
185 | template <class NodeT> |
186 | void PrintDomTree(const DomTreeNodeBase<NodeT> *N, raw_ostream &O, |
187 | unsigned Lev) { |
188 | O.indent(2 * Lev) << "[" << Lev << "] " << N; |
189 | for (typename DomTreeNodeBase<NodeT>::const_iterator I = N->begin(), |
190 | E = N->end(); |
191 | I != E; ++I) |
192 | PrintDomTree<NodeT>(*I, O, Lev + 1); |
193 | } |
194 | |
195 | namespace DomTreeBuilder { |
196 | // The routines below are provided in a separate header but referenced here. |
197 | template <typename DomTreeT> |
198 | void Calculate(DomTreeT &DT); |
199 | |
200 | template <typename DomTreeT> |
201 | void CalculateWithUpdates(DomTreeT &DT, |
202 | ArrayRef<typename DomTreeT::UpdateType> Updates); |
203 | |
204 | template <typename DomTreeT> |
205 | void InsertEdge(DomTreeT &DT, typename DomTreeT::NodePtr From, |
206 | typename DomTreeT::NodePtr To); |
207 | |
208 | template <typename DomTreeT> |
209 | void DeleteEdge(DomTreeT &DT, typename DomTreeT::NodePtr From, |
210 | typename DomTreeT::NodePtr To); |
211 | |
212 | template <typename DomTreeT> |
213 | void ApplyUpdates(DomTreeT &DT, |
214 | GraphDiff<typename DomTreeT::NodePtr, |
215 | DomTreeT::IsPostDominator> &PreViewCFG, |
216 | GraphDiff<typename DomTreeT::NodePtr, |
217 | DomTreeT::IsPostDominator> *PostViewCFG); |
218 | |
219 | template <typename DomTreeT> |
220 | bool Verify(const DomTreeT &DT, typename DomTreeT::VerificationLevel VL); |
221 | } // namespace DomTreeBuilder |
222 | |
223 | /// Core dominator tree base class. |
224 | /// |
225 | /// This class is a generic template over graph nodes. It is instantiated for |
226 | /// various graphs in the LLVM IR or in the code generator. |
227 | template <typename NodeT, bool IsPostDom> |
228 | class DominatorTreeBase { |
229 | public: |
230 | static_assert(std::is_pointer<typename GraphTraits<NodeT *>::NodeRef>::value, |
231 | "Currently DominatorTreeBase supports only pointer nodes"); |
232 | using NodeType = NodeT; |
233 | using NodePtr = NodeT *; |
234 | using ParentPtr = decltype(std::declval<NodeT *>()->getParent()); |
235 | static_assert(std::is_pointer<ParentPtr>::value, |
236 | "Currently NodeT's parent must be a pointer type"); |
237 | using ParentType = std::remove_pointer_t<ParentPtr>; |
238 | static constexpr bool IsPostDominator = IsPostDom; |
239 | |
240 | using UpdateType = cfg::Update<NodePtr>; |
241 | using UpdateKind = cfg::UpdateKind; |
242 | static constexpr UpdateKind Insert = UpdateKind::Insert; |
243 | static constexpr UpdateKind Delete = UpdateKind::Delete; |
244 | |
245 | enum class VerificationLevel { Fast, Basic, Full }; |
246 | |
247 | protected: |
248 | // Dominators always have a single root, postdominators can have more. |
249 | SmallVector<NodeT *, IsPostDom ? 4 : 1> Roots; |
250 | |
251 | using DomTreeNodeMapType = |
252 | DenseMap<NodeT *, std::unique_ptr<DomTreeNodeBase<NodeT>>>; |
253 | DomTreeNodeMapType DomTreeNodes; |
254 | DomTreeNodeBase<NodeT> *RootNode = nullptr; |
255 | ParentPtr Parent = nullptr; |
256 | |
257 | mutable bool DFSInfoValid = false; |
258 | mutable unsigned int SlowQueries = 0; |
259 | |
260 | friend struct DomTreeBuilder::SemiNCAInfo<DominatorTreeBase>; |
261 | |
262 | public: |
263 | DominatorTreeBase() {} |
264 | |
265 | DominatorTreeBase(DominatorTreeBase &&Arg) |
266 | : Roots(std::move(Arg.Roots)), |
267 | DomTreeNodes(std::move(Arg.DomTreeNodes)), |
268 | RootNode(Arg.RootNode), |
269 | Parent(Arg.Parent), |
270 | DFSInfoValid(Arg.DFSInfoValid), |
271 | SlowQueries(Arg.SlowQueries) { |
272 | Arg.wipe(); |
273 | } |
274 | |
275 | DominatorTreeBase &operator=(DominatorTreeBase &&RHS) { |
276 | Roots = std::move(RHS.Roots); |
277 | DomTreeNodes = std::move(RHS.DomTreeNodes); |
278 | RootNode = RHS.RootNode; |
279 | Parent = RHS.Parent; |
280 | DFSInfoValid = RHS.DFSInfoValid; |
281 | SlowQueries = RHS.SlowQueries; |
282 | RHS.wipe(); |
283 | return *this; |
284 | } |
285 | |
286 | DominatorTreeBase(const DominatorTreeBase &) = delete; |
287 | DominatorTreeBase &operator=(const DominatorTreeBase &) = delete; |
288 | |
289 | /// Iteration over roots. |
290 | /// |
291 | /// This may include multiple blocks if we are computing post dominators. |
292 | /// For forward dominators, this will always be a single block (the entry |
293 | /// block). |
294 | using root_iterator = typename SmallVectorImpl<NodeT *>::iterator; |
295 | using const_root_iterator = typename SmallVectorImpl<NodeT *>::const_iterator; |
296 | |
297 | root_iterator root_begin() { return Roots.begin(); } |
298 | const_root_iterator root_begin() const { return Roots.begin(); } |
299 | root_iterator root_end() { return Roots.end(); } |
300 | const_root_iterator root_end() const { return Roots.end(); } |
301 | |
302 | size_t root_size() const { return Roots.size(); } |
303 | |
304 | iterator_range<root_iterator> roots() { |
305 | return make_range(root_begin(), root_end()); |
306 | } |
307 | iterator_range<const_root_iterator> roots() const { |
308 | return make_range(root_begin(), root_end()); |
309 | } |
310 | |
311 | /// isPostDominator - Returns true if analysis based of postdoms |
312 | /// |
