File: | lib/Transforms/Vectorize/SLPVectorizer.cpp |
Warning: | line 2382, column 32 Called C++ object pointer is null |
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 | // This pass implements the Bottom Up SLP vectorizer. It detects consecutive | |||
10 | // stores that can be put together into vector-stores. Next, it attempts to | |||
11 | // construct vectorizable tree using the use-def chains. If a profitable tree | |||
12 | // was found, the SLP vectorizer performs vectorization on the tree. | |||
13 | // | |||
14 | // The pass is inspired by the work described in the paper: | |||
15 | // "Loop-Aware SLP in GCC" by Ira Rosen, Dorit Nuzman, Ayal Zaks. | |||
16 | // | |||
17 | //===----------------------------------------------------------------------===// | |||
18 | #include "llvm/Transforms/Vectorize/SLPVectorizer.h" | |||
19 | #include "llvm/ADT/Optional.h" | |||
20 | #include "llvm/ADT/PostOrderIterator.h" | |||
21 | #include "llvm/ADT/SetVector.h" | |||
22 | #include "llvm/ADT/Statistic.h" | |||
23 | #include "llvm/Analysis/CodeMetrics.h" | |||
24 | #include "llvm/Analysis/GlobalsModRef.h" | |||
25 | #include "llvm/Analysis/LoopAccessAnalysis.h" | |||
26 | #include "llvm/Analysis/ScalarEvolutionExpressions.h" | |||
27 | #include "llvm/Analysis/ValueTracking.h" | |||
28 | #include "llvm/Analysis/VectorUtils.h" | |||
29 | #include "llvm/IR/DataLayout.h" | |||
30 | #include "llvm/IR/Dominators.h" | |||
31 | #include "llvm/IR/IRBuilder.h" | |||
32 | #include "llvm/IR/Instructions.h" | |||
33 | #include "llvm/IR/IntrinsicInst.h" | |||
34 | #include "llvm/IR/Module.h" | |||
35 | #include "llvm/IR/NoFolder.h" | |||
36 | #include "llvm/IR/Type.h" | |||
37 | #include "llvm/IR/Value.h" | |||
38 | #include "llvm/IR/Verifier.h" | |||
39 | #include "llvm/Pass.h" | |||
40 | #include "llvm/Support/CommandLine.h" | |||
41 | #include "llvm/Support/Debug.h" | |||
42 | #include "llvm/Support/GraphWriter.h" | |||
43 | #include "llvm/Support/KnownBits.h" | |||
44 | #include "llvm/Support/raw_ostream.h" | |||
45 | #include "llvm/Transforms/Utils/LoopUtils.h" | |||
46 | #include "llvm/Transforms/Vectorize.h" | |||
47 | #include <algorithm> | |||
48 | #include <memory> | |||
49 | ||||
50 | using namespace llvm; | |||
51 | using namespace slpvectorizer; | |||
52 | ||||
53 | #define SV_NAME"slp-vectorizer" "slp-vectorizer" | |||
54 | #define DEBUG_TYPE"SLP" "SLP" | |||
55 | ||||
56 | STATISTIC(NumVectorInstructions, "Number of vector instructions generated")static llvm::Statistic NumVectorInstructions = {"SLP", "NumVectorInstructions" , "Number of vector instructions generated", {0}, false}; | |||
57 | ||||
58 | static cl::opt<int> | |||
59 | SLPCostThreshold("slp-threshold", cl::init(0), cl::Hidden, | |||
60 | cl::desc("Only vectorize if you gain more than this " | |||
61 | "number ")); | |||
62 | ||||
63 | static cl::opt<bool> | |||
64 | ShouldVectorizeHor("slp-vectorize-hor", cl::init(true), cl::Hidden, | |||
65 | cl::desc("Attempt to vectorize horizontal reductions")); | |||
66 | ||||
67 | static cl::opt<bool> ShouldStartVectorizeHorAtStore( | |||
68 | "slp-vectorize-hor-store", cl::init(false), cl::Hidden, | |||
69 | cl::desc( | |||
70 | "Attempt to vectorize horizontal reductions feeding into a store")); | |||
71 | ||||
72 | static cl::opt<int> | |||
73 | MaxVectorRegSizeOption("slp-max-reg-size", cl::init(128), cl::Hidden, | |||
74 | cl::desc("Attempt to vectorize for this register size in bits")); | |||
75 | ||||
76 | /// Limits the size of scheduling regions in a block. | |||
77 | /// It avoid long compile times for _very_ large blocks where vector | |||
78 | /// instructions are spread over a wide range. | |||
79 | /// This limit is way higher than needed by real-world functions. | |||
80 | static cl::opt<int> | |||
81 | ScheduleRegionSizeBudget("slp-schedule-budget", cl::init(100000), cl::Hidden, | |||
82 | cl::desc("Limit the size of the SLP scheduling region per block")); | |||
83 | ||||
84 | static cl::opt<int> MinVectorRegSizeOption( | |||
85 | "slp-min-reg-size", cl::init(128), cl::Hidden, | |||
86 | cl::desc("Attempt to vectorize for this register size in bits")); | |||
87 | ||||
88 | static cl::opt<unsigned> RecursionMaxDepth( | |||
89 | "slp-recursion-max-depth", cl::init(12), cl::Hidden, | |||
90 | cl::desc("Limit the recursion depth when building a vectorizable tree")); | |||
91 | ||||
92 | static cl::opt<unsigned> MinTreeSize( | |||
93 | "slp-min-tree-size", cl::init(3), cl::Hidden, | |||
94 | cl::desc("Only vectorize small trees if they are fully vectorizable")); | |||
95 | ||||
96 | static cl::opt<bool> | |||
97 | ViewSLPTree("view-slp-tree", cl::Hidden, | |||
98 | cl::desc("Display the SLP trees with Graphviz")); | |||
99 | ||||
100 | // Limit the number of alias checks. The limit is chosen so that | |||
101 | // it has no negative effect on the llvm benchmarks. | |||
102 | static const unsigned AliasedCheckLimit = 10; | |||
103 | ||||
104 | // Another limit for the alias checks: The maximum distance between load/store | |||
105 | // instructions where alias checks are done. | |||
106 | // This limit is useful for very large basic blocks. | |||
107 | static const unsigned MaxMemDepDistance = 160; | |||
108 | ||||
109 | /// If the ScheduleRegionSizeBudget is exhausted, we allow small scheduling | |||
110 | /// regions to be handled. | |||
111 | static const int MinScheduleRegionSize = 16; | |||
112 | ||||
113 | /// \brief Predicate for the element types that the SLP vectorizer supports. | |||
114 | /// | |||
115 | /// The most important thing to filter here are types which are invalid in LLVM | |||
116 | /// vectors. We also filter target specific types which have absolutely no | |||
117 | /// meaningful vectorization path such as x86_fp80 and ppc_f128. This just | |||
118 | /// avoids spending time checking the cost model and realizing that they will | |||
119 | /// be inevitably scalarized. | |||
120 | static bool isValidElementType(Type *Ty) { | |||
121 | return VectorType::isValidElementType(Ty) && !Ty->isX86_FP80Ty() && | |||
122 | !Ty->isPPC_FP128Ty(); | |||
123 | } | |||
124 | ||||
125 | /// \returns true if all of the instructions in \p VL are in the same block or | |||
126 | /// false otherwise. | |||
127 | static bool allSameBlock(ArrayRef<Value *> VL) { | |||
128 | Instruction *I0 = dyn_cast<Instruction>(VL[0]); | |||
129 | if (!I0) | |||
130 | return false; | |||
131 | BasicBlock *BB = I0->getParent(); | |||
132 | for (int i = 1, e = VL.size(); i < e; i++) { | |||
133 | Instruction *I = dyn_cast<Instruction>(VL[i]); | |||
134 | if (!I) | |||
135 | return false; | |||
136 | ||||
137 | if (BB != I->getParent()) | |||
138 | return false; | |||
139 | } | |||
140 | return true; | |||
141 | } | |||
142 | ||||
143 | /// \returns True if all of the values in \p VL are constants. | |||
144 | static bool allConstant(ArrayRef<Value *> VL) { | |||
145 | for (Value *i : VL) | |||
146 | if (!isa<Constant>(i)) | |||
147 | return false; | |||
148 | return true; | |||
149 | } | |||
150 | ||||
151 | /// \returns True if all of the values in \p VL are identical. | |||
152 | static bool isSplat(ArrayRef<Value *> VL) { | |||
153 | for (unsigned i = 1, e = VL.size(); i < e; ++i) | |||
154 | if (VL[i] != VL[0]) | |||
155 | return false; | |||
156 | return true; | |||
157 | } | |||
158 | ||||
159 | ///\returns Opcode that can be clubbed with \p Op to create an alternate | |||
160 | /// sequence which can later be merged as a ShuffleVector instruction. | |||
161 | static unsigned getAltOpcode(unsigned Op) { | |||
162 | switch (Op) { | |||
163 | case Instruction::FAdd: | |||
164 | return Instruction::FSub; | |||
165 | case Instruction::FSub: | |||
166 | return Instruction::FAdd; | |||
167 | case Instruction::Add: | |||
168 | return Instruction::Sub; | |||
169 | case Instruction::Sub: | |||
170 | return Instruction::Add; | |||
171 | default: | |||
172 | return 0; | |||
173 | } | |||
174 | } | |||
175 | ||||
176 | ///\returns bool representing if Opcode \p Op can be part | |||
177 | /// of an alternate sequence which can later be merged as | |||
178 | /// a ShuffleVector instruction. | |||
179 | static bool canCombineAsAltInst(unsigned Op) { | |||
180 | return Op == Instruction::FAdd || Op == Instruction::FSub || | |||
181 | Op == Instruction::Sub || Op == Instruction::Add; | |||
182 | } | |||
183 | ||||
184 | /// \returns ShuffleVector instruction if instructions in \p VL have | |||
185 | /// alternate fadd,fsub / fsub,fadd/add,sub/sub,add sequence. | |||
186 | /// (i.e. e.g. opcodes of fadd,fsub,fadd,fsub...) | |||
187 | static unsigned isAltInst(ArrayRef<Value *> VL) { | |||
188 | Instruction *I0 = dyn_cast<Instruction>(VL[0]); | |||
189 | unsigned Opcode = I0->getOpcode(); | |||
190 | unsigned AltOpcode = getAltOpcode(Opcode); | |||
191 | for (int i = 1, e = VL.size(); i < e; i++) { | |||
192 | Instruction *I = dyn_cast<Instruction>(VL[i]); | |||
193 | if (!I || I->getOpcode() != ((i & 1) ? AltOpcode : Opcode)) | |||
194 | return 0; | |||
195 | } | |||
196 | return Instruction::ShuffleVector; | |||
197 | } | |||
198 | ||||
199 | /// \returns The opcode if all of the Instructions in \p VL have the same | |||
200 | /// opcode, or zero. | |||
201 | static unsigned getSameOpcode(ArrayRef<Value *> VL) { | |||
202 | Instruction *I0 = dyn_cast<Instruction>(VL[0]); | |||
203 | if (!I0) | |||
204 | return 0; | |||
205 | unsigned Opcode = I0->getOpcode(); | |||
206 | for (int i = 1, e = VL.size(); i < e; i++) { | |||
207 | Instruction *I = dyn_cast<Instruction>(VL[i]); | |||
208 | if (!I || Opcode != I->getOpcode()) { | |||
209 | if (canCombineAsAltInst(Opcode) && i == 1) | |||
210 | return isAltInst(VL); | |||
211 | return 0; | |||
212 | } | |||
213 | } | |||
214 | return Opcode; | |||
215 | } | |||
216 | ||||
217 | /// \returns true if all of the values in \p VL have the same type or false | |||
218 | /// otherwise. | |||
219 | static bool allSameType(ArrayRef<Value *> VL) { | |||
220 | Type *Ty = VL[0]->getType(); | |||
221 | for (int i = 1, e = VL.size(); i < e; i++) | |||
222 | if (VL[i]->getType() != Ty) | |||
223 | return false; | |||
224 | ||||
225 | return true; | |||
226 | } | |||
227 | ||||
228 | /// \returns True if Extract{Value,Element} instruction extracts element Idx. | |||
229 | static bool matchExtractIndex(Instruction *E, unsigned Idx, unsigned Opcode) { | |||
230 | assert(Opcode == Instruction::ExtractElement ||((Opcode == Instruction::ExtractElement || Opcode == Instruction ::ExtractValue) ? static_cast<void> (0) : __assert_fail ("Opcode == Instruction::ExtractElement || Opcode == Instruction::ExtractValue" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 231, __PRETTY_FUNCTION__)) | |||
231 | Opcode == Instruction::ExtractValue)((Opcode == Instruction::ExtractElement || Opcode == Instruction ::ExtractValue) ? static_cast<void> (0) : __assert_fail ("Opcode == Instruction::ExtractElement || Opcode == Instruction::ExtractValue" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 231, __PRETTY_FUNCTION__)); | |||
232 | if (Opcode == Instruction::ExtractElement) { | |||
233 | ConstantInt *CI = dyn_cast<ConstantInt>(E->getOperand(1)); | |||
234 | return CI && CI->getZExtValue() == Idx; | |||
235 | } else { | |||
236 | ExtractValueInst *EI = cast<ExtractValueInst>(E); | |||
237 | return EI->getNumIndices() == 1 && *EI->idx_begin() == Idx; | |||
238 | } | |||
239 | } | |||
240 | ||||
241 | /// \returns True if in-tree use also needs extract. This refers to | |||
242 | /// possible scalar operand in vectorized instruction. | |||
243 | static bool InTreeUserNeedToExtract(Value *Scalar, Instruction *UserInst, | |||
244 | TargetLibraryInfo *TLI) { | |||
245 | ||||
246 | unsigned Opcode = UserInst->getOpcode(); | |||
247 | switch (Opcode) { | |||
248 | case Instruction::Load: { | |||
249 | LoadInst *LI = cast<LoadInst>(UserInst); | |||
250 | return (LI->getPointerOperand() == Scalar); | |||
251 | } | |||
252 | case Instruction::Store: { | |||
253 | StoreInst *SI = cast<StoreInst>(UserInst); | |||
254 | return (SI->getPointerOperand() == Scalar); | |||
255 | } | |||
256 | case Instruction::Call: { | |||
257 | CallInst *CI = cast<CallInst>(UserInst); | |||
258 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); | |||
259 | if (hasVectorInstrinsicScalarOpd(ID, 1)) { | |||
260 | return (CI->getArgOperand(1) == Scalar); | |||
261 | } | |||
262 | LLVM_FALLTHROUGH[[clang::fallthrough]]; | |||
263 | } | |||
264 | default: | |||
265 | return false; | |||
266 | } | |||
267 | } | |||
268 | ||||
269 | /// \returns the AA location that is being access by the instruction. | |||
270 | static MemoryLocation getLocation(Instruction *I, AliasAnalysis *AA) { | |||
271 | if (StoreInst *SI = dyn_cast<StoreInst>(I)) | |||
272 | return MemoryLocation::get(SI); | |||
273 | if (LoadInst *LI = dyn_cast<LoadInst>(I)) | |||
274 | return MemoryLocation::get(LI); | |||
275 | return MemoryLocation(); | |||
276 | } | |||
277 | ||||
278 | /// \returns True if the instruction is not a volatile or atomic load/store. | |||
279 | static bool isSimple(Instruction *I) { | |||
280 | if (LoadInst *LI = dyn_cast<LoadInst>(I)) | |||
281 | return LI->isSimple(); | |||
282 | if (StoreInst *SI = dyn_cast<StoreInst>(I)) | |||
283 | return SI->isSimple(); | |||
284 | if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I)) | |||
285 | return !MI->isVolatile(); | |||
286 | return true; | |||
287 | } | |||
288 | ||||
289 | namespace llvm { | |||
290 | namespace slpvectorizer { | |||
291 | /// Bottom Up SLP Vectorizer. | |||
292 | class BoUpSLP { | |||
293 | public: | |||
294 | typedef SmallVector<Value *, 8> ValueList; | |||
295 | typedef SmallVector<Instruction *, 16> InstrList; | |||
296 | typedef SmallPtrSet<Value *, 16> ValueSet; | |||
297 | typedef SmallVector<StoreInst *, 8> StoreList; | |||
298 | typedef MapVector<Value *, SmallVector<Instruction *, 2>> | |||
299 | ExtraValueToDebugLocsMap; | |||
300 | ||||
301 | BoUpSLP(Function *Func, ScalarEvolution *Se, TargetTransformInfo *Tti, | |||
302 | TargetLibraryInfo *TLi, AliasAnalysis *Aa, LoopInfo *Li, | |||
303 | DominatorTree *Dt, AssumptionCache *AC, DemandedBits *DB, | |||
304 | const DataLayout *DL, OptimizationRemarkEmitter *ORE) | |||
305 | : NumLoadsWantToKeepOrder(0), NumLoadsWantToChangeOrder(0), F(Func), | |||
306 | SE(Se), TTI(Tti), TLI(TLi), AA(Aa), LI(Li), DT(Dt), AC(AC), DB(DB), | |||
307 | DL(DL), ORE(ORE), Builder(Se->getContext()) { | |||
308 | CodeMetrics::collectEphemeralValues(F, AC, EphValues); | |||
309 | // Use the vector register size specified by the target unless overridden | |||
310 | // by a command-line option. | |||
311 | // TODO: It would be better to limit the vectorization factor based on | |||
312 | // data type rather than just register size. For example, x86 AVX has | |||
313 | // 256-bit registers, but it does not support integer operations | |||
314 | // at that width (that requires AVX2). | |||
315 | if (MaxVectorRegSizeOption.getNumOccurrences()) | |||
316 | MaxVecRegSize = MaxVectorRegSizeOption; | |||
317 | else | |||
318 | MaxVecRegSize = TTI->getRegisterBitWidth(true); | |||
319 | ||||
320 | if (MinVectorRegSizeOption.getNumOccurrences()) | |||
321 | MinVecRegSize = MinVectorRegSizeOption; | |||
322 | else | |||
323 | MinVecRegSize = TTI->getMinVectorRegisterBitWidth(); | |||
324 | } | |||
325 | ||||
326 | /// \brief Vectorize the tree that starts with the elements in \p VL. | |||
327 | /// Returns the vectorized root. | |||
328 | Value *vectorizeTree(); | |||
329 | /// Vectorize the tree but with the list of externally used values \p | |||
330 | /// ExternallyUsedValues. Values in this MapVector can be replaced but the | |||
331 | /// generated extractvalue instructions. | |||
332 | Value *vectorizeTree(ExtraValueToDebugLocsMap &ExternallyUsedValues); | |||
333 | ||||
334 | /// \returns the cost incurred by unwanted spills and fills, caused by | |||
335 | /// holding live values over call sites. | |||
336 | int getSpillCost(); | |||
337 | ||||
338 | /// \returns the vectorization cost of the subtree that starts at \p VL. | |||
339 | /// A negative number means that this is profitable. | |||
340 | int getTreeCost(); | |||
341 | ||||
342 | /// Construct a vectorizable tree that starts at \p Roots, ignoring users for | |||
343 | /// the purpose of scheduling and extraction in the \p UserIgnoreLst. | |||
344 | void buildTree(ArrayRef<Value *> Roots, | |||
345 | ArrayRef<Value *> UserIgnoreLst = None); | |||
346 | /// Construct a vectorizable tree that starts at \p Roots, ignoring users for | |||
347 | /// the purpose of scheduling and extraction in the \p UserIgnoreLst taking | |||
348 | /// into account (anf updating it, if required) list of externally used | |||
349 | /// values stored in \p ExternallyUsedValues. | |||
350 | void buildTree(ArrayRef<Value *> Roots, | |||
351 | ExtraValueToDebugLocsMap &ExternallyUsedValues, | |||
352 | ArrayRef<Value *> UserIgnoreLst = None); | |||
353 | ||||
354 | /// Clear the internal data structures that are created by 'buildTree'. | |||
355 | void deleteTree() { | |||
356 | VectorizableTree.clear(); | |||
357 | ScalarToTreeEntry.clear(); | |||
358 | MustGather.clear(); | |||
359 | ExternalUses.clear(); | |||
360 | NumLoadsWantToKeepOrder = 0; | |||
361 | NumLoadsWantToChangeOrder = 0; | |||
362 | for (auto &Iter : BlocksSchedules) { | |||
363 | BlockScheduling *BS = Iter.second.get(); | |||
364 | BS->clear(); | |||
365 | } | |||
366 | MinBWs.clear(); | |||
367 | } | |||
368 | ||||
369 | unsigned getTreeSize() const { return VectorizableTree.size(); } | |||
370 | ||||
371 | /// \brief Perform LICM and CSE on the newly generated gather sequences. | |||
372 | void optimizeGatherSequence(); | |||
373 | ||||
374 | /// \returns true if it is beneficial to reverse the vector order. | |||
375 | bool shouldReorder() const { | |||
376 | return NumLoadsWantToChangeOrder > NumLoadsWantToKeepOrder; | |||
377 | } | |||
378 | ||||
379 | /// \return The vector element size in bits to use when vectorizing the | |||
380 | /// expression tree ending at \p V. If V is a store, the size is the width of | |||
381 | /// the stored value. Otherwise, the size is the width of the largest loaded | |||
382 | /// value reaching V. This method is used by the vectorizer to calculate | |||
383 | /// vectorization factors. | |||
384 | unsigned getVectorElementSize(Value *V); | |||
385 | ||||
386 | /// Compute the minimum type sizes required to represent the entries in a | |||
387 | /// vectorizable tree. | |||
388 | void computeMinimumValueSizes(); | |||
389 | ||||
390 | // \returns maximum vector register size as set by TTI or overridden by cl::opt. | |||
391 | unsigned getMaxVecRegSize() const { | |||
392 | return MaxVecRegSize; | |||
393 | } | |||
394 | ||||
395 | // \returns minimum vector register size as set by cl::opt. | |||
396 | unsigned getMinVecRegSize() const { | |||
397 | return MinVecRegSize; | |||
398 | } | |||
399 | ||||
400 | /// \brief Check if ArrayType or StructType is isomorphic to some VectorType. | |||
401 | /// | |||
402 | /// \returns number of elements in vector if isomorphism exists, 0 otherwise. | |||
403 | unsigned canMapToVector(Type *T, const DataLayout &DL) const; | |||
404 | ||||
405 | /// \returns True if the VectorizableTree is both tiny and not fully | |||
406 | /// vectorizable. We do not vectorize such trees. | |||
407 | bool isTreeTinyAndNotFullyVectorizable(); | |||
408 | ||||
409 | OptimizationRemarkEmitter *getORE() { return ORE; } | |||
410 | ||||
411 | private: | |||
412 | struct TreeEntry; | |||
413 | ||||
414 | /// \returns the cost of the vectorizable entry. | |||
415 | int getEntryCost(TreeEntry *E); | |||
416 | ||||
417 | /// This is the recursive part of buildTree. | |||
418 | void buildTree_rec(ArrayRef<Value *> Roots, unsigned Depth, int); | |||
419 | ||||
420 | /// \returns True if the ExtractElement/ExtractValue instructions in VL can | |||
421 | /// be vectorized to use the original vector (or aggregate "bitcast" to a vector). | |||
422 | bool canReuseExtract(ArrayRef<Value *> VL, unsigned Opcode) const; | |||
423 | ||||
424 | /// Vectorize a single entry in the tree. | |||
425 | Value *vectorizeTree(TreeEntry *E); | |||
426 | ||||
427 | /// Vectorize a single entry in the tree, starting in \p VL. | |||
428 | Value *vectorizeTree(ArrayRef<Value *> VL); | |||
429 | ||||
430 | /// \returns the pointer to the vectorized value if \p VL is already | |||
431 | /// vectorized, or NULL. They may happen in cycles. | |||
432 | Value *alreadyVectorized(ArrayRef<Value *> VL) const; | |||
433 | ||||
434 | /// \returns the scalarization cost for this type. Scalarization in this | |||
435 | /// context means the creation of vectors from a group of scalars. | |||
436 | int getGatherCost(Type *Ty); | |||
437 | ||||
438 | /// \returns the scalarization cost for this list of values. Assuming that | |||
439 | /// this subtree gets vectorized, we may need to extract the values from the | |||
440 | /// roots. This method calculates the cost of extracting the values. | |||
441 | int getGatherCost(ArrayRef<Value *> VL); | |||
442 | ||||
443 | /// \brief Set the Builder insert point to one after the last instruction in | |||
444 | /// the bundle | |||
445 | void setInsertPointAfterBundle(ArrayRef<Value *> VL); | |||
446 | ||||
447 | /// \returns a vector from a collection of scalars in \p VL. | |||
448 | Value *Gather(ArrayRef<Value *> VL, VectorType *Ty); | |||
449 | ||||
450 | /// \returns whether the VectorizableTree is fully vectorizable and will | |||
451 | /// be beneficial even the tree height is tiny. | |||
452 | bool isFullyVectorizableTinyTree(); | |||
453 | ||||
454 | /// \reorder commutative operands in alt shuffle if they result in | |||
455 | /// vectorized code. | |||
456 | void reorderAltShuffleOperands(ArrayRef<Value *> VL, | |||
457 | SmallVectorImpl<Value *> &Left, | |||
458 | SmallVectorImpl<Value *> &Right); | |||
459 | /// \reorder commutative operands to get better probability of | |||
460 | /// generating vectorized code. | |||
461 | void reorderInputsAccordingToOpcode(ArrayRef<Value *> VL, | |||
462 | SmallVectorImpl<Value *> &Left, | |||
463 | SmallVectorImpl<Value *> &Right); | |||
464 | struct TreeEntry { | |||
465 | TreeEntry(std::vector<TreeEntry> &Container) | |||
466 | : Scalars(), VectorizedValue(nullptr), NeedToGather(0), | |||
467 | Container(Container) {} | |||
468 | ||||
469 | /// \returns true if the scalars in VL are equal to this entry. | |||
470 | bool isSame(ArrayRef<Value *> VL) const { | |||
471 | assert(VL.size() == Scalars.size() && "Invalid size")((VL.size() == Scalars.size() && "Invalid size") ? static_cast <void> (0) : __assert_fail ("VL.size() == Scalars.size() && \"Invalid size\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 471, __PRETTY_FUNCTION__)); | |||
472 | return std::equal(VL.begin(), VL.end(), Scalars.begin()); | |||
473 | } | |||
474 | ||||
475 | /// A vector of scalars. | |||
476 | ValueList Scalars; | |||
477 | ||||
478 | /// The Scalars are vectorized into this value. It is initialized to Null. | |||
479 | Value *VectorizedValue; | |||
480 | ||||
481 | /// Do we need to gather this sequence ? | |||
482 | bool NeedToGather; | |||
483 | ||||
484 | /// Points back to the VectorizableTree. | |||
485 | /// | |||
486 | /// Only used for Graphviz right now. Unfortunately GraphTrait::NodeRef has | |||
487 | /// to be a pointer and needs to be able to initialize the child iterator. | |||
488 | /// Thus we need a reference back to the container to translate the indices | |||
489 | /// to entries. | |||
490 | std::vector<TreeEntry> &Container; | |||
491 | ||||
492 | /// The TreeEntry index containing the user of this entry. We can actually | |||
493 | /// have multiple users so the data structure is not truly a tree. | |||
494 | SmallVector<int, 1> UserTreeIndices; | |||
495 | }; | |||
496 | ||||
497 | /// Create a new VectorizableTree entry. | |||
498 | TreeEntry *newTreeEntry(ArrayRef<Value *> VL, bool Vectorized, | |||
499 | int &UserTreeIdx) { | |||
500 | VectorizableTree.emplace_back(VectorizableTree); | |||
501 | int idx = VectorizableTree.size() - 1; | |||
502 | TreeEntry *Last = &VectorizableTree[idx]; | |||
503 | Last->Scalars.insert(Last->Scalars.begin(), VL.begin(), VL.end()); | |||
504 | Last->NeedToGather = !Vectorized; | |||
505 | if (Vectorized) { | |||
506 | for (int i = 0, e = VL.size(); i != e; ++i) { | |||
507 | assert(!ScalarToTreeEntry.count(VL[i]) && "Scalar already in tree!")((!ScalarToTreeEntry.count(VL[i]) && "Scalar already in tree!" ) ? static_cast<void> (0) : __assert_fail ("!ScalarToTreeEntry.count(VL[i]) && \"Scalar already in tree!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 507, __PRETTY_FUNCTION__)); | |||
508 | ScalarToTreeEntry[VL[i]] = idx; | |||
509 | } | |||
510 | } else { | |||
511 | MustGather.insert(VL.begin(), VL.end()); | |||
512 | } | |||
513 | ||||
514 | if (UserTreeIdx >= 0) | |||
515 | Last->UserTreeIndices.push_back(UserTreeIdx); | |||
516 | UserTreeIdx = idx; | |||
517 | return Last; | |||
518 | } | |||
519 | ||||
520 | /// -- Vectorization State -- | |||
521 | /// Holds all of the tree entries. | |||
522 | std::vector<TreeEntry> VectorizableTree; | |||
523 | ||||
524 | /// Maps a specific scalar to its tree entry. | |||
525 | SmallDenseMap<Value*, int> ScalarToTreeEntry; | |||
526 | ||||
527 | /// A list of scalars that we found that we need to keep as scalars. | |||
528 | ValueSet MustGather; | |||
529 | ||||
530 | /// This POD struct describes one external user in the vectorized tree. | |||
531 | struct ExternalUser { | |||
532 | ExternalUser (Value *S, llvm::User *U, int L) : | |||
533 | Scalar(S), User(U), Lane(L){} | |||
534 | // Which scalar in our function. | |||
535 | Value *Scalar; | |||
536 | // Which user that uses the scalar. | |||
537 | llvm::User *User; | |||
538 | // Which lane does the scalar belong to. | |||
539 | int Lane; | |||
540 | }; | |||
541 | typedef SmallVector<ExternalUser, 16> UserList; | |||
542 | ||||
543 | /// Checks if two instructions may access the same memory. | |||
544 | /// | |||
545 | /// \p Loc1 is the location of \p Inst1. It is passed explicitly because it | |||
546 | /// is invariant in the calling loop. | |||
547 | bool isAliased(const MemoryLocation &Loc1, Instruction *Inst1, | |||
548 | Instruction *Inst2) { | |||
549 | ||||
550 | // First check if the result is already in the cache. | |||
551 | AliasCacheKey key = std::make_pair(Inst1, Inst2); | |||
552 | Optional<bool> &result = AliasCache[key]; | |||
553 | if (result.hasValue()) { | |||
554 | return result.getValue(); | |||
555 | } | |||
556 | MemoryLocation Loc2 = getLocation(Inst2, AA); | |||
557 | bool aliased = true; | |||
558 | if (Loc1.Ptr && Loc2.Ptr && isSimple(Inst1) && isSimple(Inst2)) { | |||
559 | // Do the alias check. | |||
560 | aliased = AA->alias(Loc1, Loc2); | |||
561 | } | |||
562 | // Store the result in the cache. | |||
563 | result = aliased; | |||
564 | return aliased; | |||
565 | } | |||
566 | ||||
567 | typedef std::pair<Instruction *, Instruction *> AliasCacheKey; | |||
568 | ||||
569 | /// Cache for alias results. | |||
570 | /// TODO: consider moving this to the AliasAnalysis itself. | |||
571 | DenseMap<AliasCacheKey, Optional<bool>> AliasCache; | |||
572 | ||||
573 | /// Removes an instruction from its block and eventually deletes it. | |||
574 | /// It's like Instruction::eraseFromParent() except that the actual deletion | |||
575 | /// is delayed until BoUpSLP is destructed. | |||
576 | /// This is required to ensure that there are no incorrect collisions in the | |||
577 | /// AliasCache, which can happen if a new instruction is allocated at the | |||
578 | /// same address as a previously deleted instruction. | |||
579 | void eraseInstruction(Instruction *I) { | |||
580 | I->removeFromParent(); | |||
581 | I->dropAllReferences(); | |||
582 | DeletedInstructions.emplace_back(I); | |||
583 | } | |||
584 | ||||
585 | /// Temporary store for deleted instructions. Instructions will be deleted | |||
586 | /// eventually when the BoUpSLP is destructed. | |||
587 | SmallVector<unique_value, 8> DeletedInstructions; | |||
588 | ||||
589 | /// A list of values that need to extracted out of the tree. | |||
590 | /// This list holds pairs of (Internal Scalar : External User). External User | |||
591 | /// can be nullptr, it means that this Internal Scalar will be used later, | |||
592 | /// after vectorization. | |||
593 | UserList ExternalUses; | |||
594 | ||||
595 | /// Values used only by @llvm.assume calls. | |||
596 | SmallPtrSet<const Value *, 32> EphValues; | |||
597 | ||||
598 | /// Holds all of the instructions that we gathered. | |||
599 | SetVector<Instruction *> GatherSeq; | |||
600 | /// A list of blocks that we are going to CSE. | |||
601 | SetVector<BasicBlock *> CSEBlocks; | |||
602 | ||||
603 | /// Contains all scheduling relevant data for an instruction. | |||
604 | /// A ScheduleData either represents a single instruction or a member of an | |||
605 | /// instruction bundle (= a group of instructions which is combined into a | |||
606 | /// vector instruction). | |||
607 | struct ScheduleData { | |||
608 | ||||
609 | // The initial value for the dependency counters. It means that the | |||
610 | // dependencies are not calculated yet. | |||
611 | enum { InvalidDeps = -1 }; | |||
612 | ||||
613 | ScheduleData() | |||
614 | : Inst(nullptr), FirstInBundle(nullptr), NextInBundle(nullptr), | |||
615 | NextLoadStore(nullptr), SchedulingRegionID(0), SchedulingPriority(0), | |||
616 | Dependencies(InvalidDeps), UnscheduledDeps(InvalidDeps), | |||
617 | UnscheduledDepsInBundle(InvalidDeps), IsScheduled(false) {} | |||
618 | ||||
619 | void init(int BlockSchedulingRegionID) { | |||
620 | FirstInBundle = this; | |||
621 | NextInBundle = nullptr; | |||
622 | NextLoadStore = nullptr; | |||
623 | IsScheduled = false; | |||
624 | SchedulingRegionID = BlockSchedulingRegionID; | |||
625 | UnscheduledDepsInBundle = UnscheduledDeps; | |||
626 | clearDependencies(); | |||
627 | } | |||
628 | ||||
629 | /// Returns true if the dependency information has been calculated. | |||
630 | bool hasValidDependencies() const { return Dependencies != InvalidDeps; } | |||
631 | ||||
632 | /// Returns true for single instructions and for bundle representatives | |||
633 | /// (= the head of a bundle). | |||
634 | bool isSchedulingEntity() const { return FirstInBundle == this; } | |||
635 | ||||
636 | /// Returns true if it represents an instruction bundle and not only a | |||
637 | /// single instruction. | |||
638 | bool isPartOfBundle() const { | |||
639 | return NextInBundle != nullptr || FirstInBundle != this; | |||
640 | } | |||
641 | ||||
642 | /// Returns true if it is ready for scheduling, i.e. it has no more | |||
643 | /// unscheduled depending instructions/bundles. | |||
644 | bool isReady() const { | |||
645 | assert(isSchedulingEntity() &&((isSchedulingEntity() && "can't consider non-scheduling entity for ready list" ) ? static_cast<void> (0) : __assert_fail ("isSchedulingEntity() && \"can't consider non-scheduling entity for ready list\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 646, __PRETTY_FUNCTION__)) | |||
646 | "can't consider non-scheduling entity for ready list")((isSchedulingEntity() && "can't consider non-scheduling entity for ready list" ) ? static_cast<void> (0) : __assert_fail ("isSchedulingEntity() && \"can't consider non-scheduling entity for ready list\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 646, __PRETTY_FUNCTION__)); | |||
647 | return UnscheduledDepsInBundle == 0 && !IsScheduled; | |||
648 | } | |||
649 | ||||
650 | /// Modifies the number of unscheduled dependencies, also updating it for | |||
651 | /// the whole bundle. | |||
652 | int incrementUnscheduledDeps(int Incr) { | |||
653 | UnscheduledDeps += Incr; | |||
654 | return FirstInBundle->UnscheduledDepsInBundle += Incr; | |||
655 | } | |||
656 | ||||
657 | /// Sets the number of unscheduled dependencies to the number of | |||
658 | /// dependencies. | |||
659 | void resetUnscheduledDeps() { | |||
660 | incrementUnscheduledDeps(Dependencies - UnscheduledDeps); | |||
661 | } | |||
662 | ||||
663 | /// Clears all dependency information. | |||
664 | void clearDependencies() { | |||
665 | Dependencies = InvalidDeps; | |||
666 | resetUnscheduledDeps(); | |||
667 | MemoryDependencies.clear(); | |||
