1 //===- BBVectorize.cpp - A Basic-Block Vectorizer -------------------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file implements a basic-block vectorization pass. The algorithm was
11 // inspired by that used by the Vienna MAP Vectorizor by Franchetti and Kral,
12 // et al. It works by looking for chains of pairable operations and then
13 // pairing them.
14 //
15 //===----------------------------------------------------------------------===//
16
17 #define BBV_NAME "bb-vectorize"
18 #include "llvm/Transforms/Vectorize.h"
19 #include "llvm/ADT/DenseMap.h"
20 #include "llvm/ADT/DenseSet.h"
21 #include "llvm/ADT/STLExtras.h"
22 #include "llvm/ADT/SmallSet.h"
23 #include "llvm/ADT/SmallVector.h"
24 #include "llvm/ADT/Statistic.h"
25 #include "llvm/ADT/StringExtras.h"
26 #include "llvm/Analysis/AliasAnalysis.h"
27 #include "llvm/Analysis/AliasSetTracker.h"
28 #include "llvm/Analysis/GlobalsModRef.h"
29 #include "llvm/Analysis/ScalarEvolution.h"
30 #include "llvm/Analysis/ScalarEvolutionAliasAnalysis.h"
31 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
32 #include "llvm/Analysis/TargetLibraryInfo.h"
33 #include "llvm/Analysis/TargetTransformInfo.h"
34 #include "llvm/Analysis/ValueTracking.h"
35 #include "llvm/IR/Constants.h"
36 #include "llvm/IR/DataLayout.h"
37 #include "llvm/IR/DerivedTypes.h"
38 #include "llvm/IR/Dominators.h"
39 #include "llvm/IR/Function.h"
40 #include "llvm/IR/Instructions.h"
41 #include "llvm/IR/IntrinsicInst.h"
42 #include "llvm/IR/Intrinsics.h"
43 #include "llvm/IR/LLVMContext.h"
44 #include "llvm/IR/Metadata.h"
45 #include "llvm/IR/Module.h"
46 #include "llvm/IR/Type.h"
47 #include "llvm/IR/ValueHandle.h"
48 #include "llvm/Pass.h"
49 #include "llvm/Support/CommandLine.h"
50 #include "llvm/Support/Debug.h"
51 #include "llvm/Support/raw_ostream.h"
52 #include "llvm/Transforms/Utils/Local.h"
53 #include <algorithm>
54 using namespace llvm;
55
56 #define DEBUG_TYPE BBV_NAME
57
58 static cl::opt<bool>
59 IgnoreTargetInfo("bb-vectorize-ignore-target-info", cl::init(false),
60 cl::Hidden, cl::desc("Ignore target information"));
61
62 static cl::opt<unsigned>
63 ReqChainDepth("bb-vectorize-req-chain-depth", cl::init(6), cl::Hidden,
64 cl::desc("The required chain depth for vectorization"));
65
66 static cl::opt<bool>
67 UseChainDepthWithTI("bb-vectorize-use-chain-depth", cl::init(false),
68 cl::Hidden, cl::desc("Use the chain depth requirement with"
69 " target information"));
70
71 static cl::opt<unsigned>
72 SearchLimit("bb-vectorize-search-limit", cl::init(400), cl::Hidden,
73 cl::desc("The maximum search distance for instruction pairs"));
74
75 static cl::opt<bool>
76 SplatBreaksChain("bb-vectorize-splat-breaks-chain", cl::init(false), cl::Hidden,
77 cl::desc("Replicating one element to a pair breaks the chain"));
78
79 static cl::opt<unsigned>
80 VectorBits("bb-vectorize-vector-bits", cl::init(128), cl::Hidden,
81 cl::desc("The size of the native vector registers"));
82
83 static cl::opt<unsigned>
84 MaxIter("bb-vectorize-max-iter", cl::init(0), cl::Hidden,
85 cl::desc("The maximum number of pairing iterations"));
86
87 static cl::opt<bool>
88 Pow2LenOnly("bb-vectorize-pow2-len-only", cl::init(false), cl::Hidden,
89 cl::desc("Don't try to form non-2^n-length vectors"));
90
91 static cl::opt<unsigned>
92 MaxInsts("bb-vectorize-max-instr-per-group", cl::init(500), cl::Hidden,
93 cl::desc("The maximum number of pairable instructions per group"));
94
95 static cl::opt<unsigned>
96 MaxPairs("bb-vectorize-max-pairs-per-group", cl::init(3000), cl::Hidden,
97 cl::desc("The maximum number of candidate instruction pairs per group"));
98
99 static cl::opt<unsigned>
100 MaxCandPairsForCycleCheck("bb-vectorize-max-cycle-check-pairs", cl::init(200),
101 cl::Hidden, cl::desc("The maximum number of candidate pairs with which to use"
102 " a full cycle check"));
103
104 static cl::opt<bool>
105 NoBools("bb-vectorize-no-bools", cl::init(false), cl::Hidden,
106 cl::desc("Don't try to vectorize boolean (i1) values"));
107
108 static cl::opt<bool>
109 NoInts("bb-vectorize-no-ints", cl::init(false), cl::Hidden,
110 cl::desc("Don't try to vectorize integer values"));
111
112 static cl::opt<bool>
113 NoFloats("bb-vectorize-no-floats", cl::init(false), cl::Hidden,
114 cl::desc("Don't try to vectorize floating-point values"));
115
116 // FIXME: This should default to false once pointer vector support works.
117 static cl::opt<bool>
118 NoPointers("bb-vectorize-no-pointers", cl::init(/*false*/ true), cl::Hidden,
119 cl::desc("Don't try to vectorize pointer values"));
120
121 static cl::opt<bool>
122 NoCasts("bb-vectorize-no-casts", cl::init(false), cl::Hidden,
123 cl::desc("Don't try to vectorize casting (conversion) operations"));
124
125 static cl::opt<bool>
126 NoMath("bb-vectorize-no-math", cl::init(false), cl::Hidden,
127 cl::desc("Don't try to vectorize floating-point math intrinsics"));
128
129 static cl::opt<bool>
130 NoBitManipulation("bb-vectorize-no-bitmanip", cl::init(false), cl::Hidden,
131 cl::desc("Don't try to vectorize BitManipulation intrinsics"));
132
133 static cl::opt<bool>
134 NoFMA("bb-vectorize-no-fma", cl::init(false), cl::Hidden,
135 cl::desc("Don't try to vectorize the fused-multiply-add intrinsic"));
136
137 static cl::opt<bool>
138 NoSelect("bb-vectorize-no-select", cl::init(false), cl::Hidden,
139 cl::desc("Don't try to vectorize select instructions"));
140
141 static cl::opt<bool>
142 NoCmp("bb-vectorize-no-cmp", cl::init(false), cl::Hidden,
143 cl::desc("Don't try to vectorize comparison instructions"));
144
145 static cl::opt<bool>
146 NoGEP("bb-vectorize-no-gep", cl::init(false), cl::Hidden,
147 cl::desc("Don't try to vectorize getelementptr instructions"));
148
149 static cl::opt<bool>
150 NoMemOps("bb-vectorize-no-mem-ops", cl::init(false), cl::Hidden,
151 cl::desc("Don't try to vectorize loads and stores"));
152
153 static cl::opt<bool>
154 AlignedOnly("bb-vectorize-aligned-only", cl::init(false), cl::Hidden,
155 cl::desc("Only generate aligned loads and stores"));
156
157 static cl::opt<bool>
158 NoMemOpBoost("bb-vectorize-no-mem-op-boost",
159 cl::init(false), cl::Hidden,
160 cl::desc("Don't boost the chain-depth contribution of loads and stores"));
161
162 static cl::opt<bool>
163 FastDep("bb-vectorize-fast-dep", cl::init(false), cl::Hidden,
164 cl::desc("Use a fast instruction dependency analysis"));
165
166 #ifndef NDEBUG
167 static cl::opt<bool>
168 DebugInstructionExamination("bb-vectorize-debug-instruction-examination",
169 cl::init(false), cl::Hidden,
170 cl::desc("When debugging is enabled, output information on the"
171 " instruction-examination process"));
172 static cl::opt<bool>
173 DebugCandidateSelection("bb-vectorize-debug-candidate-selection",
174 cl::init(false), cl::Hidden,
175 cl::desc("When debugging is enabled, output information on the"
176 " candidate-selection process"));
177 static cl::opt<bool>
178 DebugPairSelection("bb-vectorize-debug-pair-selection",
179 cl::init(false), cl::Hidden,
180 cl::desc("When debugging is enabled, output information on the"
181 " pair-selection process"));
182 static cl::opt<bool>
183 DebugCycleCheck("bb-vectorize-debug-cycle-check",
184 cl::init(false), cl::Hidden,
185 cl::desc("When debugging is enabled, output information on the"
186 " cycle-checking process"));
187
188 static cl::opt<bool>
189 PrintAfterEveryPair("bb-vectorize-debug-print-after-every-pair",
190 cl::init(false), cl::Hidden,
191 cl::desc("When debugging is enabled, dump the basic block after"
192 " every pair is fused"));
193 #endif
194
195 STATISTIC(NumFusedOps, "Number of operations fused by bb-vectorize");
196
197 namespace {
198 struct BBVectorize : public BasicBlockPass {
199 static char ID; // Pass identification, replacement for typeid
200
201 const VectorizeConfig Config;
202
BBVectorize__anonbe9f7d990111::BBVectorize203 BBVectorize(const VectorizeConfig &C = VectorizeConfig())
204 : BasicBlockPass(ID), Config(C) {
205 initializeBBVectorizePass(*PassRegistry::getPassRegistry());
206 }
207
BBVectorize__anonbe9f7d990111::BBVectorize208 BBVectorize(Pass *P, Function &F, const VectorizeConfig &C)
209 : BasicBlockPass(ID), Config(C) {
210 AA = &P->getAnalysis<AAResultsWrapperPass>().getAAResults();
211 DT = &P->getAnalysis<DominatorTreeWrapperPass>().getDomTree();
212 SE = &P->getAnalysis<ScalarEvolutionWrapperPass>().getSE();
213 TLI = &P->getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
214 TTI = IgnoreTargetInfo
215 ? nullptr
216 : &P->getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
217 }
218
219 typedef std::pair<Value *, Value *> ValuePair;
220 typedef std::pair<ValuePair, int> ValuePairWithCost;
221 typedef std::pair<ValuePair, size_t> ValuePairWithDepth;
222 typedef std::pair<ValuePair, ValuePair> VPPair; // A ValuePair pair
223 typedef std::pair<VPPair, unsigned> VPPairWithType;
224
225 AliasAnalysis *AA;
226 DominatorTree *DT;
227 ScalarEvolution *SE;
228 const TargetLibraryInfo *TLI;
229 const TargetTransformInfo *TTI;
230
231 // FIXME: const correct?
