File: | lib/Transforms/Vectorize/SLPVectorizer.cpp |
Warning: | line 2950, column 32 Called C++ object pointer is null |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | //===- SLPVectorizer.cpp - A bottom up SLP Vectorizer ---------------------===// | |||
2 | // | |||
3 | // The LLVM Compiler Infrastructure | |||
4 | // | |||
5 | // This file is distributed under the University of Illinois Open Source | |||
6 | // License. See LICENSE.TXT for details. | |||
7 | // | |||
8 | //===----------------------------------------------------------------------===// | |||
9 | // | |||
10 | // This pass implements the Bottom Up SLP vectorizer. It detects consecutive | |||
11 | // stores that can be put together into vector-stores. Next, it attempts to | |||
12 | // construct vectorizable tree using the use-def chains. If a profitable tree | |||
13 | // was found, the SLP vectorizer performs vectorization on the tree. | |||
14 | // | |||
15 | // The pass is inspired by the work described in the paper: | |||
16 | // "Loop-Aware SLP in GCC" by Ira Rosen, Dorit Nuzman, Ayal Zaks. | |||
17 | // | |||
18 | //===----------------------------------------------------------------------===// | |||
19 | ||||
20 | #include "llvm/Transforms/Vectorize/SLPVectorizer.h" | |||
21 | #include "llvm/ADT/ArrayRef.h" | |||
22 | #include "llvm/ADT/DenseMap.h" | |||
23 | #include "llvm/ADT/DenseSet.h" | |||
24 | #include "llvm/ADT/MapVector.h" | |||
25 | #include "llvm/ADT/None.h" | |||
26 | #include "llvm/ADT/Optional.h" | |||
27 | #include "llvm/ADT/PostOrderIterator.h" | |||
28 | #include "llvm/ADT/STLExtras.h" | |||
29 | #include "llvm/ADT/SetVector.h" | |||
30 | #include "llvm/ADT/SmallPtrSet.h" | |||
31 | #include "llvm/ADT/SmallSet.h" | |||
32 | #include "llvm/ADT/SmallVector.h" | |||
33 | #include "llvm/ADT/Statistic.h" | |||
34 | #include "llvm/ADT/iterator.h" | |||
35 | #include "llvm/ADT/iterator_range.h" | |||
36 | #include "llvm/Analysis/AliasAnalysis.h" | |||
37 | #include "llvm/Analysis/CodeMetrics.h" | |||
38 | #include "llvm/Analysis/DemandedBits.h" | |||
39 | #include "llvm/Analysis/GlobalsModRef.h" | |||
40 | #include "llvm/Analysis/LoopAccessAnalysis.h" | |||
41 | #include "llvm/Analysis/LoopInfo.h" | |||
42 | #include "llvm/Analysis/MemoryLocation.h" | |||
43 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" | |||
44 | #include "llvm/Analysis/ScalarEvolution.h" | |||
45 | #include "llvm/Analysis/ScalarEvolutionExpressions.h" | |||
46 | #include "llvm/Analysis/TargetLibraryInfo.h" | |||
47 | #include "llvm/Analysis/TargetTransformInfo.h" | |||
48 | #include "llvm/Analysis/ValueTracking.h" | |||
49 | #include "llvm/Analysis/VectorUtils.h" | |||
50 | #include "llvm/IR/Attributes.h" | |||
51 | #include "llvm/IR/BasicBlock.h" | |||
52 | #include "llvm/IR/Constant.h" | |||
53 | #include "llvm/IR/Constants.h" | |||
54 | #include "llvm/IR/DataLayout.h" | |||
55 | #include "llvm/IR/DebugLoc.h" | |||
56 | #include "llvm/IR/DerivedTypes.h" | |||
57 | #include "llvm/IR/Dominators.h" | |||
58 | #include "llvm/IR/Function.h" | |||
59 | #include "llvm/IR/IRBuilder.h" | |||
60 | #include "llvm/IR/InstrTypes.h" | |||
61 | #include "llvm/IR/Instruction.h" | |||
62 | #include "llvm/IR/Instructions.h" | |||
63 | #include "llvm/IR/IntrinsicInst.h" | |||
64 | #include "llvm/IR/Intrinsics.h" | |||
65 | #include "llvm/IR/Module.h" | |||
66 | #include "llvm/IR/NoFolder.h" | |||
67 | #include "llvm/IR/Operator.h" | |||
68 | #include "llvm/IR/PassManager.h" | |||
69 | #include "llvm/IR/PatternMatch.h" | |||
70 | #include "llvm/IR/Type.h" | |||
71 | #include "llvm/IR/Use.h" | |||
72 | #include "llvm/IR/User.h" | |||
73 | #include "llvm/IR/Value.h" | |||
74 | #include "llvm/IR/ValueHandle.h" | |||
75 | #include "llvm/IR/Verifier.h" | |||
76 | #include "llvm/Pass.h" | |||
77 | #include "llvm/Support/Casting.h" | |||
78 | #include "llvm/Support/CommandLine.h" | |||
79 | #include "llvm/Support/Compiler.h" | |||
80 | #include "llvm/Support/DOTGraphTraits.h" | |||
81 | #include "llvm/Support/Debug.h" | |||
82 | #include "llvm/Support/ErrorHandling.h" | |||
83 | #include "llvm/Support/GraphWriter.h" | |||
84 | #include "llvm/Support/KnownBits.h" | |||
85 | #include "llvm/Support/MathExtras.h" | |||
86 | #include "llvm/Support/raw_ostream.h" | |||
87 | #include "llvm/Transforms/Utils/LoopUtils.h" | |||
88 | #include "llvm/Transforms/Vectorize.h" | |||
89 | #include <algorithm> | |||
90 | #include <cassert> | |||
91 | #include <cstdint> | |||
92 | #include <iterator> | |||
93 | #include <memory> | |||
94 | #include <set> | |||
95 | #include <string> | |||
96 | #include <tuple> | |||
97 | #include <utility> | |||
98 | #include <vector> | |||
99 | ||||
100 | using namespace llvm; | |||
101 | using namespace llvm::PatternMatch; | |||
102 | using namespace slpvectorizer; | |||
103 | ||||
104 | #define SV_NAME"slp-vectorizer" "slp-vectorizer" | |||
105 | #define DEBUG_TYPE"SLP" "SLP" | |||
106 | ||||
107 | STATISTIC(NumVectorInstructions, "Number of vector instructions generated")static llvm::Statistic NumVectorInstructions = {"SLP", "NumVectorInstructions" , "Number of vector instructions generated", {0}, {false}}; | |||
108 | ||||
109 | static cl::opt<int> | |||
110 | SLPCostThreshold("slp-threshold", cl::init(0), cl::Hidden, | |||
111 | cl::desc("Only vectorize if you gain more than this " | |||
112 | "number ")); | |||
113 | ||||
114 | static cl::opt<bool> | |||
115 | ShouldVectorizeHor("slp-vectorize-hor", cl::init(true), cl::Hidden, | |||
116 | cl::desc("Attempt to vectorize horizontal reductions")); | |||
117 | ||||
118 | static cl::opt<bool> ShouldStartVectorizeHorAtStore( | |||
119 | "slp-vectorize-hor-store", cl::init(false), cl::Hidden, | |||
120 | cl::desc( | |||
121 | "Attempt to vectorize horizontal reductions feeding into a store")); | |||
122 | ||||
123 | static cl::opt<int> | |||
124 | MaxVectorRegSizeOption("slp-max-reg-size", cl::init(128), cl::Hidden, | |||
125 | cl::desc("Attempt to vectorize for this register size in bits")); | |||
126 | ||||
127 | /// Limits the size of scheduling regions in a block. | |||
128 | /// It avoid long compile times for _very_ large blocks where vector | |||
129 | /// instructions are spread over a wide range. | |||
130 | /// This limit is way higher than needed by real-world functions. | |||
131 | static cl::opt<int> | |||
132 | ScheduleRegionSizeBudget("slp-schedule-budget", cl::init(100000), cl::Hidden, | |||
133 | cl::desc("Limit the size of the SLP scheduling region per block")); | |||
134 | ||||
135 | static cl::opt<int> MinVectorRegSizeOption( | |||
136 | "slp-min-reg-size", cl::init(128), cl::Hidden, | |||
137 | cl::desc("Attempt to vectorize for this register size in bits")); | |||
138 | ||||
139 | static cl::opt<unsigned> RecursionMaxDepth( | |||
140 | "slp-recursion-max-depth", cl::init(12), cl::Hidden, | |||
141 | cl::desc("Limit the recursion depth when building a vectorizable tree")); | |||
142 | ||||
143 | static cl::opt<unsigned> MinTreeSize( | |||
144 | "slp-min-tree-size", cl::init(3), cl::Hidden, | |||
145 | cl::desc("Only vectorize small trees if they are fully vectorizable")); | |||
146 | ||||
147 | static cl::opt<bool> | |||
148 | ViewSLPTree("view-slp-tree", cl::Hidden, | |||
149 | cl::desc("Display the SLP trees with Graphviz")); | |||
150 | ||||
151 | // Limit the number of alias checks. The limit is chosen so that | |||
152 | // it has no negative effect on the llvm benchmarks. | |||
153 | static const unsigned AliasedCheckLimit = 10; | |||
154 | ||||
155 | // Another limit for the alias checks: The maximum distance between load/store | |||
156 | // instructions where alias checks are done. | |||
157 | // This limit is useful for very large basic blocks. | |||
158 | static const unsigned MaxMemDepDistance = 160; | |||
159 | ||||
160 | /// If the ScheduleRegionSizeBudget is exhausted, we allow small scheduling | |||
161 | /// regions to be handled. | |||
162 | static const int MinScheduleRegionSize = 16; | |||
163 | ||||
164 | /// \brief Predicate for the element types that the SLP vectorizer supports. | |||
165 | /// | |||
166 | /// The most important thing to filter here are types which are invalid in LLVM | |||
167 | /// vectors. We also filter target specific types which have absolutely no | |||
168 | /// meaningful vectorization path such as x86_fp80 and ppc_f128. This just | |||
169 | /// avoids spending time checking the cost model and realizing that they will | |||
170 | /// be inevitably scalarized. | |||
171 | static bool isValidElementType(Type *Ty) { | |||
172 | return VectorType::isValidElementType(Ty) && !Ty->isX86_FP80Ty() && | |||
173 | !Ty->isPPC_FP128Ty(); | |||
174 | } | |||
175 | ||||
176 | /// \returns true if all of the instructions in \p VL are in the same block or | |||
177 | /// false otherwise. | |||
178 | static bool allSameBlock(ArrayRef<Value *> VL) { | |||
179 | Instruction *I0 = dyn_cast<Instruction>(VL[0]); | |||
180 | if (!I0) | |||
181 | return false; | |||
182 | BasicBlock *BB = I0->getParent(); | |||
183 | for (int i = 1, e = VL.size(); i < e; i++) { | |||
184 | Instruction *I = dyn_cast<Instruction>(VL[i]); | |||
185 | if (!I) | |||
186 | return false; | |||
187 | ||||
188 | if (BB != I->getParent()) | |||
189 | return false; | |||
190 | } | |||
191 | return true; | |||
192 | } | |||
193 | ||||
194 | /// \returns True if all of the values in \p VL are constants. | |||
195 | static bool allConstant(ArrayRef<Value *> VL) { | |||
196 | for (Value *i : VL) | |||
197 | if (!isa<Constant>(i)) | |||
198 | return false; | |||
199 | return true; | |||
200 | } | |||
201 | ||||
202 | /// \returns True if all of the values in \p VL are identical. | |||
203 | static bool isSplat(ArrayRef<Value *> VL) { | |||
204 | for (unsigned i = 1, e = VL.size(); i < e; ++i) | |||
205 | if (VL[i] != VL[0]) | |||
206 | return false; | |||
207 | return true; | |||
208 | } | |||
209 | ||||
210 | /// Checks if the vector of instructions can be represented as a shuffle, like: | |||
211 | /// %x0 = extractelement <4 x i8> %x, i32 0 | |||
212 | /// %x3 = extractelement <4 x i8> %x, i32 3 | |||
213 | /// %y1 = extractelement <4 x i8> %y, i32 1 | |||
214 | /// %y2 = extractelement <4 x i8> %y, i32 2 | |||
215 | /// %x0x0 = mul i8 %x0, %x0 | |||
216 | /// %x3x3 = mul i8 %x3, %x3 | |||
217 | /// %y1y1 = mul i8 %y1, %y1 | |||
218 | /// %y2y2 = mul i8 %y2, %y2 | |||
219 | /// %ins1 = insertelement <4 x i8> undef, i8 %x0x0, i32 0 | |||
220 | /// %ins2 = insertelement <4 x i8> %ins1, i8 %x3x3, i32 1 | |||
221 | /// %ins3 = insertelement <4 x i8> %ins2, i8 %y1y1, i32 2 | |||
222 | /// %ins4 = insertelement <4 x i8> %ins3, i8 %y2y2, i32 3 | |||
223 | /// ret <4 x i8> %ins4 | |||
224 | /// can be transformed into: | |||
225 | /// %1 = shufflevector <4 x i8> %x, <4 x i8> %y, <4 x i32> <i32 0, i32 3, i32 5, | |||
226 | /// i32 6> | |||
227 | /// %2 = mul <4 x i8> %1, %1 | |||
228 | /// ret <4 x i8> %2 | |||
229 | /// We convert this initially to something like: | |||
230 | /// %x0 = extractelement <4 x i8> %x, i32 0 | |||
231 | /// %x3 = extractelement <4 x i8> %x, i32 3 | |||
232 | /// %y1 = extractelement <4 x i8> %y, i32 1 | |||
233 | /// %y2 = extractelement <4 x i8> %y, i32 2 | |||
234 | /// %1 = insertelement <4 x i8> undef, i8 %x0, i32 0 | |||
235 | /// %2 = insertelement <4 x i8> %1, i8 %x3, i32 1 | |||
236 | /// %3 = insertelement <4 x i8> %2, i8 %y1, i32 2 | |||
237 | /// %4 = insertelement <4 x i8> %3, i8 %y2, i32 3 | |||
238 | /// %5 = mul <4 x i8> %4, %4 | |||
239 | /// %6 = extractelement <4 x i8> %5, i32 0 | |||
240 | /// %ins1 = insertelement <4 x i8> undef, i8 %6, i32 0 | |||
241 | /// %7 = extractelement <4 x i8> %5, i32 1 | |||
242 | /// %ins2 = insertelement <4 x i8> %ins1, i8 %7, i32 1 | |||
243 | /// %8 = extractelement <4 x i8> %5, i32 2 | |||
244 | /// %ins3 = insertelement <4 x i8> %ins2, i8 %8, i32 2 | |||
245 | /// %9 = extractelement <4 x i8> %5, i32 3 | |||
246 | /// %ins4 = insertelement <4 x i8> %ins3, i8 %9, i32 3 | |||
247 | /// ret <4 x i8> %ins4 | |||
248 | /// InstCombiner transforms this into a shuffle and vector mul | |||
249 | static Optional<TargetTransformInfo::ShuffleKind> | |||
250 | isShuffle(ArrayRef<Value *> VL) { | |||
251 | auto *EI0 = cast<ExtractElementInst>(VL[0]); | |||
252 | unsigned Size = EI0->getVectorOperandType()->getVectorNumElements(); | |||
253 | Value *Vec1 = nullptr; | |||
254 | Value *Vec2 = nullptr; | |||
255 | enum ShuffleMode {Unknown, FirstAlternate, SecondAlternate, Permute}; | |||
256 | ShuffleMode CommonShuffleMode = Unknown; | |||
257 | for (unsigned I = 0, E = VL.size(); I < E; ++I) { | |||
258 | auto *EI = cast<ExtractElementInst>(VL[I]); | |||
259 | auto *Vec = EI->getVectorOperand(); | |||
260 | // All vector operands must have the same number of vector elements. | |||
261 | if (Vec->getType()->getVectorNumElements() != Size) | |||
262 | return None; | |||
263 | auto *Idx = dyn_cast<ConstantInt>(EI->getIndexOperand()); | |||
264 | if (!Idx) | |||
265 | return None; | |||
266 | // Undefined behavior if Idx is negative or >= Size. | |||
267 | if (Idx->getValue().uge(Size)) | |||
268 | continue; | |||
269 | unsigned IntIdx = Idx->getValue().getZExtValue(); | |||
270 | // We can extractelement from undef vector. | |||
271 | if (isa<UndefValue>(Vec)) | |||
272 | continue; | |||
273 | // For correct shuffling we have to have at most 2 different vector operands | |||
274 | // in all extractelement instructions. | |||
275 | if (Vec1 && Vec2 && Vec != Vec1 && Vec != Vec2) | |||
276 | return None; | |||
277 | if (CommonShuffleMode == Permute) | |||
278 | continue; | |||
279 | // If the extract index is not the same as the operation number, it is a | |||
280 | // permutation. | |||
281 | if (IntIdx != I) { | |||
282 | CommonShuffleMode = Permute; | |||
283 | continue; | |||
284 | } | |||
285 | // Check the shuffle mode for the current operation. | |||
286 | if (!Vec1) | |||
287 | Vec1 = Vec; | |||
288 | else if (Vec != Vec1) | |||
289 | Vec2 = Vec; | |||
290 | // Example: shufflevector A, B, <0,5,2,7> | |||
291 | // I is odd and IntIdx for A == I - FirstAlternate shuffle. | |||
292 | // I is even and IntIdx for B == I - FirstAlternate shuffle. | |||
293 | // Example: shufflevector A, B, <4,1,6,3> | |||
294 | // I is even and IntIdx for A == I - SecondAlternate shuffle. | |||
295 | // I is odd and IntIdx for B == I - SecondAlternate shuffle. | |||
296 | const bool IIsEven = I & 1; | |||
297 | const bool CurrVecIsA = Vec == Vec1; | |||
298 | const bool IIsOdd = !IIsEven; | |||
299 | const bool CurrVecIsB = !CurrVecIsA; | |||
300 | ShuffleMode CurrentShuffleMode = | |||
301 | ((IIsOdd && CurrVecIsA) || (IIsEven && CurrVecIsB)) ? FirstAlternate | |||
302 | : SecondAlternate; | |||
303 | // Common mode is not set or the same as the shuffle mode of the current | |||
304 | // operation - alternate. | |||
305 | if (CommonShuffleMode == Unknown) | |||
306 | CommonShuffleMode = CurrentShuffleMode; | |||
307 | // Common shuffle mode is not the same as the shuffle mode of the current | |||
308 | // operation - permutation. | |||
309 | if (CommonShuffleMode != CurrentShuffleMode) | |||
310 | CommonShuffleMode = Permute; | |||
311 | } | |||
312 | // If we're not crossing lanes in different vectors, consider it as blending. | |||
313 | if ((CommonShuffleMode == FirstAlternate || | |||
314 | CommonShuffleMode == SecondAlternate) && | |||
315 | Vec2) | |||
316 | return TargetTransformInfo::SK_Alternate; | |||
317 | // If Vec2 was never used, we have a permutation of a single vector, otherwise | |||
318 | // we have permutation of 2 vectors. | |||
319 | return Vec2 ? TargetTransformInfo::SK_PermuteTwoSrc | |||
320 | : TargetTransformInfo::SK_PermuteSingleSrc; | |||
321 | } | |||
322 | ||||
323 | ///\returns Opcode that can be clubbed with \p Op to create an alternate | |||
324 | /// sequence which can later be merged as a ShuffleVector instruction. | |||
325 | static unsigned getAltOpcode(unsigned Op) { | |||
326 | switch (Op) { | |||
327 | case Instruction::FAdd: | |||
328 | return Instruction::FSub; | |||
329 | case Instruction::FSub: | |||
330 | return Instruction::FAdd; | |||
331 | case Instruction::Add: | |||
332 | return Instruction::Sub; | |||
333 | case Instruction::Sub: | |||
334 | return Instruction::Add; | |||
335 | default: | |||
336 | return 0; | |||
337 | } | |||
338 | } | |||
339 | ||||
340 | static bool isOdd(unsigned Value) { | |||
341 | return Value & 1; | |||
342 | } | |||
343 | ||||
344 | static bool sameOpcodeOrAlt(unsigned Opcode, unsigned AltOpcode, | |||
345 | unsigned CheckedOpcode) { | |||
346 | return Opcode == CheckedOpcode || AltOpcode == CheckedOpcode; | |||
347 | } | |||
348 | ||||
349 | /// Chooses the correct key for scheduling data. If \p Op has the same (or | |||
350 | /// alternate) opcode as \p OpValue, the key is \p Op. Otherwise the key is \p | |||
351 | /// OpValue. | |||
352 | static Value *isOneOf(Value *OpValue, Value *Op) { | |||
353 | auto *I = dyn_cast<Instruction>(Op); | |||
354 | if (!I) | |||
355 | return OpValue; | |||
356 | auto *OpInst = cast<Instruction>(OpValue); | |||
357 | unsigned OpInstOpcode = OpInst->getOpcode(); | |||
358 | unsigned IOpcode = I->getOpcode(); | |||
359 | if (sameOpcodeOrAlt(OpInstOpcode, getAltOpcode(OpInstOpcode), IOpcode)) | |||
360 | return Op; | |||
361 | return OpValue; | |||
362 | } | |||
363 | ||||
364 | namespace { | |||
365 | ||||
366 | /// Contains data for the instructions going to be vectorized. | |||
367 | struct RawInstructionsData { | |||
368 | /// Main Opcode of the instructions going to be vectorized. | |||
369 | unsigned Opcode = 0; | |||
370 | ||||
371 | /// The list of instructions have some instructions with alternate opcodes. | |||
372 | bool HasAltOpcodes = false; | |||
373 | }; | |||
374 | ||||
375 | } // end anonymous namespace | |||
376 | ||||
377 | /// Checks the list of the vectorized instructions \p VL and returns info about | |||
378 | /// this list. | |||
379 | static RawInstructionsData getMainOpcode(ArrayRef<Value *> VL) { | |||
380 | auto *I0 = dyn_cast<Instruction>(VL[0]); | |||
381 | if (!I0) | |||
382 | return {}; | |||
383 | RawInstructionsData Res; | |||
384 | unsigned Opcode = I0->getOpcode(); | |||
385 | // Walk through the list of the vectorized instructions | |||
386 | // in order to check its structure described by RawInstructionsData. | |||
387 | for (unsigned Cnt = 0, E = VL.size(); Cnt != E; ++Cnt) { | |||
388 | auto *I = dyn_cast<Instruction>(VL[Cnt]); | |||
389 | if (!I) | |||
390 | return {}; | |||
391 | if (Opcode != I->getOpcode()) | |||
392 | Res.HasAltOpcodes = true; | |||
393 | } | |||
394 | Res.Opcode = Opcode; | |||
395 | return Res; | |||
396 | } | |||
397 | ||||
398 | namespace { | |||
399 | ||||
400 | /// Main data required for vectorization of instructions. | |||
401 | struct InstructionsState { | |||
402 | /// The very first instruction in the list with the main opcode. | |||
403 | Value *OpValue = nullptr; | |||
404 | ||||
405 | /// The main opcode for the list of instructions. | |||
406 | unsigned Opcode = 0; | |||
407 | ||||
408 | /// Some of the instructions in the list have alternate opcodes. | |||
409 | bool IsAltShuffle = false; | |||
410 | ||||
411 | InstructionsState() = default; | |||
412 | InstructionsState(Value *OpValue, unsigned Opcode, bool IsAltShuffle) | |||
413 | : OpValue(OpValue), Opcode(Opcode), IsAltShuffle(IsAltShuffle) {} | |||
414 | }; | |||
415 | ||||
416 | } // end anonymous namespace | |||
417 | ||||
418 | /// \returns analysis of the Instructions in \p VL described in | |||
419 | /// InstructionsState, the Opcode that we suppose the whole list | |||
420 | /// could be vectorized even if its structure is diverse. | |||
421 | static InstructionsState getSameOpcode(ArrayRef<Value *> VL) { | |||
422 | auto Res = getMainOpcode(VL); | |||
423 | unsigned Opcode = Res.Opcode; | |||
424 | if (!Res.HasAltOpcodes) | |||
425 | return InstructionsState(VL[0], Opcode, false); | |||
426 | auto *OpInst = cast<Instruction>(VL[0]); | |||
427 | unsigned AltOpcode = getAltOpcode(Opcode); | |||
428 | // Examine each element in the list instructions VL to determine | |||
429 | // if some operations there could be considered as an alternative | |||
430 | // (for example as subtraction relates to addition operation). | |||
431 | for (int Cnt = 0, E = VL.size(); Cnt < E; Cnt++) { | |||
432 | auto *I = cast<Instruction>(VL[Cnt]); | |||
433 | unsigned InstOpcode = I->getOpcode(); | |||
434 | if ((Res.HasAltOpcodes && | |||
435 | InstOpcode != (isOdd(Cnt) ? AltOpcode : Opcode)) || | |||
436 | (!Res.HasAltOpcodes && InstOpcode != Opcode)) { | |||
437 | return InstructionsState(OpInst, 0, false); | |||
438 | } | |||
439 | } | |||
440 | return InstructionsState(OpInst, Opcode, Res.HasAltOpcodes); | |||
441 | } | |||
442 | ||||
443 | /// \returns true if all of the values in \p VL have the same type or false | |||
444 | /// otherwise. | |||
445 | static bool allSameType(ArrayRef<Value *> VL) { | |||
446 | Type *Ty = VL[0]->getType(); | |||
447 | for (int i = 1, e = VL.size(); i < e; i++) | |||
448 | if (VL[i]->getType() != Ty) | |||
449 | return false; | |||
450 | ||||
451 | return true; | |||
452 | } | |||
453 | ||||
454 | /// \returns True if Extract{Value,Element} instruction extracts element Idx. | |||
455 | static Optional<unsigned> getExtractIndex(Instruction *E) { | |||
456 | unsigned Opcode = E->getOpcode(); | |||
457 | assert((Opcode == Instruction::ExtractElement ||(static_cast <bool> ((Opcode == Instruction::ExtractElement || Opcode == Instruction::ExtractValue) && "Expected extractelement or extractvalue instruction." ) ? void (0) : __assert_fail ("(Opcode == Instruction::ExtractElement || Opcode == Instruction::ExtractValue) && \"Expected extractelement or extractvalue instruction.\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 459, __extension__ __PRETTY_FUNCTION__)) | |||
458 | Opcode == Instruction::ExtractValue) &&(static_cast <bool> ((Opcode == Instruction::ExtractElement || Opcode == Instruction::ExtractValue) && "Expected extractelement or extractvalue instruction." ) ? void (0) : __assert_fail ("(Opcode == Instruction::ExtractElement || Opcode == Instruction::ExtractValue) && \"Expected extractelement or extractvalue instruction.\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 459, __extension__ __PRETTY_FUNCTION__)) | |||
459 | "Expected extractelement or extractvalue instruction.")(static_cast <bool> ((Opcode == Instruction::ExtractElement || Opcode == Instruction::ExtractValue) && "Expected extractelement or extractvalue instruction." ) ? void (0) : __assert_fail ("(Opcode == Instruction::ExtractElement || Opcode == Instruction::ExtractValue) && \"Expected extractelement or extractvalue instruction.\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 459, __extension__ __PRETTY_FUNCTION__)); | |||
460 | if (Opcode == Instruction::ExtractElement) { | |||
461 | auto *CI = dyn_cast<ConstantInt>(E->getOperand(1)); | |||
462 | if (!CI) | |||
463 | return None; | |||
464 | return CI->getZExtValue(); | |||
465 | } | |||
466 | ExtractValueInst *EI = cast<ExtractValueInst>(E); | |||
467 | if (EI->getNumIndices() != 1) | |||
468 | return None; | |||
469 | return *EI->idx_begin(); | |||
470 | } | |||
471 | ||||
472 | /// \returns True if in-tree use also needs extract. This refers to | |||
473 | /// possible scalar operand in vectorized instruction. | |||
474 | static bool InTreeUserNeedToExtract(Value *Scalar, Instruction *UserInst, | |||
475 | TargetLibraryInfo *TLI) { | |||
476 | unsigned Opcode = UserInst->getOpcode(); | |||
477 | switch (Opcode) { | |||
478 | case Instruction::Load: { | |||
479 | LoadInst *LI = cast<LoadInst>(UserInst); | |||
480 | return (LI->getPointerOperand() == Scalar); | |||
481 | } | |||
482 | case Instruction::Store: { | |||
483 | StoreInst *SI = cast<StoreInst>(UserInst); | |||
484 | return (SI->getPointerOperand() == Scalar); | |||
485 | } | |||
486 | case Instruction::Call: { | |||
487 | CallInst *CI = cast<CallInst>(UserInst); | |||
488 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); | |||
489 | if (hasVectorInstrinsicScalarOpd(ID, 1)) { | |||
490 | return (CI->getArgOperand(1) == Scalar); | |||
491 | } | |||
492 | LLVM_FALLTHROUGH[[clang::fallthrough]]; | |||
493 | } | |||
494 | default: | |||
495 | return false; | |||
496 | } | |||
497 | } | |||
498 | ||||
499 | /// \returns the AA location that is being access by the instruction. | |||
500 | static MemoryLocation getLocation(Instruction *I, AliasAnalysis *AA) { | |||
501 | if (StoreInst *SI = dyn_cast<StoreInst>(I)) | |||
502 | return MemoryLocation::get(SI); | |||
503 | if (LoadInst *LI = dyn_cast<LoadInst>(I)) | |||
504 | return MemoryLocation::get(LI); | |||
505 | return MemoryLocation(); | |||
506 | } | |||
507 | ||||
508 | /// \returns True if the instruction is not a volatile or atomic load/store. | |||
509 | static bool isSimple(Instruction *I) { | |||
510 | if (LoadInst *LI = dyn_cast<LoadInst>(I)) | |||
511 | return LI->isSimple(); | |||
512 | if (StoreInst *SI = dyn_cast<StoreInst>(I)) | |||
513 | return SI->isSimple(); | |||
514 | if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I)) | |||
515 | return !MI->isVolatile(); | |||
516 | return true; | |||
517 | } | |||
518 | ||||
519 | namespace llvm { | |||
520 | ||||
521 | namespace slpvectorizer { | |||
522 | ||||
523 | /// Bottom Up SLP Vectorizer. | |||
524 | class BoUpSLP { | |||
525 | public: | |||
526 | using ValueList = SmallVector<Value *, 8>; | |||
527 | using InstrList = SmallVector<Instruction *, 16>; | |||
528 | using ValueSet = SmallPtrSet<Value *, 16>; | |||
529 | using StoreList = SmallVector<StoreInst *, 8>; | |||
530 | using ExtraValueToDebugLocsMap = | |||
531 | MapVector<Value *, SmallVector<Instruction *, 2>>; | |||
532 | ||||
533 | BoUpSLP(Function *Func, ScalarEvolution *Se, TargetTransformInfo *Tti, | |||
534 | TargetLibraryInfo *TLi, AliasAnalysis *Aa, LoopInfo *Li, | |||
535 | DominatorTree *Dt, AssumptionCache *AC, DemandedBits *DB, | |||
536 | const DataLayout *DL, OptimizationRemarkEmitter *ORE) | |||
537 | : F(Func), SE(Se), TTI(Tti), TLI(TLi), AA(Aa), LI(Li), DT(Dt), AC(AC), | |||
538 | DB(DB), DL(DL), ORE(ORE), Builder(Se->getContext()) { | |||
539 | CodeMetrics::collectEphemeralValues(F, AC, EphValues); | |||
540 | // Use the vector register size specified by the target unless overridden | |||
541 | // by a command-line option. | |||
542 | // TODO: It would be better to limit the vectorization factor based on | |||
543 | // data type rather than just register size. For example, x86 AVX has | |||
544 | // 256-bit registers, but it does not support integer operations | |||
545 | // at that width (that requires AVX2). | |||
546 | if (MaxVectorRegSizeOption.getNumOccurrences()) | |||
547 | MaxVecRegSize = MaxVectorRegSizeOption; | |||
548 | else | |||
549 | MaxVecRegSize = TTI->getRegisterBitWidth(true); | |||
550 | ||||
551 | if (MinVectorRegSizeOption.getNumOccurrences()) | |||
552 | MinVecRegSize = MinVectorRegSizeOption; | |||
553 | else | |||
554 | MinVecRegSize = TTI->getMinVectorRegisterBitWidth(); | |||
555 | } | |||
556 | ||||
557 | /// \brief Vectorize the tree that starts with the elements in \p VL. | |||
558 | /// Returns the vectorized root. | |||
559 | Value *vectorizeTree(); | |||
560 | ||||
561 | /// Vectorize the tree but with the list of externally used values \p | |||
562 | /// ExternallyUsedValues. Values in this MapVector can be replaced but the | |||
563 | /// generated extractvalue instructions. | |||
564 | Value *vectorizeTree(ExtraValueToDebugLocsMap &ExternallyUsedValues); | |||
565 | ||||
566 | /// \returns the cost incurred by unwanted spills and fills, caused by | |||
567 | /// holding live values over call sites. | |||
568 | int getSpillCost(); | |||
569 | ||||
570 | /// \returns the vectorization cost of the subtree that starts at \p VL. | |||
571 | /// A negative number means that this is profitable. | |||
572 | int getTreeCost(); | |||
573 | ||||
574 | /// Construct a vectorizable tree that starts at \p Roots, ignoring users for | |||
575 | /// the purpose of scheduling and extraction in the \p UserIgnoreLst. | |||
576 | void buildTree(ArrayRef<Value *> Roots, | |||
577 | ArrayRef<Value *> UserIgnoreLst = None); | |||
578 | ||||
579 | /// Construct a vectorizable tree that starts at \p Roots, ignoring users for | |||
580 | /// the purpose of scheduling and extraction in the \p UserIgnoreLst taking | |||
581 | /// into account (anf updating it, if required) list of externally used | |||
582 | /// values stored in \p ExternallyUsedValues. | |||
583 | void buildTree(ArrayRef<Value *> Roots, | |||
584 | ExtraValueToDebugLocsMap &ExternallyUsedValues, | |||
585 | ArrayRef<Value *> UserIgnoreLst = None); | |||
586 | ||||
587 | /// Clear the internal data structures that are created by 'buildTree'. | |||
588 | void deleteTree() { | |||
589 | VectorizableTree.clear(); | |||
590 | ScalarToTreeEntry.clear(); | |||
591 | MustGather.clear(); | |||
592 | ExternalUses.clear(); | |||
593 | NumOpsWantToKeepOrder.clear(); | |||
594 | NumOpsWantToKeepOriginalOrder = 0; | |||
595 | for (auto &Iter : BlocksSchedules) { | |||
596 | BlockScheduling *BS = Iter.second.get(); | |||
597 | BS->clear(); | |||
598 | } | |||
599 | MinBWs.clear(); | |||
600 | } | |||
601 | ||||
602 | unsigned getTreeSize() const { return VectorizableTree.size(); } | |||
603 | ||||
604 | /// \brief Perform LICM and CSE on the newly generated gather sequences. | |||
605 | void optimizeGatherSequence(); | |||
606 | ||||
607 | /// \returns The best order of instructions for vectorization. | |||
608 | Optional<ArrayRef<unsigned>> bestOrder() const { | |||
609 | auto I = std::max_element( | |||
610 | NumOpsWantToKeepOrder.begin(), NumOpsWantToKeepOrder.end(), | |||
611 | [](const decltype(NumOpsWantToKeepOrder)::value_type &D1, | |||
612 | const decltype(NumOpsWantToKeepOrder)::value_type &D2) { | |||
613 | return D1.second < D2.second; | |||
614 | }); | |||
615 | if (I == NumOpsWantToKeepOrder.end() || | |||
616 | I->getSecond() <= NumOpsWantToKeepOriginalOrder) | |||
617 | return None; | |||
618 | ||||
619 | return makeArrayRef(I->getFirst()); | |||
620 | } | |||
621 | ||||
622 | /// \return The vector element size in bits to use when vectorizing the | |||
623 | /// expression tree ending at \p V. If V is a store, the size is the width of | |||
624 | /// the stored value. Otherwise, the size is the width of the largest loaded | |||
625 | /// value reaching V. This method is used by the vectorizer to calculate | |||
626 | /// vectorization factors. | |||
627 | unsigned getVectorElementSize(Value *V); | |||
628 | ||||
629 | /// Compute the minimum type sizes required to represent the entries in a | |||
630 | /// vectorizable tree. | |||
631 | void computeMinimumValueSizes(); | |||
632 | ||||
633 | // \returns maximum vector register size as set by TTI or overridden by cl::opt. | |||
634 | unsigned getMaxVecRegSize() const { | |||
635 | return MaxVecRegSize; | |||
636 | } | |||
637 | ||||
638 | // \returns minimum vector register size as set by cl::opt. | |||
639 | unsigned getMinVecRegSize() const { | |||
640 | return MinVecRegSize; | |||
641 | } | |||
642 | ||||
643 | /// \brief Check if ArrayType or StructType is isomorphic to some VectorType. | |||
644 | /// | |||
645 | /// \returns number of elements in vector if isomorphism exists, 0 otherwise. | |||
646 | unsigned canMapToVector(Type *T, const DataLayout &DL) const; | |||
647 | ||||
648 | /// \returns True if the VectorizableTree is both tiny and not fully | |||
649 | /// vectorizable. We do not vectorize such trees. | |||
650 | bool isTreeTinyAndNotFullyVectorizable(); | |||
651 | ||||
652 | OptimizationRemarkEmitter *getORE() { return ORE; } | |||
653 | ||||
654 | private: | |||
655 | struct TreeEntry; | |||
656 | ||||
657 | /// Checks if all users of \p I are the part of the vectorization tree. | |||
658 | bool areAllUsersVectorized(Instruction *I) const; | |||
659 | ||||
660 | /// \returns the cost of the vectorizable entry. | |||
661 | int getEntryCost(TreeEntry *E); | |||
662 | ||||
663 | /// This is the recursive part of buildTree. | |||
664 | void buildTree_rec(ArrayRef<Value *> Roots, unsigned Depth, int); | |||
665 | ||||
666 | /// \returns true if the ExtractElement/ExtractValue instructions in \p VL can | |||
667 | /// be vectorized to use the original vector (or aggregate "bitcast" to a | |||
668 | /// vector) and sets \p CurrentOrder to the identity permutation; otherwise | |||
669 | /// returns false, setting \p CurrentOrder to either an empty vector or a | |||
670 | /// non-identity permutation that allows to reuse extract instructions. | |||
671 | bool canReuseExtract(ArrayRef<Value *> VL, Value *OpValue, | |||
672 | SmallVectorImpl<unsigned> &CurrentOrder) const; | |||
673 | ||||
674 | /// Vectorize a single entry in the tree. | |||
675 | Value *vectorizeTree(TreeEntry *E); | |||
676 | ||||
677 | /// Vectorize a single entry in the tree, starting in \p VL. | |||
678 | Value *vectorizeTree(ArrayRef<Value *> VL); | |||
679 | ||||
680 | /// \returns the scalarization cost for this type. Scalarization in this | |||
681 | /// context means the creation of vectors from a group of scalars. | |||
682 | int getGatherCost(Type *Ty, const DenseSet<unsigned> &ShuffledIndices); | |||
683 | ||||
684 | /// \returns the scalarization cost for this list of values. Assuming that | |||
685 | /// this subtree gets vectorized, we may need to extract the values from the | |||
686 | /// roots. This method calculates the cost of extracting the values. | |||
687 | int getGatherCost(ArrayRef<Value *> VL); | |||
688 | ||||
689 | /// \brief Set the Builder insert point to one after the last instruction in | |||
690 | /// the bundle | |||
691 | void setInsertPointAfterBundle(ArrayRef<Value *> VL, Value *OpValue); | |||
692 | ||||
693 | /// \returns a vector from a collection of scalars in \p VL. | |||
694 | Value *Gather(ArrayRef<Value *> VL, VectorType *Ty); | |||
695 | ||||
696 | /// \returns whether the VectorizableTree is fully vectorizable and will | |||
697 | /// be beneficial even the tree height is tiny. | |||
698 | bool isFullyVectorizableTinyTree(); | |||
699 | ||||
700 | /// \reorder commutative operands in alt shuffle if they result in | |||
701 | /// vectorized code. | |||
702 | void reorderAltShuffleOperands(unsigned Opcode, ArrayRef<Value *> VL, | |||
703 | SmallVectorImpl<Value *> &Left, | |||
704 | SmallVectorImpl<Value *> &Right); | |||
705 | ||||
706 | /// \reorder commutative operands to get better probability of | |||
707 | /// generating vectorized code. | |||
708 | void reorderInputsAccordingToOpcode(unsigned Opcode, ArrayRef<Value *> VL, | |||
709 | SmallVectorImpl<Value *> &Left, | |||
710 | SmallVectorImpl<Value *> &Right); | |||
711 | struct TreeEntry { | |||
712 | TreeEntry(std::vector<TreeEntry> &Container) : Container(Container) {} | |||
713 | ||||
714 | /// \returns true if the scalars in VL are equal to this entry. | |||
715 | bool isSame(ArrayRef<Value *> VL) const { | |||
716 | if (VL.size() == Scalars.size()) | |||
717 | return std::equal(VL.begin(), VL.end(), Scalars.begin()); | |||
718 | return VL.size() == ReuseShuffleIndices.size() && | |||
719 | std::equal( | |||
720 | VL.begin(), VL.end(), ReuseShuffleIndices.begin(), | |||
721 | [this](Value *V, unsigned Idx) { return V == Scalars[Idx]; }); | |||
722 | } | |||
723 | ||||
724 | /// A vector of scalars. | |||
725 | ValueList Scalars; | |||
726 | ||||
727 | /// The Scalars are vectorized into this value. It is initialized to Null. | |||
728 | Value *VectorizedValue = nullptr; | |||
729 | ||||
730 | /// Do we need to gather this sequence ? | |||
731 | bool NeedToGather = false; | |||
732 | ||||
733 | /// Does this sequence require some shuffling? | |||
734 | SmallVector<unsigned, 4> ReuseShuffleIndices; | |||
735 | ||||
736 | /// Does this entry require reordering? | |||
737 | ArrayRef<unsigned> ReorderIndices; | |||
738 | ||||
739 | /// Points back to the VectorizableTree. | |||
740 | /// | |||
741 | /// Only used for Graphviz right now. Unfortunately GraphTrait::NodeRef has | |||
742 | /// to be a pointer and needs to be able to initialize the child iterator. | |||
743 | /// Thus we need a reference back to the container to translate the indices | |||
744 | /// to entries. | |||
745 | std::vector<TreeEntry> &Container; | |||
746 | ||||
747 | /// The TreeEntry index containing the user of this entry. We can actually | |||
748 | /// have multiple users so the data structure is not truly a tree. | |||
749 | SmallVector<int, 1> UserTreeIndices; | |||
750 | }; | |||
751 | ||||
752 | /// Create a new VectorizableTree entry. | |||
753 | void newTreeEntry(ArrayRef<Value *> VL, bool Vectorized, int &UserTreeIdx, | |||
754 | ArrayRef<unsigned> ReuseShuffleIndices = None, | |||
755 | ArrayRef<unsigned> ReorderIndices = None) { | |||
756 | VectorizableTree.emplace_back(VectorizableTree); | |||
757 | int idx = VectorizableTree.size() - 1; | |||
758 | TreeEntry *Last = &VectorizableTree[idx]; | |||
759 | Last->Scalars.insert(Last->Scalars.begin(), VL.begin(), VL.end()); | |||
760 | Last->NeedToGather = !Vectorized; | |||
761 | Last->ReuseShuffleIndices.append(ReuseShuffleIndices.begin(), | |||
762 | ReuseShuffleIndices.end()); | |||
763 | Last->ReorderIndices = ReorderIndices; | |||
764 | if (Vectorized) { | |||
765 | for (int i = 0, e = VL.size(); i != e; ++i) { | |||
766 | assert(!getTreeEntry(VL[i]) && "Scalar already in tree!")(static_cast <bool> (!getTreeEntry(VL[i]) && "Scalar already in tree!" ) ? void (0) : __assert_fail ("!getTreeEntry(VL[i]) && \"Scalar already in tree!\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 766, __extension__ __PRETTY_FUNCTION__)); | |||
767 | ScalarToTreeEntry[VL[i]] = idx; | |||
768 | } | |||
769 | } else { | |||
770 | MustGather.insert(VL.begin(), VL.end()); | |||
771 | } | |||
772 | ||||
773 | if (UserTreeIdx >= 0) | |||
774 | Last->UserTreeIndices.push_back(UserTreeIdx); | |||
775 | UserTreeIdx = idx; | |||
776 | } | |||
777 | ||||
778 | /// -- Vectorization State -- | |||
779 | /// Holds all of the tree entries. | |||
780 | std::vector<TreeEntry> VectorizableTree; | |||
781 | ||||
782 | TreeEntry *getTreeEntry(Value *V) { | |||
783 | auto I = ScalarToTreeEntry.find(V); | |||
784 | if (I != ScalarToTreeEntry.end()) | |||
785 | return &VectorizableTree[I->second]; | |||
786 | return nullptr; | |||
787 | } | |||
788 | ||||
789 | /// Maps a specific scalar to its tree entry. | |||
790 | SmallDenseMap<Value*, int> ScalarToTreeEntry; | |||
791 | ||||
792 | /// A list of scalars that we found that we need to keep as scalars. | |||
793 | ValueSet MustGather; | |||
794 | ||||
795 | /// This POD struct describes one external user in the vectorized tree. | |||
796 | struct ExternalUser { | |||
797 | ExternalUser(Value *S, llvm::User *U, int L) | |||
798 | : Scalar(S), User(U), Lane(L) {} | |||
799 | ||||
800 | // Which scalar in our function. | |||
801 | Value *Scalar; | |||
802 | ||||
803 | // Which user that uses the scalar. | |||
804 | llvm::User *User; | |||
805 | ||||
806 | // Which lane does the scalar belong to. | |||
807 | int Lane; | |||
808 | }; | |||
809 | using UserList = SmallVector<ExternalUser, 16>; | |||
810 | ||||
811 | /// Checks if two instructions may access the same memory. | |||
812 | /// | |||
813 | /// \p Loc1 is the location of \p Inst1. It is passed explicitly because it | |||
814 | /// is invariant in the calling loop. | |||
815 | bool isAliased(const MemoryLocation &Loc1, Instruction *Inst1, | |||
816 | Instruction *Inst2) { | |||
817 | // First check if the result is already in the cache. | |||
818 | AliasCacheKey key = std::make_pair(Inst1, Inst2); | |||
819 | Optional<bool> &result = AliasCache[key]; | |||
820 | if (result.hasValue()) { | |||
821 | return result.getValue(); | |||
822 | } | |||
823 | MemoryLocation Loc2 = getLocation(Inst2, AA); | |||
824 | bool aliased = true; | |||
825 | if (Loc1.Ptr && Loc2.Ptr && isSimple(Inst1) && isSimple(Inst2)) { | |||
826 | // Do the alias check. | |||
827 | aliased = AA->alias(Loc1, Loc2); | |||
828 | } | |||
829 | // Store the result in the cache. | |||
830 | result = aliased; | |||
831 | return aliased; | |||
832 | } | |||
833 | ||||
834 | using AliasCacheKey = std::pair<Instruction *, Instruction *>; | |||
835 | ||||
836 | /// Cache for alias results. | |||
837 | /// TODO: consider moving this to the AliasAnalysis itself. | |||
838 | DenseMap<AliasCacheKey, Optional<bool>> AliasCache; | |||
839 | ||||
840 | /// Removes an instruction from its block and eventually deletes it. | |||
841 | /// It's like Instruction::eraseFromParent() except that the actual deletion | |||
842 | /// is delayed until BoUpSLP is destructed. | |||
843 | /// This is required to ensure that there are no incorrect collisions in the | |||
844 | /// AliasCache, which can happen if a new instruction is allocated at the | |||
845 | /// same address as a previously deleted instruction. | |||
846 | void eraseInstruction(Instruction *I) { | |||
847 | I->removeFromParent(); | |||
848 | I->dropAllReferences(); | |||
849 | DeletedInstructions.emplace_back(I); | |||
850 | } | |||
851 | ||||
852 | /// Temporary store for deleted instructions. Instructions will be deleted | |||
853 | /// eventually when the BoUpSLP is destructed. | |||
854 | SmallVector<unique_value, 8> DeletedInstructions; | |||
855 | ||||
856 | /// A list of values that need to extracted out of the tree. | |||
857 | /// This list holds pairs of (Internal Scalar : External User). External User | |||
858 | /// can be nullptr, it means that this Internal Scalar will be used later, | |||
859 | /// after vectorization. | |||
860 | UserList ExternalUses; | |||
861 | ||||
862 | /// Values used only by @llvm.assume calls. | |||
863 | SmallPtrSet<const Value *, 32> EphValues; | |||
864 | ||||
865 | /// Holds all of the instructions that we gathered. | |||