313 | bool isPostDominator() const { return IsPostDominator; } |
314 | |
315 | /// compare - Return false if the other dominator tree base matches this |
316 | /// dominator tree base. Otherwise return true. |
317 | bool compare(const DominatorTreeBase &Other) const { |
318 | if (Parent != Other.Parent) return true; |
319 | |
320 | if (Roots.size() != Other.Roots.size()) |
321 | return true; |
322 | |
323 | if (!std::is_permutation(Roots.begin(), Roots.end(), Other.Roots.begin())) |
324 | return true; |
325 | |
326 | const DomTreeNodeMapType &OtherDomTreeNodes = Other.DomTreeNodes; |
327 | if (DomTreeNodes.size() != OtherDomTreeNodes.size()) |
328 | return true; |
329 | |
330 | for (const auto &DomTreeNode : DomTreeNodes) { |
331 | NodeT *BB = DomTreeNode.first; |
332 | typename DomTreeNodeMapType::const_iterator OI = |
333 | OtherDomTreeNodes.find(BB); |
334 | if (OI == OtherDomTreeNodes.end()) |
335 | return true; |
336 | |
337 | DomTreeNodeBase<NodeT> &MyNd = *DomTreeNode.second; |
338 | DomTreeNodeBase<NodeT> &OtherNd = *OI->second; |
339 | |
340 | if (MyNd.compare(&OtherNd)) |
341 | return true; |
342 | } |
343 | |
344 | return false; |
345 | } |
346 | |
347 | /// getNode - return the (Post)DominatorTree node for the specified basic |
348 | /// block. This is the same as using operator[] on this class. The result |
349 | /// may (but is not required to) be null for a forward (backwards) |
350 | /// statically unreachable block. |
351 | DomTreeNodeBase<NodeT> *getNode(const NodeT *BB) const { |
352 | auto I = DomTreeNodes.find(BB); |
353 | if (I != DomTreeNodes.end()) |
354 | return I->second.get(); |
355 | return nullptr; |
356 | } |
357 | |
358 | /// See getNode. |
359 | DomTreeNodeBase<NodeT> *operator[](const NodeT *BB) const { |
360 | return getNode(BB); |
361 | } |
362 | |
363 | /// getRootNode - This returns the entry node for the CFG of the function. If |
364 | /// this tree represents the post-dominance relations for a function, however, |
365 | /// this root may be a node with the block == NULL. This is the case when |
366 | /// there are multiple exit nodes from a particular function. Consumers of |
367 | /// post-dominance information must be capable of dealing with this |
368 | /// possibility. |
369 | /// |
370 | DomTreeNodeBase<NodeT> *getRootNode() { return RootNode; } |
371 | const DomTreeNodeBase<NodeT> *getRootNode() const { return RootNode; } |
372 | |
373 | /// Get all nodes dominated by R, including R itself. |
374 | void getDescendants(NodeT *R, SmallVectorImpl<NodeT *> &Result) const { |
375 | Result.clear(); |
376 | const DomTreeNodeBase<NodeT> *RN = getNode(R); |
377 | if (!RN) |
378 | return; // If R is unreachable, it will not be present in the DOM tree. |
379 | SmallVector<const DomTreeNodeBase<NodeT> *, 8> WL; |
380 | WL.push_back(RN); |
381 | |
382 | while (!WL.empty()) { |
383 | const DomTreeNodeBase<NodeT> *N = WL.pop_back_val(); |
384 | Result.push_back(N->getBlock()); |
385 | WL.append(N->begin(), N->end()); |
386 | } |
387 | } |
388 | |
389 | /// properlyDominates - Returns true iff A dominates B and A != B. |
390 | /// Note that this is not a constant time operation! |
391 | /// |
392 | bool properlyDominates(const DomTreeNodeBase<NodeT> *A, |
393 | const DomTreeNodeBase<NodeT> *B) const { |
394 | if (!A || !B) |
395 | return false; |
396 | if (A == B) |
397 | return false; |
398 | return dominates(A, B); |
399 | } |
400 | |
401 | bool properlyDominates(const NodeT *A, const NodeT *B) const; |
402 | |
403 | /// isReachableFromEntry - Return true if A is dominated by the entry |
404 | /// block of the function containing it. |
405 | bool isReachableFromEntry(const NodeT *A) const { |
406 | assert(!this->isPostDominator() &&((!this->isPostDominator() && "This is not implemented for post dominators" ) ? static_cast<void> (0) : __assert_fail ("!this->isPostDominator() && \"This is not implemented for post dominators\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/include/llvm/Support/GenericDomTree.h" , 407, __PRETTY_FUNCTION__)) |
407 | "This is not implemented for post dominators")((!this->isPostDominator() && "This is not implemented for post dominators" ) ? static_cast<void> (0) : __assert_fail ("!this->isPostDominator() && \"This is not implemented for post dominators\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/include/llvm/Support/GenericDomTree.h" , 407, __PRETTY_FUNCTION__)); |
408 | return isReachableFromEntry(getNode(const_cast<NodeT *>(A))); |
409 | } |
410 | |
411 | bool isReachableFromEntry(const DomTreeNodeBase<NodeT> *A) const { return A; } |
412 | |
413 | /// dominates - Returns true iff A dominates B. Note that this is not a |
414 | /// constant time operation! |
415 | /// |
416 | bool dominates(const DomTreeNodeBase<NodeT> *A, |
417 | const DomTreeNodeBase<NodeT> *B) const { |
418 | // A node trivially dominates itself. |
419 | if (B == A) |
420 | return true; |
421 | |
422 | // An unreachable node is dominated by anything. |
423 | if (!isReachableFromEntry(B)) |
424 | return true; |
425 | |
426 | // And dominates nothing. |
427 | if (!isReachableFromEntry(A)) |