668 | } | |||
669 | ||||
670 | void dump(raw_ostream &os) const { | |||
671 | if (!isSchedulingEntity()) { | |||
672 | os << "/ " << *Inst; | |||
673 | } else if (NextInBundle) { | |||
674 | os << '[' << *Inst; | |||
675 | ScheduleData *SD = NextInBundle; | |||
676 | while (SD) { | |||
677 | os << ';' << *SD->Inst; | |||
678 | SD = SD->NextInBundle; | |||
679 | } | |||
680 | os << ']'; | |||
681 | } else { | |||
682 | os << *Inst; | |||
683 | } | |||
684 | } | |||
685 | ||||
686 | Instruction *Inst; | |||
687 | ||||
688 | /// Points to the head in an instruction bundle (and always to this for | |||
689 | /// single instructions). | |||
690 | ScheduleData *FirstInBundle; | |||
691 | ||||
692 | /// Single linked list of all instructions in a bundle. Null if it is a | |||
693 | /// single instruction. | |||
694 | ScheduleData *NextInBundle; | |||
695 | ||||
696 | /// Single linked list of all memory instructions (e.g. load, store, call) | |||
697 | /// in the block - until the end of the scheduling region. | |||
698 | ScheduleData *NextLoadStore; | |||
699 | ||||
700 | /// The dependent memory instructions. | |||
701 | /// This list is derived on demand in calculateDependencies(). | |||
702 | SmallVector<ScheduleData *, 4> MemoryDependencies; | |||
703 | ||||
704 | /// This ScheduleData is in the current scheduling region if this matches | |||
705 | /// the current SchedulingRegionID of BlockScheduling. | |||
706 | int SchedulingRegionID; | |||
707 | ||||
708 | /// Used for getting a "good" final ordering of instructions. | |||
709 | int SchedulingPriority; | |||
710 | ||||
711 | /// The number of dependencies. Constitutes of the number of users of the | |||
712 | /// instruction plus the number of dependent memory instructions (if any). | |||
713 | /// This value is calculated on demand. | |||
714 | /// If InvalidDeps, the number of dependencies is not calculated yet. | |||
715 | /// | |||
716 | int Dependencies; | |||
717 | ||||
718 | /// The number of dependencies minus the number of dependencies of scheduled | |||
719 | /// instructions. As soon as this is zero, the instruction/bundle gets ready | |||
720 | /// for scheduling. | |||
721 | /// Note that this is negative as long as Dependencies is not calculated. | |||
722 | int UnscheduledDeps; | |||
723 | ||||
724 | /// The sum of UnscheduledDeps in a bundle. Equals to UnscheduledDeps for | |||
725 | /// single instructions. | |||
726 | int UnscheduledDepsInBundle; | |||
727 | ||||
728 | /// True if this instruction is scheduled (or considered as scheduled in the | |||
729 | /// dry-run). | |||
730 | bool IsScheduled; | |||
731 | }; | |||
732 | ||||
733 | #ifndef NDEBUG | |||
734 | friend inline raw_ostream &operator<<(raw_ostream &os, | |||
735 | const BoUpSLP::ScheduleData &SD) { | |||
736 | SD.dump(os); | |||
737 | return os; | |||
738 | } | |||
739 | #endif | |||
740 | friend struct GraphTraits<BoUpSLP *>; | |||
741 | friend struct DOTGraphTraits<BoUpSLP *>; | |||
742 | ||||
743 | /// Contains all scheduling data for a basic block. | |||
744 | /// | |||
745 | struct BlockScheduling { | |||
746 | ||||
747 | BlockScheduling(BasicBlock *BB) | |||
748 | : BB(BB), ChunkSize(BB->size()), ChunkPos(ChunkSize), | |||
749 | ScheduleStart(nullptr), ScheduleEnd(nullptr), | |||
750 | FirstLoadStoreInRegion(nullptr), LastLoadStoreInRegion(nullptr), | |||
751 | ScheduleRegionSize(0), | |||
752 | ScheduleRegionSizeLimit(ScheduleRegionSizeBudget), | |||
753 | // Make sure that the initial SchedulingRegionID is greater than the | |||
754 | // initial SchedulingRegionID in ScheduleData (which is 0). | |||
755 | SchedulingRegionID(1) {} | |||
756 | ||||
757 | void clear() { | |||
758 | ReadyInsts.clear(); | |||
759 | ScheduleStart = nullptr; | |||
760 | ScheduleEnd = nullptr; | |||
761 | FirstLoadStoreInRegion = nullptr; | |||
762 | LastLoadStoreInRegion = nullptr; | |||
763 | ||||
764 | // Reduce the maximum schedule region size by the size of the | |||
765 | // previous scheduling run. | |||
766 | ScheduleRegionSizeLimit -= ScheduleRegionSize; | |||
767 | if (ScheduleRegionSizeLimit < MinScheduleRegionSize) | |||
768 | ScheduleRegionSizeLimit = MinScheduleRegionSize; | |||
769 | ScheduleRegionSize = 0; | |||
770 | ||||
771 | // Make a new scheduling region, i.e. all existing ScheduleData is not | |||
772 | // in the new region yet. | |||
773 | ++SchedulingRegionID; | |||
774 | } | |||
775 | ||||
776 | ScheduleData *getScheduleData(Value *V) { | |||
777 | ScheduleData *SD = ScheduleDataMap[V]; | |||
778 | if (SD && SD->SchedulingRegionID == SchedulingRegionID) | |||
779 | return SD; | |||
780 | return nullptr; | |||
781 | } | |||
782 | ||||
783 | bool isInSchedulingRegion(ScheduleData *SD) { | |||
784 | return SD->SchedulingRegionID == SchedulingRegionID; | |||
785 | } | |||
786 | ||||
787 | /// Marks an instruction as scheduled and puts all dependent ready | |||
788 | /// instructions into the ready-list. | |||
789 | template <typename ReadyListType> | |||
790 | void schedule(ScheduleData *SD, ReadyListType &ReadyList) { | |||
791 | SD->IsScheduled = true; | |||
792 | DEBUG(dbgs() << "SLP: schedule " << *SD << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: schedule " << *SD << "\n"; } } while (false); | |||
793 | ||||
794 | ScheduleData *BundleMember = SD; | |||
795 | while (BundleMember) { | |||
796 | // Handle the def-use chain dependencies. | |||
797 | for (Use &U : BundleMember->Inst->operands()) { | |||
798 | ScheduleData *OpDef = getScheduleData(U.get()); | |||
799 | if (OpDef && OpDef->hasValidDependencies() && | |||
800 | OpDef->incrementUnscheduledDeps(-1) == 0) { | |||
801 | // There are no more unscheduled dependencies after decrementing, | |||
802 | // so we can put the dependent instruction into the ready list. | |||
803 | ScheduleData *DepBundle = OpDef->FirstInBundle; | |||
804 | assert(!DepBundle->IsScheduled &&((!DepBundle->IsScheduled && "already scheduled bundle gets ready" ) ? static_cast<void> (0) : __assert_fail ("!DepBundle->IsScheduled && \"already scheduled bundle gets ready\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 805, __PRETTY_FUNCTION__)) | |||
805 | "already scheduled bundle gets ready")((!DepBundle->IsScheduled && "already scheduled bundle gets ready" ) ? static_cast<void> (0) : __assert_fail ("!DepBundle->IsScheduled && \"already scheduled bundle gets ready\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 805, __PRETTY_FUNCTION__)); | |||
806 | ReadyList.insert(DepBundle); | |||
807 | DEBUG(dbgs() << "SLP: gets ready (def): " << *DepBundle << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: gets ready (def): " << *DepBundle << "\n"; } } while (false); | |||
808 | } | |||
809 | } | |||
810 | // Handle the memory dependencies. | |||
811 | for (ScheduleData *MemoryDepSD : BundleMember->MemoryDependencies) { | |||
812 | if (MemoryDepSD->incrementUnscheduledDeps(-1) == 0) { | |||
813 | // There are no more unscheduled dependencies after decrementing, | |||
814 | // so we can put the dependent instruction into the ready list. | |||
815 | ScheduleData *DepBundle = MemoryDepSD->FirstInBundle; | |||
816 | assert(!DepBundle->IsScheduled &&((!DepBundle->IsScheduled && "already scheduled bundle gets ready" ) ? static_cast<void> (0) : __assert_fail ("!DepBundle->IsScheduled && \"already scheduled bundle gets ready\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 817, __PRETTY_FUNCTION__)) | |||
817 | "already scheduled bundle gets ready")((!DepBundle->IsScheduled && "already scheduled bundle gets ready" ) ? static_cast<void> (0) : __assert_fail ("!DepBundle->IsScheduled && \"already scheduled bundle gets ready\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 817, __PRETTY_FUNCTION__)); | |||
818 | ReadyList.insert(DepBundle); | |||
819 | DEBUG(dbgs() << "SLP: gets ready (mem): " << *DepBundle << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: gets ready (mem): " << *DepBundle << "\n"; } } while (false); | |||
820 | } | |||
821 | } | |||
822 | BundleMember = BundleMember->NextInBundle; | |||
823 | } | |||
824 | } | |||
825 | ||||
826 | /// Put all instructions into the ReadyList which are ready for scheduling. | |||
827 | template <typename ReadyListType> | |||
828 | void initialFillReadyList(ReadyListType &ReadyList) { | |||
829 | for (auto *I = ScheduleStart; I != ScheduleEnd; I = I->getNextNode()) { | |||
830 | ScheduleData *SD = getScheduleData(I); | |||
831 | if (SD->isSchedulingEntity() && SD->isReady()) { | |||
832 | ReadyList.insert(SD); | |||
833 | 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); | |||
834 | } | |||
835 | } | |||
836 | } | |||
837 | ||||
838 | /// Checks if a bundle of instructions can be scheduled, i.e. has no | |||
839 | /// cyclic dependencies. This is only a dry-run, no instructions are | |||
840 | /// actually moved at this stage. | |||
841 | bool tryScheduleBundle(ArrayRef<Value *> VL, BoUpSLP *SLP); | |||
842 | ||||
843 | /// Un-bundles a group of instructions. | |||
844 | void cancelScheduling(ArrayRef<Value *> VL); | |||
845 | ||||
846 | /// Extends the scheduling region so that V is inside the region. | |||
847 | /// \returns true if the region size is within the limit. | |||
848 | bool extendSchedulingRegion(Value *V); | |||
849 | ||||
850 | /// Initialize the ScheduleData structures for new instructions in the | |||
851 | /// scheduling region. | |||
852 | void initScheduleData(Instruction *FromI, Instruction *ToI, | |||
853 | ScheduleData *PrevLoadStore, | |||
854 | ScheduleData *NextLoadStore); | |||
855 | ||||
856 | /// Updates the dependency information of a bundle and of all instructions/ | |||
857 | /// bundles which depend on the original bundle. | |||
858 | void calculateDependencies(ScheduleData *SD, bool InsertInReadyList, | |||
859 | BoUpSLP *SLP); | |||
860 | ||||
861 | /// Sets all instruction in the scheduling region to un-scheduled. | |||
862 | void resetSchedule(); | |||
863 | ||||
864 | BasicBlock *BB; | |||
865 | ||||
866 | /// Simple memory allocation for ScheduleData. | |||
867 | std::vector<std::unique_ptr<ScheduleData[]>> ScheduleDataChunks; | |||
868 | ||||
869 | /// The size of a ScheduleData array in ScheduleDataChunks. | |||
870 | int ChunkSize; | |||
871 | ||||
872 | /// The allocator position in the current chunk, which is the last entry | |||
873 | /// of ScheduleDataChunks. | |||
874 | int ChunkPos; | |||
875 | ||||
876 | /// Attaches ScheduleData to Instruction. | |||
877 | /// Note that the mapping survives during all vectorization iterations, i.e. | |||
878 | /// ScheduleData structures are recycled. | |||
879 | DenseMap<Value *, ScheduleData *> ScheduleDataMap; | |||
880 | ||||
881 | struct ReadyList : SmallVector<ScheduleData *, 8> { | |||
882 | void insert(ScheduleData *SD) { push_back(SD); } | |||
883 | }; | |||
884 | ||||
885 | /// The ready-list for scheduling (only used for the dry-run). | |||
886 | ReadyList ReadyInsts; | |||
887 | ||||
888 | /// The first instruction of the scheduling region. | |||
889 | Instruction *ScheduleStart; | |||
890 | ||||
891 | /// The first instruction _after_ the scheduling region. | |||
892 | Instruction *ScheduleEnd; | |||
893 | ||||
894 | /// The first memory accessing instruction in the scheduling region | |||
895 | /// (can be null). | |||
896 | ScheduleData *FirstLoadStoreInRegion; | |||
897 | ||||
898 | /// The last memory accessing instruction in the scheduling region | |||
899 | /// (can be null). | |||
900 | ScheduleData *LastLoadStoreInRegion; | |||
901 | ||||
902 | /// The current size of the scheduling region. | |||
903 | int ScheduleRegionSize; | |||
904 | ||||
905 | /// The maximum size allowed for the scheduling region. | |||
906 | int ScheduleRegionSizeLimit; | |||
907 | ||||
908 | /// The ID of the scheduling region. For a new vectorization iteration this | |||
909 | /// is incremented which "removes" all ScheduleData from the region. | |||
910 | int SchedulingRegionID; | |||
911 | }; | |||
912 | ||||
913 | /// Attaches the BlockScheduling structures to basic blocks. | |||
914 | MapVector<BasicBlock *, std::unique_ptr<BlockScheduling>> BlocksSchedules; | |||
915 | ||||
916 | /// Performs the "real" scheduling. Done before vectorization is actually | |||
917 | /// performed in a basic block. | |||
918 | void scheduleBlock(BlockScheduling *BS); | |||
919 | ||||
920 | /// List of users to ignore during scheduling and that don't need extracting. | |||
921 | ArrayRef<Value *> UserIgnoreList; | |||
922 | ||||
923 | // Number of load bundles that contain consecutive loads. | |||
924 | int NumLoadsWantToKeepOrder; | |||
925 | ||||
926 | // Number of load bundles that contain consecutive loads in reversed order. | |||
927 | int NumLoadsWantToChangeOrder; | |||
928 | ||||
929 | // Analysis and block reference. | |||
930 | Function *F; | |||
931 | ScalarEvolution *SE; | |||
932 | TargetTransformInfo *TTI; | |||
933 | TargetLibraryInfo *TLI; | |||
934 | AliasAnalysis *AA; | |||
935 | LoopInfo *LI; | |||
936 | DominatorTree *DT; | |||
937 | AssumptionCache *AC; | |||
938 | DemandedBits *DB; | |||
939 | const DataLayout *DL; | |||
940 | OptimizationRemarkEmitter *ORE; | |||
941 | ||||
942 | unsigned MaxVecRegSize; // This is set by TTI or overridden by cl::opt. | |||
943 | unsigned MinVecRegSize; // Set by cl::opt (default: 128). | |||
944 | /// Instruction builder to construct the vectorized tree. | |||
945 | IRBuilder<> Builder; | |||
946 | ||||
947 | /// A map of scalar integer values to the smallest bit width with which they | |||
948 | /// can legally be represented. The values map to (width, signed) pairs, | |||
949 | /// where "width" indicates the minimum bit width and "signed" is True if the | |||
950 | /// value must be signed-extended, rather than zero-extended, back to its | |||
951 | /// original width. | |||
952 | MapVector<Value *, std::pair<uint64_t, bool>> MinBWs; | |||
953 | }; | |||
954 | } // end namespace slpvectorizer | |||
955 | ||||
956 | template <> struct GraphTraits<BoUpSLP *> { | |||
957 | typedef BoUpSLP::TreeEntry TreeEntry; | |||
958 | ||||
959 | /// NodeRef has to be a pointer per the GraphWriter. | |||
960 | typedef TreeEntry *NodeRef; | |||
961 | ||||
962 | /// \brief Add the VectorizableTree to the index iterator to be able to return | |||
963 | /// TreeEntry pointers. | |||
964 | struct ChildIteratorType | |||
965 | : public iterator_adaptor_base<ChildIteratorType, | |||
966 | SmallVector<int, 1>::iterator> { | |||
967 | ||||
968 | std::vector<TreeEntry> &VectorizableTree; | |||
969 | ||||
970 | ChildIteratorType(SmallVector<int, 1>::iterator W, | |||
971 | std::vector<TreeEntry> &VT) | |||
972 | : ChildIteratorType::iterator_adaptor_base(W), VectorizableTree(VT) {} | |||
973 | ||||
974 | NodeRef operator*() { return &VectorizableTree[*I]; } | |||
975 | }; | |||
976 | ||||
977 | static NodeRef getEntryNode(BoUpSLP &R) { return &R.VectorizableTree[0]; } | |||
978 | ||||
979 | static ChildIteratorType child_begin(NodeRef N) { | |||
980 | return {N->UserTreeIndices.begin(), N->Container}; | |||
981 | } | |||
982 | static ChildIteratorType child_end(NodeRef N) { | |||
983 | return {N->UserTreeIndices.end(), N->Container}; | |||
984 | } | |||
985 | ||||
986 | /// For the node iterator we just need to turn the TreeEntry iterator into a | |||
987 | /// TreeEntry* iterator so that it dereferences to NodeRef. | |||
988 | typedef pointer_iterator<std::vector<TreeEntry>::iterator> nodes_iterator; | |||
989 | ||||
990 | static nodes_iterator nodes_begin(BoUpSLP *R) { | |||
991 | return nodes_iterator(R->VectorizableTree.begin()); | |||
992 | } | |||
993 | static nodes_iterator nodes_end(BoUpSLP *R) { | |||
994 | return nodes_iterator(R->VectorizableTree.end()); | |||
995 | } | |||
996 | ||||
997 | static unsigned size(BoUpSLP *R) { return R->VectorizableTree.size(); } | |||
998 | }; | |||
999 | ||||
1000 | template <> struct DOTGraphTraits<BoUpSLP *> : public DefaultDOTGraphTraits { | |||
1001 | typedef BoUpSLP::TreeEntry TreeEntry; | |||
1002 | ||||
1003 | DOTGraphTraits(bool isSimple = false) : DefaultDOTGraphTraits(isSimple) {} | |||
1004 | ||||
1005 | std::string getNodeLabel(const TreeEntry *Entry, const BoUpSLP *R) { | |||
1006 | std::string Str; | |||
1007 | raw_string_ostream OS(Str); | |||
1008 | if (isSplat(Entry->Scalars)) { | |||
1009 | OS << "<splat> " << *Entry->Scalars[0]; | |||
1010 | return Str; | |||
1011 | } | |||
1012 | for (auto V : Entry->Scalars) { | |||
1013 | OS << *V; | |||
1014 | if (std::any_of( | |||
1015 | R->ExternalUses.begin(), R->ExternalUses.end(), | |||
1016 | [&](const BoUpSLP::ExternalUser &EU) { return EU.Scalar == V; })) | |||
1017 | OS << " <extract>"; | |||
1018 | OS << "\n"; | |||
1019 | } | |||
1020 | return Str; | |||
1021 | } | |||
1022 | ||||
1023 | static std::string getNodeAttributes(const TreeEntry *Entry, | |||
1024 | const BoUpSLP *) { | |||
1025 | if (Entry->NeedToGather) | |||
1026 | return "color=red"; | |||
1027 | return ""; | |||
1028 | } | |||
1029 | }; | |||
1030 | ||||
1031 | } // end namespace llvm | |||
1032 | ||||
1033 | void BoUpSLP::buildTree(ArrayRef<Value *> Roots, | |||
1034 | ArrayRef<Value *> UserIgnoreLst) { | |||
1035 | ExtraValueToDebugLocsMap ExternallyUsedValues; | |||
1036 | buildTree(Roots, ExternallyUsedValues, UserIgnoreLst); | |||
1037 | } | |||
1038 | void BoUpSLP::buildTree(ArrayRef<Value *> Roots, | |||
1039 | ExtraValueToDebugLocsMap &ExternallyUsedValues, | |||
1040 | ArrayRef<Value *> UserIgnoreLst) { | |||
1041 | deleteTree(); | |||
1042 | UserIgnoreList = UserIgnoreLst; | |||
1043 | if (!allSameType(Roots)) | |||
1044 | return; | |||
1045 | buildTree_rec(Roots, 0, -1); | |||
1046 | ||||
1047 | // Collect the values that we need to extract from the tree. | |||
1048 | for (TreeEntry &EIdx : VectorizableTree) { | |||
1049 | TreeEntry *Entry = &EIdx; | |||
1050 | ||||
1051 | // For each lane: | |||
1052 | for (int Lane = 0, LE = Entry->Scalars.size(); Lane != LE; ++Lane) { | |||
1053 | Value *Scalar = Entry->Scalars[Lane]; | |||
1054 | ||||
1055 | // No need to handle users of gathered values. | |||
1056 | if (Entry->NeedToGather) | |||
1057 | continue; | |||
1058 | ||||
1059 | // Check if the scalar is externally used as an extra arg. | |||
1060 | auto ExtI = ExternallyUsedValues.find(Scalar); | |||
1061 | if (ExtI != ExternallyUsedValues.end()) { | |||
1062 | 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) | |||
1063 | 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); | |||
1064 | ExternalUses.emplace_back(Scalar, nullptr, Lane); | |||
1065 | continue; | |||
1066 | } | |||
1067 | for (User *U : Scalar->users()) { | |||
1068 | DEBUG(dbgs() << "SLP: Checking user:" << *U << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Checking user:" << *U << ".\n"; } } while (false); | |||
1069 | ||||
1070 | Instruction *UserInst = dyn_cast<Instruction>(U); | |||
1071 | if (!UserInst) | |||
1072 | continue; | |||
1073 | ||||
1074 | // Skip in-tree scalars that become vectors | |||
1075 | if (ScalarToTreeEntry.count(U)) { | |||
1076 | int Idx = ScalarToTreeEntry[U]; | |||
1077 | TreeEntry *UseEntry = &VectorizableTree[Idx]; | |||
1078 | Value *UseScalar = UseEntry->Scalars[0]; | |||
1079 | // Some in-tree scalars will remain as scalar in vectorized | |||
1080 | // instructions. If that is the case, the one in Lane 0 will | |||
1081 | // be used. | |||
1082 | if (UseScalar != U || | |||
1083 | !InTreeUserNeedToExtract(Scalar, UserInst, TLI)) { | |||
1084 | 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) | |||
1085 | << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: \tInternal user will be removed:" << *U << ".\n"; } } while (false); | |||
1086 | assert(!VectorizableTree[Idx].NeedToGather && "Bad state")((!VectorizableTree[Idx].NeedToGather && "Bad state") ? static_cast<void> (0) : __assert_fail ("!VectorizableTree[Idx].NeedToGather && \"Bad state\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1086, __PRETTY_FUNCTION__)); | |||
1087 | continue; | |||
1088 | } | |||
1089 | } | |||
1090 | ||||
1091 | // Ignore users in the user ignore list. | |||
1092 | if (is_contained(UserIgnoreList, UserInst)) | |||
1093 | continue; | |||
1094 | ||||
1095 | 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) | |||
1096 | 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); | |||
1097 | ExternalUses.push_back(ExternalUser(Scalar, U, Lane)); | |||
1098 | } | |||
1099 | } | |||
1100 | } | |||
1101 | } | |||
1102 | ||||
1103 | void BoUpSLP::buildTree_rec(ArrayRef<Value *> VL, unsigned Depth, | |||
1104 | int UserTreeIdx) { | |||
1105 | bool isAltShuffle = false; | |||
1106 | assert((allConstant(VL) || allSameType(VL)) && "Invalid types!")(((allConstant(VL) || allSameType(VL)) && "Invalid types!" ) ? static_cast<void> (0) : __assert_fail ("(allConstant(VL) || allSameType(VL)) && \"Invalid types!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1106, __PRETTY_FUNCTION__)); | |||
1107 | ||||
1108 | if (Depth == RecursionMaxDepth) { | |||
1109 | 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); | |||
1110 | newTreeEntry(VL, false, UserTreeIdx); | |||
1111 | return; | |||
1112 | } | |||
1113 | ||||
1114 | // Don't handle vectors. | |||
1115 | if (VL[0]->getType()->isVectorTy()) { | |||
1116 | 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); | |||
1117 | newTreeEntry(VL, false, UserTreeIdx); | |||
1118 | return; | |||
1119 | } | |||
1120 | ||||
1121 | if (StoreInst *SI = dyn_cast<StoreInst>(VL[0])) | |||
1122 | if (SI->getValueOperand()->getType()->isVectorTy()) { | |||
1123 | 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); | |||
1124 | newTreeEntry(VL, false, UserTreeIdx); | |||
1125 | return; | |||
1126 | } | |||
1127 | unsigned Opcode = getSameOpcode(VL); | |||
1128 | ||||
1129 | // Check that this shuffle vector refers to the alternate | |||
1130 | // sequence of opcodes. | |||
1131 | if (Opcode == Instruction::ShuffleVector) { | |||
1132 | Instruction *I0 = dyn_cast<Instruction>(VL[0]); | |||
1133 | unsigned Op = I0->getOpcode(); | |||
1134 | if (Op != Instruction::ShuffleVector) | |||
1135 | isAltShuffle = true; | |||
1136 | } | |||
1137 | ||||
1138 | // If all of the operands are identical or constant we have a simple solution. | |||
1139 | if (allConstant(VL) || isSplat(VL) || !allSameBlock(VL) || !Opcode) { | |||
1140 | 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); | |||
1141 | newTreeEntry(VL, false, UserTreeIdx); | |||
1142 | return; | |||
1143 | } | |||
1144 | ||||
1145 | // We now know that this is a vector of instructions of the same type from | |||
1146 | // the same block. | |||
1147 | ||||
1148 | // Don't vectorize ephemeral values. | |||
1149 | for (unsigned i = 0, e = VL.size(); i != e; ++i) { | |||
1150 | if (EphValues.count(VL[i])) { | |||
1151 | 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) | |||
1152 | ") is ephemeral.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: The instruction (" << * VL[i] << ") is ephemeral.\n"; } } while (false); | |||
1153 | newTreeEntry(VL, false, UserTreeIdx); | |||
1154 | return; | |||
1155 | } | |||
1156 | } | |||
1157 | ||||
1158 | // Check if this is a duplicate of another entry. | |||
1159 | if (ScalarToTreeEntry.count(VL[0])) { | |||
1160 | int Idx = ScalarToTreeEntry[VL[0]]; | |||
1161 | TreeEntry *E = &VectorizableTree[Idx]; | |||
1162 | for (unsigned i = 0, e = VL.size(); i != e; ++i) { | |||
1163 | DEBUG(dbgs() << "SLP: \tChecking bundle: " << *VL[i] << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: \tChecking bundle: " << *VL[i] << ".\n"; } } while (false); | |||
1164 | if (E->Scalars[i] != VL[i]) { | |||
1165 | 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); | |||
1166 | newTreeEntry(VL, false, UserTreeIdx); | |||
1167 | return; | |||
1168 | } | |||
1169 | } | |||
1170 | // Record the reuse of the tree node. FIXME, currently this is only used to | |||
1171 | // properly draw the graph rather than for the actual vectorization. | |||
1172 | E->UserTreeIndices.push_back(UserTreeIdx); | |||
1173 | DEBUG(dbgs() << "SLP: Perfect diamond merge at " << *VL[0] << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Perfect diamond merge at " << *VL[0] << ".\n"; } } while (false); | |||
1174 | return; | |||
1175 | } | |||
1176 | ||||
1177 | // Check that none of the instructions in the bundle are already in the tree. | |||
1178 | for (unsigned i = 0, e = VL.size(); i != e; ++i) { | |||
1179 | if (ScalarToTreeEntry.count(VL[i])) { | |||
1180 | 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) | |||
1181 | ") 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); | |||
1182 | newTreeEntry(VL, false, UserTreeIdx); | |||
1183 | return; | |||
1184 | } | |||
1185 | } | |||
1186 | ||||
1187 | // If any of the scalars is marked as a value that needs to stay scalar then | |||
1188 | // we need to gather the scalars. | |||
1189 | for (unsigned i = 0, e = VL.size(); i != e; ++i) { | |||
1190 | if (MustGather.count(VL[i])) { | |||
1191 | 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); | |||
1192 | newTreeEntry(VL, false, UserTreeIdx); | |||
1193 | return; | |||
1194 | } | |||
1195 | } | |||
1196 | ||||
1197 | // Check that all of the users of the scalars that we want to vectorize are | |||
1198 | // schedulable. | |||
1199 | Instruction *VL0 = cast<Instruction>(VL[0]); | |||
1200 | BasicBlock *BB = cast<Instruction>(VL0)->getParent(); | |||
1201 | ||||
1202 | if (!DT->isReachableFromEntry(BB)) { | |||
1203 | // Don't go into unreachable blocks. They may contain instructions with | |||
1204 | // dependency cycles which confuse the final scheduling. | |||
1205 | DEBUG(dbgs() << "SLP: bundle in unreachable block.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: bundle in unreachable block.\n" ; } } while (false); | |||
1206 | newTreeEntry(VL, false, UserTreeIdx); | |||
1207 | return; | |||
1208 | } | |||
1209 | ||||
1210 | // Check that every instructions appears once in this bundle. | |||
1211 | for (unsigned i = 0, e = VL.size(); i < e; ++i) | |||
1212 | for (unsigned j = i+1; j < e; ++j) | |||
1213 | if (VL[i] == VL[j]) { | |||
1214 | 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); | |||
1215 | newTreeEntry(VL, false, UserTreeIdx); | |||
1216 | return; | |||
1217 | } | |||
1218 | ||||
1219 | auto &BSRef = BlocksSchedules[BB]; | |||
1220 | if (!BSRef) { | |||
1221 | BSRef = llvm::make_unique<BlockScheduling>(BB); | |||
1222 | } | |||
1223 | BlockScheduling &BS = *BSRef.get(); | |||
1224 | ||||
1225 | if (!BS.tryScheduleBundle(VL, this)) { | |||
1226 | 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); | |||
1227 | assert((!BS.getScheduleData(VL[0]) ||(((!BS.getScheduleData(VL[0]) || !BS.getScheduleData(VL[0])-> isPartOfBundle()) && "tryScheduleBundle should cancelScheduling on failure" ) ? static_cast<void> (0) : __assert_fail ("(!BS.getScheduleData(VL[0]) || !BS.getScheduleData(VL[0])->isPartOfBundle()) && \"tryScheduleBundle should cancelScheduling on failure\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1229, __PRETTY_FUNCTION__)) | |||
1228 | !BS.getScheduleData(VL[0])->isPartOfBundle()) &&(((!BS.getScheduleData(VL[0]) || !BS.getScheduleData(VL[0])-> isPartOfBundle()) && "tryScheduleBundle should cancelScheduling on failure" ) ? static_cast<void> (0) : __assert_fail ("(!BS.getScheduleData(VL[0]) || !BS.getScheduleData(VL[0])->isPartOfBundle()) && \"tryScheduleBundle should cancelScheduling on failure\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1229, __PRETTY_FUNCTION__)) | |||
1229 | "tryScheduleBundle should cancelScheduling on failure")(((!BS.getScheduleData(VL[0]) || !BS.getScheduleData(VL[0])-> isPartOfBundle()) && "tryScheduleBundle should cancelScheduling on failure" ) ? static_cast<void> (0) : __assert_fail ("(!BS.getScheduleData(VL[0]) || !BS.getScheduleData(VL[0])->isPartOfBundle()) && \"tryScheduleBundle should cancelScheduling on failure\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1229, __PRETTY_FUNCTION__)); | |||
1230 | newTreeEntry(VL, false, UserTreeIdx); | |||
1231 | return; | |||
1232 | } | |||
1233 | 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); | |||
1234 | ||||
1235 | switch (Opcode) { | |||
1236 | case Instruction::PHI: { | |||
1237 | PHINode *PH = dyn_cast<PHINode>(VL0); | |||
1238 | ||||
1239 | // Check for terminator values (e.g. invoke). | |||
1240 | for (unsigned j = 0; j < VL.size(); ++j) | |||
1241 | for (unsigned i = 0, e = PH->getNumIncomingValues(); i < e; ++i) { | |||
1242 | TerminatorInst *Term = dyn_cast<TerminatorInst>( | |||
1243 | cast<PHINode>(VL[j])->getIncomingValueForBlock(PH->getIncomingBlock(i))); | |||
1244 | if (Term) { | |||
1245 | 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); | |||
1246 | BS.cancelScheduling(VL); | |||
1247 | newTreeEntry(VL, false, UserTreeIdx); | |||
1248 | return; | |||
1249 | } | |||
1250 | } | |||
1251 | ||||
1252 | newTreeEntry(VL, true, UserTreeIdx); | |||
1253 | 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); | |||
1254 | ||||
1255 | for (unsigned i = 0, e = PH->getNumIncomingValues(); i < e; ++i) { | |||
1256 | ValueList Operands; | |||
1257 | // Prepare the operand vector. | |||
1258 | for (Value *j : VL) | |||
1259 | Operands.push_back(cast<PHINode>(j)->getIncomingValueForBlock( | |||
1260 | PH->getIncomingBlock(i))); | |||
1261 | ||||
1262 | buildTree_rec(Operands, Depth + 1, UserTreeIdx); | |||
1263 | } | |||
1264 | return; | |||
1265 | } | |||
1266 | case Instruction::ExtractValue: | |||
1267 | case Instruction::ExtractElement: { | |||
1268 | bool Reuse = canReuseExtract(VL, Opcode); | |||
1269 | if (Reuse) { | |||
1270 | DEBUG(dbgs() << "SLP: Reusing extract sequence.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Reusing extract sequence.\n" ; } } while (false); | |||
1271 | } else { | |||
1272 | BS.cancelScheduling(VL); | |||
1273 | } | |||
1274 | newTreeEntry(VL, Reuse, UserTreeIdx); | |||
1275 | return; | |||
1276 | } | |||
1277 | case Instruction::Load: { | |||
1278 | // Check that a vectorized load would load the same memory as a scalar | |||
1279 | // load. | |||
1280 | // For example we don't want vectorize loads that are smaller than 8 bit. | |||
1281 | // Even though we have a packed struct {<i2, i2, i2, i2>} LLVM treats | |||
1282 | // loading/storing it as an i8 struct. If we vectorize loads/stores from | |||
1283 | // such a struct we read/write packed bits disagreeing with the | |||
1284 | // unvectorized version. | |||
1285 | Type *ScalarTy = VL[0]->getType(); | |||
1286 | ||||
1287 | if (DL->getTypeSizeInBits(ScalarTy) != | |||
1288 | DL->getTypeAllocSizeInBits(ScalarTy)) { | |||
1289 | BS.cancelScheduling(VL); | |||
1290 | newTreeEntry(VL, false, UserTreeIdx); | |||
1291 | 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); | |||
1292 | return; | |||
1293 | } | |||
1294 | ||||
1295 | // Make sure all loads in the bundle are simple - we can't vectorize | |||
1296 | // atomic or volatile loads. | |||
1297 | for (unsigned i = 0, e = VL.size() - 1; i < e; ++i) { | |||
1298 | LoadInst *L = cast<LoadInst>(VL[i]); | |||
1299 | if (!L->isSimple()) { | |||
1300 | BS.cancelScheduling(VL); | |||
1301 | newTreeEntry(VL, false, UserTreeIdx); | |||
1302 | DEBUG(dbgs() << "SLP: Gathering non-simple loads.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering non-simple loads.\n" ; } } while (false); | |||
1303 | return; | |||
1304 | } | |||
1305 | } | |||
1306 | ||||
1307 | // Check if the loads are consecutive, reversed, or neither. | |||
1308 | // TODO: What we really want is to sort the loads, but for now, check | |||
1309 | // the two likely directions. | |||
1310 | bool Consecutive = true; | |||
1311 | bool ReverseConsecutive = true; | |||
1312 | for (unsigned i = 0, e = VL.size() - 1; i < e; ++i) { | |||
1313 | if (!isConsecutiveAccess(VL[i], VL[i + 1], *DL, *SE)) { | |||
1314 | Consecutive = false; | |||
1315 | break; | |||
1316 | } else { | |||
1317 | ReverseConsecutive = false; | |||
1318 | } | |||
1319 | } | |||
1320 | ||||
1321 | if (Consecutive) { | |||
1322 | ++NumLoadsWantToKeepOrder; | |||
1323 | newTreeEntry(VL, true, UserTreeIdx); | |||
1324 | 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); | |||
1325 | return; | |||
1326 | } | |||
1327 | ||||
1328 | // If none of the load pairs were consecutive when checked in order, | |||
1329 | // check the reverse order. | |||
1330 | if (ReverseConsecutive) | |||
1331 | for (unsigned i = VL.size() - 1; i > 0; --i) | |||