866 | SetVector<Instruction *> GatherSeq; | |||
867 | ||||
868 | /// A list of blocks that we are going to CSE. | |||
869 | SetVector<BasicBlock *> CSEBlocks; | |||
870 | ||||
871 | /// Contains all scheduling relevant data for an instruction. | |||
872 | /// A ScheduleData either represents a single instruction or a member of an | |||
873 | /// instruction bundle (= a group of instructions which is combined into a | |||
874 | /// vector instruction). | |||
875 | struct ScheduleData { | |||
876 | // The initial value for the dependency counters. It means that the | |||
877 | // dependencies are not calculated yet. | |||
878 | enum { InvalidDeps = -1 }; | |||
879 | ||||
880 | ScheduleData() = default; | |||
881 | ||||
882 | void init(int BlockSchedulingRegionID, Value *OpVal) { | |||
883 | FirstInBundle = this; | |||
884 | NextInBundle = nullptr; | |||
885 | NextLoadStore = nullptr; | |||
886 | IsScheduled = false; | |||
887 | SchedulingRegionID = BlockSchedulingRegionID; | |||
888 | UnscheduledDepsInBundle = UnscheduledDeps; | |||
889 | clearDependencies(); | |||
890 | OpValue = OpVal; | |||
891 | } | |||
892 | ||||
893 | /// Returns true if the dependency information has been calculated. | |||
894 | bool hasValidDependencies() const { return Dependencies != InvalidDeps; } | |||
895 | ||||
896 | /// Returns true for single instructions and for bundle representatives | |||
897 | /// (= the head of a bundle). | |||
898 | bool isSchedulingEntity() const { return FirstInBundle == this; } | |||
899 | ||||
900 | /// Returns true if it represents an instruction bundle and not only a | |||
901 | /// single instruction. | |||
902 | bool isPartOfBundle() const { | |||
903 | return NextInBundle != nullptr || FirstInBundle != this; | |||
904 | } | |||
905 | ||||
906 | /// Returns true if it is ready for scheduling, i.e. it has no more | |||
907 | /// unscheduled depending instructions/bundles. | |||
908 | bool isReady() const { | |||
909 | assert(isSchedulingEntity() &&(static_cast <bool> (isSchedulingEntity() && "can't consider non-scheduling entity for ready list" ) ? void (0) : __assert_fail ("isSchedulingEntity() && \"can't consider non-scheduling entity for ready list\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 910, __extension__ __PRETTY_FUNCTION__)) | |||
910 | "can't consider non-scheduling entity for ready list")(static_cast <bool> (isSchedulingEntity() && "can't consider non-scheduling entity for ready list" ) ? void (0) : __assert_fail ("isSchedulingEntity() && \"can't consider non-scheduling entity for ready list\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 910, __extension__ __PRETTY_FUNCTION__)); | |||
911 | return UnscheduledDepsInBundle == 0 && !IsScheduled; | |||
912 | } | |||
913 | ||||
914 | /// Modifies the number of unscheduled dependencies, also updating it for | |||
915 | /// the whole bundle. | |||
916 | int incrementUnscheduledDeps(int Incr) { | |||
917 | UnscheduledDeps += Incr; | |||
918 | return FirstInBundle->UnscheduledDepsInBundle += Incr; | |||
919 | } | |||
920 | ||||
921 | /// Sets the number of unscheduled dependencies to the number of | |||
922 | /// dependencies. | |||
923 | void resetUnscheduledDeps() { | |||
924 | incrementUnscheduledDeps(Dependencies - UnscheduledDeps); | |||
925 | } | |||
926 | ||||
927 | /// Clears all dependency information. | |||
928 | void clearDependencies() { | |||
929 | Dependencies = InvalidDeps; | |||
930 | resetUnscheduledDeps(); | |||
931 | MemoryDependencies.clear(); | |||
932 | } | |||
933 | ||||
934 | void dump(raw_ostream &os) const { | |||
935 | if (!isSchedulingEntity()) { | |||
936 | os << "/ " << *Inst; | |||
937 | } else if (NextInBundle) { | |||
938 | os << '[' << *Inst; | |||
939 | ScheduleData *SD = NextInBundle; | |||
940 | while (SD) { | |||
941 | os << ';' << *SD->Inst; | |||
942 | SD = SD->NextInBundle; | |||
943 | } | |||
944 | os << ']'; | |||
945 | } else { | |||
946 | os << *Inst; | |||
947 | } | |||
948 | } | |||
949 | ||||
950 | Instruction *Inst = nullptr; | |||
951 | ||||
952 | /// Points to the head in an instruction bundle (and always to this for | |||
953 | /// single instructions). | |||
954 | ScheduleData *FirstInBundle = nullptr; | |||
955 | ||||
956 | /// Single linked list of all instructions in a bundle. Null if it is a | |||
957 | /// single instruction. | |||
958 | ScheduleData *NextInBundle = nullptr; | |||
959 | ||||
960 | /// Single linked list of all memory instructions (e.g. load, store, call) | |||
961 | /// in the block - until the end of the scheduling region. | |||
962 | ScheduleData *NextLoadStore = nullptr; | |||
963 | ||||
964 | /// The dependent memory instructions. | |||
965 | /// This list is derived on demand in calculateDependencies(). | |||
966 | SmallVector<ScheduleData *, 4> MemoryDependencies; | |||
967 | ||||
968 | /// This ScheduleData is in the current scheduling region if this matches | |||
969 | /// the current SchedulingRegionID of BlockScheduling. | |||
970 | int SchedulingRegionID = 0; | |||
971 | ||||
972 | /// Used for getting a "good" final ordering of instructions. | |||
973 | int SchedulingPriority = 0; | |||
974 | ||||
975 | /// The number of dependencies. Constitutes of the number of users of the | |||
976 | /// instruction plus the number of dependent memory instructions (if any). | |||
977 | /// This value is calculated on demand. | |||
978 | /// If InvalidDeps, the number of dependencies is not calculated yet. | |||
979 | int Dependencies = InvalidDeps; | |||
980 | ||||
981 | /// The number of dependencies minus the number of dependencies of scheduled | |||
982 | /// instructions. As soon as this is zero, the instruction/bundle gets ready | |||
983 | /// for scheduling. | |||
984 | /// Note that this is negative as long as Dependencies is not calculated. | |||
985 | int UnscheduledDeps = InvalidDeps; | |||
986 | ||||
987 | /// The sum of UnscheduledDeps in a bundle. Equals to UnscheduledDeps for | |||
988 | /// single instructions. | |||
989 | int UnscheduledDepsInBundle = InvalidDeps; | |||
990 | ||||
991 | /// True if this instruction is scheduled (or considered as scheduled in the | |||
992 | /// dry-run). | |||
993 | bool IsScheduled = false; | |||
994 | ||||
995 | /// Opcode of the current instruction in the schedule data. | |||
996 | Value *OpValue = nullptr; | |||
997 | }; | |||
998 | ||||
999 | #ifndef NDEBUG | |||
1000 | friend inline raw_ostream &operator<<(raw_ostream &os, | |||
1001 | const BoUpSLP::ScheduleData &SD) { | |||
1002 | SD.dump(os); | |||
1003 | return os; | |||
1004 | } | |||
1005 | #endif | |||
1006 | ||||
1007 | friend struct GraphTraits<BoUpSLP *>; | |||
1008 | friend struct DOTGraphTraits<BoUpSLP *>; | |||
1009 | ||||
1010 | /// Contains all scheduling data for a basic block. | |||
1011 | struct BlockScheduling { | |||
1012 | BlockScheduling(BasicBlock *BB) | |||
1013 | : BB(BB), ChunkSize(BB->size()), ChunkPos(ChunkSize) {} | |||
1014 | ||||
1015 | void clear() { | |||
1016 | ReadyInsts.clear(); | |||
1017 | ScheduleStart = nullptr; | |||
1018 | ScheduleEnd = nullptr; | |||
1019 | FirstLoadStoreInRegion = nullptr; | |||
1020 | LastLoadStoreInRegion = nullptr; | |||
1021 | ||||
1022 | // Reduce the maximum schedule region size by the size of the | |||
1023 | // previous scheduling run. | |||
1024 | ScheduleRegionSizeLimit -= ScheduleRegionSize; | |||
1025 | if (ScheduleRegionSizeLimit < MinScheduleRegionSize) | |||
1026 | ScheduleRegionSizeLimit = MinScheduleRegionSize; | |||
1027 | ScheduleRegionSize = 0; | |||
1028 | ||||
1029 | // Make a new scheduling region, i.e. all existing ScheduleData is not | |||
1030 | // in the new region yet. | |||
1031 | ++SchedulingRegionID; | |||
1032 | } | |||
1033 | ||||
1034 | ScheduleData *getScheduleData(Value *V) { | |||
1035 | ScheduleData *SD = ScheduleDataMap[V]; | |||
1036 | if (SD && SD->SchedulingRegionID == SchedulingRegionID) | |||
1037 | return SD; | |||
1038 | return nullptr; | |||
1039 | } | |||
1040 | ||||
1041 | ScheduleData *getScheduleData(Value *V, Value *Key) { | |||
1042 | if (V == Key) | |||
1043 | return getScheduleData(V); | |||
1044 | auto I = ExtraScheduleDataMap.find(V); | |||
1045 | if (I != ExtraScheduleDataMap.end()) { | |||
1046 | ScheduleData *SD = I->second[Key]; | |||
1047 | if (SD && SD->SchedulingRegionID == SchedulingRegionID) | |||
1048 | return SD; | |||
1049 | } | |||
1050 | return nullptr; | |||
1051 | } | |||
1052 | ||||
1053 | bool isInSchedulingRegion(ScheduleData *SD) { | |||
1054 | return SD->SchedulingRegionID == SchedulingRegionID; | |||
1055 | } | |||
1056 | ||||
1057 | /// Marks an instruction as scheduled and puts all dependent ready | |||
1058 | /// instructions into the ready-list. | |||
1059 | template <typename ReadyListType> | |||
1060 | void schedule(ScheduleData *SD, ReadyListType &ReadyList) { | |||
1061 | SD->IsScheduled = true; | |||
1062 | DEBUG(dbgs() << "SLP: schedule " << *SD << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: schedule " << *SD << "\n"; } } while (false); | |||
1063 | ||||
1064 | ScheduleData *BundleMember = SD; | |||
1065 | while (BundleMember) { | |||
1066 | if (BundleMember->Inst != BundleMember->OpValue) { | |||
1067 | BundleMember = BundleMember->NextInBundle; | |||
1068 | continue; | |||
1069 | } | |||
1070 | // Handle the def-use chain dependencies. | |||
1071 | for (Use &U : BundleMember->Inst->operands()) { | |||
1072 | auto *I = dyn_cast<Instruction>(U.get()); | |||
1073 | if (!I) | |||
1074 | continue; | |||
1075 | doForAllOpcodes(I, [&ReadyList](ScheduleData *OpDef) { | |||
1076 | if (OpDef && OpDef->hasValidDependencies() && | |||
1077 | OpDef->incrementUnscheduledDeps(-1) == 0) { | |||
1078 | // There are no more unscheduled dependencies after | |||
1079 | // decrementing, so we can put the dependent instruction | |||
1080 | // into the ready list. | |||
1081 | ScheduleData *DepBundle = OpDef->FirstInBundle; | |||
1082 | assert(!DepBundle->IsScheduled &&(static_cast <bool> (!DepBundle->IsScheduled && "already scheduled bundle gets ready") ? void (0) : __assert_fail ("!DepBundle->IsScheduled && \"already scheduled bundle gets ready\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1083, __extension__ __PRETTY_FUNCTION__)) | |||
1083 | "already scheduled bundle gets ready")(static_cast <bool> (!DepBundle->IsScheduled && "already scheduled bundle gets ready") ? void (0) : __assert_fail ("!DepBundle->IsScheduled && \"already scheduled bundle gets ready\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1083, __extension__ __PRETTY_FUNCTION__)); | |||
1084 | ReadyList.insert(DepBundle); | |||
1085 | DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: gets ready (def): " << *DepBundle << "\n"; } } while (false) | |||
1086 | << "SLP: gets ready (def): " << *DepBundle << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: gets ready (def): " << *DepBundle << "\n"; } } while (false); | |||
1087 | } | |||
1088 | }); | |||
1089 | } | |||
1090 | // Handle the memory dependencies. | |||
1091 | for (ScheduleData *MemoryDepSD : BundleMember->MemoryDependencies) { | |||
1092 | if (MemoryDepSD->incrementUnscheduledDeps(-1) == 0) { | |||
1093 | // There are no more unscheduled dependencies after decrementing, | |||
1094 | // so we can put the dependent instruction into the ready list. | |||
1095 | ScheduleData *DepBundle = MemoryDepSD->FirstInBundle; | |||
1096 | assert(!DepBundle->IsScheduled &&(static_cast <bool> (!DepBundle->IsScheduled && "already scheduled bundle gets ready") ? void (0) : __assert_fail ("!DepBundle->IsScheduled && \"already scheduled bundle gets ready\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1097, __extension__ __PRETTY_FUNCTION__)) | |||
1097 | "already scheduled bundle gets ready")(static_cast <bool> (!DepBundle->IsScheduled && "already scheduled bundle gets ready") ? void (0) : __assert_fail ("!DepBundle->IsScheduled && \"already scheduled bundle gets ready\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1097, __extension__ __PRETTY_FUNCTION__)); | |||
1098 | ReadyList.insert(DepBundle); | |||
1099 | DEBUG(dbgs() << "SLP: gets ready (mem): " << *DepBundledo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: gets ready (mem): " << *DepBundle << "\n"; } } while (false) | |||
1100 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: gets ready (mem): " << *DepBundle << "\n"; } } while (false); | |||
1101 | } | |||
1102 | } | |||
1103 | BundleMember = BundleMember->NextInBundle; | |||
1104 | } | |||
1105 | } | |||
1106 | ||||
1107 | void doForAllOpcodes(Value *V, | |||
1108 | function_ref<void(ScheduleData *SD)> Action) { | |||
1109 | if (ScheduleData *SD = getScheduleData(V)) | |||
1110 | Action(SD); | |||
1111 | auto I = ExtraScheduleDataMap.find(V); | |||
1112 | if (I != ExtraScheduleDataMap.end()) | |||
1113 | for (auto &P : I->second) | |||
1114 | if (P.second->SchedulingRegionID == SchedulingRegionID) | |||
1115 | Action(P.second); | |||
1116 | } | |||
1117 | ||||
1118 | /// Put all instructions into the ReadyList which are ready for scheduling. | |||
1119 | template <typename ReadyListType> | |||
1120 | void initialFillReadyList(ReadyListType &ReadyList) { | |||
1121 | for (auto *I = ScheduleStart; I != ScheduleEnd; I = I->getNextNode()) { | |||
1122 | doForAllOpcodes(I, [&](ScheduleData *SD) { | |||
1123 | if (SD->isSchedulingEntity() && SD->isReady()) { | |||
1124 | ReadyList.insert(SD); | |||
1125 | DEBUG(dbgs() << "SLP: initially in ready list: " << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: initially in ready list: " << *I << "\n"; } } while (false); | |||
1126 | } | |||
1127 | }); | |||
1128 | } | |||
1129 | } | |||
1130 | ||||
1131 | /// Checks if a bundle of instructions can be scheduled, i.e. has no | |||
1132 | /// cyclic dependencies. This is only a dry-run, no instructions are | |||
1133 | /// actually moved at this stage. | |||
1134 | bool tryScheduleBundle(ArrayRef<Value *> VL, BoUpSLP *SLP, Value *OpValue); | |||
1135 | ||||
1136 | /// Un-bundles a group of instructions. | |||
1137 | void cancelScheduling(ArrayRef<Value *> VL, Value *OpValue); | |||
1138 | ||||
1139 | /// Allocates schedule data chunk. | |||
1140 | ScheduleData *allocateScheduleDataChunks(); | |||
1141 | ||||
1142 | /// Extends the scheduling region so that V is inside the region. | |||
1143 | /// \returns true if the region size is within the limit. | |||
1144 | bool extendSchedulingRegion(Value *V, Value *OpValue); | |||
1145 | ||||
1146 | /// Initialize the ScheduleData structures for new instructions in the | |||
1147 | /// scheduling region. | |||
1148 | void initScheduleData(Instruction *FromI, Instruction *ToI, | |||
1149 | ScheduleData *PrevLoadStore, | |||
1150 | ScheduleData *NextLoadStore); | |||
1151 | ||||
1152 | /// Updates the dependency information of a bundle and of all instructions/ | |||
1153 | /// bundles which depend on the original bundle. | |||
1154 | void calculateDependencies(ScheduleData *SD, bool InsertInReadyList, | |||
1155 | BoUpSLP *SLP); | |||
1156 | ||||
1157 | /// Sets all instruction in the scheduling region to un-scheduled. | |||
1158 | void resetSchedule(); | |||
1159 | ||||
1160 | BasicBlock *BB; | |||
1161 | ||||
1162 | /// Simple memory allocation for ScheduleData. | |||
1163 | std::vector<std::unique_ptr<ScheduleData[]>> ScheduleDataChunks; | |||
1164 | ||||
1165 | /// The size of a ScheduleData array in ScheduleDataChunks. | |||
1166 | int ChunkSize; | |||
1167 | ||||
1168 | /// The allocator position in the current chunk, which is the last entry | |||
1169 | /// of ScheduleDataChunks. | |||
1170 | int ChunkPos; | |||
1171 | ||||
1172 | /// Attaches ScheduleData to Instruction. | |||
1173 | /// Note that the mapping survives during all vectorization iterations, i.e. | |||
1174 | /// ScheduleData structures are recycled. | |||
1175 | DenseMap<Value *, ScheduleData *> ScheduleDataMap; | |||
1176 | ||||
1177 | /// Attaches ScheduleData to Instruction with the leading key. | |||
1178 | DenseMap<Value *, SmallDenseMap<Value *, ScheduleData *>> | |||
1179 | ExtraScheduleDataMap; | |||
1180 | ||||
1181 | struct ReadyList : SmallVector<ScheduleData *, 8> { | |||
1182 | void insert(ScheduleData *SD) { push_back(SD); } | |||
1183 | }; | |||
1184 | ||||
1185 | /// The ready-list for scheduling (only used for the dry-run). | |||
1186 | ReadyList ReadyInsts; | |||
1187 | ||||
1188 | /// The first instruction of the scheduling region. | |||
1189 | Instruction *ScheduleStart = nullptr; | |||
1190 | ||||
1191 | /// The first instruction _after_ the scheduling region. | |||
1192 | Instruction *ScheduleEnd = nullptr; | |||
1193 | ||||
1194 | /// The first memory accessing instruction in the scheduling region | |||
1195 | /// (can be null). | |||
1196 | ScheduleData *FirstLoadStoreInRegion = nullptr; | |||
1197 | ||||
1198 | /// The last memory accessing instruction in the scheduling region | |||
1199 | /// (can be null). | |||
1200 | ScheduleData *LastLoadStoreInRegion = nullptr; | |||
1201 | ||||
1202 | /// The current size of the scheduling region. | |||
1203 | int ScheduleRegionSize = 0; | |||
1204 | ||||
1205 | /// The maximum size allowed for the scheduling region. | |||
1206 | int ScheduleRegionSizeLimit = ScheduleRegionSizeBudget; | |||
1207 | ||||
1208 | /// The ID of the scheduling region. For a new vectorization iteration this | |||
1209 | /// is incremented which "removes" all ScheduleData from the region. | |||
1210 | // Make sure that the initial SchedulingRegionID is greater than the | |||
1211 | // initial SchedulingRegionID in ScheduleData (which is 0). | |||
1212 | int SchedulingRegionID = 1; | |||
1213 | }; | |||
1214 | ||||
1215 | /// Attaches the BlockScheduling structures to basic blocks. | |||
1216 | MapVector<BasicBlock *, std::unique_ptr<BlockScheduling>> BlocksSchedules; | |||
1217 | ||||
1218 | /// Performs the "real" scheduling. Done before vectorization is actually | |||
1219 | /// performed in a basic block. | |||
1220 | void scheduleBlock(BlockScheduling *BS); | |||
1221 | ||||
1222 | /// List of users to ignore during scheduling and that don't need extracting. | |||
1223 | ArrayRef<Value *> UserIgnoreList; | |||
1224 | ||||
1225 | using OrdersType = SmallVector<unsigned, 4>; | |||
1226 | /// A DenseMapInfo implementation for holding DenseMaps and DenseSets of | |||
1227 | /// sorted SmallVectors of unsigned. | |||
1228 | struct OrdersTypeDenseMapInfo { | |||
1229 | static OrdersType getEmptyKey() { | |||
1230 | OrdersType V; | |||
1231 | V.push_back(~1U); | |||
1232 | return V; | |||
1233 | } | |||
1234 | ||||
1235 | static OrdersType getTombstoneKey() { | |||
1236 | OrdersType V; | |||
1237 | V.push_back(~2U); | |||
1238 | return V; | |||
1239 | } | |||
1240 | ||||
1241 | static unsigned getHashValue(const OrdersType &V) { | |||
1242 | return static_cast<unsigned>(hash_combine_range(V.begin(), V.end())); | |||
1243 | } | |||
1244 | ||||
1245 | static bool isEqual(const OrdersType &LHS, const OrdersType &RHS) { | |||
1246 | return LHS == RHS; | |||
1247 | } | |||
1248 | }; | |||
1249 | ||||
1250 | /// Contains orders of operations along with the number of bundles that have | |||
1251 | /// operations in this order. It stores only those orders that require | |||
1252 | /// reordering, if reordering is not required it is counted using \a | |||
1253 | /// NumOpsWantToKeepOriginalOrder. | |||
1254 | DenseMap<OrdersType, unsigned, OrdersTypeDenseMapInfo> NumOpsWantToKeepOrder; | |||
1255 | /// Number of bundles that do not require reordering. | |||
1256 | unsigned NumOpsWantToKeepOriginalOrder = 0; | |||
1257 | ||||
1258 | // Analysis and block reference. | |||
1259 | Function *F; | |||
1260 | ScalarEvolution *SE; | |||
1261 | TargetTransformInfo *TTI; | |||
1262 | TargetLibraryInfo *TLI; | |||
1263 | AliasAnalysis *AA; | |||
1264 | LoopInfo *LI; | |||
1265 | DominatorTree *DT; | |||
1266 | AssumptionCache *AC; | |||
1267 | DemandedBits *DB; | |||
1268 | const DataLayout *DL; | |||
1269 | OptimizationRemarkEmitter *ORE; | |||
1270 | ||||
1271 | unsigned MaxVecRegSize; // This is set by TTI or overridden by cl::opt. | |||
1272 | unsigned MinVecRegSize; // Set by cl::opt (default: 128). | |||
1273 | ||||
1274 | /// Instruction builder to construct the vectorized tree. | |||
1275 | IRBuilder<> Builder; | |||
1276 | ||||
1277 | /// A map of scalar integer values to the smallest bit width with which they | |||
1278 | /// can legally be represented. The values map to (width, signed) pairs, | |||
1279 | /// where "width" indicates the minimum bit width and "signed" is True if the | |||
1280 | /// value must be signed-extended, rather than zero-extended, back to its | |||
1281 | /// original width. | |||
1282 | MapVector<Value *, std::pair<uint64_t, bool>> MinBWs; | |||
1283 | }; | |||
1284 | ||||
1285 | } // end namespace slpvectorizer | |||
1286 | ||||
1287 | template <> struct GraphTraits<BoUpSLP *> { | |||
1288 | using TreeEntry = BoUpSLP::TreeEntry; | |||
1289 | ||||
1290 | /// NodeRef has to be a pointer per the GraphWriter. | |||
1291 | using NodeRef = TreeEntry *; | |||
1292 | ||||
1293 | /// \brief Add the VectorizableTree to the index iterator to be able to return | |||
1294 | /// TreeEntry pointers. | |||
1295 | struct ChildIteratorType | |||
1296 | : public iterator_adaptor_base<ChildIteratorType, | |||
1297 | SmallVector<int, 1>::iterator> { | |||
1298 | std::vector<TreeEntry> &VectorizableTree; | |||
1299 | ||||
1300 | ChildIteratorType(SmallVector<int, 1>::iterator W, | |||
1301 | std::vector<TreeEntry> &VT) | |||
1302 | : ChildIteratorType::iterator_adaptor_base(W), VectorizableTree(VT) {} | |||
1303 | ||||
1304 | NodeRef operator*() { return &VectorizableTree[*I]; } | |||
1305 | }; | |||
1306 | ||||
1307 | static NodeRef getEntryNode(BoUpSLP &R) { return &R.VectorizableTree[0]; } | |||
1308 | ||||
1309 | static ChildIteratorType child_begin(NodeRef N) { | |||
1310 | return {N->UserTreeIndices.begin(), N->Container}; | |||
1311 | } | |||
1312 | ||||
1313 | static ChildIteratorType child_end(NodeRef N) { | |||
1314 | return {N->UserTreeIndices.end(), N->Container}; | |||
1315 | } | |||
1316 | ||||
1317 | /// For the node iterator we just need to turn the TreeEntry iterator into a | |||
1318 | /// TreeEntry* iterator so that it dereferences to NodeRef. | |||
1319 | using nodes_iterator = pointer_iterator<std::vector<TreeEntry>::iterator>; | |||
1320 | ||||
1321 | static nodes_iterator nodes_begin(BoUpSLP *R) { | |||
1322 | return nodes_iterator(R->VectorizableTree.begin()); | |||
1323 | } | |||
1324 | ||||
1325 | static nodes_iterator nodes_end(BoUpSLP *R) { | |||
1326 | return nodes_iterator(R->VectorizableTree.end()); | |||
1327 | } | |||
1328 | ||||
1329 | static unsigned size(BoUpSLP *R) { return R->VectorizableTree.size(); } | |||
1330 | }; | |||
1331 | ||||
1332 | template <> struct DOTGraphTraits<BoUpSLP *> : public DefaultDOTGraphTraits { | |||
1333 | using TreeEntry = BoUpSLP::TreeEntry; | |||
1334 | ||||
1335 | DOTGraphTraits(bool isSimple = false) : DefaultDOTGraphTraits(isSimple) {} | |||
1336 | ||||
1337 | std::string getNodeLabel(const TreeEntry *Entry, const BoUpSLP *R) { | |||
1338 | std::string Str; | |||
1339 | raw_string_ostream OS(Str); | |||
1340 | if (isSplat(Entry->Scalars)) { | |||
1341 | OS << "<splat> " << *Entry->Scalars[0]; | |||
1342 | return Str; | |||
1343 | } | |||
1344 | for (auto V : Entry->Scalars) { | |||
1345 | OS << *V; | |||
1346 | if (std::any_of( | |||
1347 | R->ExternalUses.begin(), R->ExternalUses.end(), | |||
1348 | [&](const BoUpSLP::ExternalUser &EU) { return EU.Scalar == V; })) | |||
1349 | OS << " <extract>"; | |||
1350 | OS << "\n"; | |||
1351 | } | |||
1352 | return Str; | |||
1353 | } | |||
1354 | ||||
1355 | static std::string getNodeAttributes(const TreeEntry *Entry, | |||
1356 | const BoUpSLP *) { | |||
1357 | if (Entry->NeedToGather) | |||
1358 | return "color=red"; | |||
1359 | return ""; | |||
1360 | } | |||
1361 | }; | |||
1362 | ||||
1363 | } // end namespace llvm | |||
1364 | ||||
1365 | void BoUpSLP::buildTree(ArrayRef<Value *> Roots, | |||
1366 | ArrayRef<Value *> UserIgnoreLst) { | |||
1367 | ExtraValueToDebugLocsMap ExternallyUsedValues; | |||
1368 | buildTree(Roots, ExternallyUsedValues, UserIgnoreLst); | |||
1369 | } | |||
1370 | ||||
1371 | void BoUpSLP::buildTree(ArrayRef<Value *> Roots, | |||
1372 | ExtraValueToDebugLocsMap &ExternallyUsedValues, | |||
1373 | ArrayRef<Value *> UserIgnoreLst) { | |||
1374 | deleteTree(); | |||
1375 | UserIgnoreList = UserIgnoreLst; | |||
1376 | if (!allSameType(Roots)) | |||
1377 | return; | |||
1378 | buildTree_rec(Roots, 0, -1); | |||
1379 | ||||
1380 | // Collect the values that we need to extract from the tree. | |||
1381 | for (TreeEntry &EIdx : VectorizableTree) { | |||
1382 | TreeEntry *Entry = &EIdx; | |||
1383 | ||||
1384 | // No need to handle users of gathered values. | |||
1385 | if (Entry->NeedToGather) | |||
1386 | continue; | |||
1387 | ||||
1388 | // For each lane: | |||
1389 | for (int Lane = 0, LE = Entry->Scalars.size(); Lane != LE; ++Lane) { | |||
1390 | Value *Scalar = Entry->Scalars[Lane]; | |||
1391 | int FoundLane = Lane; | |||
1392 | if (!Entry->ReuseShuffleIndices.empty()) { | |||
1393 | FoundLane = | |||
1394 | std::distance(Entry->ReuseShuffleIndices.begin(), | |||
1395 | llvm::find(Entry->ReuseShuffleIndices, FoundLane)); | |||
1396 | } | |||
1397 | ||||
1398 | // Check if the scalar is externally used as an extra arg. | |||
1399 | auto ExtI = ExternallyUsedValues.find(Scalar); | |||
1400 | if (ExtI != ExternallyUsedValues.end()) { | |||
1401 | 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) | |||
1402 | 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); | |||
1403 | ExternalUses.emplace_back(Scalar, nullptr, FoundLane); | |||
1404 | } | |||
1405 | for (User *U : Scalar->users()) { | |||
1406 | DEBUG(dbgs() << "SLP: Checking user:" << *U << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Checking user:" << *U << ".\n"; } } while (false); | |||
1407 | ||||
1408 | Instruction *UserInst = dyn_cast<Instruction>(U); | |||
1409 | if (!UserInst) | |||
1410 | continue; | |||
1411 | ||||
1412 | // Skip in-tree scalars that become vectors | |||
1413 | if (TreeEntry *UseEntry = getTreeEntry(U)) { | |||
1414 | Value *UseScalar = UseEntry->Scalars[0]; | |||
1415 | // Some in-tree scalars will remain as scalar in vectorized | |||
1416 | // instructions. If that is the case, the one in Lane 0 will | |||
1417 | // be used. | |||
1418 | if (UseScalar != U || | |||
1419 | !InTreeUserNeedToExtract(Scalar, UserInst, TLI)) { | |||
1420 | 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) | |||
1421 | << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: \tInternal user will be removed:" << *U << ".\n"; } } while (false); | |||
1422 | assert(!UseEntry->NeedToGather && "Bad state")(static_cast <bool> (!UseEntry->NeedToGather && "Bad state") ? void (0) : __assert_fail ("!UseEntry->NeedToGather && \"Bad state\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1422, __extension__ __PRETTY_FUNCTION__)); | |||
1423 | continue; | |||
1424 | } | |||
1425 | } | |||
1426 | ||||
1427 | // Ignore users in the user ignore list. | |||
1428 | if (is_contained(UserIgnoreList, UserInst)) | |||
1429 | continue; | |||
1430 | ||||
1431 | 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) | |||
1432 | 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); | |||
1433 | ExternalUses.push_back(ExternalUser(Scalar, U, FoundLane)); | |||
1434 | } | |||
1435 | } | |||
1436 | } | |||
1437 | } | |||
1438 | ||||
1439 | void BoUpSLP::buildTree_rec(ArrayRef<Value *> VL, unsigned Depth, | |||
1440 | int UserTreeIdx) { | |||
1441 | assert((allConstant(VL) || allSameType(VL)) && "Invalid types!")(static_cast <bool> ((allConstant(VL) || allSameType(VL )) && "Invalid types!") ? void (0) : __assert_fail ("(allConstant(VL) || allSameType(VL)) && \"Invalid types!\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1441, __extension__ __PRETTY_FUNCTION__)); | |||
1442 | ||||
1443 | InstructionsState S = getSameOpcode(VL); | |||
1444 | if (Depth == RecursionMaxDepth) { | |||
1445 | 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); | |||
1446 | newTreeEntry(VL, false, UserTreeIdx); | |||
1447 | return; | |||
1448 | } | |||
1449 | ||||
1450 | // Don't handle vectors. | |||
1451 | if (S.OpValue->getType()->isVectorTy()) { | |||
1452 | 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); | |||
1453 | newTreeEntry(VL, false, UserTreeIdx); | |||
1454 | return; | |||
1455 | } | |||
1456 | ||||
1457 | if (StoreInst *SI = dyn_cast<StoreInst>(S.OpValue)) | |||
1458 | if (SI->getValueOperand()->getType()->isVectorTy()) { | |||
1459 | 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); | |||
1460 | newTreeEntry(VL, false, UserTreeIdx); | |||
1461 | return; | |||
1462 | } | |||
1463 | ||||
1464 | // If all of the operands are identical or constant we have a simple solution. | |||
1465 | if (allConstant(VL) || isSplat(VL) || !allSameBlock(VL) || !S.Opcode) { | |||
1466 | 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); | |||
1467 | newTreeEntry(VL, false, UserTreeIdx); | |||
1468 | return; | |||
1469 | } | |||
1470 | ||||
1471 | // We now know that this is a vector of instructions of the same type from | |||
1472 | // the same block. | |||
1473 | ||||
1474 | // Don't vectorize ephemeral values. | |||
1475 | for (unsigned i = 0, e = VL.size(); i != e; ++i) { | |||
1476 | if (EphValues.count(VL[i])) { | |||
1477 | DEBUG(dbgs() << "SLP: The instruction (" << *VL[i] <<do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: The instruction (" << * VL[i] << ") is ephemeral.\n"; } } while (false) | |||
1478 | ") is ephemeral.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: The instruction (" << * VL[i] << ") is ephemeral.\n"; } } while (false); | |||
1479 | newTreeEntry(VL, false, UserTreeIdx); | |||
1480 | return; | |||
1481 | } | |||
1482 | } | |||
1483 | ||||
1484 | // Check if this is a duplicate of another entry. | |||
1485 | if (TreeEntry *E = getTreeEntry(S.OpValue)) { | |||
1486 | DEBUG(dbgs() << "SLP: \tChecking bundle: " << *S.OpValue << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: \tChecking bundle: " << *S.OpValue << ".\n"; } } while (false); | |||
1487 | if (!E->isSame(VL)) { | |||
1488 | 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); | |||
1489 | newTreeEntry(VL, false, UserTreeIdx); | |||
1490 | return; | |||
1491 | } | |||
1492 | // Record the reuse of the tree node. FIXME, currently this is only used to | |||
1493 | // properly draw the graph rather than for the actual vectorization. | |||
1494 | E->UserTreeIndices.push_back(UserTreeIdx); | |||
1495 | DEBUG(dbgs() << "SLP: Perfect diamond merge at " << *S.OpValue << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Perfect diamond merge at " << *S.OpValue << ".\n"; } } while (false); | |||
1496 | return; | |||
1497 | } | |||
1498 | ||||
1499 | // Check that none of the instructions in the bundle are already in the tree. | |||
1500 | for (unsigned i = 0, e = VL.size(); i != e; ++i) { | |||
1501 | auto *I = dyn_cast<Instruction>(VL[i]); | |||
1502 | if (!I) | |||
1503 | continue; | |||
1504 | if (getTreeEntry(I)) { | |||
1505 | DEBUG(dbgs() << "SLP: The instruction (" << *VL[i] <<do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: The instruction (" << * VL[i] << ") is already in tree.\n"; } } while (false) | |||
1506 | ") is already in tree.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: The instruction (" << * VL[i] << ") is already in tree.\n"; } } while (false); | |||
1507 | newTreeEntry(VL, false, UserTreeIdx); | |||
1508 | return; | |||
1509 | } | |||
1510 | } | |||
1511 | ||||
1512 | // If any of the scalars is marked as a value that needs to stay scalar, then | |||
1513 | // we need to gather the scalars. | |||
1514 | for (unsigned i = 0, e = VL.size(); i != e; ++i) { | |||
1515 | if (MustGather.count(VL[i])) { | |||
1516 | 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); | |||
1517 | newTreeEntry(VL, false, UserTreeIdx); | |||
1518 | return; | |||
1519 | } | |||
1520 | } | |||
1521 | ||||
1522 | // Check that all of the users of the scalars that we want to vectorize are | |||
1523 | // schedulable. | |||
1524 | auto *VL0 = cast<Instruction>(S.OpValue); | |||
1525 | BasicBlock *BB = VL0->getParent(); | |||
1526 | ||||
1527 | if (!DT->isReachableFromEntry(BB)) { | |||
1528 | // Don't go into unreachable blocks. They may contain instructions with | |||
1529 | // dependency cycles which confuse the final scheduling. | |||
1530 | DEBUG(dbgs() << "SLP: bundle in unreachable block.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: bundle in unreachable block.\n" ; } } while (false); | |||
1531 | newTreeEntry(VL, false, UserTreeIdx); | |||
1532 | return; | |||
1533 | } | |||
1534 | ||||
1535 | // Check that every instruction appears once in this bundle. | |||
1536 | SmallVector<unsigned, 4> ReuseShuffleIndicies; | |||
1537 | SmallVector<Value *, 4> UniqueValues; | |||
1538 | DenseMap<Value *, unsigned> UniquePositions; | |||
1539 | for (Value *V : VL) { | |||
1540 | auto Res = UniquePositions.try_emplace(V, UniqueValues.size()); | |||
1541 | ReuseShuffleIndicies.emplace_back(Res.first->second); | |||
1542 | if (Res.second) | |||
1543 | UniqueValues.emplace_back(V); | |||
1544 | } | |||
1545 | if (UniqueValues.size() == VL.size()) { | |||
1546 | ReuseShuffleIndicies.clear(); | |||
1547 | } else { | |||
1548 | DEBUG(dbgs() << "SLP: Shuffle for reused scalars.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Shuffle for reused scalars.\n" ; } } while (false); | |||
1549 | if (UniqueValues.size() <= 1 || !llvm::isPowerOf2_32(UniqueValues.size())) { | |||
1550 | 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); | |||
1551 | newTreeEntry(VL, false, UserTreeIdx); | |||
1552 | return; | |||
1553 | } | |||
1554 | VL = UniqueValues; | |||
1555 | } | |||
1556 | ||||
1557 | auto &BSRef = BlocksSchedules[BB]; | |||
1558 | if (!BSRef) | |||
1559 | BSRef = llvm::make_unique<BlockScheduling>(BB); | |||
1560 | ||||
1561 | BlockScheduling &BS = *BSRef.get(); | |||
1562 | ||||
1563 | if (!BS.tryScheduleBundle(VL, this, VL0)) { | |||
1564 | 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); | |||
1565 | assert((!BS.getScheduleData(VL0) ||(static_cast <bool> ((!BS.getScheduleData(VL0) || !BS.getScheduleData (VL0)->isPartOfBundle()) && "tryScheduleBundle should cancelScheduling on failure" ) ? void (0) : __assert_fail ("(!BS.getScheduleData(VL0) || !BS.getScheduleData(VL0)->isPartOfBundle()) && \"tryScheduleBundle should cancelScheduling on failure\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1567, __extension__ __PRETTY_FUNCTION__)) | |||
1566 | !BS.getScheduleData(VL0)->isPartOfBundle()) &&(static_cast <bool> ((!BS.getScheduleData(VL0) || !BS.getScheduleData (VL0)->isPartOfBundle()) && "tryScheduleBundle should cancelScheduling on failure" ) ? void (0) : __assert_fail ("(!BS.getScheduleData(VL0) || !BS.getScheduleData(VL0)->isPartOfBundle()) && \"tryScheduleBundle should cancelScheduling on failure\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1567, __extension__ __PRETTY_FUNCTION__)) | |||
1567 | "tryScheduleBundle should cancelScheduling on failure")(static_cast <bool> ((!BS.getScheduleData(VL0) || !BS.getScheduleData (VL0)->isPartOfBundle()) && "tryScheduleBundle should cancelScheduling on failure" ) ? void (0) : __assert_fail ("(!BS.getScheduleData(VL0) || !BS.getScheduleData(VL0)->isPartOfBundle()) && \"tryScheduleBundle should cancelScheduling on failure\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1567, __extension__ __PRETTY_FUNCTION__)); | |||
1568 | newTreeEntry(VL, false, UserTreeIdx, ReuseShuffleIndicies); | |||
1569 | return; | |||
1570 | } | |||
1571 | 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); | |||
1572 | ||||
1573 | unsigned ShuffleOrOp = S.IsAltShuffle ? | |||
1574 | (unsigned) Instruction::ShuffleVector : S.Opcode; | |||
1575 | switch (ShuffleOrOp) { | |||
1576 | case Instruction::PHI: { | |||
1577 | PHINode *PH = dyn_cast<PHINode>(VL0); | |||
1578 | ||||
1579 | // Check for terminator values (e.g. invoke). | |||
1580 | for (unsigned j = 0; j < VL.size(); ++j) | |||
1581 | for (unsigned i = 0, e = PH->getNumIncomingValues(); i < e; ++i) { | |||
1582 | TerminatorInst *Term = dyn_cast<TerminatorInst>( | |||
1583 | cast<PHINode>(VL[j])->getIncomingValueForBlock(PH->getIncomingBlock(i))); | |||
1584 | if (Term) { | |||
1585 | DEBUG(dbgs() << "SLP: Need to swizzle PHINodes (TerminatorInst use).\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Need to swizzle PHINodes (TerminatorInst use).\n" ; } } while (false); | |||
1586 | BS.cancelScheduling(VL, VL0); | |||
1587 | newTreeEntry(VL, false, UserTreeIdx, ReuseShuffleIndicies); | |||
1588 | return; | |||
1589 | } | |||
1590 | } | |||
1591 | ||||
1592 | newTreeEntry(VL, true, UserTreeIdx, ReuseShuffleIndicies); | |||
1593 | 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); | |||
1594 | ||||
1595 | for (unsigned i = 0, e = PH->getNumIncomingValues(); i < e; ++i) { | |||
1596 | ValueList Operands; | |||
1597 | // Prepare the operand vector. | |||
1598 | for (Value *j : VL) | |||
1599 | Operands.push_back(cast<PHINode>(j)->getIncomingValueForBlock( | |||
1600 | PH->getIncomingBlock(i))); | |||
1601 | ||||
1602 | buildTree_rec(Operands, Depth + 1, UserTreeIdx); | |||
1603 | } | |||
1604 | return; | |||
1605 | } | |||
1606 | case Instruction::ExtractValue: | |||
1607 | case Instruction::ExtractElement: { | |||
1608 | OrdersType CurrentOrder; | |||
1609 | bool Reuse = canReuseExtract(VL, VL0, CurrentOrder); | |||
1610 | if (Reuse) { | |||
1611 | 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); | |||
1612 | ++NumOpsWantToKeepOriginalOrder; | |||
1613 | newTreeEntry(VL, /*Vectorized=*/true, UserTreeIdx, | |||
1614 | ReuseShuffleIndicies); | |||
1615 | return; | |||
1616 | } | |||
1617 | if (!CurrentOrder.empty()) { | |||
1618 | 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) | |||
1619 | 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) | |||
1620 | "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) | |||
1621 | 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) | |||
1622 | 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) | |||
1623 | 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) | |||
1624 | })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); | |||
1625 | // Insert new order with initial value 0, if it does not exist, | |||
1626 | // otherwise return the iterator to the existing one. | |||
1627 | auto StoredCurrentOrderAndNum = | |||
1628 | NumOpsWantToKeepOrder.try_emplace(CurrentOrder).first; | |||
1629 | ++StoredCurrentOrderAndNum->getSecond(); | |||
1630 | newTreeEntry(VL, /*Vectorized=*/true, UserTreeIdx, ReuseShuffleIndicies, | |||
1631 | StoredCurrentOrderAndNum->getFirst()); | |||
1632 | return; | |||
1633 | } | |||
1634 | DEBUG(dbgs() << "SLP: Gather extract sequence.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gather extract sequence.\n"; } } while (false); | |||
1635 | newTreeEntry(VL, /*Vectorized=*/false, UserTreeIdx, ReuseShuffleIndicies); | |||
1636 | BS.cancelScheduling(VL, VL0); | |||
1637 | return; | |||
1638 | } | |||
1639 | case Instruction::Load: { | |||
1640 | // Check that a vectorized load would load the same memory as a scalar | |||
1641 | // load. For example, we don't want to vectorize loads that are smaller | |||
1642 | // than 8-bit. Even though we have a packed struct {<i2, i2, i2, i2>} LLVM | |||
1643 | // treats loading/storing it as an i8 struct. If we vectorize loads/stores | |||
1644 | // from such a struct, we read/write packed bits disagreeing with the | |||
1645 | // unvectorized version. | |||
1646 | Type *ScalarTy = VL0->getType(); | |||
1647 | ||||
1648 | if (DL->getTypeSizeInBits(ScalarTy) != | |||
1649 | DL->getTypeAllocSizeInBits(ScalarTy)) { | |||
1650 | BS.cancelScheduling(VL, VL0); | |||
1651 | newTreeEntry(VL, false, UserTreeIdx, ReuseShuffleIndicies); | |||
1652 | 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); | |||
1653 | return; | |||
1654 | } | |||
1655 | ||||
1656 | // Make sure all loads in the bundle are simple - we can't vectorize | |||
1657 | // atomic or volatile loads. | |||
1658 | SmallVector<Value *, 4> PointerOps(VL.size()); | |||
1659 | auto POIter = PointerOps.begin(); | |||
1660 | for (Value *V : VL) { | |||
1661 | auto *L = cast<LoadInst>(V); | |||
1662 | if (!L->isSimple()) { | |||
1663 | BS.cancelScheduling(VL, VL0); | |||
1664 | newTreeEntry(VL, false, UserTreeIdx, ReuseShuffleIndicies); | |||
1665 | DEBUG(dbgs() << "SLP: Gathering non-simple loads.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering non-simple loads.\n" ; } } while (false); | |||
1666 | return; | |||
1667 | } | |||
1668 | *POIter = L->getPointerOperand(); | |||
1669 | ++POIter; | |||
1670 | } | |||
1671 | ||||
1672 | OrdersType CurrentOrder; | |||
1673 | // Check the order of pointer operands. | |||
1674 | if (llvm::sortPtrAccesses(PointerOps, *DL, *SE, CurrentOrder)) { | |||
1675 | Value *Ptr0; | |||
1676 | Value *PtrN; | |||
1677 | if (CurrentOrder.empty()) { | |||
1678 | Ptr0 = PointerOps.front(); | |||
1679 | PtrN = PointerOps.back(); | |||
1680 | } else { | |||
1681 | Ptr0 = PointerOps[CurrentOrder.front()]; | |||
1682 | PtrN = PointerOps[CurrentOrder.back()]; | |||
1683 | } | |||
1684 | const SCEV *Scev0 = SE->getSCEV(Ptr0); | |||
1685 | const SCEV *ScevN = SE->getSCEV(PtrN); | |||
1686 | const auto *Diff = | |||
1687 | dyn_cast<SCEVConstant>(SE->getMinusSCEV(ScevN, Scev0)); | |||
1688 | uint64_t Size = DL->getTypeAllocSize(ScalarTy); | |||
1689 | // Check that the sorted loads are consecutive. | |||
1690 | if (Diff && Diff->getAPInt().getZExtValue() == (VL.size() - 1) * Size) { | |||
1691 | if (CurrentOrder.empty()) { | |||
1692 | // Original loads are consecutive and does not require reordering. | |||
1693 | ++NumOpsWantToKeepOriginalOrder; | |||
1694 | newTreeEntry(VL, /*Vectorized=*/true, UserTreeIdx, | |||