428 | return false; |
429 | |
430 | if (B->getIDom() == A) return true; |
431 | |
432 | if (A->getIDom() == B) return false; |
433 | |
434 | // A can only dominate B if it is higher in the tree. |
435 | if (A->getLevel() >= B->getLevel()) return false; |
436 | |
437 | // Compare the result of the tree walk and the dfs numbers, if expensive |
438 | // checks are enabled. |
439 | #ifdef EXPENSIVE_CHECKS |
440 | assert((!DFSInfoValid ||(((!DFSInfoValid || (dominatedBySlowTreeWalk(A, B) == B->DominatedBy (A))) && "Tree walk disagrees with dfs numbers!") ? static_cast <void> (0) : __assert_fail ("(!DFSInfoValid || (dominatedBySlowTreeWalk(A, B) == B->DominatedBy(A))) && \"Tree walk disagrees with dfs numbers!\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/include/llvm/Support/GenericDomTree.h" , 442, __PRETTY_FUNCTION__)) |
441 | (dominatedBySlowTreeWalk(A, B) == B->DominatedBy(A))) &&(((!DFSInfoValid || (dominatedBySlowTreeWalk(A, B) == B->DominatedBy (A))) && "Tree walk disagrees with dfs numbers!") ? static_cast <void> (0) : __assert_fail ("(!DFSInfoValid || (dominatedBySlowTreeWalk(A, B) == B->DominatedBy(A))) && \"Tree walk disagrees with dfs numbers!\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/include/llvm/Support/GenericDomTree.h" , 442, __PRETTY_FUNCTION__)) |
442 | "Tree walk disagrees with dfs numbers!")(((!DFSInfoValid || (dominatedBySlowTreeWalk(A, B) == B->DominatedBy (A))) && "Tree walk disagrees with dfs numbers!") ? static_cast <void> (0) : __assert_fail ("(!DFSInfoValid || (dominatedBySlowTreeWalk(A, B) == B->DominatedBy(A))) && \"Tree walk disagrees with dfs numbers!\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/include/llvm/Support/GenericDomTree.h" , 442, __PRETTY_FUNCTION__)); |
443 | #endif |
444 | |
445 | if (DFSInfoValid) |
446 | return B->DominatedBy(A); |
447 | |
448 | // If we end up with too many slow queries, just update the |
449 | // DFS numbers on the theory that we are going to keep querying. |
450 | SlowQueries++; |
451 | if (SlowQueries > 32) { |
452 | updateDFSNumbers(); |
453 | return B->DominatedBy(A); |
454 | } |
455 | |
456 | return dominatedBySlowTreeWalk(A, B); |
457 | } |
458 | |
459 | bool dominates(const NodeT *A, const NodeT *B) const; |
460 | |
461 | NodeT *getRoot() const { |
462 | assert(this->Roots.size() == 1 && "Should always have entry node!")((this->Roots.size() == 1 && "Should always have entry node!" ) ? static_cast<void> (0) : __assert_fail ("this->Roots.size() == 1 && \"Should always have entry node!\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/include/llvm/Support/GenericDomTree.h" , 462, __PRETTY_FUNCTION__)); |
463 | return this->Roots[0]; |
464 | } |
465 | |
466 | /// findNearestCommonDominator - Find nearest common dominator basic block |
467 | /// for basic block A and B. If there is no such block then return nullptr. |
468 | NodeT *findNearestCommonDominator(NodeT *A, NodeT *B) const { |
469 | assert(A && B && "Pointers are not valid")((A && B && "Pointers are not valid") ? static_cast <void> (0) : __assert_fail ("A && B && \"Pointers are not valid\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/include/llvm/Support/GenericDomTree.h" , 469, __PRETTY_FUNCTION__)); |
470 | assert(A->getParent() == B->getParent() &&((A->getParent() == B->getParent() && "Two blocks are not in same function" ) ? static_cast<void> (0) : __assert_fail ("A->getParent() == B->getParent() && \"Two blocks are not in same function\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/include/llvm/Support/GenericDomTree.h" , 471, __PRETTY_FUNCTION__)) |
471 | "Two blocks are not in same function")((A->getParent() == B->getParent() && "Two blocks are not in same function" ) ? static_cast<void> (0) : __assert_fail ("A->getParent() == B->getParent() && \"Two blocks are not in same function\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/include/llvm/Support/GenericDomTree.h" , 471, __PRETTY_FUNCTION__)); |
472 | |
473 | // If either A or B is a entry block then it is nearest common dominator |
474 | // (for forward-dominators). |
475 | if (!isPostDominator()) { |
476 | NodeT &Entry = A->getParent()->front(); |
477 | if (A == &Entry || B == &Entry) |
478 | return &Entry; |
479 | } |
480 | |
481 | DomTreeNodeBase<NodeT> *NodeA = getNode(A); |
482 | DomTreeNodeBase<NodeT> *NodeB = getNode(B); |
483 | |
484 | if (!NodeA || !NodeB) return nullptr; |
485 | |
486 | // Use level information to go up the tree until the levels match. Then |
487 | // continue going up til we arrive at the same node. |
488 | while (NodeA && NodeA != NodeB) { |
489 | if (NodeA->getLevel() < NodeB->getLevel()) std::swap(NodeA, NodeB); |
490 | |
491 | NodeA = NodeA->IDom; |
492 | } |
493 | |
494 | return NodeA ? NodeA->getBlock() : nullptr; |
495 | } |
496 | |
497 | const NodeT *findNearestCommonDominator(const NodeT *A, |
498 | const NodeT *B) const { |
499 | // Cast away the const qualifiers here. This is ok since |
500 | // const is re-introduced on the return type. |
501 | return findNearestCommonDominator(const_cast<NodeT *>(A), |
502 | const_cast<NodeT *>(B)); |
503 | } |
504 | |