1332 | if (!isConsecutiveAccess(VL[i], VL[i - 1], *DL, *SE)) { | |||
1333 | ReverseConsecutive = false; | |||
1334 | break; | |||
1335 | } | |||
1336 | ||||
1337 | BS.cancelScheduling(VL); | |||
1338 | newTreeEntry(VL, false, UserTreeIdx); | |||
1339 | ||||
1340 | if (ReverseConsecutive) { | |||
1341 | ++NumLoadsWantToChangeOrder; | |||
1342 | DEBUG(dbgs() << "SLP: Gathering reversed loads.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering reversed loads.\n" ; } } while (false); | |||
1343 | } else { | |||
1344 | DEBUG(dbgs() << "SLP: Gathering non-consecutive loads.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering non-consecutive loads.\n" ; } } while (false); | |||
1345 | } | |||
1346 | return; | |||
1347 | } | |||
1348 | case Instruction::ZExt: | |||
1349 | case Instruction::SExt: | |||
1350 | case Instruction::FPToUI: | |||
1351 | case Instruction::FPToSI: | |||
1352 | case Instruction::FPExt: | |||
1353 | case Instruction::PtrToInt: | |||
1354 | case Instruction::IntToPtr: | |||
1355 | case Instruction::SIToFP: | |||
1356 | case Instruction::UIToFP: | |||
1357 | case Instruction::Trunc: | |||
1358 | case Instruction::FPTrunc: | |||
1359 | case Instruction::BitCast: { | |||
1360 | Type *SrcTy = VL0->getOperand(0)->getType(); | |||
1361 | for (unsigned i = 0; i < VL.size(); ++i) { | |||
1362 | Type *Ty = cast<Instruction>(VL[i])->getOperand(0)->getType(); | |||
1363 | if (Ty != SrcTy || !isValidElementType(Ty)) { | |||
1364 | BS.cancelScheduling(VL); | |||
1365 | newTreeEntry(VL, false, UserTreeIdx); | |||
1366 | 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); | |||
1367 | return; | |||
1368 | } | |||
1369 | } | |||
1370 | newTreeEntry(VL, true, UserTreeIdx); | |||
1371 | 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); | |||
1372 | ||||
1373 | for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) { | |||
1374 | ValueList Operands; | |||
1375 | // Prepare the operand vector. | |||
1376 | for (Value *j : VL) | |||
1377 | Operands.push_back(cast<Instruction>(j)->getOperand(i)); | |||
1378 | ||||
1379 | buildTree_rec(Operands, Depth + 1, UserTreeIdx); | |||
1380 | } | |||
1381 | return; | |||
1382 | } | |||
1383 | case Instruction::ICmp: | |||
1384 | case Instruction::FCmp: { | |||
1385 | // Check that all of the compares have the same predicate. | |||
1386 | CmpInst::Predicate P0 = cast<CmpInst>(VL0)->getPredicate(); | |||
1387 | Type *ComparedTy = cast<Instruction>(VL[0])->getOperand(0)->getType(); | |||
1388 | for (unsigned i = 1, e = VL.size(); i < e; ++i) { | |||
1389 | CmpInst *Cmp = cast<CmpInst>(VL[i]); | |||
1390 | if (Cmp->getPredicate() != P0 || | |||
1391 | Cmp->getOperand(0)->getType() != ComparedTy) { | |||
1392 | BS.cancelScheduling(VL); | |||
1393 | newTreeEntry(VL, false, UserTreeIdx); | |||
1394 | 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); | |||
1395 | return; | |||
1396 | } | |||
1397 | } | |||
1398 | ||||
1399 | newTreeEntry(VL, true, UserTreeIdx); | |||
1400 | 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); | |||
1401 | ||||
1402 | for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) { | |||
1403 | ValueList Operands; | |||
1404 | // Prepare the operand vector. | |||
1405 | for (Value *j : VL) | |||
1406 | Operands.push_back(cast<Instruction>(j)->getOperand(i)); | |||
1407 | ||||
1408 | buildTree_rec(Operands, Depth + 1, UserTreeIdx); | |||
1409 | } | |||
1410 | return; | |||
1411 | } | |||
1412 | case Instruction::Select: | |||
1413 | case Instruction::Add: | |||
1414 | case Instruction::FAdd: | |||
1415 | case Instruction::Sub: | |||
1416 | case Instruction::FSub: | |||
1417 | case Instruction::Mul: | |||
1418 | case Instruction::FMul: | |||
1419 | case Instruction::UDiv: | |||
1420 | case Instruction::SDiv: | |||
1421 | case Instruction::FDiv: | |||
1422 | case Instruction::URem: | |||
1423 | case Instruction::SRem: | |||
1424 | case Instruction::FRem: | |||
1425 | case Instruction::Shl: | |||
1426 | case Instruction::LShr: | |||
1427 | case Instruction::AShr: | |||
1428 | case Instruction::And: | |||
1429 | case Instruction::Or: | |||
1430 | case Instruction::Xor: { | |||
1431 | newTreeEntry(VL, true, UserTreeIdx); | |||
1432 | 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); | |||
1433 | ||||
1434 | // Sort operands of the instructions so that each side is more likely to | |||
1435 | // have the same opcode. | |||
1436 | if (isa<BinaryOperator>(VL0) && VL0->isCommutative()) { | |||
1437 | ValueList Left, Right; | |||
1438 | reorderInputsAccordingToOpcode(VL, Left, Right); | |||
1439 | buildTree_rec(Left, Depth + 1, UserTreeIdx); | |||
1440 | buildTree_rec(Right, Depth + 1, UserTreeIdx); | |||
1441 | return; | |||
1442 | } | |||
1443 | ||||
1444 | for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) { | |||
1445 | ValueList Operands; | |||
1446 | // Prepare the operand vector. | |||
1447 | for (Value *j : VL) | |||
1448 | Operands.push_back(cast<Instruction>(j)->getOperand(i)); | |||
1449 | ||||
1450 | buildTree_rec(Operands, Depth + 1, UserTreeIdx); | |||
1451 | } | |||
1452 | return; | |||
1453 | } | |||
1454 | case Instruction::GetElementPtr: { | |||
1455 | // We don't combine GEPs with complicated (nested) indexing. | |||
1456 | for (unsigned j = 0; j < VL.size(); ++j) { | |||
1457 | if (cast<Instruction>(VL[j])->getNumOperands() != 2) { | |||
1458 | 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); | |||
1459 | BS.cancelScheduling(VL); | |||
1460 | newTreeEntry(VL, false, UserTreeIdx); | |||
1461 | return; | |||
1462 | } | |||
1463 | } | |||
1464 | ||||
1465 | // We can't combine several GEPs into one vector if they operate on | |||
1466 | // different types. | |||
1467 | Type *Ty0 = cast<Instruction>(VL0)->getOperand(0)->getType(); | |||
1468 | for (unsigned j = 0; j < VL.size(); ++j) { | |||
1469 | Type *CurTy = cast<Instruction>(VL[j])->getOperand(0)->getType(); | |||
1470 | if (Ty0 != CurTy) { | |||
1471 | 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); | |||
1472 | BS.cancelScheduling(VL); | |||
1473 | newTreeEntry(VL, false, UserTreeIdx); | |||
1474 | return; | |||
1475 | } | |||
1476 | } | |||
1477 | ||||
1478 | // We don't combine GEPs with non-constant indexes. | |||
1479 | for (unsigned j = 0; j < VL.size(); ++j) { | |||
1480 | auto Op = cast<Instruction>(VL[j])->getOperand(1); | |||
1481 | if (!isa<ConstantInt>(Op)) { | |||
1482 | DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: not-vectorizable GEP (non-constant indexes).\n" ; } } while (false) | |||
1483 | 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); | |||
1484 | BS.cancelScheduling(VL); | |||
1485 | newTreeEntry(VL, false, UserTreeIdx); | |||
1486 | return; | |||
1487 | } | |||
1488 | } | |||
1489 | ||||
1490 | newTreeEntry(VL, true, UserTreeIdx); | |||
1491 | 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); | |||
1492 | for (unsigned i = 0, e = 2; i < e; ++i) { | |||
1493 | ValueList Operands; | |||
1494 | // Prepare the operand vector. | |||
1495 | for (Value *j : VL) | |||
1496 | Operands.push_back(cast<Instruction>(j)->getOperand(i)); | |||
1497 | ||||
1498 | buildTree_rec(Operands, Depth + 1, UserTreeIdx); | |||
1499 | } | |||
1500 | return; | |||
1501 | } | |||
1502 | case Instruction::Store: { | |||
1503 | // Check if the stores are consecutive or of we need to swizzle them. | |||
1504 | for (unsigned i = 0, e = VL.size() - 1; i < e; ++i) | |||
1505 | if (!isConsecutiveAccess(VL[i], VL[i + 1], *DL, *SE)) { | |||
1506 | BS.cancelScheduling(VL); | |||
1507 | newTreeEntry(VL, false, UserTreeIdx); | |||
1508 | DEBUG(dbgs() << "SLP: Non-consecutive store.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Non-consecutive store.\n"; } } while (false); | |||
1509 | return; | |||
1510 | } | |||
1511 | ||||
1512 | newTreeEntry(VL, true, UserTreeIdx); | |||
1513 | 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); | |||
1514 | ||||
1515 | ValueList Operands; | |||
1516 | for (Value *j : VL) | |||
1517 | Operands.push_back(cast<Instruction>(j)->getOperand(0)); | |||
1518 | ||||
1519 | buildTree_rec(Operands, Depth + 1, UserTreeIdx); | |||
1520 | return; | |||
1521 | } | |||
1522 | case Instruction::Call: { | |||
1523 | // Check if the calls are all to the same vectorizable intrinsic. | |||
1524 | CallInst *CI = cast<CallInst>(VL[0]); | |||
1525 | // Check if this is an Intrinsic call or something that can be | |||
1526 | // represented by an intrinsic call | |||
1527 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); | |||
1528 | if (!isTriviallyVectorizable(ID)) { | |||
1529 | BS.cancelScheduling(VL); | |||
1530 | newTreeEntry(VL, false, UserTreeIdx); | |||
1531 | DEBUG(dbgs() << "SLP: Non-vectorizable call.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Non-vectorizable call.\n"; } } while (false); | |||
1532 | return; | |||
1533 | } | |||
1534 | Function *Int = CI->getCalledFunction(); | |||
1535 | Value *A1I = nullptr; | |||
1536 | if (hasVectorInstrinsicScalarOpd(ID, 1)) | |||
1537 | A1I = CI->getArgOperand(1); | |||
1538 | for (unsigned i = 1, e = VL.size(); i != e; ++i) { | |||
1539 | CallInst *CI2 = dyn_cast<CallInst>(VL[i]); | |||
1540 | if (!CI2 || CI2->getCalledFunction() != Int || | |||
1541 | getVectorIntrinsicIDForCall(CI2, TLI) != ID || | |||
1542 | !CI->hasIdenticalOperandBundleSchema(*CI2)) { | |||
1543 | BS.cancelScheduling(VL); | |||
1544 | newTreeEntry(VL, false, UserTreeIdx); | |||
1545 | 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 ) | |||
1546 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: mismatched calls:" << * CI << "!=" << *VL[i] << "\n"; } } while (false ); | |||
1547 | return; | |||
1548 | } | |||
1549 | // ctlz,cttz and powi are special intrinsics whose second argument | |||
1550 | // should be same in order for them to be vectorized. | |||
1551 | if (hasVectorInstrinsicScalarOpd(ID, 1)) { | |||
1552 | Value *A1J = CI2->getArgOperand(1); | |||
1553 | if (A1I != A1J) { | |||
1554 | BS.cancelScheduling(VL); | |||
1555 | newTreeEntry(VL, false, UserTreeIdx); | |||
1556 | 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) | |||
1557 | << " argument "<< A1I<<"!=" << A1Jdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: mismatched arguments in call:" << *CI << " argument "<< A1I<<"!=" << A1J << "\n"; } } while (false) | |||
1558 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: mismatched arguments in call:" << *CI << " argument "<< A1I<<"!=" << A1J << "\n"; } } while (false); | |||
1559 | return; | |||
1560 | } | |||
1561 | } | |||
1562 | // Verify that the bundle operands are identical between the two calls. | |||
1563 | if (CI->hasOperandBundles() && | |||
1564 | !std::equal(CI->op_begin() + CI->getBundleOperandsStartIndex(), | |||
1565 | CI->op_begin() + CI->getBundleOperandsEndIndex(), | |||
1566 | CI2->op_begin() + CI2->getBundleOperandsStartIndex())) { | |||
1567 | BS.cancelScheduling(VL); | |||
1568 | newTreeEntry(VL, false, UserTreeIdx); | |||
1569 | 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) | |||
1570 | << *VL[i] << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: mismatched bundle operands in calls:" << *CI << "!=" << *VL[i] << '\n'; } } while (false); | |||
1571 | return; | |||
1572 | } | |||
1573 | } | |||
1574 | ||||
1575 | newTreeEntry(VL, true, UserTreeIdx); | |||
1576 | for (unsigned i = 0, e = CI->getNumArgOperands(); i != e; ++i) { | |||
1577 | ValueList Operands; | |||
1578 | // Prepare the operand vector. | |||
1579 | for (Value *j : VL) { | |||
1580 | CallInst *CI2 = dyn_cast<CallInst>(j); | |||
1581 | Operands.push_back(CI2->getArgOperand(i)); | |||
1582 | } | |||
1583 | buildTree_rec(Operands, Depth + 1, UserTreeIdx); | |||
1584 | } | |||
1585 | return; | |||
1586 | } | |||
1587 | case Instruction::ShuffleVector: { | |||
1588 | // If this is not an alternate sequence of opcode like add-sub | |||
1589 | // then do not vectorize this instruction. | |||
1590 | if (!isAltShuffle) { | |||
1591 | BS.cancelScheduling(VL); | |||
1592 | newTreeEntry(VL, false, UserTreeIdx); | |||
1593 | DEBUG(dbgs() << "SLP: ShuffleVector are not vectorized.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: ShuffleVector are not vectorized.\n" ; } } while (false); | |||
1594 | return; | |||
1595 | } | |||
1596 | newTreeEntry(VL, true, UserTreeIdx); | |||
1597 | DEBUG(dbgs() << "SLP: added a ShuffleVector op.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: added a ShuffleVector op.\n" ; } } while (false); | |||
1598 | ||||
1599 | // Reorder operands if reordering would enable vectorization. | |||
1600 | if (isa<BinaryOperator>(VL0)) { | |||
1601 | ValueList Left, Right; | |||
1602 | reorderAltShuffleOperands(VL, Left, Right); | |||
1603 | buildTree_rec(Left, Depth + 1, UserTreeIdx); | |||
1604 | buildTree_rec(Right, Depth + 1, UserTreeIdx); | |||
1605 | return; | |||
1606 | } | |||
1607 | ||||
1608 | for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) { | |||
1609 | ValueList Operands; | |||
1610 | // Prepare the operand vector. | |||
1611 | for (Value *j : VL) | |||
1612 | Operands.push_back(cast<Instruction>(j)->getOperand(i)); | |||
1613 | ||||
1614 | buildTree_rec(Operands, Depth + 1, UserTreeIdx); | |||
1615 | } | |||
1616 | return; | |||
1617 | } | |||
1618 | default: | |||
1619 | BS.cancelScheduling(VL); | |||
1620 | newTreeEntry(VL, false, UserTreeIdx); | |||
1621 | DEBUG(dbgs() << "SLP: Gathering unknown instruction.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering unknown instruction.\n" ; } } while (false); | |||
1622 | return; | |||
1623 | } | |||
1624 | } | |||
1625 | ||||
1626 | unsigned BoUpSLP::canMapToVector(Type *T, const DataLayout &DL) const { | |||
1627 | unsigned N; | |||
1628 | Type *EltTy; | |||
1629 | auto *ST = dyn_cast<StructType>(T); | |||
1630 | if (ST) { | |||
1631 | N = ST->getNumElements(); | |||
1632 | EltTy = *ST->element_begin(); | |||
1633 | } else { | |||
1634 | N = cast<ArrayType>(T)->getNumElements(); | |||
1635 | EltTy = cast<ArrayType>(T)->getElementType(); | |||
1636 | } | |||
1637 | if (!isValidElementType(EltTy)) | |||
1638 | return 0; | |||
1639 | uint64_t VTSize = DL.getTypeStoreSizeInBits(VectorType::get(EltTy, N)); | |||
1640 | if (VTSize < MinVecRegSize || VTSize > MaxVecRegSize || VTSize != DL.getTypeStoreSizeInBits(T)) | |||
1641 | return 0; | |||
1642 | if (ST) { | |||
1643 | // Check that struct is homogeneous. | |||
1644 | for (const auto *Ty : ST->elements()) | |||
1645 | if (Ty != EltTy) | |||
1646 | return 0; | |||
1647 | } | |||
1648 | return N; | |||
1649 | } | |||
1650 | ||||
1651 | bool BoUpSLP::canReuseExtract(ArrayRef<Value *> VL, unsigned Opcode) const { | |||
1652 | assert(Opcode == Instruction::ExtractElement ||((Opcode == Instruction::ExtractElement || Opcode == Instruction ::ExtractValue) ? static_cast<void> (0) : __assert_fail ("Opcode == Instruction::ExtractElement || Opcode == Instruction::ExtractValue" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1653, __PRETTY_FUNCTION__)) | |||
1653 | Opcode == Instruction::ExtractValue)((Opcode == Instruction::ExtractElement || Opcode == Instruction ::ExtractValue) ? static_cast<void> (0) : __assert_fail ("Opcode == Instruction::ExtractElement || Opcode == Instruction::ExtractValue" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1653, __PRETTY_FUNCTION__)); | |||
1654 | assert(Opcode == getSameOpcode(VL) && "Invalid opcode")((Opcode == getSameOpcode(VL) && "Invalid opcode") ? static_cast <void> (0) : __assert_fail ("Opcode == getSameOpcode(VL) && \"Invalid opcode\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1654, __PRETTY_FUNCTION__)); | |||
1655 | // Check if all of the extracts come from the same vector and from the | |||
1656 | // correct offset. | |||
1657 | Value *VL0 = VL[0]; | |||
1658 | Instruction *E0 = cast<Instruction>(VL0); | |||
1659 | Value *Vec = E0->getOperand(0); | |||
1660 | ||||
1661 | // We have to extract from a vector/aggregate with the same number of elements. | |||
1662 | unsigned NElts; | |||
1663 | if (Opcode == Instruction::ExtractValue) { | |||
1664 | const DataLayout &DL = E0->getModule()->getDataLayout(); | |||
1665 | NElts = canMapToVector(Vec->getType(), DL); | |||
1666 | if (!NElts) | |||
1667 | return false; | |||
1668 | // Check if load can be rewritten as load of vector. | |||
1669 | LoadInst *LI = dyn_cast<LoadInst>(Vec); | |||
1670 | if (!LI || !LI->isSimple() || !LI->hasNUses(VL.size())) | |||
1671 | return false; | |||
1672 | } else { | |||
1673 | NElts = Vec->getType()->getVectorNumElements(); | |||
1674 | } | |||
1675 | ||||
1676 | if (NElts != VL.size()) | |||
1677 | return false; | |||
1678 | ||||
1679 | // Check that all of the indices extract from the correct offset. | |||
1680 | if (!matchExtractIndex(E0, 0, Opcode)) | |||
1681 | return false; | |||
1682 | ||||
1683 | for (unsigned i = 1, e = VL.size(); i < e; ++i) { | |||
1684 | Instruction *E = cast<Instruction>(VL[i]); | |||
1685 | if (!matchExtractIndex(E, i, Opcode)) | |||
1686 | return false; | |||
1687 | if (E->getOperand(0) != Vec) | |||
1688 | return false; | |||
1689 | } | |||
1690 | ||||
1691 | return true; | |||
1692 | } | |||
1693 | ||||
1694 | int BoUpSLP::getEntryCost(TreeEntry *E) { | |||
1695 | ArrayRef<Value*> VL = E->Scalars; | |||
1696 | ||||
1697 | Type *ScalarTy = VL[0]->getType(); | |||
1698 | if (StoreInst *SI = dyn_cast<StoreInst>(VL[0])) | |||
1699 | ScalarTy = SI->getValueOperand()->getType(); | |||
1700 | else if (CmpInst *CI = dyn_cast<CmpInst>(VL[0])) | |||
1701 | ScalarTy = CI->getOperand(0)->getType(); | |||
1702 | VectorType *VecTy = VectorType::get(ScalarTy, VL.size()); | |||
1703 | ||||
1704 | // If we have computed a smaller type for the expression, update VecTy so | |||
1705 | // that the costs will be accurate. | |||
1706 | if (MinBWs.count(VL[0])) | |||
1707 | VecTy = VectorType::get( | |||
1708 | IntegerType::get(F->getContext(), MinBWs[VL[0]].first), VL.size()); | |||
1709 | ||||
1710 | if (E->NeedToGather) { | |||
1711 | if (allConstant(VL)) | |||
1712 | return 0; | |||
1713 | if (isSplat(VL)) { | |||
1714 | return TTI->getShuffleCost(TargetTransformInfo::SK_Broadcast, VecTy, 0); | |||
1715 | } | |||
1716 | return getGatherCost(E->Scalars); | |||
1717 | } | |||
1718 | unsigned Opcode = getSameOpcode(VL); | |||
1719 | assert(Opcode && allSameType(VL) && allSameBlock(VL) && "Invalid VL")((Opcode && allSameType(VL) && allSameBlock(VL ) && "Invalid VL") ? static_cast<void> (0) : __assert_fail ("Opcode && allSameType(VL) && allSameBlock(VL) && \"Invalid VL\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1719, __PRETTY_FUNCTION__)); | |||
1720 | Instruction *VL0 = cast<Instruction>(VL[0]); | |||
1721 | switch (Opcode) { | |||
1722 | case Instruction::PHI: { | |||
1723 | return 0; | |||
1724 | } | |||
1725 | case Instruction::ExtractValue: | |||
1726 | case Instruction::ExtractElement: { | |||
1727 | if (canReuseExtract(VL, Opcode)) { | |||
1728 | int DeadCost = 0; | |||
1729 | for (unsigned i = 0, e = VL.size(); i < e; ++i) { | |||
1730 | Instruction *E = cast<Instruction>(VL[i]); | |||
1731 | // If all users are going to be vectorized, instruction can be | |||
1732 | // considered as dead. | |||
1733 | // The same, if have only one user, it will be vectorized for sure. | |||
1734 | if (E->hasOneUse() || | |||
1735 | std::all_of(E->user_begin(), E->user_end(), [this](User *U) { | |||
1736 | return ScalarToTreeEntry.count(U) > 0; | |||
1737 | })) | |||
1738 | // Take credit for instruction that will become dead. | |||
1739 | DeadCost += | |||
1740 | TTI->getVectorInstrCost(Instruction::ExtractElement, VecTy, i); | |||
1741 | } | |||
1742 | return -DeadCost; | |||
1743 | } | |||
1744 | return getGatherCost(VecTy); | |||
1745 | } | |||
1746 | case Instruction::ZExt: | |||
1747 | case Instruction::SExt: | |||
1748 | case Instruction::FPToUI: | |||
1749 | case Instruction::FPToSI: | |||
1750 | case Instruction::FPExt: | |||
1751 | case Instruction::PtrToInt: | |||
1752 | case Instruction::IntToPtr: | |||
1753 | case Instruction::SIToFP: | |||
1754 | case Instruction::UIToFP: | |||
1755 | case Instruction::Trunc: | |||
1756 | case Instruction::FPTrunc: | |||
1757 | case Instruction::BitCast: { | |||
1758 | Type *SrcTy = VL0->getOperand(0)->getType(); | |||
1759 | ||||
1760 | // Calculate the cost of this instruction. | |||
1761 | int ScalarCost = VL.size() * TTI->getCastInstrCost(VL0->getOpcode(), | |||
1762 | VL0->getType(), SrcTy, VL0); | |||
1763 | ||||
1764 | VectorType *SrcVecTy = VectorType::get(SrcTy, VL.size()); | |||
1765 | int VecCost = TTI->getCastInstrCost(VL0->getOpcode(), VecTy, SrcVecTy, VL0); | |||
1766 | return VecCost - ScalarCost; | |||
1767 | } | |||
1768 | case Instruction::FCmp: | |||
1769 | case Instruction::ICmp: | |||
1770 | case Instruction::Select: { | |||
1771 | // Calculate the cost of this instruction. | |||
1772 | VectorType *MaskTy = VectorType::get(Builder.getInt1Ty(), VL.size()); | |||
1773 | int ScalarCost = VecTy->getNumElements() * | |||
1774 | TTI->getCmpSelInstrCost(Opcode, ScalarTy, Builder.getInt1Ty(), VL0); | |||
1775 | int VecCost = TTI->getCmpSelInstrCost(Opcode, VecTy, MaskTy, VL0); | |||
1776 | return VecCost - ScalarCost; | |||
1777 | } | |||
1778 | case Instruction::Add: | |||
1779 | case Instruction::FAdd: | |||
1780 | case Instruction::Sub: | |||
1781 | case Instruction::FSub: | |||
1782 | case Instruction::Mul: | |||
1783 | case Instruction::FMul: | |||
1784 | case Instruction::UDiv: | |||
1785 | case Instruction::SDiv: | |||
1786 | case Instruction::FDiv: | |||
1787 | case Instruction::URem: | |||
1788 | case Instruction::SRem: | |||
1789 | case Instruction::FRem: | |||
1790 | case Instruction::Shl: | |||
1791 | case Instruction::LShr: | |||
1792 | case Instruction::AShr: | |||
1793 | case Instruction::And: | |||
1794 | case Instruction::Or: | |||
1795 | case Instruction::Xor: { | |||
1796 | // Certain instructions can be cheaper to vectorize if they have a | |||
1797 | // constant second vector operand. | |||
1798 | TargetTransformInfo::OperandValueKind Op1VK = | |||
1799 | TargetTransformInfo::OK_AnyValue; | |||
1800 | TargetTransformInfo::OperandValueKind Op2VK = | |||
1801 | TargetTransformInfo::OK_UniformConstantValue; | |||
1802 | TargetTransformInfo::OperandValueProperties Op1VP = | |||
1803 | TargetTransformInfo::OP_None; | |||
1804 | TargetTransformInfo::OperandValueProperties Op2VP = | |||
1805 | TargetTransformInfo::OP_None; | |||
1806 | ||||
1807 | // If all operands are exactly the same ConstantInt then set the | |||
1808 | // operand kind to OK_UniformConstantValue. | |||
1809 | // If instead not all operands are constants, then set the operand kind | |||
1810 | // to OK_AnyValue. If all operands are constants but not the same, | |||
1811 | // then set the operand kind to OK_NonUniformConstantValue. | |||
1812 | ConstantInt *CInt = nullptr; | |||
1813 | for (unsigned i = 0; i < VL.size(); ++i) { | |||
1814 | const Instruction *I = cast<Instruction>(VL[i]); | |||
1815 | if (!isa<ConstantInt>(I->getOperand(1))) { | |||
1816 | Op2VK = TargetTransformInfo::OK_AnyValue; | |||
1817 | break; | |||
1818 | } | |||
1819 | if (i == 0) { | |||
1820 | CInt = cast<ConstantInt>(I->getOperand(1)); | |||
1821 | continue; | |||
1822 | } | |||
1823 | if (Op2VK == TargetTransformInfo::OK_UniformConstantValue && | |||
1824 | CInt != cast<ConstantInt>(I->getOperand(1))) | |||
1825 | Op2VK = TargetTransformInfo::OK_NonUniformConstantValue; | |||
1826 | } | |||
1827 | // FIXME: Currently cost of model modification for division by power of | |||
1828 | // 2 is handled for X86 and AArch64. Add support for other targets. | |||
1829 | if (Op2VK == TargetTransformInfo::OK_UniformConstantValue && CInt && | |||
1830 | CInt->getValue().isPowerOf2()) | |||
1831 | Op2VP = TargetTransformInfo::OP_PowerOf2; | |||
1832 | ||||
1833 | SmallVector<const Value *, 4> Operands(VL0->operand_values()); | |||
1834 | int ScalarCost = | |||
1835 | VecTy->getNumElements() * | |||
1836 | TTI->getArithmeticInstrCost(Opcode, ScalarTy, Op1VK, Op2VK, Op1VP, | |||
1837 | Op2VP, Operands); | |||
1838 | int VecCost = TTI->getArithmeticInstrCost(Opcode, VecTy, Op1VK, Op2VK, | |||
1839 | Op1VP, Op2VP, Operands); | |||
1840 | return VecCost - ScalarCost; | |||
1841 | } | |||
1842 | case Instruction::GetElementPtr: { | |||
1843 | TargetTransformInfo::OperandValueKind Op1VK = | |||
1844 | TargetTransformInfo::OK_AnyValue; | |||
1845 | TargetTransformInfo::OperandValueKind Op2VK = | |||
1846 | TargetTransformInfo::OK_UniformConstantValue; | |||
1847 | ||||
1848 | int ScalarCost = | |||
1849 | VecTy->getNumElements() * | |||
1850 | TTI->getArithmeticInstrCost(Instruction::Add, ScalarTy, Op1VK, Op2VK); | |||
1851 | int VecCost = | |||
1852 | TTI->getArithmeticInstrCost(Instruction::Add, VecTy, Op1VK, Op2VK); | |||
1853 | ||||
1854 | return VecCost - ScalarCost; | |||
1855 | } | |||
1856 | case Instruction::Load: { | |||
1857 | // Cost of wide load - cost of scalar loads. | |||
1858 | unsigned alignment = dyn_cast<LoadInst>(VL0)->getAlignment(); | |||
1859 | int ScalarLdCost = VecTy->getNumElements() * | |||
1860 | TTI->getMemoryOpCost(Instruction::Load, ScalarTy, alignment, 0, VL0); | |||
1861 | int VecLdCost = TTI->getMemoryOpCost(Instruction::Load, | |||
1862 | VecTy, alignment, 0, VL0); | |||
1863 | return VecLdCost - ScalarLdCost; | |||
1864 | } | |||
1865 | case Instruction::Store: { | |||
1866 | // We know that we can merge the stores. Calculate the cost. | |||
1867 | unsigned alignment = dyn_cast<StoreInst>(VL0)->getAlignment(); | |||
1868 | int ScalarStCost = VecTy->getNumElements() * | |||
1869 | TTI->getMemoryOpCost(Instruction::Store, ScalarTy, alignment, 0, VL0); | |||
1870 | int VecStCost = TTI->getMemoryOpCost(Instruction::Store, | |||
1871 | VecTy, alignment, 0, VL0); | |||
1872 | return VecStCost - ScalarStCost; | |||
1873 | } | |||
1874 | case Instruction::Call: { | |||
1875 | CallInst *CI = cast<CallInst>(VL0); | |||
1876 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); | |||
1877 | ||||
1878 | // Calculate the cost of the scalar and vector calls. | |||
1879 | SmallVector<Type*, 4> ScalarTys; | |||
1880 | for (unsigned op = 0, opc = CI->getNumArgOperands(); op!= opc; ++op) | |||
1881 | ScalarTys.push_back(CI->getArgOperand(op)->getType()); | |||
1882 | ||||
1883 | FastMathFlags FMF; | |||
1884 | if (auto *FPMO = dyn_cast<FPMathOperator>(CI)) | |||
1885 | FMF = FPMO->getFastMathFlags(); | |||
1886 | ||||
1887 | int ScalarCallCost = VecTy->getNumElements() * | |||
1888 | TTI->getIntrinsicInstrCost(ID, ScalarTy, ScalarTys, FMF); | |||
1889 | ||||
1890 | SmallVector<Value *, 4> Args(CI->arg_operands()); | |||
1891 | int VecCallCost = TTI->getIntrinsicInstrCost(ID, CI->getType(), Args, FMF, | |||
1892 | VecTy->getNumElements()); | |||
1893 | ||||
1894 | 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) | |||
1895 | << " (" << VecCallCost << "-" << ScalarCallCost << ")"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Call cost "<< VecCallCost - ScalarCallCost << " (" << VecCallCost << "-" << ScalarCallCost << ")" << " for " << *CI << "\n"; } } while (false) | |||
1896 | << " for " << *CI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Call cost "<< VecCallCost - ScalarCallCost << " (" << VecCallCost << "-" << ScalarCallCost << ")" << " for " << *CI << "\n"; } } while (false); | |||
1897 | ||||
1898 | return VecCallCost - ScalarCallCost; | |||
1899 | } | |||
1900 | case Instruction::ShuffleVector: { | |||
1901 | TargetTransformInfo::OperandValueKind Op1VK = | |||
1902 | TargetTransformInfo::OK_AnyValue; | |||
1903 | TargetTransformInfo::OperandValueKind Op2VK = | |||
1904 | TargetTransformInfo::OK_AnyValue; | |||
1905 | int ScalarCost = 0; | |||
1906 | int VecCost = 0; | |||
1907 | for (Value *i : VL) { | |||
1908 | Instruction *I = cast<Instruction>(i); | |||
1909 | if (!I) | |||
1910 | break; | |||
1911 | ScalarCost += | |||
1912 | TTI->getArithmeticInstrCost(I->getOpcode(), ScalarTy, Op1VK, Op2VK); | |||
1913 | } | |||
1914 | // VecCost is equal to sum of the cost of creating 2 vectors | |||
1915 | // and the cost of creating shuffle. | |||
1916 | Instruction *I0 = cast<Instruction>(VL[0]); | |||
1917 | VecCost = | |||
1918 | TTI->getArithmeticInstrCost(I0->getOpcode(), VecTy, Op1VK, Op2VK); | |||
1919 | Instruction *I1 = cast<Instruction>(VL[1]); | |||
1920 | VecCost += | |||
1921 | TTI->getArithmeticInstrCost(I1->getOpcode(), VecTy, Op1VK, Op2VK); | |||
1922 | VecCost += | |||
1923 | TTI->getShuffleCost(TargetTransformInfo::SK_Alternate, VecTy, 0); | |||
1924 | return VecCost - ScalarCost; | |||
1925 | } | |||
1926 | default: | |||
1927 | llvm_unreachable("Unknown instruction")::llvm::llvm_unreachable_internal("Unknown instruction", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1927); | |||
1928 | } | |||
1929 | } | |||
1930 | ||||
1931 | bool BoUpSLP::isFullyVectorizableTinyTree() { | |||
1932 | 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) | |||
1933 | 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); | |||
1934 | ||||
1935 | // We only handle trees of heights 1 and 2. | |||
1936 | if (VectorizableTree.size() == 1 && !VectorizableTree[0].NeedToGather) | |||
1937 | return true; | |||
1938 | ||||
1939 | if (VectorizableTree.size() != 2) | |||
1940 | return false; | |||
1941 | ||||
1942 | // Handle splat and all-constants stores. | |||
1943 | if (!VectorizableTree[0].NeedToGather && | |||
1944 | (allConstant(VectorizableTree[1].Scalars) || | |||
1945 | isSplat(VectorizableTree[1].Scalars))) | |||
1946 | return true; | |||
1947 | ||||
1948 | // Gathering cost would be too much for tiny trees. | |||
1949 | if (VectorizableTree[0].NeedToGather || VectorizableTree[1].NeedToGather) | |||
1950 | return false; | |||
1951 | ||||
1952 | return true; | |||
1953 | } | |||
1954 | ||||
1955 | bool BoUpSLP::isTreeTinyAndNotFullyVectorizable() { | |||
1956 | ||||
1957 | // We can vectorize the tree if its size is greater than or equal to the | |||
1958 | // minimum size specified by the MinTreeSize command line option. | |||
1959 | if (VectorizableTree.size() >= MinTreeSize) | |||
1960 | return false; | |||
1961 | ||||
1962 | // If we have a tiny tree (a tree whose size is less than MinTreeSize), we | |||
1963 | // can vectorize it if we can prove it fully vectorizable. | |||
1964 | if (isFullyVectorizableTinyTree()) | |||
1965 | return false; | |||
1966 | ||||
1967 | assert(VectorizableTree.empty()((VectorizableTree.empty() ? ExternalUses.empty() : true && "We shouldn't have any external users") ? static_cast<void > (0) : __assert_fail ("VectorizableTree.empty() ? ExternalUses.empty() : true && \"We shouldn't have any external users\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1969, __PRETTY_FUNCTION__)) | |||
1968 | ? ExternalUses.empty()((VectorizableTree.empty() ? ExternalUses.empty() : true && "We shouldn't have any external users") ? static_cast<void > (0) : __assert_fail ("VectorizableTree.empty() ? ExternalUses.empty() : true && \"We shouldn't have any external users\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1969, __PRETTY_FUNCTION__)) | |||
1969 | : true && "We shouldn't have any external users")((VectorizableTree.empty() ? ExternalUses.empty() : true && "We shouldn't have any external users") ? static_cast<void > (0) : __assert_fail ("VectorizableTree.empty() ? ExternalUses.empty() : true && \"We shouldn't have any external users\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1969, __PRETTY_FUNCTION__)); | |||
1970 | ||||
1971 | // Otherwise, we can't vectorize the tree. It is both tiny and not fully | |||
1972 | // vectorizable. | |||
1973 | return true; | |||
1974 | } | |||
1975 | ||||
1976 | int BoUpSLP::getSpillCost() { | |||
1977 | // Walk from the bottom of the tree to the top, tracking which values are | |||
1978 | // live. When we see a call instruction that is not part of our tree, | |||
1979 | // query TTI to see if there is a cost to keeping values live over it | |||
1980 | // (for example, if spills and fills are required). | |||
1981 | unsigned BundleWidth = VectorizableTree.front().Scalars.size(); | |||
1982 | int Cost = 0; | |||
1983 | ||||
1984 | SmallPtrSet<Instruction*, 4> LiveValues; | |||
1985 | Instruction *PrevInst = nullptr; | |||
1986 | ||||
1987 | for (const auto &N : VectorizableTree) { | |||
1988 | Instruction *Inst = dyn_cast<Instruction>(N.Scalars[0]); | |||
1989 | if (!Inst) | |||
1990 | continue; | |||
1991 | ||||
1992 | if (!PrevInst) { | |||