232
233 bool vectorizePairs(BasicBlock &BB, bool NonPow2Len = false);
234
235 bool getCandidatePairs(BasicBlock &BB,
236 BasicBlock::iterator &Start,
237 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
238 DenseSet<ValuePair> &FixedOrderPairs,
239 DenseMap<ValuePair, int> &CandidatePairCostSavings,
240 std::vector<Value *> &PairableInsts, bool NonPow2Len);
241
242 // FIXME: The current implementation does not account for pairs that
243 // are connected in multiple ways. For example:
244 // C1 = A1 / A2; C2 = A2 / A1 (which may be both direct and a swap)
245 enum PairConnectionType {
246 PairConnectionDirect,
247 PairConnectionSwap,
248 PairConnectionSplat
249 };
250
251 void computeConnectedPairs(
252 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
253 DenseSet<ValuePair> &CandidatePairsSet,
254 std::vector<Value *> &PairableInsts,
255 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
256 DenseMap<VPPair, unsigned> &PairConnectionTypes);
257
258 void buildDepMap(BasicBlock &BB,
259 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
260 std::vector<Value *> &PairableInsts,
261 DenseSet<ValuePair> &PairableInstUsers);
262
263 void choosePairs(DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
264 DenseSet<ValuePair> &CandidatePairsSet,
265 DenseMap<ValuePair, int> &CandidatePairCostSavings,
266 std::vector<Value *> &PairableInsts,
267 DenseSet<ValuePair> &FixedOrderPairs,
268 DenseMap<VPPair, unsigned> &PairConnectionTypes,
269 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
270 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps,
271 DenseSet<ValuePair> &PairableInstUsers,
272 DenseMap<Value *, Value *>& ChosenPairs);
273
274 void fuseChosenPairs(BasicBlock &BB,
275 std::vector<Value *> &PairableInsts,
276 DenseMap<Value *, Value *>& ChosenPairs,
277 DenseSet<ValuePair> &FixedOrderPairs,
278 DenseMap<VPPair, unsigned> &PairConnectionTypes,
279 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
280 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps);
281
282
283 bool isInstVectorizable(Instruction *I, bool &IsSimpleLoadStore);
284
285 bool areInstsCompatible(Instruction *I, Instruction *J,
286 bool IsSimpleLoadStore, bool NonPow2Len,
287 int &CostSavings, int &FixedOrder);
288
289 bool trackUsesOfI(DenseSet<Value *> &Users,
290 AliasSetTracker &WriteSet, Instruction *I,
291 Instruction *J, bool UpdateUsers = true,
292 DenseSet<ValuePair> *LoadMoveSetPairs = nullptr);
293
294 void computePairsConnectedTo(
295 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
296 DenseSet<ValuePair> &CandidatePairsSet,
297 std::vector<Value *> &PairableInsts,
298 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
299 DenseMap<VPPair, unsigned> &PairConnectionTypes,
300 ValuePair P);
301
302 bool pairsConflict(ValuePair P, ValuePair Q,
303 DenseSet<ValuePair> &PairableInstUsers,
304 DenseMap<ValuePair, std::vector<ValuePair> >
305 *PairableInstUserMap = nullptr,
306 DenseSet<VPPair> *PairableInstUserPairSet = nullptr);
307
308 bool pairWillFormCycle(ValuePair P,
309 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUsers,
310 DenseSet<ValuePair> &CurrentPairs);
311
312 void pruneDAGFor(
313 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
314 std::vector<Value *> &PairableInsts,
315 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
316 DenseSet<ValuePair> &PairableInstUsers,
317 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap,
318 DenseSet<VPPair> &PairableInstUserPairSet,
319 DenseMap<Value *, Value *> &ChosenPairs,
320 DenseMap<ValuePair, size_t> &DAG,
321 DenseSet<ValuePair> &PrunedDAG, ValuePair J,
322 bool UseCycleCheck);
323
324 void buildInitialDAGFor(
325 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
326 DenseSet<ValuePair> &CandidatePairsSet,
327 std::vector<Value *> &PairableInsts,
328 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
329 DenseSet<ValuePair> &PairableInstUsers,
330 DenseMap<Value *, Value *> &ChosenPairs,
331 DenseMap<ValuePair, size_t> &DAG, ValuePair J);
332
333 void findBestDAGFor(
334 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
335 DenseSet<ValuePair> &CandidatePairsSet,
336 DenseMap<ValuePair, int> &CandidatePairCostSavings,
337 std::vector<Value *> &PairableInsts,
338 DenseSet<ValuePair> &FixedOrderPairs,
339 DenseMap<VPPair, unsigned> &PairConnectionTypes,
340 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
341 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps,
342 DenseSet<ValuePair> &PairableInstUsers,
343 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap,
344 DenseSet<VPPair> &PairableInstUserPairSet,
345 DenseMap<Value *, Value *> &ChosenPairs,
346 DenseSet<ValuePair> &BestDAG, size_t &BestMaxDepth,
347 int &BestEffSize, Value *II, std::vector<Value *>&JJ,
348 bool UseCycleCheck);
349
350 Value *getReplacementPointerInput(LLVMContext& Context, Instruction *I,
351 Instruction *J, unsigned o);
352
353 void fillNewShuffleMask(LLVMContext& Context, Instruction *J,
354 unsigned MaskOffset, unsigned NumInElem,
355 unsigned NumInElem1, unsigned IdxOffset,
356 std::vector<Constant*> &Mask);
357
358 Value *getReplacementShuffleMask(LLVMContext& Context, Instruction *I,
359 Instruction *J);
360
361 bool expandIEChain(LLVMContext& Context, Instruction *I, Instruction *J,
362 unsigned o, Value *&LOp, unsigned numElemL,
363 Type *ArgTypeL, Type *ArgTypeR, bool IBeforeJ,
364 unsigned IdxOff = 0);
365
366 Value *getReplacementInput(LLVMContext& Context, Instruction *I,
367 Instruction *J, unsigned o, bool IBeforeJ);
368
369 void getReplacementInputsForPair(LLVMContext& Context, Instruction *I,
370 Instruction *J, SmallVectorImpl<Value *> &ReplacedOperands,
371 bool IBeforeJ);
372
373 void replaceOutputsOfPair(LLVMContext& Context, Instruction *I,
374 Instruction *J, Instruction *K,
375 Instruction *&InsertionPt, Instruction *&K1,
376 Instruction *&K2);
377
378 void collectPairLoadMoveSet(BasicBlock &BB,
379 DenseMap<Value *, Value *> &ChosenPairs,
380 DenseMap<Value *, std::vector<Value *> > &LoadMoveSet,
381 DenseSet<ValuePair> &LoadMoveSetPairs,
382 Instruction *I);
383
384 void collectLoadMoveSet(BasicBlock &BB,
385 std::vector<Value *> &PairableInsts,
386 DenseMap<Value *, Value *> &ChosenPairs,
387 DenseMap<Value *, std::vector<Value *> > &LoadMoveSet,
388 DenseSet<ValuePair> &LoadMoveSetPairs);
389
390 bool canMoveUsesOfIAfterJ(BasicBlock &BB,
391 DenseSet<ValuePair> &LoadMoveSetPairs,
392 Instruction *I, Instruction *J);
393
394 void moveUsesOfIAfterJ(BasicBlock &BB,
395 DenseSet<ValuePair> &LoadMoveSetPairs,
396 Instruction *&InsertionPt,
397 Instruction *I, Instruction *J);
398
vectorizeBB__anonbe9f7d990111::BBVectorize399 bool vectorizeBB(BasicBlock &BB) {
400 if (skipOptnoneFunction(BB))
401 return false;
402 if (!DT->isReachableFromEntry(&BB)) {
403 DEBUG(dbgs() << "BBV: skipping unreachable " << BB.getName() <<
404 " in " << BB.getParent()->getName() << "\n");
405 return false;
406 }
407
408 DEBUG(if (TTI) dbgs() << "BBV: using target information\n");
409
410 bool changed = false;
411 // Iterate a sufficient number of times to merge types of size 1 bit,
412 // then 2 bits, then 4, etc. up to half of the target vector width of the
413 // target vector register.
414 unsigned n = 1;
415 for (unsigned v = 2;
416 (TTI || v <= Config.VectorBits) &&
417 (!Config.MaxIter || n <= Config.MaxIter);
418 v *= 2, ++n) {
419 DEBUG(dbgs() << "BBV: fusing loop #" << n <<
420 " for " << BB.getName() << " in " <<
421 BB.getParent()->getName() << "...\n");
422 if (vectorizePairs(BB))
423 changed = true;
424 else
425 break;
426 }
427
428 if (changed && !Pow2LenOnly) {
429 ++n;
430 for (; !Config.MaxIter || n <= Config.MaxIter; ++n) {
431 DEBUG(dbgs() << "BBV: fusing for non-2^n-length vectors loop #: " <<
432 n << " for " << BB.getName() << " in " <<
433 BB.getParent()->getName() << "...\n");
434 if (!vectorizePairs(BB, true)) break;
435 }
436 }
437
438 DEBUG(dbgs() << "BBV: done!\n");
439 return changed;
440 }
441
runOnBasicBlock__anonbe9f7d990111::BBVectorize442 bool runOnBasicBlock(BasicBlock &BB) override {
443 // OptimizeNone check deferred to vectorizeBB().
444
445 AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
446 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
447 SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
448 TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
449 TTI = IgnoreTargetInfo
450 ? nullptr
451 : &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(
452 *BB.getParent());
453
454 return vectorizeBB(BB);
455 }
456
getAnalysisUsage__anonbe9f7d990111::BBVectorize457 void getAnalysisUsage(AnalysisUsage &AU) const override {
458 BasicBlockPass::getAnalysisUsage(AU);
459 AU.addRequired<AAResultsWrapperPass>();
460 AU.addRequired<DominatorTreeWrapperPass>();
461 AU.addRequired<ScalarEvolutionWrapperPass>();
462 AU.addRequired<TargetLibraryInfoWrapperPass>();
463 AU.addRequired<TargetTransformInfoWrapperPass>();
464 AU.addPreserved<DominatorTreeWrapperPass>();
465 AU.addPreserved<GlobalsAAWrapperPass>();
466 AU.addPreserved<ScalarEvolutionWrapperPass>();
467 AU.addPreserved<SCEVAAWrapperPass>();
468 AU.setPreservesCFG();
469 }
470
getVecTypeForPair__anonbe9f7d990111::BBVectorize471 static inline VectorType *getVecTypeForPair(Type *ElemTy, Type *Elem2Ty) {
472 assert(ElemTy->getScalarType() == Elem2Ty->getScalarType() &&
473 "Cannot form vector from incompatible scalar types");
474 Type *STy = ElemTy->getScalarType();
475
476 unsigned numElem;
477 if (VectorType *VTy = dyn_cast<VectorType>(ElemTy)) {
478 numElem = VTy->getNumElements();
479 } else {
480 numElem = 1;
481 }
482
483 if (VectorType *VTy = dyn_cast<VectorType>(Elem2Ty)) {
484 numElem += VTy->getNumElements();
485 } else {
486 numElem += 1;
487 }
488
489 return VectorType::get(STy, numElem);
490 }
491
getInstructionTypes__anonbe9f7d990111::BBVectorize492 static inline void getInstructionTypes(Instruction *I,
493 Type *&T1, Type *&T2) {
494 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
495 // For stores, it is the value type, not the pointer type that matters
496 // because the value is what will come from a vector register.
497
498 Value *IVal = SI->getValueOperand();
499 T1 = IVal->getType();
500 } else {
501 T1 = I->getType();
502 }
503
504 if (CastInst *CI = dyn_cast<CastInst>(I))
505 T2 = CI->getSrcTy();
506 else
507 T2 = T1;
508
509 if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
510 T2 = SI->getCondition()->getType();
511 } else if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(I)) {
512 T2 = SI->getOperand(0)->getType();
513 } else if (CmpInst *CI = dyn_cast<CmpInst>(I)) {
514 T2 = CI->getOperand(0)->getType();
515 }
516 }
517
518 // Returns the weight associated with the provided value. A chain of
519 // candidate pairs has a length given by the sum of the weights of its
520 // members (one weight per pair; the weight of each member of the pair
521 // is assumed to be the same). This length is then compared to the
522 // chain-length threshold to determine if a given chain is significant
523 // enough to be vectorized. The length is also used in comparing
524 // candidate chains where longer chains are considered to be better.
525 // Note: when this function returns 0, the resulting instructions are
526 // not actually fused.
getDepthFactor__anonbe9f7d990111::BBVectorize527 inline size_t getDepthFactor(Value *V) {
528 // InsertElement and ExtractElement have a depth factor of zero. This is
529 // for two reasons: First, they cannot be usefully fused. Second, because
530 // the pass generates a lot of these, they can confuse the simple metric
531 // used to compare the dags in the next iteration. Thus, giving them a
532 // weight of zero allows the pass to essentially ignore them in
533 // subsequent iterations when looking for vectorization opportunities
534 // while still tracking dependency chains that flow through those
535 // instructions.
536 if (isa<InsertElementInst>(V) || isa<ExtractElementInst>(V))
537 return 0;
538
539 // Give a load or store half of the required depth so that load/store
540 // pairs will vectorize.
541 if (!Config.NoMemOpBoost && (isa<LoadInst>(V) || isa<StoreInst>(V)))
542 return Config.ReqChainDepth/2;
543
544 return 1;
545 }
546
547 // Returns the cost of the provided instruction using TTI.