505 | bool isVirtualRoot(const DomTreeNodeBase<NodeT> *A) const { |
506 | return isPostDominator() && !A->getBlock(); |
507 | } |
508 | |
509 | //===--------------------------------------------------------------------===// |
510 | // API to update (Post)DominatorTree information based on modifications to |
511 | // the CFG... |
512 | |
513 | /// Inform the dominator tree about a sequence of CFG edge insertions and |
514 | /// deletions and perform a batch update on the tree. |
515 | /// |
516 | /// This function should be used when there were multiple CFG updates after |
517 | /// the last dominator tree update. It takes care of performing the updates |
518 | /// in sync with the CFG and optimizes away the redundant operations that |
519 | /// cancel each other. |
520 | /// The functions expects the sequence of updates to be balanced. Eg.: |
521 | /// - {{Insert, A, B}, {Delete, A, B}, {Insert, A, B}} is fine, because |
522 | /// logically it results in a single insertions. |
523 | /// - {{Insert, A, B}, {Insert, A, B}} is invalid, because it doesn't make |
524 | /// sense to insert the same edge twice. |
525 | /// |
526 | /// What's more, the functions assumes that it's safe to ask every node in the |
527 | /// CFG about its children and inverse children. This implies that deletions |
528 | /// of CFG edges must not delete the CFG nodes before calling this function. |
529 | /// |
530 | /// The applyUpdates function can reorder the updates and remove redundant |
531 | /// ones internally. The batch updater is also able to detect sequences of |
532 | /// zero and exactly one update -- it's optimized to do less work in these |
533 | /// cases. |
534 | /// |
535 | /// Note that for postdominators it automatically takes care of applying |
536 | /// updates on reverse edges internally (so there's no need to swap the |
537 | /// From and To pointers when constructing DominatorTree::UpdateType). |
538 | /// The type of updates is the same for DomTreeBase<T> and PostDomTreeBase<T> |
539 | /// with the same template parameter T. |
540 | /// |
541 | /// \param Updates An unordered sequence of updates to perform. The current |
542 | /// CFG and the reverse of these updates provides the pre-view of the CFG. |
543 | /// |
544 | void applyUpdates(ArrayRef<UpdateType> Updates) { |
545 | GraphDiff<NodePtr, IsPostDominator> PreViewCFG( |
546 | Updates, /*ReverseApplyUpdates=*/true); |
547 | DomTreeBuilder::ApplyUpdates(*this, PreViewCFG, nullptr); |
548 | } |
549 | |
550 | /// \param Updates An unordered sequence of updates to perform. The current |
551 | /// CFG and the reverse of these updates provides the pre-view of the CFG. |
552 | /// \param PostViewUpdates An unordered sequence of update to perform in order |
553 | /// to obtain a post-view of the CFG. The DT will be updates assuming the |
554 | /// obtained PostViewCFG is the desired end state. |
555 | void applyUpdates(ArrayRef<UpdateType> Updates, |
556 | ArrayRef<UpdateType> PostViewUpdates) { |
557 | // GraphDiff<NodePtr, IsPostDom> *PostViewCFG = nullptr) { |
558 | if (Updates.empty()) { |
559 | GraphDiff<NodePtr, IsPostDom> PostViewCFG(PostViewUpdates); |
560 | DomTreeBuilder::ApplyUpdates(*this, PostViewCFG, &PostViewCFG); |
561 | } else { |
562 | // TODO: |
563 | // PreViewCFG needs to merge Updates and PostViewCFG. The updates in |
564 | // Updates need to be reversed, and match the direction in PostViewCFG. |
565 | // Normally, a PostViewCFG is created without reversing updates, so one |
566 | // of the internal vectors needs reversing in order to do the |
567 | // legalization of the merged vector of updates. |
568 | llvm_unreachable("Currently unsupported to update given a set of "::llvm::llvm_unreachable_internal("Currently unsupported to update given a set of " "updates towards a PostView", "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/include/llvm/Support/GenericDomTree.h" , 569) |
569 | "updates towards a PostView")::llvm::llvm_unreachable_internal("Currently unsupported to update given a set of " "updates towards a PostView", "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/include/llvm/Support/GenericDomTree.h" , 569); |
570 | } |
571 | } |
572 | |
573 | /// Inform the dominator tree about a CFG edge insertion and update the tree. |
574 | /// |
575 | /// This function has to be called just before or just after making the update |
576 | /// on the actual CFG. There cannot be any other updates that the dominator |
577 | /// tree doesn't know about. |
578 | /// |
579 | /// Note that for postdominators it automatically takes care of inserting |
580 | /// a reverse edge internally (so there's no need to swap the parameters). |
581 | /// |
582 | void insertEdge(NodeT *From, NodeT *To) { |
583 | assert(From)((From) ? static_cast<void> (0) : __assert_fail ("From" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/include/llvm/Support/GenericDomTree.h" , 583, __PRETTY_FUNCTION__)); |
584 | assert(To)((To) ? static_cast<void> (0) : __assert_fail ("To", "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/include/llvm/Support/GenericDomTree.h" , 584, __PRETTY_FUNCTION__)); |