1695 | ReuseShuffleIndicies); | |||
1696 | 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); | |||
1697 | } else { | |||
1698 | // Need to reorder. | |||
1699 | auto I = NumOpsWantToKeepOrder.try_emplace(CurrentOrder).first; | |||
1700 | ++I->getSecond(); | |||
1701 | newTreeEntry(VL, /*Vectorized=*/true, UserTreeIdx, | |||
1702 | ReuseShuffleIndicies, I->getFirst()); | |||
1703 | 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); | |||
1704 | } | |||
1705 | return; | |||
1706 | } | |||
1707 | } | |||
1708 | ||||
1709 | DEBUG(dbgs() << "SLP: Gathering non-consecutive loads.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering non-consecutive loads.\n" ; } } while (false); | |||
1710 | BS.cancelScheduling(VL, VL0); | |||
1711 | newTreeEntry(VL, false, UserTreeIdx, ReuseShuffleIndicies); | |||
1712 | return; | |||
1713 | } | |||
1714 | case Instruction::ZExt: | |||
1715 | case Instruction::SExt: | |||
1716 | case Instruction::FPToUI: | |||
1717 | case Instruction::FPToSI: | |||
1718 | case Instruction::FPExt: | |||
1719 | case Instruction::PtrToInt: | |||
1720 | case Instruction::IntToPtr: | |||
1721 | case Instruction::SIToFP: | |||
1722 | case Instruction::UIToFP: | |||
1723 | case Instruction::Trunc: | |||
1724 | case Instruction::FPTrunc: | |||
1725 | case Instruction::BitCast: { | |||
1726 | Type *SrcTy = VL0->getOperand(0)->getType(); | |||
1727 | for (unsigned i = 0; i < VL.size(); ++i) { | |||
1728 | Type *Ty = cast<Instruction>(VL[i])->getOperand(0)->getType(); | |||
1729 | if (Ty != SrcTy || !isValidElementType(Ty)) { | |||
1730 | BS.cancelScheduling(VL, VL0); | |||
1731 | newTreeEntry(VL, false, UserTreeIdx, ReuseShuffleIndicies); | |||
1732 | DEBUG(dbgs() << "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); | |||
1733 | return; | |||
1734 | } | |||
1735 | } | |||
1736 | newTreeEntry(VL, true, UserTreeIdx, ReuseShuffleIndicies); | |||
1737 | 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); | |||
1738 | ||||
1739 | for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) { | |||
1740 | ValueList Operands; | |||
1741 | // Prepare the operand vector. | |||
1742 | for (Value *j : VL) | |||
1743 | Operands.push_back(cast<Instruction>(j)->getOperand(i)); | |||
1744 | ||||
1745 | buildTree_rec(Operands, Depth + 1, UserTreeIdx); | |||
1746 | } | |||
1747 | return; | |||
1748 | } | |||
1749 | case Instruction::ICmp: | |||
1750 | case Instruction::FCmp: { | |||
1751 | // Check that all of the compares have the same predicate. | |||
1752 | CmpInst::Predicate P0 = cast<CmpInst>(VL0)->getPredicate(); | |||
1753 | Type *ComparedTy = VL0->getOperand(0)->getType(); | |||
1754 | for (unsigned i = 1, e = VL.size(); i < e; ++i) { | |||
1755 | CmpInst *Cmp = cast<CmpInst>(VL[i]); | |||
1756 | if (Cmp->getPredicate() != P0 || | |||
1757 | Cmp->getOperand(0)->getType() != ComparedTy) { | |||
1758 | BS.cancelScheduling(VL, VL0); | |||
1759 | newTreeEntry(VL, false, UserTreeIdx, ReuseShuffleIndicies); | |||
1760 | DEBUG(dbgs() << "SLP: Gathering cmp with different predicate.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering cmp with different predicate.\n" ; } } while (false); | |||
1761 | return; | |||
1762 | } | |||
1763 | } | |||
1764 | ||||
1765 | newTreeEntry(VL, true, UserTreeIdx, ReuseShuffleIndicies); | |||
1766 | 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); | |||
1767 | ||||
1768 | for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) { | |||
1769 | ValueList Operands; | |||
1770 | // Prepare the operand vector. | |||
1771 | for (Value *j : VL) | |||
1772 | Operands.push_back(cast<Instruction>(j)->getOperand(i)); | |||
1773 | ||||
1774 | buildTree_rec(Operands, Depth + 1, UserTreeIdx); | |||
1775 | } | |||
1776 | return; | |||
1777 | } | |||
1778 | case Instruction::Select: | |||
1779 | case Instruction::Add: | |||
1780 | case Instruction::FAdd: | |||
1781 | case Instruction::Sub: | |||
1782 | case Instruction::FSub: | |||
1783 | case Instruction::Mul: | |||
1784 | case Instruction::FMul: | |||
1785 | case Instruction::UDiv: | |||
1786 | case Instruction::SDiv: | |||
1787 | case Instruction::FDiv: | |||
1788 | case Instruction::URem: | |||
1789 | case Instruction::SRem: | |||
1790 | case Instruction::FRem: | |||
1791 | case Instruction::Shl: | |||
1792 | case Instruction::LShr: | |||
1793 | case Instruction::AShr: | |||
1794 | case Instruction::And: | |||
1795 | case Instruction::Or: | |||
1796 | case Instruction::Xor: | |||
1797 | newTreeEntry(VL, true, UserTreeIdx, ReuseShuffleIndicies); | |||
1798 | DEBUG(dbgs() << "SLP: added a vector of bin op.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: added a vector of bin op.\n" ; } } while (false); | |||
1799 | ||||
1800 | // Sort operands of the instructions so that each side is more likely to | |||
1801 | // have the same opcode. | |||
1802 | if (isa<BinaryOperator>(VL0) && VL0->isCommutative()) { | |||
1803 | ValueList Left, Right; | |||
1804 | reorderInputsAccordingToOpcode(S.Opcode, VL, Left, Right); | |||
1805 | buildTree_rec(Left, Depth + 1, UserTreeIdx); | |||
1806 | buildTree_rec(Right, Depth + 1, UserTreeIdx); | |||
1807 | return; | |||
1808 | } | |||
1809 | ||||
1810 | for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) { | |||
1811 | ValueList Operands; | |||
1812 | // Prepare the operand vector. | |||
1813 | for (Value *j : VL) | |||
1814 | Operands.push_back(cast<Instruction>(j)->getOperand(i)); | |||
1815 | ||||
1816 | buildTree_rec(Operands, Depth + 1, UserTreeIdx); | |||
1817 | } | |||
1818 | return; | |||
1819 | ||||
1820 | case Instruction::GetElementPtr: { | |||
1821 | // We don't combine GEPs with complicated (nested) indexing. | |||
1822 | for (unsigned j = 0; j < VL.size(); ++j) { | |||
1823 | if (cast<Instruction>(VL[j])->getNumOperands() != 2) { | |||
1824 | 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); | |||
1825 | BS.cancelScheduling(VL, VL0); | |||
1826 | newTreeEntry(VL, false, UserTreeIdx, ReuseShuffleIndicies); | |||
1827 | return; | |||
1828 | } | |||
1829 | } | |||
1830 | ||||
1831 | // We can't combine several GEPs into one vector if they operate on | |||
1832 | // different types. | |||
1833 | Type *Ty0 = VL0->getOperand(0)->getType(); | |||
1834 | for (unsigned j = 0; j < VL.size(); ++j) { | |||
1835 | Type *CurTy = cast<Instruction>(VL[j])->getOperand(0)->getType(); | |||
1836 | if (Ty0 != CurTy) { | |||
1837 | DEBUG(dbgs() << "SLP: not-vectorizable GEP (different types).\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: not-vectorizable GEP (different types).\n" ; } } while (false); | |||
1838 | BS.cancelScheduling(VL, VL0); | |||
1839 | newTreeEntry(VL, false, UserTreeIdx, ReuseShuffleIndicies); | |||
1840 | return; | |||
1841 | } | |||
1842 | } | |||
1843 | ||||
1844 | // We don't combine GEPs with non-constant indexes. | |||
1845 | for (unsigned j = 0; j < VL.size(); ++j) { | |||
1846 | auto Op = cast<Instruction>(VL[j])->getOperand(1); | |||
1847 | if (!isa<ConstantInt>(Op)) { | |||
1848 | DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: not-vectorizable GEP (non-constant indexes).\n" ; } } while (false) | |||
1849 | dbgs() << "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); | |||
1850 | BS.cancelScheduling(VL, VL0); | |||
1851 | newTreeEntry(VL, false, UserTreeIdx, ReuseShuffleIndicies); | |||
1852 | return; | |||
1853 | } | |||
1854 | } | |||
1855 | ||||
1856 | newTreeEntry(VL, true, UserTreeIdx, ReuseShuffleIndicies); | |||
1857 | 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); | |||
1858 | for (unsigned i = 0, e = 2; i < e; ++i) { | |||
1859 | ValueList Operands; | |||
1860 | // Prepare the operand vector. | |||
1861 | for (Value *j : VL) | |||
1862 | Operands.push_back(cast<Instruction>(j)->getOperand(i)); | |||
1863 | ||||
1864 | buildTree_rec(Operands, Depth + 1, UserTreeIdx); | |||
1865 | } | |||
1866 | return; | |||
1867 | } | |||
1868 | case Instruction::Store: { | |||
1869 | // Check if the stores are consecutive or of we need to swizzle them. | |||
1870 | for (unsigned i = 0, e = VL.size() - 1; i < e; ++i) | |||
1871 | if (!isConsecutiveAccess(VL[i], VL[i + 1], *DL, *SE)) { | |||
1872 | BS.cancelScheduling(VL, VL0); | |||
1873 | newTreeEntry(VL, false, UserTreeIdx, ReuseShuffleIndicies); | |||
1874 | DEBUG(dbgs() << "SLP: Non-consecutive store.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Non-consecutive store.\n"; } } while (false); | |||
1875 | return; | |||
1876 | } | |||
1877 | ||||
1878 | newTreeEntry(VL, true, UserTreeIdx, ReuseShuffleIndicies); | |||
1879 | 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); | |||
1880 | ||||
1881 | ValueList Operands; | |||
1882 | for (Value *j : VL) | |||
1883 | Operands.push_back(cast<Instruction>(j)->getOperand(0)); | |||
1884 | ||||
1885 | buildTree_rec(Operands, Depth + 1, UserTreeIdx); | |||
1886 | return; | |||
1887 | } | |||
1888 | case Instruction::Call: { | |||
1889 | // Check if the calls are all to the same vectorizable intrinsic. | |||
1890 | CallInst *CI = cast<CallInst>(VL0); | |||
1891 | // Check if this is an Intrinsic call or something that can be | |||
1892 | // represented by an intrinsic call | |||
1893 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); | |||
1894 | if (!isTriviallyVectorizable(ID)) { | |||
1895 | BS.cancelScheduling(VL, VL0); | |||
1896 | newTreeEntry(VL, false, UserTreeIdx, ReuseShuffleIndicies); | |||
1897 | DEBUG(dbgs() << "SLP: Non-vectorizable call.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Non-vectorizable call.\n"; } } while (false); | |||
1898 | return; | |||
1899 | } | |||
1900 | Function *Int = CI->getCalledFunction(); | |||
1901 | Value *A1I = nullptr; | |||
1902 | if (hasVectorInstrinsicScalarOpd(ID, 1)) | |||
1903 | A1I = CI->getArgOperand(1); | |||
1904 | for (unsigned i = 1, e = VL.size(); i != e; ++i) { | |||
1905 | CallInst *CI2 = dyn_cast<CallInst>(VL[i]); | |||
1906 | if (!CI2 || CI2->getCalledFunction() != Int || | |||
1907 | getVectorIntrinsicIDForCall(CI2, TLI) != ID || | |||
1908 | !CI->hasIdenticalOperandBundleSchema(*CI2)) { | |||
1909 | BS.cancelScheduling(VL, VL0); | |||
1910 | newTreeEntry(VL, false, UserTreeIdx, ReuseShuffleIndicies); | |||
1911 | DEBUG(dbgs() << "SLP: mismatched calls:" << *CI << "!=" << *VL[i]do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: mismatched calls:" << * CI << "!=" << *VL[i] << "\n"; } } while (false ) | |||
1912 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: mismatched calls:" << * CI << "!=" << *VL[i] << "\n"; } } while (false ); | |||
1913 | return; | |||
1914 | } | |||
1915 | // ctlz,cttz and powi are special intrinsics whose second argument | |||
1916 | // should be same in order for them to be vectorized. | |||
1917 | if (hasVectorInstrinsicScalarOpd(ID, 1)) { | |||
1918 | Value *A1J = CI2->getArgOperand(1); | |||
1919 | if (A1I != A1J) { | |||
1920 | BS.cancelScheduling(VL, VL0); | |||
1921 | newTreeEntry(VL, false, UserTreeIdx, ReuseShuffleIndicies); | |||
1922 | DEBUG(dbgs() << "SLP: mismatched arguments in call:" << *CIdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: mismatched arguments in call:" << *CI << " argument "<< A1I<<"!=" << A1J << "\n"; } } while (false) | |||
1923 | << " argument "<< A1I<<"!=" << A1Jdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: mismatched arguments in call:" << *CI << " argument "<< A1I<<"!=" << A1J << "\n"; } } while (false) | |||
1924 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: mismatched arguments in call:" << *CI << " argument "<< A1I<<"!=" << A1J << "\n"; } } while (false); | |||
1925 | return; | |||
1926 | } | |||
1927 | } | |||
1928 | // Verify that the bundle operands are identical between the two calls. | |||
1929 | if (CI->hasOperandBundles() && | |||
1930 | !std::equal(CI->op_begin() + CI->getBundleOperandsStartIndex(), | |||
1931 | CI->op_begin() + CI->getBundleOperandsEndIndex(), | |||
1932 | CI2->op_begin() + CI2->getBundleOperandsStartIndex())) { | |||
1933 | BS.cancelScheduling(VL, VL0); | |||
1934 | newTreeEntry(VL, false, UserTreeIdx, ReuseShuffleIndicies); | |||
1935 | DEBUG(dbgs() << "SLP: mismatched bundle operands in calls:" << *CI << "!="do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: mismatched bundle operands in calls:" << *CI << "!=" << *VL[i] << '\n'; } } while (false) | |||
1936 | << *VL[i] << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: mismatched bundle operands in calls:" << *CI << "!=" << *VL[i] << '\n'; } } while (false); | |||
1937 | return; | |||
1938 | } | |||
1939 | } | |||
1940 | ||||
1941 | newTreeEntry(VL, true, UserTreeIdx, ReuseShuffleIndicies); | |||
1942 | for (unsigned i = 0, e = CI->getNumArgOperands(); i != e; ++i) { | |||
1943 | ValueList Operands; | |||
1944 | // Prepare the operand vector. | |||
1945 | for (Value *j : VL) { | |||
1946 | CallInst *CI2 = dyn_cast<CallInst>(j); | |||
1947 | Operands.push_back(CI2->getArgOperand(i)); | |||
1948 | } | |||
1949 | buildTree_rec(Operands, Depth + 1, UserTreeIdx); | |||
1950 | } | |||
1951 | return; | |||
1952 | } | |||
1953 | case Instruction::ShuffleVector: | |||
1954 | // If this is not an alternate sequence of opcode like add-sub | |||
1955 | // then do not vectorize this instruction. | |||
1956 | if (!S.IsAltShuffle) { | |||
1957 | BS.cancelScheduling(VL, VL0); | |||
1958 | newTreeEntry(VL, false, UserTreeIdx, ReuseShuffleIndicies); | |||
1959 | DEBUG(dbgs() << "SLP: ShuffleVector are not vectorized.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: ShuffleVector are not vectorized.\n" ; } } while (false); | |||
1960 | return; | |||
1961 | } | |||
1962 | newTreeEntry(VL, true, UserTreeIdx, ReuseShuffleIndicies); | |||
1963 | DEBUG(dbgs() << "SLP: added a ShuffleVector op.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: added a ShuffleVector op.\n" ; } } while (false); | |||
1964 | ||||
1965 | // Reorder operands if reordering would enable vectorization. | |||
1966 | if (isa<BinaryOperator>(VL0)) { | |||
1967 | ValueList Left, Right; | |||
1968 | reorderAltShuffleOperands(S.Opcode, VL, Left, Right); | |||
1969 | buildTree_rec(Left, Depth + 1, UserTreeIdx); | |||
1970 | buildTree_rec(Right, Depth + 1, UserTreeIdx); | |||
1971 | return; | |||
1972 | } | |||
1973 | ||||
1974 | for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) { | |||
1975 | ValueList Operands; | |||
1976 | // Prepare the operand vector. | |||
1977 | for (Value *j : VL) | |||
1978 | Operands.push_back(cast<Instruction>(j)->getOperand(i)); | |||
1979 | ||||
1980 | buildTree_rec(Operands, Depth + 1, UserTreeIdx); | |||
1981 | } | |||
1982 | return; | |||
1983 | ||||
1984 | default: | |||
1985 | BS.cancelScheduling(VL, VL0); | |||
1986 | newTreeEntry(VL, false, UserTreeIdx, ReuseShuffleIndicies); | |||
1987 | DEBUG(dbgs() << "SLP: Gathering unknown instruction.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering unknown instruction.\n" ; } } while (false); | |||
1988 | return; | |||
1989 | } | |||
1990 | } | |||
1991 | ||||
1992 | unsigned BoUpSLP::canMapToVector(Type *T, const DataLayout &DL) const { | |||
1993 | unsigned N; | |||
1994 | Type *EltTy; | |||
1995 | auto *ST = dyn_cast<StructType>(T); | |||
1996 | if (ST) { | |||
1997 | N = ST->getNumElements(); | |||
1998 | EltTy = *ST->element_begin(); | |||
1999 | } else { | |||
2000 | N = cast<ArrayType>(T)->getNumElements(); | |||
2001 | EltTy = cast<ArrayType>(T)->getElementType(); | |||
2002 | } | |||
2003 | if (!isValidElementType(EltTy)) | |||
2004 | return 0; | |||
2005 | uint64_t VTSize = DL.getTypeStoreSizeInBits(VectorType::get(EltTy, N)); | |||
2006 | if (VTSize < MinVecRegSize || VTSize > MaxVecRegSize || VTSize != DL.getTypeStoreSizeInBits(T)) | |||
2007 | return 0; | |||
2008 | if (ST) { | |||
2009 | // Check that struct is homogeneous. | |||
2010 | for (const auto *Ty : ST->elements()) | |||
2011 | if (Ty != EltTy) | |||
2012 | return 0; | |||
2013 | } | |||
2014 | return N; | |||
2015 | } | |||
2016 | ||||
2017 | bool BoUpSLP::canReuseExtract(ArrayRef<Value *> VL, Value *OpValue, | |||
2018 | SmallVectorImpl<unsigned> &CurrentOrder) const { | |||
2019 | Instruction *E0 = cast<Instruction>(OpValue); | |||
2020 | assert(E0->getOpcode() == Instruction::ExtractElement ||(static_cast <bool> (E0->getOpcode() == Instruction:: ExtractElement || E0->getOpcode() == Instruction::ExtractValue ) ? void (0) : __assert_fail ("E0->getOpcode() == Instruction::ExtractElement || E0->getOpcode() == Instruction::ExtractValue" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2021, __extension__ __PRETTY_FUNCTION__)) | |||
2021 | E0->getOpcode() == Instruction::ExtractValue)(static_cast <bool> (E0->getOpcode() == Instruction:: ExtractElement || E0->getOpcode() == Instruction::ExtractValue ) ? void (0) : __assert_fail ("E0->getOpcode() == Instruction::ExtractElement || E0->getOpcode() == Instruction::ExtractValue" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2021, __extension__ __PRETTY_FUNCTION__)); | |||
2022 | assert(E0->getOpcode() == getSameOpcode(VL).Opcode && "Invalid opcode")(static_cast <bool> (E0->getOpcode() == getSameOpcode (VL).Opcode && "Invalid opcode") ? void (0) : __assert_fail ("E0->getOpcode() == getSameOpcode(VL).Opcode && \"Invalid opcode\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2022, __extension__ __PRETTY_FUNCTION__)); | |||
2023 | // Check if all of the extracts come from the same vector and from the | |||
2024 | // correct offset. | |||
2025 | Value *Vec = E0->getOperand(0); | |||
2026 | ||||
2027 | CurrentOrder.clear(); | |||
2028 | ||||
2029 | // We have to extract from a vector/aggregate with the same number of elements. | |||
2030 | unsigned NElts; | |||
2031 | if (E0->getOpcode() == Instruction::ExtractValue) { | |||
2032 | const DataLayout &DL = E0->getModule()->getDataLayout(); | |||
2033 | NElts = canMapToVector(Vec->getType(), DL); | |||
2034 | if (!NElts) | |||
2035 | return false; | |||
2036 | // Check if load can be rewritten as load of vector. | |||
2037 | LoadInst *LI = dyn_cast<LoadInst>(Vec); | |||
2038 | if (!LI || !LI->isSimple() || !LI->hasNUses(VL.size())) | |||
2039 | return false; | |||
2040 | } else { | |||
2041 | NElts = Vec->getType()->getVectorNumElements(); | |||
2042 | } | |||
2043 | ||||
2044 | if (NElts != VL.size()) | |||
2045 | return false; | |||
2046 | ||||
2047 | // Check that all of the indices extract from the correct offset. | |||
2048 | bool ShouldKeepOrder = true; | |||
2049 | unsigned E = VL.size(); | |||
2050 | // Assign to all items the initial value E + 1 so we can check if the extract | |||
2051 | // instruction index was used already. | |||
2052 | // Also, later we can check that all the indices are used and we have a | |||
2053 | // consecutive access in the extract instructions, by checking that no | |||
2054 | // element of CurrentOrder still has value E + 1. | |||
2055 | CurrentOrder.assign(E, E + 1); | |||
2056 | unsigned I = 0; | |||
2057 | for (; I < E; ++I) { | |||
2058 | auto *Inst = cast<Instruction>(VL[I]); | |||
2059 | if (Inst->getOperand(0) != Vec) | |||
2060 | break; | |||
2061 | Optional<unsigned> Idx = getExtractIndex(Inst); | |||
2062 | if (!Idx) | |||
2063 | break; | |||
2064 | const unsigned ExtIdx = *Idx; | |||
2065 | if (ExtIdx != I) { | |||
2066 | if (ExtIdx >= E || CurrentOrder[ExtIdx] != E + 1) | |||
2067 | break; | |||
2068 | ShouldKeepOrder = false; | |||
2069 | CurrentOrder[ExtIdx] = I; | |||
2070 | } else { | |||
2071 | if (CurrentOrder[I] != E + 1) | |||
2072 | break; | |||
2073 | CurrentOrder[I] = I; | |||
2074 | } | |||
2075 | } | |||
2076 | if (I < E) { | |||
2077 | CurrentOrder.clear(); | |||
2078 | return false; | |||
2079 | } | |||
2080 | ||||
2081 | return ShouldKeepOrder; | |||
2082 | } | |||
2083 | ||||
2084 | bool BoUpSLP::areAllUsersVectorized(Instruction *I) const { | |||
2085 | return I->hasOneUse() || | |||
2086 | std::all_of(I->user_begin(), I->user_end(), [this](User *U) { | |||
2087 | return ScalarToTreeEntry.count(U) > 0; | |||
2088 | }); | |||
2089 | } | |||
2090 | ||||
2091 | int BoUpSLP::getEntryCost(TreeEntry *E) { | |||
2092 | ArrayRef<Value*> VL = E->Scalars; | |||
2093 | ||||
2094 | Type *ScalarTy = VL[0]->getType(); | |||
2095 | if (StoreInst *SI = dyn_cast<StoreInst>(VL[0])) | |||
2096 | ScalarTy = SI->getValueOperand()->getType(); | |||
2097 | else if (CmpInst *CI = dyn_cast<CmpInst>(VL[0])) | |||
2098 | ScalarTy = CI->getOperand(0)->getType(); | |||
2099 | VectorType *VecTy = VectorType::get(ScalarTy, VL.size()); | |||
2100 | ||||
2101 | // If we have computed a smaller type for the expression, update VecTy so | |||
2102 | // that the costs will be accurate. | |||
2103 | if (MinBWs.count(VL[0])) | |||
2104 | VecTy = VectorType::get( | |||
2105 | IntegerType::get(F->getContext(), MinBWs[VL[0]].first), VL.size()); | |||
2106 | ||||
2107 | unsigned ReuseShuffleNumbers = E->ReuseShuffleIndices.size(); | |||
2108 | bool NeedToShuffleReuses = !E->ReuseShuffleIndices.empty(); | |||
2109 | int ReuseShuffleCost = 0; | |||
2110 | if (NeedToShuffleReuses) { | |||
2111 | ReuseShuffleCost = | |||
2112 | TTI->getShuffleCost(TargetTransformInfo::SK_PermuteSingleSrc, VecTy); | |||
2113 | } | |||
2114 | if (E->NeedToGather) { | |||
2115 | if (allConstant(VL)) | |||
2116 | return 0; | |||
2117 | if (isSplat(VL)) { | |||
2118 | return ReuseShuffleCost + | |||
2119 | TTI->getShuffleCost(TargetTransformInfo::SK_Broadcast, VecTy, 0); | |||
2120 | } | |||
2121 | if (getSameOpcode(VL).Opcode == Instruction::ExtractElement && | |||
2122 | allSameType(VL) && allSameBlock(VL)) { | |||
2123 | Optional<TargetTransformInfo::ShuffleKind> ShuffleKind = isShuffle(VL); | |||
2124 | if (ShuffleKind.hasValue()) { | |||
2125 | int Cost = TTI->getShuffleCost(ShuffleKind.getValue(), VecTy); | |||
2126 | for (auto *V : VL) { | |||
2127 | // If all users of instruction are going to be vectorized and this | |||
2128 | // instruction itself is not going to be vectorized, consider this | |||
2129 | // instruction as dead and remove its cost from the final cost of the | |||
2130 | // vectorized tree. | |||
2131 | if (areAllUsersVectorized(cast<Instruction>(V)) && | |||
2132 | !ScalarToTreeEntry.count(V)) { | |||
2133 | auto *IO = cast<ConstantInt>( | |||
2134 | cast<ExtractElementInst>(V)->getIndexOperand()); | |||
2135 | Cost -= TTI->getVectorInstrCost(Instruction::ExtractElement, VecTy, | |||
2136 | IO->getZExtValue()); | |||
2137 | } | |||
2138 | } | |||
2139 | return ReuseShuffleCost + Cost; | |||
2140 | } | |||
2141 | } | |||
2142 | return ReuseShuffleCost + getGatherCost(VL); | |||
2143 | } | |||
2144 | InstructionsState S = getSameOpcode(VL); | |||
2145 | assert(S.Opcode && allSameType(VL) && allSameBlock(VL) && "Invalid VL")(static_cast <bool> (S.Opcode && allSameType(VL ) && allSameBlock(VL) && "Invalid VL") ? void (0) : __assert_fail ("S.Opcode && allSameType(VL) && allSameBlock(VL) && \"Invalid VL\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2145, __extension__ __PRETTY_FUNCTION__)); | |||
2146 | Instruction *VL0 = cast<Instruction>(S.OpValue); | |||
2147 | unsigned ShuffleOrOp = S.IsAltShuffle ? | |||
2148 | (unsigned) Instruction::ShuffleVector : S.Opcode; | |||
2149 | switch (ShuffleOrOp) { | |||
2150 | case Instruction::PHI: | |||
2151 | return 0; | |||
2152 | ||||
2153 | case Instruction::ExtractValue: | |||
2154 | case Instruction::ExtractElement: | |||
2155 | if (NeedToShuffleReuses) { | |||
2156 | unsigned Idx = 0; | |||
2157 | for (unsigned I : E->ReuseShuffleIndices) { | |||
2158 | if (ShuffleOrOp == Instruction::ExtractElement) { | |||
2159 | auto *IO = cast<ConstantInt>( | |||
2160 | cast<ExtractElementInst>(VL[I])->getIndexOperand()); | |||
2161 | Idx = IO->getZExtValue(); | |||
2162 | ReuseShuffleCost -= TTI->getVectorInstrCost( | |||
2163 | Instruction::ExtractElement, VecTy, Idx); | |||
2164 | } else { | |||
2165 | ReuseShuffleCost -= TTI->getVectorInstrCost( | |||
2166 | Instruction::ExtractElement, VecTy, Idx); | |||
2167 | ++Idx; | |||
2168 | } | |||
2169 | } | |||
2170 | Idx = ReuseShuffleNumbers; | |||
2171 | for (Value *V : VL) { | |||
2172 | if (ShuffleOrOp == Instruction::ExtractElement) { | |||
2173 | auto *IO = cast<ConstantInt>( | |||
2174 | cast<ExtractElementInst>(V)->getIndexOperand()); | |||
2175 | Idx = IO->getZExtValue(); | |||
2176 | } else { | |||
2177 | --Idx; | |||
2178 | } | |||
2179 | ReuseShuffleCost += | |||
2180 | TTI->getVectorInstrCost(Instruction::ExtractElement, VecTy, Idx); | |||
2181 | } | |||
2182 | } | |||
2183 | if (!E->NeedToGather) { | |||
2184 | int DeadCost = ReuseShuffleCost; | |||
2185 | if (!E->ReorderIndices.empty()) { | |||
2186 | // TODO: Merge this shuffle with the ReuseShuffleCost. | |||
2187 | DeadCost += TTI->getShuffleCost( | |||
2188 | TargetTransformInfo::SK_PermuteSingleSrc, VecTy); | |||
2189 | } | |||
2190 | for (unsigned i = 0, e = VL.size(); i < e; ++i) { | |||
2191 | Instruction *E = cast<Instruction>(VL[i]); | |||
2192 | // If all users are going to be vectorized, instruction can be | |||
2193 | // considered as dead. | |||
2194 | // The same, if have only one user, it will be vectorized for sure. | |||
2195 | if (areAllUsersVectorized(E)) | |||
2196 | // Take credit for instruction that will become dead. | |||
2197 | DeadCost -= | |||
2198 | TTI->getVectorInstrCost(Instruction::ExtractElement, VecTy, i); | |||
2199 | } | |||
2200 | return DeadCost; | |||
2201 | } | |||
2202 | return ReuseShuffleCost + getGatherCost(VL); | |||
2203 | ||||
2204 | case Instruction::ZExt: | |||
2205 | case Instruction::SExt: | |||
2206 | case Instruction::FPToUI: | |||
2207 | case Instruction::FPToSI: | |||
2208 | case Instruction::FPExt: | |||
2209 | case Instruction::PtrToInt: | |||
2210 | case Instruction::IntToPtr: | |||
2211 | case Instruction::SIToFP: | |||
2212 | case Instruction::UIToFP: | |||
2213 | case Instruction::Trunc: | |||
2214 | case Instruction::FPTrunc: | |||
2215 | case Instruction::BitCast: { | |||
2216 | Type *SrcTy = VL0->getOperand(0)->getType(); | |||
2217 | if (NeedToShuffleReuses) { | |||
2218 | ReuseShuffleCost -= | |||
2219 | (ReuseShuffleNumbers - VL.size()) * | |||
2220 | TTI->getCastInstrCost(S.Opcode, ScalarTy, SrcTy, VL0); | |||
2221 | } | |||
2222 | ||||
2223 | // Calculate the cost of this instruction. | |||
2224 | int ScalarCost = VL.size() * TTI->getCastInstrCost(VL0->getOpcode(), | |||
2225 | VL0->getType(), SrcTy, VL0); | |||
2226 | ||||
2227 | VectorType *SrcVecTy = VectorType::get(SrcTy, VL.size()); | |||
2228 | int VecCost = 0; | |||
2229 | // Check if the values are candidates to demote. | |||
2230 | if (!MinBWs.count(VL0) || VecTy != SrcVecTy) { | |||
2231 | VecCost = ReuseShuffleCost + | |||
2232 | TTI->getCastInstrCost(VL0->getOpcode(), VecTy, SrcVecTy, VL0); | |||
2233 | } | |||
2234 | return VecCost - ScalarCost; | |||
2235 | } | |||
2236 | case Instruction::FCmp: | |||
2237 | case Instruction::ICmp: | |||
2238 | case Instruction::Select: { | |||
2239 | // Calculate the cost of this instruction. | |||
2240 | if (NeedToShuffleReuses) { | |||
2241 | ReuseShuffleCost -= (ReuseShuffleNumbers - VL.size()) * | |||
2242 | TTI->getCmpSelInstrCost(S.Opcode, ScalarTy, | |||
2243 | Builder.getInt1Ty(), VL0); | |||
2244 | } | |||
2245 | VectorType *MaskTy = VectorType::get(Builder.getInt1Ty(), VL.size()); | |||
2246 | int ScalarCost = VecTy->getNumElements() * | |||
2247 | TTI->getCmpSelInstrCost(S.Opcode, ScalarTy, Builder.getInt1Ty(), VL0); | |||
2248 | int VecCost = TTI->getCmpSelInstrCost(S.Opcode, VecTy, MaskTy, VL0); | |||
2249 | return ReuseShuffleCost + VecCost - ScalarCost; | |||
2250 | } | |||
2251 | case Instruction::Add: | |||
2252 | case Instruction::FAdd: | |||
2253 | case Instruction::Sub: | |||
2254 | case Instruction::FSub: | |||
2255 | case Instruction::Mul: | |||
2256 | case Instruction::FMul: | |||
2257 | case Instruction::UDiv: | |||
2258 | case Instruction::SDiv: | |||
2259 | case Instruction::FDiv: | |||
2260 | case Instruction::URem: | |||
2261 | case Instruction::SRem: | |||
2262 | case Instruction::FRem: | |||
2263 | case Instruction::Shl: | |||
2264 | case Instruction::LShr: | |||
2265 | case Instruction::AShr: | |||
2266 | case Instruction::And: | |||
2267 | case Instruction::Or: | |||
2268 | case Instruction::Xor: { | |||
2269 | // Certain instructions can be cheaper to vectorize if they have a | |||
2270 | // constant second vector operand. | |||
2271 | TargetTransformInfo::OperandValueKind Op1VK = | |||
2272 | TargetTransformInfo::OK_AnyValue; | |||
2273 | TargetTransformInfo::OperandValueKind Op2VK = | |||
2274 | TargetTransformInfo::OK_UniformConstantValue; | |||
2275 | TargetTransformInfo::OperandValueProperties Op1VP = | |||
2276 | TargetTransformInfo::OP_None; | |||
2277 | TargetTransformInfo::OperandValueProperties Op2VP = | |||
2278 | TargetTransformInfo::OP_None; | |||
2279 | ||||
2280 | // If all operands are exactly the same ConstantInt then set the | |||
2281 | // operand kind to OK_UniformConstantValue. | |||
2282 | // If instead not all operands are constants, then set the operand kind | |||
2283 | // to OK_AnyValue. If all operands are constants but not the same, | |||
2284 | // then set the operand kind to OK_NonUniformConstantValue. | |||
2285 | ConstantInt *CInt = nullptr; | |||
2286 | for (unsigned i = 0; i < VL.size(); ++i) { | |||
2287 | const Instruction *I = cast<Instruction>(VL[i]); | |||
2288 | if (!isa<ConstantInt>(I->getOperand(1))) { | |||
2289 | Op2VK = TargetTransformInfo::OK_AnyValue; | |||
2290 | break; | |||
2291 | } | |||
2292 | if (i == 0) { | |||
2293 | CInt = cast<ConstantInt>(I->getOperand(1)); | |||
2294 | continue; | |||
2295 | } | |||
2296 | if (Op2VK == TargetTransformInfo::OK_UniformConstantValue && | |||
2297 | CInt != cast<ConstantInt>(I->getOperand(1))) | |||
2298 | Op2VK = TargetTransformInfo::OK_NonUniformConstantValue; | |||
2299 | } | |||
2300 | // FIXME: Currently cost of model modification for division by power of | |||
2301 | // 2 is handled for X86 and AArch64. Add support for other targets. | |||
2302 | if (Op2VK == TargetTransformInfo::OK_UniformConstantValue && CInt && | |||
2303 | CInt->getValue().isPowerOf2()) | |||
2304 | Op2VP = TargetTransformInfo::OP_PowerOf2; | |||
2305 | ||||
2306 | SmallVector<const Value *, 4> Operands(VL0->operand_values()); | |||
2307 | if (NeedToShuffleReuses) { | |||
2308 | ReuseShuffleCost -= | |||
2309 | (ReuseShuffleNumbers - VL.size()) * | |||
2310 | TTI->getArithmeticInstrCost(S.Opcode, ScalarTy, Op1VK, Op2VK, Op1VP, | |||
2311 | Op2VP, Operands); | |||
2312 | } | |||
2313 | int ScalarCost = | |||
2314 | VecTy->getNumElements() * | |||
2315 | TTI->getArithmeticInstrCost(S.Opcode, ScalarTy, Op1VK, Op2VK, Op1VP, | |||
2316 | Op2VP, Operands); | |||
2317 | int VecCost = TTI->getArithmeticInstrCost(S.Opcode, VecTy, Op1VK, Op2VK, | |||
2318 | Op1VP, Op2VP, Operands); | |||
2319 | return ReuseShuffleCost + VecCost - ScalarCost; | |||
2320 | } | |||
2321 | case Instruction::GetElementPtr: { | |||
2322 | TargetTransformInfo::OperandValueKind Op1VK = | |||
2323 | TargetTransformInfo::OK_AnyValue; | |||
2324 | TargetTransformInfo::OperandValueKind Op2VK = | |||
2325 | TargetTransformInfo::OK_UniformConstantValue; | |||
2326 | ||||
2327 | if (NeedToShuffleReuses) { | |||
2328 | ReuseShuffleCost -= (ReuseShuffleNumbers - VL.size()) * | |||
2329 | TTI->getArithmeticInstrCost(Instruction::Add, | |||
2330 | ScalarTy, Op1VK, Op2VK); | |||
2331 | } | |||
2332 | int ScalarCost = | |||
2333 | VecTy->getNumElements() * | |||
2334 | TTI->getArithmeticInstrCost(Instruction::Add, ScalarTy, Op1VK, Op2VK); | |||
2335 | int VecCost = | |||
2336 | TTI->getArithmeticInstrCost(Instruction::Add, VecTy, Op1VK, Op2VK); | |||
2337 | ||||
2338 | return ReuseShuffleCost + VecCost - ScalarCost; | |||
2339 | } | |||
2340 | case Instruction::Load: { | |||
2341 | // Cost of wide load - cost of scalar loads. | |||
2342 | unsigned alignment = dyn_cast<LoadInst>(VL0)->getAlignment(); | |||
2343 | if (NeedToShuffleReuses) { | |||
2344 | ReuseShuffleCost -= (ReuseShuffleNumbers - VL.size()) * | |||
2345 | TTI->getMemoryOpCost(Instruction::Load, ScalarTy, | |||
2346 | alignment, 0, VL0); | |||
2347 | } | |||
2348 | int ScalarLdCost = VecTy->getNumElements() * | |||
2349 | TTI->getMemoryOpCost(Instruction::Load, ScalarTy, alignment, 0, VL0); | |||
2350 | int VecLdCost = TTI->getMemoryOpCost(Instruction::Load, | |||
2351 | VecTy, alignment, 0, VL0); | |||
2352 | if (!E->ReorderIndices.empty()) { | |||
2353 | // TODO: Merge this shuffle with the ReuseShuffleCost. | |||
2354 | VecLdCost += TTI->getShuffleCost( | |||
2355 | TargetTransformInfo::SK_PermuteSingleSrc, VecTy); | |||
2356 | } | |||
2357 | return ReuseShuffleCost + VecLdCost - ScalarLdCost; | |||
2358 | } | |||
2359 | case Instruction::Store: { | |||
2360 | // We know that we can merge the stores. Calculate the cost. | |||
2361 | unsigned alignment = dyn_cast<StoreInst>(VL0)->getAlignment(); | |||
2362 | if (NeedToShuffleReuses) { | |||
2363 | ReuseShuffleCost -= (ReuseShuffleNumbers - VL.size()) * | |||
2364 | TTI->getMemoryOpCost(Instruction::Store, ScalarTy, | |||
2365 | alignment, 0, VL0); | |||
2366 | } | |||
2367 | int ScalarStCost = VecTy->getNumElements() * | |||
2368 | TTI->getMemoryOpCost(Instruction::Store, ScalarTy, alignment, 0, VL0); | |||
2369 | int VecStCost = TTI->getMemoryOpCost(Instruction::Store, | |||
2370 | VecTy, alignment, 0, VL0); | |||
2371 | return ReuseShuffleCost + VecStCost - ScalarStCost; | |||
2372 | } | |||
2373 | case Instruction::Call: { | |||
2374 | CallInst *CI = cast<CallInst>(VL0); | |||
2375 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); | |||
2376 | ||||
2377 | // Calculate the cost of the scalar and vector calls. | |||
2378 | SmallVector<Type*, 4> ScalarTys; | |||
2379 | for (unsigned op = 0, opc = CI->getNumArgOperands(); op!= opc; ++op) | |||
2380 | ScalarTys.push_back(CI->getArgOperand(op)->getType()); | |||
2381 | ||||
2382 | FastMathFlags FMF; | |||
2383 | if (auto *FPMO = dyn_cast<FPMathOperator>(CI)) | |||
2384 | FMF = FPMO->getFastMathFlags(); | |||
2385 | ||||
2386 | if (NeedToShuffleReuses) { | |||
2387 | ReuseShuffleCost -= | |||
2388 | (ReuseShuffleNumbers - VL.size()) * | |||
2389 | TTI->getIntrinsicInstrCost(ID, ScalarTy, ScalarTys, FMF); | |||
2390 | } | |||
2391 | int ScalarCallCost = VecTy->getNumElements() * | |||
2392 | TTI->getIntrinsicInstrCost(ID, ScalarTy, ScalarTys, FMF); | |||
2393 | ||||
2394 | SmallVector<Value *, 4> Args(CI->arg_operands()); | |||
2395 | int VecCallCost = TTI->getIntrinsicInstrCost(ID, CI->getType(), Args, FMF, | |||
2396 | VecTy->getNumElements()); | |||
2397 | ||||
2398 | 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) | |||
2399 | << " (" << VecCallCost << "-" << ScalarCallCost << ")"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Call cost "<< VecCallCost - ScalarCallCost << " (" << VecCallCost << "-" << ScalarCallCost << ")" << " for " << *CI << "\n"; } } while (false) | |||
2400 | << " for " << *CI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Call cost "<< VecCallCost - ScalarCallCost << " (" << VecCallCost << "-" << ScalarCallCost << ")" << " for " << *CI << "\n"; } } while (false); | |||
2401 | ||||
2402 | return ReuseShuffleCost + VecCallCost - ScalarCallCost; | |||
2403 | } | |||
2404 | case Instruction::ShuffleVector: { | |||
2405 | TargetTransformInfo::OperandValueKind Op1VK = | |||
2406 | TargetTransformInfo::OK_AnyValue; | |||
2407 | TargetTransformInfo::OperandValueKind Op2VK = | |||
2408 | TargetTransformInfo::OK_AnyValue; | |||
2409 | int ScalarCost = 0; | |||
2410 | if (NeedToShuffleReuses) { | |||
2411 | for (unsigned Idx : E->ReuseShuffleIndices) { | |||
2412 | Instruction *I = cast<Instruction>(VL[Idx]); | |||
2413 | if (!I) | |||
2414 | continue; | |||
2415 | ReuseShuffleCost -= TTI->getArithmeticInstrCost( | |||
2416 | I->getOpcode(), ScalarTy, Op1VK, Op2VK); | |||
2417 | } | |||
2418 | for (Value *V : VL) { | |||
2419 | Instruction *I = cast<Instruction>(V); | |||
2420 | if (!I) | |||
2421 | continue; | |||
2422 | ReuseShuffleCost += TTI->getArithmeticInstrCost( | |||
2423 | I->getOpcode(), ScalarTy, Op1VK, Op2VK); | |||
2424 | } | |||
2425 | } | |||
2426 | int VecCost = 0; | |||
2427 | for (Value *i : VL) { | |||
2428 | Instruction *I = cast<Instruction>(i); | |||
2429 | if (!I) | |||
2430 | break; | |||
2431 | ScalarCost += | |||
2432 | TTI->getArithmeticInstrCost(I->getOpcode(), ScalarTy, Op1VK, Op2VK); | |||
2433 | } | |||
2434 | // VecCost is equal to sum of the cost of creating 2 vectors | |||
2435 | // and the cost of creating shuffle. | |||
2436 | Instruction *I0 = cast<Instruction>(VL[0]); | |||
2437 | VecCost = | |||
2438 | TTI->getArithmeticInstrCost(I0->getOpcode(), VecTy, Op1VK, Op2VK); | |||
2439 | Instruction *I1 = cast<Instruction>(VL[1]); | |||
2440 | VecCost += | |||
2441 | TTI->getArithmeticInstrCost(I1->getOpcode(), VecTy, Op1VK, Op2VK); | |||
2442 | VecCost += | |||
2443 | TTI->getShuffleCost(TargetTransformInfo::SK_Alternate, VecTy, 0); | |||
2444 | return ReuseShuffleCost + VecCost - ScalarCost; | |||
2445 | } | |||
2446 | default: | |||
2447 | llvm_unreachable("Unknown instruction")::llvm::llvm_unreachable_internal("Unknown instruction", "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2447); | |||
2448 | } | |||
2449 | } | |||
2450 | ||||
2451 | bool BoUpSLP::isFullyVectorizableTinyTree() { | |||
2452 | 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) | |||
2453 | 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); | |||
2454 | ||||
2455 | // We only handle trees of heights 1 and 2. | |||
2456 | if (VectorizableTree.size() == 1 && !VectorizableTree[0].NeedToGather) | |||
2457 | return true; | |||
2458 | ||||
2459 | if (VectorizableTree.size() != 2) | |||
2460 | return false; | |||
2461 | ||||
2462 | // Handle splat and all-constants stores. | |||
2463 | if (!VectorizableTree[0].NeedToGather && | |||
2464 | (allConstant(VectorizableTree[1].Scalars) || | |||
2465 | isSplat(VectorizableTree[1].Scalars))) | |||
2466 | return true; | |||
2467 | ||||
2468 | // Gathering cost would be too much for tiny trees. | |||
2469 | if (VectorizableTree[0].NeedToGather || VectorizableTree[1].NeedToGather) | |||
2470 | return false; | |||
2471 | ||||
2472 | return true; | |||
2473 | } | |||
2474 | ||||
2475 | bool BoUpSLP::isTreeTinyAndNotFullyVectorizable() { | |||
2476 | // We can vectorize the tree if its size is greater than or equal to the | |||
2477 | // minimum size specified by the MinTreeSize command line option. | |||
2478 | if (VectorizableTree.size() >= MinTreeSize) | |||
2479 | return false; | |||
2480 | ||||
2481 | // If we have a tiny tree (a tree whose size is less than MinTreeSize), we | |||
2482 | // can vectorize it if we can prove it fully vectorizable. | |||
2483 | if (isFullyVectorizableTinyTree()) | |||
2484 | return false; | |||
2485 | ||||
2486 | assert(VectorizableTree.empty()(static_cast <bool> (VectorizableTree.empty() ? ExternalUses .empty() : true && "We shouldn't have any external users" ) ? void (0) : __assert_fail ("VectorizableTree.empty() ? ExternalUses.empty() : true && \"We shouldn't have any external users\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2488, __extension__ __PRETTY_FUNCTION__)) | |||
2487 | ? ExternalUses.empty()(static_cast <bool> (VectorizableTree.empty() ? ExternalUses .empty() : true && "We shouldn't have any external users" ) ? void (0) : __assert_fail ("VectorizableTree.empty() ? ExternalUses.empty() : true && \"We shouldn't have any external users\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2488, __extension__ __PRETTY_FUNCTION__)) | |||
2488 | : true && "We shouldn't have any external users")(static_cast <bool> (VectorizableTree.empty() ? ExternalUses .empty() : true && "We shouldn't have any external users" ) ? void (0) : __assert_fail ("VectorizableTree.empty() ? ExternalUses.empty() : true && \"We shouldn't have any external users\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2488, __extension__ __PRETTY_FUNCTION__)); | |||
2489 | ||||
2490 | // Otherwise, we can't vectorize the tree. It is both tiny and not fully | |||
2491 | // vectorizable. | |||
2492 | return true; | |||
2493 | } | |||
2494 | ||||
2495 | int BoUpSLP::getSpillCost() { | |||
2496 | // Walk from the bottom of the tree to the top, tracking which values are | |||
2497 | // live. When we see a call instruction that is not part of our tree, | |||
2498 | // query TTI to see if there is a cost to keeping values live over it | |||
2499 | // (for example, if spills and fills are required). | |||
2500 | unsigned BundleWidth = VectorizableTree.front().Scalars.size(); | |||
2501 | int Cost = 0; | |||
2502 | ||||
2503 | SmallPtrSet<Instruction*, 4> LiveValues; | |||
2504 | Instruction *PrevInst = nullptr; | |||
2505 | ||||
2506 | for (const auto &N : VectorizableTree) { | |||
2507 | Instruction *Inst = dyn_cast<Instruction>(N.Scalars[0]); | |||
2508 | if (!Inst) | |||
2509 | continue; | |||
2510 | ||||
2511 | if (!PrevInst) { | |||
2512 | PrevInst = Inst; | |||
2513 | continue; | |||
2514 | } | |||
2515 | ||||
2516 | // Update LiveValues. | |||
2517 | LiveValues.erase(PrevInst); | |||
2518 | for (auto &J : PrevInst->operands()) { | |||
2519 | if (isa<Instruction>(&*J) && getTreeEntry(&*J)) | |||
2520 | LiveValues.insert(cast<Instruction>(&*J)); | |||
2521 | } | |||
2522 | ||||
2523 | 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) | |||
2524 | 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) | |||
2525 | 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) | |||
2526 | 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) | |||
2527 | 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) | |||
2528 | 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) | |||
2529 | })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); | |||
2530 | ||||
2531 | // Now find the sequence of instructions between PrevInst and Inst. | |||
2532 | BasicBlock::reverse_iterator InstIt = ++Inst->getIterator().getReverse(), | |||
2533 | PrevInstIt = | |||