548 // This does not handle loads and stores.
getInstrCost__anonbe9f7d990111::BBVectorize549 unsigned getInstrCost(unsigned Opcode, Type *T1, Type *T2,
550 TargetTransformInfo::OperandValueKind Op1VK =
551 TargetTransformInfo::OK_AnyValue,
552 TargetTransformInfo::OperandValueKind Op2VK =
553 TargetTransformInfo::OK_AnyValue) {
554 switch (Opcode) {
555 default: break;
556 case Instruction::GetElementPtr:
557 // We mark this instruction as zero-cost because scalar GEPs are usually
558 // lowered to the instruction addressing mode. At the moment we don't
559 // generate vector GEPs.
560 return 0;
561 case Instruction::Br:
562 return TTI->getCFInstrCost(Opcode);
563 case Instruction::PHI:
564 return 0;
565 case Instruction::Add:
566 case Instruction::FAdd:
567 case Instruction::Sub:
568 case Instruction::FSub:
569 case Instruction::Mul:
570 case Instruction::FMul:
571 case Instruction::UDiv:
572 case Instruction::SDiv:
573 case Instruction::FDiv:
574 case Instruction::URem:
575 case Instruction::SRem:
576 case Instruction::FRem:
577 case Instruction::Shl:
578 case Instruction::LShr:
579 case Instruction::AShr:
580 case Instruction::And:
581 case Instruction::Or:
582 case Instruction::Xor:
583 return TTI->getArithmeticInstrCost(Opcode, T1, Op1VK, Op2VK);
584 case Instruction::Select:
585 case Instruction::ICmp:
586 case Instruction::FCmp:
587 return TTI->getCmpSelInstrCost(Opcode, T1, T2);
588 case Instruction::ZExt:
589 case Instruction::SExt:
590 case Instruction::FPToUI:
591 case Instruction::FPToSI:
592 case Instruction::FPExt:
593 case Instruction::PtrToInt:
594 case Instruction::IntToPtr:
595 case Instruction::SIToFP:
596 case Instruction::UIToFP:
597 case Instruction::Trunc:
598 case Instruction::FPTrunc:
599 case Instruction::BitCast:
600 case Instruction::ShuffleVector:
601 return TTI->getCastInstrCost(Opcode, T1, T2);
602 }
603
604 return 1;
605 }
606
607 // This determines the relative offset of two loads or stores, returning
608 // true if the offset could be determined to be some constant value.
609 // For example, if OffsetInElmts == 1, then J accesses the memory directly
610 // after I; if OffsetInElmts == -1 then I accesses the memory
611 // directly after J.
getPairPtrInfo__anonbe9f7d990111::BBVectorize612 bool getPairPtrInfo(Instruction *I, Instruction *J,
613 Value *&IPtr, Value *&JPtr, unsigned &IAlignment, unsigned &JAlignment,
614 unsigned &IAddressSpace, unsigned &JAddressSpace,
615 int64_t &OffsetInElmts, bool ComputeOffset = true) {
616 OffsetInElmts = 0;
617 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
618 LoadInst *LJ = cast<LoadInst>(J);
619 IPtr = LI->getPointerOperand();
620 JPtr = LJ->getPointerOperand();
621 IAlignment = LI->getAlignment();
622 JAlignment = LJ->getAlignment();
623 IAddressSpace = LI->getPointerAddressSpace();
624 JAddressSpace = LJ->getPointerAddressSpace();
625 } else {
626 StoreInst *SI = cast<StoreInst>(I), *SJ = cast<StoreInst>(J);
627 IPtr = SI->getPointerOperand();
628 JPtr = SJ->getPointerOperand();
629 IAlignment = SI->getAlignment();
630 JAlignment = SJ->getAlignment();
631 IAddressSpace = SI->getPointerAddressSpace();
632 JAddressSpace = SJ->getPointerAddressSpace();
633 }
634
635 if (!ComputeOffset)
636 return true;
637
638 const SCEV *IPtrSCEV = SE->getSCEV(IPtr);
639 const SCEV *JPtrSCEV = SE->getSCEV(JPtr);
640
641 // If this is a trivial offset, then we'll get something like
642 // 1*sizeof(type). With target data, which we need anyway, this will get
643 // constant folded into a number.
644 const SCEV *OffsetSCEV = SE->getMinusSCEV(JPtrSCEV, IPtrSCEV);
645 if (const SCEVConstant *ConstOffSCEV =
646 dyn_cast<SCEVConstant>(OffsetSCEV)) {
647 ConstantInt *IntOff = ConstOffSCEV->getValue();
648 int64_t Offset = IntOff->getSExtValue();
649 const DataLayout &DL = I->getModule()->getDataLayout();
650 Type *VTy = IPtr->getType()->getPointerElementType();
651 int64_t VTyTSS = (int64_t)DL.getTypeStoreSize(VTy);
652
653 Type *VTy2 = JPtr->getType()->getPointerElementType();
654 if (VTy != VTy2 && Offset < 0) {
655 int64_t VTy2TSS = (int64_t)DL.getTypeStoreSize(VTy2);
656 OffsetInElmts = Offset/VTy2TSS;
657 return (std::abs(Offset) % VTy2TSS) == 0;
658 }
659
660 OffsetInElmts = Offset/VTyTSS;
661 return (std::abs(Offset) % VTyTSS) == 0;
662 }
663
664 return false;
665 }
666
667 // Returns true if the provided CallInst represents an intrinsic that can
668 // be vectorized.
isVectorizableIntrinsic__anonbe9f7d990111::BBVectorize669 bool isVectorizableIntrinsic(CallInst* I) {
670 Function *F = I->getCalledFunction();
671 if (!F) return false;
672
673 Intrinsic::ID IID = F->getIntrinsicID();
674 if (!IID) return false;
675
676 switch(IID) {
677 default:
678 return false;
679 case Intrinsic::sqrt:
680 case Intrinsic::powi:
681 case Intrinsic::sin:
682 case Intrinsic::cos:
683 case Intrinsic::log:
684 case Intrinsic::log2:
685 case Intrinsic::log10:
686 case Intrinsic::exp:
687 case Intrinsic::exp2:
688 case Intrinsic::pow:
689 case Intrinsic::round:
690 case Intrinsic::copysign:
691 case Intrinsic::ceil:
692 case Intrinsic::nearbyint:
693 case Intrinsic::rint:
694 case Intrinsic::trunc:
695 case Intrinsic::floor:
696 case Intrinsic::fabs:
697 case Intrinsic::minnum:
698 case Intrinsic::maxnum:
699 return Config.VectorizeMath;
700 case Intrinsic::bswap:
701 case Intrinsic::ctpop:
702 case Intrinsic::ctlz:
703 case Intrinsic::cttz:
704 return Config.VectorizeBitManipulations;
705 case Intrinsic::fma:
706 case Intrinsic::fmuladd:
707 return Config.VectorizeFMA;
708 }
709 }
710
isPureIEChain__anonbe9f7d990111::BBVectorize711 bool isPureIEChain(InsertElementInst *IE) {
712 InsertElementInst *IENext = IE;
713 do {
714 if (!isa<UndefValue>(IENext->getOperand(0)) &&
715 !isa<InsertElementInst>(IENext->getOperand(0))) {
716 return false;
717 }
718 } while ((IENext =
719 dyn_cast<InsertElementInst>(IENext->getOperand(0))));
720
721 return true;
722 }
723 };
724
725 // This function implements one vectorization iteration on the provided
726 // basic block. It returns true if the block is changed.
vectorizePairs(BasicBlock & BB,bool NonPow2Len)727 bool BBVectorize::vectorizePairs(BasicBlock &BB, bool NonPow2Len) {
728 bool ShouldContinue;
729 BasicBlock::iterator Start = BB.getFirstInsertionPt();
730
731 std::vector<Value *> AllPairableInsts;
732 DenseMap<Value *, Value *> AllChosenPairs;
733 DenseSet<ValuePair> AllFixedOrderPairs;
734 DenseMap<VPPair, unsigned> AllPairConnectionTypes;
735 DenseMap<ValuePair, std::vector<ValuePair> > AllConnectedPairs,
736 AllConnectedPairDeps;
737
738 do {
739 std::vector<Value *> PairableInsts;
740 DenseMap<Value *, std::vector<Value *> > CandidatePairs;
741 DenseSet<ValuePair> FixedOrderPairs;
742 DenseMap<ValuePair, int> CandidatePairCostSavings;
743 ShouldContinue = getCandidatePairs(BB, Start, CandidatePairs,
744 FixedOrderPairs,
745 CandidatePairCostSavings,
746 PairableInsts, NonPow2Len);
747 if (PairableInsts.empty()) continue;
748
749 // Build the candidate pair set for faster lookups.
750 DenseSet<ValuePair> CandidatePairsSet;
751 for (DenseMap<Value *, std::vector<Value *> >::iterator I =
752 CandidatePairs.begin(), E = CandidatePairs.end(); I != E; ++I)
753 for (std::vector<Value *>::iterator J = I->second.begin(),
754 JE = I->second.end(); J != JE; ++J)
755 CandidatePairsSet.insert(ValuePair(I->first, *J));
756
757 // Now we have a map of all of the pairable instructions and we need to
758 // select the best possible pairing. A good pairing is one such that the
759 // users of the pair are also paired. This defines a (directed) forest
760 // over the pairs such that two pairs are connected iff the second pair
761 // uses the first.
762
763 // Note that it only matters that both members of the second pair use some
764 // element of the first pair (to allow for splatting).
765
766 DenseMap<ValuePair, std::vector<ValuePair> > ConnectedPairs,
767 ConnectedPairDeps;
768 DenseMap<VPPair, unsigned> PairConnectionTypes;
769 computeConnectedPairs(CandidatePairs, CandidatePairsSet,
770 PairableInsts, ConnectedPairs, PairConnectionTypes);
771 if (ConnectedPairs.empty()) continue;
772
773 for (DenseMap<ValuePair, std::vector<ValuePair> >::iterator
774 I = ConnectedPairs.begin(), IE = ConnectedPairs.end();
775 I != IE; ++I)
776 for (std::vector<ValuePair>::iterator J = I->second.begin(),
777 JE = I->second.end(); J != JE; ++J)
778 ConnectedPairDeps[*J].push_back(I->first);
779
780 // Build the pairable-instruction dependency map
781 DenseSet<ValuePair> PairableInstUsers;
782 buildDepMap(BB, CandidatePairs, PairableInsts, PairableInstUsers);
783
784 // There is now a graph of the connected pairs. For each variable, pick
785 // the pairing with the largest dag meeting the depth requirement on at
786 // least one branch. Then select all pairings that are part of that dag
787 // and remove them from the list of available pairings and pairable
788 // variables.
789
790 DenseMap<Value *, Value *> ChosenPairs;
791 choosePairs(CandidatePairs, CandidatePairsSet,
792 CandidatePairCostSavings,
793 PairableInsts, FixedOrderPairs, PairConnectionTypes,
794 ConnectedPairs, ConnectedPairDeps,
795 PairableInstUsers, ChosenPairs);
796
797 if (ChosenPairs.empty()) continue;
798 AllPairableInsts.insert(AllPairableInsts.end(), PairableInsts.begin(),
799 PairableInsts.end());
800 AllChosenPairs.insert(ChosenPairs.begin(), ChosenPairs.end());
801
802 // Only for the chosen pairs, propagate information on fixed-order pairs,
803 // pair connections, and their types to the data structures used by the
804 // pair fusion procedures.
805 for (DenseMap<Value *, Value *>::iterator I = ChosenPairs.begin(),
806 IE = ChosenPairs.end(); I != IE; ++I) {
807 if (FixedOrderPairs.count(*I))
808 AllFixedOrderPairs.insert(*I);
809 else if (FixedOrderPairs.count(ValuePair(I->second, I->first)))
810 AllFixedOrderPairs.insert(ValuePair(I->second, I->first));
811
812 for (DenseMap<Value *, Value *>::iterator J = ChosenPairs.begin();
813 J != IE; ++J) {
814 DenseMap<VPPair, unsigned>::iterator K =
815 PairConnectionTypes.find(VPPair(*I, *J));
816 if (K != PairConnectionTypes.end()) {
817 AllPairConnectionTypes.insert(*K);
818 } else {
819 K = PairConnectionTypes.find(VPPair(*J, *I));
820 if (K != PairConnectionTypes.end())
821 AllPairConnectionTypes.insert(*K);
822 }
823 }
824 }
825
826 for (DenseMap<ValuePair, std::vector<ValuePair> >::iterator
827 I = ConnectedPairs.begin(), IE = ConnectedPairs.end();
828 I != IE; ++I)
829 for (std::vector<ValuePair>::iterator J = I->second.begin(),
830 JE = I->second.end(); J != JE; ++J)
831 if (AllPairConnectionTypes.count(VPPair(I->first, *J))) {
832 AllConnectedPairs[I->first].push_back(*J);
833 AllConnectedPairDeps[*J].push_back(I->first);
834 }
835 } while (ShouldContinue);
836
837 if (AllChosenPairs.empty()) return false;
838 NumFusedOps += AllChosenPairs.size();
839
840 // A set of pairs has now been selected. It is now necessary to replace the
841 // paired instructions with vector instructions. For this procedure each
842 // operand must be replaced with a vector operand. This vector is formed
843 // by using build_vector on the old operands. The replaced values are then
844 // replaced with a vector_extract on the result. Subsequent optimization
845 // passes should coalesce the build/extract combinations.