585 | assert(From->getParent() == Parent)((From->getParent() == Parent) ? static_cast<void> ( 0) : __assert_fail ("From->getParent() == Parent", "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/include/llvm/Support/GenericDomTree.h" , 585, __PRETTY_FUNCTION__)); |
586 | assert(To->getParent() == Parent)((To->getParent() == Parent) ? static_cast<void> (0) : __assert_fail ("To->getParent() == Parent", "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/include/llvm/Support/GenericDomTree.h" , 586, __PRETTY_FUNCTION__)); |
587 | DomTreeBuilder::InsertEdge(*this, From, To); |
588 | } |
589 | |
590 | /// Inform the dominator tree about a CFG edge deletion and update the tree. |
591 | /// |
592 | /// This function has to be called just after making the update on the actual |
593 | /// CFG. An internal functions checks if the edge doesn't exist in the CFG in |
594 | /// DEBUG mode. There cannot be any other updates that the |
595 | /// dominator tree doesn't know about. |
596 | /// |
597 | /// Note that for postdominators it automatically takes care of deleting |
598 | /// a reverse edge internally (so there's no need to swap the parameters). |
599 | /// |
600 | void deleteEdge(NodeT *From, NodeT *To) { |
601 | assert(From)((From) ? static_cast<void> (0) : __assert_fail ("From" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/include/llvm/Support/GenericDomTree.h" , 601, __PRETTY_FUNCTION__)); |
602 | assert(To)((To) ? static_cast<void> (0) : __assert_fail ("To", "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/include/llvm/Support/GenericDomTree.h" , 602, __PRETTY_FUNCTION__)); |
603 | assert(From->getParent() == Parent)((From->getParent() == Parent) ? static_cast<void> ( 0) : __assert_fail ("From->getParent() == Parent", "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/include/llvm/Support/GenericDomTree.h" , 603, __PRETTY_FUNCTION__)); |
604 | assert(To->getParent() == Parent)((To->getParent() == Parent) ? static_cast<void> (0) : __assert_fail ("To->getParent() == Parent", "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/include/llvm/Support/GenericDomTree.h" , 604, __PRETTY_FUNCTION__)); |
605 | DomTreeBuilder::DeleteEdge(*this, From, To); |
606 | } |
607 | |
608 | /// Add a new node to the dominator tree information. |
609 | /// |
610 | /// This creates a new node as a child of DomBB dominator node, linking it |
611 | /// into the children list of the immediate dominator. |
612 | /// |
613 | /// \param BB New node in CFG. |
614 | /// \param DomBB CFG node that is dominator for BB. |
615 | /// \returns New dominator tree node that represents new CFG node. |
616 | /// |
617 | DomTreeNodeBase<NodeT> *addNewBlock(NodeT *BB, NodeT *DomBB) { |
618 | assert(getNode(BB) == nullptr && "Block already in dominator tree!")((getNode(BB) == nullptr && "Block already in dominator tree!" ) ? static_cast<void> (0) : __assert_fail ("getNode(BB) == nullptr && \"Block already in dominator tree!\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/include/llvm/Support/GenericDomTree.h" , 618, __PRETTY_FUNCTION__)); |
619 | DomTreeNodeBase<NodeT> *IDomNode = getNode(DomBB); |
620 | assert(IDomNode && "Not immediate dominator specified for block!")((IDomNode && "Not immediate dominator specified for block!" ) ? static_cast<void> (0) : __assert_fail ("IDomNode && \"Not immediate dominator specified for block!\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/include/llvm/Support/GenericDomTree.h" , 620, __PRETTY_FUNCTION__)); |
621 | DFSInfoValid = false; |
622 | return createChild(BB, IDomNode); |
623 | } |
624 | |
625 | /// Add a new node to the forward dominator tree and make it a new root. |
626 | /// |
627 | /// \param BB New node in CFG. |
628 | /// \returns New dominator tree node that represents new CFG node. |
629 | /// |
630 | DomTreeNodeBase<NodeT> *setNewRoot(NodeT *BB) { |
631 | assert(getNode(BB) == nullptr && "Block already in dominator tree!")((getNode(BB) == nullptr && "Block already in dominator tree!" ) ? static_cast<void> (0) : __assert_fail ("getNode(BB) == nullptr && \"Block already in dominator tree!\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/include/llvm/Support/GenericDomTree.h" , 631, __PRETTY_FUNCTION__)); |
632 | assert(!this->isPostDominator() &&((!this->isPostDominator() && "Cannot change root of post-dominator tree" ) ? static_cast<void> (0) : __assert_fail ("!this->isPostDominator() && \"Cannot change root of post-dominator tree\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/include/llvm/Support/GenericDomTree.h" , 633, __PRETTY_FUNCTION__)) |
633 | "Cannot change root of post-dominator tree")((!this->isPostDominator() && "Cannot change root of post-dominator tree" ) ? static_cast<void> (0) : __assert_fail ("!this->isPostDominator() && \"Cannot change root of post-dominator tree\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/include/llvm/Support/GenericDomTree.h" , 633, __PRETTY_FUNCTION__)); |
634 | DFSInfoValid = false; |
635 | DomTreeNodeBase<NodeT> *NewNode = createNode(BB); |
636 | if (Roots.empty()) { |
637 | addRoot(BB); |
638 | } else { |