1993 | PrevInst = Inst; | |||
1994 | continue; | |||
1995 | } | |||
1996 | ||||
1997 | // Update LiveValues. | |||
1998 | LiveValues.erase(PrevInst); | |||
1999 | for (auto &J : PrevInst->operands()) { | |||
2000 | if (isa<Instruction>(&*J) && ScalarToTreeEntry.count(&*J)) | |||
2001 | LiveValues.insert(cast<Instruction>(&*J)); | |||
2002 | } | |||
2003 | ||||
2004 | 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) | |||
2005 | 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) | |||
2006 | 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) | |||
2007 | 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) | |||
2008 | 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) | |||
2009 | 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) | |||
2010 | )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); | |||
2011 | ||||
2012 | // Now find the sequence of instructions between PrevInst and Inst. | |||
2013 | BasicBlock::reverse_iterator InstIt = ++Inst->getIterator().getReverse(), | |||
2014 | PrevInstIt = | |||
2015 | PrevInst->getIterator().getReverse(); | |||
2016 | while (InstIt != PrevInstIt) { | |||
2017 | if (PrevInstIt == PrevInst->getParent()->rend()) { | |||
2018 | PrevInstIt = Inst->getParent()->rbegin(); | |||
2019 | continue; | |||
2020 | } | |||
2021 | ||||
2022 | if (isa<CallInst>(&*PrevInstIt) && &*PrevInstIt != PrevInst) { | |||
2023 | SmallVector<Type*, 4> V; | |||
2024 | for (auto *II : LiveValues) | |||
2025 | V.push_back(VectorType::get(II->getType(), BundleWidth)); | |||
2026 | Cost += TTI->getCostOfKeepingLiveOverCall(V); | |||
2027 | } | |||
2028 | ||||
2029 | ++PrevInstIt; | |||
2030 | } | |||
2031 | ||||
2032 | PrevInst = Inst; | |||
2033 | } | |||
2034 | ||||
2035 | return Cost; | |||
2036 | } | |||
2037 | ||||
2038 | int BoUpSLP::getTreeCost() { | |||
2039 | int Cost = 0; | |||
2040 | 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) | |||
2041 | VectorizableTree.size() << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Calculating cost for tree of size " << VectorizableTree.size() << ".\n"; } } while ( false); | |||
2042 | ||||
2043 | unsigned BundleWidth = VectorizableTree[0].Scalars.size(); | |||
2044 | ||||
2045 | for (TreeEntry &TE : VectorizableTree) { | |||
2046 | int C = getEntryCost(&TE); | |||
2047 | 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) | |||
2048 | << *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); | |||
2049 | Cost += C; | |||
2050 | } | |||
2051 | ||||
2052 | SmallSet<Value *, 16> ExtractCostCalculated; | |||
2053 | int ExtractCost = 0; | |||
2054 | for (ExternalUser &EU : ExternalUses) { | |||
2055 | // We only add extract cost once for the same scalar. | |||
2056 | if (!ExtractCostCalculated.insert(EU.Scalar).second) | |||
2057 | continue; | |||
2058 | ||||
2059 | // Uses by ephemeral values are free (because the ephemeral value will be | |||
2060 | // removed prior to code generation, and so the extraction will be | |||
2061 | // removed as well). | |||
2062 | if (EphValues.count(EU.User)) | |||
2063 | continue; | |||
2064 | ||||
2065 | // If we plan to rewrite the tree in a smaller type, we will need to sign | |||
2066 | // extend the extracted value back to the original type. Here, we account | |||
2067 | // for the extract and the added cost of the sign extend if needed. | |||
2068 | auto *VecTy = VectorType::get(EU.Scalar->getType(), BundleWidth); | |||
2069 | auto *ScalarRoot = VectorizableTree[0].Scalars[0]; | |||
2070 | if (MinBWs.count(ScalarRoot)) { | |||
2071 | auto *MinTy = IntegerType::get(F->getContext(), MinBWs[ScalarRoot].first); | |||
2072 | auto Extend = | |||
2073 | MinBWs[ScalarRoot].second ? Instruction::SExt : Instruction::ZExt; | |||
2074 | VecTy = VectorType::get(MinTy, BundleWidth); | |||
2075 | ExtractCost += TTI->getExtractWithExtendCost(Extend, EU.Scalar->getType(), | |||
2076 | VecTy, EU.Lane); | |||
2077 | } else { | |||
2078 | ExtractCost += | |||
2079 | TTI->getVectorInstrCost(Instruction::ExtractElement, VecTy, EU.Lane); | |||
2080 | } | |||
2081 | } | |||
2082 | ||||
2083 | int SpillCost = getSpillCost(); | |||
2084 | Cost += SpillCost + ExtractCost; | |||
2085 | ||||
2086 | std::string Str; | |||
2087 | { | |||
2088 | raw_string_ostream OS(Str); | |||
2089 | OS << "SLP: Spill Cost = " << SpillCost << ".\n" | |||
2090 | << "SLP: Extract Cost = " << ExtractCost << ".\n" | |||
2091 | << "SLP: Total Cost = " << Cost << ".\n"; | |||
2092 | } | |||
2093 | DEBUG(dbgs() << Str)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << Str; } } while (false); | |||
2094 | ||||
2095 | if (ViewSLPTree) | |||
2096 | ViewGraph(this, "SLP" + F->getName(), false, Str); | |||
2097 | ||||
2098 | return Cost; | |||
2099 | } | |||
2100 | ||||
2101 | int BoUpSLP::getGatherCost(Type *Ty) { | |||
2102 | int Cost = 0; | |||
2103 | for (unsigned i = 0, e = cast<VectorType>(Ty)->getNumElements(); i < e; ++i) | |||
2104 | Cost += TTI->getVectorInstrCost(Instruction::InsertElement, Ty, i); | |||
2105 | return Cost; | |||
2106 | } | |||
2107 | ||||
2108 | int BoUpSLP::getGatherCost(ArrayRef<Value *> VL) { | |||
2109 | // Find the type of the operands in VL. | |||
2110 | Type *ScalarTy = VL[0]->getType(); | |||
2111 | if (StoreInst *SI = dyn_cast<StoreInst>(VL[0])) | |||
2112 | ScalarTy = SI->getValueOperand()->getType(); | |||
2113 | VectorType *VecTy = VectorType::get(ScalarTy, VL.size()); | |||
2114 | // Find the cost of inserting/extracting values from the vector. | |||
2115 | return getGatherCost(VecTy); | |||
2116 | } | |||
2117 | ||||
2118 | // Reorder commutative operations in alternate shuffle if the resulting vectors | |||
2119 | // are consecutive loads. This would allow us to vectorize the tree. | |||
2120 | // If we have something like- | |||
2121 | // load a[0] - load b[0] | |||
2122 | // load b[1] + load a[1] | |||
2123 | // load a[2] - load b[2] | |||
2124 | // load a[3] + load b[3] | |||
2125 | // Reordering the second load b[1] load a[1] would allow us to vectorize this | |||
2126 | // code. | |||
2127 | void BoUpSLP::reorderAltShuffleOperands(ArrayRef<Value *> VL, | |||
2128 | SmallVectorImpl<Value *> &Left, | |||
2129 | SmallVectorImpl<Value *> &Right) { | |||
2130 | // Push left and right operands of binary operation into Left and Right | |||
2131 | for (Value *i : VL) { | |||
2132 | Left.push_back(cast<Instruction>(i)->getOperand(0)); | |||
2133 | Right.push_back(cast<Instruction>(i)->getOperand(1)); | |||
2134 | } | |||
2135 | ||||
2136 | // Reorder if we have a commutative operation and consecutive access | |||
2137 | // are on either side of the alternate instructions. | |||
2138 | for (unsigned j = 0; j < VL.size() - 1; ++j) { | |||
2139 | if (LoadInst *L = dyn_cast<LoadInst>(Left[j])) { | |||
2140 | if (LoadInst *L1 = dyn_cast<LoadInst>(Right[j + 1])) { | |||
2141 | Instruction *VL1 = cast<Instruction>(VL[j]); | |||
2142 | Instruction *VL2 = cast<Instruction>(VL[j + 1]); | |||
2143 | if (VL1->isCommutative() && isConsecutiveAccess(L, L1, *DL, *SE)) { | |||
2144 | std::swap(Left[j], Right[j]); | |||
2145 | continue; | |||
2146 | } else if (VL2->isCommutative() && | |||
2147 | isConsecutiveAccess(L, L1, *DL, *SE)) { | |||
2148 | std::swap(Left[j + 1], Right[j + 1]); | |||
2149 | continue; | |||
2150 | } | |||
2151 | // else unchanged | |||
2152 | } | |||
2153 | } | |||
2154 | if (LoadInst *L = dyn_cast<LoadInst>(Right[j])) { | |||
2155 | if (LoadInst *L1 = dyn_cast<LoadInst>(Left[j + 1])) { | |||
2156 | Instruction *VL1 = cast<Instruction>(VL[j]); | |||
2157 | Instruction *VL2 = cast<Instruction>(VL[j + 1]); | |||
2158 | if (VL1->isCommutative() && isConsecutiveAccess(L, L1, *DL, *SE)) { | |||
2159 | std::swap(Left[j], Right[j]); | |||
2160 | continue; | |||
2161 | } else if (VL2->isCommutative() && | |||
2162 | isConsecutiveAccess(L, L1, *DL, *SE)) { | |||
2163 | std::swap(Left[j + 1], Right[j + 1]); | |||
2164 | continue; | |||
2165 | } | |||
2166 | // else unchanged | |||
2167 | } | |||
2168 | } | |||
2169 | } | |||
2170 | } | |||
2171 | ||||
2172 | // Return true if I should be commuted before adding it's left and right | |||
2173 | // operands to the arrays Left and Right. | |||
2174 | // | |||
2175 | // The vectorizer is trying to either have all elements one side being | |||
2176 | // instruction with the same opcode to enable further vectorization, or having | |||
2177 | // a splat to lower the vectorizing cost. | |||
2178 | static bool shouldReorderOperands(int i, Instruction &I, | |||
2179 | SmallVectorImpl<Value *> &Left, | |||
2180 | SmallVectorImpl<Value *> &Right, | |||
2181 | bool AllSameOpcodeLeft, | |||
2182 | bool AllSameOpcodeRight, bool SplatLeft, | |||
2183 | bool SplatRight) { | |||
2184 | Value *VLeft = I.getOperand(0); | |||
2185 | Value *VRight = I.getOperand(1); | |||
2186 | // If we have "SplatRight", try to see if commuting is needed to preserve it. | |||
2187 | if (SplatRight) { | |||
2188 | if (VRight == Right[i - 1]) | |||
2189 | // Preserve SplatRight | |||
2190 | return false; | |||
2191 | if (VLeft == Right[i - 1]) { | |||
2192 | // Commuting would preserve SplatRight, but we don't want to break | |||
2193 | // SplatLeft either, i.e. preserve the original order if possible. | |||
2194 | // (FIXME: why do we care?) | |||
2195 | if (SplatLeft && VLeft == Left[i - 1]) | |||
2196 | return false; | |||
2197 | return true; | |||
2198 | } | |||
2199 | } | |||
2200 | // Symmetrically handle Right side. | |||
2201 | if (SplatLeft) { | |||
2202 | if (VLeft == Left[i - 1]) | |||
2203 | // Preserve SplatLeft | |||
2204 | return false; | |||
2205 | if (VRight == Left[i - 1]) | |||
2206 | return true; | |||
2207 | } | |||
2208 | ||||
2209 | Instruction *ILeft = dyn_cast<Instruction>(VLeft); | |||
2210 | Instruction *IRight = dyn_cast<Instruction>(VRight); | |||
2211 | ||||
2212 | // If we have "AllSameOpcodeRight", try to see if the left operands preserves | |||
2213 | // it and not the right, in this case we want to commute. | |||
2214 | if (AllSameOpcodeRight) { | |||
2215 | unsigned RightPrevOpcode = cast<Instruction>(Right[i - 1])->getOpcode(); | |||
2216 | if (IRight && RightPrevOpcode == IRight->getOpcode()) | |||
2217 | // Do not commute, a match on the right preserves AllSameOpcodeRight | |||
2218 | return false; | |||
2219 | if (ILeft && RightPrevOpcode == ILeft->getOpcode()) { | |||
2220 | // We have a match and may want to commute, but first check if there is | |||
2221 | // not also a match on the existing operands on the Left to preserve | |||
2222 | // AllSameOpcodeLeft, i.e. preserve the original order if possible. | |||
2223 | // (FIXME: why do we care?) | |||
2224 | if (AllSameOpcodeLeft && ILeft && | |||
2225 | cast<Instruction>(Left[i - 1])->getOpcode() == ILeft->getOpcode()) | |||
2226 | return false; | |||
2227 | return true; | |||
2228 | } | |||
2229 | } | |||
2230 | // Symmetrically handle Left side. | |||
2231 | if (AllSameOpcodeLeft) { | |||
2232 | unsigned LeftPrevOpcode = cast<Instruction>(Left[i - 1])->getOpcode(); | |||
2233 | if (ILeft && LeftPrevOpcode == ILeft->getOpcode()) | |||
2234 | return false; | |||
2235 | if (IRight && LeftPrevOpcode == IRight->getOpcode()) | |||
2236 | return true; | |||
2237 | } | |||
2238 | return false; | |||
2239 | } | |||
2240 | ||||
2241 | void BoUpSLP::reorderInputsAccordingToOpcode(ArrayRef<Value *> VL, | |||
2242 | SmallVectorImpl<Value *> &Left, | |||
2243 | SmallVectorImpl<Value *> &Right) { | |||
2244 | ||||
2245 | if (VL.size()) { | |||
2246 | // Peel the first iteration out of the loop since there's nothing | |||
2247 | // interesting to do anyway and it simplifies the checks in the loop. | |||
2248 | auto VLeft = cast<Instruction>(VL[0])->getOperand(0); | |||
2249 | auto VRight = cast<Instruction>(VL[0])->getOperand(1); | |||
2250 | if (!isa<Instruction>(VRight) && isa<Instruction>(VLeft)) | |||
2251 | // Favor having instruction to the right. FIXME: why? | |||
2252 | std::swap(VLeft, VRight); | |||
2253 | Left.push_back(VLeft); | |||
2254 | Right.push_back(VRight); | |||
2255 | } | |||
2256 | ||||
2257 | // Keep track if we have instructions with all the same opcode on one side. | |||
2258 | bool AllSameOpcodeLeft = isa<Instruction>(Left[0]); | |||
2259 | bool AllSameOpcodeRight = isa<Instruction>(Right[0]); | |||
2260 | // Keep track if we have one side with all the same value (broadcast). | |||
2261 | bool SplatLeft = true; | |||
2262 | bool SplatRight = true; | |||
2263 | ||||
2264 | for (unsigned i = 1, e = VL.size(); i != e; ++i) { | |||
2265 | Instruction *I = cast<Instruction>(VL[i]); | |||
2266 | assert(I->isCommutative() && "Can only process commutative instruction")((I->isCommutative() && "Can only process commutative instruction" ) ? static_cast<void> (0) : __assert_fail ("I->isCommutative() && \"Can only process commutative instruction\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2266, __PRETTY_FUNCTION__)); | |||
2267 | // Commute to favor either a splat or maximizing having the same opcodes on | |||
2268 | // one side. | |||
2269 | if (shouldReorderOperands(i, *I, Left, Right, AllSameOpcodeLeft, | |||
2270 | AllSameOpcodeRight, SplatLeft, SplatRight)) { | |||
2271 | Left.push_back(I->getOperand(1)); | |||
2272 | Right.push_back(I->getOperand(0)); | |||
2273 | } else { | |||
2274 | Left.push_back(I->getOperand(0)); | |||
2275 | Right.push_back(I->getOperand(1)); | |||
2276 | } | |||
2277 | // Update Splat* and AllSameOpcode* after the insertion. | |||
2278 | SplatRight = SplatRight && (Right[i - 1] == Right[i]); | |||
2279 | SplatLeft = SplatLeft && (Left[i - 1] == Left[i]); | |||
2280 | AllSameOpcodeLeft = AllSameOpcodeLeft && isa<Instruction>(Left[i]) && | |||
2281 | (cast<Instruction>(Left[i - 1])->getOpcode() == | |||
2282 | cast<Instruction>(Left[i])->getOpcode()); | |||
2283 | AllSameOpcodeRight = AllSameOpcodeRight && isa<Instruction>(Right[i]) && | |||
2284 | (cast<Instruction>(Right[i - 1])->getOpcode() == | |||
2285 | cast<Instruction>(Right[i])->getOpcode()); | |||
2286 | } | |||
2287 | ||||
2288 | // If one operand end up being broadcast, return this operand order. | |||
2289 | if (SplatRight || SplatLeft) | |||
2290 | return; | |||
2291 | ||||
2292 | // Finally check if we can get longer vectorizable chain by reordering | |||
2293 | // without breaking the good operand order detected above. | |||
2294 | // E.g. If we have something like- | |||
2295 | // load a[0] load b[0] | |||
2296 | // load b[1] load a[1] | |||
2297 | // load a[2] load b[2] | |||
2298 | // load a[3] load b[3] | |||
2299 | // Reordering the second load b[1] load a[1] would allow us to vectorize | |||
2300 | // this code and we still retain AllSameOpcode property. | |||
2301 | // FIXME: This load reordering might break AllSameOpcode in some rare cases | |||
2302 | // such as- | |||
2303 | // add a[0],c[0] load b[0] | |||
2304 | // add a[1],c[2] load b[1] | |||
2305 | // b[2] load b[2] | |||
2306 | // add a[3],c[3] load b[3] | |||
2307 | for (unsigned j = 0; j < VL.size() - 1; ++j) { | |||
2308 | if (LoadInst *L = dyn_cast<LoadInst>(Left[j])) { | |||
2309 | if (LoadInst *L1 = dyn_cast<LoadInst>(Right[j + 1])) { | |||
2310 | if (isConsecutiveAccess(L, L1, *DL, *SE)) { | |||
2311 | std::swap(Left[j + 1], Right[j + 1]); | |||
2312 | continue; | |||
2313 | } | |||
2314 | } | |||
2315 | } | |||
2316 | if (LoadInst *L = dyn_cast<LoadInst>(Right[j])) { | |||
2317 | if (LoadInst *L1 = dyn_cast<LoadInst>(Left[j + 1])) { | |||
2318 | if (isConsecutiveAccess(L, L1, *DL, *SE)) { | |||
2319 | std::swap(Left[j + 1], Right[j + 1]); | |||
2320 | continue; | |||
2321 | } | |||
2322 | } | |||
2323 | } | |||
2324 | // else unchanged | |||
2325 | } | |||
2326 | } | |||
2327 | ||||
2328 | void BoUpSLP::setInsertPointAfterBundle(ArrayRef<Value *> VL) { | |||
2329 | ||||
2330 | // Get the basic block this bundle is in. All instructions in the bundle | |||
2331 | // should be in this block. | |||
2332 | auto *Front = cast<Instruction>(VL.front()); | |||
2333 | auto *BB = Front->getParent(); | |||
2334 | assert(all_of(make_range(VL.begin(), VL.end()), [&](Value *V) -> bool {((all_of(make_range(VL.begin(), VL.end()), [&](Value *V) -> bool { return cast<Instruction>(V)->getParent() == BB ; })) ? static_cast<void> (0) : __assert_fail ("all_of(make_range(VL.begin(), VL.end()), [&](Value *V) -> bool { return cast<Instruction>(V)->getParent() == BB; })" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2336, __PRETTY_FUNCTION__)) | |||
2335 | return cast<Instruction>(V)->getParent() == BB;((all_of(make_range(VL.begin(), VL.end()), [&](Value *V) -> bool { return cast<Instruction>(V)->getParent() == BB ; })) ? static_cast<void> (0) : __assert_fail ("all_of(make_range(VL.begin(), VL.end()), [&](Value *V) -> bool { return cast<Instruction>(V)->getParent() == BB; })" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2336, __PRETTY_FUNCTION__)) | |||
2336 | }))((all_of(make_range(VL.begin(), VL.end()), [&](Value *V) -> bool { return cast<Instruction>(V)->getParent() == BB ; })) ? static_cast<void> (0) : __assert_fail ("all_of(make_range(VL.begin(), VL.end()), [&](Value *V) -> bool { return cast<Instruction>(V)->getParent() == BB; })" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2336, __PRETTY_FUNCTION__)); | |||
2337 | ||||
2338 | // The last instruction in the bundle in program order. | |||
2339 | Instruction *LastInst = nullptr; | |||
2340 | ||||
2341 | // Find the last instruction. The common case should be that BB has been | |||
2342 | // scheduled, and the last instruction is VL.back(). So we start with | |||
2343 | // VL.back() and iterate over schedule data until we reach the end of the | |||
2344 | // bundle. The end of the bundle is marked by null ScheduleData. | |||
2345 | if (BlocksSchedules.count(BB)) { | |||
2346 | auto *Bundle = BlocksSchedules[BB]->getScheduleData(VL.back()); | |||
2347 | if (Bundle && Bundle->isPartOfBundle()) | |||
2348 | for (; Bundle; Bundle = Bundle->NextInBundle) | |||
2349 | LastInst = Bundle->Inst; | |||
2350 | } | |||
2351 | ||||
2352 | // LastInst can still be null at this point if there's either not an entry | |||
2353 | // for BB in BlocksSchedules or there's no ScheduleData available for | |||
2354 | // VL.back(). This can be the case if buildTree_rec aborts for various | |||
2355 | // reasons (e.g., the maximum recursion depth is reached, the maximum region | |||
2356 | // size is reached, etc.). ScheduleData is initialized in the scheduling | |||
2357 | // "dry-run". | |||
2358 | // | |||
2359 | // If this happens, we can still find the last instruction by brute force. We | |||
2360 | // iterate forwards from Front (inclusive) until we either see all | |||
2361 | // instructions in the bundle or reach the end of the block. If Front is the | |||
2362 | // last instruction in program order, LastInst will be set to Front, and we | |||
2363 | // will visit all the remaining instructions in the block. | |||
2364 | // | |||
2365 | // One of the reasons we exit early from buildTree_rec is to place an upper | |||
2366 | // bound on compile-time. Thus, taking an additional compile-time hit here is | |||
2367 | // not ideal. However, this should be exceedingly rare since it requires that | |||
2368 | // we both exit early from buildTree_rec and that the bundle be out-of-order | |||
2369 | // (causing us to iterate all the way to the end of the block). | |||
2370 | if (!LastInst) { | |||
2371 | SmallPtrSet<Value *, 16> Bundle(VL.begin(), VL.end()); | |||
2372 | for (auto &I : make_range(BasicBlock::iterator(Front), BB->end())) { | |||
2373 | if (Bundle.erase(&I)) | |||
2374 | LastInst = &I; | |||
2375 | if (Bundle.empty()) | |||
2376 | break; | |||
2377 | } | |||
2378 | } | |||
2379 | ||||
2380 | // Set the insertion point after the last instruction in the bundle. Set the | |||
2381 | // debug location to Front. | |||
2382 | Builder.SetInsertPoint(BB, ++LastInst->getIterator()); | |||
| ||||
2383 | Builder.SetCurrentDebugLocation(Front->getDebugLoc()); | |||
2384 | } | |||
2385 | ||||
2386 | Value *BoUpSLP::Gather(ArrayRef<Value *> VL, VectorType *Ty) { | |||
2387 | Value *Vec = UndefValue::get(Ty); | |||
2388 | // Generate the 'InsertElement' instruction. | |||
2389 | for (unsigned i = 0; i < Ty->getNumElements(); ++i) { | |||
2390 | Vec = Builder.CreateInsertElement(Vec, VL[i], Builder.getInt32(i)); | |||
2391 | if (Instruction *Insrt = dyn_cast<Instruction>(Vec)) { | |||
2392 | GatherSeq.insert(Insrt); | |||
2393 | CSEBlocks.insert(Insrt->getParent()); | |||
2394 | ||||
2395 | // Add to our 'need-to-extract' list. | |||
2396 | if (ScalarToTreeEntry.count(VL[i])) { | |||
2397 | int Idx = ScalarToTreeEntry[VL[i]]; | |||
2398 | TreeEntry *E = &VectorizableTree[Idx]; | |||
2399 | // Find which lane we need to extract. | |||
2400 | int FoundLane = -1; | |||
2401 | for (unsigned Lane = 0, LE = VL.size(); Lane != LE; ++Lane) { | |||
2402 | // Is this the lane of the scalar that we are looking for ? | |||
2403 | if (E->Scalars[Lane] == VL[i]) { | |||
2404 | FoundLane = Lane; | |||
2405 | break; | |||
2406 | } | |||
2407 | } | |||
2408 | assert(FoundLane >= 0 && "Could not find the correct lane")((FoundLane >= 0 && "Could not find the correct lane" ) ? static_cast<void> (0) : __assert_fail ("FoundLane >= 0 && \"Could not find the correct lane\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2408, __PRETTY_FUNCTION__)); | |||
2409 | ExternalUses.push_back(ExternalUser(VL[i], Insrt, FoundLane)); | |||
2410 | } | |||
2411 | } | |||
2412 | } | |||
2413 | ||||
2414 | return Vec; | |||
2415 | } | |||
2416 | ||||
2417 | Value *BoUpSLP::alreadyVectorized(ArrayRef<Value *> VL) const { | |||
2418 | SmallDenseMap<Value*, int>::const_iterator Entry | |||
2419 | = ScalarToTreeEntry.find(VL[0]); | |||
2420 | if (Entry != ScalarToTreeEntry.end()) { | |||
2421 | int Idx = Entry->second; | |||
2422 | const TreeEntry *En = &VectorizableTree[Idx]; | |||
2423 | if (En->isSame(VL) && En->VectorizedValue) | |||
2424 | return En->VectorizedValue; | |||
2425 | } | |||
2426 | return nullptr; | |||
2427 | } | |||
2428 | ||||
2429 | Value *BoUpSLP::vectorizeTree(ArrayRef<Value *> VL) { | |||
2430 | if (ScalarToTreeEntry.count(VL[0])) { | |||
2431 | int Idx = ScalarToTreeEntry[VL[0]]; | |||
2432 | TreeEntry *E = &VectorizableTree[Idx]; | |||
2433 | if (E->isSame(VL)) | |||
2434 | return vectorizeTree(E); | |||
2435 | } | |||
2436 | ||||
2437 | Type *ScalarTy = VL[0]->getType(); | |||
2438 | if (StoreInst *SI = dyn_cast<StoreInst>(VL[0])) | |||
2439 | ScalarTy = SI->getValueOperand()->getType(); | |||
2440 | VectorType *VecTy = VectorType::get(ScalarTy, VL.size()); | |||
2441 | ||||
2442 | return Gather(VL, VecTy); | |||
2443 | } | |||
2444 | ||||
2445 | Value *BoUpSLP::vectorizeTree(TreeEntry *E) { | |||
2446 | IRBuilder<>::InsertPointGuard Guard(Builder); | |||
2447 | ||||
2448 | if (E->VectorizedValue) { | |||
| ||||
2449 | 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); | |||
2450 | return E->VectorizedValue; | |||
2451 | } | |||
2452 | ||||
2453 | Instruction *VL0 = cast<Instruction>(E->Scalars[0]); | |||
2454 | Type *ScalarTy = VL0->getType(); | |||
2455 | if (StoreInst *SI = dyn_cast<StoreInst>(VL0)) | |||
2456 | ScalarTy = SI->getValueOperand()->getType(); | |||
2457 | VectorType *VecTy = VectorType::get(ScalarTy, E->Scalars.size()); | |||
2458 | ||||
2459 | if (E->NeedToGather) { | |||
2460 | setInsertPointAfterBundle(E->Scalars); | |||
2461 | auto *V = Gather(E->Scalars, VecTy); | |||
2462 | E->VectorizedValue = V; | |||
2463 | return V; | |||
2464 | } | |||
2465 | ||||
2466 | unsigned Opcode = getSameOpcode(E->Scalars); | |||
2467 | ||||
2468 | switch (Opcode) { | |||
2469 | case Instruction::PHI: { | |||
2470 | PHINode *PH = dyn_cast<PHINode>(VL0); | |||
2471 | Builder.SetInsertPoint(PH->getParent()->getFirstNonPHI()); | |||
2472 | Builder.SetCurrentDebugLocation(PH->getDebugLoc()); | |||
2473 | PHINode *NewPhi = Builder.CreatePHI(VecTy, PH->getNumIncomingValues()); | |||
2474 | E->VectorizedValue = NewPhi; | |||
2475 | ||||
2476 | // PHINodes may have multiple entries from the same block. We want to | |||
2477 | // visit every block once. | |||
2478 | SmallSet<BasicBlock*, 4> VisitedBBs; | |||
2479 | ||||
2480 | for (unsigned i = 0, e = PH->getNumIncomingValues(); i < e; ++i) { | |||
2481 | ValueList Operands; | |||
2482 | BasicBlock *IBB = PH->getIncomingBlock(i); | |||
2483 | ||||
2484 | if (!VisitedBBs.insert(IBB).second) { | |||
2485 | NewPhi->addIncoming(NewPhi->getIncomingValueForBlock(IBB), IBB); | |||
2486 | continue; | |||
2487 | } | |||
2488 | ||||
2489 | // Prepare the operand vector. | |||
2490 | for (Value *V : E->Scalars) | |||
2491 | Operands.push_back(cast<PHINode>(V)->getIncomingValueForBlock(IBB)); | |||
2492 | ||||
2493 | Builder.SetInsertPoint(IBB->getTerminator()); | |||
2494 | Builder.SetCurrentDebugLocation(PH->getDebugLoc()); | |||
2495 | Value *Vec = vectorizeTree(Operands); | |||
2496 | NewPhi->addIncoming(Vec, IBB); | |||
2497 | } | |||
2498 | ||||
2499 | assert(NewPhi->getNumIncomingValues() == PH->getNumIncomingValues() &&((NewPhi->getNumIncomingValues() == PH->getNumIncomingValues () && "Invalid number of incoming values") ? static_cast <void> (0) : __assert_fail ("NewPhi->getNumIncomingValues() == PH->getNumIncomingValues() && \"Invalid number of incoming values\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2500, __PRETTY_FUNCTION__)) | |||
2500 | "Invalid number of incoming values")((NewPhi->getNumIncomingValues() == PH->getNumIncomingValues () && "Invalid number of incoming values") ? static_cast <void> (0) : __assert_fail ("NewPhi->getNumIncomingValues() == PH->getNumIncomingValues() && \"Invalid number of incoming values\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2500, __PRETTY_FUNCTION__)); | |||
2501 | return NewPhi; | |||
2502 | } | |||
2503 | ||||
2504 | case Instruction::ExtractElement: { | |||
2505 | if (canReuseExtract(E->Scalars, Instruction::ExtractElement)) { | |||
2506 | Value *V = VL0->getOperand(0); | |||
2507 | E->VectorizedValue = V; | |||
2508 | return V; | |||
2509 | } | |||
2510 | setInsertPointAfterBundle(E->Scalars); | |||
2511 | auto *V = Gather(E->Scalars, VecTy); | |||
2512 | E->VectorizedValue = V; | |||
2513 | return V; | |||
2514 | } | |||
2515 | case Instruction::ExtractValue: { | |||
2516 | if (canReuseExtract(E->Scalars, Instruction::ExtractValue)) { | |||
2517 | LoadInst *LI = cast<LoadInst>(VL0->getOperand(0)); | |||
2518 | Builder.SetInsertPoint(LI); | |||
2519 | PointerType *PtrTy = PointerType::get(VecTy, LI->getPointerAddressSpace()); | |||
2520 | Value *Ptr = Builder.CreateBitCast(LI->getOperand(0), PtrTy); | |||
2521 | LoadInst *V = Builder.CreateAlignedLoad(Ptr, LI->getAlignment()); | |||
2522 | E->VectorizedValue = V; | |||
2523 | return propagateMetadata(V, E->Scalars); | |||
2524 | } | |||
2525 | setInsertPointAfterBundle(E->Scalars); | |||
2526 | auto *V = Gather(E->Scalars, VecTy); | |||
2527 | E->VectorizedValue = V; | |||
2528 | return V; | |||
2529 | } | |||
2530 | case Instruction::ZExt: | |||
2531 | case Instruction::SExt: | |||
2532 | case Instruction::FPToUI: | |||
2533 | case Instruction::FPToSI: | |||
2534 | case Instruction::FPExt: | |||
2535 | case Instruction::PtrToInt: | |||
2536 | case Instruction::IntToPtr: | |||
2537 | case Instruction::SIToFP: | |||
2538 | case Instruction::UIToFP: | |||
2539 | case Instruction::Trunc: | |||
2540 | case Instruction::FPTrunc: | |||
2541 | case Instruction::BitCast: { | |||
2542 | ValueList INVL; | |||
2543 | for (Value *V : E->Scalars) | |||
2544 | INVL.push_back(cast<Instruction>(V)->getOperand(0)); | |||
2545 | ||||
2546 | setInsertPointAfterBundle(E->Scalars); | |||
2547 | ||||
2548 | Value *InVec = vectorizeTree(INVL); | |||
2549 | ||||
2550 | if (Value *V = alreadyVectorized(E->Scalars)) | |||
2551 | return V; | |||
2552 | ||||
2553 | CastInst *CI = dyn_cast<CastInst>(VL0); | |||
2554 | Value *V = Builder.CreateCast(CI->getOpcode(), InVec, VecTy); | |||
2555 | E->VectorizedValue = V; | |||
2556 | ++NumVectorInstructions; | |||
2557 | return V; | |||
2558 | } | |||
2559 | case Instruction::FCmp: | |||
2560 | case Instruction::ICmp: { | |||
2561 | ValueList LHSV, RHSV; | |||
2562 | for (Value *V : E->Scalars) { | |||
2563 | LHSV.push_back(cast<Instruction>(V)->getOperand(0)); | |||
2564 | RHSV.push_back(cast<Instruction>(V)->getOperand(1)); | |||
2565 | } | |||
2566 | ||||
2567 | setInsertPointAfterBundle(E->Scalars); | |||
2568 | ||||
2569 | Value *L = vectorizeTree(LHSV); | |||
2570 | Value *R = vectorizeTree(RHSV); | |||
2571 | ||||
2572 | if (Value *V = alreadyVectorized(E->Scalars)) | |||
2573 | return V; | |||
2574 | ||||
2575 | CmpInst::Predicate P0 = cast<CmpInst>(VL0)->getPredicate(); | |||
2576 | Value *V; | |||
2577 | if (Opcode == Instruction::FCmp) | |||
2578 | V = Builder.CreateFCmp(P0, L, R); | |||
2579 | else | |||
2580 | V = Builder.CreateICmp(P0, L, R); | |||
2581 | ||||
2582 | E->VectorizedValue = V; | |||
2583 | propagateIRFlags(E->VectorizedValue, E->Scalars); | |||
2584 | ++NumVectorInstructions; | |||
2585 | return V; | |||
2586 | } | |||
2587 | case Instruction::Select: { | |||
2588 | ValueList TrueVec, FalseVec, CondVec; | |||
2589 | for (Value *V : E->Scalars) { | |||
2590 | CondVec.push_back(cast<Instruction>(V)->getOperand(0)); | |||
2591 | TrueVec.push_back(cast<Instruction>(V)->getOperand(1)); | |||
2592 | FalseVec.push_back(cast<Instruction>(V)->getOperand(2)); | |||
2593 | } | |||
2594 | ||||
2595 | setInsertPointAfterBundle(E->Scalars); | |||
2596 | ||||
2597 | Value *Cond = vectorizeTree(CondVec); | |||
2598 | Value *True = vectorizeTree(TrueVec); | |||
2599 | Value *False = vectorizeTree(FalseVec); | |||
2600 | ||||
2601 | if (Value *V = alreadyVectorized(E->Scalars)) | |||
2602 | return V; | |||
2603 | ||||
2604 | Value *V = Builder.CreateSelect(Cond, True, False); | |||
2605 | E->VectorizedValue = V; | |||
2606 | ++NumVectorInstructions; | |||
2607 | return V; | |||
2608 | } | |||
2609 | case Instruction::Add: | |||
2610 | case Instruction::FAdd: | |||
2611 | case Instruction::Sub: | |||
2612 | case Instruction::FSub: | |||
2613 | case Instruction::Mul: | |||
2614 | case Instruction::FMul: | |||
2615 | case Instruction::UDiv: | |||
2616 | case Instruction::SDiv: | |||
2617 | case Instruction::FDiv: | |||
2618 | case Instruction::URem: | |||
2619 | case Instruction::SRem: | |||
2620 | case Instruction::FRem: | |||
2621 | case Instruction::Shl: | |||
2622 | case Instruction::LShr: | |||
2623 | case Instruction::AShr: | |||
2624 | case Instruction::And: | |||
2625 | case Instruction::Or: | |||
2626 | case Instruction::Xor: { | |||
2627 | ValueList LHSVL, RHSVL; | |||
2628 | if (isa<BinaryOperator>(VL0) && VL0->isCommutative()) | |||
2629 | reorderInputsAccordingToOpcode(E->Scalars, LHSVL, RHSVL); | |||
2630 | else | |||
2631 | for (Value *V : E->Scalars) { | |||
2632 | LHSVL.push_back(cast<Instruction>(V)->getOperand(0)); | |||
2633 | RHSVL.push_back(cast<Instruction>(V)->getOperand(1)); | |||
2634 | } | |||
2635 | ||||
2636 | setInsertPointAfterBundle(E->Scalars); | |||
2637 | ||||
2638 | Value *LHS = vectorizeTree(LHSVL); | |||
2639 | Value *RHS = vectorizeTree(RHSVL); | |||
2640 | ||||
2641 | if (Value *V = alreadyVectorized(E->Scalars)) | |||
2642 | return V; | |||
2643 | ||||
2644 | BinaryOperator *BinOp = cast<BinaryOperator>(VL0); | |||
2645 | Value *V = Builder.CreateBinOp(BinOp->getOpcode(), LHS, RHS); | |||
2646 | E->VectorizedValue = V; | |||
2647 | propagateIRFlags(E->VectorizedValue, E->Scalars); | |||
2648 | ++NumVectorInstructions; | |||
2649 | ||||
2650 | if (Instruction *I = dyn_cast<Instruction>(V)) | |||
2651 | return propagateMetadata(I, E->Scalars); | |||
2652 | ||||
2653 | return V; | |||
2654 | } | |||
2655 | case Instruction::Load: { | |||
2656 | // Loads are inserted at the head of the tree because we don't want to | |||
2657 | // sink them all the way down past store instructions. | |||
2658 | setInsertPointAfterBundle(E->Scalars); | |||
2659 | ||||
2660 | LoadInst *LI = cast<LoadInst>(VL0); | |||
2661 | Type *ScalarLoadTy = LI->getType(); | |||
2662 | unsigned AS = LI->getPointerAddressSpace(); | |||
2663 | ||||
2664 | Value *VecPtr = Builder.CreateBitCast(LI->getPointerOperand(), | |||
2665 | VecTy->getPointerTo(AS)); | |||
2666 | ||||
2667 | // The pointer operand uses an in-tree scalar so we add the new BitCast to | |||
2668 | // ExternalUses list to make sure that an extract will be generated in the | |||