846
847 fuseChosenPairs(BB, AllPairableInsts, AllChosenPairs, AllFixedOrderPairs,
848 AllPairConnectionTypes,
849 AllConnectedPairs, AllConnectedPairDeps);
850
851 // It is important to cleanup here so that future iterations of this
852 // function have less work to do.
853 (void)SimplifyInstructionsInBlock(&BB, TLI);
854 return true;
855 }
856
857 // This function returns true if the provided instruction is capable of being
858 // fused into a vector instruction. This determination is based only on the
859 // type and other attributes of the instruction.
isInstVectorizable(Instruction * I,bool & IsSimpleLoadStore)860 bool BBVectorize::isInstVectorizable(Instruction *I,
861 bool &IsSimpleLoadStore) {
862 IsSimpleLoadStore = false;
863
864 if (CallInst *C = dyn_cast<CallInst>(I)) {
865 if (!isVectorizableIntrinsic(C))
866 return false;
867 } else if (LoadInst *L = dyn_cast<LoadInst>(I)) {
868 // Vectorize simple loads if possbile:
869 IsSimpleLoadStore = L->isSimple();
870 if (!IsSimpleLoadStore || !Config.VectorizeMemOps)
871 return false;
872 } else if (StoreInst *S = dyn_cast<StoreInst>(I)) {
873 // Vectorize simple stores if possbile:
874 IsSimpleLoadStore = S->isSimple();
875 if (!IsSimpleLoadStore || !Config.VectorizeMemOps)
876 return false;
877 } else if (CastInst *C = dyn_cast<CastInst>(I)) {
878 // We can vectorize casts, but not casts of pointer types, etc.
879 if (!Config.VectorizeCasts)
880 return false;
881
882 Type *SrcTy = C->getSrcTy();
883 if (!SrcTy->isSingleValueType())
884 return false;
885
886 Type *DestTy = C->getDestTy();
887 if (!DestTy->isSingleValueType())
888 return false;
889 } else if (isa<SelectInst>(I)) {
890 if (!Config.VectorizeSelect)
891 return false;
892 } else if (isa<CmpInst>(I)) {
893 if (!Config.VectorizeCmp)
894 return false;
895 } else if (GetElementPtrInst *G = dyn_cast<GetElementPtrInst>(I)) {
896 if (!Config.VectorizeGEP)
897 return false;
898
899 // Currently, vector GEPs exist only with one index.
900 if (G->getNumIndices() != 1)
901 return false;
902 } else if (!(I->isBinaryOp() || isa<ShuffleVectorInst>(I) ||
903 isa<ExtractElementInst>(I) || isa<InsertElementInst>(I))) {
904 return false;
905 }
906
907 Type *T1, *T2;
908 getInstructionTypes(I, T1, T2);
909
910 // Not every type can be vectorized...
911 if (!(VectorType::isValidElementType(T1) || T1->isVectorTy()) ||
912 !(VectorType::isValidElementType(T2) || T2->isVectorTy()))
913 return false;
914
915 if (T1->getScalarSizeInBits() == 1) {
916 if (!Config.VectorizeBools)
917 return false;
918 } else {
919 if (!Config.VectorizeInts && T1->isIntOrIntVectorTy())
920 return false;
921 }
922
923 if (T2->getScalarSizeInBits() == 1) {
924 if (!Config.VectorizeBools)
925 return false;
926 } else {
927 if (!Config.VectorizeInts && T2->isIntOrIntVectorTy())
928 return false;
929 }
930
931 if (!Config.VectorizeFloats
932 && (T1->isFPOrFPVectorTy() || T2->isFPOrFPVectorTy()))
933 return false;
934
935 // Don't vectorize target-specific types.
936 if (T1->isX86_FP80Ty() || T1->isPPC_FP128Ty() || T1->isX86_MMXTy())
937 return false;
938 if (T2->isX86_FP80Ty() || T2->isPPC_FP128Ty() || T2->isX86_MMXTy())
939 return false;
940
941 if (!Config.VectorizePointers && (T1->getScalarType()->isPointerTy() ||
942 T2->getScalarType()->isPointerTy()))
943 return false;
944
945 if (!TTI && (T1->getPrimitiveSizeInBits() >= Config.VectorBits ||
946 T2->getPrimitiveSizeInBits() >= Config.VectorBits))
947 return false;
948
949 return true;
950 }
951
952 // This function returns true if the two provided instructions are compatible
953 // (meaning that they can be fused into a vector instruction). This assumes
954 // that I has already been determined to be vectorizable and that J is not
955 // in the use dag of I.
areInstsCompatible(Instruction * I,Instruction * J,bool IsSimpleLoadStore,bool NonPow2Len,int & CostSavings,int & FixedOrder)956 bool BBVectorize::areInstsCompatible(Instruction *I, Instruction *J,
957 bool IsSimpleLoadStore, bool NonPow2Len,
958 int &CostSavings, int &FixedOrder) {
959 DEBUG(if (DebugInstructionExamination) dbgs() << "BBV: looking at " << *I <<
960 " <-> " << *J << "\n");
961
962 CostSavings = 0;
963 FixedOrder = 0;
964
965 // Loads and stores can be merged if they have different alignments,
966 // but are otherwise the same.
967 if (!J->isSameOperationAs(I, Instruction::CompareIgnoringAlignment |
968 (NonPow2Len ? Instruction::CompareUsingScalarTypes : 0)))
969 return false;
970
971 Type *IT1, *IT2, *JT1, *JT2;
972 getInstructionTypes(I, IT1, IT2);
973 getInstructionTypes(J, JT1, JT2);
974 unsigned MaxTypeBits = std::max(
975 IT1->getPrimitiveSizeInBits() + JT1->getPrimitiveSizeInBits(),
976 IT2->getPrimitiveSizeInBits() + JT2->getPrimitiveSizeInBits());
977 if (!TTI && MaxTypeBits > Config.VectorBits)
978 return false;
979
980 // FIXME: handle addsub-type operations!
981
982 if (IsSimpleLoadStore) {
983 Value *IPtr, *JPtr;
984 unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace;
985 int64_t OffsetInElmts = 0;
986 if (getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment,
987 IAddressSpace, JAddressSpace, OffsetInElmts) &&
988 std::abs(OffsetInElmts) == 1) {
989 FixedOrder = (int) OffsetInElmts;
990 unsigned BottomAlignment = IAlignment;
991 if (OffsetInElmts < 0) BottomAlignment = JAlignment;
992
993 Type *aTypeI = isa<StoreInst>(I) ?
994 cast<StoreInst>(I)->getValueOperand()->getType() : I->getType();
995 Type *aTypeJ = isa<StoreInst>(J) ?
996 cast<StoreInst>(J)->getValueOperand()->getType() : J->getType();
997 Type *VType = getVecTypeForPair(aTypeI, aTypeJ);
998
999 if (Config.AlignedOnly) {
1000 // An aligned load or store is possible only if the instruction
1001 // with the lower offset has an alignment suitable for the
1002 // vector type.
1003 const DataLayout &DL = I->getModule()->getDataLayout();
1004 unsigned VecAlignment = DL.getPrefTypeAlignment(VType);
1005 if (BottomAlignment < VecAlignment)
1006 return false;
1007 }
1008
1009 if (TTI) {
1010 unsigned ICost = TTI->getMemoryOpCost(I->getOpcode(), aTypeI,
1011 IAlignment, IAddressSpace);
1012 unsigned JCost = TTI->getMemoryOpCost(J->getOpcode(), aTypeJ,
1013 JAlignment, JAddressSpace);
1014 unsigned VCost = TTI->getMemoryOpCost(I->getOpcode(), VType,
1015 BottomAlignment,
1016 IAddressSpace);
1017
1018 ICost += TTI->getAddressComputationCost(aTypeI);
1019 JCost += TTI->getAddressComputationCost(aTypeJ);
1020 VCost += TTI->getAddressComputationCost(VType);
1021
1022 if (VCost > ICost + JCost)
1023 return false;
1024
1025 // We don't want to fuse to a type that will be split, even
1026 // if the two input types will also be split and there is no other
1027 // associated cost.
1028 unsigned VParts = TTI->getNumberOfParts(VType);
1029 if (VParts > 1)
1030 return false;
1031 else if (!VParts && VCost == ICost + JCost)
1032 return false;
1033
1034 CostSavings = ICost + JCost - VCost;
1035 }
1036 } else {
1037 return false;
1038 }
1039 } else if (TTI) {
1040 unsigned ICost = getInstrCost(I->getOpcode(), IT1, IT2);
1041 unsigned JCost = getInstrCost(J->getOpcode(), JT1, JT2);
1042 Type *VT1 = getVecTypeForPair(IT1, JT1),
1043 *VT2 = getVecTypeForPair(IT2, JT2);
1044 TargetTransformInfo::OperandValueKind Op1VK =
1045 TargetTransformInfo::OK_AnyValue;
1046 TargetTransformInfo::OperandValueKind Op2VK =
1047 TargetTransformInfo::OK_AnyValue;
1048
1049 // On some targets (example X86) the cost of a vector shift may vary
1050 // depending on whether the second operand is a Uniform or
1051 // NonUniform Constant.
1052 switch (I->getOpcode()) {
1053 default : break;
1054 case Instruction::Shl:
1055 case Instruction::LShr:
1056 case Instruction::AShr:
1057
1058 // If both I and J are scalar shifts by constant, then the
1059 // merged vector shift count would be either a constant splat value
1060 // or a non-uniform vector of constants.
1061 if (ConstantInt *CII = dyn_cast<ConstantInt>(I->getOperand(1))) {
1062 if (ConstantInt *CIJ = dyn_cast<ConstantInt>(J->getOperand(1)))
1063 Op2VK = CII == CIJ ? TargetTransformInfo::OK_UniformConstantValue :
1064 TargetTransformInfo::OK_NonUniformConstantValue;
1065 } else {
1066 // Check for a splat of a constant or for a non uniform vector
1067 // of constants.
1068 Value *IOp = I->getOperand(1);
1069 Value *JOp = J->getOperand(1);
1070 if ((isa<ConstantVector>(IOp) || isa<ConstantDataVector>(IOp)) &&
1071 (isa<ConstantVector>(JOp) || isa<ConstantDataVector>(JOp))) {
1072 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
1073 Constant *SplatValue = cast<Constant>(IOp)->getSplatValue();
1074 if (SplatValue != nullptr &&
1075 SplatValue == cast<Constant>(JOp)->getSplatValue())
1076 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
1077 }
1078 }
1079 }
1080
1081 // Note that this procedure is incorrect for insert and extract element
1082 // instructions (because combining these often results in a shuffle),
1083 // but this cost is ignored (because insert and extract element
1084 // instructions are assigned a zero depth factor and are not really
1085 // fused in general).
1086 unsigned VCost = getInstrCost(I->getOpcode(), VT1, VT2, Op1VK, Op2VK);
1087
1088 if (VCost > ICost + JCost)
1089 return false;
1090
1091 // We don't want to fuse to a type that will be split, even
1092 // if the two input types will also be split and there is no other
1093 // associated cost.
1094 unsigned VParts1 = TTI->getNumberOfParts(VT1),
1095 VParts2 = TTI->getNumberOfParts(VT2);
1096 if (VParts1 > 1 || VParts2 > 1)
1097 return false;
1098 else if ((!VParts1 || !VParts2) && VCost == ICost + JCost)
1099 return false;
1100
1101 CostSavings = ICost + JCost - VCost;
1102 }
1103
1104 // The powi,ctlz,cttz intrinsics are special because only the first
1105 // argument is vectorized, the second arguments must be equal.