639 | assert(Roots.size() == 1)((Roots.size() == 1) ? static_cast<void> (0) : __assert_fail ("Roots.size() == 1", "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/include/llvm/Support/GenericDomTree.h" , 639, __PRETTY_FUNCTION__)); |
640 | NodeT *OldRoot = Roots.front(); |
641 | auto &OldNode = DomTreeNodes[OldRoot]; |
642 | OldNode = NewNode->addChild(std::move(DomTreeNodes[OldRoot])); |
643 | OldNode->IDom = NewNode; |
644 | OldNode->UpdateLevel(); |
645 | Roots[0] = BB; |
646 | } |
647 | return RootNode = NewNode; |
648 | } |
649 | |
650 | /// changeImmediateDominator - This method is used to update the dominator |
651 | /// tree information when a node's immediate dominator changes. |
652 | /// |
653 | void changeImmediateDominator(DomTreeNodeBase<NodeT> *N, |
654 | DomTreeNodeBase<NodeT> *NewIDom) { |
655 | assert(N && NewIDom && "Cannot change null node pointers!")((N && NewIDom && "Cannot change null node pointers!" ) ? static_cast<void> (0) : __assert_fail ("N && NewIDom && \"Cannot change null node pointers!\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/include/llvm/Support/GenericDomTree.h" , 655, __PRETTY_FUNCTION__)); |
656 | DFSInfoValid = false; |
657 | N->setIDom(NewIDom); |
658 | } |
659 | |
660 | void changeImmediateDominator(NodeT *BB, NodeT *NewBB) { |
661 | changeImmediateDominator(getNode(BB), getNode(NewBB)); |
662 | } |
663 | |
664 | /// eraseNode - Removes a node from the dominator tree. Block must not |
665 | /// dominate any other blocks. Removes node from its immediate dominator's |
666 | /// children list. Deletes dominator node associated with basic block BB. |
667 | void eraseNode(NodeT *BB) { |
668 | DomTreeNodeBase<NodeT> *Node = getNode(BB); |
669 | assert(Node && "Removing node that isn't in dominator tree.")((Node && "Removing node that isn't in dominator tree." ) ? static_cast<void> (0) : __assert_fail ("Node && \"Removing node that isn't in dominator tree.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/include/llvm/Support/GenericDomTree.h" , 669, __PRETTY_FUNCTION__)); |
670 | assert(Node->isLeaf() && "Node is not a leaf node.")((Node->isLeaf() && "Node is not a leaf node.") ? static_cast <void> (0) : __assert_fail ("Node->isLeaf() && \"Node is not a leaf node.\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/include/llvm/Support/GenericDomTree.h" , 670, __PRETTY_FUNCTION__)); |
671 | |
672 | DFSInfoValid = false; |
673 | |
674 | // Remove node from immediate dominator's children list. |
675 | DomTreeNodeBase<NodeT> *IDom = Node->getIDom(); |
676 | if (IDom) { |
677 | const auto I = find(IDom->Children, Node); |
678 | assert(I != IDom->Children.end() &&((I != IDom->Children.end() && "Not in immediate dominator children set!" ) ? static_cast<void> (0) : __assert_fail ("I != IDom->Children.end() && \"Not in immediate dominator children set!\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/include/llvm/Support/GenericDomTree.h" , 679, __PRETTY_FUNCTION__)) |
679 | "Not in immediate dominator children set!")((I != IDom->Children.end() && "Not in immediate dominator children set!" ) ? static_cast<void> (0) : __assert_fail ("I != IDom->Children.end() && \"Not in immediate dominator children set!\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/include/llvm/Support/GenericDomTree.h" , 679, __PRETTY_FUNCTION__)); |
680 | // I am no longer your child... |
681 | IDom->Children.erase(I); |
682 | } |
683 | |
684 | DomTreeNodes.erase(BB); |
685 | |
686 | if (!IsPostDom) return; |
687 | |
688 | // Remember to update PostDominatorTree roots. |
689 | auto RIt = llvm::find(Roots, BB); |
690 | if (RIt != Roots.end()) { |
691 | std::swap(*RIt, Roots.back()); |
692 | Roots.pop_back(); |
693 | } |
694 | } |
695 | |
696 | /// splitBlock - BB is split and now it has one successor. Update dominator |
697 | /// tree to reflect this change. |
698 | void splitBlock(NodeT *NewBB) { |
699 | if (IsPostDominator) |
700 | Split<Inverse<NodeT *>>(NewBB); |
701 | else |
702 | Split<NodeT *>(NewBB); |
703 | } |
704 | |
705 | /// print - Convert to human readable form |
706 | /// |
707 | void print(raw_ostream &O) const { |
708 | O << "=============================--------------------------------\n"; |
709 | if (IsPostDominator) |
710 | O << "Inorder PostDominator Tree: "; |
711 | else |
712 | O << "Inorder Dominator Tree: "; |
713 | if (!DFSInfoValid) |
714 | O << "DFSNumbers invalid: " << SlowQueries << " slow queries."; |
715 | O << "\n"; |
716 | |
717 | // The postdom tree can have a null root if there are no returns. |
718 | if (getRootNode()) PrintDomTree<NodeT>(getRootNode(), O, 1); |
719 | O << "Roots: "; |
720 | for (const NodePtr Block : Roots) { |
721 | Block->printAsOperand(O, false); |
722 | O << " "; |
723 | } |
724 | O << "\n"; |
725 | } |
726 | |
727 | public: |
728 | /// updateDFSNumbers - Assign In and Out numbers to the nodes while walking |
729 | /// dominator tree in dfs order. |
730 | void updateDFSNumbers() const { |
731 | if (DFSInfoValid) { |
732 | SlowQueries = 0; |
733 | return; |
734 | } |
735 | |
736 | SmallVector<std::pair<const DomTreeNodeBase<NodeT> *, |
737 | typename DomTreeNodeBase<NodeT>::const_iterator>, |