2669 | // future. | |||
2670 | if (ScalarToTreeEntry.count(LI->getPointerOperand())) | |||
2671 | ExternalUses.push_back( | |||
2672 | ExternalUser(LI->getPointerOperand(), cast<User>(VecPtr), 0)); | |||
2673 | ||||
2674 | unsigned Alignment = LI->getAlignment(); | |||
2675 | LI = Builder.CreateLoad(VecPtr); | |||
2676 | if (!Alignment) { | |||
2677 | Alignment = DL->getABITypeAlignment(ScalarLoadTy); | |||
2678 | } | |||
2679 | LI->setAlignment(Alignment); | |||
2680 | E->VectorizedValue = LI; | |||
2681 | ++NumVectorInstructions; | |||
2682 | return propagateMetadata(LI, E->Scalars); | |||
2683 | } | |||
2684 | case Instruction::Store: { | |||
2685 | StoreInst *SI = cast<StoreInst>(VL0); | |||
2686 | unsigned Alignment = SI->getAlignment(); | |||
2687 | unsigned AS = SI->getPointerAddressSpace(); | |||
2688 | ||||
2689 | ValueList ValueOp; | |||
2690 | for (Value *V : E->Scalars) | |||
2691 | ValueOp.push_back(cast<StoreInst>(V)->getValueOperand()); | |||
2692 | ||||
2693 | setInsertPointAfterBundle(E->Scalars); | |||
2694 | ||||
2695 | Value *VecValue = vectorizeTree(ValueOp); | |||
2696 | Value *VecPtr = Builder.CreateBitCast(SI->getPointerOperand(), | |||
2697 | VecTy->getPointerTo(AS)); | |||
2698 | StoreInst *S = Builder.CreateStore(VecValue, VecPtr); | |||
2699 | ||||
2700 | // The pointer operand uses an in-tree scalar so we add the new BitCast to | |||
2701 | // ExternalUses list to make sure that an extract will be generated in the | |||
2702 | // future. | |||
2703 | if (ScalarToTreeEntry.count(SI->getPointerOperand())) | |||
2704 | ExternalUses.push_back( | |||
2705 | ExternalUser(SI->getPointerOperand(), cast<User>(VecPtr), 0)); | |||
2706 | ||||
2707 | if (!Alignment) { | |||
2708 | Alignment = DL->getABITypeAlignment(SI->getValueOperand()->getType()); | |||
2709 | } | |||
2710 | S->setAlignment(Alignment); | |||
2711 | E->VectorizedValue = S; | |||
2712 | ++NumVectorInstructions; | |||
2713 | return propagateMetadata(S, E->Scalars); | |||
2714 | } | |||
2715 | case Instruction::GetElementPtr: { | |||
2716 | setInsertPointAfterBundle(E->Scalars); | |||
2717 | ||||
2718 | ValueList Op0VL; | |||
2719 | for (Value *V : E->Scalars) | |||
2720 | Op0VL.push_back(cast<GetElementPtrInst>(V)->getOperand(0)); | |||
2721 | ||||
2722 | Value *Op0 = vectorizeTree(Op0VL); | |||
2723 | ||||
2724 | std::vector<Value *> OpVecs; | |||
2725 | for (int j = 1, e = cast<GetElementPtrInst>(VL0)->getNumOperands(); j < e; | |||
2726 | ++j) { | |||
2727 | ValueList OpVL; | |||
2728 | for (Value *V : E->Scalars) | |||
2729 | OpVL.push_back(cast<GetElementPtrInst>(V)->getOperand(j)); | |||
2730 | ||||
2731 | Value *OpVec = vectorizeTree(OpVL); | |||
2732 | OpVecs.push_back(OpVec); | |||
2733 | } | |||
2734 | ||||
2735 | Value *V = Builder.CreateGEP( | |||
2736 | cast<GetElementPtrInst>(VL0)->getSourceElementType(), Op0, OpVecs); | |||
2737 | E->VectorizedValue = V; | |||
2738 | ++NumVectorInstructions; | |||
2739 | ||||
2740 | if (Instruction *I = dyn_cast<Instruction>(V)) | |||
2741 | return propagateMetadata(I, E->Scalars); | |||
2742 | ||||
2743 | return V; | |||
2744 | } | |||
2745 | case Instruction::Call: { | |||
2746 | CallInst *CI = cast<CallInst>(VL0); | |||
2747 | setInsertPointAfterBundle(E->Scalars); | |||
2748 | Function *FI; | |||
2749 | Intrinsic::ID IID = Intrinsic::not_intrinsic; | |||
2750 | Value *ScalarArg = nullptr; | |||
2751 | if (CI && (FI = CI->getCalledFunction())) { | |||
2752 | IID = FI->getIntrinsicID(); | |||
2753 | } | |||
2754 | std::vector<Value *> OpVecs; | |||
2755 | for (int j = 0, e = CI->getNumArgOperands(); j < e; ++j) { | |||
2756 | ValueList OpVL; | |||
2757 | // ctlz,cttz and powi are special intrinsics whose second argument is | |||
2758 | // a scalar. This argument should not be vectorized. | |||
2759 | if (hasVectorInstrinsicScalarOpd(IID, 1) && j == 1) { | |||
2760 | CallInst *CEI = cast<CallInst>(E->Scalars[0]); | |||
2761 | ScalarArg = CEI->getArgOperand(j); | |||
2762 | OpVecs.push_back(CEI->getArgOperand(j)); | |||
2763 | continue; | |||
2764 | } | |||
2765 | for (Value *V : E->Scalars) { | |||
2766 | CallInst *CEI = cast<CallInst>(V); | |||
2767 | OpVL.push_back(CEI->getArgOperand(j)); | |||
2768 | } | |||
2769 | ||||
2770 | Value *OpVec = vectorizeTree(OpVL); | |||
2771 | DEBUG(dbgs() << "SLP: OpVec[" << j << "]: " << *OpVec << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: OpVec[" << j << "]: " << *OpVec << "\n"; } } while (false); | |||
2772 | OpVecs.push_back(OpVec); | |||
2773 | } | |||
2774 | ||||
2775 | Module *M = F->getParent(); | |||
2776 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); | |||
2777 | Type *Tys[] = { VectorType::get(CI->getType(), E->Scalars.size()) }; | |||
2778 | Function *CF = Intrinsic::getDeclaration(M, ID, Tys); | |||
2779 | SmallVector<OperandBundleDef, 1> OpBundles; | |||
2780 | CI->getOperandBundlesAsDefs(OpBundles); | |||
2781 | Value *V = Builder.CreateCall(CF, OpVecs, OpBundles); | |||
2782 | ||||
2783 | // The scalar argument uses an in-tree scalar so we add the new vectorized | |||
2784 | // call to ExternalUses list to make sure that an extract will be | |||
2785 | // generated in the future. | |||
2786 | if (ScalarArg && ScalarToTreeEntry.count(ScalarArg)) | |||
2787 | ExternalUses.push_back(ExternalUser(ScalarArg, cast<User>(V), 0)); | |||
2788 | ||||
2789 | E->VectorizedValue = V; | |||
2790 | propagateIRFlags(E->VectorizedValue, E->Scalars); | |||
2791 | ++NumVectorInstructions; | |||
2792 | return V; | |||
2793 | } | |||
2794 | case Instruction::ShuffleVector: { | |||
2795 | ValueList LHSVL, RHSVL; | |||
2796 | assert(isa<BinaryOperator>(VL0) && "Invalid Shuffle Vector Operand")((isa<BinaryOperator>(VL0) && "Invalid Shuffle Vector Operand" ) ? static_cast<void> (0) : __assert_fail ("isa<BinaryOperator>(VL0) && \"Invalid Shuffle Vector Operand\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2796, __PRETTY_FUNCTION__)); | |||
2797 | reorderAltShuffleOperands(E->Scalars, LHSVL, RHSVL); | |||
2798 | setInsertPointAfterBundle(E->Scalars); | |||
2799 | ||||
2800 | Value *LHS = vectorizeTree(LHSVL); | |||
2801 | Value *RHS = vectorizeTree(RHSVL); | |||
2802 | ||||
2803 | if (Value *V = alreadyVectorized(E->Scalars)) | |||
2804 | return V; | |||
2805 | ||||
2806 | // Create a vector of LHS op1 RHS | |||
2807 | BinaryOperator *BinOp0 = cast<BinaryOperator>(VL0); | |||
2808 | Value *V0 = Builder.CreateBinOp(BinOp0->getOpcode(), LHS, RHS); | |||
2809 | ||||
2810 | // Create a vector of LHS op2 RHS | |||
2811 | Instruction *VL1 = cast<Instruction>(E->Scalars[1]); | |||
2812 | BinaryOperator *BinOp1 = cast<BinaryOperator>(VL1); | |||
2813 | Value *V1 = Builder.CreateBinOp(BinOp1->getOpcode(), LHS, RHS); | |||
2814 | ||||
2815 | // Create shuffle to take alternate operations from the vector. | |||
2816 | // Also, gather up odd and even scalar ops to propagate IR flags to | |||
2817 | // each vector operation. | |||
2818 | ValueList OddScalars, EvenScalars; | |||
2819 | unsigned e = E->Scalars.size(); | |||
2820 | SmallVector<Constant *, 8> Mask(e); | |||
2821 | for (unsigned i = 0; i < e; ++i) { | |||
2822 | if (i & 1) { | |||
2823 | Mask[i] = Builder.getInt32(e + i); | |||
2824 | OddScalars.push_back(E->Scalars[i]); | |||
2825 | } else { | |||
2826 | Mask[i] = Builder.getInt32(i); | |||
2827 | EvenScalars.push_back(E->Scalars[i]); | |||
2828 | } | |||
2829 | } | |||
2830 | ||||
2831 | Value *ShuffleMask = ConstantVector::get(Mask); | |||
2832 | propagateIRFlags(V0, EvenScalars); | |||
2833 | propagateIRFlags(V1, OddScalars); | |||
2834 | ||||
2835 | Value *V = Builder.CreateShuffleVector(V0, V1, ShuffleMask); | |||
2836 | E->VectorizedValue = V; | |||
2837 | ++NumVectorInstructions; | |||
2838 | if (Instruction *I = dyn_cast<Instruction>(V)) | |||
2839 | return propagateMetadata(I, E->Scalars); | |||
2840 | ||||
2841 | return V; | |||
2842 | } | |||
2843 | default: | |||
2844 | llvm_unreachable("unknown inst")::llvm::llvm_unreachable_internal("unknown inst", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2844); | |||
2845 | } | |||
2846 | return nullptr; | |||
2847 | } | |||
2848 | ||||
2849 | Value *BoUpSLP::vectorizeTree() { | |||
2850 | ExtraValueToDebugLocsMap ExternallyUsedValues; | |||
2851 | return vectorizeTree(ExternallyUsedValues); | |||
2852 | } | |||
2853 | ||||
2854 | Value * | |||
2855 | BoUpSLP::vectorizeTree(ExtraValueToDebugLocsMap &ExternallyUsedValues) { | |||
2856 | ||||
2857 | // All blocks must be scheduled before any instructions are inserted. | |||
2858 | for (auto &BSIter : BlocksSchedules) { | |||
2859 | scheduleBlock(BSIter.second.get()); | |||
2860 | } | |||
2861 | ||||
2862 | Builder.SetInsertPoint(&F->getEntryBlock().front()); | |||
2863 | auto *VectorRoot = vectorizeTree(&VectorizableTree[0]); | |||
2864 | ||||
2865 | // If the vectorized tree can be rewritten in a smaller type, we truncate the | |||
2866 | // vectorized root. InstCombine will then rewrite the entire expression. We | |||
2867 | // sign extend the extracted values below. | |||
2868 | auto *ScalarRoot = VectorizableTree[0].Scalars[0]; | |||
2869 | if (MinBWs.count(ScalarRoot)) { | |||
2870 | if (auto *I = dyn_cast<Instruction>(VectorRoot)) | |||
2871 | Builder.SetInsertPoint(&*++BasicBlock::iterator(I)); | |||
2872 | auto BundleWidth = VectorizableTree[0].Scalars.size(); | |||
2873 | auto *MinTy = IntegerType::get(F->getContext(), MinBWs[ScalarRoot].first); | |||
2874 | auto *VecTy = VectorType::get(MinTy, BundleWidth); | |||
2875 | auto *Trunc = Builder.CreateTrunc(VectorRoot, VecTy); | |||
2876 | VectorizableTree[0].VectorizedValue = Trunc; | |||
2877 | } | |||
2878 | ||||
2879 | DEBUG(dbgs() << "SLP: Extracting " << ExternalUses.size() << " values .\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Extracting " << ExternalUses .size() << " values .\n"; } } while (false); | |||
2880 | ||||
2881 | // If necessary, sign-extend or zero-extend ScalarRoot to the larger type | |||
2882 | // specified by ScalarType. | |||
2883 | auto extend = [&](Value *ScalarRoot, Value *Ex, Type *ScalarType) { | |||
2884 | if (!MinBWs.count(ScalarRoot)) | |||
2885 | return Ex; | |||
2886 | if (MinBWs[ScalarRoot].second) | |||
2887 | return Builder.CreateSExt(Ex, ScalarType); | |||
2888 | return Builder.CreateZExt(Ex, ScalarType); | |||
2889 | }; | |||
2890 | ||||
2891 | // Extract all of the elements with the external uses. | |||
2892 | for (const auto &ExternalUse : ExternalUses) { | |||
2893 | Value *Scalar = ExternalUse.Scalar; | |||
2894 | llvm::User *User = ExternalUse.User; | |||
2895 | ||||
2896 | // Skip users that we already RAUW. This happens when one instruction | |||
2897 | // has multiple uses of the same value. | |||
2898 | if (User && !is_contained(Scalar->users(), User)) | |||
2899 | continue; | |||
2900 | assert(ScalarToTreeEntry.count(Scalar) && "Invalid scalar")((ScalarToTreeEntry.count(Scalar) && "Invalid scalar" ) ? static_cast<void> (0) : __assert_fail ("ScalarToTreeEntry.count(Scalar) && \"Invalid scalar\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2900, __PRETTY_FUNCTION__)); | |||
2901 | ||||
2902 | int Idx = ScalarToTreeEntry[Scalar]; | |||
2903 | TreeEntry *E = &VectorizableTree[Idx]; | |||
2904 | assert(!E->NeedToGather && "Extracting from a gather list")((!E->NeedToGather && "Extracting from a gather list" ) ? static_cast<void> (0) : __assert_fail ("!E->NeedToGather && \"Extracting from a gather list\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2904, __PRETTY_FUNCTION__)); | |||
2905 | ||||
2906 | Value *Vec = E->VectorizedValue; | |||
2907 | assert(Vec && "Can't find vectorizable value")((Vec && "Can't find vectorizable value") ? static_cast <void> (0) : __assert_fail ("Vec && \"Can't find vectorizable value\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2907, __PRETTY_FUNCTION__)); | |||
2908 | ||||
2909 | Value *Lane = Builder.getInt32(ExternalUse.Lane); | |||
2910 | // If User == nullptr, the Scalar is used as extra arg. Generate | |||
2911 | // ExtractElement instruction and update the record for this scalar in | |||
2912 | // ExternallyUsedValues. | |||
2913 | if (!User) { | |||
2914 | assert(ExternallyUsedValues.count(Scalar) &&((ExternallyUsedValues.count(Scalar) && "Scalar with nullptr as an external user must be registered in " "ExternallyUsedValues map") ? static_cast<void> (0) : __assert_fail ("ExternallyUsedValues.count(Scalar) && \"Scalar with nullptr as an external user must be registered in \" \"ExternallyUsedValues map\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2916, __PRETTY_FUNCTION__)) | |||
2915 | "Scalar with nullptr as an external user must be registered in "((ExternallyUsedValues.count(Scalar) && "Scalar with nullptr as an external user must be registered in " "ExternallyUsedValues map") ? static_cast<void> (0) : __assert_fail ("ExternallyUsedValues.count(Scalar) && \"Scalar with nullptr as an external user must be registered in \" \"ExternallyUsedValues map\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2916, __PRETTY_FUNCTION__)) | |||
2916 | "ExternallyUsedValues map")((ExternallyUsedValues.count(Scalar) && "Scalar with nullptr as an external user must be registered in " "ExternallyUsedValues map") ? static_cast<void> (0) : __assert_fail ("ExternallyUsedValues.count(Scalar) && \"Scalar with nullptr as an external user must be registered in \" \"ExternallyUsedValues map\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2916, __PRETTY_FUNCTION__)); | |||
2917 | if (auto *VecI = dyn_cast<Instruction>(Vec)) { | |||
2918 | Builder.SetInsertPoint(VecI->getParent(), | |||
2919 | std::next(VecI->getIterator())); | |||
2920 | } else { | |||
2921 | Builder.SetInsertPoint(&F->getEntryBlock().front()); | |||
2922 | } | |||
2923 | Value *Ex = Builder.CreateExtractElement(Vec, Lane); | |||
2924 | Ex = extend(ScalarRoot, Ex, Scalar->getType()); | |||
2925 | CSEBlocks.insert(cast<Instruction>(Scalar)->getParent()); | |||
2926 | auto &Locs = ExternallyUsedValues[Scalar]; | |||
2927 | ExternallyUsedValues.insert({Ex, Locs}); | |||
2928 | ExternallyUsedValues.erase(Scalar); | |||
2929 | continue; | |||
2930 | } | |||
2931 | ||||
2932 | // Generate extracts for out-of-tree users. | |||
2933 | // Find the insertion point for the extractelement lane. | |||
2934 | if (auto *VecI = dyn_cast<Instruction>(Vec)) { | |||
2935 | if (PHINode *PH = dyn_cast<PHINode>(User)) { | |||
2936 | for (int i = 0, e = PH->getNumIncomingValues(); i != e; ++i) { | |||
2937 | if (PH->getIncomingValue(i) == Scalar) { | |||
2938 | TerminatorInst *IncomingTerminator = | |||
2939 | PH->getIncomingBlock(i)->getTerminator(); | |||
2940 | if (isa<CatchSwitchInst>(IncomingTerminator)) { | |||
2941 | Builder.SetInsertPoint(VecI->getParent(), | |||
2942 | std::next(VecI->getIterator())); | |||
2943 | } else { | |||
2944 | Builder.SetInsertPoint(PH->getIncomingBlock(i)->getTerminator()); | |||
2945 | } | |||
2946 | Value *Ex = Builder.CreateExtractElement(Vec, Lane); | |||
2947 | Ex = extend(ScalarRoot, Ex, Scalar->getType()); | |||
2948 | CSEBlocks.insert(PH->getIncomingBlock(i)); | |||
2949 | PH->setOperand(i, Ex); | |||
2950 | } | |||
2951 | } | |||
2952 | } else { | |||
2953 | Builder.SetInsertPoint(cast<Instruction>(User)); | |||
2954 | Value *Ex = Builder.CreateExtractElement(Vec, Lane); | |||
2955 | Ex = extend(ScalarRoot, Ex, Scalar->getType()); | |||
2956 | CSEBlocks.insert(cast<Instruction>(User)->getParent()); | |||
2957 | User->replaceUsesOfWith(Scalar, Ex); | |||
2958 | } | |||
2959 | } else { | |||
2960 | Builder.SetInsertPoint(&F->getEntryBlock().front()); | |||
2961 | Value *Ex = Builder.CreateExtractElement(Vec, Lane); | |||
2962 | Ex = extend(ScalarRoot, Ex, Scalar->getType()); | |||
2963 | CSEBlocks.insert(&F->getEntryBlock()); | |||
2964 | User->replaceUsesOfWith(Scalar, Ex); | |||
2965 | } | |||
2966 | ||||
2967 | DEBUG(dbgs() << "SLP: Replaced:" << *User << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Replaced:" << *User << ".\n"; } } while (false); | |||
2968 | } | |||
2969 | ||||
2970 | // For each vectorized value: | |||
2971 | for (TreeEntry &EIdx : VectorizableTree) { | |||
2972 | TreeEntry *Entry = &EIdx; | |||
2973 | ||||
2974 | // For each lane: | |||
2975 | for (int Lane = 0, LE = Entry->Scalars.size(); Lane != LE; ++Lane) { | |||
2976 | Value *Scalar = Entry->Scalars[Lane]; | |||
2977 | // No need to handle users of gathered values. | |||
2978 | if (Entry->NeedToGather) | |||
2979 | continue; | |||
2980 | ||||
2981 | assert(Entry->VectorizedValue && "Can't find vectorizable value")((Entry->VectorizedValue && "Can't find vectorizable value" ) ? static_cast<void> (0) : __assert_fail ("Entry->VectorizedValue && \"Can't find vectorizable value\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2981, __PRETTY_FUNCTION__)); | |||
2982 | ||||
2983 | Type *Ty = Scalar->getType(); | |||
2984 | if (!Ty->isVoidTy()) { | |||
2985 | #ifndef NDEBUG | |||
2986 | for (User *U : Scalar->users()) { | |||
2987 | DEBUG(dbgs() << "SLP: \tvalidating user:" << *U << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: \tvalidating user:" << *U << ".\n"; } } while (false); | |||
2988 | ||||
2989 | assert((ScalarToTreeEntry.count(U) ||(((ScalarToTreeEntry.count(U) || is_contained(UserIgnoreList, U)) && "Replacing out-of-tree value with undef") ? static_cast <void> (0) : __assert_fail ("(ScalarToTreeEntry.count(U) || is_contained(UserIgnoreList, U)) && \"Replacing out-of-tree value with undef\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2992, __PRETTY_FUNCTION__)) | |||
2990 | // It is legal to replace users in the ignorelist by undef.(((ScalarToTreeEntry.count(U) || is_contained(UserIgnoreList, U)) && "Replacing out-of-tree value with undef") ? static_cast <void> (0) : __assert_fail ("(ScalarToTreeEntry.count(U) || is_contained(UserIgnoreList, U)) && \"Replacing out-of-tree value with undef\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2992, __PRETTY_FUNCTION__)) | |||
2991 | is_contained(UserIgnoreList, U)) &&(((ScalarToTreeEntry.count(U) || is_contained(UserIgnoreList, U)) && "Replacing out-of-tree value with undef") ? static_cast <void> (0) : __assert_fail ("(ScalarToTreeEntry.count(U) || is_contained(UserIgnoreList, U)) && \"Replacing out-of-tree value with undef\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2992, __PRETTY_FUNCTION__)) | |||
2992 | "Replacing out-of-tree value with undef")(((ScalarToTreeEntry.count(U) || is_contained(UserIgnoreList, U)) && "Replacing out-of-tree value with undef") ? static_cast <void> (0) : __assert_fail ("(ScalarToTreeEntry.count(U) || is_contained(UserIgnoreList, U)) && \"Replacing out-of-tree value with undef\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2992, __PRETTY_FUNCTION__)); | |||
2993 | } | |||
2994 | #endif | |||
2995 | Value *Undef = UndefValue::get(Ty); | |||
2996 | Scalar->replaceAllUsesWith(Undef); | |||
2997 | } | |||
2998 | DEBUG(dbgs() << "SLP: \tErasing scalar:" << *Scalar << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: \tErasing scalar:" << * Scalar << ".\n"; } } while (false); | |||
2999 | eraseInstruction(cast<Instruction>(Scalar)); | |||
3000 | } | |||
3001 | } | |||
3002 | ||||
3003 | Builder.ClearInsertionPoint(); | |||
3004 | ||||
3005 | return VectorizableTree[0].VectorizedValue; | |||
3006 | } | |||
3007 | ||||
3008 | void BoUpSLP::optimizeGatherSequence() { | |||
3009 | DEBUG(dbgs() << "SLP: Optimizing " << GatherSeq.size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Optimizing " << GatherSeq .size() << " gather sequences instructions.\n"; } } while (false) | |||
3010 | << " gather sequences instructions.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Optimizing " << GatherSeq .size() << " gather sequences instructions.\n"; } } while (false); | |||
3011 | // LICM InsertElementInst sequences. | |||
3012 | for (Instruction *it : GatherSeq) { | |||
3013 | InsertElementInst *Insert = dyn_cast<InsertElementInst>(it); | |||
3014 | ||||
3015 | if (!Insert) | |||
3016 | continue; | |||
3017 | ||||
3018 | // Check if this block is inside a loop. | |||
3019 | Loop *L = LI->getLoopFor(Insert->getParent()); | |||
3020 | if (!L) | |||
3021 | continue; | |||
3022 | ||||
3023 | // Check if it has a preheader. | |||
3024 | BasicBlock *PreHeader = L->getLoopPreheader(); | |||
3025 | if (!PreHeader) | |||
3026 | continue; | |||
3027 | ||||
3028 | // If the vector or the element that we insert into it are | |||
3029 | // instructions that are defined in this basic block then we can't | |||
3030 | // hoist this instruction. | |||
3031 | Instruction *CurrVec = dyn_cast<Instruction>(Insert->getOperand(0)); | |||
3032 | Instruction *NewElem = dyn_cast<Instruction>(Insert->getOperand(1)); | |||
3033 | if (CurrVec && L->contains(CurrVec)) | |||
3034 | continue; | |||
3035 | if (NewElem && L->contains(NewElem)) | |||
3036 | continue; | |||
3037 | ||||
3038 | // We can hoist this instruction. Move it to the pre-header. | |||
3039 | Insert->moveBefore(PreHeader->getTerminator()); | |||
3040 | } | |||
3041 | ||||
3042 | // Make a list of all reachable blocks in our CSE queue. | |||
3043 | SmallVector<const DomTreeNode *, 8> CSEWorkList; | |||
3044 | CSEWorkList.reserve(CSEBlocks.size()); | |||
3045 | for (BasicBlock *BB : CSEBlocks) | |||
3046 | if (DomTreeNode *N = DT->getNode(BB)) { | |||
3047 | assert(DT->isReachableFromEntry(N))((DT->isReachableFromEntry(N)) ? static_cast<void> ( 0) : __assert_fail ("DT->isReachableFromEntry(N)", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3047, __PRETTY_FUNCTION__)); | |||
3048 | CSEWorkList.push_back(N); | |||
3049 | } | |||
3050 | ||||
3051 | // Sort blocks by domination. This ensures we visit a block after all blocks | |||
3052 | // dominating it are visited. | |||
3053 | std::stable_sort(CSEWorkList.begin(), CSEWorkList.end(), | |||
3054 | [this](const DomTreeNode *A, const DomTreeNode *B) { | |||
3055 | return DT->properlyDominates(A, B); | |||
3056 | }); | |||
3057 | ||||
3058 | // Perform O(N^2) search over the gather sequences and merge identical | |||
3059 | // instructions. TODO: We can further optimize this scan if we split the | |||
3060 | // instructions into different buckets based on the insert lane. | |||
3061 | SmallVector<Instruction *, 16> Visited; | |||
3062 | for (auto I = CSEWorkList.begin(), E = CSEWorkList.end(); I != E; ++I) { | |||
3063 | assert((I == CSEWorkList.begin() || !DT->dominates(*I, *std::prev(I))) &&(((I == CSEWorkList.begin() || !DT->dominates(*I, *std::prev (I))) && "Worklist not sorted properly!") ? static_cast <void> (0) : __assert_fail ("(I == CSEWorkList.begin() || !DT->dominates(*I, *std::prev(I))) && \"Worklist not sorted properly!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3064, __PRETTY_FUNCTION__)) | |||
3064 | "Worklist not sorted properly!")(((I == CSEWorkList.begin() || !DT->dominates(*I, *std::prev (I))) && "Worklist not sorted properly!") ? static_cast <void> (0) : __assert_fail ("(I == CSEWorkList.begin() || !DT->dominates(*I, *std::prev(I))) && \"Worklist not sorted properly!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3064, __PRETTY_FUNCTION__)); | |||
3065 | BasicBlock *BB = (*I)->getBlock(); | |||
3066 | // For all instructions in blocks containing gather sequences: | |||
3067 | for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e;) { | |||
3068 | Instruction *In = &*it++; | |||
3069 | if (!isa<InsertElementInst>(In) && !isa<ExtractElementInst>(In)) | |||
3070 | continue; | |||
3071 | ||||
3072 | // Check if we can replace this instruction with any of the | |||
3073 | // visited instructions. | |||
3074 | for (Instruction *v : Visited) { | |||
3075 | if (In->isIdenticalTo(v) && | |||
3076 | DT->dominates(v->getParent(), In->getParent())) { | |||
3077 | In->replaceAllUsesWith(v); | |||
3078 | eraseInstruction(In); | |||
3079 | In = nullptr; | |||
3080 | break; | |||
3081 | } | |||
3082 | } | |||
3083 | if (In) { | |||
3084 | assert(!is_contained(Visited, In))((!is_contained(Visited, In)) ? static_cast<void> (0) : __assert_fail ("!is_contained(Visited, In)", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3084, __PRETTY_FUNCTION__)); | |||
3085 | Visited.push_back(In); | |||
3086 | } | |||
3087 | } | |||
3088 | } | |||
3089 | CSEBlocks.clear(); | |||
3090 | GatherSeq.clear(); | |||
3091 | } | |||
3092 | ||||
3093 | // Groups the instructions to a bundle (which is then a single scheduling entity) | |||
3094 | // and schedules instructions until the bundle gets ready. | |||
3095 | bool BoUpSLP::BlockScheduling::tryScheduleBundle(ArrayRef<Value *> VL, | |||
3096 | BoUpSLP *SLP) { | |||
3097 | if (isa<PHINode>(VL[0])) | |||
3098 | return true; | |||
3099 | ||||
3100 | // Initialize the instruction bundle. | |||
3101 | Instruction *OldScheduleEnd = ScheduleEnd; | |||
3102 | ScheduleData *PrevInBundle = nullptr; | |||
3103 | ScheduleData *Bundle = nullptr; | |||
3104 | bool ReSchedule = false; | |||
3105 | DEBUG(dbgs() << "SLP: bundle: " << *VL[0] << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: bundle: " << *VL[0] << "\n"; } } while (false); | |||
3106 | ||||
3107 | // Make sure that the scheduling region contains all | |||
3108 | // instructions of the bundle. | |||
3109 | for (Value *V : VL) { | |||
3110 | if (!extendSchedulingRegion(V)) | |||
3111 | return false; | |||
3112 | } | |||
3113 | ||||
3114 | for (Value *V : VL) { | |||
3115 | ScheduleData *BundleMember = getScheduleData(V); | |||
3116 | assert(BundleMember &&((BundleMember && "no ScheduleData for bundle member (maybe not in same basic block)" ) ? static_cast<void> (0) : __assert_fail ("BundleMember && \"no ScheduleData for bundle member (maybe not in same basic block)\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3117, __PRETTY_FUNCTION__)) | |||
3117 | "no ScheduleData for bundle member (maybe not in same basic block)")((BundleMember && "no ScheduleData for bundle member (maybe not in same basic block)" ) ? static_cast<void> (0) : __assert_fail ("BundleMember && \"no ScheduleData for bundle member (maybe not in same basic block)\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3117, __PRETTY_FUNCTION__)); | |||
3118 | if (BundleMember->IsScheduled) { | |||
3119 | // A bundle member was scheduled as single instruction before and now | |||
3120 | // needs to be scheduled as part of the bundle. We just get rid of the | |||
3121 | // existing schedule. | |||
3122 | 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) | |||
3123 | << " was already scheduled\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: reset schedule because " << *BundleMember << " was already scheduled\n"; } } while (false); | |||
3124 | ReSchedule = true; | |||
3125 | } | |||
3126 | assert(BundleMember->isSchedulingEntity() &&((BundleMember->isSchedulingEntity() && "bundle member already part of other bundle" ) ? static_cast<void> (0) : __assert_fail ("BundleMember->isSchedulingEntity() && \"bundle member already part of other bundle\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3127, __PRETTY_FUNCTION__)) | |||
3127 | "bundle member already part of other bundle")((BundleMember->isSchedulingEntity() && "bundle member already part of other bundle" ) ? static_cast<void> (0) : __assert_fail ("BundleMember->isSchedulingEntity() && \"bundle member already part of other bundle\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3127, __PRETTY_FUNCTION__)); | |||
3128 | if (PrevInBundle) { | |||
3129 | PrevInBundle->NextInBundle = BundleMember; | |||
3130 | } else { | |||
3131 | Bundle = BundleMember; | |||
3132 | } | |||
3133 | BundleMember->UnscheduledDepsInBundle = 0; | |||
3134 | Bundle->UnscheduledDepsInBundle += BundleMember->UnscheduledDeps; | |||
3135 | ||||
3136 | // Group the instructions to a bundle. | |||
3137 | BundleMember->FirstInBundle = Bundle; | |||
3138 | PrevInBundle = BundleMember; | |||
3139 | } | |||
3140 | if (ScheduleEnd != OldScheduleEnd) { | |||
3141 | // The scheduling region got new instructions at the lower end (or it is a | |||
3142 | // new region for the first bundle). This makes it necessary to | |||
3143 | // recalculate all dependencies. | |||
3144 | // It is seldom that this needs to be done a second time after adding the | |||
3145 | // initial bundle to the region. | |||
3146 | for (auto *I = ScheduleStart; I != ScheduleEnd; I = I->getNextNode()) { | |||
3147 | ScheduleData *SD = getScheduleData(I); | |||
3148 | SD->clearDependencies(); | |||
3149 | } | |||
3150 | ReSchedule = true; | |||
3151 | } | |||
3152 | if (ReSchedule) { | |||
3153 | resetSchedule(); | |||
3154 | initialFillReadyList(ReadyInsts); | |||
3155 | } | |||
3156 | ||||
3157 | 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) | |||
3158 | << BB->getName() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: try schedule bundle " << *Bundle << " in block " << BB->getName() << "\n"; } } while (false); | |||
3159 | ||||
3160 | calculateDependencies(Bundle, true, SLP); | |||
3161 | ||||
3162 | // Now try to schedule the new bundle. As soon as the bundle is "ready" it | |||
3163 | // means that there are no cyclic dependencies and we can schedule it. | |||
3164 | // Note that's important that we don't "schedule" the bundle yet (see | |||
3165 | // cancelScheduling). | |||
3166 | while (!Bundle->isReady() && !ReadyInsts.empty()) { | |||
3167 | ||||
3168 | ScheduleData *pickedSD = ReadyInsts.back(); | |||
3169 | ReadyInsts.pop_back(); | |||
3170 | ||||
3171 | if (pickedSD->isSchedulingEntity() && pickedSD->isReady()) { | |||
3172 | schedule(pickedSD, ReadyInsts); | |||
3173 | } | |||
3174 | } | |||
3175 | if (!Bundle->isReady()) { | |||
3176 | cancelScheduling(VL); | |||
3177 | return false; | |||
3178 | } | |||
3179 | return true; | |||
3180 | } | |||
3181 | ||||
3182 | void BoUpSLP::BlockScheduling::cancelScheduling(ArrayRef<Value *> VL) { | |||
3183 | if (isa<PHINode>(VL[0])) | |||
3184 | return; | |||
3185 | ||||
3186 | ScheduleData *Bundle = getScheduleData(VL[0]); | |||
3187 | DEBUG(dbgs() << "SLP: cancel scheduling of " << *Bundle << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: cancel scheduling of " << *Bundle << "\n"; } } while (false); | |||
3188 | assert(!Bundle->IsScheduled &&((!Bundle->IsScheduled && "Can't cancel bundle which is already scheduled" ) ? static_cast<void> (0) : __assert_fail ("!Bundle->IsScheduled && \"Can't cancel bundle which is already scheduled\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3189, __PRETTY_FUNCTION__)) | |||
3189 | "Can't cancel bundle which is already scheduled")((!Bundle->IsScheduled && "Can't cancel bundle which is already scheduled" ) ? static_cast<void> (0) : __assert_fail ("!Bundle->IsScheduled && \"Can't cancel bundle which is already scheduled\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3189, __PRETTY_FUNCTION__)); | |||
3190 | assert(Bundle->isSchedulingEntity() && Bundle->isPartOfBundle() &&((Bundle->isSchedulingEntity() && Bundle->isPartOfBundle () && "tried to unbundle something which is not a bundle" ) ? static_cast<void> (0) : __assert_fail ("Bundle->isSchedulingEntity() && Bundle->isPartOfBundle() && \"tried to unbundle something which is not a bundle\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3191, __PRETTY_FUNCTION__)) | |||
3191 | "tried to unbundle something which is not a bundle")((Bundle->isSchedulingEntity() && Bundle->isPartOfBundle () && "tried to unbundle something which is not a bundle" ) ? static_cast<void> (0) : __assert_fail ("Bundle->isSchedulingEntity() && Bundle->isPartOfBundle() && \"tried to unbundle something which is not a bundle\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3191, __PRETTY_FUNCTION__)); | |||
3192 | ||||
3193 | // Un-bundle: make single instructions out of the bundle. | |||
3194 | ScheduleData *BundleMember = Bundle; | |||
3195 | while (BundleMember) { | |||
3196 | assert(BundleMember->FirstInBundle == Bundle && "corrupt bundle links")((BundleMember->FirstInBundle == Bundle && "corrupt bundle links" ) ? static_cast<void> (0) : __assert_fail ("BundleMember->FirstInBundle == Bundle && \"corrupt bundle links\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3196, __PRETTY_FUNCTION__)); | |||
3197 | BundleMember->FirstInBundle = BundleMember; | |||
3198 | ScheduleData *Next = BundleMember->NextInBundle; | |||
3199 | BundleMember->NextInBundle = nullptr; | |||
3200 | BundleMember->UnscheduledDepsInBundle = BundleMember->UnscheduledDeps; | |||