1106 CallInst *CI = dyn_cast<CallInst>(I);
1107 Function *FI;
1108 if (CI && (FI = CI->getCalledFunction())) {
1109 Intrinsic::ID IID = FI->getIntrinsicID();
1110 if (IID == Intrinsic::powi || IID == Intrinsic::ctlz ||
1111 IID == Intrinsic::cttz) {
1112 Value *A1I = CI->getArgOperand(1),
1113 *A1J = cast<CallInst>(J)->getArgOperand(1);
1114 const SCEV *A1ISCEV = SE->getSCEV(A1I),
1115 *A1JSCEV = SE->getSCEV(A1J);
1116 return (A1ISCEV == A1JSCEV);
1117 }
1118
1119 if (IID && TTI) {
1120 SmallVector<Type*, 4> Tys;
1121 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
1122 Tys.push_back(CI->getArgOperand(i)->getType());
1123 unsigned ICost = TTI->getIntrinsicInstrCost(IID, IT1, Tys);
1124
1125 Tys.clear();
1126 CallInst *CJ = cast<CallInst>(J);
1127 for (unsigned i = 0, ie = CJ->getNumArgOperands(); i != ie; ++i)
1128 Tys.push_back(CJ->getArgOperand(i)->getType());
1129 unsigned JCost = TTI->getIntrinsicInstrCost(IID, JT1, Tys);
1130
1131 Tys.clear();
1132 assert(CI->getNumArgOperands() == CJ->getNumArgOperands() &&
1133 "Intrinsic argument counts differ");
1134 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
1135 if ((IID == Intrinsic::powi || IID == Intrinsic::ctlz ||
1136 IID == Intrinsic::cttz) && i == 1)
1137 Tys.push_back(CI->getArgOperand(i)->getType());
1138 else
1139 Tys.push_back(getVecTypeForPair(CI->getArgOperand(i)->getType(),
1140 CJ->getArgOperand(i)->getType()));
1141 }
1142
1143 Type *RetTy = getVecTypeForPair(IT1, JT1);
1144 unsigned VCost = TTI->getIntrinsicInstrCost(IID, RetTy, Tys);
1145
1146 if (VCost > ICost + JCost)
1147 return false;
1148
1149 // We don't want to fuse to a type that will be split, even
1150 // if the two input types will also be split and there is no other
1151 // associated cost.
1152 unsigned RetParts = TTI->getNumberOfParts(RetTy);
1153 if (RetParts > 1)
1154 return false;
1155 else if (!RetParts && VCost == ICost + JCost)
1156 return false;
1157
1158 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
1159 if (!Tys[i]->isVectorTy())
1160 continue;
1161
1162 unsigned NumParts = TTI->getNumberOfParts(Tys[i]);
1163 if (NumParts > 1)
1164 return false;
1165 else if (!NumParts && VCost == ICost + JCost)
1166 return false;
1167 }
1168
1169 CostSavings = ICost + JCost - VCost;
1170 }
1171 }
1172
1173 return true;
1174 }
1175
1176 // Figure out whether or not J uses I and update the users and write-set
1177 // structures associated with I. Specifically, Users represents the set of
1178 // instructions that depend on I. WriteSet represents the set
1179 // of memory locations that are dependent on I. If UpdateUsers is true,
1180 // and J uses I, then Users is updated to contain J and WriteSet is updated
1181 // to contain any memory locations to which J writes. The function returns
1182 // true if J uses I. By default, alias analysis is used to determine
1183 // whether J reads from memory that overlaps with a location in WriteSet.
1184 // If LoadMoveSet is not null, then it is a previously-computed map
1185 // where the key is the memory-based user instruction and the value is
1186 // the instruction to be compared with I. So, if LoadMoveSet is provided,
1187 // then the alias analysis is not used. This is necessary because this
1188 // function is called during the process of moving instructions during
1189 // vectorization and the results of the alias analysis are not stable during
1190 // that process.
trackUsesOfI(DenseSet<Value * > & Users,AliasSetTracker & WriteSet,Instruction * I,Instruction * J,bool UpdateUsers,DenseSet<ValuePair> * LoadMoveSetPairs)1191 bool BBVectorize::trackUsesOfI(DenseSet<Value *> &Users,
1192 AliasSetTracker &WriteSet, Instruction *I,
1193 Instruction *J, bool UpdateUsers,
1194 DenseSet<ValuePair> *LoadMoveSetPairs) {
1195 bool UsesI = false;
1196
1197 // This instruction may already be marked as a user due, for example, to
1198 // being a member of a selected pair.
1199 if (Users.count(J))
1200 UsesI = true;
1201
1202 if (!UsesI)
1203 for (User::op_iterator JU = J->op_begin(), JE = J->op_end();
1204 JU != JE; ++JU) {
1205 Value *V = *JU;
1206 if (I == V || Users.count(V)) {
1207 UsesI = true;
1208 break;
1209 }
1210 }
1211 if (!UsesI && J->mayReadFromMemory()) {
1212 if (LoadMoveSetPairs) {
1213 UsesI = LoadMoveSetPairs->count(ValuePair(J, I));
1214 } else {
1215 for (AliasSetTracker::iterator W = WriteSet.begin(),
1216 WE = WriteSet.end(); W != WE; ++W) {
1217 if (W->aliasesUnknownInst(J, *AA)) {
1218 UsesI = true;
1219 break;
1220 }
1221 }
1222 }
1223 }
1224
1225 if (UsesI && UpdateUsers) {
1226 if (J->mayWriteToMemory()) WriteSet.add(J);
1227 Users.insert(J);
1228 }
1229
1230 return UsesI;
1231 }
1232
1233 // This function iterates over all instruction pairs in the provided
1234 // basic block and collects all candidate pairs for vectorization.
getCandidatePairs(BasicBlock & BB,BasicBlock::iterator & Start,DenseMap<Value *,std::vector<Value * >> & CandidatePairs,DenseSet<ValuePair> & FixedOrderPairs,DenseMap<ValuePair,int> & CandidatePairCostSavings,std::vector<Value * > & PairableInsts,bool NonPow2Len)1235 bool BBVectorize::getCandidatePairs(BasicBlock &BB,
1236 BasicBlock::iterator &Start,
1237 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1238 DenseSet<ValuePair> &FixedOrderPairs,
1239 DenseMap<ValuePair, int> &CandidatePairCostSavings,
1240 std::vector<Value *> &PairableInsts, bool NonPow2Len) {
1241 size_t TotalPairs = 0;
1242 BasicBlock::iterator E = BB.end();
1243 if (Start == E) return false;
1244
1245 bool ShouldContinue = false, IAfterStart = false;
1246 for (BasicBlock::iterator I = Start++; I != E; ++I) {
1247 if (I == Start) IAfterStart = true;
1248
1249 bool IsSimpleLoadStore;
1250 if (!isInstVectorizable(&*I, IsSimpleLoadStore))
1251 continue;
1252
1253 // Look for an instruction with which to pair instruction *I...
1254 DenseSet<Value *> Users;
1255 AliasSetTracker WriteSet(*AA);
1256 if (I->mayWriteToMemory())
1257 WriteSet.add(&*I);
1258
1259 bool JAfterStart = IAfterStart;
1260 BasicBlock::iterator J = std::next(I);
1261 for (unsigned ss = 0; J != E && ss <= Config.SearchLimit; ++J, ++ss) {
1262 if (&*J == Start)
1263 JAfterStart = true;
1264
1265 // Determine if J uses I, if so, exit the loop.
1266 bool UsesI = trackUsesOfI(Users, WriteSet, &*I, &*J, !Config.FastDep);
1267 if (Config.FastDep) {
1268 // Note: For this heuristic to be effective, independent operations
1269 // must tend to be intermixed. This is likely to be true from some
1270 // kinds of grouped loop unrolling (but not the generic LLVM pass),
1271 // but otherwise may require some kind of reordering pass.
1272
1273 // When using fast dependency analysis,
1274 // stop searching after first use:
1275 if (UsesI) break;
1276 } else {
1277 if (UsesI) continue;
1278 }
1279
1280 // J does not use I, and comes before the first use of I, so it can be
1281 // merged with I if the instructions are compatible.
1282 int CostSavings, FixedOrder;
1283 if (!areInstsCompatible(&*I, &*J, IsSimpleLoadStore, NonPow2Len,
1284 CostSavings, FixedOrder))
1285 continue;
1286
1287 // J is a candidate for merging with I.
1288 if (PairableInsts.empty() ||
1289 PairableInsts[PairableInsts.size() - 1] != &*I) {
1290 PairableInsts.push_back(&*I);
1291 }
1292
1293 CandidatePairs[&*I].push_back(&*J);
1294 ++TotalPairs;
1295 if (TTI)
1296 CandidatePairCostSavings.insert(
1297 ValuePairWithCost(ValuePair(&*I, &*J), CostSavings));
1298
1299 if (FixedOrder == 1)
1300 FixedOrderPairs.insert(ValuePair(&*I, &*J));
1301 else if (FixedOrder == -1)
1302 FixedOrderPairs.insert(ValuePair(&*J, &*I));
1303
1304 // The next call to this function must start after the last instruction
1305 // selected during this invocation.
1306 if (JAfterStart) {
1307 Start = std::next(J);
1308 IAfterStart = JAfterStart = false;
1309 }
1310
1311 DEBUG(if (DebugCandidateSelection) dbgs() << "BBV: candidate pair "
1312 << *I << " <-> " << *J << " (cost savings: " <<
1313 CostSavings << ")\n");
1314
1315 // If we have already found too many pairs, break here and this function
1316 // will be called again starting after the last instruction selected
1317 // during this invocation.
1318 if (PairableInsts.size() >= Config.MaxInsts ||
1319 TotalPairs >= Config.MaxPairs) {
1320 ShouldContinue = true;
1321 break;
1322 }
1323 }
1324
1325 if (ShouldContinue)
1326 break;
1327 }
1328
1329 DEBUG(dbgs() << "BBV: found " << PairableInsts.size()
1330 << " instructions with candidate pairs\n");
1331
1332 return ShouldContinue;
1333 }
1334
1335 // Finds candidate pairs connected to the pair P = <PI, PJ>. This means that
1336 // it looks for pairs such that both members have an input which is an
1337 // output of PI or PJ.
computePairsConnectedTo(DenseMap<Value *,std::vector<Value * >> & CandidatePairs,DenseSet<ValuePair> & CandidatePairsSet,std::vector<Value * > & PairableInsts,DenseMap<ValuePair,std::vector<ValuePair>> & ConnectedPairs,DenseMap<VPPair,unsigned> & PairConnectionTypes,ValuePair P)1338 void BBVectorize::computePairsConnectedTo(
1339 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1340 DenseSet<ValuePair> &CandidatePairsSet,
1341 std::vector<Value *> &PairableInsts,
1342 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
1343 DenseMap<VPPair, unsigned> &PairConnectionTypes,
1344 ValuePair P) {
1345 StoreInst *SI, *SJ;
1346
1347 // For each possible pairing for this variable, look at the uses of
1348 // the first value...
1349 for (Value::user_iterator I = P.first->user_begin(),
1350 E = P.first->user_end();
1351 I != E; ++I) {
1352 User *UI = *I;
1353 if (isa<LoadInst>(UI)) {
1354 // A pair cannot be connected to a load because the load only takes one
1355 // operand (the address) and it is a scalar even after vectorization.
1356 continue;
1357 } else if ((SI = dyn_cast<StoreInst>(UI)) &&
1358 P.first == SI->getPointerOperand()) {
1359 // Similarly, a pair cannot be connected to a store through its
1360 // pointer operand.
1361 continue;
1362 }
1363
1364 // For each use of the first variable, look for uses of the second
1365 // variable...