738 | 32> WorkStack; |
739 | |
740 | const DomTreeNodeBase<NodeT> *ThisRoot = getRootNode(); |
741 | assert((!Parent || ThisRoot) && "Empty constructed DomTree")(((!Parent || ThisRoot) && "Empty constructed DomTree" ) ? static_cast<void> (0) : __assert_fail ("(!Parent || ThisRoot) && \"Empty constructed DomTree\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/include/llvm/Support/GenericDomTree.h" , 741, __PRETTY_FUNCTION__)); |
742 | if (!ThisRoot) |
743 | return; |
744 | |
745 | // Both dominators and postdominators have a single root node. In the case |
746 | // case of PostDominatorTree, this node is a virtual root. |
747 | WorkStack.push_back({ThisRoot, ThisRoot->begin()}); |
748 | |
749 | unsigned DFSNum = 0; |
750 | ThisRoot->DFSNumIn = DFSNum++; |
751 | |
752 | while (!WorkStack.empty()) { |
753 | const DomTreeNodeBase<NodeT> *Node = WorkStack.back().first; |
754 | const auto ChildIt = WorkStack.back().second; |
755 | |
756 | // If we visited all of the children of this node, "recurse" back up the |
757 | // stack setting the DFOutNum. |
758 | if (ChildIt == Node->end()) { |
759 | Node->DFSNumOut = DFSNum++; |
760 | WorkStack.pop_back(); |
761 | } else { |
762 | // Otherwise, recursively visit this child. |
763 | const DomTreeNodeBase<NodeT> *Child = *ChildIt; |
764 | ++WorkStack.back().second; |
765 | |
766 | WorkStack.push_back({Child, Child->begin()}); |
767 | Child->DFSNumIn = DFSNum++; |
768 | } |
769 | } |
770 | |
771 | SlowQueries = 0; |
772 | DFSInfoValid = true; |
773 | } |
774 | |
775 | /// recalculate - compute a dominator tree for the given function |
776 | void recalculate(ParentType &Func) { |
777 | Parent = &Func; |
778 | DomTreeBuilder::Calculate(*this); |
779 | } |
780 | |
781 | void recalculate(ParentType &Func, ArrayRef<UpdateType> Updates) { |
782 | Parent = &Func; |
783 | DomTreeBuilder::CalculateWithUpdates(*this, Updates); |
784 | } |
785 | |
786 | /// verify - checks if the tree is correct. There are 3 level of verification: |
787 | /// - Full -- verifies if the tree is correct by making sure all the |
788 | /// properties (including the parent and the sibling property) |
789 | /// hold. |
790 | /// Takes O(N^3) time. |
791 | /// |
792 | /// - Basic -- checks if the tree is correct, but compares it to a freshly |
793 | /// constructed tree instead of checking the sibling property. |
794 | /// Takes O(N^2) time. |
795 | /// |
796 | /// - Fast -- checks basic tree structure and compares it with a freshly |
797 | /// constructed tree. |
798 | /// Takes O(N^2) time worst case, but is faster in practise (same |
799 | /// as tree construction). |
800 | bool verify(VerificationLevel VL = VerificationLevel::Full) const { |
801 | return DomTreeBuilder::Verify(*this, VL); |
802 | } |
803 | |
804 | void reset() { |
805 | DomTreeNodes.clear(); |
806 | Roots.clear(); |
807 | RootNode = nullptr; |
808 | Parent = nullptr; |
809 | DFSInfoValid = false; |
810 | SlowQueries = 0; |
811 | } |
812 | |
813 | protected: |
814 | void addRoot(NodeT *BB) { this->Roots.push_back(BB); } |
815 | |
816 | DomTreeNodeBase<NodeT> *createChild(NodeT *BB, DomTreeNodeBase<NodeT> *IDom) { |
817 | return (DomTreeNodes[BB] = IDom->addChild( |
818 | std::make_unique<DomTreeNodeBase<NodeT>>(BB, IDom))) |
819 | .get(); |
820 | } |
821 | |
822 | DomTreeNodeBase<NodeT> *createNode(NodeT *BB) { |
823 | return (DomTreeNodes[BB] = |
824 | std::make_unique<DomTreeNodeBase<NodeT>>(BB, nullptr)) |
825 | .get(); |
826 | } |
827 | |
828 | // NewBB is split and now it has one successor. Update dominator tree to |
829 | // reflect this change. |
830 | template <class N> |
831 | void Split(typename GraphTraits<N>::NodeRef NewBB) { |
832 | using GraphT = GraphTraits<N>; |
833 | using NodeRef = typename GraphT::NodeRef; |
834 | assert(std::distance(GraphT::child_begin(NewBB),((std::distance(GraphT::child_begin(NewBB), GraphT::child_end (NewBB)) == 1 && "NewBB should have a single successor!" ) ? static_cast<void> (0) : __assert_fail ("std::distance(GraphT::child_begin(NewBB), GraphT::child_end(NewBB)) == 1 && \"NewBB should have a single successor!\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/include/llvm/Support/GenericDomTree.h" , 836, __PRETTY_FUNCTION__)) |
835 | GraphT::child_end(NewBB)) == 1 &&((std::distance(GraphT::child_begin(NewBB), GraphT::child_end (NewBB)) == 1 && "NewBB should have a single successor!" ) ? static_cast<void> (0) : __assert_fail ("std::distance(GraphT::child_begin(NewBB), GraphT::child_end(NewBB)) == 1 && \"NewBB should have a single successor!\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/include/llvm/Support/GenericDomTree.h" , 836, __PRETTY_FUNCTION__)) |
836 | "NewBB should have a single successor!")((std::distance(GraphT::child_begin(NewBB), GraphT::child_end (NewBB)) == 1 && "NewBB should have a single successor!" ) ? static_cast<void> (0) : __assert_fail ("std::distance(GraphT::child_begin(NewBB), GraphT::child_end(NewBB)) == 1 && \"NewBB should have a single successor!\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/include/llvm/Support/GenericDomTree.h" , 836, __PRETTY_FUNCTION__)); |