2534 | PrevInst->getIterator().getReverse(); | |||
2535 | while (InstIt != PrevInstIt) { | |||
2536 | if (PrevInstIt == PrevInst->getParent()->rend()) { | |||
2537 | PrevInstIt = Inst->getParent()->rbegin(); | |||
2538 | continue; | |||
2539 | } | |||
2540 | ||||
2541 | if (isa<CallInst>(&*PrevInstIt) && &*PrevInstIt != PrevInst) { | |||
2542 | SmallVector<Type*, 4> V; | |||
2543 | for (auto *II : LiveValues) | |||
2544 | V.push_back(VectorType::get(II->getType(), BundleWidth)); | |||
2545 | Cost += TTI->getCostOfKeepingLiveOverCall(V); | |||
2546 | } | |||
2547 | ||||
2548 | ++PrevInstIt; | |||
2549 | } | |||
2550 | ||||
2551 | PrevInst = Inst; | |||
2552 | } | |||
2553 | ||||
2554 | return Cost; | |||
2555 | } | |||
2556 | ||||
2557 | int BoUpSLP::getTreeCost() { | |||
2558 | int Cost = 0; | |||
2559 | 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) | |||
2560 | VectorizableTree.size() << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Calculating cost for tree of size " << VectorizableTree.size() << ".\n"; } } while ( false); | |||
2561 | ||||
2562 | unsigned BundleWidth = VectorizableTree[0].Scalars.size(); | |||
2563 | ||||
2564 | for (unsigned I = 0, E = VectorizableTree.size(); I < E; ++I) { | |||
2565 | TreeEntry &TE = VectorizableTree[I]; | |||
2566 | ||||
2567 | // We create duplicate tree entries for gather sequences that have multiple | |||
2568 | // uses. However, we should not compute the cost of duplicate sequences. | |||
2569 | // For example, if we have a build vector (i.e., insertelement sequence) | |||
2570 | // that is used by more than one vector instruction, we only need to | |||
2571 | // compute the cost of the insertelement instructions once. The redundent | |||
2572 | // instructions will be eliminated by CSE. | |||
2573 | // | |||
2574 | // We should consider not creating duplicate tree entries for gather | |||
2575 | // sequences, and instead add additional edges to the tree representing | |||
2576 | // their uses. Since such an approach results in fewer total entries, | |||
2577 | // existing heuristics based on tree size may yeild different results. | |||
2578 | // | |||
2579 | if (TE.NeedToGather && | |||
2580 | std::any_of(std::next(VectorizableTree.begin(), I + 1), | |||
2581 | VectorizableTree.end(), [TE](TreeEntry &Entry) { | |||
2582 | return Entry.NeedToGather && Entry.isSame(TE.Scalars); | |||
2583 | })) | |||
2584 | continue; | |||
2585 | ||||
2586 | int C = getEntryCost(&TE); | |||
2587 | DEBUG(dbgs() << "SLP: Adding cost " << C << " for bundle that starts with "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Adding cost " << C << " for bundle that starts with " << *TE.Scalars[0] << ".\n"; } } while (false) | |||
2588 | << *TE.Scalars[0] << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Adding cost " << C << " for bundle that starts with " << *TE.Scalars[0] << ".\n"; } } while (false); | |||
2589 | Cost += C; | |||
2590 | } | |||
2591 | ||||
2592 | SmallSet<Value *, 16> ExtractCostCalculated; | |||
2593 | int ExtractCost = 0; | |||
2594 | for (ExternalUser &EU : ExternalUses) { | |||
2595 | // We only add extract cost once for the same scalar. | |||
2596 | if (!ExtractCostCalculated.insert(EU.Scalar).second) | |||
2597 | continue; | |||
2598 | ||||
2599 | // Uses by ephemeral values are free (because the ephemeral value will be | |||
2600 | // removed prior to code generation, and so the extraction will be | |||
2601 | // removed as well). | |||
2602 | if (EphValues.count(EU.User)) | |||
2603 | continue; | |||
2604 | ||||
2605 | // If we plan to rewrite the tree in a smaller type, we will need to sign | |||
2606 | // extend the extracted value back to the original type. Here, we account | |||
2607 | // for the extract and the added cost of the sign extend if needed. | |||
2608 | auto *VecTy = VectorType::get(EU.Scalar->getType(), BundleWidth); | |||
2609 | auto *ScalarRoot = VectorizableTree[0].Scalars[0]; | |||
2610 | if (MinBWs.count(ScalarRoot)) { | |||
2611 | auto *MinTy = IntegerType::get(F->getContext(), MinBWs[ScalarRoot].first); | |||
2612 | auto Extend = | |||
2613 | MinBWs[ScalarRoot].second ? Instruction::SExt : Instruction::ZExt; | |||
2614 | VecTy = VectorType::get(MinTy, BundleWidth); | |||
2615 | ExtractCost += TTI->getExtractWithExtendCost(Extend, EU.Scalar->getType(), | |||
2616 | VecTy, EU.Lane); | |||
2617 | } else { | |||
2618 | ExtractCost += | |||
2619 | TTI->getVectorInstrCost(Instruction::ExtractElement, VecTy, EU.Lane); | |||
2620 | } | |||
2621 | } | |||
2622 | ||||
2623 | int SpillCost = getSpillCost(); | |||
2624 | Cost += SpillCost + ExtractCost; | |||
2625 | ||||
2626 | std::string Str; | |||
2627 | { | |||
2628 | raw_string_ostream OS(Str); | |||
2629 | OS << "SLP: Spill Cost = " << SpillCost << ".\n" | |||
2630 | << "SLP: Extract Cost = " << ExtractCost << ".\n" | |||
2631 | << "SLP: Total Cost = " << Cost << ".\n"; | |||
2632 | } | |||
2633 | DEBUG(dbgs() << Str)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << Str; } } while (false); | |||
2634 | ||||
2635 | if (ViewSLPTree) | |||
2636 | ViewGraph(this, "SLP" + F->getName(), false, Str); | |||
2637 | ||||
2638 | return Cost; | |||
2639 | } | |||
2640 | ||||
2641 | int BoUpSLP::getGatherCost(Type *Ty, | |||
2642 | const DenseSet<unsigned> &ShuffledIndices) { | |||
2643 | int Cost = 0; | |||
2644 | for (unsigned i = 0, e = cast<VectorType>(Ty)->getNumElements(); i < e; ++i) | |||
2645 | if (!ShuffledIndices.count(i)) | |||
2646 | Cost += TTI->getVectorInstrCost(Instruction::InsertElement, Ty, i); | |||
2647 | if (!ShuffledIndices.empty()) | |||
2648 | Cost += TTI->getShuffleCost(TargetTransformInfo::SK_PermuteSingleSrc, Ty); | |||
2649 | return Cost; | |||
2650 | } | |||
2651 | ||||
2652 | int BoUpSLP::getGatherCost(ArrayRef<Value *> VL) { | |||
2653 | // Find the type of the operands in VL. | |||
2654 | Type *ScalarTy = VL[0]->getType(); | |||
2655 | if (StoreInst *SI = dyn_cast<StoreInst>(VL[0])) | |||
2656 | ScalarTy = SI->getValueOperand()->getType(); | |||
2657 | VectorType *VecTy = VectorType::get(ScalarTy, VL.size()); | |||
2658 | // Find the cost of inserting/extracting values from the vector. | |||
2659 | // Check if the same elements are inserted several times and count them as | |||
2660 | // shuffle candidates. | |||
2661 | DenseSet<unsigned> ShuffledElements; | |||
2662 | DenseSet<Value *> UniqueElements; | |||
2663 | // Iterate in reverse order to consider insert elements with the high cost. | |||
2664 | for (unsigned I = VL.size(); I > 0; --I) { | |||
2665 | unsigned Idx = I - 1; | |||
2666 | if (!UniqueElements.insert(VL[Idx]).second) | |||
2667 | ShuffledElements.insert(Idx); | |||
2668 | } | |||
2669 | return getGatherCost(VecTy, ShuffledElements); | |||
2670 | } | |||
2671 | ||||
2672 | // Reorder commutative operations in alternate shuffle if the resulting vectors | |||
2673 | // are consecutive loads. This would allow us to vectorize the tree. | |||
2674 | // If we have something like- | |||
2675 | // load a[0] - load b[0] | |||
2676 | // load b[1] + load a[1] | |||
2677 | // load a[2] - load b[2] | |||
2678 | // load a[3] + load b[3] | |||
2679 | // Reordering the second load b[1] load a[1] would allow us to vectorize this | |||
2680 | // code. | |||
2681 | void BoUpSLP::reorderAltShuffleOperands(unsigned Opcode, ArrayRef<Value *> VL, | |||
2682 | SmallVectorImpl<Value *> &Left, | |||
2683 | SmallVectorImpl<Value *> &Right) { | |||
2684 | // Push left and right operands of binary operation into Left and Right | |||
2685 | unsigned AltOpcode = getAltOpcode(Opcode); | |||
2686 | (void)AltOpcode; | |||
2687 | for (Value *V : VL) { | |||
2688 | auto *I = cast<Instruction>(V); | |||
2689 | assert(sameOpcodeOrAlt(Opcode, AltOpcode, I->getOpcode()) &&(static_cast <bool> (sameOpcodeOrAlt(Opcode, AltOpcode, I->getOpcode()) && "Incorrect instruction in vector" ) ? void (0) : __assert_fail ("sameOpcodeOrAlt(Opcode, AltOpcode, I->getOpcode()) && \"Incorrect instruction in vector\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2690, __extension__ __PRETTY_FUNCTION__)) | |||
2690 | "Incorrect instruction in vector")(static_cast <bool> (sameOpcodeOrAlt(Opcode, AltOpcode, I->getOpcode()) && "Incorrect instruction in vector" ) ? void (0) : __assert_fail ("sameOpcodeOrAlt(Opcode, AltOpcode, I->getOpcode()) && \"Incorrect instruction in vector\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2690, __extension__ __PRETTY_FUNCTION__)); | |||
2691 | Left.push_back(I->getOperand(0)); | |||
2692 | Right.push_back(I->getOperand(1)); | |||
2693 | } | |||
2694 | ||||
2695 | // Reorder if we have a commutative operation and consecutive access | |||
2696 | // are on either side of the alternate instructions. | |||
2697 | for (unsigned j = 0; j < VL.size() - 1; ++j) { | |||
2698 | if (LoadInst *L = dyn_cast<LoadInst>(Left[j])) { | |||
2699 | if (LoadInst *L1 = dyn_cast<LoadInst>(Right[j + 1])) { | |||
2700 | Instruction *VL1 = cast<Instruction>(VL[j]); | |||
2701 | Instruction *VL2 = cast<Instruction>(VL[j + 1]); | |||
2702 | if (VL1->isCommutative() && isConsecutiveAccess(L, L1, *DL, *SE)) { | |||
2703 | std::swap(Left[j], Right[j]); | |||
2704 | continue; | |||
2705 | } else if (VL2->isCommutative() && | |||
2706 | isConsecutiveAccess(L, L1, *DL, *SE)) { | |||
2707 | std::swap(Left[j + 1], Right[j + 1]); | |||
2708 | continue; | |||
2709 | } | |||
2710 | // else unchanged | |||
2711 | } | |||
2712 | } | |||
2713 | if (LoadInst *L = dyn_cast<LoadInst>(Right[j])) { | |||
2714 | if (LoadInst *L1 = dyn_cast<LoadInst>(Left[j + 1])) { | |||
2715 | Instruction *VL1 = cast<Instruction>(VL[j]); | |||
2716 | Instruction *VL2 = cast<Instruction>(VL[j + 1]); | |||
2717 | if (VL1->isCommutative() && isConsecutiveAccess(L, L1, *DL, *SE)) { | |||
2718 | std::swap(Left[j], Right[j]); | |||
2719 | continue; | |||
2720 | } else if (VL2->isCommutative() && | |||
2721 | isConsecutiveAccess(L, L1, *DL, *SE)) { | |||
2722 | std::swap(Left[j + 1], Right[j + 1]); | |||
2723 | continue; | |||
2724 | } | |||
2725 | // else unchanged | |||
2726 | } | |||
2727 | } | |||
2728 | } | |||
2729 | } | |||
2730 | ||||
2731 | // Return true if I should be commuted before adding it's left and right | |||
2732 | // operands to the arrays Left and Right. | |||
2733 | // | |||
2734 | // The vectorizer is trying to either have all elements one side being | |||
2735 | // instruction with the same opcode to enable further vectorization, or having | |||
2736 | // a splat to lower the vectorizing cost. | |||
2737 | static bool shouldReorderOperands( | |||
2738 | int i, unsigned Opcode, Instruction &I, ArrayRef<Value *> Left, | |||
2739 | ArrayRef<Value *> Right, bool AllSameOpcodeLeft, bool AllSameOpcodeRight, | |||
2740 | bool SplatLeft, bool SplatRight, Value *&VLeft, Value *&VRight) { | |||
2741 | VLeft = I.getOperand(0); | |||
2742 | VRight = I.getOperand(1); | |||
2743 | // If we have "SplatRight", try to see if commuting is needed to preserve it. | |||
2744 | if (SplatRight) { | |||
2745 | if (VRight == Right[i - 1]) | |||
2746 | // Preserve SplatRight | |||
2747 | return false; | |||
2748 | if (VLeft == Right[i - 1]) { | |||
2749 | // Commuting would preserve SplatRight, but we don't want to break | |||
2750 | // SplatLeft either, i.e. preserve the original order if possible. | |||
2751 | // (FIXME: why do we care?) | |||
2752 | if (SplatLeft && VLeft == Left[i - 1]) | |||
2753 | return false; | |||
2754 | return true; | |||
2755 | } | |||
2756 | } | |||
2757 | // Symmetrically handle Right side. | |||
2758 | if (SplatLeft) { | |||
2759 | if (VLeft == Left[i - 1]) | |||
2760 | // Preserve SplatLeft | |||
2761 | return false; | |||
2762 | if (VRight == Left[i - 1]) | |||
2763 | return true; | |||
2764 | } | |||
2765 | ||||
2766 | Instruction *ILeft = dyn_cast<Instruction>(VLeft); | |||
2767 | Instruction *IRight = dyn_cast<Instruction>(VRight); | |||
2768 | ||||
2769 | // If we have "AllSameOpcodeRight", try to see if the left operands preserves | |||
2770 | // it and not the right, in this case we want to commute. | |||
2771 | if (AllSameOpcodeRight) { | |||
2772 | unsigned RightPrevOpcode = cast<Instruction>(Right[i - 1])->getOpcode(); | |||
2773 | if (IRight && RightPrevOpcode == IRight->getOpcode()) | |||
2774 | // Do not commute, a match on the right preserves AllSameOpcodeRight | |||
2775 | return false; | |||
2776 | if (ILeft && RightPrevOpcode == ILeft->getOpcode()) { | |||
2777 | // We have a match and may want to commute, but first check if there is | |||
2778 | // not also a match on the existing operands on the Left to preserve | |||
2779 | // AllSameOpcodeLeft, i.e. preserve the original order if possible. | |||
2780 | // (FIXME: why do we care?) | |||
2781 | if (AllSameOpcodeLeft && ILeft && | |||
2782 | cast<Instruction>(Left[i - 1])->getOpcode() == ILeft->getOpcode()) | |||
2783 | return false; | |||
2784 | return true; | |||
2785 | } | |||
2786 | } | |||
2787 | // Symmetrically handle Left side. | |||
2788 | if (AllSameOpcodeLeft) { | |||
2789 | unsigned LeftPrevOpcode = cast<Instruction>(Left[i - 1])->getOpcode(); | |||
2790 | if (ILeft && LeftPrevOpcode == ILeft->getOpcode()) | |||
2791 | return false; | |||
2792 | if (IRight && LeftPrevOpcode == IRight->getOpcode()) | |||
2793 | return true; | |||
2794 | } | |||
2795 | return false; | |||
2796 | } | |||
2797 | ||||
2798 | void BoUpSLP::reorderInputsAccordingToOpcode(unsigned Opcode, | |||
2799 | ArrayRef<Value *> VL, | |||
2800 | SmallVectorImpl<Value *> &Left, | |||
2801 | SmallVectorImpl<Value *> &Right) { | |||
2802 | if (!VL.empty()) { | |||
2803 | // Peel the first iteration out of the loop since there's nothing | |||
2804 | // interesting to do anyway and it simplifies the checks in the loop. | |||
2805 | auto *I = cast<Instruction>(VL[0]); | |||
2806 | Value *VLeft = I->getOperand(0); | |||
2807 | Value *VRight = I->getOperand(1); | |||
2808 | if (!isa<Instruction>(VRight) && isa<Instruction>(VLeft)) | |||
2809 | // Favor having instruction to the right. FIXME: why? | |||
2810 | std::swap(VLeft, VRight); | |||
2811 | Left.push_back(VLeft); | |||
2812 | Right.push_back(VRight); | |||
2813 | } | |||
2814 | ||||
2815 | // Keep track if we have instructions with all the same opcode on one side. | |||
2816 | bool AllSameOpcodeLeft = isa<Instruction>(Left[0]); | |||
2817 | bool AllSameOpcodeRight = isa<Instruction>(Right[0]); | |||
2818 | // Keep track if we have one side with all the same value (broadcast). | |||
2819 | bool SplatLeft = true; | |||
2820 | bool SplatRight = true; | |||
2821 | ||||
2822 | for (unsigned i = 1, e = VL.size(); i != e; ++i) { | |||
2823 | Instruction *I = cast<Instruction>(VL[i]); | |||
2824 | assert(((I->getOpcode() == Opcode && I->isCommutative()) ||(static_cast <bool> (((I->getOpcode() == Opcode && I->isCommutative()) || (I->getOpcode() != Opcode && Instruction::isCommutative(Opcode))) && "Can only process commutative instruction" ) ? void (0) : __assert_fail ("((I->getOpcode() == Opcode && I->isCommutative()) || (I->getOpcode() != Opcode && Instruction::isCommutative(Opcode))) && \"Can only process commutative instruction\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2826, __extension__ __PRETTY_FUNCTION__)) | |||
2825 | (I->getOpcode() != Opcode && Instruction::isCommutative(Opcode))) &&(static_cast <bool> (((I->getOpcode() == Opcode && I->isCommutative()) || (I->getOpcode() != Opcode && Instruction::isCommutative(Opcode))) && "Can only process commutative instruction" ) ? void (0) : __assert_fail ("((I->getOpcode() == Opcode && I->isCommutative()) || (I->getOpcode() != Opcode && Instruction::isCommutative(Opcode))) && \"Can only process commutative instruction\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2826, __extension__ __PRETTY_FUNCTION__)) | |||
2826 | "Can only process commutative instruction")(static_cast <bool> (((I->getOpcode() == Opcode && I->isCommutative()) || (I->getOpcode() != Opcode && Instruction::isCommutative(Opcode))) && "Can only process commutative instruction" ) ? void (0) : __assert_fail ("((I->getOpcode() == Opcode && I->isCommutative()) || (I->getOpcode() != Opcode && Instruction::isCommutative(Opcode))) && \"Can only process commutative instruction\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2826, __extension__ __PRETTY_FUNCTION__)); | |||
2827 | // Commute to favor either a splat or maximizing having the same opcodes on | |||
2828 | // one side. | |||
2829 | Value *VLeft; | |||
2830 | Value *VRight; | |||
2831 | if (shouldReorderOperands(i, Opcode, *I, Left, Right, AllSameOpcodeLeft, | |||
2832 | AllSameOpcodeRight, SplatLeft, SplatRight, VLeft, | |||
2833 | VRight)) { | |||
2834 | Left.push_back(VRight); | |||
2835 | Right.push_back(VLeft); | |||
2836 | } else { | |||
2837 | Left.push_back(VLeft); | |||
2838 | Right.push_back(VRight); | |||
2839 | } | |||
2840 | // Update Splat* and AllSameOpcode* after the insertion. | |||
2841 | SplatRight = SplatRight && (Right[i - 1] == Right[i]); | |||
2842 | SplatLeft = SplatLeft && (Left[i - 1] == Left[i]); | |||
2843 | AllSameOpcodeLeft = AllSameOpcodeLeft && isa<Instruction>(Left[i]) && | |||
2844 | (cast<Instruction>(Left[i - 1])->getOpcode() == | |||
2845 | cast<Instruction>(Left[i])->getOpcode()); | |||
2846 | AllSameOpcodeRight = AllSameOpcodeRight && isa<Instruction>(Right[i]) && | |||
2847 | (cast<Instruction>(Right[i - 1])->getOpcode() == | |||
2848 | cast<Instruction>(Right[i])->getOpcode()); | |||
2849 | } | |||
2850 | ||||
2851 | // If one operand end up being broadcast, return this operand order. | |||
2852 | if (SplatRight || SplatLeft) | |||
2853 | return; | |||
2854 | ||||
2855 | // Finally check if we can get longer vectorizable chain by reordering | |||
2856 | // without breaking the good operand order detected above. | |||
2857 | // E.g. If we have something like- | |||
2858 | // load a[0] load b[0] | |||
2859 | // load b[1] load a[1] | |||
2860 | // load a[2] load b[2] | |||
2861 | // load a[3] load b[3] | |||
2862 | // Reordering the second load b[1] load a[1] would allow us to vectorize | |||
2863 | // this code and we still retain AllSameOpcode property. | |||
2864 | // FIXME: This load reordering might break AllSameOpcode in some rare cases | |||
2865 | // such as- | |||
2866 | // add a[0],c[0] load b[0] | |||
2867 | // add a[1],c[2] load b[1] | |||
2868 | // b[2] load b[2] | |||
2869 | // add a[3],c[3] load b[3] | |||
2870 | for (unsigned j = 0; j < VL.size() - 1; ++j) { | |||
2871 | if (LoadInst *L = dyn_cast<LoadInst>(Left[j])) { | |||
2872 | if (LoadInst *L1 = dyn_cast<LoadInst>(Right[j + 1])) { | |||
2873 | if (isConsecutiveAccess(L, L1, *DL, *SE)) { | |||
2874 | std::swap(Left[j + 1], Right[j + 1]); | |||
2875 | continue; | |||
2876 | } | |||
2877 | } | |||
2878 | } | |||
2879 | if (LoadInst *L = dyn_cast<LoadInst>(Right[j])) { | |||
2880 | if (LoadInst *L1 = dyn_cast<LoadInst>(Left[j + 1])) { | |||
2881 | if (isConsecutiveAccess(L, L1, *DL, *SE)) { | |||
2882 | std::swap(Left[j + 1], Right[j + 1]); | |||
2883 | continue; | |||
2884 | } | |||
2885 | } | |||
2886 | } | |||
2887 | // else unchanged | |||
2888 | } | |||
2889 | } | |||
2890 | ||||
2891 | void BoUpSLP::setInsertPointAfterBundle(ArrayRef<Value *> VL, Value *OpValue) { | |||
2892 | // Get the basic block this bundle is in. All instructions in the bundle | |||
2893 | // should be in this block. | |||
2894 | auto *Front = cast<Instruction>(OpValue); | |||
2895 | auto *BB = Front->getParent(); | |||
2896 | const unsigned Opcode = cast<Instruction>(OpValue)->getOpcode(); | |||
2897 | const unsigned AltOpcode = getAltOpcode(Opcode); | |||
2898 | assert(llvm::all_of(make_range(VL.begin(), VL.end()), [=](Value *V) -> bool {(static_cast <bool> (llvm::all_of(make_range(VL.begin() , VL.end()), [=](Value *V) -> bool { return !sameOpcodeOrAlt (Opcode, AltOpcode, cast<Instruction>(V)->getOpcode( )) || cast<Instruction>(V)->getParent() == BB; })) ? void (0) : __assert_fail ("llvm::all_of(make_range(VL.begin(), VL.end()), [=](Value *V) -> bool { return !sameOpcodeOrAlt(Opcode, AltOpcode, cast<Instruction>(V)->getOpcode()) || cast<Instruction>(V)->getParent() == BB; })" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2902, __extension__ __PRETTY_FUNCTION__)) | |||
2899 | return !sameOpcodeOrAlt(Opcode, AltOpcode,(static_cast <bool> (llvm::all_of(make_range(VL.begin() , VL.end()), [=](Value *V) -> bool { return !sameOpcodeOrAlt (Opcode, AltOpcode, cast<Instruction>(V)->getOpcode( )) || cast<Instruction>(V)->getParent() == BB; })) ? void (0) : __assert_fail ("llvm::all_of(make_range(VL.begin(), VL.end()), [=](Value *V) -> bool { return !sameOpcodeOrAlt(Opcode, AltOpcode, cast<Instruction>(V)->getOpcode()) || cast<Instruction>(V)->getParent() == BB; })" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2902, __extension__ __PRETTY_FUNCTION__)) | |||
2900 | cast<Instruction>(V)->getOpcode()) ||(static_cast <bool> (llvm::all_of(make_range(VL.begin() , VL.end()), [=](Value *V) -> bool { return !sameOpcodeOrAlt (Opcode, AltOpcode, cast<Instruction>(V)->getOpcode( )) || cast<Instruction>(V)->getParent() == BB; })) ? void (0) : __assert_fail ("llvm::all_of(make_range(VL.begin(), VL.end()), [=](Value *V) -> bool { return !sameOpcodeOrAlt(Opcode, AltOpcode, cast<Instruction>(V)->getOpcode()) || cast<Instruction>(V)->getParent() == BB; })" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2902, __extension__ __PRETTY_FUNCTION__)) | |||
2901 | cast<Instruction>(V)->getParent() == BB;(static_cast <bool> (llvm::all_of(make_range(VL.begin() , VL.end()), [=](Value *V) -> bool { return !sameOpcodeOrAlt (Opcode, AltOpcode, cast<Instruction>(V)->getOpcode( )) || cast<Instruction>(V)->getParent() == BB; })) ? void (0) : __assert_fail ("llvm::all_of(make_range(VL.begin(), VL.end()), [=](Value *V) -> bool { return !sameOpcodeOrAlt(Opcode, AltOpcode, cast<Instruction>(V)->getOpcode()) || cast<Instruction>(V)->getParent() == BB; })" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2902, __extension__ __PRETTY_FUNCTION__)) | |||
2902 | }))(static_cast <bool> (llvm::all_of(make_range(VL.begin() , VL.end()), [=](Value *V) -> bool { return !sameOpcodeOrAlt (Opcode, AltOpcode, cast<Instruction>(V)->getOpcode( )) || cast<Instruction>(V)->getParent() == BB; })) ? void (0) : __assert_fail ("llvm::all_of(make_range(VL.begin(), VL.end()), [=](Value *V) -> bool { return !sameOpcodeOrAlt(Opcode, AltOpcode, cast<Instruction>(V)->getOpcode()) || cast<Instruction>(V)->getParent() == BB; })" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2902, __extension__ __PRETTY_FUNCTION__)); | |||
2903 | ||||
2904 | // The last instruction in the bundle in program order. | |||
2905 | Instruction *LastInst = nullptr; | |||
2906 | ||||
2907 | // Find the last instruction. The common case should be that BB has been | |||
2908 | // scheduled, and the last instruction is VL.back(). So we start with | |||
2909 | // VL.back() and iterate over schedule data until we reach the end of the | |||
2910 | // bundle. The end of the bundle is marked by null ScheduleData. | |||
2911 | if (BlocksSchedules.count(BB)) { | |||
2912 | auto *Bundle = | |||
2913 | BlocksSchedules[BB]->getScheduleData(isOneOf(OpValue, VL.back())); | |||
2914 | if (Bundle && Bundle->isPartOfBundle()) | |||
2915 | for (; Bundle; Bundle = Bundle->NextInBundle) | |||
2916 | if (Bundle->OpValue == Bundle->Inst) | |||
2917 | LastInst = Bundle->Inst; | |||
2918 | } | |||
2919 | ||||
2920 | // LastInst can still be null at this point if there's either not an entry | |||
2921 | // for BB in BlocksSchedules or there's no ScheduleData available for | |||
2922 | // VL.back(). This can be the case if buildTree_rec aborts for various | |||
2923 | // reasons (e.g., the maximum recursion depth is reached, the maximum region | |||
2924 | // size is reached, etc.). ScheduleData is initialized in the scheduling | |||
2925 | // "dry-run". | |||
2926 | // | |||
2927 | // If this happens, we can still find the last instruction by brute force. We | |||
2928 | // iterate forwards from Front (inclusive) until we either see all | |||
2929 | // instructions in the bundle or reach the end of the block. If Front is the | |||
2930 | // last instruction in program order, LastInst will be set to Front, and we | |||
2931 | // will visit all the remaining instructions in the block. | |||
2932 | // | |||
2933 | // One of the reasons we exit early from buildTree_rec is to place an upper | |||
2934 | // bound on compile-time. Thus, taking an additional compile-time hit here is | |||
2935 | // not ideal. However, this should be exceedingly rare since it requires that | |||
2936 | // we both exit early from buildTree_rec and that the bundle be out-of-order | |||
2937 | // (causing us to iterate all the way to the end of the block). | |||
2938 | if (!LastInst) { | |||
2939 | SmallPtrSet<Value *, 16> Bundle(VL.begin(), VL.end()); | |||
2940 | for (auto &I : make_range(BasicBlock::iterator(Front), BB->end())) { | |||
2941 | if (Bundle.erase(&I) && sameOpcodeOrAlt(Opcode, AltOpcode, I.getOpcode())) | |||
2942 | LastInst = &I; | |||
2943 | if (Bundle.empty()) | |||
2944 | break; | |||
2945 | } | |||
2946 | } | |||
2947 | ||||
2948 | // Set the insertion point after the last instruction in the bundle. Set the | |||
2949 | // debug location to Front. | |||
2950 | Builder.SetInsertPoint(BB, ++LastInst->getIterator()); | |||
| ||||
2951 | Builder.SetCurrentDebugLocation(Front->getDebugLoc()); | |||
2952 | } | |||
2953 | ||||
2954 | Value *BoUpSLP::Gather(ArrayRef<Value *> VL, VectorType *Ty) { | |||
2955 | Value *Vec = UndefValue::get(Ty); | |||
2956 | // Generate the 'InsertElement' instruction. | |||
2957 | for (unsigned i = 0; i < Ty->getNumElements(); ++i) { | |||
2958 | Vec = Builder.CreateInsertElement(Vec, VL[i], Builder.getInt32(i)); | |||
2959 | if (Instruction *Insrt = dyn_cast<Instruction>(Vec)) { | |||
2960 | GatherSeq.insert(Insrt); | |||
2961 | CSEBlocks.insert(Insrt->getParent()); | |||
2962 | ||||
2963 | // Add to our 'need-to-extract' list. | |||
2964 | if (TreeEntry *E = getTreeEntry(VL[i])) { | |||
2965 | // Find which lane we need to extract. | |||
2966 | int FoundLane = -1; | |||
2967 | for (unsigned Lane = 0, LE = E->Scalars.size(); Lane != LE; ++Lane) { | |||
2968 | // Is this the lane of the scalar that we are looking for ? | |||
2969 | if (E->Scalars[Lane] == VL[i]) { | |||
2970 | FoundLane = Lane; | |||
2971 | break; | |||
2972 | } | |||
2973 | } | |||
2974 | assert(FoundLane >= 0 && "Could not find the correct lane")(static_cast <bool> (FoundLane >= 0 && "Could not find the correct lane" ) ? void (0) : __assert_fail ("FoundLane >= 0 && \"Could not find the correct lane\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2974, __extension__ __PRETTY_FUNCTION__)); | |||
2975 | if (!E->ReuseShuffleIndices.empty()) { | |||
2976 | FoundLane = | |||
2977 | std::distance(E->ReuseShuffleIndices.begin(), | |||
2978 | llvm::find(E->ReuseShuffleIndices, FoundLane)); | |||
2979 | } | |||
2980 | ExternalUses.push_back(ExternalUser(VL[i], Insrt, FoundLane)); | |||
2981 | } | |||
2982 | } | |||
2983 | } | |||
2984 | ||||
2985 | return Vec; | |||
2986 | } | |||
2987 | ||||
2988 | Value *BoUpSLP::vectorizeTree(ArrayRef<Value *> VL) { | |||
2989 | InstructionsState S = getSameOpcode(VL); | |||
2990 | if (S.Opcode) { | |||
2991 | if (TreeEntry *E = getTreeEntry(S.OpValue)) { | |||
2992 | if (E->isSame(VL)) { | |||
2993 | Value *V = vectorizeTree(E); | |||
2994 | if (VL.size() == E->Scalars.size() && !E->ReuseShuffleIndices.empty()) { | |||
2995 | // We need to get the vectorized value but without shuffle. | |||
2996 | if (auto *SV = dyn_cast<ShuffleVectorInst>(V)) { | |||
2997 | V = SV->getOperand(0); | |||
2998 | } else { | |||
2999 | // Reshuffle to get only unique values. | |||
3000 | SmallVector<unsigned, 4> UniqueIdxs; | |||
3001 | SmallSet<unsigned, 4> UsedIdxs; | |||
3002 | for(unsigned Idx : E->ReuseShuffleIndices) | |||
3003 | if (UsedIdxs.insert(Idx).second) | |||
3004 | UniqueIdxs.emplace_back(Idx); | |||
3005 | V = Builder.CreateShuffleVector(V, UndefValue::get(V->getType()), | |||
3006 | UniqueIdxs); | |||
3007 | } | |||
3008 | } | |||
3009 | return V; | |||
3010 | } | |||
3011 | } | |||
3012 | } | |||
3013 | ||||
3014 | Type *ScalarTy = S.OpValue->getType(); | |||
3015 | if (StoreInst *SI = dyn_cast<StoreInst>(S.OpValue)) | |||
3016 | ScalarTy = SI->getValueOperand()->getType(); | |||
3017 | ||||
3018 | // Check that every instruction appears once in this bundle. | |||
3019 | SmallVector<unsigned, 4> ReuseShuffleIndicies; | |||
3020 | SmallVector<Value *, 4> UniqueValues; | |||
3021 | if (VL.size() > 2) { | |||
3022 | DenseMap<Value *, unsigned> UniquePositions; | |||
3023 | for (Value *V : VL) { | |||
3024 | auto Res = UniquePositions.try_emplace(V, UniqueValues.size()); | |||
3025 | ReuseShuffleIndicies.emplace_back(Res.first->second); | |||
3026 | if (Res.second || isa<Constant>(V)) | |||
3027 | UniqueValues.emplace_back(V); | |||
3028 | } | |||
3029 | // Do not shuffle single element or if number of unique values is not power | |||
3030 | // of 2. | |||
3031 | if (UniqueValues.size() == VL.size() || UniqueValues.size() <= 1 || | |||
3032 | !llvm::isPowerOf2_32(UniqueValues.size())) | |||
3033 | ReuseShuffleIndicies.clear(); | |||
3034 | else | |||
3035 | VL = UniqueValues; | |||
3036 | } | |||
3037 | VectorType *VecTy = VectorType::get(ScalarTy, VL.size()); | |||
3038 | ||||
3039 | Value *V = Gather(VL, VecTy); | |||
3040 | if (!ReuseShuffleIndicies.empty()) { | |||
3041 | V = Builder.CreateShuffleVector(V, UndefValue::get(VecTy), | |||
3042 | ReuseShuffleIndicies, "shuffle"); | |||
3043 | if (auto *I = dyn_cast<Instruction>(V)) { | |||
3044 | GatherSeq.insert(I); | |||
3045 | CSEBlocks.insert(I->getParent()); | |||
3046 | } | |||
3047 | } | |||
3048 | return V; | |||
3049 | } | |||
3050 | ||||
3051 | static void inversePermutation(ArrayRef<unsigned> Indices, | |||
3052 | SmallVectorImpl<unsigned> &Mask) { | |||
3053 | Mask.clear(); | |||
3054 | const unsigned E = Indices.size(); | |||
3055 | Mask.resize(E); | |||
3056 | for (unsigned I = 0; I < E; ++I) | |||
3057 | Mask[Indices[I]] = I; | |||
3058 | } | |||
3059 | ||||
3060 | Value *BoUpSLP::vectorizeTree(TreeEntry *E) { | |||
3061 | IRBuilder<>::InsertPointGuard Guard(Builder); | |||
3062 | ||||
3063 | if (E->VectorizedValue) { | |||
| ||||
3064 | 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); | |||
3065 | return E->VectorizedValue; | |||
3066 | } | |||
3067 | ||||
3068 | InstructionsState S = getSameOpcode(E->Scalars); | |||
3069 | Instruction *VL0 = cast<Instruction>(E->Scalars[0]); | |||
3070 | Type *ScalarTy = VL0->getType(); | |||
3071 | if (StoreInst *SI = dyn_cast<StoreInst>(VL0)) | |||
3072 | ScalarTy = SI->getValueOperand()->getType(); | |||
3073 | VectorType *VecTy = VectorType::get(ScalarTy, E->Scalars.size()); | |||
3074 | ||||
3075 | bool NeedToShuffleReuses = !E->ReuseShuffleIndices.empty(); | |||
3076 | ||||
3077 | if (E->NeedToGather) { | |||
3078 | setInsertPointAfterBundle(E->Scalars, VL0); | |||
3079 | auto *V = Gather(E->Scalars, VecTy); | |||
3080 | if (NeedToShuffleReuses) { | |||
3081 | V = Builder.CreateShuffleVector(V, UndefValue::get(VecTy), | |||
3082 | E->ReuseShuffleIndices, "shuffle"); | |||
3083 | if (auto *I = dyn_cast<Instruction>(V)) { | |||
3084 | GatherSeq.insert(I); | |||
3085 | CSEBlocks.insert(I->getParent()); | |||
3086 | } | |||
3087 | } | |||
3088 | E->VectorizedValue = V; | |||
3089 | return V; | |||
3090 | } | |||
3091 | ||||
3092 | unsigned ShuffleOrOp = S.IsAltShuffle ? | |||
3093 | (unsigned) Instruction::ShuffleVector : S.Opcode; | |||
3094 | switch (ShuffleOrOp) { | |||
3095 | case Instruction::PHI: { | |||
3096 | PHINode *PH = dyn_cast<PHINode>(VL0); | |||
3097 | Builder.SetInsertPoint(PH->getParent()->getFirstNonPHI()); | |||
3098 | Builder.SetCurrentDebugLocation(PH->getDebugLoc()); | |||
3099 | PHINode *NewPhi = Builder.CreatePHI(VecTy, PH->getNumIncomingValues()); | |||
3100 | Value *V = NewPhi; | |||
3101 | if (NeedToShuffleReuses) { | |||
3102 | V = Builder.CreateShuffleVector(V, UndefValue::get(VecTy), | |||
3103 | E->ReuseShuffleIndices, "shuffle"); | |||
3104 | } | |||
3105 | E->VectorizedValue = V; | |||
3106 | ||||
3107 | // PHINodes may have multiple entries from the same block. We want to | |||
3108 | // visit every block once. | |||
3109 | SmallSet<BasicBlock*, 4> VisitedBBs; | |||
3110 | ||||
3111 | for (unsigned i = 0, e = PH->getNumIncomingValues(); i < e; ++i) { | |||
3112 | ValueList Operands; | |||
3113 | BasicBlock *IBB = PH->getIncomingBlock(i); | |||
3114 | ||||
3115 | if (!VisitedBBs.insert(IBB).second) { | |||
3116 | NewPhi->addIncoming(NewPhi->getIncomingValueForBlock(IBB), IBB); | |||
3117 | continue; | |||
3118 | } | |||
3119 | ||||
3120 | // Prepare the operand vector. | |||
3121 | for (Value *V : E->Scalars) | |||
3122 | Operands.push_back(cast<PHINode>(V)->getIncomingValueForBlock(IBB)); | |||
3123 | ||||
3124 | Builder.SetInsertPoint(IBB->getTerminator()); | |||
3125 | Builder.SetCurrentDebugLocation(PH->getDebugLoc()); | |||
3126 | Value *Vec = vectorizeTree(Operands); | |||
3127 | NewPhi->addIncoming(Vec, IBB); | |||
3128 | } | |||
3129 | ||||
3130 | assert(NewPhi->getNumIncomingValues() == PH->getNumIncomingValues() &&(static_cast <bool> (NewPhi->getNumIncomingValues() == PH->getNumIncomingValues() && "Invalid number of incoming values" ) ? void (0) : __assert_fail ("NewPhi->getNumIncomingValues() == PH->getNumIncomingValues() && \"Invalid number of incoming values\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3131, __extension__ __PRETTY_FUNCTION__)) | |||
3131 | "Invalid number of incoming values")(static_cast <bool> (NewPhi->getNumIncomingValues() == PH->getNumIncomingValues() && "Invalid number of incoming values" ) ? void (0) : __assert_fail ("NewPhi->getNumIncomingValues() == PH->getNumIncomingValues() && \"Invalid number of incoming values\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3131, __extension__ __PRETTY_FUNCTION__)); | |||
3132 | return V; | |||
3133 | } | |||
3134 | ||||
3135 | case Instruction::ExtractElement: { | |||
3136 | if (!E->NeedToGather) { | |||
3137 | Value *V = VL0->getOperand(0); | |||
3138 | if (!E->ReorderIndices.empty()) { | |||
3139 | OrdersType Mask; | |||
3140 | inversePermutation(E->ReorderIndices, Mask); | |||
3141 | Builder.SetInsertPoint(VL0); | |||
3142 | V = Builder.CreateShuffleVector(V, UndefValue::get(VecTy), Mask, | |||
3143 | "reorder_shuffle"); | |||
3144 | } | |||
3145 | if (NeedToShuffleReuses) { | |||
3146 | // TODO: Merge this shuffle with the ReorderShuffleMask. | |||
3147 | if (!E->ReorderIndices.empty()) | |||
3148 | Builder.SetInsertPoint(VL0); | |||
3149 | V = Builder.CreateShuffleVector(V, UndefValue::get(VecTy), | |||
3150 | E->ReuseShuffleIndices, "shuffle"); | |||
3151 | } | |||
3152 | E->VectorizedValue = V; | |||
3153 | return V; | |||
3154 | } | |||
3155 | setInsertPointAfterBundle(E->Scalars, VL0); | |||
3156 | auto *V = Gather(E->Scalars, VecTy); | |||
3157 | if (NeedToShuffleReuses) { | |||
3158 | V = Builder.CreateShuffleVector(V, UndefValue::get(VecTy), | |||
3159 | E->ReuseShuffleIndices, "shuffle"); | |||
3160 | if (auto *I = dyn_cast<Instruction>(V)) { | |||
3161 | GatherSeq.insert(I); | |||
3162 | CSEBlocks.insert(I->getParent()); | |||
3163 | } | |||
3164 | } | |||
3165 | E->VectorizedValue = V; | |||
3166 | return V; | |||
3167 | } | |||
3168 | case Instruction::ExtractValue: { | |||
3169 | if (!E->NeedToGather) { | |||
3170 | LoadInst *LI = cast<LoadInst>(VL0->getOperand(0)); | |||
3171 | Builder.SetInsertPoint(LI); | |||
3172 | PointerType *PtrTy = PointerType::get(VecTy, LI->getPointerAddressSpace()); | |||
3173 | Value *Ptr = Builder.CreateBitCast(LI->getOperand(0), PtrTy); | |||
3174 | LoadInst *V = Builder.CreateAlignedLoad(Ptr, LI->getAlignment()); | |||
3175 | Value *NewV = propagateMetadata(V, E->Scalars); | |||
3176 | if (!E->ReorderIndices.empty()) { | |||
3177 | OrdersType Mask; | |||
3178 | inversePermutation(E->ReorderIndices, Mask); | |||
3179 | NewV = Builder.CreateShuffleVector(NewV, UndefValue::get(VecTy), Mask, | |||
3180 | "reorder_shuffle"); | |||
3181 | } | |||
3182 | if (NeedToShuffleReuses) { | |||
3183 | // TODO: Merge this shuffle with the ReorderShuffleMask. | |||
3184 | NewV = Builder.CreateShuffleVector( | |||
3185 | NewV, UndefValue::get(VecTy), E->ReuseShuffleIndices, "shuffle"); | |||
3186 | } | |||
3187 | E->VectorizedValue = NewV; | |||
3188 | return NewV; | |||
3189 | } | |||
3190 | setInsertPointAfterBundle(E->Scalars, VL0); | |||
3191 | auto *V = Gather(E->Scalars, VecTy); | |||
3192 | if (NeedToShuffleReuses) { | |||
3193 | V = Builder.CreateShuffleVector(V, UndefValue::get(VecTy), | |||
3194 | E->ReuseShuffleIndices, "shuffle"); | |||
3195 | if (auto *I = dyn_cast<Instruction>(V)) { | |||
3196 | GatherSeq.insert(I); | |||
3197 | CSEBlocks.insert(I->getParent()); | |||
3198 | } | |||
3199 | } | |||
3200 | E->VectorizedValue = V; | |||
3201 | return V; | |||
3202 | } | |||
3203 | case Instruction::ZExt: | |||
3204 | case Instruction::SExt: | |||
3205 | case Instruction::FPToUI: | |||
3206 | case Instruction::FPToSI: | |||
3207 | case Instruction::FPExt: | |||
3208 | case Instruction::PtrToInt: | |||
3209 | case Instruction::IntToPtr: | |||
3210 | case Instruction::SIToFP: | |||
3211 | case Instruction::UIToFP: | |||
3212 | case Instruction::Trunc: | |||
3213 | case Instruction::FPTrunc: | |||
3214 | case Instruction::BitCast: { | |||
3215 | ValueList INVL; | |||
3216 | for (Value *V : E->Scalars) | |||
3217 | INVL.push_back(cast<Instruction>(V)->getOperand(0)); | |||
3218 | ||||
3219 | setInsertPointAfterBundle(E->Scalars, VL0); | |||
3220 | ||||
3221 | Value *InVec = vectorizeTree(INVL); | |||
3222 | ||||
3223 | if (E->VectorizedValue) { | |||
3224 | DEBUG(dbgs() << "SLP: Diamond merged for " << *VL0 << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Diamond merged for " << *VL0 << ".\n"; } } while (false); | |||
3225 | return E->VectorizedValue; | |||
3226 | } | |||
3227 | ||||
3228 | CastInst *CI = dyn_cast<CastInst>(VL0); | |||
3229 | Value *V = Builder.CreateCast(CI->getOpcode(), InVec, VecTy); | |||
3230 | if (NeedToShuffleReuses) { | |||
3231 | V = Builder.CreateShuffleVector(V, UndefValue::get(VecTy), | |||
3232 | E->ReuseShuffleIndices, "shuffle"); | |||
3233 | } | |||
3234 | E->VectorizedValue = V; | |||
3235 | ++NumVectorInstructions; | |||
3236 | return V; | |||
3237 | } | |||
3238 | case Instruction::FCmp: | |||
3239 | case Instruction::ICmp: { | |||
3240 | ValueList LHSV, RHSV; | |||
3241 | for (Value *V : E->Scalars) { | |||
3242 | LHSV.push_back(cast<Instruction>(V)->getOperand(0)); | |||
3243 | RHSV.push_back(cast<Instruction>(V)->getOperand(1)); | |||
3244 | } | |||
3245 | ||||
3246 | setInsertPointAfterBundle(E->Scalars, VL0); | |||
3247 | ||||
3248 | Value *L = vectorizeTree(LHSV); | |||
3249 | Value *R = vectorizeTree(RHSV); | |||
3250 | ||||
3251 | if (E->VectorizedValue) { | |||
3252 | DEBUG(dbgs() << "SLP: Diamond merged for " << *VL0 << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Diamond merged for " << *VL0 << ".\n"; } } while (false); | |||
3253 | return E->VectorizedValue; | |||
3254 | } | |||
3255 | ||||
3256 | CmpInst::Predicate P0 = cast<CmpInst>(VL0)->getPredicate(); | |||
3257 | Value *V; | |||
3258 | if (S.Opcode == Instruction::FCmp) | |||
3259 | V = Builder.CreateFCmp(P0, L, R); | |||
3260 | else | |||
3261 | V = Builder.CreateICmp(P0, L, R); | |||
3262 | ||||
3263 | propagateIRFlags(V, E->Scalars, VL0); | |||
3264 | if (NeedToShuffleReuses) { | |||
3265 | V = Builder.CreateShuffleVector(V, UndefValue::get(VecTy), | |||
3266 | E->ReuseShuffleIndices, "shuffle"); | |||
3267 | } | |||
3268 | E->VectorizedValue = V; | |||
3269 | ++NumVectorInstructions; | |||
3270 | return V; | |||
3271 | } | |||
3272 | case Instruction::Select: { | |||
3273 | ValueList TrueVec, FalseVec, CondVec; | |||
3274 | for (Value *V : E->Scalars) { | |||
3275 | CondVec.push_back(cast<Instruction>(V)->getOperand(0)); | |||
3276 | TrueVec.push_back(cast<Instruction>(V)->getOperand(1)); | |||
3277 | FalseVec.push_back(cast<Instruction>(V)->getOperand(2)); | |||
3278 | } | |||
3279 | ||||
3280 | setInsertPointAfterBundle(E->Scalars, VL0); | |||
3281 | ||||
3282 | Value *Cond = vectorizeTree(CondVec); | |||
3283 | Value *True = vectorizeTree(TrueVec); | |||
3284 | Value *False = vectorizeTree(FalseVec); | |||
3285 | ||||
3286 | if (E->VectorizedValue) { | |||
3287 | DEBUG(dbgs() << "SLP: Diamond merged for " << *VL0 << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Diamond merged for " << *VL0 << ".\n"; } } while (false); | |||
3288 | return E->VectorizedValue; | |||
3289 | } | |||
3290 | ||||
3291 | Value *V = Builder.CreateSelect(Cond, True, False); | |||
3292 | if (NeedToShuffleReuses) { | |||
3293 | V = Builder.CreateShuffleVector(V, UndefValue::get(VecTy), | |||
3294 | E->ReuseShuffleIndices, "shuffle"); | |||
3295 | } | |||
3296 | E->VectorizedValue = V; | |||
3297 | ++NumVectorInstructions; | |||
3298 | return V; | |||
3299 | } | |||
3300 | case Instruction::Add: | |||
3301 | case Instruction::FAdd: | |||
3302 | case Instruction::Sub: | |||
3303 | case Instruction::FSub: | |||
3304 | case Instruction::Mul: | |||
3305 | case Instruction::FMul: | |||
3306 | case Instruction::UDiv: | |||
3307 | case Instruction::SDiv: | |||
3308 | case Instruction::FDiv: | |||
3309 | case Instruction::URem: | |||
3310 | case Instruction::SRem: | |||
3311 | case Instruction::FRem: | |||
3312 | case Instruction::Shl: | |||
3313 | case Instruction::LShr: | |||
3314 | case Instruction::AShr: | |||
3315 | case Instruction::And: | |||
3316 | case Instruction::Or: | |||
3317 | case Instruction::Xor: { | |||
3318 | ValueList LHSVL, RHSVL; | |||
3319 | if (isa<BinaryOperator>(VL0) && VL0->isCommutative()) | |||
3320 | reorderInputsAccordingToOpcode(S.Opcode, E->Scalars, LHSVL, | |||
3321 | RHSVL); | |||
3322 | else | |||
3323 | for (Value *V : E->Scalars) { | |||