1366 for (User *UJ : P.second->users()) {
1367 if ((SJ = dyn_cast<StoreInst>(UJ)) &&
1368 P.second == SJ->getPointerOperand())
1369 continue;
1370
1371 // Look for <I, J>:
1372 if (CandidatePairsSet.count(ValuePair(UI, UJ))) {
1373 VPPair VP(P, ValuePair(UI, UJ));
1374 ConnectedPairs[VP.first].push_back(VP.second);
1375 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionDirect));
1376 }
1377
1378 // Look for <J, I>:
1379 if (CandidatePairsSet.count(ValuePair(UJ, UI))) {
1380 VPPair VP(P, ValuePair(UJ, UI));
1381 ConnectedPairs[VP.first].push_back(VP.second);
1382 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSwap));
1383 }
1384 }
1385
1386 if (Config.SplatBreaksChain) continue;
1387 // Look for cases where just the first value in the pair is used by
1388 // both members of another pair (splatting).
1389 for (Value::user_iterator J = P.first->user_begin(); J != E; ++J) {
1390 User *UJ = *J;
1391 if ((SJ = dyn_cast<StoreInst>(UJ)) &&
1392 P.first == SJ->getPointerOperand())
1393 continue;
1394
1395 if (CandidatePairsSet.count(ValuePair(UI, UJ))) {
1396 VPPair VP(P, ValuePair(UI, UJ));
1397 ConnectedPairs[VP.first].push_back(VP.second);
1398 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSplat));
1399 }
1400 }
1401 }
1402
1403 if (Config.SplatBreaksChain) return;
1404 // Look for cases where just the second value in the pair is used by
1405 // both members of another pair (splatting).
1406 for (Value::user_iterator I = P.second->user_begin(),
1407 E = P.second->user_end();
1408 I != E; ++I) {
1409 User *UI = *I;
1410 if (isa<LoadInst>(UI))
1411 continue;
1412 else if ((SI = dyn_cast<StoreInst>(UI)) &&
1413 P.second == SI->getPointerOperand())
1414 continue;
1415
1416 for (Value::user_iterator J = P.second->user_begin(); J != E; ++J) {
1417 User *UJ = *J;
1418 if ((SJ = dyn_cast<StoreInst>(UJ)) &&
1419 P.second == SJ->getPointerOperand())
1420 continue;
1421
1422 if (CandidatePairsSet.count(ValuePair(UI, UJ))) {
1423 VPPair VP(P, ValuePair(UI, UJ));
1424 ConnectedPairs[VP.first].push_back(VP.second);
1425 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSplat));
1426 }
1427 }
1428 }
1429 }
1430
1431 // This function figures out which pairs are connected. Two pairs are
1432 // connected if some output of the first pair forms an input to both members
1433 // of the second pair.
computeConnectedPairs(DenseMap<Value *,std::vector<Value * >> & CandidatePairs,DenseSet<ValuePair> & CandidatePairsSet,std::vector<Value * > & PairableInsts,DenseMap<ValuePair,std::vector<ValuePair>> & ConnectedPairs,DenseMap<VPPair,unsigned> & PairConnectionTypes)1434 void BBVectorize::computeConnectedPairs(
1435 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1436 DenseSet<ValuePair> &CandidatePairsSet,
1437 std::vector<Value *> &PairableInsts,
1438 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
1439 DenseMap<VPPair, unsigned> &PairConnectionTypes) {
1440 for (std::vector<Value *>::iterator PI = PairableInsts.begin(),
1441 PE = PairableInsts.end(); PI != PE; ++PI) {
1442 DenseMap<Value *, std::vector<Value *> >::iterator PP =
1443 CandidatePairs.find(*PI);
1444 if (PP == CandidatePairs.end())
1445 continue;
1446
1447 for (std::vector<Value *>::iterator P = PP->second.begin(),
1448 E = PP->second.end(); P != E; ++P)
1449 computePairsConnectedTo(CandidatePairs, CandidatePairsSet,
1450 PairableInsts, ConnectedPairs,
1451 PairConnectionTypes, ValuePair(*PI, *P));
1452 }
1453
1454 DEBUG(size_t TotalPairs = 0;
1455 for (DenseMap<ValuePair, std::vector<ValuePair> >::iterator I =
1456 ConnectedPairs.begin(), IE = ConnectedPairs.end(); I != IE; ++I)
1457 TotalPairs += I->second.size();
1458 dbgs() << "BBV: found " << TotalPairs
1459 << " pair connections.\n");
1460 }
1461
1462 // This function builds a set of use tuples such that <A, B> is in the set
1463 // if B is in the use dag of A. If B is in the use dag of A, then B
1464 // depends on the output of A.
buildDepMap(BasicBlock & BB,DenseMap<Value *,std::vector<Value * >> & CandidatePairs,std::vector<Value * > & PairableInsts,DenseSet<ValuePair> & PairableInstUsers)1465 void BBVectorize::buildDepMap(
1466 BasicBlock &BB,
1467 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1468 std::vector<Value *> &PairableInsts,
1469 DenseSet<ValuePair> &PairableInstUsers) {
1470 DenseSet<Value *> IsInPair;
1471 for (DenseMap<Value *, std::vector<Value *> >::iterator C =
1472 CandidatePairs.begin(), E = CandidatePairs.end(); C != E; ++C) {
1473 IsInPair.insert(C->first);
1474 IsInPair.insert(C->second.begin(), C->second.end());
1475 }
1476
1477 // Iterate through the basic block, recording all users of each
1478 // pairable instruction.
1479
1480 BasicBlock::iterator E = BB.end(), EL =
1481 BasicBlock::iterator(cast<Instruction>(PairableInsts.back()));
1482 for (BasicBlock::iterator I = BB.getFirstInsertionPt(); I != E; ++I) {
1483 if (IsInPair.find(&*I) == IsInPair.end())
1484 continue;
1485
1486 DenseSet<Value *> Users;
1487 AliasSetTracker WriteSet(*AA);
1488 if (I->mayWriteToMemory())
1489 WriteSet.add(&*I);
1490
1491 for (BasicBlock::iterator J = std::next(I); J != E; ++J) {
1492 (void)trackUsesOfI(Users, WriteSet, &*I, &*J);
1493
1494 if (J == EL)
1495 break;
1496 }
1497
1498 for (DenseSet<Value *>::iterator U = Users.begin(), E = Users.end();
1499 U != E; ++U) {
1500 if (IsInPair.find(*U) == IsInPair.end()) continue;
1501 PairableInstUsers.insert(ValuePair(&*I, *U));
1502 }
1503
1504 if (I == EL)
1505 break;
1506 }
1507 }
1508
1509 // Returns true if an input to pair P is an output of pair Q and also an
1510 // input of pair Q is an output of pair P. If this is the case, then these
1511 // two pairs cannot be simultaneously fused.
pairsConflict(ValuePair P,ValuePair Q,DenseSet<ValuePair> & PairableInstUsers,DenseMap<ValuePair,std::vector<ValuePair>> * PairableInstUserMap,DenseSet<VPPair> * PairableInstUserPairSet)1512 bool BBVectorize::pairsConflict(ValuePair P, ValuePair Q,
1513 DenseSet<ValuePair> &PairableInstUsers,
1514 DenseMap<ValuePair, std::vector<ValuePair> > *PairableInstUserMap,
1515 DenseSet<VPPair> *PairableInstUserPairSet) {
1516 // Two pairs are in conflict if they are mutual Users of eachother.
1517 bool QUsesP = PairableInstUsers.count(ValuePair(P.first, Q.first)) ||
1518 PairableInstUsers.count(ValuePair(P.first, Q.second)) ||
1519 PairableInstUsers.count(ValuePair(P.second, Q.first)) ||
1520 PairableInstUsers.count(ValuePair(P.second, Q.second));
1521 bool PUsesQ = PairableInstUsers.count(ValuePair(Q.first, P.first)) ||
1522 PairableInstUsers.count(ValuePair(Q.first, P.second)) ||
1523 PairableInstUsers.count(ValuePair(Q.second, P.first)) ||
1524 PairableInstUsers.count(ValuePair(Q.second, P.second));
1525 if (PairableInstUserMap) {
1526 // FIXME: The expensive part of the cycle check is not so much the cycle
1527 // check itself but this edge insertion procedure. This needs some
1528 // profiling and probably a different data structure.
1529 if (PUsesQ) {
1530 if (PairableInstUserPairSet->insert(VPPair(Q, P)).second)
1531 (*PairableInstUserMap)[Q].push_back(P);
1532 }
1533 if (QUsesP) {
1534 if (PairableInstUserPairSet->insert(VPPair(P, Q)).second)
1535 (*PairableInstUserMap)[P].push_back(Q);
1536 }
1537 }
1538
1539 return (QUsesP && PUsesQ);
1540 }
1541
1542 // This function walks the use graph of current pairs to see if, starting
1543 // from P, the walk returns to P.
pairWillFormCycle(ValuePair P,DenseMap<ValuePair,std::vector<ValuePair>> & PairableInstUserMap,DenseSet<ValuePair> & CurrentPairs)1544 bool BBVectorize::pairWillFormCycle(ValuePair P,
1545 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap,
1546 DenseSet<ValuePair> &CurrentPairs) {
1547 DEBUG(if (DebugCycleCheck)
1548 dbgs() << "BBV: starting cycle check for : " << *P.first << " <-> "
1549 << *P.second << "\n");
1550 // A lookup table of visisted pairs is kept because the PairableInstUserMap
1551 // contains non-direct associations.
1552 DenseSet<ValuePair> Visited;
1553 SmallVector<ValuePair, 32> Q;
1554 // General depth-first post-order traversal:
1555 Q.push_back(P);
1556 do {
1557 ValuePair QTop = Q.pop_back_val();
1558 Visited.insert(QTop);
1559
1560 DEBUG(if (DebugCycleCheck)
1561 dbgs() << "BBV: cycle check visiting: " << *QTop.first << " <-> "
1562 << *QTop.second << "\n");
1563 DenseMap<ValuePair, std::vector<ValuePair> >::iterator QQ =
1564 PairableInstUserMap.find(QTop);
1565 if (QQ == PairableInstUserMap.end())
1566 continue;
1567
1568 for (std::vector<ValuePair>::iterator C = QQ->second.begin(),
1569 CE = QQ->second.end(); C != CE; ++C) {
1570 if (*C == P) {
1571 DEBUG(dbgs()
1572 << "BBV: rejected to prevent non-trivial cycle formation: "
1573 << QTop.first << " <-> " << C->second << "\n");
1574 return true;
1575 }
1576
1577 if (CurrentPairs.count(*C) && !Visited.count(*C))
1578 Q.push_back(*C);
1579 }
1580 } while (!Q.empty());
1581
1582 return false;
1583 }
1584
1585 // This function builds the initial dag of connected pairs with the
1586 // pair J at the root.
buildInitialDAGFor(DenseMap<Value *,std::vector<Value * >> & CandidatePairs,DenseSet<ValuePair> & CandidatePairsSet,std::vector<Value * > & PairableInsts,DenseMap<ValuePair,std::vector<ValuePair>> & ConnectedPairs,DenseSet<ValuePair> & PairableInstUsers,DenseMap<Value *,Value * > & ChosenPairs,DenseMap<ValuePair,size_t> & DAG,ValuePair J)1587 void BBVectorize::buildInitialDAGFor(
1588 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1589 DenseSet<ValuePair> &CandidatePairsSet,
1590 std::vector<Value *> &PairableInsts,
1591 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
1592 DenseSet<ValuePair> &PairableInstUsers,
1593 DenseMap<Value *, Value *> &ChosenPairs,
1594 DenseMap<ValuePair, size_t> &DAG, ValuePair J) {
1595 // Each of these pairs is viewed as the root node of a DAG. The DAG
1596 // is then walked (depth-first). As this happens, we keep track of
1597 // the pairs that compose the DAG and the maximum depth of the DAG.