837 | NodeRef NewBBSucc = *GraphT::child_begin(NewBB); |
838 | |
839 | SmallVector<NodeRef, 4> PredBlocks; |
840 | for (auto Pred : children<Inverse<N>>(NewBB)) |
841 | PredBlocks.push_back(Pred); |
842 | |
843 | assert(!PredBlocks.empty() && "No predblocks?")((!PredBlocks.empty() && "No predblocks?") ? static_cast <void> (0) : __assert_fail ("!PredBlocks.empty() && \"No predblocks?\"" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/include/llvm/Support/GenericDomTree.h" , 843, __PRETTY_FUNCTION__)); |
844 | |
845 | bool NewBBDominatesNewBBSucc = true; |
846 | for (auto Pred : children<Inverse<N>>(NewBBSucc)) { |
847 | if (Pred != NewBB && !dominates(NewBBSucc, Pred) && |
848 | isReachableFromEntry(Pred)) { |
849 | NewBBDominatesNewBBSucc = false; |
850 | break; |
851 | } |
852 | } |
853 | |
854 | // Find NewBB's immediate dominator and create new dominator tree node for |
855 | // NewBB. |
856 | NodeT *NewBBIDom = nullptr; |
857 | unsigned i = 0; |
858 | for (i = 0; i < PredBlocks.size(); ++i) |
859 | if (isReachableFromEntry(PredBlocks[i])) { |
860 | NewBBIDom = PredBlocks[i]; |
861 | break; |
862 | } |
863 | |
864 | // It's possible that none of the predecessors of NewBB are reachable; |
865 | // in that case, NewBB itself is unreachable, so nothing needs to be |
866 | // changed. |
867 | if (!NewBBIDom) return; |
868 | |
869 | for (i = i + 1; i < PredBlocks.size(); ++i) { |
870 | if (isReachableFromEntry(PredBlocks[i])) |
871 | NewBBIDom = findNearestCommonDominator(NewBBIDom, PredBlocks[i]); |
872 | } |
873 | |
874 | // Create the new dominator tree node... and set the idom of NewBB. |
875 | DomTreeNodeBase<NodeT> *NewBBNode = addNewBlock(NewBB, NewBBIDom); |
876 | |
877 | // If NewBB strictly dominates other blocks, then it is now the immediate |
878 | // dominator of NewBBSucc. Update the dominator tree as appropriate. |
879 | if (NewBBDominatesNewBBSucc) { |
880 | DomTreeNodeBase<NodeT> *NewBBSuccNode = getNode(NewBBSucc); |
881 | changeImmediateDominator(NewBBSuccNode, NewBBNode); |
882 | } |
883 | } |
884 | |
885 | private: |
886 | bool dominatedBySlowTreeWalk(const DomTreeNodeBase<NodeT> *A, |
887 | const DomTreeNodeBase<NodeT> *B) const { |
888 | assert(A != B)((A != B) ? static_cast<void> (0) : __assert_fail ("A != B" , "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/include/llvm/Support/GenericDomTree.h" , 888, __PRETTY_FUNCTION__)); |
889 | assert(isReachableFromEntry(B))((isReachableFromEntry(B)) ? static_cast<void> (0) : __assert_fail ("isReachableFromEntry(B)", "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/include/llvm/Support/GenericDomTree.h" , 889, __PRETTY_FUNCTION__)); |
890 | assert(isReachableFromEntry(A))((isReachableFromEntry(A)) ? static_cast<void> (0) : __assert_fail ("isReachableFromEntry(A)", "/build/llvm-toolchain-snapshot-12~++20200927111121+5811d723998/llvm/include/llvm/Support/GenericDomTree.h" , 890, __PRETTY_FUNCTION__)); |
891 | |
892 | const unsigned ALevel = A->getLevel(); |
893 | const DomTreeNodeBase<NodeT> *IDom; |
894 | |
895 | // Don't walk nodes above A's subtree. When we reach A's level, we must |
896 | // either find A or be in some other subtree not dominated by A. |
897 | while ((IDom = B->getIDom()) != nullptr && IDom->getLevel() >= ALevel) |
898 | B = IDom; // Walk up the tree |
899 | |
900 | return B == A; |
901 | } |
902 | |
903 | /// Wipe this tree's state without releasing any resources. |
904 | /// |
905 | /// This is essentially a post-move helper only. It leaves the object in an |
906 | /// assignable and destroyable state, but otherwise invalid. |
907 | void wipe() { |
908 | DomTreeNodes.clear(); |
909 | RootNode = nullptr; |
910 | Parent = nullptr; |
911 | } |
912 | }; |
913 | |
914 | template <typename T> |
915 | using DomTreeBase = DominatorTreeBase<T, false>; |
916 | |
917 | template <typename T> |
918 | using PostDomTreeBase = DominatorTreeBase<T, true>; |
919 | |
920 | // These two functions are declared out of line as a workaround for building |
921 | // with old (< r147295) versions of clang because of pr11642. |
922 | template <typename NodeT, bool IsPostDom> |
923 | bool DominatorTreeBase<NodeT, IsPostDom>::dominates(const NodeT *A, |
924 | const NodeT *B) const { |
925 | if (A == B) |
926 | return true; |
927 | |
928 | // Cast away the const qualifiers here. This is ok since |
929 | // this function doesn't actually return the values returned |
930 | // from getNode. |
931 | return dominates(getNode(const_cast<NodeT *>(A)), |
932 | getNode(const_cast<NodeT *>(B))); |
933 | } |
934 | template <typename NodeT, bool IsPostDom> |
935 | bool DominatorTreeBase<NodeT, IsPostDom>::properlyDominates( |
936 | const NodeT *A, const NodeT *B) const { |
937 | if (A == B) |
938 | return false; |
939 | |
940 | // Cast away the const qualifiers here. This is ok since |
941 | // this function doesn't actually return the values returned |
942 | // from getNode. |
943 | return dominates(getNode(const_cast<NodeT *>(A)), |
944 | getNode(const_cast<NodeT *>(B))); |
945 | } |
946 | |
947 | } // end namespace llvm |
948 | |
949 | #endif // LLVM_SUPPORT_GENERICDOMTREE_H |