3324 | auto *I = cast<Instruction>(V); | |||
3325 | LHSVL.push_back(I->getOperand(0)); | |||
3326 | RHSVL.push_back(I->getOperand(1)); | |||
3327 | } | |||
3328 | ||||
3329 | setInsertPointAfterBundle(E->Scalars, VL0); | |||
3330 | ||||
3331 | Value *LHS = vectorizeTree(LHSVL); | |||
3332 | Value *RHS = vectorizeTree(RHSVL); | |||
3333 | ||||
3334 | if (E->VectorizedValue) { | |||
3335 | DEBUG(dbgs() << "SLP: Diamond merged for " << *VL0 << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Diamond merged for " << *VL0 << ".\n"; } } while (false); | |||
3336 | return E->VectorizedValue; | |||
3337 | } | |||
3338 | ||||
3339 | Value *V = Builder.CreateBinOp( | |||
3340 | static_cast<Instruction::BinaryOps>(S.Opcode), LHS, RHS); | |||
3341 | propagateIRFlags(V, E->Scalars, VL0); | |||
3342 | if (auto *I = dyn_cast<Instruction>(V)) | |||
3343 | V = propagateMetadata(I, E->Scalars); | |||
3344 | ||||
3345 | if (NeedToShuffleReuses) { | |||
3346 | V = Builder.CreateShuffleVector(V, UndefValue::get(VecTy), | |||
3347 | E->ReuseShuffleIndices, "shuffle"); | |||
3348 | } | |||
3349 | E->VectorizedValue = V; | |||
3350 | ++NumVectorInstructions; | |||
3351 | ||||
3352 | return V; | |||
3353 | } | |||
3354 | case Instruction::Load: { | |||
3355 | // Loads are inserted at the head of the tree because we don't want to | |||
3356 | // sink them all the way down past store instructions. | |||
3357 | bool IsReorder = !E->ReorderIndices.empty(); | |||
3358 | if (IsReorder) | |||
3359 | VL0 = cast<Instruction>(E->Scalars[E->ReorderIndices.front()]); | |||
3360 | setInsertPointAfterBundle(E->Scalars, VL0); | |||
3361 | ||||
3362 | LoadInst *LI = cast<LoadInst>(VL0); | |||
3363 | Type *ScalarLoadTy = LI->getType(); | |||
3364 | unsigned AS = LI->getPointerAddressSpace(); | |||
3365 | ||||
3366 | Value *VecPtr = Builder.CreateBitCast(LI->getPointerOperand(), | |||
3367 | VecTy->getPointerTo(AS)); | |||
3368 | ||||
3369 | // The pointer operand uses an in-tree scalar so we add the new BitCast to | |||
3370 | // ExternalUses list to make sure that an extract will be generated in the | |||
3371 | // future. | |||
3372 | Value *PO = LI->getPointerOperand(); | |||
3373 | if (getTreeEntry(PO)) | |||
3374 | ExternalUses.push_back(ExternalUser(PO, cast<User>(VecPtr), 0)); | |||
3375 | ||||
3376 | unsigned Alignment = LI->getAlignment(); | |||
3377 | LI = Builder.CreateLoad(VecPtr); | |||
3378 | if (!Alignment) { | |||
3379 | Alignment = DL->getABITypeAlignment(ScalarLoadTy); | |||
3380 | } | |||
3381 | LI->setAlignment(Alignment); | |||
3382 | Value *V = propagateMetadata(LI, E->Scalars); | |||
3383 | if (IsReorder) { | |||
3384 | OrdersType Mask; | |||
3385 | inversePermutation(E->ReorderIndices, Mask); | |||
3386 | V = Builder.CreateShuffleVector(V, UndefValue::get(V->getType()), | |||
3387 | Mask, "reorder_shuffle"); | |||
3388 | } | |||
3389 | if (NeedToShuffleReuses) { | |||
3390 | // TODO: Merge this shuffle with the ReorderShuffleMask. | |||
3391 | V = Builder.CreateShuffleVector(V, UndefValue::get(VecTy), | |||
3392 | E->ReuseShuffleIndices, "shuffle"); | |||
3393 | } | |||
3394 | E->VectorizedValue = V; | |||
3395 | ++NumVectorInstructions; | |||
3396 | return V; | |||
3397 | } | |||
3398 | case Instruction::Store: { | |||
3399 | StoreInst *SI = cast<StoreInst>(VL0); | |||
3400 | unsigned Alignment = SI->getAlignment(); | |||
3401 | unsigned AS = SI->getPointerAddressSpace(); | |||
3402 | ||||
3403 | ValueList ScalarStoreValues; | |||
3404 | for (Value *V : E->Scalars) | |||
3405 | ScalarStoreValues.push_back(cast<StoreInst>(V)->getValueOperand()); | |||
3406 | ||||
3407 | setInsertPointAfterBundle(E->Scalars, VL0); | |||
3408 | ||||
3409 | Value *VecValue = vectorizeTree(ScalarStoreValues); | |||
3410 | Value *ScalarPtr = SI->getPointerOperand(); | |||
3411 | Value *VecPtr = Builder.CreateBitCast(ScalarPtr, VecTy->getPointerTo(AS)); | |||
3412 | StoreInst *S = Builder.CreateStore(VecValue, VecPtr); | |||
3413 | ||||
3414 | // The pointer operand uses an in-tree scalar, so add the new BitCast to | |||
3415 | // ExternalUses to make sure that an extract will be generated in the | |||
3416 | // future. | |||
3417 | if (getTreeEntry(ScalarPtr)) | |||
3418 | ExternalUses.push_back(ExternalUser(ScalarPtr, cast<User>(VecPtr), 0)); | |||
3419 | ||||
3420 | if (!Alignment) | |||
3421 | Alignment = DL->getABITypeAlignment(SI->getValueOperand()->getType()); | |||
3422 | ||||
3423 | S->setAlignment(Alignment); | |||
3424 | Value *V = propagateMetadata(S, E->Scalars); | |||
3425 | if (NeedToShuffleReuses) { | |||
3426 | V = Builder.CreateShuffleVector(V, UndefValue::get(VecTy), | |||
3427 | E->ReuseShuffleIndices, "shuffle"); | |||
3428 | } | |||
3429 | E->VectorizedValue = V; | |||
3430 | ++NumVectorInstructions; | |||
3431 | return V; | |||
3432 | } | |||
3433 | case Instruction::GetElementPtr: { | |||
3434 | setInsertPointAfterBundle(E->Scalars, VL0); | |||
3435 | ||||
3436 | ValueList Op0VL; | |||
3437 | for (Value *V : E->Scalars) | |||
3438 | Op0VL.push_back(cast<GetElementPtrInst>(V)->getOperand(0)); | |||
3439 | ||||
3440 | Value *Op0 = vectorizeTree(Op0VL); | |||
3441 | ||||
3442 | std::vector<Value *> OpVecs; | |||
3443 | for (int j = 1, e = cast<GetElementPtrInst>(VL0)->getNumOperands(); j < e; | |||
3444 | ++j) { | |||
3445 | ValueList OpVL; | |||
3446 | for (Value *V : E->Scalars) | |||
3447 | OpVL.push_back(cast<GetElementPtrInst>(V)->getOperand(j)); | |||
3448 | ||||
3449 | Value *OpVec = vectorizeTree(OpVL); | |||
3450 | OpVecs.push_back(OpVec); | |||
3451 | } | |||
3452 | ||||
3453 | Value *V = Builder.CreateGEP( | |||
3454 | cast<GetElementPtrInst>(VL0)->getSourceElementType(), Op0, OpVecs); | |||
3455 | if (Instruction *I = dyn_cast<Instruction>(V)) | |||
3456 | V = propagateMetadata(I, E->Scalars); | |||
3457 | ||||
3458 | if (NeedToShuffleReuses) { | |||
3459 | V = Builder.CreateShuffleVector(V, UndefValue::get(VecTy), | |||
3460 | E->ReuseShuffleIndices, "shuffle"); | |||
3461 | } | |||
3462 | E->VectorizedValue = V; | |||
3463 | ++NumVectorInstructions; | |||
3464 | ||||
3465 | return V; | |||
3466 | } | |||
3467 | case Instruction::Call: { | |||
3468 | CallInst *CI = cast<CallInst>(VL0); | |||
3469 | setInsertPointAfterBundle(E->Scalars, VL0); | |||
3470 | Function *FI; | |||
3471 | Intrinsic::ID IID = Intrinsic::not_intrinsic; | |||
3472 | Value *ScalarArg = nullptr; | |||
3473 | if (CI && (FI = CI->getCalledFunction())) { | |||
3474 | IID = FI->getIntrinsicID(); | |||
3475 | } | |||
3476 | std::vector<Value *> OpVecs; | |||
3477 | for (int j = 0, e = CI->getNumArgOperands(); j < e; ++j) { | |||
3478 | ValueList OpVL; | |||
3479 | // ctlz,cttz and powi are special intrinsics whose second argument is | |||
3480 | // a scalar. This argument should not be vectorized. | |||
3481 | if (hasVectorInstrinsicScalarOpd(IID, 1) && j == 1) { | |||
3482 | CallInst *CEI = cast<CallInst>(VL0); | |||
3483 | ScalarArg = CEI->getArgOperand(j); | |||
3484 | OpVecs.push_back(CEI->getArgOperand(j)); | |||
3485 | continue; | |||
3486 | } | |||
3487 | for (Value *V : E->Scalars) { | |||
3488 | CallInst *CEI = cast<CallInst>(V); | |||
3489 | OpVL.push_back(CEI->getArgOperand(j)); | |||
3490 | } | |||
3491 | ||||
3492 | Value *OpVec = vectorizeTree(OpVL); | |||
3493 | DEBUG(dbgs() << "SLP: OpVec[" << j << "]: " << *OpVec << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: OpVec[" << j << "]: " << *OpVec << "\n"; } } while (false); | |||
3494 | OpVecs.push_back(OpVec); | |||
3495 | } | |||
3496 | ||||
3497 | Module *M = F->getParent(); | |||
3498 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); | |||
3499 | Type *Tys[] = { VectorType::get(CI->getType(), E->Scalars.size()) }; | |||
3500 | Function *CF = Intrinsic::getDeclaration(M, ID, Tys); | |||
3501 | SmallVector<OperandBundleDef, 1> OpBundles; | |||
3502 | CI->getOperandBundlesAsDefs(OpBundles); | |||
3503 | Value *V = Builder.CreateCall(CF, OpVecs, OpBundles); | |||
3504 | ||||
3505 | // The scalar argument uses an in-tree scalar so we add the new vectorized | |||
3506 | // call to ExternalUses list to make sure that an extract will be | |||
3507 | // generated in the future. | |||
3508 | if (ScalarArg && getTreeEntry(ScalarArg)) | |||
3509 | ExternalUses.push_back(ExternalUser(ScalarArg, cast<User>(V), 0)); | |||
3510 | ||||
3511 | propagateIRFlags(V, E->Scalars, VL0); | |||
3512 | if (NeedToShuffleReuses) { | |||
3513 | V = Builder.CreateShuffleVector(V, UndefValue::get(VecTy), | |||
3514 | E->ReuseShuffleIndices, "shuffle"); | |||
3515 | } | |||
3516 | E->VectorizedValue = V; | |||
3517 | ++NumVectorInstructions; | |||
3518 | return V; | |||
3519 | } | |||
3520 | case Instruction::ShuffleVector: { | |||
3521 | ValueList LHSVL, RHSVL; | |||
3522 | assert(Instruction::isBinaryOp(S.Opcode) &&(static_cast <bool> (Instruction::isBinaryOp(S.Opcode) && "Invalid Shuffle Vector Operand") ? void (0) : __assert_fail ("Instruction::isBinaryOp(S.Opcode) && \"Invalid Shuffle Vector Operand\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3523, __extension__ __PRETTY_FUNCTION__)) | |||
3523 | "Invalid Shuffle Vector Operand")(static_cast <bool> (Instruction::isBinaryOp(S.Opcode) && "Invalid Shuffle Vector Operand") ? void (0) : __assert_fail ("Instruction::isBinaryOp(S.Opcode) && \"Invalid Shuffle Vector Operand\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3523, __extension__ __PRETTY_FUNCTION__)); | |||
3524 | reorderAltShuffleOperands(S.Opcode, E->Scalars, LHSVL, RHSVL); | |||
3525 | setInsertPointAfterBundle(E->Scalars, VL0); | |||
3526 | ||||
3527 | Value *LHS = vectorizeTree(LHSVL); | |||
3528 | Value *RHS = vectorizeTree(RHSVL); | |||
3529 | ||||
3530 | if (E->VectorizedValue) { | |||
3531 | DEBUG(dbgs() << "SLP: Diamond merged for " << *VL0 << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Diamond merged for " << *VL0 << ".\n"; } } while (false); | |||
3532 | return E->VectorizedValue; | |||
3533 | } | |||
3534 | ||||
3535 | // Create a vector of LHS op1 RHS | |||
3536 | Value *V0 = Builder.CreateBinOp( | |||
3537 | static_cast<Instruction::BinaryOps>(S.Opcode), LHS, RHS); | |||
3538 | ||||
3539 | unsigned AltOpcode = getAltOpcode(S.Opcode); | |||
3540 | // Create a vector of LHS op2 RHS | |||
3541 | Value *V1 = Builder.CreateBinOp( | |||
3542 | static_cast<Instruction::BinaryOps>(AltOpcode), LHS, RHS); | |||
3543 | ||||
3544 | // Create shuffle to take alternate operations from the vector. | |||
3545 | // Also, gather up odd and even scalar ops to propagate IR flags to | |||
3546 | // each vector operation. | |||
3547 | ValueList OddScalars, EvenScalars; | |||
3548 | unsigned e = E->Scalars.size(); | |||
3549 | SmallVector<Constant *, 8> Mask(e); | |||
3550 | for (unsigned i = 0; i < e; ++i) { | |||
3551 | if (isOdd(i)) { | |||
3552 | Mask[i] = Builder.getInt32(e + i); | |||
3553 | OddScalars.push_back(E->Scalars[i]); | |||
3554 | } else { | |||
3555 | Mask[i] = Builder.getInt32(i); | |||
3556 | EvenScalars.push_back(E->Scalars[i]); | |||
3557 | } | |||
3558 | } | |||
3559 | ||||
3560 | Value *ShuffleMask = ConstantVector::get(Mask); | |||
3561 | propagateIRFlags(V0, EvenScalars); | |||
3562 | propagateIRFlags(V1, OddScalars); | |||
3563 | ||||
3564 | Value *V = Builder.CreateShuffleVector(V0, V1, ShuffleMask); | |||
3565 | if (Instruction *I = dyn_cast<Instruction>(V)) | |||
3566 | V = propagateMetadata(I, E->Scalars); | |||
3567 | if (NeedToShuffleReuses) { | |||
3568 | V = Builder.CreateShuffleVector(V, UndefValue::get(VecTy), | |||
3569 | E->ReuseShuffleIndices, "shuffle"); | |||
3570 | } | |||
3571 | E->VectorizedValue = V; | |||
3572 | ++NumVectorInstructions; | |||
3573 | ||||
3574 | return V; | |||
3575 | } | |||
3576 | default: | |||
3577 | llvm_unreachable("unknown inst")::llvm::llvm_unreachable_internal("unknown inst", "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3577); | |||
3578 | } | |||
3579 | return nullptr; | |||
3580 | } | |||
3581 | ||||
3582 | Value *BoUpSLP::vectorizeTree() { | |||
3583 | ExtraValueToDebugLocsMap ExternallyUsedValues; | |||
3584 | return vectorizeTree(ExternallyUsedValues); | |||
3585 | } | |||
3586 | ||||
3587 | Value * | |||
3588 | BoUpSLP::vectorizeTree(ExtraValueToDebugLocsMap &ExternallyUsedValues) { | |||
3589 | // All blocks must be scheduled before any instructions are inserted. | |||
3590 | for (auto &BSIter : BlocksSchedules) { | |||
3591 | scheduleBlock(BSIter.second.get()); | |||
3592 | } | |||
3593 | ||||
3594 | Builder.SetInsertPoint(&F->getEntryBlock().front()); | |||
3595 | auto *VectorRoot = vectorizeTree(&VectorizableTree[0]); | |||
3596 | ||||
3597 | // If the vectorized tree can be rewritten in a smaller type, we truncate the | |||
3598 | // vectorized root. InstCombine will then rewrite the entire expression. We | |||
3599 | // sign extend the extracted values below. | |||
3600 | auto *ScalarRoot = VectorizableTree[0].Scalars[0]; | |||
3601 | if (MinBWs.count(ScalarRoot)) { | |||
3602 | if (auto *I = dyn_cast<Instruction>(VectorRoot)) | |||
3603 | Builder.SetInsertPoint(&*++BasicBlock::iterator(I)); | |||
3604 | auto BundleWidth = VectorizableTree[0].Scalars.size(); | |||
3605 | auto *MinTy = IntegerType::get(F->getContext(), MinBWs[ScalarRoot].first); | |||
3606 | auto *VecTy = VectorType::get(MinTy, BundleWidth); | |||
3607 | auto *Trunc = Builder.CreateTrunc(VectorRoot, VecTy); | |||
3608 | VectorizableTree[0].VectorizedValue = Trunc; | |||
3609 | } | |||
3610 | ||||
3611 | DEBUG(dbgs() << "SLP: Extracting " << ExternalUses.size() << " values .\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Extracting " << ExternalUses .size() << " values .\n"; } } while (false); | |||
3612 | ||||
3613 | // If necessary, sign-extend or zero-extend ScalarRoot to the larger type | |||
3614 | // specified by ScalarType. | |||
3615 | auto extend = [&](Value *ScalarRoot, Value *Ex, Type *ScalarType) { | |||
3616 | if (!MinBWs.count(ScalarRoot)) | |||
3617 | return Ex; | |||
3618 | if (MinBWs[ScalarRoot].second) | |||
3619 | return Builder.CreateSExt(Ex, ScalarType); | |||
3620 | return Builder.CreateZExt(Ex, ScalarType); | |||
3621 | }; | |||
3622 | ||||
3623 | // Extract all of the elements with the external uses. | |||
3624 | for (const auto &ExternalUse : ExternalUses) { | |||
3625 | Value *Scalar = ExternalUse.Scalar; | |||
3626 | llvm::User *User = ExternalUse.User; | |||
3627 | ||||
3628 | // Skip users that we already RAUW. This happens when one instruction | |||
3629 | // has multiple uses of the same value. | |||
3630 | if (User && !is_contained(Scalar->users(), User)) | |||
3631 | continue; | |||
3632 | TreeEntry *E = getTreeEntry(Scalar); | |||
3633 | assert(E && "Invalid scalar")(static_cast <bool> (E && "Invalid scalar") ? void (0) : __assert_fail ("E && \"Invalid scalar\"", "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3633, __extension__ __PRETTY_FUNCTION__)); | |||
3634 | assert(!E->NeedToGather && "Extracting from a gather list")(static_cast <bool> (!E->NeedToGather && "Extracting from a gather list" ) ? void (0) : __assert_fail ("!E->NeedToGather && \"Extracting from a gather list\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3634, __extension__ __PRETTY_FUNCTION__)); | |||
3635 | ||||
3636 | Value *Vec = E->VectorizedValue; | |||
3637 | assert(Vec && "Can't find vectorizable value")(static_cast <bool> (Vec && "Can't find vectorizable value" ) ? void (0) : __assert_fail ("Vec && \"Can't find vectorizable value\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3637, __extension__ __PRETTY_FUNCTION__)); | |||
3638 | ||||
3639 | Value *Lane = Builder.getInt32(ExternalUse.Lane); | |||
3640 | // If User == nullptr, the Scalar is used as extra arg. Generate | |||
3641 | // ExtractElement instruction and update the record for this scalar in | |||
3642 | // ExternallyUsedValues. | |||
3643 | if (!User) { | |||
3644 | assert(ExternallyUsedValues.count(Scalar) &&(static_cast <bool> (ExternallyUsedValues.count(Scalar) && "Scalar with nullptr as an external user must be registered in " "ExternallyUsedValues map") ? void (0) : __assert_fail ("ExternallyUsedValues.count(Scalar) && \"Scalar with nullptr as an external user must be registered in \" \"ExternallyUsedValues map\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3646, __extension__ __PRETTY_FUNCTION__)) | |||
3645 | "Scalar with nullptr as an external user must be registered in "(static_cast <bool> (ExternallyUsedValues.count(Scalar) && "Scalar with nullptr as an external user must be registered in " "ExternallyUsedValues map") ? void (0) : __assert_fail ("ExternallyUsedValues.count(Scalar) && \"Scalar with nullptr as an external user must be registered in \" \"ExternallyUsedValues map\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3646, __extension__ __PRETTY_FUNCTION__)) | |||
3646 | "ExternallyUsedValues map")(static_cast <bool> (ExternallyUsedValues.count(Scalar) && "Scalar with nullptr as an external user must be registered in " "ExternallyUsedValues map") ? void (0) : __assert_fail ("ExternallyUsedValues.count(Scalar) && \"Scalar with nullptr as an external user must be registered in \" \"ExternallyUsedValues map\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3646, __extension__ __PRETTY_FUNCTION__)); | |||
3647 | if (auto *VecI = dyn_cast<Instruction>(Vec)) { | |||
3648 | Builder.SetInsertPoint(VecI->getParent(), | |||
3649 | std::next(VecI->getIterator())); | |||
3650 | } else { | |||
3651 | Builder.SetInsertPoint(&F->getEntryBlock().front()); | |||
3652 | } | |||
3653 | Value *Ex = Builder.CreateExtractElement(Vec, Lane); | |||
3654 | Ex = extend(ScalarRoot, Ex, Scalar->getType()); | |||
3655 | CSEBlocks.insert(cast<Instruction>(Scalar)->getParent()); | |||
3656 | auto &Locs = ExternallyUsedValues[Scalar]; | |||
3657 | ExternallyUsedValues.insert({Ex, Locs}); | |||
3658 | ExternallyUsedValues.erase(Scalar); | |||
3659 | continue; | |||
3660 | } | |||
3661 | ||||
3662 | // Generate extracts for out-of-tree users. | |||
3663 | // Find the insertion point for the extractelement lane. | |||
3664 | if (auto *VecI = dyn_cast<Instruction>(Vec)) { | |||
3665 | if (PHINode *PH = dyn_cast<PHINode>(User)) { | |||
3666 | for (int i = 0, e = PH->getNumIncomingValues(); i != e; ++i) { | |||
3667 | if (PH->getIncomingValue(i) == Scalar) { | |||
3668 | TerminatorInst *IncomingTerminator = | |||
3669 | PH->getIncomingBlock(i)->getTerminator(); | |||
3670 | if (isa<CatchSwitchInst>(IncomingTerminator)) { | |||
3671 | Builder.SetInsertPoint(VecI->getParent(), | |||
3672 | std::next(VecI->getIterator())); | |||
3673 | } else { | |||
3674 | Builder.SetInsertPoint(PH->getIncomingBlock(i)->getTerminator()); | |||
3675 | } | |||
3676 | Value *Ex = Builder.CreateExtractElement(Vec, Lane); | |||
3677 | Ex = extend(ScalarRoot, Ex, Scalar->getType()); | |||
3678 | CSEBlocks.insert(PH->getIncomingBlock(i)); | |||
3679 | PH->setOperand(i, Ex); | |||
3680 | } | |||
3681 | } | |||
3682 | } else { | |||
3683 | Builder.SetInsertPoint(cast<Instruction>(User)); | |||
3684 | Value *Ex = Builder.CreateExtractElement(Vec, Lane); | |||
3685 | Ex = extend(ScalarRoot, Ex, Scalar->getType()); | |||
3686 | CSEBlocks.insert(cast<Instruction>(User)->getParent()); | |||
3687 | User->replaceUsesOfWith(Scalar, Ex); | |||
3688 | } | |||
3689 | } else { | |||
3690 | Builder.SetInsertPoint(&F->getEntryBlock().front()); | |||
3691 | Value *Ex = Builder.CreateExtractElement(Vec, Lane); | |||
3692 | Ex = extend(ScalarRoot, Ex, Scalar->getType()); | |||
3693 | CSEBlocks.insert(&F->getEntryBlock()); | |||
3694 | User->replaceUsesOfWith(Scalar, Ex); | |||
3695 | } | |||
3696 | ||||
3697 | DEBUG(dbgs() << "SLP: Replaced:" << *User << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Replaced:" << *User << ".\n"; } } while (false); | |||
3698 | } | |||
3699 | ||||
3700 | // For each vectorized value: | |||
3701 | for (TreeEntry &EIdx : VectorizableTree) { | |||
3702 | TreeEntry *Entry = &EIdx; | |||
3703 | ||||
3704 | // No need to handle users of gathered values. | |||
3705 | if (Entry->NeedToGather) | |||
3706 | continue; | |||
3707 | ||||
3708 | assert(Entry->VectorizedValue && "Can't find vectorizable value")(static_cast <bool> (Entry->VectorizedValue && "Can't find vectorizable value") ? void (0) : __assert_fail ( "Entry->VectorizedValue && \"Can't find vectorizable value\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3708, __extension__ __PRETTY_FUNCTION__)); | |||
3709 | ||||
3710 | // For each lane: | |||
3711 | for (int Lane = 0, LE = Entry->Scalars.size(); Lane != LE; ++Lane) { | |||
3712 | Value *Scalar = Entry->Scalars[Lane]; | |||
3713 | ||||
3714 | Type *Ty = Scalar->getType(); | |||
3715 | if (!Ty->isVoidTy()) { | |||
3716 | #ifndef NDEBUG | |||
3717 | for (User *U : Scalar->users()) { | |||
3718 | DEBUG(dbgs() << "SLP: \tvalidating user:" << *U << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: \tvalidating user:" << *U << ".\n"; } } while (false); | |||
3719 | ||||
3720 | // It is legal to replace users in the ignorelist by undef. | |||
3721 | assert((getTreeEntry(U) || is_contained(UserIgnoreList, U)) &&(static_cast <bool> ((getTreeEntry(U) || is_contained(UserIgnoreList , U)) && "Replacing out-of-tree value with undef") ? void (0) : __assert_fail ("(getTreeEntry(U) || is_contained(UserIgnoreList, U)) && \"Replacing out-of-tree value with undef\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3722, __extension__ __PRETTY_FUNCTION__)) | |||
3722 | "Replacing out-of-tree value with undef")(static_cast <bool> ((getTreeEntry(U) || is_contained(UserIgnoreList , U)) && "Replacing out-of-tree value with undef") ? void (0) : __assert_fail ("(getTreeEntry(U) || is_contained(UserIgnoreList, U)) && \"Replacing out-of-tree value with undef\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3722, __extension__ __PRETTY_FUNCTION__)); | |||
3723 | } | |||
3724 | #endif | |||
3725 | Value *Undef = UndefValue::get(Ty); | |||
3726 | Scalar->replaceAllUsesWith(Undef); | |||
3727 | } | |||
3728 | DEBUG(dbgs() << "SLP: \tErasing scalar:" << *Scalar << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: \tErasing scalar:" << * Scalar << ".\n"; } } while (false); | |||
3729 | eraseInstruction(cast<Instruction>(Scalar)); | |||
3730 | } | |||
3731 | } | |||
3732 | ||||
3733 | Builder.ClearInsertionPoint(); | |||
3734 | ||||
3735 | return VectorizableTree[0].VectorizedValue; | |||
3736 | } | |||
3737 | ||||
3738 | void BoUpSLP::optimizeGatherSequence() { | |||
3739 | DEBUG(dbgs() << "SLP: Optimizing " << GatherSeq.size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Optimizing " << GatherSeq .size() << " gather sequences instructions.\n"; } } while (false) | |||
3740 | << " gather sequences instructions.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Optimizing " << GatherSeq .size() << " gather sequences instructions.\n"; } } while (false); | |||
3741 | // LICM InsertElementInst sequences. | |||
3742 | for (Instruction *I : GatherSeq) { | |||
3743 | if (!isa<InsertElementInst>(I) && !isa<ShuffleVectorInst>(I)) | |||
3744 | continue; | |||
3745 | ||||
3746 | // Check if this block is inside a loop. | |||
3747 | Loop *L = LI->getLoopFor(I->getParent()); | |||
3748 | if (!L) | |||
3749 | continue; | |||
3750 | ||||
3751 | // Check if it has a preheader. | |||
3752 | BasicBlock *PreHeader = L->getLoopPreheader(); | |||
3753 | if (!PreHeader) | |||
3754 | continue; | |||
3755 | ||||
3756 | // If the vector or the element that we insert into it are | |||
3757 | // instructions that are defined in this basic block then we can't | |||
3758 | // hoist this instruction. | |||
3759 | auto *Op0 = dyn_cast<Instruction>(I->getOperand(0)); | |||
3760 | auto *Op1 = dyn_cast<Instruction>(I->getOperand(1)); | |||
3761 | if (Op0 && L->contains(Op0)) | |||
3762 | continue; | |||
3763 | if (Op1 && L->contains(Op1)) | |||
3764 | continue; | |||
3765 | ||||
3766 | // We can hoist this instruction. Move it to the pre-header. | |||
3767 | I->moveBefore(PreHeader->getTerminator()); | |||
3768 | } | |||
3769 | ||||
3770 | // Make a list of all reachable blocks in our CSE queue. | |||
3771 | SmallVector<const DomTreeNode *, 8> CSEWorkList; | |||
3772 | CSEWorkList.reserve(CSEBlocks.size()); | |||
3773 | for (BasicBlock *BB : CSEBlocks) | |||
3774 | if (DomTreeNode *N = DT->getNode(BB)) { | |||
3775 | assert(DT->isReachableFromEntry(N))(static_cast <bool> (DT->isReachableFromEntry(N)) ? void (0) : __assert_fail ("DT->isReachableFromEntry(N)", "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3775, __extension__ __PRETTY_FUNCTION__)); | |||
3776 | CSEWorkList.push_back(N); | |||
3777 | } | |||
3778 | ||||
3779 | // Sort blocks by domination. This ensures we visit a block after all blocks | |||
3780 | // dominating it are visited. | |||
3781 | std::stable_sort(CSEWorkList.begin(), CSEWorkList.end(), | |||
3782 | [this](const DomTreeNode *A, const DomTreeNode *B) { | |||
3783 | return DT->properlyDominates(A, B); | |||
3784 | }); | |||
3785 | ||||
3786 | // Perform O(N^2) search over the gather sequences and merge identical | |||
3787 | // instructions. TODO: We can further optimize this scan if we split the | |||
3788 | // instructions into different buckets based on the insert lane. | |||
3789 | SmallVector<Instruction *, 16> Visited; | |||
3790 | for (auto I = CSEWorkList.begin(), E = CSEWorkList.end(); I != E; ++I) { | |||
3791 | assert((I == CSEWorkList.begin() || !DT->dominates(*I, *std::prev(I))) &&(static_cast <bool> ((I == CSEWorkList.begin() || !DT-> dominates(*I, *std::prev(I))) && "Worklist not sorted properly!" ) ? void (0) : __assert_fail ("(I == CSEWorkList.begin() || !DT->dominates(*I, *std::prev(I))) && \"Worklist not sorted properly!\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3792, __extension__ __PRETTY_FUNCTION__)) | |||
3792 | "Worklist not sorted properly!")(static_cast <bool> ((I == CSEWorkList.begin() || !DT-> dominates(*I, *std::prev(I))) && "Worklist not sorted properly!" ) ? void (0) : __assert_fail ("(I == CSEWorkList.begin() || !DT->dominates(*I, *std::prev(I))) && \"Worklist not sorted properly!\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3792, __extension__ __PRETTY_FUNCTION__)); | |||
3793 | BasicBlock *BB = (*I)->getBlock(); | |||
3794 | // For all instructions in blocks containing gather sequences: | |||
3795 | for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e;) { | |||
3796 | Instruction *In = &*it++; | |||
3797 | if (!isa<InsertElementInst>(In) && !isa<ExtractElementInst>(In)) | |||
3798 | continue; | |||
3799 | ||||
3800 | // Check if we can replace this instruction with any of the | |||
3801 | // visited instructions. | |||
3802 | for (Instruction *v : Visited) { | |||
3803 | if (In->isIdenticalTo(v) && | |||
3804 | DT->dominates(v->getParent(), In->getParent())) { | |||
3805 | In->replaceAllUsesWith(v); | |||
3806 | eraseInstruction(In); | |||
3807 | In = nullptr; | |||
3808 | break; | |||
3809 | } | |||
3810 | } | |||
3811 | if (In) { | |||
3812 | assert(!is_contained(Visited, In))(static_cast <bool> (!is_contained(Visited, In)) ? void (0) : __assert_fail ("!is_contained(Visited, In)", "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3812, __extension__ __PRETTY_FUNCTION__)); | |||
3813 | Visited.push_back(In); | |||
3814 | } | |||
3815 | } | |||
3816 | } | |||
3817 | CSEBlocks.clear(); | |||
3818 | GatherSeq.clear(); | |||
3819 | } | |||
3820 | ||||
3821 | // Groups the instructions to a bundle (which is then a single scheduling entity) | |||
3822 | // and schedules instructions until the bundle gets ready. | |||
3823 | bool BoUpSLP::BlockScheduling::tryScheduleBundle(ArrayRef<Value *> VL, | |||
3824 | BoUpSLP *SLP, Value *OpValue) { | |||
3825 | if (isa<PHINode>(OpValue)) | |||
3826 | return true; | |||
3827 | ||||
3828 | // Initialize the instruction bundle. | |||
3829 | Instruction *OldScheduleEnd = ScheduleEnd; | |||
3830 | ScheduleData *PrevInBundle = nullptr; | |||
3831 | ScheduleData *Bundle = nullptr; | |||
3832 | bool ReSchedule = false; | |||
3833 | DEBUG(dbgs() << "SLP: bundle: " << *OpValue << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: bundle: " << *OpValue << "\n"; } } while (false); | |||
3834 | ||||
3835 | // Make sure that the scheduling region contains all | |||
3836 | // instructions of the bundle. | |||
3837 | for (Value *V : VL) { | |||
3838 | if (!extendSchedulingRegion(V, OpValue)) | |||
3839 | return false; | |||
3840 | } | |||
3841 | ||||
3842 | for (Value *V : VL) { | |||
3843 | ScheduleData *BundleMember = getScheduleData(V); | |||
3844 | assert(BundleMember &&(static_cast <bool> (BundleMember && "no ScheduleData for bundle member (maybe not in same basic block)" ) ? void (0) : __assert_fail ("BundleMember && \"no ScheduleData for bundle member (maybe not in same basic block)\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3845, __extension__ __PRETTY_FUNCTION__)) | |||
3845 | "no ScheduleData for bundle member (maybe not in same basic block)")(static_cast <bool> (BundleMember && "no ScheduleData for bundle member (maybe not in same basic block)" ) ? void (0) : __assert_fail ("BundleMember && \"no ScheduleData for bundle member (maybe not in same basic block)\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3845, __extension__ __PRETTY_FUNCTION__)); | |||
3846 | if (BundleMember->IsScheduled) { | |||
3847 | // A bundle member was scheduled as single instruction before and now | |||
3848 | // needs to be scheduled as part of the bundle. We just get rid of the | |||
3849 | // existing schedule. | |||
3850 | 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) | |||
3851 | << " was already scheduled\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: reset schedule because " << *BundleMember << " was already scheduled\n"; } } while (false); | |||
3852 | ReSchedule = true; | |||
3853 | } | |||
3854 | assert(BundleMember->isSchedulingEntity() &&(static_cast <bool> (BundleMember->isSchedulingEntity () && "bundle member already part of other bundle") ? void (0) : __assert_fail ("BundleMember->isSchedulingEntity() && \"bundle member already part of other bundle\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3855, __extension__ __PRETTY_FUNCTION__)) | |||
3855 | "bundle member already part of other bundle")(static_cast <bool> (BundleMember->isSchedulingEntity () && "bundle member already part of other bundle") ? void (0) : __assert_fail ("BundleMember->isSchedulingEntity() && \"bundle member already part of other bundle\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3855, __extension__ __PRETTY_FUNCTION__)); | |||
3856 | if (PrevInBundle) { | |||
3857 | PrevInBundle->NextInBundle = BundleMember; | |||
3858 | } else { | |||
3859 | Bundle = BundleMember; | |||
3860 | } | |||
3861 | BundleMember->UnscheduledDepsInBundle = 0; | |||
3862 | Bundle->UnscheduledDepsInBundle += BundleMember->UnscheduledDeps; | |||
3863 | ||||
3864 | // Group the instructions to a bundle. | |||
3865 | BundleMember->FirstInBundle = Bundle; | |||
3866 | PrevInBundle = BundleMember; | |||
3867 | } | |||
3868 | if (ScheduleEnd != OldScheduleEnd) { | |||
3869 | // The scheduling region got new instructions at the lower end (or it is a | |||
3870 | // new region for the first bundle). This makes it necessary to | |||
3871 | // recalculate all dependencies. | |||
3872 | // It is seldom that this needs to be done a second time after adding the | |||
3873 | // initial bundle to the region. | |||
3874 | for (auto *I = ScheduleStart; I != ScheduleEnd; I = I->getNextNode()) { | |||
3875 | doForAllOpcodes(I, [](ScheduleData *SD) { | |||
3876 | SD->clearDependencies(); | |||
3877 | }); | |||
3878 | } | |||
3879 | ReSchedule = true; | |||
3880 | } | |||
3881 | if (ReSchedule) { | |||
3882 | resetSchedule(); | |||
3883 | initialFillReadyList(ReadyInsts); | |||
3884 | } | |||
3885 | ||||
3886 | 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) | |||
3887 | << BB->getName() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: try schedule bundle " << *Bundle << " in block " << BB->getName() << "\n"; } } while (false); | |||
3888 | ||||
3889 | calculateDependencies(Bundle, true, SLP); | |||
3890 | ||||
3891 | // Now try to schedule the new bundle. As soon as the bundle is "ready" it | |||
3892 | // means that there are no cyclic dependencies and we can schedule it. | |||
3893 | // Note that's important that we don't "schedule" the bundle yet (see | |||
3894 | // cancelScheduling). | |||
3895 | while (!Bundle->isReady() && !ReadyInsts.empty()) { | |||
3896 | ||||
3897 | ScheduleData *pickedSD = ReadyInsts.back(); | |||
3898 | ReadyInsts.pop_back(); | |||
3899 | ||||
3900 | if (pickedSD->isSchedulingEntity() && pickedSD->isReady()) { | |||
3901 | schedule(pickedSD, ReadyInsts); | |||
3902 | } | |||
3903 | } | |||
3904 | if (!Bundle->isReady()) { | |||
3905 | cancelScheduling(VL, OpValue); | |||
3906 | return false; | |||
3907 | } | |||
3908 | return true; | |||
3909 | } | |||
3910 | ||||
3911 | void BoUpSLP::BlockScheduling::cancelScheduling(ArrayRef<Value *> VL, | |||
3912 | Value *OpValue) { | |||
3913 | if (isa<PHINode>(OpValue)) | |||
3914 | return; | |||
3915 | ||||
3916 | ScheduleData *Bundle = getScheduleData(OpValue); | |||
3917 | DEBUG(dbgs() << "SLP: cancel scheduling of " << *Bundle << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: cancel scheduling of " << *Bundle << "\n"; } } while (false); | |||
3918 | assert(!Bundle->IsScheduled &&(static_cast <bool> (!Bundle->IsScheduled && "Can't cancel bundle which is already scheduled") ? void (0) : __assert_fail ("!Bundle->IsScheduled && \"Can't cancel bundle which is already scheduled\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3919, __extension__ __PRETTY_FUNCTION__)) | |||
3919 | "Can't cancel bundle which is already scheduled")(static_cast <bool> (!Bundle->IsScheduled && "Can't cancel bundle which is already scheduled") ? void (0) : __assert_fail ("!Bundle->IsScheduled && \"Can't cancel bundle which is already scheduled\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3919, __extension__ __PRETTY_FUNCTION__)); | |||
3920 | assert(Bundle->isSchedulingEntity() && Bundle->isPartOfBundle() &&(static_cast <bool> (Bundle->isSchedulingEntity() && Bundle->isPartOfBundle() && "tried to unbundle something which is not a bundle" ) ? void (0) : __assert_fail ("Bundle->isSchedulingEntity() && Bundle->isPartOfBundle() && \"tried to unbundle something which is not a bundle\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3921, __extension__ __PRETTY_FUNCTION__)) | |||
3921 | "tried to unbundle something which is not a bundle")(static_cast <bool> (Bundle->isSchedulingEntity() && Bundle->isPartOfBundle() && "tried to unbundle something which is not a bundle" ) ? void (0) : __assert_fail ("Bundle->isSchedulingEntity() && Bundle->isPartOfBundle() && \"tried to unbundle something which is not a bundle\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3921, __extension__ __PRETTY_FUNCTION__)); | |||
3922 | ||||
3923 | // Un-bundle: make single instructions out of the bundle. | |||
3924 | ScheduleData *BundleMember = Bundle; | |||
3925 | while (BundleMember) { | |||
3926 | assert(BundleMember->FirstInBundle == Bundle && "corrupt bundle links")(static_cast <bool> (BundleMember->FirstInBundle == Bundle && "corrupt bundle links") ? void (0) : __assert_fail ("BundleMember->FirstInBundle == Bundle && \"corrupt bundle links\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3926, __extension__ __PRETTY_FUNCTION__)); | |||
3927 | BundleMember->FirstInBundle = BundleMember; | |||
3928 | ScheduleData *Next = BundleMember->NextInBundle; | |||
3929 | BundleMember->NextInBundle = nullptr; | |||
3930 | BundleMember->UnscheduledDepsInBundle = BundleMember->UnscheduledDeps; | |||
3931 | if (BundleMember->UnscheduledDepsInBundle == 0) { | |||
3932 | ReadyInsts.insert(BundleMember); | |||
3933 | } | |||
3934 | BundleMember = Next; | |||
3935 | } | |||
3936 | } | |||
3937 | ||||
3938 | BoUpSLP::ScheduleData *BoUpSLP::BlockScheduling::allocateScheduleDataChunks() { | |||
3939 | // Allocate a new ScheduleData for the instruction. | |||
3940 | if (ChunkPos >= ChunkSize) { | |||
3941 | ScheduleDataChunks.push_back(llvm::make_unique<ScheduleData[]>(ChunkSize)); | |||
3942 | ChunkPos = 0; | |||
3943 | } | |||
3944 | return &(ScheduleDataChunks.back()[ChunkPos++]); | |||
3945 | } | |||
3946 | ||||
3947 | bool BoUpSLP::BlockScheduling::extendSchedulingRegion(Value *V, | |||
3948 | Value *OpValue) { | |||
3949 | if (getScheduleData(V, isOneOf(OpValue, V))) | |||
3950 | return true; | |||
3951 | Instruction *I = dyn_cast<Instruction>(V); | |||
3952 | assert(I && "bundle member must be an instruction")(static_cast <bool> (I && "bundle member must be an instruction" ) ? void (0) : __assert_fail ("I && \"bundle member must be an instruction\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3952, __extension__ __PRETTY_FUNCTION__)); | |||
3953 | assert(!isa<PHINode>(I) && "phi nodes don't need to be scheduled")(static_cast <bool> (!isa<PHINode>(I) && "phi nodes don't need to be scheduled" ) ? void (0) : __assert_fail ("!isa<PHINode>(I) && \"phi nodes don't need to be scheduled\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3953, __extension__ __PRETTY_FUNCTION__)); | |||
3954 | auto &&CheckSheduleForI = [this, OpValue](Instruction *I) -> bool { | |||
3955 | ScheduleData *ISD = getScheduleData(I); | |||
3956 | if (!ISD) | |||
3957 | return false; | |||
3958 | assert(isInSchedulingRegion(ISD) &&(static_cast <bool> (isInSchedulingRegion(ISD) && "ScheduleData not in scheduling region") ? void (0) : __assert_fail ("isInSchedulingRegion(ISD) && \"ScheduleData not in scheduling region\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3959, __extension__ __PRETTY_FUNCTION__)) | |||
3959 | "ScheduleData not in scheduling region")(static_cast <bool> (isInSchedulingRegion(ISD) && "ScheduleData not in scheduling region") ? void (0) : __assert_fail ("isInSchedulingRegion(ISD) && \"ScheduleData not in scheduling region\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3959, __extension__ __PRETTY_FUNCTION__)); | |||
3960 | ScheduleData *SD = allocateScheduleDataChunks(); | |||
3961 | SD->Inst = I; | |||
3962 | SD->init(SchedulingRegionID, OpValue); | |||
3963 | ExtraScheduleDataMap[I][OpValue] = SD; | |||
3964 | return true; | |||
3965 | }; | |||
3966 | if (CheckSheduleForI(I)) | |||
3967 | return true; | |||
3968 | if (!ScheduleStart) { | |||
3969 | // It's the first instruction in the new region. | |||
3970 | initScheduleData(I, I->getNextNode(), nullptr, nullptr); | |||
3971 | ScheduleStart = I; | |||
3972 | ScheduleEnd = I->getNextNode(); | |||
3973 | if (isOneOf(OpValue, I) != I) | |||
3974 | CheckSheduleForI(I); | |||
3975 | assert(ScheduleEnd && "tried to vectorize a TerminatorInst?")(static_cast <bool> (ScheduleEnd && "tried to vectorize a TerminatorInst?" ) ? void (0) : __assert_fail ("ScheduleEnd && \"tried to vectorize a TerminatorInst?\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3975, __extension__ __PRETTY_FUNCTION__)); | |||
3976 | 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); | |||
3977 | return true; | |||
3978 | } | |||
3979 | // Search up and down at the same time, because we don't know if the new | |||
3980 | // instruction is above or below the existing scheduling region. | |||
3981 | BasicBlock::reverse_iterator UpIter = | |||
3982 | ++ScheduleStart->getIterator().getReverse(); | |||
3983 | BasicBlock::reverse_iterator UpperEnd = BB->rend(); | |||
3984 | BasicBlock::iterator DownIter = ScheduleEnd->getIterator(); | |||
3985 | BasicBlock::iterator LowerEnd = BB->end(); | |||
3986 | while (true) { | |||
3987 | if (++ScheduleRegionSize > ScheduleRegionSizeLimit) { | |||
3988 | 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); | |||
3989 | return false; | |||
3990 | } | |||
3991 | ||||
3992 | if (UpIter != UpperEnd) { | |||
3993 | if (&*UpIter == I) { | |||
3994 | initScheduleData(I, ScheduleStart, nullptr, FirstLoadStoreInRegion); | |||
3995 | ScheduleStart = I; | |||
3996 | if (isOneOf(OpValue, I) != I) | |||
3997 | CheckSheduleForI(I); | |||
3998 | DEBUG(dbgs() << "SLP: extend schedule region start to " << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: extend schedule region start to " << *I << "\n"; } } while (false); | |||
3999 | return true; | |||
4000 | } | |||
4001 | UpIter++; | |||
4002 | } | |||
4003 | if (DownIter != LowerEnd) { | |||
4004 | if (&*DownIter == I) { | |||
4005 | initScheduleData(ScheduleEnd, I->getNextNode(), LastLoadStoreInRegion, | |||
4006 | nullptr); | |||
4007 | ScheduleEnd = I->getNextNode(); | |||
4008 | if (isOneOf(OpValue, I) != I) | |||
4009 | CheckSheduleForI(I); | |||
4010 | assert(ScheduleEnd && "tried to vectorize a TerminatorInst?")(static_cast <bool> (ScheduleEnd && "tried to vectorize a TerminatorInst?" ) ? void (0) : __assert_fail ("ScheduleEnd && \"tried to vectorize a TerminatorInst?\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4010, __extension__ __PRETTY_FUNCTION__)); | |||
4011 | DEBUG(dbgs() << "SLP: extend schedule region end to " << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: extend schedule region end to " << *I << "\n"; } } while (false); | |||
4012 | return true; | |||
4013 | } | |||
4014 | DownIter++; | |||
4015 | } | |||
4016 | assert((UpIter != UpperEnd || DownIter != LowerEnd) &&(static_cast <bool> ((UpIter != UpperEnd || DownIter != LowerEnd) && "instruction not found in block") ? void (0) : __assert_fail ("(UpIter != UpperEnd || DownIter != LowerEnd) && \"instruction not found in block\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4017, __extension__ __PRETTY_FUNCTION__)) | |||
4017 | "instruction not found in block")(static_cast <bool> ((UpIter != UpperEnd || DownIter != LowerEnd) && "instruction not found in block") ? void (0) : __assert_fail ("(UpIter != UpperEnd || DownIter != LowerEnd) && \"instruction not found in block\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4017, __extension__ __PRETTY_FUNCTION__)); | |||