1598 SmallVector<ValuePairWithDepth, 32> Q;
1599 // General depth-first post-order traversal:
1600 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first)));
1601 do {
1602 ValuePairWithDepth QTop = Q.back();
1603
1604 // Push each child onto the queue:
1605 bool MoreChildren = false;
1606 size_t MaxChildDepth = QTop.second;
1607 DenseMap<ValuePair, std::vector<ValuePair> >::iterator QQ =
1608 ConnectedPairs.find(QTop.first);
1609 if (QQ != ConnectedPairs.end())
1610 for (std::vector<ValuePair>::iterator k = QQ->second.begin(),
1611 ke = QQ->second.end(); k != ke; ++k) {
1612 // Make sure that this child pair is still a candidate:
1613 if (CandidatePairsSet.count(*k)) {
1614 DenseMap<ValuePair, size_t>::iterator C = DAG.find(*k);
1615 if (C == DAG.end()) {
1616 size_t d = getDepthFactor(k->first);
1617 Q.push_back(ValuePairWithDepth(*k, QTop.second+d));
1618 MoreChildren = true;
1619 } else {
1620 MaxChildDepth = std::max(MaxChildDepth, C->second);
1621 }
1622 }
1623 }
1624
1625 if (!MoreChildren) {
1626 // Record the current pair as part of the DAG:
1627 DAG.insert(ValuePairWithDepth(QTop.first, MaxChildDepth));
1628 Q.pop_back();
1629 }
1630 } while (!Q.empty());
1631 }
1632
1633 // Given some initial dag, prune it by removing conflicting pairs (pairs
1634 // that cannot be simultaneously chosen for vectorization).
pruneDAGFor(DenseMap<Value *,std::vector<Value * >> & CandidatePairs,std::vector<Value * > & PairableInsts,DenseMap<ValuePair,std::vector<ValuePair>> & ConnectedPairs,DenseSet<ValuePair> & PairableInstUsers,DenseMap<ValuePair,std::vector<ValuePair>> & PairableInstUserMap,DenseSet<VPPair> & PairableInstUserPairSet,DenseMap<Value *,Value * > & ChosenPairs,DenseMap<ValuePair,size_t> & DAG,DenseSet<ValuePair> & PrunedDAG,ValuePair J,bool UseCycleCheck)1635 void BBVectorize::pruneDAGFor(
1636 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1637 std::vector<Value *> &PairableInsts,
1638 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
1639 DenseSet<ValuePair> &PairableInstUsers,
1640 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap,
1641 DenseSet<VPPair> &PairableInstUserPairSet,
1642 DenseMap<Value *, Value *> &ChosenPairs,
1643 DenseMap<ValuePair, size_t> &DAG,
1644 DenseSet<ValuePair> &PrunedDAG, ValuePair J,
1645 bool UseCycleCheck) {
1646 SmallVector<ValuePairWithDepth, 32> Q;
1647 // General depth-first post-order traversal:
1648 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first)));
1649 do {
1650 ValuePairWithDepth QTop = Q.pop_back_val();
1651 PrunedDAG.insert(QTop.first);
1652
1653 // Visit each child, pruning as necessary...
1654 SmallVector<ValuePairWithDepth, 8> BestChildren;
1655 DenseMap<ValuePair, std::vector<ValuePair> >::iterator QQ =
1656 ConnectedPairs.find(QTop.first);
1657 if (QQ == ConnectedPairs.end())
1658 continue;
1659
1660 for (std::vector<ValuePair>::iterator K = QQ->second.begin(),
1661 KE = QQ->second.end(); K != KE; ++K) {
1662 DenseMap<ValuePair, size_t>::iterator C = DAG.find(*K);
1663 if (C == DAG.end()) continue;
1664
1665 // This child is in the DAG, now we need to make sure it is the
1666 // best of any conflicting children. There could be multiple
1667 // conflicting children, so first, determine if we're keeping
1668 // this child, then delete conflicting children as necessary.
1669
1670 // It is also necessary to guard against pairing-induced
1671 // dependencies. Consider instructions a .. x .. y .. b
1672 // such that (a,b) are to be fused and (x,y) are to be fused
1673 // but a is an input to x and b is an output from y. This
1674 // means that y cannot be moved after b but x must be moved
1675 // after b for (a,b) to be fused. In other words, after
1676 // fusing (a,b) we have y .. a/b .. x where y is an input
1677 // to a/b and x is an output to a/b: x and y can no longer
1678 // be legally fused. To prevent this condition, we must
1679 // make sure that a child pair added to the DAG is not
1680 // both an input and output of an already-selected pair.
1681
1682 // Pairing-induced dependencies can also form from more complicated
1683 // cycles. The pair vs. pair conflicts are easy to check, and so
1684 // that is done explicitly for "fast rejection", and because for
1685 // child vs. child conflicts, we may prefer to keep the current
1686 // pair in preference to the already-selected child.
1687 DenseSet<ValuePair> CurrentPairs;
1688
1689 bool CanAdd = true;
1690 for (SmallVectorImpl<ValuePairWithDepth>::iterator C2
1691 = BestChildren.begin(), E2 = BestChildren.end();
1692 C2 != E2; ++C2) {
1693 if (C2->first.first == C->first.first ||
1694 C2->first.first == C->first.second ||
1695 C2->first.second == C->first.first ||
1696 C2->first.second == C->first.second ||
1697 pairsConflict(C2->first, C->first, PairableInstUsers,
1698 UseCycleCheck ? &PairableInstUserMap : nullptr,
1699 UseCycleCheck ? &PairableInstUserPairSet
1700 : nullptr)) {
1701 if (C2->second >= C->second) {
1702 CanAdd = false;
1703 break;
1704 }
1705
1706 CurrentPairs.insert(C2->first);
1707 }
1708 }
1709 if (!CanAdd) continue;
1710
1711 // Even worse, this child could conflict with another node already
1712 // selected for the DAG. If that is the case, ignore this child.
1713 for (DenseSet<ValuePair>::iterator T = PrunedDAG.begin(),
1714 E2 = PrunedDAG.end(); T != E2; ++T) {
1715 if (T->first == C->first.first ||
1716 T->first == C->first.second ||
1717 T->second == C->first.first ||
1718 T->second == C->first.second ||
1719 pairsConflict(*T, C->first, PairableInstUsers,
1720 UseCycleCheck ? &PairableInstUserMap : nullptr,
1721 UseCycleCheck ? &PairableInstUserPairSet
1722 : nullptr)) {
1723 CanAdd = false;
1724 break;
1725 }
1726
1727 CurrentPairs.insert(*T);
1728 }
1729 if (!CanAdd) continue;
1730
1731 // And check the queue too...
1732 for (SmallVectorImpl<ValuePairWithDepth>::iterator C2 = Q.begin(),
1733 E2 = Q.end(); C2 != E2; ++C2) {
1734 if (C2->first.first == C->first.first ||
1735 C2->first.first == C->first.second ||
1736 C2->first.second == C->first.first ||
1737 C2->first.second == C->first.second ||
1738 pairsConflict(C2->first, C->first, PairableInstUsers,
1739 UseCycleCheck ? &PairableInstUserMap : nullptr,
1740 UseCycleCheck ? &PairableInstUserPairSet
1741 : nullptr)) {
1742 CanAdd = false;
1743 break;
1744 }
1745
1746 CurrentPairs.insert(C2->first);
1747 }
1748 if (!CanAdd) continue;
1749
1750 // Last but not least, check for a conflict with any of the
1751 // already-chosen pairs.
1752 for (DenseMap<Value *, Value *>::iterator C2 =
1753 ChosenPairs.begin(), E2 = ChosenPairs.end();
1754 C2 != E2; ++C2) {
1755 if (pairsConflict(*C2, C->first, PairableInstUsers,
1756 UseCycleCheck ? &PairableInstUserMap : nullptr,
1757 UseCycleCheck ? &PairableInstUserPairSet
1758 : nullptr)) {
1759 CanAdd = false;
1760 break;
1761 }
1762
1763 CurrentPairs.insert(*C2);
1764 }
1765 if (!CanAdd) continue;
1766
1767 // To check for non-trivial cycles formed by the addition of the
1768 // current pair we've formed a list of all relevant pairs, now use a
1769 // graph walk to check for a cycle. We start from the current pair and
1770 // walk the use dag to see if we again reach the current pair. If we
1771 // do, then the current pair is rejected.
1772
1773 // FIXME: It may be more efficient to use a topological-ordering
1774 // algorithm to improve the cycle check. This should be investigated.
1775 if (UseCycleCheck &&
1776 pairWillFormCycle(C->first, PairableInstUserMap, CurrentPairs))
1777 continue;
1778
1779 // This child can be added, but we may have chosen it in preference
1780 // to an already-selected child. Check for this here, and if a
1781 // conflict is found, then remove the previously-selected child
1782 // before adding this one in its place.
1783 for (SmallVectorImpl<ValuePairWithDepth>::iterator C2
1784 = BestChildren.begin(); C2 != BestChildren.end();) {
1785 if (C2->first.first == C->first.first ||
1786 C2->first.first == C->first.second ||
1787 C2->first.second == C->first.first ||
1788 C2->first.second == C->first.second ||
1789 pairsConflict(C2->first, C->first, PairableInstUsers))
1790 C2 = BestChildren.erase(C2);
1791 else
1792 ++C2;
1793 }
1794
1795 BestChildren.push_back(ValuePairWithDepth(C->first, C->second));
1796 }
1797
1798 for (SmallVectorImpl<ValuePairWithDepth>::iterator C
1799 = BestChildren.begin(), E2 = BestChildren.end();
1800 C != E2; ++C) {
1801 size_t DepthF = getDepthFactor(C->first.first);
1802 Q.push_back(ValuePairWithDepth(C->first, QTop.second+DepthF));
1803 }
1804 } while (!Q.empty());
1805 }
1806
1807 // This function finds the best dag of mututally-compatible connected
1808 // pairs, given the choice of root pairs as an iterator range.
findBestDAGFor(DenseMap<Value *,std::vector<Value * >> & CandidatePairs,DenseSet<ValuePair> & CandidatePairsSet,DenseMap<ValuePair,int> & CandidatePairCostSavings,std::vector<Value * > & PairableInsts,DenseSet<ValuePair> & FixedOrderPairs,DenseMap<VPPair,unsigned> & PairConnectionTypes,DenseMap<ValuePair,std::vector<ValuePair>> & ConnectedPairs,DenseMap<ValuePair,std::vector<ValuePair>> & ConnectedPairDeps,DenseSet<ValuePair> & PairableInstUsers,DenseMap<ValuePair,std::vector<ValuePair>> & PairableInstUserMap,DenseSet<VPPair> & PairableInstUserPairSet,DenseMap<Value *,Value * > & ChosenPairs,DenseSet<ValuePair> & BestDAG,size_t & BestMaxDepth,int & BestEffSize,Value * II,std::vector<Value * > & JJ,bool UseCycleCheck)1809 void BBVectorize::findBestDAGFor(
1810 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1811 DenseSet<ValuePair> &CandidatePairsSet,
1812 DenseMap<ValuePair, int> &CandidatePairCostSavings,
1813 std::vector<Value *> &PairableInsts,
1814 DenseSet<ValuePair> &FixedOrderPairs,
1815 DenseMap<VPPair, unsigned> &PairConnectionTypes,
1816 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
1817 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps,
1818 DenseSet<ValuePair> &PairableInstUsers,
1819 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap,
1820 DenseSet<VPPair> &PairableInstUserPairSet,
1821 DenseMap<Value *, Value *> &ChosenPairs,
1822 DenseSet<ValuePair> &BestDAG, size_t &BestMaxDepth,
1823 int &BestEffSize, Value *II, std::vector<Value *>&JJ,
1824 bool UseCycleCheck) {
1825 for (std::vector<Value *>::iterator J = JJ.begin(), JE = JJ.end();
1826 J != JE; ++J) {
1827 ValuePair IJ(II, *J);
1828 if (!CandidatePairsSet.count(IJ))
1829 continue;
1830
1831 // Before going any further, make sure that this pair does not
1832 // conflict with any already-selected pairs (see comment below
1833 // near the DAG pruning for more details).
1834 DenseSet<ValuePair> ChosenPairSet;
1835 bool DoesConflict = false;
1836 for (DenseMap<Value *, Value *>::iterator C = ChosenPairs.begin(),
1837 E = ChosenPairs.end(); C != E; ++C) {
1838 if (pairsConflict(*C, IJ, PairableInstUsers,
1839 UseCycleCheck ? &PairableInstUserMap : nullptr,
1840 UseCycleCheck ? &PairableInstUserPairSet : nullptr)) {
1841 DoesConflict = true;
1842 break;
1843 }
1844
1845 ChosenPairSet.insert(*C);
1846 }
1847 if (DoesConflict) continue;
1848
1849 if (UseCycleCheck &&
1850 pairWillFormCycle(IJ, PairableInstUserMap, ChosenPairSet))
1851 continue;
1852
1853 DenseMap<ValuePair, size_t> DAG;
1854 buildInitialDAGFor(CandidatePairs, CandidatePairsSet,
1855 PairableInsts, ConnectedPairs,
1856 PairableInstUsers, ChosenPairs, DAG, IJ);
1857
1858 // Because we'll keep the child with the largest depth, the largest
1859 // depth is still the same in the unpruned DAG.