3201 | if (BundleMember->UnscheduledDepsInBundle == 0) { | |||
3202 | ReadyInsts.insert(BundleMember); | |||
3203 | } | |||
3204 | BundleMember = Next; | |||
3205 | } | |||
3206 | } | |||
3207 | ||||
3208 | bool BoUpSLP::BlockScheduling::extendSchedulingRegion(Value *V) { | |||
3209 | if (getScheduleData(V)) | |||
3210 | return true; | |||
3211 | Instruction *I = dyn_cast<Instruction>(V); | |||
3212 | assert(I && "bundle member must be an instruction")((I && "bundle member must be an instruction") ? static_cast <void> (0) : __assert_fail ("I && \"bundle member must be an instruction\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3212, __PRETTY_FUNCTION__)); | |||
3213 | assert(!isa<PHINode>(I) && "phi nodes don't need to be scheduled")((!isa<PHINode>(I) && "phi nodes don't need to be scheduled" ) ? static_cast<void> (0) : __assert_fail ("!isa<PHINode>(I) && \"phi nodes don't need to be scheduled\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3213, __PRETTY_FUNCTION__)); | |||
3214 | if (!ScheduleStart) { | |||
3215 | // It's the first instruction in the new region. | |||
3216 | initScheduleData(I, I->getNextNode(), nullptr, nullptr); | |||
3217 | ScheduleStart = I; | |||
3218 | ScheduleEnd = I->getNextNode(); | |||
3219 | assert(ScheduleEnd && "tried to vectorize a TerminatorInst?")((ScheduleEnd && "tried to vectorize a TerminatorInst?" ) ? static_cast<void> (0) : __assert_fail ("ScheduleEnd && \"tried to vectorize a TerminatorInst?\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3219, __PRETTY_FUNCTION__)); | |||
3220 | 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); | |||
3221 | return true; | |||
3222 | } | |||
3223 | // Search up and down at the same time, because we don't know if the new | |||
3224 | // instruction is above or below the existing scheduling region. | |||
3225 | BasicBlock::reverse_iterator UpIter = | |||
3226 | ++ScheduleStart->getIterator().getReverse(); | |||
3227 | BasicBlock::reverse_iterator UpperEnd = BB->rend(); | |||
3228 | BasicBlock::iterator DownIter = ScheduleEnd->getIterator(); | |||
3229 | BasicBlock::iterator LowerEnd = BB->end(); | |||
3230 | for (;;) { | |||
3231 | if (++ScheduleRegionSize > ScheduleRegionSizeLimit) { | |||
3232 | 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); | |||
3233 | return false; | |||
3234 | } | |||
3235 | ||||
3236 | if (UpIter != UpperEnd) { | |||
3237 | if (&*UpIter == I) { | |||
3238 | initScheduleData(I, ScheduleStart, nullptr, FirstLoadStoreInRegion); | |||
3239 | ScheduleStart = I; | |||
3240 | 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); | |||
3241 | return true; | |||
3242 | } | |||
3243 | UpIter++; | |||
3244 | } | |||
3245 | if (DownIter != LowerEnd) { | |||
3246 | if (&*DownIter == I) { | |||
3247 | initScheduleData(ScheduleEnd, I->getNextNode(), LastLoadStoreInRegion, | |||
3248 | nullptr); | |||
3249 | ScheduleEnd = I->getNextNode(); | |||
3250 | assert(ScheduleEnd && "tried to vectorize a TerminatorInst?")((ScheduleEnd && "tried to vectorize a TerminatorInst?" ) ? static_cast<void> (0) : __assert_fail ("ScheduleEnd && \"tried to vectorize a TerminatorInst?\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3250, __PRETTY_FUNCTION__)); | |||
3251 | 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); | |||
3252 | return true; | |||
3253 | } | |||
3254 | DownIter++; | |||
3255 | } | |||
3256 | assert((UpIter != UpperEnd || DownIter != LowerEnd) &&(((UpIter != UpperEnd || DownIter != LowerEnd) && "instruction not found in block" ) ? static_cast<void> (0) : __assert_fail ("(UpIter != UpperEnd || DownIter != LowerEnd) && \"instruction not found in block\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3257, __PRETTY_FUNCTION__)) | |||
3257 | "instruction not found in block")(((UpIter != UpperEnd || DownIter != LowerEnd) && "instruction not found in block" ) ? static_cast<void> (0) : __assert_fail ("(UpIter != UpperEnd || DownIter != LowerEnd) && \"instruction not found in block\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3257, __PRETTY_FUNCTION__)); | |||
3258 | } | |||
3259 | return true; | |||
3260 | } | |||
3261 | ||||
3262 | void BoUpSLP::BlockScheduling::initScheduleData(Instruction *FromI, | |||
3263 | Instruction *ToI, | |||
3264 | ScheduleData *PrevLoadStore, | |||
3265 | ScheduleData *NextLoadStore) { | |||
3266 | ScheduleData *CurrentLoadStore = PrevLoadStore; | |||
3267 | for (Instruction *I = FromI; I != ToI; I = I->getNextNode()) { | |||
3268 | ScheduleData *SD = ScheduleDataMap[I]; | |||
3269 | if (!SD) { | |||
3270 | // Allocate a new ScheduleData for the instruction. | |||
3271 | if (ChunkPos >= ChunkSize) { | |||
3272 | ScheduleDataChunks.push_back( | |||
3273 | llvm::make_unique<ScheduleData[]>(ChunkSize)); | |||
3274 | ChunkPos = 0; | |||
3275 | } | |||
3276 | SD = &(ScheduleDataChunks.back()[ChunkPos++]); | |||
3277 | ScheduleDataMap[I] = SD; | |||
3278 | SD->Inst = I; | |||
3279 | } | |||
3280 | assert(!isInSchedulingRegion(SD) &&((!isInSchedulingRegion(SD) && "new ScheduleData already in scheduling region" ) ? static_cast<void> (0) : __assert_fail ("!isInSchedulingRegion(SD) && \"new ScheduleData already in scheduling region\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3281, __PRETTY_FUNCTION__)) | |||
3281 | "new ScheduleData already in scheduling region")((!isInSchedulingRegion(SD) && "new ScheduleData already in scheduling region" ) ? static_cast<void> (0) : __assert_fail ("!isInSchedulingRegion(SD) && \"new ScheduleData already in scheduling region\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3281, __PRETTY_FUNCTION__)); | |||
3282 | SD->init(SchedulingRegionID); | |||
3283 | ||||
3284 | if (I->mayReadOrWriteMemory()) { | |||
3285 | // Update the linked list of memory accessing instructions. | |||
3286 | if (CurrentLoadStore) { | |||
3287 | CurrentLoadStore->NextLoadStore = SD; | |||
3288 | } else { | |||
3289 | FirstLoadStoreInRegion = SD; | |||
3290 | } | |||
3291 | CurrentLoadStore = SD; | |||
3292 | } | |||
3293 | } | |||
3294 | if (NextLoadStore) { | |||
3295 | if (CurrentLoadStore) | |||
3296 | CurrentLoadStore->NextLoadStore = NextLoadStore; | |||
3297 | } else { | |||
3298 | LastLoadStoreInRegion = CurrentLoadStore; | |||
3299 | } | |||
3300 | } | |||
3301 | ||||
3302 | void BoUpSLP::BlockScheduling::calculateDependencies(ScheduleData *SD, | |||
3303 | bool InsertInReadyList, | |||
3304 | BoUpSLP *SLP) { | |||
3305 | assert(SD->isSchedulingEntity())((SD->isSchedulingEntity()) ? static_cast<void> (0) : __assert_fail ("SD->isSchedulingEntity()", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3305, __PRETTY_FUNCTION__)); | |||
3306 | ||||
3307 | SmallVector<ScheduleData *, 10> WorkList; | |||
3308 | WorkList.push_back(SD); | |||
3309 | ||||
3310 | while (!WorkList.empty()) { | |||
3311 | ScheduleData *SD = WorkList.back(); | |||
3312 | WorkList.pop_back(); | |||
3313 | ||||
3314 | ScheduleData *BundleMember = SD; | |||
3315 | while (BundleMember) { | |||
3316 | assert(isInSchedulingRegion(BundleMember))((isInSchedulingRegion(BundleMember)) ? static_cast<void> (0) : __assert_fail ("isInSchedulingRegion(BundleMember)", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3316, __PRETTY_FUNCTION__)); | |||
3317 | if (!BundleMember->hasValidDependencies()) { | |||
3318 | ||||
3319 | DEBUG(dbgs() << "SLP: update deps of " << *BundleMember << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: update deps of " << *BundleMember << "\n"; } } while (false); | |||
3320 | BundleMember->Dependencies = 0; | |||
3321 | BundleMember->resetUnscheduledDeps(); | |||
3322 | ||||
3323 | // Handle def-use chain dependencies. | |||
3324 | for (User *U : BundleMember->Inst->users()) { | |||
3325 | if (isa<Instruction>(U)) { | |||
3326 | ScheduleData *UseSD = getScheduleData(U); | |||
3327 | if (UseSD && isInSchedulingRegion(UseSD->FirstInBundle)) { | |||
3328 | BundleMember->Dependencies++; | |||
3329 | ScheduleData *DestBundle = UseSD->FirstInBundle; | |||
3330 | if (!DestBundle->IsScheduled) | |||
3331 | BundleMember->incrementUnscheduledDeps(1); | |||
3332 | if (!DestBundle->hasValidDependencies()) | |||
3333 | WorkList.push_back(DestBundle); | |||
3334 | } | |||
3335 | } else { | |||
3336 | // I'm not sure if this can ever happen. But we need to be safe. | |||
3337 | // This lets the instruction/bundle never be scheduled and | |||
3338 | // eventually disable vectorization. | |||
3339 | BundleMember->Dependencies++; | |||
3340 | BundleMember->incrementUnscheduledDeps(1); | |||
3341 | } | |||
3342 | } | |||
3343 | ||||
3344 | // Handle the memory dependencies. | |||
3345 | ScheduleData *DepDest = BundleMember->NextLoadStore; | |||
3346 | if (DepDest) { | |||
3347 | Instruction *SrcInst = BundleMember->Inst; | |||
3348 | MemoryLocation SrcLoc = getLocation(SrcInst, SLP->AA); | |||
3349 | bool SrcMayWrite = BundleMember->Inst->mayWriteToMemory(); | |||
3350 | unsigned numAliased = 0; | |||
3351 | unsigned DistToSrc = 1; | |||
3352 | ||||
3353 | while (DepDest) { | |||
3354 | assert(isInSchedulingRegion(DepDest))((isInSchedulingRegion(DepDest)) ? static_cast<void> (0 ) : __assert_fail ("isInSchedulingRegion(DepDest)", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3354, __PRETTY_FUNCTION__)); | |||
3355 | ||||
3356 | // We have two limits to reduce the complexity: | |||
3357 | // 1) AliasedCheckLimit: It's a small limit to reduce calls to | |||
3358 | // SLP->isAliased (which is the expensive part in this loop). | |||
3359 | // 2) MaxMemDepDistance: It's for very large blocks and it aborts | |||
3360 | // the whole loop (even if the loop is fast, it's quadratic). | |||
3361 | // It's important for the loop break condition (see below) to | |||
3362 | // check this limit even between two read-only instructions. | |||
3363 | if (DistToSrc >= MaxMemDepDistance || | |||
3364 | ((SrcMayWrite || DepDest->Inst->mayWriteToMemory()) && | |||
3365 | (numAliased >= AliasedCheckLimit || | |||
3366 | SLP->isAliased(SrcLoc, SrcInst, DepDest->Inst)))) { | |||
3367 | ||||
3368 | // We increment the counter only if the locations are aliased | |||
3369 | // (instead of counting all alias checks). This gives a better | |||
3370 | // balance between reduced runtime and accurate dependencies. | |||
3371 | numAliased++; | |||
3372 | ||||
3373 | DepDest->MemoryDependencies.push_back(BundleMember); | |||
3374 | BundleMember->Dependencies++; | |||
3375 | ScheduleData *DestBundle = DepDest->FirstInBundle; | |||
3376 | if (!DestBundle->IsScheduled) { | |||
3377 | BundleMember->incrementUnscheduledDeps(1); | |||
3378 | } | |||
3379 | if (!DestBundle->hasValidDependencies()) { | |||
3380 | WorkList.push_back(DestBundle); | |||
3381 | } | |||
3382 | } | |||
3383 | DepDest = DepDest->NextLoadStore; | |||
3384 | ||||
3385 | // Example, explaining the loop break condition: Let's assume our | |||
3386 | // starting instruction is i0 and MaxMemDepDistance = 3. | |||
3387 | // | |||
3388 | // +--------v--v--v | |||
3389 | // i0,i1,i2,i3,i4,i5,i6,i7,i8 | |||
3390 | // +--------^--^--^ | |||
3391 | // | |||
3392 | // MaxMemDepDistance let us stop alias-checking at i3 and we add | |||
3393 | // dependencies from i0 to i3,i4,.. (even if they are not aliased). | |||
3394 | // Previously we already added dependencies from i3 to i6,i7,i8 | |||
3395 | // (because of MaxMemDepDistance). As we added a dependency from | |||
3396 | // i0 to i3, we have transitive dependencies from i0 to i6,i7,i8 | |||
3397 | // and we can abort this loop at i6. | |||
3398 | if (DistToSrc >= 2 * MaxMemDepDistance) | |||
3399 | break; | |||
3400 | DistToSrc++; | |||
3401 | } | |||
3402 | } | |||
3403 | } | |||
3404 | BundleMember = BundleMember->NextInBundle; | |||
3405 | } | |||
3406 | if (InsertInReadyList && SD->isReady()) { | |||
3407 | ReadyInsts.push_back(SD); | |||
3408 | 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); | |||
3409 | } | |||
3410 | } | |||
3411 | } | |||
3412 | ||||
3413 | void BoUpSLP::BlockScheduling::resetSchedule() { | |||
3414 | assert(ScheduleStart &&((ScheduleStart && "tried to reset schedule on block which has not been scheduled" ) ? static_cast<void> (0) : __assert_fail ("ScheduleStart && \"tried to reset schedule on block which has not been scheduled\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3415, __PRETTY_FUNCTION__)) | |||
3415 | "tried to reset schedule on block which has not been scheduled")((ScheduleStart && "tried to reset schedule on block which has not been scheduled" ) ? static_cast<void> (0) : __assert_fail ("ScheduleStart && \"tried to reset schedule on block which has not been scheduled\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3415, __PRETTY_FUNCTION__)); | |||
3416 | for (Instruction *I = ScheduleStart; I != ScheduleEnd; I = I->getNextNode()) { | |||
3417 | ScheduleData *SD = getScheduleData(I); | |||
3418 | assert(isInSchedulingRegion(SD))((isInSchedulingRegion(SD)) ? static_cast<void> (0) : __assert_fail ("isInSchedulingRegion(SD)", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3418, __PRETTY_FUNCTION__)); | |||
3419 | SD->IsScheduled = false; | |||
3420 | SD->resetUnscheduledDeps(); | |||
3421 | } | |||
3422 | ReadyInsts.clear(); | |||
3423 | } | |||
3424 | ||||
3425 | void BoUpSLP::scheduleBlock(BlockScheduling *BS) { | |||
3426 | ||||
3427 | if (!BS->ScheduleStart) | |||
3428 | return; | |||
3429 | ||||
3430 | 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); | |||
3431 | ||||
3432 | BS->resetSchedule(); | |||
3433 | ||||
3434 | // For the real scheduling we use a more sophisticated ready-list: it is | |||
3435 | // sorted by the original instruction location. This lets the final schedule | |||
3436 | // be as close as possible to the original instruction order. | |||
3437 | struct ScheduleDataCompare { | |||
3438 | bool operator()(ScheduleData *SD1, ScheduleData *SD2) const { | |||
3439 | return SD2->SchedulingPriority < SD1->SchedulingPriority; | |||
3440 | } | |||
3441 | }; | |||
3442 | std::set<ScheduleData *, ScheduleDataCompare> ReadyInsts; | |||
3443 | ||||
3444 | // Ensure that all dependency data is updated and fill the ready-list with | |||
3445 | // initial instructions. | |||
3446 | int Idx = 0; | |||
3447 | int NumToSchedule = 0; | |||
3448 | for (auto *I = BS->ScheduleStart; I != BS->ScheduleEnd; | |||
3449 | I = I->getNextNode()) { | |||
3450 | ScheduleData *SD = BS->getScheduleData(I); | |||
3451 | assert(((SD->isPartOfBundle() == (ScalarToTreeEntry.count(SD-> Inst) != 0) && "scheduler and vectorizer have different opinion on what is a bundle" ) ? static_cast<void> (0) : __assert_fail ("SD->isPartOfBundle() == (ScalarToTreeEntry.count(SD->Inst) != 0) && \"scheduler and vectorizer have different opinion on what is a bundle\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3453, __PRETTY_FUNCTION__)) | |||
3452 | SD->isPartOfBundle() == (ScalarToTreeEntry.count(SD->Inst) != 0) &&((SD->isPartOfBundle() == (ScalarToTreeEntry.count(SD-> Inst) != 0) && "scheduler and vectorizer have different opinion on what is a bundle" ) ? static_cast<void> (0) : __assert_fail ("SD->isPartOfBundle() == (ScalarToTreeEntry.count(SD->Inst) != 0) && \"scheduler and vectorizer have different opinion on what is a bundle\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3453, __PRETTY_FUNCTION__)) | |||
3453 | "scheduler and vectorizer have different opinion on what is a bundle")((SD->isPartOfBundle() == (ScalarToTreeEntry.count(SD-> Inst) != 0) && "scheduler and vectorizer have different opinion on what is a bundle" ) ? static_cast<void> (0) : __assert_fail ("SD->isPartOfBundle() == (ScalarToTreeEntry.count(SD->Inst) != 0) && \"scheduler and vectorizer have different opinion on what is a bundle\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3453, __PRETTY_FUNCTION__)); | |||
3454 | SD->FirstInBundle->SchedulingPriority = Idx++; | |||
3455 | if (SD->isSchedulingEntity()) { | |||
3456 | BS->calculateDependencies(SD, false, this); | |||
3457 | NumToSchedule++; | |||
3458 | } | |||
3459 | } | |||
3460 | BS->initialFillReadyList(ReadyInsts); | |||
3461 | ||||
3462 | Instruction *LastScheduledInst = BS->ScheduleEnd; | |||
3463 | ||||
3464 | // Do the "real" scheduling. | |||
3465 | while (!ReadyInsts.empty()) { | |||
3466 | ScheduleData *picked = *ReadyInsts.begin(); | |||
3467 | ReadyInsts.erase(ReadyInsts.begin()); | |||
3468 | ||||
3469 | // Move the scheduled instruction(s) to their dedicated places, if not | |||
3470 | // there yet. | |||
3471 | ScheduleData *BundleMember = picked; | |||
3472 | while (BundleMember) { | |||
3473 | Instruction *pickedInst = BundleMember->Inst; | |||
3474 | if (LastScheduledInst->getNextNode() != pickedInst) { | |||
3475 | BS->BB->getInstList().remove(pickedInst); | |||
3476 | BS->BB->getInstList().insert(LastScheduledInst->getIterator(), | |||
3477 | pickedInst); | |||
3478 | } | |||
3479 | LastScheduledInst = pickedInst; | |||
3480 | BundleMember = BundleMember->NextInBundle; | |||
3481 | } | |||
3482 | ||||
3483 | BS->schedule(picked, ReadyInsts); | |||
3484 | NumToSchedule--; | |||
3485 | } | |||
3486 | assert(NumToSchedule == 0 && "could not schedule all instructions")((NumToSchedule == 0 && "could not schedule all instructions" ) ? static_cast<void> (0) : __assert_fail ("NumToSchedule == 0 && \"could not schedule all instructions\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3486, __PRETTY_FUNCTION__)); | |||
3487 | ||||
3488 | // Avoid duplicate scheduling of the block. | |||
3489 | BS->ScheduleStart = nullptr; | |||
3490 | } | |||
3491 | ||||
3492 | unsigned BoUpSLP::getVectorElementSize(Value *V) { | |||
3493 | // If V is a store, just return the width of the stored value without | |||
3494 | // traversing the expression tree. This is the common case. | |||
3495 | if (auto *Store = dyn_cast<StoreInst>(V)) | |||
3496 | return DL->getTypeSizeInBits(Store->getValueOperand()->getType()); | |||
3497 | ||||
3498 | // If V is not a store, we can traverse the expression tree to find loads | |||
3499 | // that feed it. The type of the loaded value may indicate a more suitable | |||
3500 | // width than V's type. We want to base the vector element size on the width | |||
3501 | // of memory operations where possible. | |||
3502 | SmallVector<Instruction *, 16> Worklist; | |||
3503 | SmallPtrSet<Instruction *, 16> Visited; | |||
3504 | if (auto *I = dyn_cast<Instruction>(V)) | |||
3505 | Worklist.push_back(I); | |||
3506 | ||||
3507 | // Traverse the expression tree in bottom-up order looking for loads. If we | |||
3508 | // encounter an instruciton we don't yet handle, we give up. | |||
3509 | auto MaxWidth = 0u; | |||
3510 | auto FoundUnknownInst = false; | |||
3511 | while (!Worklist.empty() && !FoundUnknownInst) { | |||
3512 | auto *I = Worklist.pop_back_val(); | |||
3513 | Visited.insert(I); | |||
3514 | ||||
3515 | // We should only be looking at scalar instructions here. If the current | |||
3516 | // instruction has a vector type, give up. | |||
3517 | auto *Ty = I->getType(); | |||
3518 | if (isa<VectorType>(Ty)) | |||
3519 | FoundUnknownInst = true; | |||
3520 | ||||
3521 | // If the current instruction is a load, update MaxWidth to reflect the | |||
3522 | // width of the loaded value. | |||
3523 | else if (isa<LoadInst>(I)) | |||
3524 | MaxWidth = std::max<unsigned>(MaxWidth, DL->getTypeSizeInBits(Ty)); | |||
3525 | ||||
3526 | // Otherwise, we need to visit the operands of the instruction. We only | |||
3527 | // handle the interesting cases from buildTree here. If an operand is an | |||
3528 | // instruction we haven't yet visited, we add it to the worklist. | |||
3529 | else if (isa<PHINode>(I) || isa<CastInst>(I) || isa<GetElementPtrInst>(I) || | |||
3530 | isa<CmpInst>(I) || isa<SelectInst>(I) || isa<BinaryOperator>(I)) { | |||
3531 | for (Use &U : I->operands()) | |||
3532 | if (auto *J = dyn_cast<Instruction>(U.get())) | |||
3533 | if (!Visited.count(J)) | |||
3534 | Worklist.push_back(J); | |||
3535 | } | |||
3536 | ||||
3537 | // If we don't yet handle the instruction, give up. | |||
3538 | else | |||
3539 | FoundUnknownInst = true; | |||
3540 | } | |||
3541 | ||||
3542 | // If we didn't encounter a memory access in the expression tree, or if we | |||
3543 | // gave up for some reason, just return the width of V. | |||
3544 | if (!MaxWidth || FoundUnknownInst) | |||
3545 | return DL->getTypeSizeInBits(V->getType()); | |||
3546 | ||||
3547 | // Otherwise, return the maximum width we found. | |||
3548 | return MaxWidth; | |||
3549 | } | |||
3550 | ||||
3551 | // Determine if a value V in a vectorizable expression Expr can be demoted to a | |||
3552 | // smaller type with a truncation. We collect the values that will be demoted | |||
3553 | // in ToDemote and additional roots that require investigating in Roots. | |||
3554 | static bool collectValuesToDemote(Value *V, SmallPtrSetImpl<Value *> &Expr, | |||
3555 | SmallVectorImpl<Value *> &ToDemote, | |||
3556 | SmallVectorImpl<Value *> &Roots) { | |||
3557 | ||||
3558 | // We can always demote constants. | |||
3559 | if (isa<Constant>(V)) { | |||
3560 | ToDemote.push_back(V); | |||
3561 | return true; | |||
3562 | } | |||
3563 | ||||
3564 | // If the value is not an instruction in the expression with only one use, it | |||
3565 | // cannot be demoted. | |||
3566 | auto *I = dyn_cast<Instruction>(V); | |||
3567 | if (!I || !I->hasOneUse() || !Expr.count(I)) | |||
3568 | return false; | |||
3569 | ||||
3570 | switch (I->getOpcode()) { | |||
3571 | ||||
3572 | // We can always demote truncations and extensions. Since truncations can | |||
3573 | // seed additional demotion, we save the truncated value. | |||
3574 | case Instruction::Trunc: | |||
3575 | Roots.push_back(I->getOperand(0)); | |||
3576 | case Instruction::ZExt: | |||
3577 | case Instruction::SExt: | |||
3578 | break; | |||
3579 | ||||
3580 | // We can demote certain binary operations if we can demote both of their | |||
3581 | // operands. | |||
3582 | case Instruction::Add: | |||
3583 | case Instruction::Sub: | |||
3584 | case Instruction::Mul: | |||
3585 | case Instruction::And: | |||
3586 | case Instruction::Or: | |||
3587 | case Instruction::Xor: | |||
3588 | if (!collectValuesToDemote(I->getOperand(0), Expr, ToDemote, Roots) || | |||
3589 | !collectValuesToDemote(I->getOperand(1), Expr, ToDemote, Roots)) | |||
3590 | return false; | |||
3591 | break; | |||
3592 | ||||
3593 | // We can demote selects if we can demote their true and false values. | |||
3594 | case Instruction::Select: { | |||
3595 | SelectInst *SI = cast<SelectInst>(I); | |||
3596 | if (!collectValuesToDemote(SI->getTrueValue(), Expr, ToDemote, Roots) || | |||
3597 | !collectValuesToDemote(SI->getFalseValue(), Expr, ToDemote, Roots)) | |||
3598 | return false; | |||
3599 | break; | |||
3600 | } | |||
3601 | ||||
3602 | // We can demote phis if we can demote all their incoming operands. Note that | |||
3603 | // we don't need to worry about cycles since we ensure single use above. | |||
3604 | case Instruction::PHI: { | |||
3605 | PHINode *PN = cast<PHINode>(I); | |||
3606 | for (Value *IncValue : PN->incoming_values()) | |||
3607 | if (!collectValuesToDemote(IncValue, Expr, ToDemote, Roots)) | |||
3608 | return false; | |||
3609 | break; | |||
3610 | } | |||
3611 | ||||
3612 | // Otherwise, conservatively give up. | |||
3613 | default: | |||
3614 | return false; | |||
3615 | } | |||
3616 | ||||
3617 | // Record the value that we can demote. | |||
3618 | ToDemote.push_back(V); | |||
3619 | return true; | |||
3620 | } | |||
3621 | ||||
3622 | void BoUpSLP::computeMinimumValueSizes() { | |||
3623 | // If there are no external uses, the expression tree must be rooted by a | |||
3624 | // store. We can't demote in-memory values, so there is nothing to do here. | |||
3625 | if (ExternalUses.empty()) | |||
3626 | return; | |||
3627 | ||||
3628 | // We only attempt to truncate integer expressions. | |||
3629 | auto &TreeRoot = VectorizableTree[0].Scalars; | |||
3630 | auto *TreeRootIT = dyn_cast<IntegerType>(TreeRoot[0]->getType()); | |||
3631 | if (!TreeRootIT) | |||
3632 | return; | |||
3633 | ||||
3634 | // If the expression is not rooted by a store, these roots should have | |||
3635 | // external uses. We will rely on InstCombine to rewrite the expression in | |||
3636 | // the narrower type. However, InstCombine only rewrites single-use values. | |||
3637 | // This means that if a tree entry other than a root is used externally, it | |||
3638 | // must have multiple uses and InstCombine will not rewrite it. The code | |||
3639 | // below ensures that only the roots are used externally. | |||
3640 | SmallPtrSet<Value *, 32> Expr(TreeRoot.begin(), TreeRoot.end()); | |||
3641 | for (auto &EU : ExternalUses) | |||
3642 | if (!Expr.erase(EU.Scalar)) | |||
3643 | return; | |||
3644 | if (!Expr.empty()) | |||
3645 | return; | |||
3646 | ||||
3647 | // Collect the scalar values of the vectorizable expression. We will use this | |||
3648 | // context to determine which values can be demoted. If we see a truncation, | |||
3649 | // we mark it as seeding another demotion. | |||
3650 | for (auto &Entry : VectorizableTree) | |||
3651 | Expr.insert(Entry.Scalars.begin(), Entry.Scalars.end()); | |||
3652 | ||||
3653 | // Ensure the roots of the vectorizable tree don't form a cycle. They must | |||
3654 | // have a single external user that is not in the vectorizable tree. | |||
3655 | for (auto *Root : TreeRoot) | |||
3656 | if (!Root->hasOneUse() || Expr.count(*Root->user_begin())) | |||
3657 | return; | |||
3658 | ||||
3659 | // Conservatively determine if we can actually truncate the roots of the | |||
3660 | // expression. Collect the values that can be demoted in ToDemote and | |||
3661 | // additional roots that require investigating in Roots. | |||
3662 | SmallVector<Value *, 32> ToDemote; | |||
3663 | SmallVector<Value *, 4> Roots; | |||
3664 | for (auto *Root : TreeRoot) | |||
3665 | if (!collectValuesToDemote(Root, Expr, ToDemote, Roots)) | |||
3666 | return; | |||
3667 | ||||
3668 | // The maximum bit width required to represent all the values that can be | |||
3669 | // demoted without loss of precision. It would be safe to truncate the roots | |||
3670 | // of the expression to this width. | |||
3671 | auto MaxBitWidth = 8u; | |||
3672 | ||||
3673 | // We first check if all the bits of the roots are demanded. If they're not, | |||
3674 | // we can truncate the roots to this narrower type. | |||
3675 | for (auto *Root : TreeRoot) { | |||
3676 | auto Mask = DB->getDemandedBits(cast<Instruction>(Root)); | |||
3677 | MaxBitWidth = std::max<unsigned>( | |||
3678 | Mask.getBitWidth() - Mask.countLeadingZeros(), MaxBitWidth); | |||
3679 | } | |||
3680 | ||||
3681 | // True if the roots can be zero-extended back to their original type, rather | |||
3682 | // than sign-extended. We know that if the leading bits are not demanded, we | |||
3683 | // can safely zero-extend. So we initialize IsKnownPositive to True. | |||
3684 | bool IsKnownPositive = true; | |||
3685 | ||||
3686 | // If all the bits of the roots are demanded, we can try a little harder to | |||
3687 | // compute a narrower type. This can happen, for example, if the roots are | |||
3688 | // getelementptr indices. InstCombine promotes these indices to the pointer | |||
3689 | // width. Thus, all their bits are technically demanded even though the | |||
3690 | // address computation might be vectorized in a smaller type. | |||
3691 | // | |||
3692 | // We start by looking at each entry that can be demoted. We compute the | |||
3693 | // maximum bit width required to store the scalar by using ValueTracking to | |||
3694 | // compute the number of high-order bits we can truncate. | |||
3695 | if (MaxBitWidth == DL->getTypeSizeInBits(TreeRoot[0]->getType())) { | |||
3696 | MaxBitWidth = 8u; | |||
3697 | ||||
3698 | // Determine if the sign bit of all the roots is known to be zero. If not, | |||
3699 | // IsKnownPositive is set to False. | |||
3700 | IsKnownPositive = all_of(TreeRoot, [&](Value *R) { | |||
3701 | KnownBits Known = computeKnownBits(R, *DL); | |||
3702 | return Known.isNonNegative(); | |||
3703 | }); | |||
3704 | ||||
3705 | // Determine the maximum number of bits required to store the scalar | |||
3706 | // values. | |||
3707 | for (auto *Scalar : ToDemote) { | |||
3708 | auto NumSignBits = ComputeNumSignBits(Scalar, *DL, 0, AC, 0, DT); | |||
3709 | auto NumTypeBits = DL->getTypeSizeInBits(Scalar->getType()); | |||
3710 | MaxBitWidth = std::max<unsigned>(NumTypeBits - NumSignBits, MaxBitWidth); | |||
3711 | } | |||
3712 | ||||
3713 | // If we can't prove that the sign bit is zero, we must add one to the | |||
3714 | // maximum bit width to account for the unknown sign bit. This preserves | |||
3715 | // the existing sign bit so we can safely sign-extend the root back to the | |||
3716 | // original type. Otherwise, if we know the sign bit is zero, we will | |||
3717 | // zero-extend the root instead. | |||
3718 | // | |||
3719 | // FIXME: This is somewhat suboptimal, as there will be cases where adding | |||
3720 | // one to the maximum bit width will yield a larger-than-necessary | |||
3721 | // type. In general, we need to add an extra bit only if we can't | |||
3722 | // prove that the upper bit of the original type is equal to the | |||
3723 | // upper bit of the proposed smaller type. If these two bits are the | |||
3724 | // same (either zero or one) we know that sign-extending from the | |||
3725 | // smaller type will result in the same value. Here, since we can't | |||
3726 | // yet prove this, we are just making the proposed smaller type | |||
3727 | // larger to ensure correctness. | |||
3728 | if (!IsKnownPositive) | |||
3729 | ++MaxBitWidth; | |||
3730 | } | |||
3731 | ||||
3732 | // Round MaxBitWidth up to the next power-of-two. | |||
3733 | if (!isPowerOf2_64(MaxBitWidth)) | |||
3734 | MaxBitWidth = NextPowerOf2(MaxBitWidth); | |||
3735 | ||||
3736 | // If the maximum bit width we compute is less than the with of the roots' | |||
3737 | // type, we can proceed with the narrowing. Otherwise, do nothing. | |||
3738 | if (MaxBitWidth >= TreeRootIT->getBitWidth()) | |||
3739 | return; | |||
3740 | ||||
3741 | // If we can truncate the root, we must collect additional values that might | |||
3742 | // be demoted as a result. That is, those seeded by truncations we will | |||
3743 | // modify. | |||
3744 | while (!Roots.empty()) | |||
3745 | collectValuesToDemote(Roots.pop_back_val(), Expr, ToDemote, Roots); | |||
3746 | ||||
3747 | // Finally, map the values we can demote to the maximum bit with we computed. | |||
3748 | for (auto *Scalar : ToDemote) | |||
3749 | MinBWs[Scalar] = std::make_pair(MaxBitWidth, !IsKnownPositive); | |||
3750 | } | |||
3751 | ||||
3752 | namespace { | |||
3753 | /// The SLPVectorizer Pass. | |||
3754 | struct SLPVectorizer : public FunctionPass { | |||
3755 | SLPVectorizerPass Impl; | |||
3756 | ||||
3757 | /// Pass identification, replacement for typeid | |||
3758 | static char ID; | |||
3759 | ||||
3760 | explicit SLPVectorizer() : FunctionPass(ID) { | |||
3761 | initializeSLPVectorizerPass(*PassRegistry::getPassRegistry()); | |||
3762 | } | |||
3763 | ||||
3764 | ||||
3765 | bool doInitialization(Module &M) override { | |||
3766 | return false; | |||
3767 | } | |||
3768 | ||||
3769 | bool runOnFunction(Function &F) override { | |||
3770 | if (skipFunction(F)) | |||
3771 | return false; | |||
3772 | ||||
3773 | auto *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE(); | |||
3774 | auto *TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F); | |||
3775 | auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>(); | |||
3776 | auto *TLI = TLIP ? &TLIP->getTLI() : nullptr; | |||
3777 | auto *AA = &getAnalysis<AAResultsWrapperPass>().getAAResults(); | |||
3778 | auto *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); | |||
3779 | auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); | |||
3780 | auto *AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F); | |||
3781 | auto *DB = &getAnalysis<DemandedBitsWrapperPass>().getDemandedBits(); | |||
3782 | auto *ORE = &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE(); | |||
3783 | ||||
3784 | return Impl.runImpl(F, SE, TTI, TLI, AA, LI, DT, AC, DB, ORE); | |||
3785 | } | |||
3786 | ||||
3787 | void getAnalysisUsage(AnalysisUsage &AU) const override { | |||
3788 | FunctionPass::getAnalysisUsage(AU); | |||