4018 | } | |||
4019 | return true; | |||
4020 | } | |||
4021 | ||||
4022 | void BoUpSLP::BlockScheduling::initScheduleData(Instruction *FromI, | |||
4023 | Instruction *ToI, | |||
4024 | ScheduleData *PrevLoadStore, | |||
4025 | ScheduleData *NextLoadStore) { | |||
4026 | ScheduleData *CurrentLoadStore = PrevLoadStore; | |||
4027 | for (Instruction *I = FromI; I != ToI; I = I->getNextNode()) { | |||
4028 | ScheduleData *SD = ScheduleDataMap[I]; | |||
4029 | if (!SD) { | |||
4030 | SD = allocateScheduleDataChunks(); | |||
4031 | ScheduleDataMap[I] = SD; | |||
4032 | SD->Inst = I; | |||
4033 | } | |||
4034 | assert(!isInSchedulingRegion(SD) &&(static_cast <bool> (!isInSchedulingRegion(SD) && "new ScheduleData already in scheduling region") ? void (0) : __assert_fail ("!isInSchedulingRegion(SD) && \"new ScheduleData already in scheduling region\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4035, __extension__ __PRETTY_FUNCTION__)) | |||
4035 | "new ScheduleData already in scheduling region")(static_cast <bool> (!isInSchedulingRegion(SD) && "new ScheduleData already in scheduling region") ? void (0) : __assert_fail ("!isInSchedulingRegion(SD) && \"new ScheduleData already in scheduling region\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4035, __extension__ __PRETTY_FUNCTION__)); | |||
4036 | SD->init(SchedulingRegionID, I); | |||
4037 | ||||
4038 | if (I->mayReadOrWriteMemory() && | |||
4039 | (!isa<IntrinsicInst>(I) || | |||
4040 | cast<IntrinsicInst>(I)->getIntrinsicID() != Intrinsic::sideeffect)) { | |||
4041 | // Update the linked list of memory accessing instructions. | |||
4042 | if (CurrentLoadStore) { | |||
4043 | CurrentLoadStore->NextLoadStore = SD; | |||
4044 | } else { | |||
4045 | FirstLoadStoreInRegion = SD; | |||
4046 | } | |||
4047 | CurrentLoadStore = SD; | |||
4048 | } | |||
4049 | } | |||
4050 | if (NextLoadStore) { | |||
4051 | if (CurrentLoadStore) | |||
4052 | CurrentLoadStore->NextLoadStore = NextLoadStore; | |||
4053 | } else { | |||
4054 | LastLoadStoreInRegion = CurrentLoadStore; | |||
4055 | } | |||
4056 | } | |||
4057 | ||||
4058 | void BoUpSLP::BlockScheduling::calculateDependencies(ScheduleData *SD, | |||
4059 | bool InsertInReadyList, | |||
4060 | BoUpSLP *SLP) { | |||
4061 | assert(SD->isSchedulingEntity())(static_cast <bool> (SD->isSchedulingEntity()) ? void (0) : __assert_fail ("SD->isSchedulingEntity()", "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4061, __extension__ __PRETTY_FUNCTION__)); | |||
4062 | ||||
4063 | SmallVector<ScheduleData *, 10> WorkList; | |||
4064 | WorkList.push_back(SD); | |||
4065 | ||||
4066 | while (!WorkList.empty()) { | |||
4067 | ScheduleData *SD = WorkList.back(); | |||
4068 | WorkList.pop_back(); | |||
4069 | ||||
4070 | ScheduleData *BundleMember = SD; | |||
4071 | while (BundleMember) { | |||
4072 | assert(isInSchedulingRegion(BundleMember))(static_cast <bool> (isInSchedulingRegion(BundleMember) ) ? void (0) : __assert_fail ("isInSchedulingRegion(BundleMember)" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4072, __extension__ __PRETTY_FUNCTION__)); | |||
4073 | if (!BundleMember->hasValidDependencies()) { | |||
4074 | ||||
4075 | DEBUG(dbgs() << "SLP: update deps of " << *BundleMember << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: update deps of " << *BundleMember << "\n"; } } while (false); | |||
4076 | BundleMember->Dependencies = 0; | |||
4077 | BundleMember->resetUnscheduledDeps(); | |||
4078 | ||||
4079 | // Handle def-use chain dependencies. | |||
4080 | if (BundleMember->OpValue != BundleMember->Inst) { | |||
4081 | ScheduleData *UseSD = getScheduleData(BundleMember->Inst); | |||
4082 | if (UseSD && isInSchedulingRegion(UseSD->FirstInBundle)) { | |||
4083 | BundleMember->Dependencies++; | |||
4084 | ScheduleData *DestBundle = UseSD->FirstInBundle; | |||
4085 | if (!DestBundle->IsScheduled) | |||
4086 | BundleMember->incrementUnscheduledDeps(1); | |||
4087 | if (!DestBundle->hasValidDependencies()) | |||
4088 | WorkList.push_back(DestBundle); | |||
4089 | } | |||
4090 | } else { | |||
4091 | for (User *U : BundleMember->Inst->users()) { | |||
4092 | if (isa<Instruction>(U)) { | |||
4093 | ScheduleData *UseSD = getScheduleData(U); | |||
4094 | if (UseSD && isInSchedulingRegion(UseSD->FirstInBundle)) { | |||
4095 | BundleMember->Dependencies++; | |||
4096 | ScheduleData *DestBundle = UseSD->FirstInBundle; | |||
4097 | if (!DestBundle->IsScheduled) | |||
4098 | BundleMember->incrementUnscheduledDeps(1); | |||
4099 | if (!DestBundle->hasValidDependencies()) | |||
4100 | WorkList.push_back(DestBundle); | |||
4101 | } | |||
4102 | } else { | |||
4103 | // I'm not sure if this can ever happen. But we need to be safe. | |||
4104 | // This lets the instruction/bundle never be scheduled and | |||
4105 | // eventually disable vectorization. | |||
4106 | BundleMember->Dependencies++; | |||
4107 | BundleMember->incrementUnscheduledDeps(1); | |||
4108 | } | |||
4109 | } | |||
4110 | } | |||
4111 | ||||
4112 | // Handle the memory dependencies. | |||
4113 | ScheduleData *DepDest = BundleMember->NextLoadStore; | |||
4114 | if (DepDest) { | |||
4115 | Instruction *SrcInst = BundleMember->Inst; | |||
4116 | MemoryLocation SrcLoc = getLocation(SrcInst, SLP->AA); | |||
4117 | bool SrcMayWrite = BundleMember->Inst->mayWriteToMemory(); | |||
4118 | unsigned numAliased = 0; | |||
4119 | unsigned DistToSrc = 1; | |||
4120 | ||||
4121 | while (DepDest) { | |||
4122 | assert(isInSchedulingRegion(DepDest))(static_cast <bool> (isInSchedulingRegion(DepDest)) ? void (0) : __assert_fail ("isInSchedulingRegion(DepDest)", "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4122, __extension__ __PRETTY_FUNCTION__)); | |||
4123 | ||||
4124 | // We have two limits to reduce the complexity: | |||
4125 | // 1) AliasedCheckLimit: It's a small limit to reduce calls to | |||
4126 | // SLP->isAliased (which is the expensive part in this loop). | |||
4127 | // 2) MaxMemDepDistance: It's for very large blocks and it aborts | |||
4128 | // the whole loop (even if the loop is fast, it's quadratic). | |||
4129 | // It's important for the loop break condition (see below) to | |||
4130 | // check this limit even between two read-only instructions. | |||
4131 | if (DistToSrc >= MaxMemDepDistance || | |||
4132 | ((SrcMayWrite || DepDest->Inst->mayWriteToMemory()) && | |||
4133 | (numAliased >= AliasedCheckLimit || | |||
4134 | SLP->isAliased(SrcLoc, SrcInst, DepDest->Inst)))) { | |||
4135 | ||||
4136 | // We increment the counter only if the locations are aliased | |||
4137 | // (instead of counting all alias checks). This gives a better | |||
4138 | // balance between reduced runtime and accurate dependencies. | |||
4139 | numAliased++; | |||
4140 | ||||
4141 | DepDest->MemoryDependencies.push_back(BundleMember); | |||
4142 | BundleMember->Dependencies++; | |||
4143 | ScheduleData *DestBundle = DepDest->FirstInBundle; | |||
4144 | if (!DestBundle->IsScheduled) { | |||
4145 | BundleMember->incrementUnscheduledDeps(1); | |||
4146 | } | |||
4147 | if (!DestBundle->hasValidDependencies()) { | |||
4148 | WorkList.push_back(DestBundle); | |||
4149 | } | |||
4150 | } | |||
4151 | DepDest = DepDest->NextLoadStore; | |||
4152 | ||||
4153 | // Example, explaining the loop break condition: Let's assume our | |||
4154 | // starting instruction is i0 and MaxMemDepDistance = 3. | |||
4155 | // | |||
4156 | // +--------v--v--v | |||
4157 | // i0,i1,i2,i3,i4,i5,i6,i7,i8 | |||
4158 | // +--------^--^--^ | |||
4159 | // | |||
4160 | // MaxMemDepDistance let us stop alias-checking at i3 and we add | |||
4161 | // dependencies from i0 to i3,i4,.. (even if they are not aliased). | |||
4162 | // Previously we already added dependencies from i3 to i6,i7,i8 | |||
4163 | // (because of MaxMemDepDistance). As we added a dependency from | |||
4164 | // i0 to i3, we have transitive dependencies from i0 to i6,i7,i8 | |||
4165 | // and we can abort this loop at i6. | |||
4166 | if (DistToSrc >= 2 * MaxMemDepDistance) | |||
4167 | break; | |||
4168 | DistToSrc++; | |||
4169 | } | |||
4170 | } | |||
4171 | } | |||
4172 | BundleMember = BundleMember->NextInBundle; | |||
4173 | } | |||
4174 | if (InsertInReadyList && SD->isReady()) { | |||
4175 | ReadyInsts.push_back(SD); | |||
4176 | DEBUG(dbgs() << "SLP: gets ready on update: " << *SD->Inst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: gets ready on update: " << *SD->Inst << "\n"; } } while (false); | |||
4177 | } | |||
4178 | } | |||
4179 | } | |||
4180 | ||||
4181 | void BoUpSLP::BlockScheduling::resetSchedule() { | |||
4182 | assert(ScheduleStart &&(static_cast <bool> (ScheduleStart && "tried to reset schedule on block which has not been scheduled" ) ? void (0) : __assert_fail ("ScheduleStart && \"tried to reset schedule on block which has not been scheduled\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4183, __extension__ __PRETTY_FUNCTION__)) | |||
4183 | "tried to reset schedule on block which has not been scheduled")(static_cast <bool> (ScheduleStart && "tried to reset schedule on block which has not been scheduled" ) ? void (0) : __assert_fail ("ScheduleStart && \"tried to reset schedule on block which has not been scheduled\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4183, __extension__ __PRETTY_FUNCTION__)); | |||
4184 | for (Instruction *I = ScheduleStart; I != ScheduleEnd; I = I->getNextNode()) { | |||
4185 | doForAllOpcodes(I, [&](ScheduleData *SD) { | |||
4186 | assert(isInSchedulingRegion(SD) &&(static_cast <bool> (isInSchedulingRegion(SD) && "ScheduleData not in scheduling region") ? void (0) : __assert_fail ("isInSchedulingRegion(SD) && \"ScheduleData not in scheduling region\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4187, __extension__ __PRETTY_FUNCTION__)) | |||
4187 | "ScheduleData not in scheduling region")(static_cast <bool> (isInSchedulingRegion(SD) && "ScheduleData not in scheduling region") ? void (0) : __assert_fail ("isInSchedulingRegion(SD) && \"ScheduleData not in scheduling region\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4187, __extension__ __PRETTY_FUNCTION__)); | |||
4188 | SD->IsScheduled = false; | |||
4189 | SD->resetUnscheduledDeps(); | |||
4190 | }); | |||
4191 | } | |||
4192 | ReadyInsts.clear(); | |||
4193 | } | |||
4194 | ||||
4195 | void BoUpSLP::scheduleBlock(BlockScheduling *BS) { | |||
4196 | if (!BS->ScheduleStart) | |||
4197 | return; | |||
4198 | ||||
4199 | 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); | |||
4200 | ||||
4201 | BS->resetSchedule(); | |||
4202 | ||||
4203 | // For the real scheduling we use a more sophisticated ready-list: it is | |||
4204 | // sorted by the original instruction location. This lets the final schedule | |||
4205 | // be as close as possible to the original instruction order. | |||
4206 | struct ScheduleDataCompare { | |||
4207 | bool operator()(ScheduleData *SD1, ScheduleData *SD2) const { | |||
4208 | return SD2->SchedulingPriority < SD1->SchedulingPriority; | |||
4209 | } | |||
4210 | }; | |||
4211 | std::set<ScheduleData *, ScheduleDataCompare> ReadyInsts; | |||
4212 | ||||
4213 | // Ensure that all dependency data is updated and fill the ready-list with | |||
4214 | // initial instructions. | |||
4215 | int Idx = 0; | |||
4216 | int NumToSchedule = 0; | |||
4217 | for (auto *I = BS->ScheduleStart; I != BS->ScheduleEnd; | |||
4218 | I = I->getNextNode()) { | |||
4219 | BS->doForAllOpcodes(I, [this, &Idx, &NumToSchedule, BS](ScheduleData *SD) { | |||
4220 | assert(SD->isPartOfBundle() ==(static_cast <bool> (SD->isPartOfBundle() == (getTreeEntry (SD->Inst) != nullptr) && "scheduler and vectorizer bundle mismatch" ) ? void (0) : __assert_fail ("SD->isPartOfBundle() == (getTreeEntry(SD->Inst) != nullptr) && \"scheduler and vectorizer bundle mismatch\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4222, __extension__ __PRETTY_FUNCTION__)) | |||
4221 | (getTreeEntry(SD->Inst) != nullptr) &&(static_cast <bool> (SD->isPartOfBundle() == (getTreeEntry (SD->Inst) != nullptr) && "scheduler and vectorizer bundle mismatch" ) ? void (0) : __assert_fail ("SD->isPartOfBundle() == (getTreeEntry(SD->Inst) != nullptr) && \"scheduler and vectorizer bundle mismatch\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4222, __extension__ __PRETTY_FUNCTION__)) | |||
4222 | "scheduler and vectorizer bundle mismatch")(static_cast <bool> (SD->isPartOfBundle() == (getTreeEntry (SD->Inst) != nullptr) && "scheduler and vectorizer bundle mismatch" ) ? void (0) : __assert_fail ("SD->isPartOfBundle() == (getTreeEntry(SD->Inst) != nullptr) && \"scheduler and vectorizer bundle mismatch\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4222, __extension__ __PRETTY_FUNCTION__)); | |||
4223 | SD->FirstInBundle->SchedulingPriority = Idx++; | |||
4224 | if (SD->isSchedulingEntity()) { | |||
4225 | BS->calculateDependencies(SD, false, this); | |||
4226 | NumToSchedule++; | |||
4227 | } | |||
4228 | }); | |||
4229 | } | |||
4230 | BS->initialFillReadyList(ReadyInsts); | |||
4231 | ||||
4232 | Instruction *LastScheduledInst = BS->ScheduleEnd; | |||
4233 | ||||
4234 | // Do the "real" scheduling. | |||
4235 | while (!ReadyInsts.empty()) { | |||
4236 | ScheduleData *picked = *ReadyInsts.begin(); | |||
4237 | ReadyInsts.erase(ReadyInsts.begin()); | |||
4238 | ||||
4239 | // Move the scheduled instruction(s) to their dedicated places, if not | |||
4240 | // there yet. | |||
4241 | ScheduleData *BundleMember = picked; | |||
4242 | while (BundleMember) { | |||
4243 | Instruction *pickedInst = BundleMember->Inst; | |||
4244 | if (LastScheduledInst->getNextNode() != pickedInst) { | |||
4245 | BS->BB->getInstList().remove(pickedInst); | |||
4246 | BS->BB->getInstList().insert(LastScheduledInst->getIterator(), | |||
4247 | pickedInst); | |||
4248 | } | |||
4249 | LastScheduledInst = pickedInst; | |||
4250 | BundleMember = BundleMember->NextInBundle; | |||
4251 | } | |||
4252 | ||||
4253 | BS->schedule(picked, ReadyInsts); | |||
4254 | NumToSchedule--; | |||
4255 | } | |||
4256 | assert(NumToSchedule == 0 && "could not schedule all instructions")(static_cast <bool> (NumToSchedule == 0 && "could not schedule all instructions" ) ? void (0) : __assert_fail ("NumToSchedule == 0 && \"could not schedule all instructions\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4256, __extension__ __PRETTY_FUNCTION__)); | |||
4257 | ||||
4258 | // Avoid duplicate scheduling of the block. | |||
4259 | BS->ScheduleStart = nullptr; | |||
4260 | } | |||
4261 | ||||
4262 | unsigned BoUpSLP::getVectorElementSize(Value *V) { | |||
4263 | // If V is a store, just return the width of the stored value without | |||
4264 | // traversing the expression tree. This is the common case. | |||
4265 | if (auto *Store = dyn_cast<StoreInst>(V)) | |||
4266 | return DL->getTypeSizeInBits(Store->getValueOperand()->getType()); | |||
4267 | ||||
4268 | // If V is not a store, we can traverse the expression tree to find loads | |||
4269 | // that feed it. The type of the loaded value may indicate a more suitable | |||
4270 | // width than V's type. We want to base the vector element size on the width | |||
4271 | // of memory operations where possible. | |||
4272 | SmallVector<Instruction *, 16> Worklist; | |||
4273 | SmallPtrSet<Instruction *, 16> Visited; | |||
4274 | if (auto *I = dyn_cast<Instruction>(V)) | |||
4275 | Worklist.push_back(I); | |||
4276 | ||||
4277 | // Traverse the expression tree in bottom-up order looking for loads. If we | |||
4278 | // encounter an instruciton we don't yet handle, we give up. | |||
4279 | auto MaxWidth = 0u; | |||
4280 | auto FoundUnknownInst = false; | |||
4281 | while (!Worklist.empty() && !FoundUnknownInst) { | |||
4282 | auto *I = Worklist.pop_back_val(); | |||
4283 | Visited.insert(I); | |||
4284 | ||||
4285 | // We should only be looking at scalar instructions here. If the current | |||
4286 | // instruction has a vector type, give up. | |||
4287 | auto *Ty = I->getType(); | |||
4288 | if (isa<VectorType>(Ty)) | |||
4289 | FoundUnknownInst = true; | |||
4290 | ||||
4291 | // If the current instruction is a load, update MaxWidth to reflect the | |||
4292 | // width of the loaded value. | |||
4293 | else if (isa<LoadInst>(I)) | |||
4294 | MaxWidth = std::max<unsigned>(MaxWidth, DL->getTypeSizeInBits(Ty)); | |||
4295 | ||||
4296 | // Otherwise, we need to visit the operands of the instruction. We only | |||
4297 | // handle the interesting cases from buildTree here. If an operand is an | |||
4298 | // instruction we haven't yet visited, we add it to the worklist. | |||
4299 | else if (isa<PHINode>(I) || isa<CastInst>(I) || isa<GetElementPtrInst>(I) || | |||
4300 | isa<CmpInst>(I) || isa<SelectInst>(I) || isa<BinaryOperator>(I)) { | |||
4301 | for (Use &U : I->operands()) | |||
4302 | if (auto *J = dyn_cast<Instruction>(U.get())) | |||
4303 | if (!Visited.count(J)) | |||
4304 | Worklist.push_back(J); | |||
4305 | } | |||
4306 | ||||
4307 | // If we don't yet handle the instruction, give up. | |||
4308 | else | |||
4309 | FoundUnknownInst = true; | |||
4310 | } | |||
4311 | ||||
4312 | // If we didn't encounter a memory access in the expression tree, or if we | |||
4313 | // gave up for some reason, just return the width of V. | |||
4314 | if (!MaxWidth || FoundUnknownInst) | |||
4315 | return DL->getTypeSizeInBits(V->getType()); | |||
4316 | ||||
4317 | // Otherwise, return the maximum width we found. | |||
4318 | return MaxWidth; | |||
4319 | } | |||
4320 | ||||
4321 | // Determine if a value V in a vectorizable expression Expr can be demoted to a | |||
4322 | // smaller type with a truncation. We collect the values that will be demoted | |||
4323 | // in ToDemote and additional roots that require investigating in Roots. | |||
4324 | static bool collectValuesToDemote(Value *V, SmallPtrSetImpl<Value *> &Expr, | |||
4325 | SmallVectorImpl<Value *> &ToDemote, | |||
4326 | SmallVectorImpl<Value *> &Roots) { | |||
4327 | // We can always demote constants. | |||
4328 | if (isa<Constant>(V)) { | |||
4329 | ToDemote.push_back(V); | |||
4330 | return true; | |||
4331 | } | |||
4332 | ||||
4333 | // If the value is not an instruction in the expression with only one use, it | |||
4334 | // cannot be demoted. | |||
4335 | auto *I = dyn_cast<Instruction>(V); | |||
4336 | if (!I || !I->hasOneUse() || !Expr.count(I)) | |||
4337 | return false; | |||
4338 | ||||
4339 | switch (I->getOpcode()) { | |||
4340 | ||||
4341 | // We can always demote truncations and extensions. Since truncations can | |||
4342 | // seed additional demotion, we save the truncated value. | |||
4343 | case Instruction::Trunc: | |||
4344 | Roots.push_back(I->getOperand(0)); | |||
4345 | break; | |||
4346 | case Instruction::ZExt: | |||
4347 | case Instruction::SExt: | |||
4348 | break; | |||
4349 | ||||
4350 | // We can demote certain binary operations if we can demote both of their | |||
4351 | // operands. | |||
4352 | case Instruction::Add: | |||
4353 | case Instruction::Sub: | |||
4354 | case Instruction::Mul: | |||
4355 | case Instruction::And: | |||
4356 | case Instruction::Or: | |||
4357 | case Instruction::Xor: | |||
4358 | if (!collectValuesToDemote(I->getOperand(0), Expr, ToDemote, Roots) || | |||
4359 | !collectValuesToDemote(I->getOperand(1), Expr, ToDemote, Roots)) | |||
4360 | return false; | |||
4361 | break; | |||
4362 | ||||
4363 | // We can demote selects if we can demote their true and false values. | |||
4364 | case Instruction::Select: { | |||
4365 | SelectInst *SI = cast<SelectInst>(I); | |||
4366 | if (!collectValuesToDemote(SI->getTrueValue(), Expr, ToDemote, Roots) || | |||
4367 | !collectValuesToDemote(SI->getFalseValue(), Expr, ToDemote, Roots)) | |||
4368 | return false; | |||
4369 | break; | |||
4370 | } | |||
4371 | ||||
4372 | // We can demote phis if we can demote all their incoming operands. Note that | |||
4373 | // we don't need to worry about cycles since we ensure single use above. | |||
4374 | case Instruction::PHI: { | |||
4375 | PHINode *PN = cast<PHINode>(I); | |||
4376 | for (Value *IncValue : PN->incoming_values()) | |||
4377 | if (!collectValuesToDemote(IncValue, Expr, ToDemote, Roots)) | |||
4378 | return false; | |||
4379 | break; | |||
4380 | } | |||
4381 | ||||
4382 | // Otherwise, conservatively give up. | |||
4383 | default: | |||
4384 | return false; | |||
4385 | } | |||
4386 | ||||
4387 | // Record the value that we can demote. | |||
4388 | ToDemote.push_back(V); | |||
4389 | return true; | |||
4390 | } | |||
4391 | ||||
4392 | void BoUpSLP::computeMinimumValueSizes() { | |||
4393 | // If there are no external uses, the expression tree must be rooted by a | |||
4394 | // store. We can't demote in-memory values, so there is nothing to do here. | |||
4395 | if (ExternalUses.empty()) | |||
4396 | return; | |||
4397 | ||||
4398 | // We only attempt to truncate integer expressions. | |||
4399 | auto &TreeRoot = VectorizableTree[0].Scalars; | |||
4400 | auto *TreeRootIT = dyn_cast<IntegerType>(TreeRoot[0]->getType()); | |||
4401 | if (!TreeRootIT) | |||
4402 | return; | |||
4403 | ||||
4404 | // If the expression is not rooted by a store, these roots should have | |||
4405 | // external uses. We will rely on InstCombine to rewrite the expression in | |||
4406 | // the narrower type. However, InstCombine only rewrites single-use values. | |||
4407 | // This means that if a tree entry other than a root is used externally, it | |||
4408 | // must have multiple uses and InstCombine will not rewrite it. The code | |||
4409 | // below ensures that only the roots are used externally. | |||
4410 | SmallPtrSet<Value *, 32> Expr(TreeRoot.begin(), TreeRoot.end()); | |||
4411 | for (auto &EU : ExternalUses) | |||
4412 | if (!Expr.erase(EU.Scalar)) | |||
4413 | return; | |||
4414 | if (!Expr.empty()) | |||
4415 | return; | |||
4416 | ||||
4417 | // Collect the scalar values of the vectorizable expression. We will use this | |||
4418 | // context to determine which values can be demoted. If we see a truncation, | |||
4419 | // we mark it as seeding another demotion. | |||
4420 | for (auto &Entry : VectorizableTree) | |||
4421 | Expr.insert(Entry.Scalars.begin(), Entry.Scalars.end()); | |||
4422 | ||||
4423 | // Ensure the roots of the vectorizable tree don't form a cycle. They must | |||
4424 | // have a single external user that is not in the vectorizable tree. | |||
4425 | for (auto *Root : TreeRoot) | |||
4426 | if (!Root->hasOneUse() || Expr.count(*Root->user_begin())) | |||
4427 | return; | |||
4428 | ||||
4429 | // Conservatively determine if we can actually truncate the roots of the | |||
4430 | // expression. Collect the values that can be demoted in ToDemote and | |||
4431 | // additional roots that require investigating in Roots. | |||
4432 | SmallVector<Value *, 32> ToDemote; | |||
4433 | SmallVector<Value *, 4> Roots; | |||
4434 | for (auto *Root : TreeRoot) | |||
4435 | if (!collectValuesToDemote(Root, Expr, ToDemote, Roots)) | |||
4436 | return; | |||
4437 | ||||
4438 | // The maximum bit width required to represent all the values that can be | |||
4439 | // demoted without loss of precision. It would be safe to truncate the roots | |||
4440 | // of the expression to this width. | |||
4441 | auto MaxBitWidth = 8u; | |||
4442 | ||||
4443 | // We first check if all the bits of the roots are demanded. If they're not, | |||
4444 | // we can truncate the roots to this narrower type. | |||
4445 | for (auto *Root : TreeRoot) { | |||
4446 | auto Mask = DB->getDemandedBits(cast<Instruction>(Root)); | |||
4447 | MaxBitWidth = std::max<unsigned>( | |||
4448 | Mask.getBitWidth() - Mask.countLeadingZeros(), MaxBitWidth); | |||
4449 | } | |||
4450 | ||||
4451 | // True if the roots can be zero-extended back to their original type, rather | |||
4452 | // than sign-extended. We know that if the leading bits are not demanded, we | |||
4453 | // can safely zero-extend. So we initialize IsKnownPositive to True. | |||
4454 | bool IsKnownPositive = true; | |||
4455 | ||||
4456 | // If all the bits of the roots are demanded, we can try a little harder to | |||
4457 | // compute a narrower type. This can happen, for example, if the roots are | |||
4458 | // getelementptr indices. InstCombine promotes these indices to the pointer | |||
4459 | // width. Thus, all their bits are technically demanded even though the | |||
4460 | // address computation might be vectorized in a smaller type. | |||
4461 | // | |||
4462 | // We start by looking at each entry that can be demoted. We compute the | |||
4463 | // maximum bit width required to store the scalar by using ValueTracking to | |||
4464 | // compute the number of high-order bits we can truncate. | |||
4465 | if (MaxBitWidth == DL->getTypeSizeInBits(TreeRoot[0]->getType()) && | |||
4466 | llvm::all_of(TreeRoot, [](Value *R) { | |||
4467 | assert(R->hasOneUse() && "Root should have only one use!")(static_cast <bool> (R->hasOneUse() && "Root should have only one use!" ) ? void (0) : __assert_fail ("R->hasOneUse() && \"Root should have only one use!\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4467, __extension__ __PRETTY_FUNCTION__)); | |||
4468 | return isa<GetElementPtrInst>(R->user_back()); | |||
4469 | })) { | |||
4470 | MaxBitWidth = 8u; | |||
4471 | ||||
4472 | // Determine if the sign bit of all the roots is known to be zero. If not, | |||
4473 | // IsKnownPositive is set to False. | |||
4474 | IsKnownPositive = llvm::all_of(TreeRoot, [&](Value *R) { | |||
4475 | KnownBits Known = computeKnownBits(R, *DL); | |||
4476 | return Known.isNonNegative(); | |||
4477 | }); | |||
4478 | ||||
4479 | // Determine the maximum number of bits required to store the scalar | |||
4480 | // values. | |||
4481 | for (auto *Scalar : ToDemote) { | |||
4482 | auto NumSignBits = ComputeNumSignBits(Scalar, *DL, 0, AC, nullptr, DT); | |||
4483 | auto NumTypeBits = DL->getTypeSizeInBits(Scalar->getType()); | |||
4484 | MaxBitWidth = std::max<unsigned>(NumTypeBits - NumSignBits, MaxBitWidth); | |||
4485 | } | |||
4486 | ||||
4487 | // If we can't prove that the sign bit is zero, we must add one to the | |||
4488 | // maximum bit width to account for the unknown sign bit. This preserves | |||
4489 | // the existing sign bit so we can safely sign-extend the root back to the | |||
4490 | // original type. Otherwise, if we know the sign bit is zero, we will | |||
4491 | // zero-extend the root instead. | |||
4492 | // | |||
4493 | // FIXME: This is somewhat suboptimal, as there will be cases where adding | |||
4494 | // one to the maximum bit width will yield a larger-than-necessary | |||
4495 | // type. In general, we need to add an extra bit only if we can't | |||
4496 | // prove that the upper bit of the original type is equal to the | |||
4497 | // upper bit of the proposed smaller type. If these two bits are the | |||
4498 | // same (either zero or one) we know that sign-extending from the | |||
4499 | // smaller type will result in the same value. Here, since we can't | |||
4500 | // yet prove this, we are just making the proposed smaller type | |||
4501 | // larger to ensure correctness. | |||
4502 | if (!IsKnownPositive) | |||
4503 | ++MaxBitWidth; | |||
4504 | } | |||
4505 | ||||
4506 | // Round MaxBitWidth up to the next power-of-two. | |||
4507 | if (!isPowerOf2_64(MaxBitWidth)) | |||
4508 | MaxBitWidth = NextPowerOf2(MaxBitWidth); | |||
4509 | ||||
4510 | // If the maximum bit width we compute is less than the with of the roots' | |||
4511 | // type, we can proceed with the narrowing. Otherwise, do nothing. | |||
4512 | if (MaxBitWidth >= TreeRootIT->getBitWidth()) | |||
4513 | return; | |||
4514 | ||||
4515 | // If we can truncate the root, we must collect additional values that might | |||
4516 | // be demoted as a result. That is, those seeded by truncations we will | |||
4517 | // modify. | |||
4518 | while (!Roots.empty()) | |||
4519 | collectValuesToDemote(Roots.pop_back_val(), Expr, ToDemote, Roots); | |||
4520 | ||||
4521 | // Finally, map the values we can demote to the maximum bit with we computed. | |||
4522 | for (auto *Scalar : ToDemote) | |||
4523 | MinBWs[Scalar] = std::make_pair(MaxBitWidth, !IsKnownPositive); | |||
4524 | } | |||
4525 | ||||
4526 | namespace { | |||
4527 | ||||
4528 | /// The SLPVectorizer Pass. | |||
4529 | struct SLPVectorizer : public FunctionPass { | |||
4530 | SLPVectorizerPass Impl; | |||
4531 | ||||
4532 | /// Pass identification, replacement for typeid | |||
4533 | static char ID; | |||
4534 | ||||
4535 | explicit SLPVectorizer() : FunctionPass(ID) { | |||
4536 | initializeSLPVectorizerPass(*PassRegistry::getPassRegistry()); | |||
4537 | } | |||
4538 | ||||
4539 | bool doInitialization(Module &M) override { | |||
4540 | return false; | |||
4541 | } | |||
4542 | ||||
4543 | bool runOnFunction(Function &F) override { | |||
4544 | if (skipFunction(F)) | |||
4545 | return false; | |||
4546 | ||||
4547 | auto *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE(); | |||
4548 | auto *TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F); | |||
4549 | auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>(); | |||
4550 | auto *TLI = TLIP ? &TLIP->getTLI() : nullptr; | |||
4551 | auto *AA = &getAnalysis<AAResultsWrapperPass>().getAAResults(); | |||
4552 | auto *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); | |||
4553 | auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); | |||
4554 | auto *AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F); | |||
4555 | auto *DB = &getAnalysis<DemandedBitsWrapperPass>().getDemandedBits(); | |||
4556 | auto *ORE = &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE(); | |||
4557 | ||||
4558 | return Impl.runImpl(F, SE, TTI, TLI, AA, LI, DT, AC, DB, ORE); | |||
4559 | } | |||
4560 | ||||
4561 | void getAnalysisUsage(AnalysisUsage &AU) const override { | |||
4562 | FunctionPass::getAnalysisUsage(AU); | |||
4563 | AU.addRequired<AssumptionCacheTracker>(); | |||
4564 | AU.addRequired<ScalarEvolutionWrapperPass>(); | |||
4565 | AU.addRequired<AAResultsWrapperPass>(); | |||
4566 | AU.addRequired<TargetTransformInfoWrapperPass>(); | |||
4567 | AU.addRequired<LoopInfoWrapperPass>(); | |||
4568 | AU.addRequired<DominatorTreeWrapperPass>(); | |||
4569 | AU.addRequired<DemandedBitsWrapperPass>(); | |||
4570 | AU.addRequired<OptimizationRemarkEmitterWrapperPass>(); | |||
4571 | AU.addPreserved<LoopInfoWrapperPass>(); | |||
4572 | AU.addPreserved<DominatorTreeWrapperPass>(); | |||
4573 | AU.addPreserved<AAResultsWrapperPass>(); | |||
4574 | AU.addPreserved<GlobalsAAWrapperPass>(); | |||
4575 | AU.setPreservesCFG(); | |||
4576 | } | |||
4577 | }; | |||
4578 | ||||
4579 | } // end anonymous namespace | |||
4580 | ||||
4581 | PreservedAnalyses SLPVectorizerPass::run(Function &F, FunctionAnalysisManager &AM) { | |||
4582 | auto *SE = &AM.getResult<ScalarEvolutionAnalysis>(F); | |||
4583 | auto *TTI = &AM.getResult<TargetIRAnalysis>(F); | |||
4584 | auto *TLI = AM.getCachedResult<TargetLibraryAnalysis>(F); | |||
4585 | auto *AA = &AM.getResult<AAManager>(F); | |||
4586 | auto *LI = &AM.getResult<LoopAnalysis>(F); | |||
4587 | auto *DT = &AM.getResult<DominatorTreeAnalysis>(F); | |||
4588 | auto *AC = &AM.getResult<AssumptionAnalysis>(F); | |||
4589 | auto *DB = &AM.getResult<DemandedBitsAnalysis>(F); | |||
4590 | auto *ORE = &AM.getResult<OptimizationRemarkEmitterAnalysis>(F); | |||
4591 | ||||
4592 | bool Changed = runImpl(F, SE, TTI, TLI, AA, LI, DT, AC, DB, ORE); | |||
4593 | if (!Changed) | |||
4594 | return PreservedAnalyses::all(); | |||
4595 | ||||
4596 | PreservedAnalyses PA; | |||
4597 | PA.preserveSet<CFGAnalyses>(); | |||
4598 | PA.preserve<AAManager>(); | |||
4599 | PA.preserve<GlobalsAA>(); | |||
4600 | return PA; | |||
4601 | } | |||
4602 | ||||
4603 | bool SLPVectorizerPass::runImpl(Function &F, ScalarEvolution *SE_, | |||
4604 | TargetTransformInfo *TTI_, | |||
4605 | TargetLibraryInfo *TLI_, AliasAnalysis *AA_, | |||
4606 | LoopInfo *LI_, DominatorTree *DT_, | |||
4607 | AssumptionCache *AC_, DemandedBits *DB_, | |||
4608 | OptimizationRemarkEmitter *ORE_) { | |||
4609 | SE = SE_; | |||
4610 | TTI = TTI_; | |||
4611 | TLI = TLI_; | |||
4612 | AA = AA_; | |||
4613 | LI = LI_; | |||
4614 | DT = DT_; | |||
4615 | AC = AC_; | |||
4616 | DB = DB_; | |||
4617 | DL = &F.getParent()->getDataLayout(); | |||
4618 | ||||
4619 | Stores.clear(); | |||
4620 | GEPs.clear(); | |||
4621 | bool Changed = false; | |||
4622 | ||||
4623 | // If the target claims to have no vector registers don't attempt | |||
4624 | // vectorization. | |||
4625 | if (!TTI->getNumberOfRegisters(true)) | |||
4626 | return false; | |||
4627 | ||||
4628 | // Don't vectorize when the attribute NoImplicitFloat is used. | |||
4629 | if (F.hasFnAttribute(Attribute::NoImplicitFloat)) | |||
4630 | return false; | |||
4631 | ||||
4632 | 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); | |||
4633 | ||||
4634 | // Use the bottom up slp vectorizer to construct chains that start with | |||
4635 | // store instructions. | |||
4636 | BoUpSLP R(&F, SE, TTI, TLI, AA, LI, DT, AC, DB, DL, ORE_); | |||
4637 | ||||
4638 | // A general note: the vectorizer must use BoUpSLP::eraseInstruction() to | |||
4639 | // delete instructions. | |||
4640 | ||||
4641 | // Scan the blocks in the function in post order. | |||
4642 | for (auto BB : post_order(&F.getEntryBlock())) { | |||
4643 | collectSeedInstructions(BB); | |||
4644 | ||||
4645 | // Vectorize trees that end at stores. | |||
4646 | if (!Stores.empty()) { | |||
4647 | 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) | |||
4648 | << " underlying objects.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Found stores for " << Stores .size() << " underlying objects.\n"; } } while (false); | |||
4649 | Changed |= vectorizeStoreChains(R); | |||
4650 | } | |||
4651 | ||||
4652 | // Vectorize trees that end at reductions. | |||
4653 | Changed |= vectorizeChainsInBlock(BB, R); | |||
4654 | ||||
4655 | // Vectorize the index computations of getelementptr instructions. This | |||
4656 | // is primarily intended to catch gather-like idioms ending at | |||
4657 | // non-consecutive loads. | |||
4658 | if (!GEPs.empty()) { | |||
4659 | 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) | |||
4660 | << " underlying objects.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Found GEPs for " << GEPs .size() << " underlying objects.\n"; } } while (false); | |||
4661 | Changed |= vectorizeGEPIndices(BB, R); | |||
4662 | } | |||
4663 | } | |||
4664 | ||||
4665 | if (Changed) { | |||
4666 | R.optimizeGatherSequence(); | |||
4667 | DEBUG(dbgs() << "SLP: vectorized \"" << F.getName() << "\"\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: vectorized \"" << F.getName () << "\"\n"; } } while (false); | |||
4668 | DEBUG(verifyFunction(F))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { verifyFunction(F); } } while (false); | |||
4669 | } | |||
4670 | return Changed; | |||
4671 | } | |||
4672 | ||||
4673 | /// \brief Check that the Values in the slice in VL array are still existent in | |||
4674 | /// the WeakTrackingVH array. | |||
4675 | /// Vectorization of part of the VL array may cause later values in the VL array | |||
4676 | /// to become invalid. We track when this has happened in the WeakTrackingVH | |||
4677 | /// array. | |||
4678 | static bool hasValueBeenRAUWed(ArrayRef<Value *> VL, | |||
4679 | ArrayRef<WeakTrackingVH> VH, unsigned SliceBegin, | |||
4680 | unsigned SliceSize) { | |||
4681 | VL = VL.slice(SliceBegin, SliceSize); | |||
4682 | VH = VH.slice(SliceBegin, SliceSize); | |||
4683 | return !std::equal(VL.begin(), VL.end(), VH.begin()); | |||
4684 | } | |||
4685 | ||||
4686 | bool SLPVectorizerPass::vectorizeStoreChain(ArrayRef<Value *> Chain, BoUpSLP &R, | |||
4687 | unsigned VecRegSize) { | |||
4688 | const unsigned ChainLen = Chain.size(); | |||
4689 | DEBUG(dbgs() << "SLP: Analyzing a store chain of length " << ChainLendo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing a store chain of length " << ChainLen << "\n"; } } while (false) | |||
4690 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing a store chain of length " << ChainLen << "\n"; } } while (false); | |||
4691 | const unsigned Sz = R.getVectorElementSize(Chain[0]); | |||
4692 | const unsigned VF = VecRegSize / Sz; | |||
4693 | ||||
4694 | if (!isPowerOf2_32(Sz) || VF < 2) | |||
4695 | return false; | |||
4696 | ||||
4697 | // Keep track of values that were deleted by vectorizing in the loop below. | |||
4698 | const SmallVector<WeakTrackingVH, 8> TrackValues(Chain.begin(), Chain.end()); | |||
4699 | ||||
4700 | bool Changed = false; | |||
4701 | // Look for profitable vectorizable trees at all offsets, starting at zero. | |||
4702 | for (unsigned i = 0, e = ChainLen; i + VF <= e; ++i) { | |||
4703 | ||||
4704 | // Check that a previous iteration of this loop did not delete the Value. | |||
4705 | if (hasValueBeenRAUWed(Chain, TrackValues, i, VF)) | |||
4706 | continue; | |||
4707 | ||||
4708 | DEBUG(dbgs() << "SLP: Analyzing " << VF << " stores at offset " << ido { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing " << VF << " stores at offset " << i << "\n"; } } while (false ) | |||
4709 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing " << VF << " stores at offset " << i << "\n"; } } while (false ); | |||
4710 | ArrayRef<Value *> Operands = Chain.slice(i, VF); | |||
4711 | ||||
4712 | R.buildTree(Operands); | |||
4713 | if (R.isTreeTinyAndNotFullyVectorizable()) | |||
4714 | continue; | |||
4715 | ||||
4716 | R.computeMinimumValueSizes(); | |||
4717 | ||||
4718 | int Cost = R.getTreeCost(); | |||
4719 | ||||
4720 | 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); | |||
4721 | if (Cost < -SLPCostThreshold) { | |||
4722 | 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); | |||
4723 | ||||
4724 | using namespace ore; | |||
4725 | ||||
4726 | R.getORE()->emit(OptimizationRemark(SV_NAME"slp-vectorizer", "StoresVectorized", | |||
4727 | cast<StoreInst>(Chain[i])) | |||
4728 | << "Stores SLP vectorized with cost " << NV("Cost", Cost) | |||
4729 | << " and with tree size " | |||
4730 | << NV("TreeSize", R.getTreeSize())); | |||
4731 | ||||
4732 | R.vectorizeTree(); | |||
4733 | ||||
4734 | // Move to the next bundle. | |||
4735 | i += VF - 1; | |||
4736 | Changed = true; | |||
4737 | } | |||
4738 | } | |||
4739 | ||||
4740 | return Changed; | |||
4741 | } | |||
4742 | ||||
4743 | bool SLPVectorizerPass::vectorizeStores(ArrayRef<StoreInst *> Stores, | |||
4744 | BoUpSLP &R) { | |||
4745 | SetVector<StoreInst *> Heads; | |||
4746 | SmallDenseSet<StoreInst *> Tails; | |||
4747 | SmallDenseMap<StoreInst *, StoreInst *> ConsecutiveChain; | |||
4748 | ||||
4749 | // We may run into multiple chains that merge into a single chain. We mark the | |||
4750 | // stores that we vectorized so that we don't visit the same store twice. | |||
4751 | BoUpSLP::ValueSet VectorizedStores; | |||
4752 | bool Changed = false; | |||
4753 | ||||
4754 | // Do a quadratic search on all of the given stores in reverse order and find | |||
4755 | // all of the pairs of stores that follow each other. | |||
4756 | SmallVector<unsigned, 16> IndexQueue; | |||
4757 | unsigned E = Stores.size(); | |||
4758 | IndexQueue.resize(E - 1); | |||
4759 | for (unsigned I = E; I > 0; --I) { | |||
4760 | unsigned Idx = I - 1; | |||
4761 | // If a store has multiple consecutive store candidates, search Stores | |||
4762 | // array according to the sequence: Idx-1, Idx+1, Idx-2, Idx+2, ... | |||
4763 | // This is because usually pairing with immediate succeeding or preceding | |||
4764 | // candidate create the best chance to find slp vectorization opportunity. | |||
4765 | unsigned Offset = 1; | |||
4766 | unsigned Cnt = 0; | |||
4767 | for (unsigned J = 0; J < E - 1; ++J, ++Offset) { | |||
4768 | if (Idx >= Offset) { | |||
4769 | IndexQueue[Cnt] = Idx - Offset; | |||
4770 | ++Cnt; | |||
4771 | } | |||
4772 | if (Idx + Offset < E) { | |||
4773 | IndexQueue[Cnt] = Idx + Offset; | |||
4774 | ++Cnt; | |||
4775 | } | |||
4776 | } | |||
4777 | ||||
4778 | for (auto K : IndexQueue) { | |||
4779 | if (isConsecutiveAccess(Stores[K], Stores[Idx], *DL, *SE)) { | |||
4780 | Tails.insert(Stores[Idx]); | |||
4781 | Heads.insert(Stores[K]); | |||
4782 | ConsecutiveChain[Stores[K]] = Stores[Idx]; | |||
4783 | break; | |||
4784 | } | |||
4785 | } | |||
4786 | } | |||
4787 | ||||
4788 | // For stores that start but don't end a link in the chain: | |||
4789 | for (auto *SI : llvm::reverse(Heads)) { | |||
4790 | if (Tails.count(SI)) | |||
4791 | continue; | |||
4792 | ||||
4793 | // We found a store instr that starts a chain. Now follow the chain and try | |||
4794 | // to vectorize it. | |||
4795 | BoUpSLP::ValueList Operands; | |||
4796 | StoreInst *I = SI; | |||
4797 | // Collect the chain into a list. | |||
4798 | while ((Tails.count(I) || Heads.count(I)) && !VectorizedStores.count(I)) { | |||
4799 | Operands.push_back(I); | |||
4800 | // Move to the next value in the chain. | |||
4801 | I = ConsecutiveChain[I]; | |||
4802 | } | |||
4803 | ||||
4804 | // FIXME: Is division-by-2 the correct step? Should we assert that the | |||
4805 | // register size is a power-of-2? | |||
4806 | for (unsigned Size = R.getMaxVecRegSize(); Size >= R.getMinVecRegSize(); | |||
4807 | Size /= 2) { | |||
4808 | if (vectorizeStoreChain(Operands, R, Size)) { | |||
4809 | // Mark the vectorized stores so that we don't vectorize them again. | |||
4810 | VectorizedStores.insert(Operands.begin(), Operands.end()); | |||
4811 | Changed = true; | |||
4812 | break; | |||
4813 | } | |||
4814 | } | |||
4815 | } | |||
4816 | ||||
4817 | return Changed; | |||
4818 | } | |||
4819 | ||||
4820 | void SLPVectorizerPass::collectSeedInstructions(BasicBlock *BB) { | |||
4821 | // Initialize the collections. We will make a single pass over the block. | |||
4822 | Stores.clear(); | |||
4823 | GEPs.clear(); | |||
4824 | ||||
4825 | // Visit the store and getelementptr instructions in BB and organize them in | |||
4826 | // Stores and GEPs according to the underlying objects of their pointer | |||
4827 | // operands. | |||
4828 | for (Instruction &I : *BB) { | |||
4829 | // Ignore store instructions that are volatile or have a pointer operand | |||
4830 | // that doesn't point to a scalar type. | |||
4831 | if (auto *SI = dyn_cast<StoreInst>(&I)) { | |||
4832 | if (!SI->isSimple()) | |||
4833 | continue; | |||
4834 | if (!isValidElementType(SI->getValueOperand()->getType())) | |||
4835 | continue; | |||
4836 | Stores[GetUnderlyingObject(SI->getPointerOperand(), *DL)].push_back(SI); | |||
4837 | } | |||
4838 | ||||
4839 | // Ignore getelementptr instructions that have more than one index, a | |||
4840 | // constant index, or a pointer operand that doesn't point to a scalar | |||
4841 | // type. | |||
4842 | else if (auto *GEP = dyn_cast<GetElementPtrInst>(&I)) { | |||
4843 | auto Idx = GEP->idx_begin()->get(); | |||
4844 | if (GEP->getNumIndices() > 1 || isa<Constant>(Idx)) | |||
4845 | continue; | |||
4846 | if (!isValidElementType(Idx->getType())) | |||
4847 | continue; | |||
4848 | if (GEP->getType()->isVectorTy()) | |||
4849 | continue; | |||
4850 | GEPs[GetUnderlyingObject(GEP->getPointerOperand(), *DL)].push_back(GEP); | |||
4851 | } | |||
4852 | } | |||
4853 | } | |||
4854 | ||||
4855 | bool SLPVectorizerPass::tryToVectorizePair(Value *A, Value *B, BoUpSLP &R) { | |||
4856 | if (!A || !B) | |||
4857 | return false; | |||
4858 | Value *VL[] = { A, B }; | |||
4859 | return tryToVectorizeList(VL, R, /*UserCost=*/0, true); | |||
4860 | } | |||
4861 | ||||
4862 | bool SLPVectorizerPass::tryToVectorizeList(ArrayRef<Value *> VL, BoUpSLP &R, | |||
4863 | int UserCost, bool AllowReorder) { | |||
4864 | if (VL.size() < 2) | |||
4865 | return false; | |||
4866 | ||||
4867 | DEBUG(dbgs() << "SLP: Trying to vectorize a list of length = " << VL.size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Trying to vectorize a list of length = " << VL.size() << ".\n"; } } while (false) | |||
4868 | << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Trying to vectorize a list of length = " << VL.size() << ".\n"; } } while (false); | |||
4869 | ||||
4870 | // Check that all of the parts are scalar instructions of the same type. | |||
4871 | Instruction *I0 = dyn_cast<Instruction>(VL[0]); | |||
4872 | if (!I0) | |||
4873 | return false; | |||
4874 | ||||
4875 | unsigned Opcode0 = I0->getOpcode(); | |||
4876 | ||||
4877 | unsigned Sz = R.getVectorElementSize(I0); | |||
4878 | unsigned MinVF = std::max(2U, R.getMinVecRegSize() / Sz); | |||
4879 | unsigned MaxVF = std::max<unsigned>(PowerOf2Floor(VL.size()), MinVF); | |||
4880 | if (MaxVF < 2) { | |||
4881 | R.getORE()->emit([&]() { | |||
4882 | return OptimizationRemarkMissed( | |||
4883 | SV_NAME"slp-vectorizer", "SmallVF", I0) | |||
4884 | << "Cannot SLP vectorize list: vectorization factor " | |||
4885 | << "less than 2 is not supported"; | |||
4886 | }); | |||
4887 | return false; | |||
4888 | } | |||
4889 | ||||
4890 | for (Value *V : VL) { | |||
4891 | Type *Ty = V->getType(); | |||
4892 | if (!isValidElementType(Ty)) { | |||
4893 | // NOTE: the following will give user internal llvm type name, which may not be useful | |||
4894 | R.getORE()->emit([&]() { | |||
4895 | std::string type_str; | |||
4896 | llvm::raw_string_ostream rso(type_str); | |||
4897 | Ty->print(rso); | |||
4898 | return OptimizationRemarkMissed( | |||
4899 | SV_NAME"slp-vectorizer", "UnsupportedType", I0) | |||
4900 | << "Cannot SLP vectorize list: type " | |||
4901 | << rso.str() + " is unsupported by vectorizer"; | |||
4902 | }); | |||
4903 | return false; | |||
4904 | } | |||
4905 | Instruction *Inst = dyn_cast<Instruction>(V); | |||
4906 | ||||
4907 | if (!Inst) | |||
4908 | return false; | |||
4909 | if (Inst->getOpcode() != Opcode0) { | |||
4910 | R.getORE()->emit([&]() { | |||
4911 | return OptimizationRemarkMissed( | |||
4912 | SV_NAME"slp-vectorizer", "InequableTypes", I0) | |||
4913 | << "Cannot SLP vectorize list: not all of the " | |||
4914 | << "parts of scalar instructions are of the same type: " | |||
4915 | << ore::NV("Instruction1Opcode", I0) << " and " | |||
4916 | << ore::NV("Instruction2Opcode", Inst); | |||
4917 | }); | |||
4918 | return false; | |||
4919 | } | |||
4920 | } | |||
4921 | ||||
4922 | bool Changed = false; | |||
4923 | bool CandidateFound = false; | |||
4924 | int MinCost = SLPCostThreshold; | |||
4925 | ||||
4926 | // Keep track of values that were deleted by vectorizing in the loop below. | |||
4927 | SmallVector<WeakTrackingVH, 8> TrackValues(VL.begin(), VL.end()); | |||
4928 | ||||
4929 | unsigned NextInst = 0, MaxInst = VL.size(); | |||
4930 | for (unsigned VF = MaxVF; NextInst + 1 < MaxInst && VF >= MinVF; | |||
4931 | VF /= 2) { | |||
4932 | // No actual vectorization should happen, if number of parts is the same as | |||
4933 | // provided vectorization factor (i.e. the scalar type is used for vector | |||
4934 | // code during codegen). | |||
4935 | auto *VecTy = VectorType::get(VL[0]->getType(), VF); | |||
4936 | if (TTI->getNumberOfParts(VecTy) == VF) | |||
4937 | continue; | |||
4938 | for (unsigned I = NextInst; I < MaxInst; ++I) { | |||
4939 | unsigned OpsWidth = 0; | |||
4940 | ||||
4941 | if (I + VF > MaxInst) | |||
4942 | OpsWidth = MaxInst - I; | |||
4943 | else | |||
4944 | OpsWidth = VF; | |||
4945 | ||||
4946 | if (!isPowerOf2_32(OpsWidth) || OpsWidth < 2) | |||
4947 | break; | |||
4948 | ||||
4949 | // Check that a previous iteration of this loop did not delete the Value. | |||
4950 | if (hasValueBeenRAUWed(VL, TrackValues, I, OpsWidth)) | |||
4951 | continue; | |||
4952 | ||||
4953 | DEBUG(dbgs() << "SLP: Analyzing " << OpsWidth << " operations "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing " << OpsWidth << " operations " << "\n"; } } while (false) | |||
4954 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing " << OpsWidth << " operations " << "\n"; } } while (false); | |||
4955 | ArrayRef<Value *> Ops = VL.slice(I, OpsWidth); | |||
4956 | ||||
4957 | R.buildTree(Ops); | |||
4958 | Optional<ArrayRef<unsigned>> Order = R.bestOrder(); | |||
4959 | // TODO: check if we can allow reordering for more cases. | |||
4960 | if (AllowReorder && Order) { | |||
4961 | // TODO: reorder tree nodes without tree rebuilding. | |||
4962 | // Conceptually, there is nothing actually preventing us from trying to | |||
4963 | // reorder a larger list. In fact, we do exactly this when vectorizing | |||
4964 | // reductions. However, at this point, we only expect to get here when | |||
4965 | // there are exactly two operations. | |||
4966 | assert(Ops.size() == 2)(static_cast <bool> (Ops.size() == 2) ? void (0) : __assert_fail ("Ops.size() == 2", "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4966, __extension__ __PRETTY_FUNCTION__)); | |||
4967 | Value *ReorderedOps[] = {Ops[1], Ops[0]}; | |||
4968 | R.buildTree(ReorderedOps, None); | |||
4969 | } | |||
4970 | if (R.isTreeTinyAndNotFullyVectorizable()) | |||
4971 | continue; | |||
4972 | ||||
4973 | R.computeMinimumValueSizes(); | |||
4974 | int Cost = R.getTreeCost() - UserCost; | |||
4975 | CandidateFound = true; | |||
4976 | MinCost = std::min(MinCost, Cost); | |||
4977 | ||||
4978 | if (Cost < -SLPCostThreshold) { | |||
4979 | 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); | |||
4980 | R.getORE()->emit(OptimizationRemark(SV_NAME"slp-vectorizer", "VectorizedList", | |||
4981 | cast<Instruction>(Ops[0])) | |||
4982 | << "SLP vectorized with cost " << ore::NV("Cost", Cost) | |||
4983 | << " and with tree size " | |||
4984 | << ore::NV("TreeSize", R.getTreeSize())); | |||
4985 | ||||
4986 | R.vectorizeTree(); | |||
4987 | // Move to the next bundle. | |||
4988 | I += VF - 1; | |||
4989 | NextInst = I + 1; | |||
4990 | Changed = true; | |||
4991 | } | |||
4992 | } | |||
4993 | } | |||
4994 | ||||
4995 | if (!Changed && CandidateFound) { | |||
4996 | R.getORE()->emit([&]() { | |||
4997 | return OptimizationRemarkMissed( | |||
4998 | SV_NAME"slp-vectorizer", "NotBeneficial", I0) | |||
4999 | << "List vectorization was possible but not beneficial with cost " | |||
5000 | << ore::NV("Cost", MinCost) << " >= " | |||
5001 | << ore::NV("Treshold", -SLPCostThreshold); | |||
5002 | }); | |||
5003 | } else if (!Changed) { | |||
5004 | R.getORE()->emit([&]() { | |||
5005 | return OptimizationRemarkMissed( | |||
5006 | SV_NAME"slp-vectorizer", "NotPossible", I0) | |||
5007 | << "Cannot SLP vectorize list: vectorization was impossible" | |||
5008 | << " with available vectorization factors"; | |||
5009 | }); | |||
5010 | } | |||
5011 | return Changed; | |||
5012 | } | |||
5013 | ||||
5014 | bool SLPVectorizerPass::tryToVectorize(Instruction *I, BoUpSLP &R) { | |||
5015 | if (!I) | |||
5016 | return false; | |||
5017 | ||||
5018 | if (!isa<BinaryOperator>(I) && !isa<CmpInst>(I)) | |||
5019 | return false; | |||
5020 | ||||
5021 | Value *P = I->getParent(); | |||
5022 | ||||
5023 | // Vectorize in current basic block only. | |||
5024 | auto *Op0 = dyn_cast<Instruction>(I->getOperand(0)); | |||
5025 | auto *Op1 = dyn_cast<Instruction>(I->getOperand(1)); | |||
5026 | if (!Op0 || !Op1 || Op0->getParent() != P || Op1->getParent() != P) | |||
5027 | return false; | |||
5028 | ||||
5029 | // Try to vectorize V. | |||
5030 | if (tryToVectorizePair(Op0, Op1, R)) | |||
5031 | return true; | |||
5032 | ||||
5033 | auto *A = dyn_cast<BinaryOperator>(Op0); | |||
5034 | auto *B = dyn_cast<BinaryOperator>(Op1); | |||
5035 | // Try to skip B. | |||
5036 | if (B && B->hasOneUse()) { | |||
5037 | auto *B0 = dyn_cast<BinaryOperator>(B->getOperand(0)); | |||
5038 | auto *B1 = dyn_cast<BinaryOperator>(B->getOperand(1)); | |||
5039 | if (B0 && B0->getParent() == P && tryToVectorizePair(A, B0, R)) | |||
5040 | return true; | |||
5041 | if (B1 && B1->getParent() == P && tryToVectorizePair(A, B1, R)) | |||
5042 | return true; | |||
5043 | } | |||
5044 | ||||
5045 | // Try to skip A. | |||
5046 | if (A && A->hasOneUse()) { | |||
5047 | auto *A0 = dyn_cast<BinaryOperator>(A->getOperand(0)); | |||
5048 | auto *A1 = dyn_cast<BinaryOperator>(A->getOperand(1)); | |||
5049 | if (A0 && A0->getParent() == P && tryToVectorizePair(A0, B, R)) | |||
5050 | return true; | |||
5051 | if (A1 && A1->getParent() == P && tryToVectorizePair(A1, B, R)) | |||
5052 | return true; | |||
5053 | } | |||
5054 | return false; | |||
5055 | } | |||
5056 | ||||
5057 | /// \brief Generate a shuffle mask to be used in a reduction tree. | |||
5058 | /// | |||
5059 | /// \param VecLen The length of the vector to be reduced. | |||
5060 | /// \param NumEltsToRdx The number of elements that should be reduced in the | |||
5061 | /// vector. | |||
5062 | /// \param IsPairwise Whether the reduction is a pairwise or splitting | |||
5063 | /// reduction. A pairwise reduction will generate a mask of | |||
5064 | /// <0,2,...> or <1,3,..> while a splitting reduction will generate | |||
5065 | /// <2,3, undef,undef> for a vector of 4 and NumElts = 2. | |||
5066 | /// \param IsLeft True will generate a mask of even elements, odd otherwise. | |||
5067 | static Value *createRdxShuffleMask(unsigned VecLen, unsigned NumEltsToRdx, | |||
5068 | bool IsPairwise, bool IsLeft, | |||
5069 | IRBuilder<> &Builder) { | |||
5070 | assert((IsPairwise || !IsLeft) && "Don't support a <0,1,undef,...> mask")(static_cast <bool> ((IsPairwise || !IsLeft) && "Don't support a <0,1,undef,...> mask") ? void (0) : __assert_fail ("(IsPairwise || !IsLeft) && \"Don't support a <0,1,undef,...> mask\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5070, __extension__ __PRETTY_FUNCTION__)); | |||
5071 | ||||
5072 | SmallVector<Constant *, 32> ShuffleMask( | |||
5073 | VecLen, UndefValue::get(Builder.getInt32Ty())); | |||
5074 | ||||
5075 | if (IsPairwise) | |||
5076 | // Build a mask of 0, 2, ... (left) or 1, 3, ... (right). | |||
5077 | for (unsigned i = 0; i != NumEltsToRdx; ++i) | |||
5078 | ShuffleMask[i] = Builder.getInt32(2 * i + !IsLeft); | |||
5079 | else | |||
5080 | // Move the upper half of the vector to the lower half. | |||
5081 | for (unsigned i = 0; i != NumEltsToRdx; ++i) | |||
5082 | ShuffleMask[i] = Builder.getInt32(NumEltsToRdx + i); | |||
5083 | ||||
5084 | return ConstantVector::get(ShuffleMask); | |||
5085 | } | |||
5086 | ||||
5087 | namespace { | |||
5088 | ||||
5089 | /// Model horizontal reductions. | |||
5090 | /// | |||
5091 | /// A horizontal reduction is a tree of reduction operations (currently add and | |||
5092 | /// fadd) that has operations that can be put into a vector as its leaf. | |||
5093 | /// For example, this tree: | |||
5094 | /// | |||
5095 | /// mul mul mul mul | |||
5096 | /// \ / \ / | |||
5097 | /// + + | |||
5098 | /// \ / | |||
5099 | /// + | |||
5100 | /// This tree has "mul" as its reduced values and "+" as its reduction | |||
5101 | /// operations. A reduction might be feeding into a store or a binary operation | |||
5102 | /// feeding a phi. | |||
5103 | /// ... | |||
5104 | /// \ / | |||
5105 | /// + | |||
5106 | /// | | |||
5107 | /// phi += | |||
5108 | /// | |||
5109 | /// Or: | |||
5110 | /// ... | |||
5111 | /// \ / | |||
5112 | /// + | |||
5113 | /// | | |||
5114 | /// *p = | |||
5115 | /// | |||
5116 | class HorizontalReduction { | |||
5117 | using ReductionOpsType = SmallVector<Value *, 16>; | |||
5118 | using ReductionOpsListType = SmallVector<ReductionOpsType, 2>; | |||
5119 | ReductionOpsListType ReductionOps; | |||
5120 | SmallVector<Value *, 32> ReducedVals; | |||
5121 | // Use map vector to make stable output. | |||
5122 | MapVector<Instruction *, Value *> ExtraArgs; | |||
5123 | ||||
5124 | /// Kind of the reduction data. | |||
5125 | enum ReductionKind { | |||
5126 | RK_None, /// Not a reduction. | |||
5127 | RK_Arithmetic, /// Binary reduction data. | |||
5128 | RK_Min, /// Minimum reduction data. | |||
5129 | RK_UMin, /// Unsigned minimum reduction data. | |||
5130 | RK_Max, /// Maximum reduction data. | |||
5131 | RK_UMax, /// Unsigned maximum reduction data. | |||
5132 | }; | |||
5133 | ||||
5134 | /// Contains info about operation, like its opcode, left and right operands. | |||
5135 | class OperationData { | |||
5136 | /// Opcode of the instruction. | |||
5137 | unsigned Opcode = 0; | |||
5138 | ||||
5139 | /// Left operand of the reduction operation. | |||
5140 | Value *LHS = nullptr; | |||
5141 | ||||
5142 | /// Right operand of the reduction operation. | |||
5143 | Value *RHS = nullptr; | |||
5144 | ||||
5145 | /// Kind of the reduction operation. | |||
5146 | ReductionKind Kind = RK_None; | |||
5147 | ||||
5148 | /// True if float point min/max reduction has no NaNs. | |||
5149 | bool NoNaN = false; | |||
5150 | ||||
5151 | /// Checks if the reduction operation can be vectorized. | |||
5152 | bool isVectorizable() const { | |||
5153 | return LHS && RHS && | |||
5154 | // We currently only support adds && min/max reductions. | |||
5155 | ((Kind == RK_Arithmetic && | |||
5156 | (Opcode == Instruction::Add || Opcode == Instruction::FAdd)) || | |||
5157 | ((Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) && | |||
5158 | (Kind == RK_Min || Kind == RK_Max)) || | |||
5159 | (Opcode == Instruction::ICmp && | |||
5160 | (Kind == RK_UMin || Kind == RK_UMax))); | |||
5161 | } | |||
5162 | ||||
5163 | /// Creates reduction operation with the current opcode. | |||
5164 | Value *createOp(IRBuilder<> &Builder, const Twine &Name) const { | |||
5165 | assert(isVectorizable() &&(static_cast <bool> (isVectorizable() && "Expected add|fadd or min/max reduction operation." ) ? void (0) : __assert_fail ("isVectorizable() && \"Expected add|fadd or min/max reduction operation.\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5166, __extension__ __PRETTY_FUNCTION__)) | |||
5166 | "Expected add|fadd or min/max reduction operation.")(static_cast <bool> (isVectorizable() && "Expected add|fadd or min/max reduction operation." ) ? void (0) : __assert_fail ("isVectorizable() && \"Expected add|fadd or min/max reduction operation.\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5166, __extension__ __PRETTY_FUNCTION__)); | |||
5167 | Value *Cmp; | |||
5168 | switch (Kind) { | |||
5169 | case RK_Arithmetic: | |||
5170 | return Builder.CreateBinOp((Instruction::BinaryOps)Opcode, LHS, RHS, | |||
5171 | Name); | |||
5172 | case RK_Min: | |||
5173 | Cmp = Opcode == Instruction::ICmp ? Builder.CreateICmpSLT(LHS, RHS) | |||
5174 | : Builder.CreateFCmpOLT(LHS, RHS); | |||
5175 | break; | |||
5176 | case RK_Max: | |||
5177 | Cmp = Opcode == Instruction::ICmp ? Builder.CreateICmpSGT(LHS, RHS) | |||
5178 | : Builder.CreateFCmpOGT(LHS, RHS); | |||
5179 | break; | |||
5180 | case RK_UMin: | |||
5181 | assert(Opcode == Instruction::ICmp && "Expected integer types.")(static_cast <bool> (Opcode == Instruction::ICmp && "Expected integer types.") ? void (0) : __assert_fail ("Opcode == Instruction::ICmp && \"Expected integer types.\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5181, __extension__ __PRETTY_FUNCTION__)); | |||
5182 | Cmp = Builder.CreateICmpULT(LHS, RHS); | |||
5183 | break; | |||
5184 | case RK_UMax: | |||
5185 | assert(Opcode == Instruction::ICmp && "Expected integer types.")(static_cast <bool> (Opcode == Instruction::ICmp && "Expected integer types.") ? void (0) : __assert_fail ("Opcode == Instruction::ICmp && \"Expected integer types.\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5185, __extension__ __PRETTY_FUNCTION__)); | |||
5186 | Cmp = Builder.CreateICmpUGT(LHS, RHS); | |||
5187 | break; | |||
5188 | case RK_None: | |||
5189 | llvm_unreachable("Unknown reduction operation.")::llvm::llvm_unreachable_internal("Unknown reduction operation." , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5189); | |||
5190 | } | |||
5191 | return Builder.CreateSelect(Cmp, LHS, RHS, Name); | |||
5192 | } | |||
5193 | ||||
5194 | public: | |||
5195 | explicit OperationData() = default; | |||
5196 | ||||
5197 | /// Construction for reduced values. They are identified by opcode only and | |||
5198 | /// don't have associated LHS/RHS values. | |||
5199 | explicit OperationData(Value *V) { | |||
5200 | if (auto *I = dyn_cast<Instruction>(V)) | |||
5201 | Opcode = I->getOpcode(); | |||
5202 | } | |||
5203 | ||||
5204 | /// Constructor for reduction operations with opcode and its left and | |||
5205 | /// right operands. | |||
5206 | OperationData(unsigned Opcode, Value *LHS, Value *RHS, ReductionKind Kind, | |||
5207 | bool NoNaN = false) | |||
5208 | : Opcode(Opcode), LHS(LHS), RHS(RHS), Kind(Kind), NoNaN(NoNaN) { | |||
5209 | assert(Kind != RK_None && "One of the reduction operations is expected.")(static_cast <bool> (Kind != RK_None && "One of the reduction operations is expected." ) ? void (0) : __assert_fail ("Kind != RK_None && \"One of the reduction operations is expected.\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5209, __extension__ __PRETTY_FUNCTION__)); | |||
5210 | } | |||
5211 | ||||
5212 | explicit operator bool() const { return Opcode; } | |||
5213 | ||||
5214 | /// Get the index of the first operand. | |||
5215 | unsigned getFirstOperandIndex() const { | |||
5216 | assert(!!*this && "The opcode is not set.")(static_cast <bool> (!!*this && "The opcode is not set." ) ? void (0) : __assert_fail ("!!*this && \"The opcode is not set.\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5216, __extension__ __PRETTY_FUNCTION__)); | |||
5217 | switch (Kind) { | |||
5218 | case RK_Min: | |||
5219 | case RK_UMin: | |||
5220 | case RK_Max: | |||
5221 | case RK_UMax: | |||
5222 | return 1; | |||
5223 | case RK_Arithmetic: | |||
5224 | case RK_None: | |||
5225 | break; | |||
5226 | } | |||
5227 | return 0; | |||
5228 | } | |||
5229 | ||||
5230 | /// Total number of operands in the reduction operation. | |||
5231 | unsigned getNumberOfOperands() const { | |||
5232 | assert(Kind != RK_None && !!*this && LHS && RHS &&(static_cast <bool> (Kind != RK_None && !!*this && LHS && RHS && "Expected reduction operation." ) ? void (0) : __assert_fail ("Kind != RK_None && !!*this && LHS && RHS && \"Expected reduction operation.\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5233, __extension__ __PRETTY_FUNCTION__)) | |||
5233 | "Expected reduction operation.")(static_cast <bool> (Kind != RK_None && !!*this && LHS && RHS && "Expected reduction operation." ) ? void (0) : __assert_fail ("Kind != RK_None && !!*this && LHS && RHS && \"Expected reduction operation.\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5233, __extension__ __PRETTY_FUNCTION__)); | |||
5234 | switch (Kind) { | |||
5235 | case RK_Arithmetic: | |||
5236 | return 2; | |||
5237 | case RK_Min: | |||
5238 | case RK_UMin: | |||
5239 | case RK_Max: | |||
5240 | case RK_UMax: | |||
5241 | return 3; | |||
5242 | case RK_None: | |||
5243 | break; | |||
5244 | } | |||
5245 | llvm_unreachable("Reduction kind is not set")::llvm::llvm_unreachable_internal("Reduction kind is not set" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5245); | |||
5246 | } | |||
5247 | ||||
5248 | /// Checks if the operation has the same parent as \p P. | |||
5249 | bool hasSameParent(Instruction *I, Value *P, bool IsRedOp) const { | |||
5250 | assert(Kind != RK_None && !!*this && LHS && RHS &&(static_cast <bool> (Kind != RK_None && !!*this && LHS && RHS && "Expected reduction operation." ) ? void (0) : __assert_fail ("Kind != RK_None && !!*this && LHS && RHS && \"Expected reduction operation.\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5251, __extension__ __PRETTY_FUNCTION__)) | |||
5251 | "Expected reduction operation.")(static_cast <bool> (Kind != RK_None && !!*this && LHS && RHS && "Expected reduction operation." ) ? void (0) : __assert_fail ("Kind != RK_None && !!*this && LHS && RHS && \"Expected reduction operation.\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5251, __extension__ __PRETTY_FUNCTION__)); | |||
5252 | if (!IsRedOp) | |||
5253 | return I->getParent() == P; | |||
5254 | switch (Kind) { | |||
5255 | case RK_Arithmetic: | |||
5256 | // Arithmetic reduction operation must be used once only. | |||
5257 | return I->getParent() == P; | |||
5258 | case RK_Min: | |||
5259 | case RK_UMin: | |||
5260 | case RK_Max: | |||
5261 | case RK_UMax: { | |||
5262 | // SelectInst must be used twice while the condition op must have single | |||
5263 | // use only. | |||
5264 | auto *Cmp = cast<Instruction>(cast<SelectInst>(I)->getCondition()); | |||
5265 | return I->getParent() == P && Cmp && Cmp->getParent() == P; | |||
5266 | } | |||
5267 | case RK_None: | |||
5268 | break; | |||
5269 | } | |||
5270 | llvm_unreachable("Reduction kind is not set")::llvm::llvm_unreachable_internal("Reduction kind is not set" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5270); | |||
5271 | } | |||
5272 | /// Expected number of uses for reduction operations/reduced values. | |||
5273 | bool hasRequiredNumberOfUses(Instruction *I, bool IsReductionOp) const { | |||
5274 | assert(Kind != RK_None && !!*this && LHS && RHS &&(static_cast <bool> (Kind != RK_None && !!*this && LHS && RHS && "Expected reduction operation." ) ? void (0) : __assert_fail ("Kind != RK_None && !!*this && LHS && RHS && \"Expected reduction operation.\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5275, __extension__ __PRETTY_FUNCTION__)) | |||
5275 | "Expected reduction operation.")(static_cast <bool> (Kind != RK_None && !!*this && LHS && RHS && "Expected reduction operation." ) ? void (0) : __assert_fail ("Kind != RK_None && !!*this && LHS && RHS && \"Expected reduction operation.\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5275, __extension__ __PRETTY_FUNCTION__)); | |||
5276 | switch (Kind) { | |||
5277 | case RK_Arithmetic: | |||
5278 | return I->hasOneUse(); | |||
5279 | case RK_Min: | |||
5280 | case RK_UMin: | |||
5281 | case RK_Max: | |||
5282 | case RK_UMax: | |||
5283 | return I->hasNUses(2) && | |||
5284 | (!IsReductionOp || | |||
5285 | cast<SelectInst>(I)->getCondition()->hasOneUse()); | |||
5286 | case RK_None: | |||
5287 | break; | |||
5288 | } | |||
5289 | llvm_unreachable("Reduction kind is not set")::llvm::llvm_unreachable_internal("Reduction kind is not set" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5289); | |||
5290 | } | |||
5291 | ||||
5292 | /// Initializes the list of reduction operations. | |||
5293 | void initReductionOps(ReductionOpsListType &ReductionOps) { | |||
5294 | assert(Kind != RK_None && !!*this && LHS && RHS &&(static_cast <bool> (Kind != RK_None && !!*this && LHS && RHS && "Expected reduction operation." ) ? void (0) : __assert_fail ("Kind != RK_None && !!*this && LHS && RHS && \"Expected reduction operation.\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5295, __extension__ __PRETTY_FUNCTION__)) | |||
5295 | "Expected reduction operation.")(static_cast <bool> (Kind != RK_None && !!*this && LHS && RHS && "Expected reduction operation." ) ? void (0) : __assert_fail ("Kind != RK_None && !!*this && LHS && RHS && \"Expected reduction operation.\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5295, __extension__ __PRETTY_FUNCTION__)); | |||
5296 | switch (Kind) { | |||
5297 | case RK_Arithmetic: | |||
5298 | ReductionOps.assign(1, ReductionOpsType()); | |||
5299 | break; | |||
5300 | case RK_Min: | |||
5301 | case RK_UMin: | |||
5302 | case RK_Max: | |||
5303 | case RK_UMax: | |||
5304 | ReductionOps.assign(2, ReductionOpsType()); | |||
5305 | break; | |||
5306 | case RK_None: | |||
5307 | llvm_unreachable("Reduction kind is not set")::llvm::llvm_unreachable_internal("Reduction kind is not set" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5307); | |||
5308 | } | |||
5309 | } | |||
5310 | /// Add all reduction operations for the reduction instruction \p I. | |||
5311 | void addReductionOps(Instruction *I, ReductionOpsListType &ReductionOps) { | |||
5312 | assert(Kind != RK_None && !!*this && LHS && RHS &&(static_cast <bool> (Kind != RK_None && !!*this && LHS && RHS && "Expected reduction operation." ) ? void (0) : __assert_fail ("Kind != RK_None && !!*this && LHS && RHS && \"Expected reduction operation.\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5313, __extension__ __PRETTY_FUNCTION__)) | |||
5313 | "Expected reduction operation.")(static_cast <bool> (Kind != RK_None && !!*this && LHS && RHS && "Expected reduction operation." ) ? void (0) : __assert_fail ("Kind != RK_None && !!*this && LHS && RHS && \"Expected reduction operation.\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5313, __extension__ __PRETTY_FUNCTION__)); | |||
5314 | switch (Kind) { | |||
5315 | case RK_Arithmetic: | |||
5316 | ReductionOps[0].emplace_back(I); | |||
5317 | break; | |||
5318 | case RK_Min: | |||
5319 | case RK_UMin: | |||
5320 | case RK_Max: | |||
5321 | case RK_UMax: | |||
5322 | ReductionOps[0].emplace_back(cast<SelectInst>(I)->getCondition()); | |||
5323 | ReductionOps[1].emplace_back(I); | |||
5324 | break; | |||
5325 | case RK_None: | |||
5326 | llvm_unreachable("Reduction kind is not set")::llvm::llvm_unreachable_internal("Reduction kind is not set" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5326); | |||
5327 | } | |||
5328 | } | |||
5329 | ||||
5330 | /// Checks if instruction is associative and can be vectorized. | |||
5331 | bool isAssociative(Instruction *I) const { | |||
5332 | assert(Kind != RK_None && *this && LHS && RHS &&(static_cast <bool> (Kind != RK_None && *this && LHS && RHS && "Expected reduction operation." ) ? void (0) : __assert_fail ("Kind != RK_None && *this && LHS && RHS && \"Expected reduction operation.\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5333, __extension__ __PRETTY_FUNCTION__)) | |||
5333 | "Expected reduction operation.")(static_cast <bool> (Kind != RK_None && *this && LHS && RHS && "Expected reduction operation." ) ? void (0) : __assert_fail ("Kind != RK_None && *this && LHS && RHS && \"Expected reduction operation.\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5333, __extension__ __PRETTY_FUNCTION__)); | |||
5334 | switch (Kind) { | |||
5335 | case RK_Arithmetic: | |||
5336 | return I->isAssociative(); | |||
5337 | case RK_Min: | |||
5338 | case RK_Max: | |||
5339 | return Opcode == Instruction::ICmp || | |||
5340 | cast<Instruction>(I->getOperand(0))->isFast(); | |||
5341 | case RK_UMin: | |||
5342 | case RK_UMax: | |||
5343 | assert(Opcode == Instruction::ICmp &&(static_cast <bool> (Opcode == Instruction::ICmp && "Only integer compare operation is expected.") ? void (0) : __assert_fail ("Opcode == Instruction::ICmp && \"Only integer compare operation is expected.\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5344, __extension__ __PRETTY_FUNCTION__)) | |||
5344 | "Only integer compare operation is expected.")(static_cast <bool> (Opcode == Instruction::ICmp && "Only integer compare operation is expected.") ? void (0) : __assert_fail ("Opcode == Instruction::ICmp && \"Only integer compare operation is expected.\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5344, __extension__ __PRETTY_FUNCTION__)); | |||
5345 | return true; | |||
5346 | case RK_None: | |||
5347 | break; | |||
5348 | } | |||
5349 | llvm_unreachable("Reduction kind is not set")::llvm::llvm_unreachable_internal("Reduction kind is not set" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5349); | |||
5350 | } | |||
5351 | ||||
5352 | /// Checks if the reduction operation can be vectorized. | |||
5353 | bool isVectorizable(Instruction *I) const { | |||
5354 | return isVectorizable() && isAssociative(I); | |||
5355 | } | |||
5356 | ||||
5357 | /// Checks if two operation data are both a reduction op or both a reduced | |||
5358 | /// value. | |||
5359 | bool operator==(const OperationData &OD) { | |||
5360 | assert(((Kind != OD.Kind) || ((!LHS == !OD.LHS) && (!RHS == !OD.RHS))) &&(static_cast <bool> (((Kind != OD.Kind) || ((!LHS == !OD .LHS) && (!RHS == !OD.RHS))) && "One of the comparing operations is incorrect." ) ? void (0) : __assert_fail ("((Kind != OD.Kind) || ((!LHS == !OD.LHS) && (!RHS == !OD.RHS))) && \"One of the comparing operations is incorrect.\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5361, __extension__ __PRETTY_FUNCTION__)) | |||
5361 | "One of the comparing operations is incorrect.")(static_cast <bool> (((Kind != OD.Kind) || ((!LHS == !OD .LHS) && (!RHS == !OD.RHS))) && "One of the comparing operations is incorrect." ) ? void (0) : __assert_fail ("((Kind != OD.Kind) || ((!LHS == !OD.LHS) && (!RHS == !OD.RHS))) && \"One of the comparing operations is incorrect.\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5361, __extension__ __PRETTY_FUNCTION__)); | |||
5362 | return this == &OD || (Kind == OD.Kind && Opcode == OD.Opcode); | |||
5363 | } | |||
5364 | bool operator!=(const OperationData &OD) { return !(*this == OD); } | |||
5365 | void clear() { | |||
5366 | Opcode = 0; | |||
5367 | LHS = nullptr; | |||
5368 | RHS = nullptr; | |||
5369 | Kind = RK_None; | |||
5370 | NoNaN = false; | |||
5371 | } | |||
5372 | ||||
5373 | /// Get the opcode of the reduction operation. | |||
5374 | unsigned getOpcode() const { | |||
5375 | assert(isVectorizable() && "Expected vectorizable operation.")(static_cast <bool> (isVectorizable() && "Expected vectorizable operation." ) ? void (0) : __assert_fail ("isVectorizable() && \"Expected vectorizable operation.\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5375, __extension__ __PRETTY_FUNCTION__)); | |||
5376 | return Opcode; | |||
5377 | } | |||
5378 | ||||
5379 | /// Get kind of reduction data. | |||
5380 | ReductionKind getKind() const { return Kind; } | |||
5381 | Value *getLHS() const { return LHS; } | |||
5382 | Value *getRHS() const { return RHS; } | |||
5383 | Type *getConditionType() const { | |||
5384 | switch (Kind) { | |||
5385 | case RK_Arithmetic: | |||
5386 | return nullptr; | |||
5387 | case RK_Min: | |||
5388 | case RK_Max: | |||
5389 | case RK_UMin: | |||
5390 | case RK_UMax: | |||
5391 | return CmpInst::makeCmpResultType(LHS->getType()); | |||
5392 | case RK_None: | |||
5393 | break; | |||
5394 | } | |||
5395 | llvm_unreachable("Reduction kind is not set")::llvm::llvm_unreachable_internal("Reduction kind is not set" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5395); | |||
5396 | } | |||
5397 | ||||
5398 | /// Creates reduction operation with the current opcode with the IR flags | |||
5399 | /// from \p ReductionOps. | |||
5400 | Value *createOp(IRBuilder<> &Builder, const Twine &Name, | |||
5401 | const ReductionOpsListType &ReductionOps) const { | |||
5402 | assert(isVectorizable() &&(static_cast <bool> (isVectorizable() && "Expected add|fadd or min/max reduction operation." ) ? void (0) : __assert_fail ("isVectorizable() && \"Expected add|fadd or min/max reduction operation.\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5403, __extension__ __PRETTY_FUNCTION__)) | |||
5403 | "Expected add|fadd or min/max reduction operation.")(static_cast <bool> (isVectorizable() && "Expected add|fadd or min/max reduction operation." ) ? void (0) : __assert_fail ("isVectorizable() && \"Expected add|fadd or min/max reduction operation.\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5403, __extension__ __PRETTY_FUNCTION__)); | |||
5404 | auto *Op = createOp(Builder, Name); | |||
5405 | switch (Kind) { | |||
5406 | case RK_Arithmetic: | |||
5407 | propagateIRFlags(Op, ReductionOps[0]); | |||
5408 | return Op; | |||
5409 | case RK_Min: | |||
5410 | case RK_Max: | |||
5411 | case RK_UMin: | |||
5412 | case RK_UMax: | |||
5413 | if (auto *SI = dyn_cast<SelectInst>(Op)) | |||
5414 | propagateIRFlags(SI->getCondition(), ReductionOps[0]); | |||
5415 | propagateIRFlags(Op, ReductionOps[1]); | |||
5416 | return Op; | |||
5417 | case RK_None: | |||
5418 | break; | |||
5419 | } | |||
5420 | llvm_unreachable("Unknown reduction operation.")::llvm::llvm_unreachable_internal("Unknown reduction operation." , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5420); | |||
5421 | } | |||
5422 | /// Creates reduction operation with the current opcode with the IR flags | |||
5423 | /// from \p I. | |||
5424 | Value *createOp(IRBuilder<> &Builder, const Twine &Name, | |||
5425 | Instruction *I) const { | |||
5426 | assert(isVectorizable() &&(static_cast <bool> (isVectorizable() && "Expected add|fadd or min/max reduction operation." ) ? void (0) : __assert_fail ("isVectorizable() && \"Expected add|fadd or min/max reduction operation.\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5427, __extension__ __PRETTY_FUNCTION__)) | |||
5427 | "Expected add|fadd or min/max reduction operation.")(static_cast <bool> (isVectorizable() && "Expected add|fadd or min/max reduction operation." ) ? void (0) : __assert_fail ("isVectorizable() && \"Expected add|fadd or min/max reduction operation.\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5427, __extension__ __PRETTY_FUNCTION__)); | |||
5428 | auto *Op = createOp(Builder, Name); | |||
5429 | switch (Kind) { | |||
5430 | case RK_Arithmetic: | |||
5431 | propagateIRFlags(Op, I); | |||
5432 | return Op; | |||
5433 | case RK_Min: | |||
5434 | case RK_Max: | |||
5435 | case RK_UMin: | |||
5436 | case RK_UMax: | |||
5437 | if (auto *SI = dyn_cast<SelectInst>(Op)) { | |||
5438 | propagateIRFlags(SI->getCondition(), | |||
5439 | cast<SelectInst>(I)->getCondition()); | |||
5440 | } | |||
5441 | propagateIRFlags(Op, I); | |||
5442 | return Op; | |||
5443 | case RK_None: | |||
5444 | break; | |||
5445 | } | |||
5446 | llvm_unreachable("Unknown reduction operation.")::llvm::llvm_unreachable_internal("Unknown reduction operation." , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5446); | |||
5447 | } | |||
5448 | ||||
5449 | TargetTransformInfo::ReductionFlags getFlags() const { | |||
5450 | TargetTransformInfo::ReductionFlags Flags; | |||
5451 | Flags.NoNaN = NoNaN; | |||
5452 | switch (Kind) { | |||
5453 | case RK_Arithmetic: | |||
5454 | break; | |||
5455 | case RK_Min: | |||
5456 | Flags.IsSigned = Opcode == Instruction::ICmp; | |||
5457 | Flags.IsMaxOp = false; | |||
5458 | break; | |||
5459 | case RK_Max: | |||
5460 | Flags.IsSigned = Opcode == Instruction::ICmp; | |||
5461 | Flags.IsMaxOp = true; | |||
5462 | break; | |||
5463 | case RK_UMin: | |||
5464 | Flags.IsSigned = false; | |||
5465 | Flags.IsMaxOp = false; | |||
5466 | break; | |||
5467 | case RK_UMax: | |||
5468 | Flags.IsSigned = false; | |||
5469 | Flags.IsMaxOp = true; | |||
5470 | break; | |||
5471 | case RK_None: | |||
5472 | llvm_unreachable("Reduction kind is not set")::llvm::llvm_unreachable_internal("Reduction kind is not set" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5472); | |||
5473 | } | |||
5474 | return Flags; | |||
5475 | } | |||
5476 | }; | |||
5477 | ||||
5478 | Instruction *ReductionRoot = nullptr; | |||
5479 | ||||
5480 | /// The operation data of the reduction operation. | |||
5481 | OperationData ReductionData; | |||
5482 | ||||
5483 | /// The operation data of the values we perform a reduction on. | |||
5484 | OperationData ReducedValueData; | |||
5485 | ||||
5486 | /// Should we model this reduction as a pairwise reduction tree or a tree that | |||
5487 | /// splits the vector in halves and adds those halves. | |||
5488 | bool IsPairwiseReduction = false; | |||
5489 | ||||
5490 | /// Checks if the ParentStackElem.first should be marked as a reduction | |||
5491 | /// operation with an extra argument or as extra argument itself. | |||
5492 | void markExtraArg(std::pair<Instruction *, unsigned> &ParentStackElem, | |||
5493 | Value *ExtraArg) { | |||
5494 | if (ExtraArgs.count(ParentStackElem.first)) { | |||
5495 | ExtraArgs[ParentStackElem.first] = nullptr; | |||
5496 | // We ran into something like: | |||
5497 | // ParentStackElem.first = ExtraArgs[ParentStackElem.first] + ExtraArg. | |||
5498 | // The whole ParentStackElem.first should be considered as an extra value | |||
5499 | // in this case. | |||
5500 | // Do not perform analysis of remaining operands of ParentStackElem.first | |||
5501 | // instruction, this whole instruction is an extra argument. | |||
5502 | ParentStackElem.second = ParentStackElem.first->getNumOperands(); | |||
5503 | } else { | |||
5504 | // We ran into something like: | |||
5505 | // ParentStackElem.first += ... + ExtraArg + ... | |||
5506 | ExtraArgs[ParentStackElem.first] = ExtraArg; | |||
5507 | } | |||
5508 | } | |||
5509 | ||||
5510 | static OperationData getOperationData(Value *V) { | |||
5511 | if (!V) | |||
5512 | return OperationData(); | |||
5513 | ||||
5514 | Value *LHS; | |||
5515 | Value *RHS; | |||
5516 | if (m_BinOp(m_Value(LHS), m_Value(RHS)).match(V)) { | |||
5517 | return OperationData(cast<BinaryOperator>(V)->getOpcode(), LHS, RHS, | |||
5518 | RK_Arithmetic); | |||
5519 | } | |||
5520 | if (auto *Select = dyn_cast<SelectInst>(V)) { | |||
5521 | // Look for a min/max pattern. | |||
5522 | if (m_UMin(m_Value(LHS), m_Value(RHS)).match(Select)) { | |||
5523 | return OperationData(Instruction::ICmp, LHS, RHS, RK_UMin); | |||
5524 | } else if (m_SMin(m_Value(LHS), m_Value(RHS)).match(Select)) { | |||
5525 | return OperationData(Instruction::ICmp, LHS, RHS, RK_Min); | |||
5526 | } else if (m_OrdFMin(m_Value(LHS), m_Value(RHS)).match(Select) || | |||
5527 | m_UnordFMin(m_Value(LHS), m_Value(RHS)).match(Select)) { | |||
5528 | return OperationData( | |||
5529 | Instruction::FCmp, LHS, RHS, RK_Min, | |||
5530 | cast<Instruction>(Select->getCondition())->hasNoNaNs()); | |||
5531 | } else if (m_UMax(m_Value(LHS), m_Value(RHS)).match(Select)) { | |||
5532 | return OperationData(Instruction::ICmp, LHS, RHS, RK_UMax); | |||
5533 | } else if (m_SMax(m_Value(LHS), m_Value(RHS)).match(Select)) { | |||
5534 | return OperationData(Instruction::ICmp, LHS, RHS, RK_Max); | |||
5535 | } else if (m_OrdFMax(m_Value(LHS), m_Value(RHS)).match(Select) || | |||
5536 | m_UnordFMax(m_Value(LHS), m_Value(RHS)).match(Select)) { | |||
5537 | return OperationData( | |||
5538 | Instruction::FCmp, LHS, RHS, RK_Max, | |||
5539 | cast<Instruction>(Select->getCondition())->hasNoNaNs()); | |||
5540 | } | |||
5541 | } | |||
5542 | return OperationData(V); | |||
5543 | } | |||
5544 | ||||
5545 | public: | |||
5546 | HorizontalReduction() = default; | |||
5547 | ||||
5548 | /// \brief Try to find a reduction tree. | |||
5549 | bool matchAssociativeReduction(PHINode *Phi, Instruction *B) { | |||
5550 | assert((!Phi || is_contained(Phi->operands(), B)) &&(static_cast <bool> ((!Phi || is_contained(Phi->operands (), B)) && "Thi phi needs to use the binary operator" ) ? void (0) : __assert_fail ("(!Phi || is_contained(Phi->operands(), B)) && \"Thi phi needs to use the binary operator\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5551, __extension__ __PRETTY_FUNCTION__)) | |||
5551 | "Thi phi needs to use the binary operator")(static_cast <bool> ((!Phi || is_contained(Phi->operands (), B)) && "Thi phi needs to use the binary operator" ) ? void (0) : __assert_fail ("(!Phi || is_contained(Phi->operands(), B)) && \"Thi phi needs to use the binary operator\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5551, __extension__ __PRETTY_FUNCTION__)); | |||
5552 | ||||
5553 | ReductionData = getOperationData(B); | |||
5554 | ||||
5555 | // We could have a initial reductions that is not an add. | |||
5556 | // r *= v1 + v2 + v3 + v4 | |||
5557 | // In such a case start looking for a tree rooted in the first '+'. | |||
5558 | if (Phi) { | |||
5559 | if (ReductionData.getLHS() == Phi) { | |||
5560 | Phi = nullptr; | |||
5561 | B = dyn_cast<Instruction>(ReductionData.getRHS()); | |||
5562 | ReductionData = getOperationData(B); | |||
5563 | } else if (ReductionData.getRHS() == Phi) { | |||
5564 | Phi = nullptr; | |||
5565 | B = dyn_cast<Instruction>(ReductionData.getLHS()); | |||
5566 | ReductionData = getOperationData(B); | |||
5567 | } | |||
5568 | } | |||
5569 | ||||
5570 | if (!ReductionData.isVectorizable(B)) | |||
5571 | return false; | |||
5572 | ||||
5573 | Type *Ty = B->getType(); | |||
5574 | if (!isValidElementType(Ty)) | |||