3789 | AU.addRequired<AssumptionCacheTracker>(); | |||
3790 | AU.addRequired<ScalarEvolutionWrapperPass>(); | |||
3791 | AU.addRequired<AAResultsWrapperPass>(); | |||
3792 | AU.addRequired<TargetTransformInfoWrapperPass>(); | |||
3793 | AU.addRequired<LoopInfoWrapperPass>(); | |||
3794 | AU.addRequired<DominatorTreeWrapperPass>(); | |||
3795 | AU.addRequired<DemandedBitsWrapperPass>(); | |||
3796 | AU.addRequired<OptimizationRemarkEmitterWrapperPass>(); | |||
3797 | AU.addPreserved<LoopInfoWrapperPass>(); | |||
3798 | AU.addPreserved<DominatorTreeWrapperPass>(); | |||
3799 | AU.addPreserved<AAResultsWrapperPass>(); | |||
3800 | AU.addPreserved<GlobalsAAWrapperPass>(); | |||
3801 | AU.setPreservesCFG(); | |||
3802 | } | |||
3803 | }; | |||
3804 | } // end anonymous namespace | |||
3805 | ||||
3806 | PreservedAnalyses SLPVectorizerPass::run(Function &F, FunctionAnalysisManager &AM) { | |||
3807 | auto *SE = &AM.getResult<ScalarEvolutionAnalysis>(F); | |||
3808 | auto *TTI = &AM.getResult<TargetIRAnalysis>(F); | |||
3809 | auto *TLI = AM.getCachedResult<TargetLibraryAnalysis>(F); | |||
3810 | auto *AA = &AM.getResult<AAManager>(F); | |||
3811 | auto *LI = &AM.getResult<LoopAnalysis>(F); | |||
3812 | auto *DT = &AM.getResult<DominatorTreeAnalysis>(F); | |||
3813 | auto *AC = &AM.getResult<AssumptionAnalysis>(F); | |||
3814 | auto *DB = &AM.getResult<DemandedBitsAnalysis>(F); | |||
3815 | auto *ORE = &AM.getResult<OptimizationRemarkEmitterAnalysis>(F); | |||
3816 | ||||
3817 | bool Changed = runImpl(F, SE, TTI, TLI, AA, LI, DT, AC, DB, ORE); | |||
3818 | if (!Changed) | |||
3819 | return PreservedAnalyses::all(); | |||
3820 | ||||
3821 | PreservedAnalyses PA; | |||
3822 | PA.preserveSet<CFGAnalyses>(); | |||
3823 | PA.preserve<AAManager>(); | |||
3824 | PA.preserve<GlobalsAA>(); | |||
3825 | return PA; | |||
3826 | } | |||
3827 | ||||
3828 | bool SLPVectorizerPass::runImpl(Function &F, ScalarEvolution *SE_, | |||
3829 | TargetTransformInfo *TTI_, | |||
3830 | TargetLibraryInfo *TLI_, AliasAnalysis *AA_, | |||
3831 | LoopInfo *LI_, DominatorTree *DT_, | |||
3832 | AssumptionCache *AC_, DemandedBits *DB_, | |||
3833 | OptimizationRemarkEmitter *ORE_) { | |||
3834 | SE = SE_; | |||
3835 | TTI = TTI_; | |||
3836 | TLI = TLI_; | |||
3837 | AA = AA_; | |||
3838 | LI = LI_; | |||
3839 | DT = DT_; | |||
3840 | AC = AC_; | |||
3841 | DB = DB_; | |||
3842 | DL = &F.getParent()->getDataLayout(); | |||
3843 | ||||
3844 | Stores.clear(); | |||
3845 | GEPs.clear(); | |||
3846 | bool Changed = false; | |||
3847 | ||||
3848 | // If the target claims to have no vector registers don't attempt | |||
3849 | // vectorization. | |||
3850 | if (!TTI->getNumberOfRegisters(true)) | |||
3851 | return false; | |||
3852 | ||||
3853 | // Don't vectorize when the attribute NoImplicitFloat is used. | |||
3854 | if (F.hasFnAttribute(Attribute::NoImplicitFloat)) | |||
3855 | return false; | |||
3856 | ||||
3857 | 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); | |||
3858 | ||||
3859 | // Use the bottom up slp vectorizer to construct chains that start with | |||
3860 | // store instructions. | |||
3861 | BoUpSLP R(&F, SE, TTI, TLI, AA, LI, DT, AC, DB, DL, ORE_); | |||
3862 | ||||
3863 | // A general note: the vectorizer must use BoUpSLP::eraseInstruction() to | |||
3864 | // delete instructions. | |||
3865 | ||||
3866 | // Scan the blocks in the function in post order. | |||
3867 | for (auto BB : post_order(&F.getEntryBlock())) { | |||
3868 | collectSeedInstructions(BB); | |||
3869 | ||||
3870 | // Vectorize trees that end at stores. | |||
3871 | if (!Stores.empty()) { | |||
3872 | 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) | |||
3873 | << " underlying objects.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Found stores for " << Stores .size() << " underlying objects.\n"; } } while (false); | |||
3874 | Changed |= vectorizeStoreChains(R); | |||
3875 | } | |||
3876 | ||||
3877 | // Vectorize trees that end at reductions. | |||
3878 | Changed |= vectorizeChainsInBlock(BB, R); | |||
3879 | ||||
3880 | // Vectorize the index computations of getelementptr instructions. This | |||
3881 | // is primarily intended to catch gather-like idioms ending at | |||
3882 | // non-consecutive loads. | |||
3883 | if (!GEPs.empty()) { | |||
3884 | 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) | |||
3885 | << " underlying objects.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Found GEPs for " << GEPs .size() << " underlying objects.\n"; } } while (false); | |||
3886 | Changed |= vectorizeGEPIndices(BB, R); | |||
3887 | } | |||
3888 | } | |||
3889 | ||||
3890 | if (Changed) { | |||
3891 | R.optimizeGatherSequence(); | |||
3892 | DEBUG(dbgs() << "SLP: vectorized \"" << F.getName() << "\"\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: vectorized \"" << F.getName () << "\"\n"; } } while (false); | |||
3893 | DEBUG(verifyFunction(F))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { verifyFunction(F); } } while (false); | |||
3894 | } | |||
3895 | return Changed; | |||
3896 | } | |||
3897 | ||||
3898 | /// \brief Check that the Values in the slice in VL array are still existent in | |||
3899 | /// the WeakTrackingVH array. | |||
3900 | /// Vectorization of part of the VL array may cause later values in the VL array | |||
3901 | /// to become invalid. We track when this has happened in the WeakTrackingVH | |||
3902 | /// array. | |||
3903 | static bool hasValueBeenRAUWed(ArrayRef<Value *> VL, | |||
3904 | ArrayRef<WeakTrackingVH> VH, unsigned SliceBegin, | |||
3905 | unsigned SliceSize) { | |||
3906 | VL = VL.slice(SliceBegin, SliceSize); | |||
3907 | VH = VH.slice(SliceBegin, SliceSize); | |||
3908 | return !std::equal(VL.begin(), VL.end(), VH.begin()); | |||
3909 | } | |||
3910 | ||||
3911 | bool SLPVectorizerPass::vectorizeStoreChain(ArrayRef<Value *> Chain, BoUpSLP &R, | |||
3912 | unsigned VecRegSize) { | |||
3913 | unsigned ChainLen = Chain.size(); | |||
3914 | 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) | |||
3915 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing a store chain of length " << ChainLen << "\n"; } } while (false); | |||
3916 | unsigned Sz = R.getVectorElementSize(Chain[0]); | |||
3917 | unsigned VF = VecRegSize / Sz; | |||
3918 | ||||
3919 | if (!isPowerOf2_32(Sz) || VF < 2) | |||
3920 | return false; | |||
3921 | ||||
3922 | // Keep track of values that were deleted by vectorizing in the loop below. | |||
3923 | SmallVector<WeakTrackingVH, 8> TrackValues(Chain.begin(), Chain.end()); | |||
3924 | ||||
3925 | bool Changed = false; | |||
3926 | // Look for profitable vectorizable trees at all offsets, starting at zero. | |||
3927 | for (unsigned i = 0, e = ChainLen; i < e; ++i) { | |||
3928 | if (i + VF > e) | |||
3929 | break; | |||
3930 | ||||
3931 | // Check that a previous iteration of this loop did not delete the Value. | |||
3932 | if (hasValueBeenRAUWed(Chain, TrackValues, i, VF)) | |||
3933 | continue; | |||
3934 | ||||
3935 | 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 ) | |||
3936 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing " << VF << " stores at offset " << i << "\n"; } } while (false ); | |||
3937 | ArrayRef<Value *> Operands = Chain.slice(i, VF); | |||
3938 | ||||
3939 | R.buildTree(Operands); | |||
3940 | if (R.isTreeTinyAndNotFullyVectorizable()) | |||
3941 | continue; | |||
3942 | ||||
3943 | R.computeMinimumValueSizes(); | |||
3944 | ||||
3945 | int Cost = R.getTreeCost(); | |||
3946 | ||||
3947 | 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); | |||
3948 | if (Cost < -SLPCostThreshold) { | |||
3949 | 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); | |||
3950 | using namespace ore; | |||
3951 | R.getORE()->emit(OptimizationRemark(SV_NAME"slp-vectorizer", "StoresVectorized", | |||
3952 | cast<StoreInst>(Chain[i])) | |||
3953 | << "Stores SLP vectorized with cost " << NV("Cost", Cost) | |||
3954 | << " and with tree size " | |||
3955 | << NV("TreeSize", R.getTreeSize())); | |||
3956 | ||||
3957 | R.vectorizeTree(); | |||
3958 | ||||
3959 | // Move to the next bundle. | |||
3960 | i += VF - 1; | |||
3961 | Changed = true; | |||
3962 | } | |||
3963 | } | |||
3964 | ||||
3965 | return Changed; | |||
3966 | } | |||
3967 | ||||
3968 | bool SLPVectorizerPass::vectorizeStores(ArrayRef<StoreInst *> Stores, | |||
3969 | BoUpSLP &R) { | |||
3970 | SetVector<StoreInst *> Heads, Tails; | |||
3971 | SmallDenseMap<StoreInst *, StoreInst *> ConsecutiveChain; | |||
3972 | ||||
3973 | // We may run into multiple chains that merge into a single chain. We mark the | |||
3974 | // stores that we vectorized so that we don't visit the same store twice. | |||
3975 | BoUpSLP::ValueSet VectorizedStores; | |||
3976 | bool Changed = false; | |||
3977 | ||||
3978 | // Do a quadratic search on all of the given stores and find | |||
3979 | // all of the pairs of stores that follow each other. | |||
3980 | SmallVector<unsigned, 16> IndexQueue; | |||
3981 | for (unsigned i = 0, e = Stores.size(); i < e; ++i) { | |||
3982 | IndexQueue.clear(); | |||
3983 | // If a store has multiple consecutive store candidates, search Stores | |||
3984 | // array according to the sequence: from i+1 to e, then from i-1 to 0. | |||
3985 | // This is because usually pairing with immediate succeeding or preceding | |||
3986 | // candidate create the best chance to find slp vectorization opportunity. | |||
3987 | unsigned j = 0; | |||
3988 | for (j = i + 1; j < e; ++j) | |||
3989 | IndexQueue.push_back(j); | |||
3990 | for (j = i; j > 0; --j) | |||
3991 | IndexQueue.push_back(j - 1); | |||
3992 | ||||
3993 | for (auto &k : IndexQueue) { | |||
3994 | if (isConsecutiveAccess(Stores[i], Stores[k], *DL, *SE)) { | |||
3995 | Tails.insert(Stores[k]); | |||
3996 | Heads.insert(Stores[i]); | |||
3997 | ConsecutiveChain[Stores[i]] = Stores[k]; | |||
3998 | break; | |||
3999 | } | |||
4000 | } | |||
4001 | } | |||
4002 | ||||
4003 | // For stores that start but don't end a link in the chain: | |||
4004 | for (SetVector<StoreInst *>::iterator it = Heads.begin(), e = Heads.end(); | |||
4005 | it != e; ++it) { | |||
4006 | if (Tails.count(*it)) | |||
4007 | continue; | |||
4008 | ||||
4009 | // We found a store instr that starts a chain. Now follow the chain and try | |||
4010 | // to vectorize it. | |||
4011 | BoUpSLP::ValueList Operands; | |||
4012 | StoreInst *I = *it; | |||
4013 | // Collect the chain into a list. | |||
4014 | while (Tails.count(I) || Heads.count(I)) { | |||
4015 | if (VectorizedStores.count(I)) | |||
4016 | break; | |||
4017 | Operands.push_back(I); | |||
4018 | // Move to the next value in the chain. | |||
4019 | I = ConsecutiveChain[I]; | |||
4020 | } | |||
4021 | ||||
4022 | // FIXME: Is division-by-2 the correct step? Should we assert that the | |||
4023 | // register size is a power-of-2? | |||
4024 | for (unsigned Size = R.getMaxVecRegSize(); Size >= R.getMinVecRegSize(); | |||
4025 | Size /= 2) { | |||
4026 | if (vectorizeStoreChain(Operands, R, Size)) { | |||
4027 | // Mark the vectorized stores so that we don't vectorize them again. | |||
4028 | VectorizedStores.insert(Operands.begin(), Operands.end()); | |||
4029 | Changed = true; | |||
4030 | break; | |||
4031 | } | |||
4032 | } | |||
4033 | } | |||
4034 | ||||
4035 | return Changed; | |||
4036 | } | |||
4037 | ||||
4038 | void SLPVectorizerPass::collectSeedInstructions(BasicBlock *BB) { | |||
4039 | ||||
4040 | // Initialize the collections. We will make a single pass over the block. | |||
4041 | Stores.clear(); | |||
4042 | GEPs.clear(); | |||
4043 | ||||
4044 | // Visit the store and getelementptr instructions in BB and organize them in | |||
4045 | // Stores and GEPs according to the underlying objects of their pointer | |||
4046 | // operands. | |||
4047 | for (Instruction &I : *BB) { | |||
4048 | ||||
4049 | // Ignore store instructions that are volatile or have a pointer operand | |||
4050 | // that doesn't point to a scalar type. | |||
4051 | if (auto *SI = dyn_cast<StoreInst>(&I)) { | |||
4052 | if (!SI->isSimple()) | |||
4053 | continue; | |||
4054 | if (!isValidElementType(SI->getValueOperand()->getType())) | |||
4055 | continue; | |||
4056 | Stores[GetUnderlyingObject(SI->getPointerOperand(), *DL)].push_back(SI); | |||
4057 | } | |||
4058 | ||||
4059 | // Ignore getelementptr instructions that have more than one index, a | |||
4060 | // constant index, or a pointer operand that doesn't point to a scalar | |||
4061 | // type. | |||
4062 | else if (auto *GEP = dyn_cast<GetElementPtrInst>(&I)) { | |||
4063 | auto Idx = GEP->idx_begin()->get(); | |||
4064 | if (GEP->getNumIndices() > 1 || isa<Constant>(Idx)) | |||
4065 | continue; | |||
4066 | if (!isValidElementType(Idx->getType())) | |||
4067 | continue; | |||
4068 | if (GEP->getType()->isVectorTy()) | |||
4069 | continue; | |||
4070 | GEPs[GetUnderlyingObject(GEP->getPointerOperand(), *DL)].push_back(GEP); | |||
4071 | } | |||
4072 | } | |||
4073 | } | |||
4074 | ||||
4075 | bool SLPVectorizerPass::tryToVectorizePair(Value *A, Value *B, BoUpSLP &R) { | |||
4076 | if (!A || !B) | |||
4077 | return false; | |||
4078 | Value *VL[] = { A, B }; | |||
4079 | return tryToVectorizeList(VL, R, None, true); | |||
4080 | } | |||
4081 | ||||
4082 | bool SLPVectorizerPass::tryToVectorizeList(ArrayRef<Value *> VL, BoUpSLP &R, | |||
4083 | ArrayRef<Value *> BuildVector, | |||
4084 | bool AllowReorder) { | |||
4085 | if (VL.size() < 2) | |||
4086 | return false; | |||
4087 | ||||
4088 | 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) | |||
4089 | << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Trying to vectorize a list of length = " << VL.size() << ".\n"; } } while (false); | |||
4090 | ||||
4091 | // Check that all of the parts are scalar instructions of the same type. | |||
4092 | Instruction *I0 = dyn_cast<Instruction>(VL[0]); | |||
4093 | if (!I0) | |||
4094 | return false; | |||
4095 | ||||
4096 | unsigned Opcode0 = I0->getOpcode(); | |||
4097 | ||||
4098 | unsigned Sz = R.getVectorElementSize(I0); | |||
4099 | unsigned MinVF = std::max(2U, R.getMinVecRegSize() / Sz); | |||
4100 | unsigned MaxVF = std::max<unsigned>(PowerOf2Floor(VL.size()), MinVF); | |||
4101 | if (MaxVF < 2) | |||
4102 | return false; | |||
4103 | ||||
4104 | for (Value *V : VL) { | |||
4105 | Type *Ty = V->getType(); | |||
4106 | if (!isValidElementType(Ty)) | |||
4107 | return false; | |||
4108 | Instruction *Inst = dyn_cast<Instruction>(V); | |||
4109 | if (!Inst || Inst->getOpcode() != Opcode0) | |||
4110 | return false; | |||
4111 | } | |||
4112 | ||||
4113 | bool Changed = false; | |||
4114 | ||||
4115 | // Keep track of values that were deleted by vectorizing in the loop below. | |||
4116 | SmallVector<WeakTrackingVH, 8> TrackValues(VL.begin(), VL.end()); | |||
4117 | ||||
4118 | unsigned NextInst = 0, MaxInst = VL.size(); | |||
4119 | for (unsigned VF = MaxVF; NextInst + 1 < MaxInst && VF >= MinVF; | |||
4120 | VF /= 2) { | |||
4121 | // No actual vectorization should happen, if number of parts is the same as | |||
4122 | // provided vectorization factor (i.e. the scalar type is used for vector | |||
4123 | // code during codegen). | |||
4124 | auto *VecTy = VectorType::get(VL[0]->getType(), VF); | |||
4125 | if (TTI->getNumberOfParts(VecTy) == VF) | |||
4126 | continue; | |||
4127 | for (unsigned I = NextInst; I < MaxInst; ++I) { | |||
4128 | unsigned OpsWidth = 0; | |||
4129 | ||||
4130 | if (I + VF > MaxInst) | |||
4131 | OpsWidth = MaxInst - I; | |||
4132 | else | |||
4133 | OpsWidth = VF; | |||
4134 | ||||
4135 | if (!isPowerOf2_32(OpsWidth) || OpsWidth < 2) | |||
4136 | break; | |||
4137 | ||||
4138 | // Check that a previous iteration of this loop did not delete the Value. | |||
4139 | if (hasValueBeenRAUWed(VL, TrackValues, I, OpsWidth)) | |||
4140 | continue; | |||
4141 | ||||
4142 | DEBUG(dbgs() << "SLP: Analyzing " << OpsWidth << " operations "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing " << OpsWidth << " operations " << "\n"; } } while (false) | |||
4143 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing " << OpsWidth << " operations " << "\n"; } } while (false); | |||
4144 | ArrayRef<Value *> Ops = VL.slice(I, OpsWidth); | |||
4145 | ||||
4146 | ArrayRef<Value *> BuildVectorSlice; | |||
4147 | if (!BuildVector.empty()) | |||
4148 | BuildVectorSlice = BuildVector.slice(I, OpsWidth); | |||
4149 | ||||
4150 | R.buildTree(Ops, BuildVectorSlice); | |||
4151 | // TODO: check if we can allow reordering for more cases. | |||
4152 | if (AllowReorder && R.shouldReorder()) { | |||
4153 | // Conceptually, there is nothing actually preventing us from trying to | |||
4154 | // reorder a larger list. In fact, we do exactly this when vectorizing | |||
4155 | // reductions. However, at this point, we only expect to get here when | |||
4156 | // there are exactly two operations. | |||
4157 | assert(Ops.size() == 2)((Ops.size() == 2) ? static_cast<void> (0) : __assert_fail ("Ops.size() == 2", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4157, __PRETTY_FUNCTION__)); | |||
4158 | assert(BuildVectorSlice.empty())((BuildVectorSlice.empty()) ? static_cast<void> (0) : __assert_fail ("BuildVectorSlice.empty()", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4158, __PRETTY_FUNCTION__)); | |||
4159 | Value *ReorderedOps[] = {Ops[1], Ops[0]}; | |||
4160 | R.buildTree(ReorderedOps, None); | |||
4161 | } | |||
4162 | if (R.isTreeTinyAndNotFullyVectorizable()) | |||
4163 | continue; | |||
4164 | ||||
4165 | R.computeMinimumValueSizes(); | |||
4166 | int Cost = R.getTreeCost(); | |||
4167 | ||||
4168 | if (Cost < -SLPCostThreshold) { | |||
4169 | 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); | |||
4170 | R.getORE()->emit(OptimizationRemark(SV_NAME"slp-vectorizer", "VectorizedList", | |||
4171 | cast<Instruction>(Ops[0])) | |||
4172 | << "SLP vectorized with cost " << ore::NV("Cost", Cost) | |||
4173 | << " and with tree size " | |||
4174 | << ore::NV("TreeSize", R.getTreeSize())); | |||
4175 | ||||
4176 | Value *VectorizedRoot = R.vectorizeTree(); | |||
4177 | ||||
4178 | // Reconstruct the build vector by extracting the vectorized root. This | |||
4179 | // way we handle the case where some elements of the vector are | |||
4180 | // undefined. | |||
4181 | // (return (inserelt <4 xi32> (insertelt undef (opd0) 0) (opd1) 2)) | |||
4182 | if (!BuildVectorSlice.empty()) { | |||
4183 | // The insert point is the last build vector instruction. The | |||
4184 | // vectorized root will precede it. This guarantees that we get an | |||
4185 | // instruction. The vectorized tree could have been constant folded. | |||
4186 | Instruction *InsertAfter = cast<Instruction>(BuildVectorSlice.back()); | |||
4187 | unsigned VecIdx = 0; | |||
4188 | for (auto &V : BuildVectorSlice) { | |||
4189 | IRBuilder<NoFolder> Builder(InsertAfter->getParent(), | |||
4190 | ++BasicBlock::iterator(InsertAfter)); | |||
4191 | Instruction *I = cast<Instruction>(V); | |||
4192 | assert(isa<InsertElementInst>(I) || isa<InsertValueInst>(I))((isa<InsertElementInst>(I) || isa<InsertValueInst> (I)) ? static_cast<void> (0) : __assert_fail ("isa<InsertElementInst>(I) || isa<InsertValueInst>(I)" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4192, __PRETTY_FUNCTION__)); | |||
4193 | Instruction *Extract = | |||
4194 | cast<Instruction>(Builder.CreateExtractElement( | |||
4195 | VectorizedRoot, Builder.getInt32(VecIdx++))); | |||
4196 | I->setOperand(1, Extract); | |||
4197 | I->removeFromParent(); | |||
4198 | I->insertAfter(Extract); | |||
4199 | InsertAfter = I; | |||
4200 | } | |||
4201 | } | |||
4202 | // Move to the next bundle. | |||
4203 | I += VF - 1; | |||
4204 | NextInst = I + 1; | |||
4205 | Changed = true; | |||
4206 | } | |||
4207 | } | |||
4208 | } | |||
4209 | ||||
4210 | return Changed; | |||
4211 | } | |||
4212 | ||||
4213 | bool SLPVectorizerPass::tryToVectorize(BinaryOperator *V, BoUpSLP &R) { | |||
4214 | if (!V) | |||
4215 | return false; | |||
4216 | ||||
4217 | Value *P = V->getParent(); | |||
4218 | ||||
4219 | // Vectorize in current basic block only. | |||
4220 | auto *Op0 = dyn_cast<Instruction>(V->getOperand(0)); | |||
4221 | auto *Op1 = dyn_cast<Instruction>(V->getOperand(1)); | |||
4222 | if (!Op0 || !Op1 || Op0->getParent() != P || Op1->getParent() != P) | |||
4223 | return false; | |||
4224 | ||||
4225 | // Try to vectorize V. | |||
4226 | if (tryToVectorizePair(Op0, Op1, R)) | |||
4227 | return true; | |||
4228 | ||||
4229 | auto *A = dyn_cast<BinaryOperator>(Op0); | |||
4230 | auto *B = dyn_cast<BinaryOperator>(Op1); | |||
4231 | // Try to skip B. | |||
4232 | if (B && B->hasOneUse()) { | |||
4233 | auto *B0 = dyn_cast<BinaryOperator>(B->getOperand(0)); | |||
4234 | auto *B1 = dyn_cast<BinaryOperator>(B->getOperand(1)); | |||
4235 | if (B0 && B0->getParent() == P && tryToVectorizePair(A, B0, R)) | |||
4236 | return true; | |||
4237 | if (B1 && B1->getParent() == P && tryToVectorizePair(A, B1, R)) | |||
4238 | return true; | |||
4239 | } | |||
4240 | ||||
4241 | // Try to skip A. | |||
4242 | if (A && A->hasOneUse()) { | |||
4243 | auto *A0 = dyn_cast<BinaryOperator>(A->getOperand(0)); | |||
4244 | auto *A1 = dyn_cast<BinaryOperator>(A->getOperand(1)); | |||
4245 | if (A0 && A0->getParent() == P && tryToVectorizePair(A0, B, R)) | |||
4246 | return true; | |||
4247 | if (A1 && A1->getParent() == P && tryToVectorizePair(A1, B, R)) | |||
4248 | return true; | |||
4249 | } | |||
4250 | return false; | |||
4251 | } | |||
4252 | ||||
4253 | /// \brief Generate a shuffle mask to be used in a reduction tree. | |||
4254 | /// | |||
4255 | /// \param VecLen The length of the vector to be reduced. | |||
4256 | /// \param NumEltsToRdx The number of elements that should be reduced in the | |||
4257 | /// vector. | |||
4258 | /// \param IsPairwise Whether the reduction is a pairwise or splitting | |||
4259 | /// reduction. A pairwise reduction will generate a mask of | |||
4260 | /// <0,2,...> or <1,3,..> while a splitting reduction will generate | |||
4261 | /// <2,3, undef,undef> for a vector of 4 and NumElts = 2. | |||
4262 | /// \param IsLeft True will generate a mask of even elements, odd otherwise. | |||
4263 | static Value *createRdxShuffleMask(unsigned VecLen, unsigned NumEltsToRdx, | |||
4264 | bool IsPairwise, bool IsLeft, | |||
4265 | IRBuilder<> &Builder) { | |||
4266 | assert((IsPairwise || !IsLeft) && "Don't support a <0,1,undef,...> mask")(((IsPairwise || !IsLeft) && "Don't support a <0,1,undef,...> mask" ) ? static_cast<void> (0) : __assert_fail ("(IsPairwise || !IsLeft) && \"Don't support a <0,1,undef,...> mask\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4266, __PRETTY_FUNCTION__)); | |||
4267 | ||||
4268 | SmallVector<Constant *, 32> ShuffleMask( | |||
4269 | VecLen, UndefValue::get(Builder.getInt32Ty())); | |||
4270 | ||||
4271 | if (IsPairwise) | |||
4272 | // Build a mask of 0, 2, ... (left) or 1, 3, ... (right). | |||
4273 | for (unsigned i = 0; i != NumEltsToRdx; ++i) | |||
4274 | ShuffleMask[i] = Builder.getInt32(2 * i + !IsLeft); | |||
4275 | else | |||
4276 | // Move the upper half of the vector to the lower half. | |||
4277 | for (unsigned i = 0; i != NumEltsToRdx; ++i) | |||
4278 | ShuffleMask[i] = Builder.getInt32(NumEltsToRdx + i); | |||
4279 | ||||
4280 | return ConstantVector::get(ShuffleMask); | |||
4281 | } | |||
4282 | ||||
4283 | namespace { | |||
4284 | /// Model horizontal reductions. | |||
4285 | /// | |||
4286 | /// A horizontal reduction is a tree of reduction operations (currently add and | |||
4287 | /// fadd) that has operations that can be put into a vector as its leaf. | |||
4288 | /// For example, this tree: | |||
4289 | /// | |||
4290 | /// mul mul mul mul | |||
4291 | /// \ / \ / | |||
4292 | /// + + | |||
4293 | /// \ / | |||
4294 | /// + | |||
4295 | /// This tree has "mul" as its reduced values and "+" as its reduction | |||
4296 | /// operations. A reduction might be feeding into a store or a binary operation | |||
4297 | /// feeding a phi. | |||
4298 | /// ... | |||
4299 | /// \ / | |||
4300 | /// + | |||
4301 | /// | | |||
4302 | /// phi += | |||
4303 | /// | |||
4304 | /// Or: | |||
4305 | /// ... | |||
4306 | /// \ / | |||
4307 | /// + | |||
4308 | /// | | |||
4309 | /// *p = | |||
4310 | /// | |||
4311 | class HorizontalReduction { | |||
4312 | SmallVector<Value *, 16> ReductionOps; | |||
4313 | SmallVector<Value *, 32> ReducedVals; | |||
4314 | // Use map vector to make stable output. | |||
4315 | MapVector<Instruction *, Value *> ExtraArgs; | |||
4316 | ||||
4317 | BinaryOperator *ReductionRoot = nullptr; | |||
4318 | ||||
4319 | /// The opcode of the reduction. | |||
4320 | Instruction::BinaryOps ReductionOpcode = Instruction::BinaryOpsEnd; | |||
4321 | /// The opcode of the values we perform a reduction on. | |||
4322 | unsigned ReducedValueOpcode = 0; | |||
4323 | /// Should we model this reduction as a pairwise reduction tree or a tree that | |||
4324 | /// splits the vector in halves and adds those halves. | |||
4325 | bool IsPairwiseReduction = false; | |||
4326 | ||||
4327 | /// Checks if the ParentStackElem.first should be marked as a reduction | |||
4328 | /// operation with an extra argument or as extra argument itself. | |||
4329 | void markExtraArg(std::pair<Instruction *, unsigned> &ParentStackElem, | |||
4330 | Value *ExtraArg) { | |||
4331 | if (ExtraArgs.count(ParentStackElem.first)) { | |||
4332 | ExtraArgs[ParentStackElem.first] = nullptr; | |||
4333 | // We ran into something like: | |||
4334 | // ParentStackElem.first = ExtraArgs[ParentStackElem.first] + ExtraArg. | |||
4335 | // The whole ParentStackElem.first should be considered as an extra value | |||
4336 | // in this case. | |||
4337 | // Do not perform analysis of remaining operands of ParentStackElem.first | |||
4338 | // instruction, this whole instruction is an extra argument. | |||
4339 | ParentStackElem.second = ParentStackElem.first->getNumOperands(); | |||
4340 | } else { | |||
4341 | // We ran into something like: | |||
4342 | // ParentStackElem.first += ... + ExtraArg + ... | |||
4343 | ExtraArgs[ParentStackElem.first] = ExtraArg; | |||
4344 | } | |||
4345 | } | |||
4346 | ||||
4347 | public: | |||
4348 | HorizontalReduction() = default; | |||
4349 | ||||
4350 | /// \brief Try to find a reduction tree. | |||
4351 | bool matchAssociativeReduction(PHINode *Phi, BinaryOperator *B) { | |||
4352 | assert((!Phi || is_contained(Phi->operands(), B)) &&(((!Phi || is_contained(Phi->operands(), B)) && "Thi phi needs to use the binary operator" ) ? static_cast<void> (0) : __assert_fail ("(!Phi || is_contained(Phi->operands(), B)) && \"Thi phi needs to use the binary operator\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4353, __PRETTY_FUNCTION__)) | |||
4353 | "Thi phi needs to use the binary operator")(((!Phi || is_contained(Phi->operands(), B)) && "Thi phi needs to use the binary operator" ) ? static_cast<void> (0) : __assert_fail ("(!Phi || is_contained(Phi->operands(), B)) && \"Thi phi needs to use the binary operator\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4353, __PRETTY_FUNCTION__)); | |||
4354 | ||||
4355 | // We could have a initial reductions that is not an add. | |||
4356 | // r *= v1 + v2 + v3 + v4 | |||
4357 | // In such a case start looking for a tree rooted in the first '+'. | |||
4358 | if (Phi) { | |||
4359 | if (B->getOperand(0) == Phi) { | |||
4360 | Phi = nullptr; | |||
4361 | B = dyn_cast<BinaryOperator>(B->getOperand(1)); | |||
4362 | } else if (B->getOperand(1) == Phi) { | |||
4363 | Phi = nullptr; | |||
4364 | B = dyn_cast<BinaryOperator>(B->getOperand(0)); | |||
4365 | } | |||
4366 | } | |||
4367 | ||||
4368 | if (!B) | |||
4369 | return false; | |||
4370 | ||||
4371 | Type *Ty = B->getType(); | |||
4372 | if (!isValidElementType(Ty)) | |||
4373 | return false; | |||
4374 | ||||
4375 | ReductionOpcode = B->getOpcode(); | |||
4376 | ReducedValueOpcode = 0; | |||
4377 | ReductionRoot = B; | |||
4378 | ||||
4379 | // We currently only support adds. | |||
4380 | if ((ReductionOpcode != Instruction::Add && | |||
4381 | ReductionOpcode != Instruction::FAdd) || | |||
4382 | !B->isAssociative()) | |||
4383 | return false; | |||
4384 | ||||
4385 | // Post order traverse the reduction tree starting at B. We only handle true | |||
4386 | // trees containing only binary operators or selects. | |||
4387 | SmallVector<std::pair<Instruction *, unsigned>, 32> Stack; | |||
4388 | Stack.push_back(std::make_pair(B, 0)); | |||
4389 | while (!Stack.empty()) { | |||
4390 | Instruction *TreeN = Stack.back().first; | |||
4391 | unsigned EdgeToVist = Stack.back().second++; | |||
4392 | bool IsReducedValue = TreeN->getOpcode() != ReductionOpcode; | |||
4393 | ||||
4394 | // Postorder vist. | |||
4395 | if (EdgeToVist == 2 || IsReducedValue) { | |||
4396 | if (IsReducedValue) | |||
4397 | ReducedVals.push_back(TreeN); | |||
4398 | else { | |||
4399 | auto I = ExtraArgs.find(TreeN); | |||
4400 | if (I != ExtraArgs.end() && !I->second) { | |||
4401 | // Check if TreeN is an extra argument of its parent operation. | |||
4402 | if (Stack.size() <= 1) { | |||
4403 | // TreeN can't be an extra argument as it is a root reduction | |||
4404 | // operation. | |||
4405 | return false; | |||
4406 | } | |||
4407 | // Yes, TreeN is an extra argument, do not add it to a list of | |||
4408 | // reduction operations. | |||
4409 | // Stack[Stack.size() - 2] always points to the parent operation. | |||
4410 | markExtraArg(Stack[Stack.size() - 2], TreeN); | |||
4411 | ExtraArgs.erase(TreeN); | |||
4412 | } else | |||
4413 | ReductionOps.push_back(TreeN); | |||
4414 | } | |||
4415 | // Retract. | |||
4416 | Stack.pop_back(); | |||
4417 | continue; | |||
4418 | } | |||
4419 | ||||
4420 | // Visit left or right. | |||
4421 | Value *NextV = TreeN->getOperand(EdgeToVist); | |||
4422 | if (NextV != Phi) { | |||
4423 | auto *I = dyn_cast<Instruction>(NextV); | |||
4424 | // Continue analysis if the next operand is a reduction operation or | |||
4425 | // (possibly) a reduced value. If the reduced value opcode is not set, | |||
4426 | // the first met operation != reduction operation is considered as the | |||
4427 | // reduced value class. | |||
4428 | if (I && (!ReducedValueOpcode || I->getOpcode() == ReducedValueOpcode || | |||
4429 | I->getOpcode() == ReductionOpcode)) { | |||
4430 | // Only handle trees in the current basic block. | |||
4431 | if (I->getParent() != B->getParent()) { | |||
4432 | // I is an extra argument for TreeN (its parent operation). | |||
4433 | markExtraArg(Stack.back(), I); | |||
4434 | continue; | |||
4435 | } | |||
4436 | ||||
4437 | // Each tree node needs to have one user except for the ultimate | |||
4438 | // reduction. | |||
4439 | if (!I->hasOneUse() && I != B) { | |||
4440 | // I is an extra argument for TreeN (its parent operation). | |||
4441 | markExtraArg(Stack.back(), I); | |||
4442 | continue; | |||
4443 | } | |||
4444 | ||||
4445 | if (I->getOpcode() == ReductionOpcode) { | |||
4446 | // We need to be able to reassociate the reduction operations. | |||
4447 | if (!I->isAssociative()) { | |||
4448 | // I is an extra argument for TreeN (its parent operation). | |||
4449 | markExtraArg(Stack.back(), I); | |||
4450 | continue; | |||
4451 | } | |||
4452 | } else if (ReducedValueOpcode && | |||
4453 | ReducedValueOpcode != I->getOpcode()) { | |||
4454 | // Make sure that the opcodes of the operations that we are going to | |||
4455 | // reduce match. | |||
4456 | // I is an extra argument for TreeN (its parent operation). | |||
4457 | markExtraArg(Stack.back(), I); | |||
4458 | continue; | |||
4459 | } else if (!ReducedValueOpcode) | |||
4460 | ReducedValueOpcode = I->getOpcode(); | |||
4461 | ||||
4462 | Stack.push_back(std::make_pair(I, 0)); | |||
4463 | continue; | |||
4464 | } | |||
4465 | } | |||
4466 | // NextV is an extra argument for TreeN (its parent operation). | |||
4467 | markExtraArg(Stack.back(), NextV); | |||
4468 | } | |||
4469 | return true; | |||
4470 | } | |||
4471 | ||||
4472 | /// \brief Attempt to vectorize the tree found by | |||
4473 | /// matchAssociativeReduction. | |||
4474 | bool tryToReduce(BoUpSLP &V, TargetTransformInfo *TTI) { | |||
4475 | if (ReducedVals.empty()) | |||
4476 | return false; | |||
4477 | ||||
4478 | // If there is a sufficient number of reduction values, reduce | |||
4479 | // to a nearby power-of-2. Can safely generate oversized | |||