5575 | return false; | |||
5576 | ||||
5577 | ReducedValueData.clear(); | |||
5578 | ReductionRoot = B; | |||
5579 | ||||
5580 | // Post order traverse the reduction tree starting at B. We only handle true | |||
5581 | // trees containing only binary operators. | |||
5582 | SmallVector<std::pair<Instruction *, unsigned>, 32> Stack; | |||
5583 | Stack.push_back(std::make_pair(B, ReductionData.getFirstOperandIndex())); | |||
5584 | ReductionData.initReductionOps(ReductionOps); | |||
5585 | while (!Stack.empty()) { | |||
5586 | Instruction *TreeN = Stack.back().first; | |||
5587 | unsigned EdgeToVist = Stack.back().second++; | |||
5588 | OperationData OpData = getOperationData(TreeN); | |||
5589 | bool IsReducedValue = OpData != ReductionData; | |||
5590 | ||||
5591 | // Postorder vist. | |||
5592 | if (IsReducedValue || EdgeToVist == OpData.getNumberOfOperands()) { | |||
5593 | if (IsReducedValue) | |||
5594 | ReducedVals.push_back(TreeN); | |||
5595 | else { | |||
5596 | auto I = ExtraArgs.find(TreeN); | |||
5597 | if (I != ExtraArgs.end() && !I->second) { | |||
5598 | // Check if TreeN is an extra argument of its parent operation. | |||
5599 | if (Stack.size() <= 1) { | |||
5600 | // TreeN can't be an extra argument as it is a root reduction | |||
5601 | // operation. | |||
5602 | return false; | |||
5603 | } | |||
5604 | // Yes, TreeN is an extra argument, do not add it to a list of | |||
5605 | // reduction operations. | |||
5606 | // Stack[Stack.size() - 2] always points to the parent operation. | |||
5607 | markExtraArg(Stack[Stack.size() - 2], TreeN); | |||
5608 | ExtraArgs.erase(TreeN); | |||
5609 | } else | |||
5610 | ReductionData.addReductionOps(TreeN, ReductionOps); | |||
5611 | } | |||
5612 | // Retract. | |||
5613 | Stack.pop_back(); | |||
5614 | continue; | |||
5615 | } | |||
5616 | ||||
5617 | // Visit left or right. | |||
5618 | Value *NextV = TreeN->getOperand(EdgeToVist); | |||
5619 | if (NextV != Phi) { | |||
5620 | auto *I = dyn_cast<Instruction>(NextV); | |||
5621 | OpData = getOperationData(I); | |||
5622 | // Continue analysis if the next operand is a reduction operation or | |||
5623 | // (possibly) a reduced value. If the reduced value opcode is not set, | |||
5624 | // the first met operation != reduction operation is considered as the | |||
5625 | // reduced value class. | |||
5626 | if (I && (!ReducedValueData || OpData == ReducedValueData || | |||
5627 | OpData == ReductionData)) { | |||
5628 | const bool IsReductionOperation = OpData == ReductionData; | |||
5629 | // Only handle trees in the current basic block. | |||
5630 | if (!ReductionData.hasSameParent(I, B->getParent(), | |||
5631 | IsReductionOperation)) { | |||
5632 | // I is an extra argument for TreeN (its parent operation). | |||
5633 | markExtraArg(Stack.back(), I); | |||
5634 | continue; | |||
5635 | } | |||
5636 | ||||
5637 | // Each tree node needs to have minimal number of users except for the | |||
5638 | // ultimate reduction. | |||
5639 | if (!ReductionData.hasRequiredNumberOfUses(I, | |||
5640 | OpData == ReductionData) && | |||
5641 | I != B) { | |||
5642 | // I is an extra argument for TreeN (its parent operation). | |||
5643 | markExtraArg(Stack.back(), I); | |||
5644 | continue; | |||
5645 | } | |||
5646 | ||||
5647 | if (IsReductionOperation) { | |||
5648 | // We need to be able to reassociate the reduction operations. | |||
5649 | if (!OpData.isAssociative(I)) { | |||
5650 | // I is an extra argument for TreeN (its parent operation). | |||
5651 | markExtraArg(Stack.back(), I); | |||
5652 | continue; | |||
5653 | } | |||
5654 | } else if (ReducedValueData && | |||
5655 | ReducedValueData != OpData) { | |||
5656 | // Make sure that the opcodes of the operations that we are going to | |||
5657 | // reduce match. | |||
5658 | // I is an extra argument for TreeN (its parent operation). | |||
5659 | markExtraArg(Stack.back(), I); | |||
5660 | continue; | |||
5661 | } else if (!ReducedValueData) | |||
5662 | ReducedValueData = OpData; | |||
5663 | ||||
5664 | Stack.push_back(std::make_pair(I, OpData.getFirstOperandIndex())); | |||
5665 | continue; | |||
5666 | } | |||
5667 | } | |||
5668 | // NextV is an extra argument for TreeN (its parent operation). | |||
5669 | markExtraArg(Stack.back(), NextV); | |||
5670 | } | |||
5671 | return true; | |||
5672 | } | |||
5673 | ||||
5674 | /// \brief Attempt to vectorize the tree found by | |||
5675 | /// matchAssociativeReduction. | |||
5676 | bool tryToReduce(BoUpSLP &V, TargetTransformInfo *TTI) { | |||
5677 | if (ReducedVals.empty()) | |||
5678 | return false; | |||
5679 | ||||
5680 | // If there is a sufficient number of reduction values, reduce | |||
5681 | // to a nearby power-of-2. Can safely generate oversized | |||
5682 | // vectors and rely on the backend to split them to legal sizes. | |||
5683 | unsigned NumReducedVals = ReducedVals.size(); | |||
5684 | if (NumReducedVals < 4) | |||
5685 | return false; | |||
5686 | ||||
5687 | unsigned ReduxWidth = PowerOf2Floor(NumReducedVals); | |||
5688 | ||||
5689 | Value *VectorizedTree = nullptr; | |||
5690 | IRBuilder<> Builder(ReductionRoot); | |||
5691 | FastMathFlags Unsafe; | |||
5692 | Unsafe.setFast(); | |||
5693 | Builder.setFastMathFlags(Unsafe); | |||
5694 | unsigned i = 0; | |||
5695 | ||||
5696 | BoUpSLP::ExtraValueToDebugLocsMap ExternallyUsedValues; | |||
5697 | // The same extra argument may be used several time, so log each attempt | |||
5698 | // to use it. | |||
5699 | for (auto &Pair : ExtraArgs) | |||
5700 | ExternallyUsedValues[Pair.second].push_back(Pair.first); | |||
5701 | SmallVector<Value *, 16> IgnoreList; | |||
5702 | for (auto &V : ReductionOps) | |||
5703 | IgnoreList.append(V.begin(), V.end()); | |||
5704 | while (i < NumReducedVals - ReduxWidth + 1 && ReduxWidth > 2) { | |||
5705 | auto VL = makeArrayRef(&ReducedVals[i], ReduxWidth); | |||
5706 | V.buildTree(VL, ExternallyUsedValues, IgnoreList); | |||
5707 | Optional<ArrayRef<unsigned>> Order = V.bestOrder(); | |||
5708 | // TODO: Handle orders of size less than number of elements in the vector. | |||
5709 | if (Order && Order->size() == VL.size()) { | |||
5710 | // TODO: reorder tree nodes without tree rebuilding. | |||
5711 | SmallVector<Value *, 4> ReorderedOps(VL.size()); | |||
5712 | llvm::transform(*Order, ReorderedOps.begin(), | |||
5713 | [VL](const unsigned Idx) { return VL[Idx]; }); | |||
5714 | V.buildTree(ReorderedOps, ExternallyUsedValues, IgnoreList); | |||
5715 | } | |||
5716 | if (V.isTreeTinyAndNotFullyVectorizable()) | |||
5717 | break; | |||
5718 | ||||
5719 | V.computeMinimumValueSizes(); | |||
5720 | ||||
5721 | // Estimate cost. | |||
5722 | int Cost = | |||
5723 | V.getTreeCost() + getReductionCost(TTI, ReducedVals[i], ReduxWidth); | |||
5724 | if (Cost >= -SLPCostThreshold) { | |||
5725 | V.getORE()->emit([&]() { | |||
5726 | return OptimizationRemarkMissed( | |||
5727 | SV_NAME"slp-vectorizer", "HorSLPNotBeneficial", cast<Instruction>(VL[0])) | |||
5728 | << "Vectorizing horizontal reduction is possible" | |||
5729 | << "but not beneficial with cost " | |||
5730 | << ore::NV("Cost", Cost) << " and threshold " | |||
5731 | << ore::NV("Threshold", -SLPCostThreshold); | |||
5732 | }); | |||
5733 | break; | |||
5734 | } | |||
5735 | ||||
5736 | DEBUG(dbgs() << "SLP: Vectorizing horizontal reduction at cost:" << Costdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Vectorizing horizontal reduction at cost:" << Cost << ". (HorRdx)\n"; } } while (false) | |||
5737 | << ". (HorRdx)\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Vectorizing horizontal reduction at cost:" << Cost << ". (HorRdx)\n"; } } while (false); | |||
5738 | V.getORE()->emit([&]() { | |||
5739 | return OptimizationRemark( | |||
5740 | SV_NAME"slp-vectorizer", "VectorizedHorizontalReduction", cast<Instruction>(VL[0])) | |||
5741 | << "Vectorized horizontal reduction with cost " | |||
5742 | << ore::NV("Cost", Cost) << " and with tree size " | |||
5743 | << ore::NV("TreeSize", V.getTreeSize()); | |||
5744 | }); | |||
5745 | ||||
5746 | // Vectorize a tree. | |||
5747 | DebugLoc Loc = cast<Instruction>(ReducedVals[i])->getDebugLoc(); | |||
5748 | Value *VectorizedRoot = V.vectorizeTree(ExternallyUsedValues); | |||
5749 | ||||
5750 | // Emit a reduction. | |||
5751 | Value *ReducedSubTree = | |||
5752 | emitReduction(VectorizedRoot, Builder, ReduxWidth, TTI); | |||
5753 | if (VectorizedTree) { | |||
5754 | Builder.SetCurrentDebugLocation(Loc); | |||
5755 | OperationData VectReductionData(ReductionData.getOpcode(), | |||
5756 | VectorizedTree, ReducedSubTree, | |||
5757 | ReductionData.getKind()); | |||
5758 | VectorizedTree = | |||
5759 | VectReductionData.createOp(Builder, "op.rdx", ReductionOps); | |||
5760 | } else | |||
5761 | VectorizedTree = ReducedSubTree; | |||
5762 | i += ReduxWidth; | |||
5763 | ReduxWidth = PowerOf2Floor(NumReducedVals - i); | |||
5764 | } | |||
5765 | ||||
5766 | if (VectorizedTree) { | |||
5767 | // Finish the reduction. | |||
5768 | for (; i < NumReducedVals; ++i) { | |||
5769 | auto *I = cast<Instruction>(ReducedVals[i]); | |||
5770 | Builder.SetCurrentDebugLocation(I->getDebugLoc()); | |||
5771 | OperationData VectReductionData(ReductionData.getOpcode(), | |||
5772 | VectorizedTree, I, | |||
5773 | ReductionData.getKind()); | |||
5774 | VectorizedTree = VectReductionData.createOp(Builder, "", ReductionOps); | |||
5775 | } | |||
5776 | for (auto &Pair : ExternallyUsedValues) { | |||
5777 | assert(!Pair.second.empty() &&(static_cast <bool> (!Pair.second.empty() && "At least one DebugLoc must be inserted" ) ? void (0) : __assert_fail ("!Pair.second.empty() && \"At least one DebugLoc must be inserted\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5778, __extension__ __PRETTY_FUNCTION__)) | |||
5778 | "At least one DebugLoc must be inserted")(static_cast <bool> (!Pair.second.empty() && "At least one DebugLoc must be inserted" ) ? void (0) : __assert_fail ("!Pair.second.empty() && \"At least one DebugLoc must be inserted\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5778, __extension__ __PRETTY_FUNCTION__)); | |||
5779 | // Add each externally used value to the final reduction. | |||
5780 | for (auto *I : Pair.second) { | |||
5781 | Builder.SetCurrentDebugLocation(I->getDebugLoc()); | |||
5782 | OperationData VectReductionData(ReductionData.getOpcode(), | |||
5783 | VectorizedTree, Pair.first, | |||
5784 | ReductionData.getKind()); | |||
5785 | VectorizedTree = VectReductionData.createOp(Builder, "op.extra", I); | |||
5786 | } | |||
5787 | } | |||
5788 | // Update users. | |||
5789 | ReductionRoot->replaceAllUsesWith(VectorizedTree); | |||
5790 | } | |||
5791 | return VectorizedTree != nullptr; | |||
5792 | } | |||
5793 | ||||
5794 | unsigned numReductionValues() const { | |||
5795 | return ReducedVals.size(); | |||
5796 | } | |||
5797 | ||||
5798 | private: | |||
5799 | /// \brief Calculate the cost of a reduction. | |||
5800 | int getReductionCost(TargetTransformInfo *TTI, Value *FirstReducedVal, | |||
5801 | unsigned ReduxWidth) { | |||
5802 | Type *ScalarTy = FirstReducedVal->getType(); | |||
5803 | Type *VecTy = VectorType::get(ScalarTy, ReduxWidth); | |||
5804 | ||||
5805 | int PairwiseRdxCost; | |||
5806 | int SplittingRdxCost; | |||
5807 | switch (ReductionData.getKind()) { | |||
5808 | case RK_Arithmetic: | |||
5809 | PairwiseRdxCost = | |||
5810 | TTI->getArithmeticReductionCost(ReductionData.getOpcode(), VecTy, | |||
5811 | /*IsPairwiseForm=*/true); | |||
5812 | SplittingRdxCost = | |||
5813 | TTI->getArithmeticReductionCost(ReductionData.getOpcode(), VecTy, | |||
5814 | /*IsPairwiseForm=*/false); | |||
5815 | break; | |||
5816 | case RK_Min: | |||
5817 | case RK_Max: | |||
5818 | case RK_UMin: | |||
5819 | case RK_UMax: { | |||
5820 | Type *VecCondTy = CmpInst::makeCmpResultType(VecTy); | |||
5821 | bool IsUnsigned = ReductionData.getKind() == RK_UMin || | |||
5822 | ReductionData.getKind() == RK_UMax; | |||
5823 | PairwiseRdxCost = | |||
5824 | TTI->getMinMaxReductionCost(VecTy, VecCondTy, | |||
5825 | /*IsPairwiseForm=*/true, IsUnsigned); | |||
5826 | SplittingRdxCost = | |||
5827 | TTI->getMinMaxReductionCost(VecTy, VecCondTy, | |||
5828 | /*IsPairwiseForm=*/false, IsUnsigned); | |||
5829 | break; | |||
5830 | } | |||
5831 | case RK_None: | |||
5832 | llvm_unreachable("Expected arithmetic or min/max reduction operation")::llvm::llvm_unreachable_internal("Expected arithmetic or min/max reduction operation" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5832); | |||
5833 | } | |||
5834 | ||||
5835 | IsPairwiseReduction = PairwiseRdxCost < SplittingRdxCost; | |||
5836 | int VecReduxCost = IsPairwiseReduction ? PairwiseRdxCost : SplittingRdxCost; | |||
5837 | ||||
5838 | int ScalarReduxCost; | |||
5839 | switch (ReductionData.getKind()) { | |||
5840 | case RK_Arithmetic: | |||
5841 | ScalarReduxCost = | |||
5842 | TTI->getArithmeticInstrCost(ReductionData.getOpcode(), ScalarTy); | |||
5843 | break; | |||
5844 | case RK_Min: | |||
5845 | case RK_Max: | |||
5846 | case RK_UMin: | |||
5847 | case RK_UMax: | |||
5848 | ScalarReduxCost = | |||
5849 | TTI->getCmpSelInstrCost(ReductionData.getOpcode(), ScalarTy) + | |||
5850 | TTI->getCmpSelInstrCost(Instruction::Select, ScalarTy, | |||
5851 | CmpInst::makeCmpResultType(ScalarTy)); | |||
5852 | break; | |||
5853 | case RK_None: | |||
5854 | llvm_unreachable("Expected arithmetic or min/max reduction operation")::llvm::llvm_unreachable_internal("Expected arithmetic or min/max reduction operation" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5854); | |||
5855 | } | |||
5856 | ScalarReduxCost *= (ReduxWidth - 1); | |||
5857 | ||||
5858 | 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) | |||
5859 | << " 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) | |||
5860 | << " (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) | |||
5861 | << (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) | |||
5862 | << " 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); | |||
5863 | ||||
5864 | return VecReduxCost - ScalarReduxCost; | |||
5865 | } | |||
5866 | ||||
5867 | /// \brief Emit a horizontal reduction of the vectorized value. | |||
5868 | Value *emitReduction(Value *VectorizedValue, IRBuilder<> &Builder, | |||
5869 | unsigned ReduxWidth, const TargetTransformInfo *TTI) { | |||
5870 | assert(VectorizedValue && "Need to have a vectorized tree node")(static_cast <bool> (VectorizedValue && "Need to have a vectorized tree node" ) ? void (0) : __assert_fail ("VectorizedValue && \"Need to have a vectorized tree node\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5870, __extension__ __PRETTY_FUNCTION__)); | |||
5871 | assert(isPowerOf2_32(ReduxWidth) &&(static_cast <bool> (isPowerOf2_32(ReduxWidth) && "We only handle power-of-two reductions for now") ? void (0) : __assert_fail ("isPowerOf2_32(ReduxWidth) && \"We only handle power-of-two reductions for now\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5872, __extension__ __PRETTY_FUNCTION__)) | |||
5872 | "We only handle power-of-two reductions for now")(static_cast <bool> (isPowerOf2_32(ReduxWidth) && "We only handle power-of-two reductions for now") ? void (0) : __assert_fail ("isPowerOf2_32(ReduxWidth) && \"We only handle power-of-two reductions for now\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5872, __extension__ __PRETTY_FUNCTION__)); | |||
5873 | ||||
5874 | if (!IsPairwiseReduction) | |||
5875 | return createSimpleTargetReduction( | |||
5876 | Builder, TTI, ReductionData.getOpcode(), VectorizedValue, | |||
5877 | ReductionData.getFlags(), ReductionOps.back()); | |||
5878 | ||||
5879 | Value *TmpVec = VectorizedValue; | |||
5880 | for (unsigned i = ReduxWidth / 2; i != 0; i >>= 1) { | |||
5881 | Value *LeftMask = | |||
5882 | createRdxShuffleMask(ReduxWidth, i, true, true, Builder); | |||
5883 | Value *RightMask = | |||
5884 | createRdxShuffleMask(ReduxWidth, i, true, false, Builder); | |||
5885 | ||||
5886 | Value *LeftShuf = Builder.CreateShuffleVector( | |||
5887 | TmpVec, UndefValue::get(TmpVec->getType()), LeftMask, "rdx.shuf.l"); | |||
5888 | Value *RightShuf = Builder.CreateShuffleVector( | |||
5889 | TmpVec, UndefValue::get(TmpVec->getType()), (RightMask), | |||
5890 | "rdx.shuf.r"); | |||
5891 | OperationData VectReductionData(ReductionData.getOpcode(), LeftShuf, | |||
5892 | RightShuf, ReductionData.getKind()); | |||
5893 | TmpVec = VectReductionData.createOp(Builder, "op.rdx", ReductionOps); | |||
5894 | } | |||
5895 | ||||
5896 | // The result is in the first element of the vector. | |||
5897 | return Builder.CreateExtractElement(TmpVec, Builder.getInt32(0)); | |||
5898 | } | |||
5899 | }; | |||
5900 | ||||
5901 | } // end anonymous namespace | |||
5902 | ||||
5903 | /// \brief Recognize construction of vectors like | |||
5904 | /// %ra = insertelement <4 x float> undef, float %s0, i32 0 | |||
5905 | /// %rb = insertelement <4 x float> %ra, float %s1, i32 1 | |||
5906 | /// %rc = insertelement <4 x float> %rb, float %s2, i32 2 | |||
5907 | /// %rd = insertelement <4 x float> %rc, float %s3, i32 3 | |||
5908 | /// starting from the last insertelement instruction. | |||
5909 | /// | |||
5910 | /// Returns true if it matches | |||
5911 | static bool findBuildVector(InsertElementInst *LastInsertElem, | |||
5912 | TargetTransformInfo *TTI, | |||
5913 | SmallVectorImpl<Value *> &BuildVectorOpds, | |||
5914 | int &UserCost) { | |||
5915 | UserCost = 0; | |||
5916 | Value *V = nullptr; | |||
5917 | do { | |||
5918 | if (auto *CI = dyn_cast<ConstantInt>(LastInsertElem->getOperand(2))) { | |||
5919 | UserCost += TTI->getVectorInstrCost(Instruction::InsertElement, | |||
5920 | LastInsertElem->getType(), | |||
5921 | CI->getZExtValue()); | |||
5922 | } | |||
5923 | BuildVectorOpds.push_back(LastInsertElem->getOperand(1)); | |||
5924 | V = LastInsertElem->getOperand(0); | |||
5925 | if (isa<UndefValue>(V)) | |||
5926 | break; | |||
5927 | LastInsertElem = dyn_cast<InsertElementInst>(V); | |||
5928 | if (!LastInsertElem || !LastInsertElem->hasOneUse()) | |||
5929 | return false; | |||
5930 | } while (true); | |||
5931 | std::reverse(BuildVectorOpds.begin(), BuildVectorOpds.end()); | |||
5932 | return true; | |||
5933 | } | |||
5934 | ||||
5935 | /// \brief Like findBuildVector, but looks for construction of aggregate. | |||
5936 | /// | |||
5937 | /// \return true if it matches. | |||
5938 | static bool findBuildAggregate(InsertValueInst *IV, | |||
5939 | SmallVectorImpl<Value *> &BuildVectorOpds) { | |||
5940 | Value *V; | |||
5941 | do { | |||
5942 | BuildVectorOpds.push_back(IV->getInsertedValueOperand()); | |||
5943 | V = IV->getAggregateOperand(); | |||
5944 | if (isa<UndefValue>(V)) | |||
5945 | break; | |||
5946 | IV = dyn_cast<InsertValueInst>(V); | |||
5947 | if (!IV || !IV->hasOneUse()) | |||
5948 | return false; | |||
5949 | } while (true); | |||
5950 | std::reverse(BuildVectorOpds.begin(), BuildVectorOpds.end()); | |||
5951 | return true; | |||
5952 | } | |||
5953 | ||||
5954 | static bool PhiTypeSorterFunc(Value *V, Value *V2) { | |||
5955 | return V->getType() < V2->getType(); | |||
5956 | } | |||
5957 | ||||
5958 | /// \brief Try and get a reduction value from a phi node. | |||
5959 | /// | |||
5960 | /// Given a phi node \p P in a block \p ParentBB, consider possible reductions | |||
5961 | /// if they come from either \p ParentBB or a containing loop latch. | |||
5962 | /// | |||
5963 | /// \returns A candidate reduction value if possible, or \code nullptr \endcode | |||
5964 | /// if not possible. | |||
5965 | static Value *getReductionValue(const DominatorTree *DT, PHINode *P, | |||
5966 | BasicBlock *ParentBB, LoopInfo *LI) { | |||
5967 | // There are situations where the reduction value is not dominated by the | |||
5968 | // reduction phi. Vectorizing such cases has been reported to cause | |||
5969 | // miscompiles. See PR25787. | |||
5970 | auto DominatedReduxValue = [&](Value *R) { | |||
5971 | return ( | |||
5972 | dyn_cast<Instruction>(R) && | |||
5973 | DT->dominates(P->getParent(), dyn_cast<Instruction>(R)->getParent())); | |||
5974 | }; | |||
5975 | ||||
5976 | Value *Rdx = nullptr; | |||
5977 | ||||
5978 | // Return the incoming value if it comes from the same BB as the phi node. | |||
5979 | if (P->getIncomingBlock(0) == ParentBB) { | |||
5980 | Rdx = P->getIncomingValue(0); | |||
5981 | } else if (P->getIncomingBlock(1) == ParentBB) { | |||
5982 | Rdx = P->getIncomingValue(1); | |||
5983 | } | |||
5984 | ||||
5985 | if (Rdx && DominatedReduxValue(Rdx)) | |||
5986 | return Rdx; | |||
5987 | ||||
5988 | // Otherwise, check whether we have a loop latch to look at. | |||
5989 | Loop *BBL = LI->getLoopFor(ParentBB); | |||
5990 | if (!BBL) | |||
5991 | return nullptr; | |||
5992 | BasicBlock *BBLatch = BBL->getLoopLatch(); | |||
5993 | if (!BBLatch) | |||
5994 | return nullptr; | |||
5995 | ||||
5996 | // There is a loop latch, return the incoming value if it comes from | |||
5997 | // that. This reduction pattern occasionally turns up. | |||
5998 | if (P->getIncomingBlock(0) == BBLatch) { | |||
5999 | Rdx = P->getIncomingValue(0); | |||
6000 | } else if (P->getIncomingBlock(1) == BBLatch) { | |||
6001 | Rdx = P->getIncomingValue(1); | |||
6002 | } | |||
6003 | ||||
6004 | if (Rdx && DominatedReduxValue(Rdx)) | |||
6005 | return Rdx; | |||
6006 | ||||
6007 | return nullptr; | |||
6008 | } | |||
6009 | ||||
6010 | /// Attempt to reduce a horizontal reduction. | |||
6011 | /// If it is legal to match a horizontal reduction feeding the phi node \a P | |||
6012 | /// with reduction operators \a Root (or one of its operands) in a basic block | |||
6013 | /// \a BB, then check if it can be done. If horizontal reduction is not found | |||
6014 | /// and root instruction is a binary operation, vectorization of the operands is | |||
6015 | /// attempted. | |||
6016 | /// \returns true if a horizontal reduction was matched and reduced or operands | |||
6017 | /// of one of the binary instruction were vectorized. | |||
6018 | /// \returns false if a horizontal reduction was not matched (or not possible) | |||
6019 | /// or no vectorization of any binary operation feeding \a Root instruction was | |||
6020 | /// performed. | |||
6021 | static bool tryToVectorizeHorReductionOrInstOperands( | |||
6022 | PHINode *P, Instruction *Root, BasicBlock *BB, BoUpSLP &R, | |||
6023 | TargetTransformInfo *TTI, | |||
6024 | const function_ref<bool(Instruction *, BoUpSLP &)> Vectorize) { | |||
6025 | if (!ShouldVectorizeHor) | |||
6026 | return false; | |||
6027 | ||||
6028 | if (!Root) | |||
6029 | return false; | |||
6030 | ||||
6031 | if (Root->getParent() != BB || isa<PHINode>(Root)) | |||
6032 | return false; | |||
6033 | // Start analysis starting from Root instruction. If horizontal reduction is | |||
6034 | // found, try to vectorize it. If it is not a horizontal reduction or | |||
6035 | // vectorization is not possible or not effective, and currently analyzed | |||
6036 | // instruction is a binary operation, try to vectorize the operands, using | |||
6037 | // pre-order DFS traversal order. If the operands were not vectorized, repeat | |||
6038 | // the same procedure considering each operand as a possible root of the | |||
6039 | // horizontal reduction. | |||
6040 | // Interrupt the process if the Root instruction itself was vectorized or all | |||
6041 | // sub-trees not higher that RecursionMaxDepth were analyzed/vectorized. | |||
6042 | SmallVector<std::pair<WeakTrackingVH, unsigned>, 8> Stack(1, {Root, 0}); | |||
6043 | SmallSet<Value *, 8> VisitedInstrs; | |||
6044 | bool Res = false; | |||
6045 | while (!Stack.empty()) { | |||
6046 | Value *V; | |||
6047 | unsigned Level; | |||
6048 | std::tie(V, Level) = Stack.pop_back_val(); | |||
6049 | if (!V) | |||
6050 | continue; | |||
6051 | auto *Inst = dyn_cast<Instruction>(V); | |||
6052 | if (!Inst) | |||
6053 | continue; | |||
6054 | auto *BI = dyn_cast<BinaryOperator>(Inst); | |||
6055 | auto *SI = dyn_cast<SelectInst>(Inst); | |||
6056 | if (BI || SI) { | |||
6057 | HorizontalReduction HorRdx; | |||
6058 | if (HorRdx.matchAssociativeReduction(P, Inst)) { | |||
6059 | if (HorRdx.tryToReduce(R, TTI)) { | |||
6060 | Res = true; | |||
6061 | // Set P to nullptr to avoid re-analysis of phi node in | |||
6062 | // matchAssociativeReduction function unless this is the root node. | |||
6063 | P = nullptr; | |||
6064 | continue; | |||
6065 | } | |||
6066 | } | |||
6067 | if (P && BI) { | |||
6068 | Inst = dyn_cast<Instruction>(BI->getOperand(0)); | |||
6069 | if (Inst == P) | |||
6070 | Inst = dyn_cast<Instruction>(BI->getOperand(1)); | |||
6071 | if (!Inst) { | |||
6072 | // Set P to nullptr to avoid re-analysis of phi node in | |||
6073 | // matchAssociativeReduction function unless this is the root node. | |||
6074 | P = nullptr; | |||
6075 | continue; | |||
6076 | } | |||
6077 | } | |||
6078 | } | |||
6079 | // Set P to nullptr to avoid re-analysis of phi node in | |||
6080 | // matchAssociativeReduction function unless this is the root node. | |||
6081 | P = nullptr; | |||
6082 | if (Vectorize(Inst, R)) { | |||
6083 | Res = true; | |||
6084 | continue; | |||
6085 | } | |||
6086 | ||||
6087 | // Try to vectorize operands. | |||
6088 | // Continue analysis for the instruction from the same basic block only to | |||
6089 | // save compile time. | |||
6090 | if (++Level < RecursionMaxDepth) | |||
6091 | for (auto *Op : Inst->operand_values()) | |||
6092 | if (VisitedInstrs.insert(Op).second) | |||
6093 | if (auto *I = dyn_cast<Instruction>(Op)) | |||
6094 | if (!isa<PHINode>(I) && I->getParent() == BB) | |||
6095 | Stack.emplace_back(Op, Level); | |||
6096 | } | |||
6097 | return Res; | |||
6098 | } | |||
6099 | ||||
6100 | bool SLPVectorizerPass::vectorizeRootInstruction(PHINode *P, Value *V, | |||
6101 | BasicBlock *BB, BoUpSLP &R, | |||
6102 | TargetTransformInfo *TTI) { | |||
6103 | if (!V) | |||
6104 | return false; | |||
6105 | auto *I = dyn_cast<Instruction>(V); | |||
6106 | if (!I) | |||
6107 | return false; | |||
6108 | ||||
6109 | if (!isa<BinaryOperator>(I)) | |||
6110 | P = nullptr; | |||
6111 | // Try to match and vectorize a horizontal reduction. | |||
6112 | auto &&ExtraVectorization = [this](Instruction *I, BoUpSLP &R) -> bool { | |||
6113 | return tryToVectorize(I, R); | |||
6114 | }; | |||
6115 | return tryToVectorizeHorReductionOrInstOperands(P, I, BB, R, TTI, | |||
6116 | ExtraVectorization); | |||
6117 | } | |||
6118 | ||||
6119 | bool SLPVectorizerPass::vectorizeInsertValueInst(InsertValueInst *IVI, | |||
6120 | BasicBlock *BB, BoUpSLP &R) { | |||
6121 | const DataLayout &DL = BB->getModule()->getDataLayout(); | |||
6122 | if (!R.canMapToVector(IVI->getType(), DL)) | |||
6123 | return false; | |||
6124 | ||||
6125 | SmallVector<Value *, 16> BuildVectorOpds; | |||
6126 | if (!findBuildAggregate(IVI, BuildVectorOpds)) | |||
6127 | return false; | |||
6128 | ||||
6129 | 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); | |||
6130 | // Aggregate value is unlikely to be processed in vector register, we need to | |||
6131 | // extract scalars into scalar registers, so NeedExtraction is set true. | |||
6132 | return tryToVectorizeList(BuildVectorOpds, R); | |||
6133 | } | |||
6134 | ||||
6135 | bool SLPVectorizerPass::vectorizeInsertElementInst(InsertElementInst *IEI, | |||
6136 | BasicBlock *BB, BoUpSLP &R) { | |||
6137 | int UserCost; | |||
6138 | SmallVector<Value *, 16> BuildVectorOpds; | |||
6139 | if (!findBuildVector(IEI, TTI, BuildVectorOpds, UserCost) || | |||
6140 | (llvm::all_of(BuildVectorOpds, | |||
6141 | [](Value *V) { return isa<ExtractElementInst>(V); }) && | |||
6142 | isShuffle(BuildVectorOpds))) | |||
6143 | return false; | |||
6144 | ||||
6145 | // Vectorize starting with the build vector operands ignoring the BuildVector | |||
6146 | // instructions for the purpose of scheduling and user extraction. | |||
6147 | return tryToVectorizeList(BuildVectorOpds, R, UserCost); | |||
6148 | } | |||
6149 | ||||
6150 | bool SLPVectorizerPass::vectorizeCmpInst(CmpInst *CI, BasicBlock *BB, | |||
6151 | BoUpSLP &R) { | |||
6152 | if (tryToVectorizePair(CI->getOperand(0), CI->getOperand(1), R)) | |||
6153 | return true; | |||
6154 | ||||
6155 | bool OpsChanged = false; | |||
6156 | for (int Idx = 0; Idx < 2; ++Idx) { | |||
6157 | OpsChanged |= | |||
6158 | vectorizeRootInstruction(nullptr, CI->getOperand(Idx), BB, R, TTI); | |||
6159 | } | |||
6160 | return OpsChanged; | |||
6161 | } | |||
6162 | ||||
6163 | bool SLPVectorizerPass::vectorizeSimpleInstructions( | |||
6164 | SmallVectorImpl<WeakVH> &Instructions, BasicBlock *BB, BoUpSLP &R) { | |||
6165 | bool OpsChanged = false; | |||
6166 | for (auto &VH : reverse(Instructions)) { | |||
6167 | auto *I = dyn_cast_or_null<Instruction>(VH); | |||
6168 | if (!I) | |||
6169 | continue; | |||
6170 | if (auto *LastInsertValue = dyn_cast<InsertValueInst>(I)) | |||
6171 | OpsChanged |= vectorizeInsertValueInst(LastInsertValue, BB, R); | |||
6172 | else if (auto *LastInsertElem = dyn_cast<InsertElementInst>(I)) | |||
6173 | OpsChanged |= vectorizeInsertElementInst(LastInsertElem, BB, R); | |||
6174 | else if (auto *CI = dyn_cast<CmpInst>(I)) | |||
6175 | OpsChanged |= vectorizeCmpInst(CI, BB, R); | |||
6176 | } | |||
6177 | Instructions.clear(); | |||
6178 | return OpsChanged; | |||
6179 | } | |||
6180 | ||||
6181 | bool SLPVectorizerPass::vectorizeChainsInBlock(BasicBlock *BB, BoUpSLP &R) { | |||
6182 | bool Changed = false; | |||
6183 | SmallVector<Value *, 4> Incoming; | |||
6184 | SmallSet<Value *, 16> VisitedInstrs; | |||
6185 | ||||
6186 | bool HaveVectorizedPhiNodes = true; | |||
6187 | while (HaveVectorizedPhiNodes) { | |||
6188 | HaveVectorizedPhiNodes = false; | |||
6189 | ||||
6190 | // Collect the incoming values from the PHIs. | |||
6191 | Incoming.clear(); | |||
6192 | for (Instruction &I : *BB) { | |||
6193 | PHINode *P = dyn_cast<PHINode>(&I); | |||
6194 | if (!P) | |||
6195 | break; | |||
6196 | ||||
6197 | if (!VisitedInstrs.count(P)) | |||
6198 | Incoming.push_back(P); | |||
6199 | } | |||
6200 | ||||
6201 | // Sort by type. | |||
6202 | std::stable_sort(Incoming.begin(), Incoming.end(), PhiTypeSorterFunc); | |||
6203 | ||||
6204 | // Try to vectorize elements base on their type. | |||
6205 | for (SmallVector<Value *, 4>::iterator IncIt = Incoming.begin(), | |||
6206 | E = Incoming.end(); | |||
6207 | IncIt != E;) { | |||
6208 | ||||
6209 | // Look for the next elements with the same type. | |||
6210 | SmallVector<Value *, 4>::iterator SameTypeIt = IncIt; | |||
6211 | while (SameTypeIt != E && | |||
6212 | (*SameTypeIt)->getType() == (*IncIt)->getType()) { | |||
6213 | VisitedInstrs.insert(*SameTypeIt); | |||
6214 | ++SameTypeIt; | |||
6215 | } | |||
6216 | ||||
6217 | // Try to vectorize them. | |||
6218 | unsigned NumElts = (SameTypeIt - IncIt); | |||
6219 | DEBUG(dbgs() << "SLP: Trying to vectorize starting at PHIs (" << NumEltsdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Trying to vectorize starting at PHIs (" << NumElts << ")\n"; } } while (false) | |||
6220 | << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Trying to vectorize starting at PHIs (" << NumElts << ")\n"; } } while (false); | |||
6221 | // The order in which the phi nodes appear in the program does not matter. | |||
6222 | // So allow tryToVectorizeList to reorder them if it is beneficial. This | |||
6223 | // is done when there are exactly two elements since tryToVectorizeList | |||
6224 | // asserts that there are only two values when AllowReorder is true. | |||
6225 | bool AllowReorder = NumElts == 2; | |||
6226 | if (NumElts > 1 && tryToVectorizeList(makeArrayRef(IncIt, NumElts), R, | |||
6227 | /*UserCost=*/0, AllowReorder)) { | |||
6228 | // Success start over because instructions might have been changed. | |||
6229 | HaveVectorizedPhiNodes = true; | |||
6230 | Changed = true; | |||
6231 | break; | |||
6232 | } | |||
6233 | ||||
6234 | // Start over at the next instruction of a different type (or the end). | |||
6235 | IncIt = SameTypeIt; | |||
6236 | } | |||
6237 | } | |||
6238 | ||||
6239 | VisitedInstrs.clear(); | |||
6240 | ||||
6241 | SmallVector<WeakVH, 8> PostProcessInstructions; | |||
6242 | SmallDenseSet<Instruction *, 4> KeyNodes; | |||
6243 | for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; it++) { | |||
6244 | // We may go through BB multiple times so skip the one we have checked. | |||
6245 | if (!VisitedInstrs.insert(&*it).second) { | |||
6246 | if (it->use_empty() && KeyNodes.count(&*it) > 0 && | |||
6247 | vectorizeSimpleInstructions(PostProcessInstructions, BB, R)) { | |||
6248 | // We would like to start over since some instructions are deleted | |||
6249 | // and the iterator may become invalid value. | |||
6250 | Changed = true; | |||
6251 | it = BB->begin(); | |||
6252 | e = BB->end(); | |||
6253 | } | |||
6254 | continue; | |||
6255 | } | |||
6256 | ||||
6257 | if (isa<DbgInfoIntrinsic>(it)) | |||
6258 | continue; | |||
6259 | ||||
6260 | // Try to vectorize reductions that use PHINodes. | |||
6261 | if (PHINode *P = dyn_cast<PHINode>(it)) { | |||
6262 | // Check that the PHI is a reduction PHI. | |||
6263 | if (P->getNumIncomingValues() != 2) | |||
6264 | return Changed; | |||
6265 | ||||
6266 | // Try to match and vectorize a horizontal reduction. | |||
6267 | if (vectorizeRootInstruction(P, getReductionValue(DT, P, BB, LI), BB, R, | |||
6268 | TTI)) { | |||
6269 | Changed = true; | |||
6270 | it = BB->begin(); | |||
6271 | e = BB->end(); | |||
6272 | continue; | |||
6273 | } | |||
6274 | continue; | |||
6275 | } | |||
6276 | ||||
6277 | // Ran into an instruction without users, like terminator, or function call | |||
6278 | // with ignored return value, store. Ignore unused instructions (basing on | |||
6279 | // instruction type, except for CallInst and InvokeInst). | |||
6280 | if (it->use_empty() && (it->getType()->isVoidTy() || isa<CallInst>(it) || | |||
6281 | isa<InvokeInst>(it))) { | |||
6282 | KeyNodes.insert(&*it); | |||
6283 | bool OpsChanged = false; | |||
6284 | if (ShouldStartVectorizeHorAtStore || !isa<StoreInst>(it)) { | |||
6285 | for (auto *V : it->operand_values()) { | |||
6286 | // Try to match and vectorize a horizontal reduction. | |||
6287 | OpsChanged |= vectorizeRootInstruction(nullptr, V, BB, R, TTI); | |||
6288 | } | |||
6289 | } | |||
6290 | // Start vectorization of post-process list of instructions from the | |||
6291 | // top-tree instructions to try to vectorize as many instructions as | |||
6292 | // possible. | |||
6293 | OpsChanged |= vectorizeSimpleInstructions(PostProcessInstructions, BB, R); | |||
6294 | if (OpsChanged) { | |||
6295 | // We would like to start over since some instructions are deleted | |||
6296 | // and the iterator may become invalid value. | |||
6297 | Changed = true; | |||
6298 | it = BB->begin(); | |||
6299 | e = BB->end(); | |||
6300 | continue; | |||
6301 | } | |||
6302 | } | |||
6303 | ||||
6304 | if (isa<InsertElementInst>(it) || isa<CmpInst>(it) || | |||
6305 | isa<InsertValueInst>(it)) | |||
6306 | PostProcessInstructions.push_back(&*it); | |||
6307 | ||||
6308 | } | |||
6309 | ||||
6310 | return Changed; | |||
6311 | } | |||
6312 | ||||
6313 | bool SLPVectorizerPass::vectorizeGEPIndices(BasicBlock *BB, BoUpSLP &R) { | |||
6314 | auto Changed = false; | |||
6315 | for (auto &Entry : GEPs) { | |||
6316 | // If the getelementptr list has fewer than two elements, there's nothing | |||
6317 | // to do. | |||
6318 | if (Entry.second.size() < 2) | |||
6319 | continue; | |||
6320 | ||||
6321 | 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 ) | |||
6322 | << 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 ); | |||
6323 | ||||
6324 | // We process the getelementptr list in chunks of 16 (like we do for | |||
6325 | // stores) to minimize compile-time. | |||
6326 | for (unsigned BI = 0, BE = Entry.second.size(); BI < BE; BI += 16) { | |||
6327 | auto Len = std::min<unsigned>(BE - BI, 16); | |||
6328 | auto GEPList = makeArrayRef(&Entry.second[BI], Len); | |||
6329 | ||||
6330 | // Initialize a set a candidate getelementptrs. Note that we use a | |||
6331 | // SetVector here to preserve program order. If the index computations | |||
6332 | // are vectorizable and begin with loads, we want to minimize the chance | |||
6333 | // of having to reorder them later. | |||
6334 | SetVector<Value *> Candidates(GEPList.begin(), GEPList.end()); | |||
6335 | ||||
6336 | // Some of the candidates may have already been vectorized after we | |||
6337 | // initially collected them. If so, the WeakTrackingVHs will have | |||
6338 | // nullified the | |||
6339 | // values, so remove them from the set of candidates. | |||
6340 | Candidates.remove(nullptr); | |||
6341 | ||||
6342 | // Remove from the set of candidates all pairs of getelementptrs with | |||
6343 | // constant differences. Such getelementptrs are likely not good | |||
6344 | // candidates for vectorization in a bottom-up phase since one can be | |||
6345 | // computed from the other. We also ensure all candidate getelementptr | |||
6346 | // indices are unique. | |||
6347 | for (int I = 0, E = GEPList.size(); I < E && Candidates.size() > 1; ++I) { | |||
6348 | auto *GEPI = cast<GetElementPtrInst>(GEPList[I]); | |||
6349 | if (!Candidates.count(GEPI)) | |||
6350 | continue; | |||
6351 | auto *SCEVI = SE->getSCEV(GEPList[I]); | |||
6352 | for (int J = I + 1; J < E && Candidates.size() > 1; ++J) { | |||
6353 | auto *GEPJ = cast<GetElementPtrInst>(GEPList[J]); | |||
6354 | auto *SCEVJ = SE->getSCEV(GEPList[J]); | |||
6355 | if (isa<SCEVConstant>(SE->getMinusSCEV(SCEVI, SCEVJ))) { | |||
6356 | Candidates.remove(GEPList[I]); | |||
6357 | Candidates.remove(GEPList[J]); | |||
6358 | } else if (GEPI->idx_begin()->get() == GEPJ->idx_begin()->get()) { | |||
6359 | Candidates.remove(GEPList[J]); | |||
6360 | } | |||
6361 | } | |||
6362 | } | |||
6363 | ||||
6364 | // We break out of the above computation as soon as we know there are | |||
6365 | // fewer than two candidates remaining. | |||
6366 | if (Candidates.size() < 2) | |||
6367 | continue; | |||
6368 | ||||
6369 | // Add the single, non-constant index of each candidate to the bundle. We | |||
6370 | // ensured the indices met these constraints when we originally collected | |||
6371 | // the getelementptrs. | |||
6372 | SmallVector<Value *, 16> Bundle(Candidates.size()); | |||
6373 | auto BundleIndex = 0u; | |||
6374 | for (auto *V : Candidates) { | |||
6375 | auto *GEP = cast<GetElementPtrInst>(V); | |||
6376 | auto *GEPIdx = GEP->idx_begin()->get(); | |||
6377 | assert(GEP->getNumIndices() == 1 || !isa<Constant>(GEPIdx))(static_cast <bool> (GEP->getNumIndices() == 1 || !isa <Constant>(GEPIdx)) ? void (0) : __assert_fail ("GEP->getNumIndices() == 1 || !isa<Constant>(GEPIdx)" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 6377, __extension__ __PRETTY_FUNCTION__)); | |||
6378 | Bundle[BundleIndex++] = GEPIdx; | |||
6379 | } | |||
6380 | ||||
6381 | // Try and vectorize the indices. We are currently only interested in | |||
6382 | // gather-like cases of the form: | |||
6383 | // | |||
6384 | // ... = g[a[0] - b[0]] + g[a[1] - b[1]] + ... | |||
6385 | // | |||
6386 | // where the loads of "a", the loads of "b", and the subtractions can be | |||
6387 | // performed in parallel. It's likely that detecting this pattern in a | |||
6388 | // bottom-up phase will be simpler and less costly than building a | |||
6389 | // full-blown top-down phase beginning at the consecutive loads. | |||
6390 | Changed |= tryToVectorizeList(Bundle, R); | |||
6391 | } | |||
6392 | } | |||
6393 | return Changed; | |||
6394 | } | |||
6395 | ||||
6396 | bool SLPVectorizerPass::vectorizeStoreChains(BoUpSLP &R) { | |||
6397 | bool Changed = false; | |||
6398 | // Attempt to sort and vectorize each of the store-groups. | |||
6399 | for (StoreListMap::iterator it = Stores.begin(), e = Stores.end(); it != e; | |||
6400 | ++it) { | |||
6401 | if (it->second.size() < 2) | |||
6402 | continue; | |||
6403 | ||||
6404 | 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 ) | |||
6405 | << 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 ); | |||
6406 | ||||
6407 | // Process the stores in chunks of 16. | |||
6408 | // TODO: The limit of 16 inhibits greater vectorization factors. | |||
6409 | // For example, AVX2 supports v32i8. Increasing this limit, however, | |||
6410 | // may cause a significant compile-time increase. | |||
6411 | for (unsigned CI = 0, CE = it->second.size(); CI < CE; CI+=16) { | |||
6412 | unsigned Len = std::min<unsigned>(CE - CI, 16); | |||
6413 | Changed |= vectorizeStores(makeArrayRef(&it->second[CI], Len), R); | |||
6414 | } | |||
6415 | } | |||
6416 | return Changed; | |||
6417 | } | |||
6418 | ||||
6419 | char SLPVectorizer::ID = 0; | |||
6420 | ||||
6421 | static const char lv_name[] = "SLP Vectorizer"; | |||
6422 | ||||
6423 | INITIALIZE_PASS_BEGIN(SLPVectorizer, SV_NAME, lv_name, false, false)static void *initializeSLPVectorizerPassOnce(PassRegistry & Registry) { | |||
6424 | INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)initializeAAResultsWrapperPassPass(Registry); | |||
6425 | INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)initializeTargetTransformInfoWrapperPassPass(Registry); | |||
6426 | INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)initializeAssumptionCacheTrackerPass(Registry); | |||
6427 | INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)initializeScalarEvolutionWrapperPassPass(Registry); | |||
6428 | INITIALIZE_PASS_DEPENDENCY(LoopSimplify)initializeLoopSimplifyPass(Registry); | |||
6429 | INITIALIZE_PASS_DEPENDENCY(DemandedBitsWrapperPass)initializeDemandedBitsWrapperPassPass(Registry); | |||
6430 | INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass)initializeOptimizationRemarkEmitterWrapperPassPass(Registry); | |||
6431 | 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)); } | |||
6432 | ||||
6433 | Pass *llvm::createSLPVectorizerPass() { return new SLPVectorizer(); } |