4480 | // vectors and rely on the backend to split them to legal sizes. | |||
4481 | unsigned NumReducedVals = ReducedVals.size(); | |||
4482 | if (NumReducedVals < 4) | |||
4483 | return false; | |||
4484 | ||||
4485 | unsigned ReduxWidth = PowerOf2Floor(NumReducedVals); | |||
4486 | ||||
4487 | Value *VectorizedTree = nullptr; | |||
4488 | IRBuilder<> Builder(ReductionRoot); | |||
4489 | FastMathFlags Unsafe; | |||
4490 | Unsafe.setUnsafeAlgebra(); | |||
4491 | Builder.setFastMathFlags(Unsafe); | |||
4492 | unsigned i = 0; | |||
4493 | ||||
4494 | BoUpSLP::ExtraValueToDebugLocsMap ExternallyUsedValues; | |||
4495 | // The same extra argument may be used several time, so log each attempt | |||
4496 | // to use it. | |||
4497 | for (auto &Pair : ExtraArgs) | |||
4498 | ExternallyUsedValues[Pair.second].push_back(Pair.first); | |||
4499 | while (i < NumReducedVals - ReduxWidth + 1 && ReduxWidth > 2) { | |||
4500 | auto VL = makeArrayRef(&ReducedVals[i], ReduxWidth); | |||
4501 | V.buildTree(VL, ExternallyUsedValues, ReductionOps); | |||
4502 | if (V.shouldReorder()) { | |||
4503 | SmallVector<Value *, 8> Reversed(VL.rbegin(), VL.rend()); | |||
4504 | V.buildTree(Reversed, ExternallyUsedValues, ReductionOps); | |||
4505 | } | |||
4506 | if (V.isTreeTinyAndNotFullyVectorizable()) | |||
4507 | break; | |||
4508 | ||||
4509 | V.computeMinimumValueSizes(); | |||
4510 | ||||
4511 | // Estimate cost. | |||
4512 | int Cost = | |||
4513 | V.getTreeCost() + getReductionCost(TTI, ReducedVals[i], ReduxWidth); | |||
4514 | if (Cost >= -SLPCostThreshold) | |||
4515 | break; | |||
4516 | ||||
4517 | 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) | |||
4518 | << ". (HorRdx)\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Vectorizing horizontal reduction at cost:" << Cost << ". (HorRdx)\n"; } } while (false); | |||
4519 | auto *I0 = cast<Instruction>(VL[0]); | |||
4520 | V.getORE()->emit( | |||
4521 | OptimizationRemark(SV_NAME"slp-vectorizer", "VectorizedHorizontalReduction", I0) | |||
4522 | << "Vectorized horizontal reduction with cost " | |||
4523 | << ore::NV("Cost", Cost) << " and with tree size " | |||
4524 | << ore::NV("TreeSize", V.getTreeSize())); | |||
4525 | ||||
4526 | // Vectorize a tree. | |||
4527 | DebugLoc Loc = cast<Instruction>(ReducedVals[i])->getDebugLoc(); | |||
4528 | Value *VectorizedRoot = V.vectorizeTree(ExternallyUsedValues); | |||
4529 | ||||
4530 | // Emit a reduction. | |||
4531 | Value *ReducedSubTree = | |||
4532 | emitReduction(VectorizedRoot, Builder, ReduxWidth, ReductionOps, TTI); | |||
4533 | if (VectorizedTree) { | |||
4534 | Builder.SetCurrentDebugLocation(Loc); | |||
4535 | VectorizedTree = Builder.CreateBinOp(ReductionOpcode, VectorizedTree, | |||
4536 | ReducedSubTree, "bin.rdx"); | |||
4537 | propagateIRFlags(VectorizedTree, ReductionOps); | |||
4538 | } else | |||
4539 | VectorizedTree = ReducedSubTree; | |||
4540 | i += ReduxWidth; | |||
4541 | ReduxWidth = PowerOf2Floor(NumReducedVals - i); | |||
4542 | } | |||
4543 | ||||
4544 | if (VectorizedTree) { | |||
4545 | // Finish the reduction. | |||
4546 | for (; i < NumReducedVals; ++i) { | |||
4547 | auto *I = cast<Instruction>(ReducedVals[i]); | |||
4548 | Builder.SetCurrentDebugLocation(I->getDebugLoc()); | |||
4549 | VectorizedTree = | |||
4550 | Builder.CreateBinOp(ReductionOpcode, VectorizedTree, I); | |||
4551 | propagateIRFlags(VectorizedTree, ReductionOps); | |||
4552 | } | |||
4553 | for (auto &Pair : ExternallyUsedValues) { | |||
4554 | assert(!Pair.second.empty() &&((!Pair.second.empty() && "At least one DebugLoc must be inserted" ) ? static_cast<void> (0) : __assert_fail ("!Pair.second.empty() && \"At least one DebugLoc must be inserted\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4555, __PRETTY_FUNCTION__)) | |||
4555 | "At least one DebugLoc must be inserted")((!Pair.second.empty() && "At least one DebugLoc must be inserted" ) ? static_cast<void> (0) : __assert_fail ("!Pair.second.empty() && \"At least one DebugLoc must be inserted\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4555, __PRETTY_FUNCTION__)); | |||
4556 | // Add each externally used value to the final reduction. | |||
4557 | for (auto *I : Pair.second) { | |||
4558 | Builder.SetCurrentDebugLocation(I->getDebugLoc()); | |||
4559 | VectorizedTree = Builder.CreateBinOp(ReductionOpcode, VectorizedTree, | |||
4560 | Pair.first, "bin.extra"); | |||
4561 | propagateIRFlags(VectorizedTree, I); | |||
4562 | } | |||
4563 | } | |||
4564 | // Update users. | |||
4565 | ReductionRoot->replaceAllUsesWith(VectorizedTree); | |||
4566 | } | |||
4567 | return VectorizedTree != nullptr; | |||
4568 | } | |||
4569 | ||||
4570 | unsigned numReductionValues() const { | |||
4571 | return ReducedVals.size(); | |||
4572 | } | |||
4573 | ||||
4574 | private: | |||
4575 | /// \brief Calculate the cost of a reduction. | |||
4576 | int getReductionCost(TargetTransformInfo *TTI, Value *FirstReducedVal, | |||
4577 | unsigned ReduxWidth) { | |||
4578 | Type *ScalarTy = FirstReducedVal->getType(); | |||
4579 | Type *VecTy = VectorType::get(ScalarTy, ReduxWidth); | |||
4580 | ||||
4581 | int PairwiseRdxCost = TTI->getReductionCost(ReductionOpcode, VecTy, true); | |||
4582 | int SplittingRdxCost = TTI->getReductionCost(ReductionOpcode, VecTy, false); | |||
4583 | ||||
4584 | IsPairwiseReduction = PairwiseRdxCost < SplittingRdxCost; | |||
4585 | int VecReduxCost = IsPairwiseReduction ? PairwiseRdxCost : SplittingRdxCost; | |||
4586 | ||||
4587 | int ScalarReduxCost = | |||
4588 | (ReduxWidth - 1) * | |||
4589 | TTI->getArithmeticInstrCost(ReductionOpcode, ScalarTy); | |||
4590 | ||||
4591 | 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) | |||
4592 | << " 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) | |||
4593 | << " (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) | |||
4594 | << (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) | |||
4595 | << " 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); | |||
4596 | ||||
4597 | return VecReduxCost - ScalarReduxCost; | |||
4598 | } | |||
4599 | ||||
4600 | /// \brief Emit a horizontal reduction of the vectorized value. | |||
4601 | Value *emitReduction(Value *VectorizedValue, IRBuilder<> &Builder, | |||
4602 | unsigned ReduxWidth, ArrayRef<Value *> RedOps, | |||
4603 | const TargetTransformInfo *TTI) { | |||
4604 | assert(VectorizedValue && "Need to have a vectorized tree node")((VectorizedValue && "Need to have a vectorized tree node" ) ? static_cast<void> (0) : __assert_fail ("VectorizedValue && \"Need to have a vectorized tree node\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4604, __PRETTY_FUNCTION__)); | |||
4605 | assert(isPowerOf2_32(ReduxWidth) &&((isPowerOf2_32(ReduxWidth) && "We only handle power-of-two reductions for now" ) ? static_cast<void> (0) : __assert_fail ("isPowerOf2_32(ReduxWidth) && \"We only handle power-of-two reductions for now\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4606, __PRETTY_FUNCTION__)) | |||
4606 | "We only handle power-of-two reductions for now")((isPowerOf2_32(ReduxWidth) && "We only handle power-of-two reductions for now" ) ? static_cast<void> (0) : __assert_fail ("isPowerOf2_32(ReduxWidth) && \"We only handle power-of-two reductions for now\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4606, __PRETTY_FUNCTION__)); | |||
4607 | ||||
4608 | if (!IsPairwiseReduction) | |||
4609 | return createSimpleTargetReduction( | |||
4610 | Builder, TTI, ReductionOpcode, VectorizedValue, | |||
4611 | TargetTransformInfo::ReductionFlags(), RedOps); | |||
4612 | ||||
4613 | Value *TmpVec = VectorizedValue; | |||
4614 | for (unsigned i = ReduxWidth / 2; i != 0; i >>= 1) { | |||
4615 | Value *LeftMask = | |||
4616 | createRdxShuffleMask(ReduxWidth, i, true, true, Builder); | |||
4617 | Value *RightMask = | |||
4618 | createRdxShuffleMask(ReduxWidth, i, true, false, Builder); | |||
4619 | ||||
4620 | Value *LeftShuf = Builder.CreateShuffleVector( | |||
4621 | TmpVec, UndefValue::get(TmpVec->getType()), LeftMask, "rdx.shuf.l"); | |||
4622 | Value *RightShuf = Builder.CreateShuffleVector( | |||
4623 | TmpVec, UndefValue::get(TmpVec->getType()), (RightMask), | |||
4624 | "rdx.shuf.r"); | |||
4625 | TmpVec = | |||
4626 | Builder.CreateBinOp(ReductionOpcode, LeftShuf, RightShuf, "bin.rdx"); | |||
4627 | propagateIRFlags(TmpVec, RedOps); | |||
4628 | } | |||
4629 | ||||
4630 | // The result is in the first element of the vector. | |||
4631 | return Builder.CreateExtractElement(TmpVec, Builder.getInt32(0)); | |||
4632 | } | |||
4633 | }; | |||
4634 | } // end anonymous namespace | |||
4635 | ||||
4636 | /// \brief Recognize construction of vectors like | |||
4637 | /// %ra = insertelement <4 x float> undef, float %s0, i32 0 | |||
4638 | /// %rb = insertelement <4 x float> %ra, float %s1, i32 1 | |||
4639 | /// %rc = insertelement <4 x float> %rb, float %s2, i32 2 | |||
4640 | /// %rd = insertelement <4 x float> %rc, float %s3, i32 3 | |||
4641 | /// | |||
4642 | /// Returns true if it matches | |||
4643 | /// | |||
4644 | static bool findBuildVector(InsertElementInst *FirstInsertElem, | |||
4645 | SmallVectorImpl<Value *> &BuildVector, | |||
4646 | SmallVectorImpl<Value *> &BuildVectorOpds) { | |||
4647 | if (!isa<UndefValue>(FirstInsertElem->getOperand(0))) | |||
4648 | return false; | |||
4649 | ||||
4650 | InsertElementInst *IE = FirstInsertElem; | |||
4651 | while (true) { | |||
4652 | BuildVector.push_back(IE); | |||
4653 | BuildVectorOpds.push_back(IE->getOperand(1)); | |||
4654 | ||||
4655 | if (IE->use_empty()) | |||
4656 | return false; | |||
4657 | ||||
4658 | InsertElementInst *NextUse = dyn_cast<InsertElementInst>(IE->user_back()); | |||
4659 | if (!NextUse) | |||
4660 | return true; | |||
4661 | ||||
4662 | // If this isn't the final use, make sure the next insertelement is the only | |||
4663 | // use. It's OK if the final constructed vector is used multiple times | |||
4664 | if (!IE->hasOneUse()) | |||
4665 | return false; | |||
4666 | ||||
4667 | IE = NextUse; | |||
4668 | } | |||
4669 | ||||
4670 | return false; | |||
4671 | } | |||
4672 | ||||
4673 | /// \brief Like findBuildVector, but looks backwards for construction of aggregate. | |||
4674 | /// | |||
4675 | /// \return true if it matches. | |||
4676 | static bool findBuildAggregate(InsertValueInst *IV, | |||
4677 | SmallVectorImpl<Value *> &BuildVector, | |||
4678 | SmallVectorImpl<Value *> &BuildVectorOpds) { | |||
4679 | Value *V; | |||
4680 | do { | |||
4681 | BuildVector.push_back(IV); | |||
4682 | BuildVectorOpds.push_back(IV->getInsertedValueOperand()); | |||
4683 | V = IV->getAggregateOperand(); | |||
4684 | if (isa<UndefValue>(V)) | |||
4685 | break; | |||
4686 | IV = dyn_cast<InsertValueInst>(V); | |||
4687 | if (!IV || !IV->hasOneUse()) | |||
4688 | return false; | |||
4689 | } while (true); | |||
4690 | std::reverse(BuildVector.begin(), BuildVector.end()); | |||
4691 | std::reverse(BuildVectorOpds.begin(), BuildVectorOpds.end()); | |||
4692 | return true; | |||
4693 | } | |||
4694 | ||||
4695 | static bool PhiTypeSorterFunc(Value *V, Value *V2) { | |||
4696 | return V->getType() < V2->getType(); | |||
4697 | } | |||
4698 | ||||
4699 | /// \brief Try and get a reduction value from a phi node. | |||
4700 | /// | |||
4701 | /// Given a phi node \p P in a block \p ParentBB, consider possible reductions | |||
4702 | /// if they come from either \p ParentBB or a containing loop latch. | |||
4703 | /// | |||
4704 | /// \returns A candidate reduction value if possible, or \code nullptr \endcode | |||
4705 | /// if not possible. | |||
4706 | static Value *getReductionValue(const DominatorTree *DT, PHINode *P, | |||
4707 | BasicBlock *ParentBB, LoopInfo *LI) { | |||
4708 | // There are situations where the reduction value is not dominated by the | |||
4709 | // reduction phi. Vectorizing such cases has been reported to cause | |||
4710 | // miscompiles. See PR25787. | |||
4711 | auto DominatedReduxValue = [&](Value *R) { | |||
4712 | return ( | |||
4713 | dyn_cast<Instruction>(R) && | |||
4714 | DT->dominates(P->getParent(), dyn_cast<Instruction>(R)->getParent())); | |||
4715 | }; | |||
4716 | ||||
4717 | Value *Rdx = nullptr; | |||
4718 | ||||
4719 | // Return the incoming value if it comes from the same BB as the phi node. | |||
4720 | if (P->getIncomingBlock(0) == ParentBB) { | |||
4721 | Rdx = P->getIncomingValue(0); | |||
4722 | } else if (P->getIncomingBlock(1) == ParentBB) { | |||
4723 | Rdx = P->getIncomingValue(1); | |||
4724 | } | |||
4725 | ||||
4726 | if (Rdx && DominatedReduxValue(Rdx)) | |||
4727 | return Rdx; | |||
4728 | ||||
4729 | // Otherwise, check whether we have a loop latch to look at. | |||
4730 | Loop *BBL = LI->getLoopFor(ParentBB); | |||
4731 | if (!BBL) | |||
4732 | return nullptr; | |||
4733 | BasicBlock *BBLatch = BBL->getLoopLatch(); | |||
4734 | if (!BBLatch) | |||
4735 | return nullptr; | |||
4736 | ||||
4737 | // There is a loop latch, return the incoming value if it comes from | |||
4738 | // that. This reduction pattern occasionally turns up. | |||
4739 | if (P->getIncomingBlock(0) == BBLatch) { | |||
4740 | Rdx = P->getIncomingValue(0); | |||
4741 | } else if (P->getIncomingBlock(1) == BBLatch) { | |||
4742 | Rdx = P->getIncomingValue(1); | |||
4743 | } | |||
4744 | ||||
4745 | if (Rdx && DominatedReduxValue(Rdx)) | |||
4746 | return Rdx; | |||
4747 | ||||
4748 | return nullptr; | |||
4749 | } | |||
4750 | ||||
4751 | /// Attempt to reduce a horizontal reduction. | |||
4752 | /// If it is legal to match a horizontal reduction feeding the phi node \a P | |||
4753 | /// with reduction operators \a Root (or one of its operands) in a basic block | |||
4754 | /// \a BB, then check if it can be done. If horizontal reduction is not found | |||
4755 | /// and root instruction is a binary operation, vectorization of the operands is | |||
4756 | /// attempted. | |||
4757 | /// \returns true if a horizontal reduction was matched and reduced or operands | |||
4758 | /// of one of the binary instruction were vectorized. | |||
4759 | /// \returns false if a horizontal reduction was not matched (or not possible) | |||
4760 | /// or no vectorization of any binary operation feeding \a Root instruction was | |||
4761 | /// performed. | |||
4762 | static bool tryToVectorizeHorReductionOrInstOperands( | |||
4763 | PHINode *P, Instruction *Root, BasicBlock *BB, BoUpSLP &R, | |||
4764 | TargetTransformInfo *TTI, | |||
4765 | const function_ref<bool(BinaryOperator *, BoUpSLP &)> Vectorize) { | |||
4766 | if (!ShouldVectorizeHor) | |||
4767 | return false; | |||
4768 | ||||
4769 | if (!Root) | |||
4770 | return false; | |||
4771 | ||||
4772 | if (Root->getParent() != BB) | |||
4773 | return false; | |||
4774 | // Start analysis starting from Root instruction. If horizontal reduction is | |||
4775 | // found, try to vectorize it. If it is not a horizontal reduction or | |||
4776 | // vectorization is not possible or not effective, and currently analyzed | |||
4777 | // instruction is a binary operation, try to vectorize the operands, using | |||
4778 | // pre-order DFS traversal order. If the operands were not vectorized, repeat | |||
4779 | // the same procedure considering each operand as a possible root of the | |||
4780 | // horizontal reduction. | |||
4781 | // Interrupt the process if the Root instruction itself was vectorized or all | |||
4782 | // sub-trees not higher that RecursionMaxDepth were analyzed/vectorized. | |||
4783 | SmallVector<std::pair<WeakTrackingVH, unsigned>, 8> Stack(1, {Root, 0}); | |||
4784 | SmallSet<Value *, 8> VisitedInstrs; | |||
4785 | bool Res = false; | |||
4786 | while (!Stack.empty()) { | |||
4787 | Value *V; | |||
4788 | unsigned Level; | |||
4789 | std::tie(V, Level) = Stack.pop_back_val(); | |||
4790 | if (!V) | |||
4791 | continue; | |||
4792 | auto *Inst = dyn_cast<Instruction>(V); | |||
4793 | if (!Inst || isa<PHINode>(Inst)) | |||
4794 | continue; | |||
4795 | if (auto *BI = dyn_cast<BinaryOperator>(Inst)) { | |||
4796 | HorizontalReduction HorRdx; | |||
4797 | if (HorRdx.matchAssociativeReduction(P, BI)) { | |||
4798 | if (HorRdx.tryToReduce(R, TTI)) { | |||
4799 | Res = true; | |||
4800 | // Set P to nullptr to avoid re-analysis of phi node in | |||
4801 | // matchAssociativeReduction function unless this is the root node. | |||
4802 | P = nullptr; | |||
4803 | continue; | |||
4804 | } | |||
4805 | } | |||
4806 | if (P) { | |||
4807 | Inst = dyn_cast<Instruction>(BI->getOperand(0)); | |||
4808 | if (Inst == P) | |||
4809 | Inst = dyn_cast<Instruction>(BI->getOperand(1)); | |||
4810 | if (!Inst) { | |||
4811 | // Set P to nullptr to avoid re-analysis of phi node in | |||
4812 | // matchAssociativeReduction function unless this is the root node. | |||
4813 | P = nullptr; | |||
4814 | continue; | |||
4815 | } | |||
4816 | } | |||
4817 | } | |||
4818 | // Set P to nullptr to avoid re-analysis of phi node in | |||
4819 | // matchAssociativeReduction function unless this is the root node. | |||
4820 | P = nullptr; | |||
4821 | if (Vectorize(dyn_cast<BinaryOperator>(Inst), R)) { | |||
4822 | Res = true; | |||
4823 | continue; | |||
4824 | } | |||
4825 | ||||
4826 | // Try to vectorize operands. | |||
4827 | if (++Level < RecursionMaxDepth) | |||
4828 | for (auto *Op : Inst->operand_values()) | |||
4829 | Stack.emplace_back(Op, Level); | |||
4830 | } | |||
4831 | return Res; | |||
4832 | } | |||
4833 | ||||
4834 | bool SLPVectorizerPass::vectorizeRootInstruction(PHINode *P, Value *V, | |||
4835 | BasicBlock *BB, BoUpSLP &R, | |||
4836 | TargetTransformInfo *TTI) { | |||
4837 | if (!V) | |||
4838 | return false; | |||
4839 | auto *I = dyn_cast<Instruction>(V); | |||
4840 | if (!I) | |||
4841 | return false; | |||
4842 | ||||
4843 | if (!isa<BinaryOperator>(I)) | |||
4844 | P = nullptr; | |||
4845 | // Try to match and vectorize a horizontal reduction. | |||
4846 | return tryToVectorizeHorReductionOrInstOperands( | |||
4847 | P, I, BB, R, TTI, [this](BinaryOperator *BI, BoUpSLP &R) -> bool { | |||
4848 | return tryToVectorize(BI, R); | |||
4849 | }); | |||
4850 | } | |||
4851 | ||||
4852 | bool SLPVectorizerPass::vectorizeChainsInBlock(BasicBlock *BB, BoUpSLP &R) { | |||
4853 | bool Changed = false; | |||
4854 | SmallVector<Value *, 4> Incoming; | |||
4855 | SmallSet<Value *, 16> VisitedInstrs; | |||
4856 | ||||
4857 | bool HaveVectorizedPhiNodes = true; | |||
4858 | while (HaveVectorizedPhiNodes) { | |||
4859 | HaveVectorizedPhiNodes = false; | |||
4860 | ||||
4861 | // Collect the incoming values from the PHIs. | |||
4862 | Incoming.clear(); | |||
4863 | for (Instruction &I : *BB) { | |||
4864 | PHINode *P = dyn_cast<PHINode>(&I); | |||
4865 | if (!P) | |||
4866 | break; | |||
4867 | ||||
4868 | if (!VisitedInstrs.count(P)) | |||
4869 | Incoming.push_back(P); | |||
4870 | } | |||
4871 | ||||
4872 | // Sort by type. | |||
4873 | std::stable_sort(Incoming.begin(), Incoming.end(), PhiTypeSorterFunc); | |||
4874 | ||||
4875 | // Try to vectorize elements base on their type. | |||
4876 | for (SmallVector<Value *, 4>::iterator IncIt = Incoming.begin(), | |||
4877 | E = Incoming.end(); | |||
4878 | IncIt != E;) { | |||
4879 | ||||
4880 | // Look for the next elements with the same type. | |||
4881 | SmallVector<Value *, 4>::iterator SameTypeIt = IncIt; | |||
4882 | while (SameTypeIt != E && | |||
4883 | (*SameTypeIt)->getType() == (*IncIt)->getType()) { | |||
4884 | VisitedInstrs.insert(*SameTypeIt); | |||
4885 | ++SameTypeIt; | |||
4886 | } | |||
4887 | ||||
4888 | // Try to vectorize them. | |||
4889 | unsigned NumElts = (SameTypeIt - IncIt); | |||
4890 | DEBUG(errs() << "SLP: Trying to vectorize starting at PHIs (" << NumElts << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { errs() << "SLP: Trying to vectorize starting at PHIs (" << NumElts << ")\n"; } } while (false); | |||
4891 | // The order in which the phi nodes appear in the program does not matter. | |||
4892 | // So allow tryToVectorizeList to reorder them if it is beneficial. This | |||
4893 | // is done when there are exactly two elements since tryToVectorizeList | |||
4894 | // asserts that there are only two values when AllowReorder is true. | |||
4895 | bool AllowReorder = NumElts == 2; | |||
4896 | if (NumElts > 1 && tryToVectorizeList(makeArrayRef(IncIt, NumElts), R, | |||
4897 | None, AllowReorder)) { | |||
4898 | // Success start over because instructions might have been changed. | |||
4899 | HaveVectorizedPhiNodes = true; | |||
4900 | Changed = true; | |||
4901 | break; | |||
4902 | } | |||
4903 | ||||
4904 | // Start over at the next instruction of a different type (or the end). | |||
4905 | IncIt = SameTypeIt; | |||
4906 | } | |||
4907 | } | |||
4908 | ||||
4909 | VisitedInstrs.clear(); | |||
4910 | ||||
4911 | for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; it++) { | |||
4912 | // We may go through BB multiple times so skip the one we have checked. | |||
4913 | if (!VisitedInstrs.insert(&*it).second) | |||
4914 | continue; | |||
4915 | ||||
4916 | if (isa<DbgInfoIntrinsic>(it)) | |||
4917 | continue; | |||
4918 | ||||
4919 | // Try to vectorize reductions that use PHINodes. | |||
4920 | if (PHINode *P = dyn_cast<PHINode>(it)) { | |||
4921 | // Check that the PHI is a reduction PHI. | |||
4922 | if (P->getNumIncomingValues() != 2) | |||
4923 | return Changed; | |||
4924 | ||||
4925 | // Try to match and vectorize a horizontal reduction. | |||
4926 | if (vectorizeRootInstruction(P, getReductionValue(DT, P, BB, LI), BB, R, | |||
4927 | TTI)) { | |||
4928 | Changed = true; | |||
4929 | it = BB->begin(); | |||
4930 | e = BB->end(); | |||
4931 | continue; | |||
4932 | } | |||
4933 | continue; | |||
4934 | } | |||
4935 | ||||
4936 | if (ShouldStartVectorizeHorAtStore) { | |||
4937 | if (StoreInst *SI = dyn_cast<StoreInst>(it)) { | |||
4938 | // Try to match and vectorize a horizontal reduction. | |||
4939 | if (vectorizeRootInstruction(nullptr, SI->getValueOperand(), BB, R, | |||
4940 | TTI)) { | |||
4941 | Changed = true; | |||
4942 | it = BB->begin(); | |||
4943 | e = BB->end(); | |||
4944 | continue; | |||
4945 | } | |||
4946 | } | |||
4947 | } | |||
4948 | ||||
4949 | // Try to vectorize horizontal reductions feeding into a return. | |||
4950 | if (ReturnInst *RI = dyn_cast<ReturnInst>(it)) { | |||
4951 | if (RI->getNumOperands() != 0) { | |||
4952 | // Try to match and vectorize a horizontal reduction. | |||
4953 | if (vectorizeRootInstruction(nullptr, RI->getOperand(0), BB, R, TTI)) { | |||
4954 | Changed = true; | |||
4955 | it = BB->begin(); | |||
4956 | e = BB->end(); | |||
4957 | continue; | |||
4958 | } | |||
4959 | } | |||
4960 | } | |||
4961 | ||||
4962 | // Try to vectorize trees that start at compare instructions. | |||
4963 | if (CmpInst *CI = dyn_cast<CmpInst>(it)) { | |||
4964 | if (tryToVectorizePair(CI->getOperand(0), CI->getOperand(1), R)) { | |||
4965 | Changed = true; | |||
4966 | // We would like to start over since some instructions are deleted | |||
4967 | // and the iterator may become invalid value. | |||
4968 | it = BB->begin(); | |||
4969 | e = BB->end(); | |||
4970 | continue; | |||
4971 | } | |||
4972 | ||||
4973 | for (int I = 0; I < 2; ++I) { | |||
4974 | if (vectorizeRootInstruction(nullptr, CI->getOperand(I), BB, R, TTI)) { | |||
4975 | Changed = true; | |||
4976 | // We would like to start over since some instructions are deleted | |||
4977 | // and the iterator may become invalid value. | |||
4978 | it = BB->begin(); | |||
4979 | e = BB->end(); | |||
4980 | break; | |||
4981 | } | |||
4982 | } | |||
4983 | continue; | |||
4984 | } | |||
4985 | ||||
4986 | // Try to vectorize trees that start at insertelement instructions. | |||
4987 | if (InsertElementInst *FirstInsertElem = dyn_cast<InsertElementInst>(it)) { | |||
4988 | SmallVector<Value *, 16> BuildVector; | |||
4989 | SmallVector<Value *, 16> BuildVectorOpds; | |||
4990 | if (!findBuildVector(FirstInsertElem, BuildVector, BuildVectorOpds)) | |||
4991 | continue; | |||
4992 | ||||
4993 | // Vectorize starting with the build vector operands ignoring the | |||
4994 | // BuildVector instructions for the purpose of scheduling and user | |||
4995 | // extraction. | |||
4996 | if (tryToVectorizeList(BuildVectorOpds, R, BuildVector)) { | |||
4997 | Changed = true; | |||
4998 | it = BB->begin(); | |||
4999 | e = BB->end(); | |||
5000 | } | |||
5001 | ||||
5002 | continue; | |||
5003 | } | |||
5004 | ||||
5005 | // Try to vectorize trees that start at insertvalue instructions feeding into | |||
5006 | // a store. | |||
5007 | if (StoreInst *SI = dyn_cast<StoreInst>(it)) { | |||
5008 | if (InsertValueInst *LastInsertValue = dyn_cast<InsertValueInst>(SI->getValueOperand())) { | |||
5009 | const DataLayout &DL = BB->getModule()->getDataLayout(); | |||
5010 | if (R.canMapToVector(SI->getValueOperand()->getType(), DL)) { | |||
5011 | SmallVector<Value *, 16> BuildVector; | |||
5012 | SmallVector<Value *, 16> BuildVectorOpds; | |||
5013 | if (!findBuildAggregate(LastInsertValue, BuildVector, BuildVectorOpds)) | |||
5014 | continue; | |||
5015 | ||||
5016 | DEBUG(dbgs() << "SLP: store of array mappable to vector: " << *SI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: store of array mappable to vector: " << *SI << "\n"; } } while (false); | |||
5017 | if (tryToVectorizeList(BuildVectorOpds, R, BuildVector, false)) { | |||
5018 | Changed = true; | |||
5019 | it = BB->begin(); | |||
5020 | e = BB->end(); | |||
5021 | } | |||
5022 | continue; | |||
5023 | } | |||
5024 | } | |||
5025 | } | |||
5026 | } | |||
5027 | ||||
5028 | return Changed; | |||
5029 | } | |||
5030 | ||||
5031 | bool SLPVectorizerPass::vectorizeGEPIndices(BasicBlock *BB, BoUpSLP &R) { | |||
5032 | auto Changed = false; | |||
5033 | for (auto &Entry : GEPs) { | |||
5034 | ||||
5035 | // If the getelementptr list has fewer than two elements, there's nothing | |||
5036 | // to do. | |||
5037 | if (Entry.second.size() < 2) | |||
5038 | continue; | |||
5039 | ||||
5040 | 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 ) | |||
5041 | << 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 ); | |||
5042 | ||||
5043 | // We process the getelementptr list in chunks of 16 (like we do for | |||
5044 | // stores) to minimize compile-time. | |||
5045 | for (unsigned BI = 0, BE = Entry.second.size(); BI < BE; BI += 16) { | |||
5046 | auto Len = std::min<unsigned>(BE - BI, 16); | |||
5047 | auto GEPList = makeArrayRef(&Entry.second[BI], Len); | |||
5048 | ||||
5049 | // Initialize a set a candidate getelementptrs. Note that we use a | |||
5050 | // SetVector here to preserve program order. If the index computations | |||
5051 | // are vectorizable and begin with loads, we want to minimize the chance | |||
5052 | // of having to reorder them later. | |||
5053 | SetVector<Value *> Candidates(GEPList.begin(), GEPList.end()); | |||
5054 | ||||
5055 | // Some of the candidates may have already been vectorized after we | |||
5056 | // initially collected them. If so, the WeakTrackingVHs will have | |||
5057 | // nullified the | |||
5058 | // values, so remove them from the set of candidates. | |||
5059 | Candidates.remove(nullptr); | |||
5060 | ||||
5061 | // Remove from the set of candidates all pairs of getelementptrs with | |||
5062 | // constant differences. Such getelementptrs are likely not good | |||
5063 | // candidates for vectorization in a bottom-up phase since one can be | |||
5064 | // computed from the other. We also ensure all candidate getelementptr | |||
5065 | // indices are unique. | |||
5066 | for (int I = 0, E = GEPList.size(); I < E && Candidates.size() > 1; ++I) { | |||
5067 | auto *GEPI = cast<GetElementPtrInst>(GEPList[I]); | |||
5068 | if (!Candidates.count(GEPI)) | |||
5069 | continue; | |||
5070 | auto *SCEVI = SE->getSCEV(GEPList[I]); | |||
5071 | for (int J = I + 1; J < E && Candidates.size() > 1; ++J) { | |||
5072 | auto *GEPJ = cast<GetElementPtrInst>(GEPList[J]); | |||
5073 | auto *SCEVJ = SE->getSCEV(GEPList[J]); | |||
5074 | if (isa<SCEVConstant>(SE->getMinusSCEV(SCEVI, SCEVJ))) { | |||
5075 | Candidates.remove(GEPList[I]); | |||
5076 | Candidates.remove(GEPList[J]); | |||
5077 | } else if (GEPI->idx_begin()->get() == GEPJ->idx_begin()->get()) { | |||
5078 | Candidates.remove(GEPList[J]); | |||
5079 | } | |||
5080 | } | |||
5081 | } | |||
5082 | ||||
5083 | // We break out of the above computation as soon as we know there are | |||
5084 | // fewer than two candidates remaining. | |||
5085 | if (Candidates.size() < 2) | |||
5086 | continue; | |||
5087 | ||||
5088 | // Add the single, non-constant index of each candidate to the bundle. We | |||
5089 | // ensured the indices met these constraints when we originally collected | |||
5090 | // the getelementptrs. | |||
5091 | SmallVector<Value *, 16> Bundle(Candidates.size()); | |||
5092 | auto BundleIndex = 0u; | |||
5093 | for (auto *V : Candidates) { | |||
5094 | auto *GEP = cast<GetElementPtrInst>(V); | |||
5095 | auto *GEPIdx = GEP->idx_begin()->get(); | |||
5096 | assert(GEP->getNumIndices() == 1 || !isa<Constant>(GEPIdx))((GEP->getNumIndices() == 1 || !isa<Constant>(GEPIdx )) ? static_cast<void> (0) : __assert_fail ("GEP->getNumIndices() == 1 || !isa<Constant>(GEPIdx)" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5096, __PRETTY_FUNCTION__)); | |||
5097 | Bundle[BundleIndex++] = GEPIdx; | |||
5098 | } | |||
5099 | ||||
5100 | // Try and vectorize the indices. We are currently only interested in | |||
5101 | // gather-like cases of the form: | |||
5102 | // | |||
5103 | // ... = g[a[0] - b[0]] + g[a[1] - b[1]] + ... | |||
5104 | // | |||
5105 | // where the loads of "a", the loads of "b", and the subtractions can be | |||
5106 | // performed in parallel. It's likely that detecting this pattern in a | |||
5107 | // bottom-up phase will be simpler and less costly than building a | |||
5108 | // full-blown top-down phase beginning at the consecutive loads. | |||
5109 | Changed |= tryToVectorizeList(Bundle, R); | |||
5110 | } | |||
5111 | } | |||
5112 | return Changed; | |||
5113 | } | |||
5114 | ||||
5115 | bool SLPVectorizerPass::vectorizeStoreChains(BoUpSLP &R) { | |||
5116 | bool Changed = false; | |||
5117 | // Attempt to sort and vectorize each of the store-groups. | |||
5118 | for (StoreListMap::iterator it = Stores.begin(), e = Stores.end(); it != e; | |||
5119 | ++it) { | |||
5120 | if (it->second.size() < 2) | |||
5121 | continue; | |||
5122 | ||||
5123 | 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 ) | |||
5124 | << 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 ); | |||
5125 | ||||
5126 | // Process the stores in chunks of 16. | |||
5127 | // TODO: The limit of 16 inhibits greater vectorization factors. | |||
5128 | // For example, AVX2 supports v32i8. Increasing this limit, however, | |||
5129 | // may cause a significant compile-time increase. | |||
5130 | for (unsigned CI = 0, CE = it->second.size(); CI < CE; CI+=16) { | |||
5131 | unsigned Len = std::min<unsigned>(CE - CI, 16); | |||
5132 | Changed |= vectorizeStores(makeArrayRef(&it->second[CI], Len), R); | |||
5133 | } | |||
5134 | } | |||
5135 | return Changed; | |||
5136 | } | |||
5137 | ||||
5138 | char SLPVectorizer::ID = 0; | |||
5139 | static const char lv_name[] = "SLP Vectorizer"; | |||
5140 | INITIALIZE_PASS_BEGIN(SLPVectorizer, SV_NAME, lv_name, false, false)static void *initializeSLPVectorizerPassOnce(PassRegistry & Registry) { | |||
5141 | INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)initializeAAResultsWrapperPassPass(Registry); | |||
5142 | INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)initializeTargetTransformInfoWrapperPassPass(Registry); | |||
5143 | INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)initializeAssumptionCacheTrackerPass(Registry); | |||
5144 | INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)initializeScalarEvolutionWrapperPassPass(Registry); | |||
5145 | INITIALIZE_PASS_DEPENDENCY(LoopSimplify)initializeLoopSimplifyPass(Registry); | |||
5146 | INITIALIZE_PASS_DEPENDENCY(DemandedBitsWrapperPass)initializeDemandedBitsWrapperPassPass(Registry); | |||
5147 | INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass)initializeOptimizationRemarkEmitterWrapperPassPass(Registry); | |||
5148 | 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)); } | |||
5149 | ||||
5150 | namespace llvm { | |||
5151 | Pass *createSLPVectorizerPass() { return new SLPVectorizer(); } | |||
5152 | } |