1860 size_t MaxDepth = DAG.lookup(IJ);
1861
1862 DEBUG(if (DebugPairSelection) dbgs() << "BBV: found DAG for pair {"
1863 << *IJ.first << " <-> " << *IJ.second << "} of depth " <<
1864 MaxDepth << " and size " << DAG.size() << "\n");
1865
1866 // At this point the DAG has been constructed, but, may contain
1867 // contradictory children (meaning that different children of
1868 // some dag node may be attempting to fuse the same instruction).
1869 // So now we walk the dag again, in the case of a conflict,
1870 // keep only the child with the largest depth. To break a tie,
1871 // favor the first child.
1872
1873 DenseSet<ValuePair> PrunedDAG;
1874 pruneDAGFor(CandidatePairs, PairableInsts, ConnectedPairs,
1875 PairableInstUsers, PairableInstUserMap,
1876 PairableInstUserPairSet,
1877 ChosenPairs, DAG, PrunedDAG, IJ, UseCycleCheck);
1878
1879 int EffSize = 0;
1880 if (TTI) {
1881 DenseSet<Value *> PrunedDAGInstrs;
1882 for (DenseSet<ValuePair>::iterator S = PrunedDAG.begin(),
1883 E = PrunedDAG.end(); S != E; ++S) {
1884 PrunedDAGInstrs.insert(S->first);
1885 PrunedDAGInstrs.insert(S->second);
1886 }
1887
1888 // The set of pairs that have already contributed to the total cost.
1889 DenseSet<ValuePair> IncomingPairs;
1890
1891 // If the cost model were perfect, this might not be necessary; but we
1892 // need to make sure that we don't get stuck vectorizing our own
1893 // shuffle chains.
1894 bool HasNontrivialInsts = false;
1895
1896 // The node weights represent the cost savings associated with
1897 // fusing the pair of instructions.
1898 for (DenseSet<ValuePair>::iterator S = PrunedDAG.begin(),
1899 E = PrunedDAG.end(); S != E; ++S) {
1900 if (!isa<ShuffleVectorInst>(S->first) &&
1901 !isa<InsertElementInst>(S->first) &&
1902 !isa<ExtractElementInst>(S->first))
1903 HasNontrivialInsts = true;
1904
1905 bool FlipOrder = false;
1906
1907 if (getDepthFactor(S->first)) {
1908 int ESContrib = CandidatePairCostSavings.find(*S)->second;
1909 DEBUG(if (DebugPairSelection) dbgs() << "\tweight {"
1910 << *S->first << " <-> " << *S->second << "} = " <<
1911 ESContrib << "\n");
1912 EffSize += ESContrib;
1913 }
1914
1915 // The edge weights contribute in a negative sense: they represent
1916 // the cost of shuffles.
1917 DenseMap<ValuePair, std::vector<ValuePair> >::iterator SS =
1918 ConnectedPairDeps.find(*S);
1919 if (SS != ConnectedPairDeps.end()) {
1920 unsigned NumDepsDirect = 0, NumDepsSwap = 0;
1921 for (std::vector<ValuePair>::iterator T = SS->second.begin(),
1922 TE = SS->second.end(); T != TE; ++T) {
1923 VPPair Q(*S, *T);
1924 if (!PrunedDAG.count(Q.second))
1925 continue;
1926 DenseMap<VPPair, unsigned>::iterator R =
1927 PairConnectionTypes.find(VPPair(Q.second, Q.first));
1928 assert(R != PairConnectionTypes.end() &&
1929 "Cannot find pair connection type");
1930 if (R->second == PairConnectionDirect)
1931 ++NumDepsDirect;
1932 else if (R->second == PairConnectionSwap)
1933 ++NumDepsSwap;
1934 }
1935
1936 // If there are more swaps than direct connections, then
1937 // the pair order will be flipped during fusion. So the real
1938 // number of swaps is the minimum number.
1939 FlipOrder = !FixedOrderPairs.count(*S) &&
1940 ((NumDepsSwap > NumDepsDirect) ||
1941 FixedOrderPairs.count(ValuePair(S->second, S->first)));
1942
1943 for (std::vector<ValuePair>::iterator T = SS->second.begin(),
1944 TE = SS->second.end(); T != TE; ++T) {
1945 VPPair Q(*S, *T);
1946 if (!PrunedDAG.count(Q.second))
1947 continue;
1948 DenseMap<VPPair, unsigned>::iterator R =
1949 PairConnectionTypes.find(VPPair(Q.second, Q.first));
1950 assert(R != PairConnectionTypes.end() &&
1951 "Cannot find pair connection type");
1952 Type *Ty1 = Q.second.first->getType(),
1953 *Ty2 = Q.second.second->getType();
1954 Type *VTy = getVecTypeForPair(Ty1, Ty2);
1955 if ((R->second == PairConnectionDirect && FlipOrder) ||
1956 (R->second == PairConnectionSwap && !FlipOrder) ||
1957 R->second == PairConnectionSplat) {
1958 int ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
1959 VTy, VTy);
1960
1961 if (VTy->getVectorNumElements() == 2) {
1962 if (R->second == PairConnectionSplat)
1963 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
1964 TargetTransformInfo::SK_Broadcast, VTy));
1965 else
1966 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
1967 TargetTransformInfo::SK_Reverse, VTy));
1968 }
1969
1970 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" <<
1971 *Q.second.first << " <-> " << *Q.second.second <<
1972 "} -> {" <<
1973 *S->first << " <-> " << *S->second << "} = " <<
1974 ESContrib << "\n");
1975 EffSize -= ESContrib;
1976 }
1977 }
1978 }
1979
1980 // Compute the cost of outgoing edges. We assume that edges outgoing
1981 // to shuffles, inserts or extracts can be merged, and so contribute
1982 // no additional cost.
1983 if (!S->first->getType()->isVoidTy()) {
1984 Type *Ty1 = S->first->getType(),
1985 *Ty2 = S->second->getType();
1986 Type *VTy = getVecTypeForPair(Ty1, Ty2);
1987
1988 bool NeedsExtraction = false;
1989 for (User *U : S->first->users()) {
1990 if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(U)) {
1991 // Shuffle can be folded if it has no other input
1992 if (isa<UndefValue>(SI->getOperand(1)))
1993 continue;
1994 }
1995 if (isa<ExtractElementInst>(U))
1996 continue;
1997 if (PrunedDAGInstrs.count(U))
1998 continue;
1999 NeedsExtraction = true;
2000 break;
2001 }
2002
2003 if (NeedsExtraction) {
2004 int ESContrib;
2005 if (Ty1->isVectorTy()) {
2006 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
2007 Ty1, VTy);
2008 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
2009 TargetTransformInfo::SK_ExtractSubvector, VTy, 0, Ty1));
2010 } else
2011 ESContrib = (int) TTI->getVectorInstrCost(
2012 Instruction::ExtractElement, VTy, 0);
2013
2014 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" <<
2015 *S->first << "} = " << ESContrib << "\n");
2016 EffSize -= ESContrib;
2017 }
2018
2019 NeedsExtraction = false;
2020 for (User *U : S->second->users()) {
2021 if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(U)) {
2022 // Shuffle can be folded if it has no other input
2023 if (isa<UndefValue>(SI->getOperand(1)))
2024 continue;
2025 }
2026 if (isa<ExtractElementInst>(U))
2027 continue;
2028 if (PrunedDAGInstrs.count(U))
2029 continue;
2030 NeedsExtraction = true;
2031 break;
2032 }
2033
2034 if (NeedsExtraction) {
2035 int ESContrib;
2036 if (Ty2->isVectorTy()) {
2037 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
2038 Ty2, VTy);
2039 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
2040 TargetTransformInfo::SK_ExtractSubvector, VTy,
2041 Ty1->isVectorTy() ? Ty1->getVectorNumElements() : 1, Ty2));
2042 } else
2043 ESContrib = (int) TTI->getVectorInstrCost(
2044 Instruction::ExtractElement, VTy, 1);
2045 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" <<
2046 *S->second << "} = " << ESContrib << "\n");
2047 EffSize -= ESContrib;
2048 }
2049 }
2050
2051 // Compute the cost of incoming edges.
2052 if (!isa<LoadInst>(S->first) && !isa<StoreInst>(S->first)) {
2053 Instruction *S1 = cast<Instruction>(S->first),
2054 *S2 = cast<Instruction>(S->second);
2055 for (unsigned o = 0; o < S1->getNumOperands(); ++o) {
2056 Value *O1 = S1->getOperand(o), *O2 = S2->getOperand(o);
2057
2058 // Combining constants into vector constants (or small vector
2059 // constants into larger ones are assumed free).
2060 if (isa<Constant>(O1) && isa<Constant>(O2))
2061 continue;
2062
2063 if (FlipOrder)
2064 std::swap(O1, O2);
2065
2066 ValuePair VP = ValuePair(O1, O2);
2067 ValuePair VPR = ValuePair(O2, O1);
2068
2069 // Internal edges are not handled here.
2070 if (PrunedDAG.count(VP) || PrunedDAG.count(VPR))
2071 continue;
2072
2073 Type *Ty1 = O1->getType(),
2074 *Ty2 = O2->getType();
2075 Type *VTy = getVecTypeForPair(Ty1, Ty2);
2076
2077 // Combining vector operations of the same type is also assumed
2078 // folded with other operations.
2079 if (Ty1 == Ty2) {
2080 // If both are insert elements, then both can be widened.
2081 InsertElementInst *IEO1 = dyn_cast<InsertElementInst>(O1),
2082 *IEO2 = dyn_cast<InsertElementInst>(O2);
2083 if (IEO1 && IEO2 && isPureIEChain(IEO1) && isPureIEChain(IEO2))
2084 continue;
2085 // If both are extract elements, and both have the same input
2086 // type, then they can be replaced with a shuffle
2087 ExtractElementInst *EIO1 = dyn_cast<ExtractElementInst>(O1),
2088 *EIO2 = dyn_cast<ExtractElementInst>(O2);
2089 if (EIO1 && EIO2 &&
2090 EIO1->getOperand(0)->getType() ==
2091 EIO2->getOperand(0)->getType())
2092 continue;
2093 // If both are a shuffle with equal operand types and only two
2094 // unqiue operands, then they can be replaced with a single
2095 // shuffle
2096 ShuffleVectorInst *SIO1 = dyn_cast<ShuffleVectorInst>(O1),
2097 *SIO2 = dyn_cast<ShuffleVectorInst>(O2);
2098 if (SIO1 && SIO2 &&
2099 SIO1->getOperand(0)->getType() ==
2100 SIO2->getOperand(0)->getType()) {
2101 SmallSet<Value *, 4> SIOps;
2102 SIOps.insert(SIO1->getOperand(0));
2103 SIOps.insert(SIO1->getOperand(1));
2104 SIOps.insert(SIO2->getOperand(0));
2105 SIOps.insert(SIO2->getOperand(1));
2106 if (SIOps.size() <= 2)
2107 continue;
2108 }
2109 }
2110
2111 int ESContrib;
2112 // This pair has already been formed.
2113 if (IncomingPairs.count(VP)) {
2114 continue;
2115 } else if (IncomingPairs.count(VPR)) {
2116 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
2117 VTy, VTy);
2118
2119 if (VTy->getVectorNumElements() == 2)
2120 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
2121 TargetTransformInfo::SK_Reverse, VTy));
2122 } else if (!Ty1->isVectorTy() && !Ty2->isVectorTy()) {
2123 ESContrib = (int) TTI->getVectorInstrCost(
2124 Instruction::InsertElement, VTy, 0);
2125 ESContrib += (int) TTI->getVectorInstrCost(
2126 Instruction::InsertElement, VTy, 1);
2127 } else if (!Ty1->isVectorTy()) {
2128 // O1 needs to be inserted into a vector of size O2, and then
2129 // both need to be shuffled together.
2130 ESContrib = (int) TTI->getVectorInstrCost(
2131 Instruction::InsertElement, Ty2, 0);
2132 ESContrib += (int) getInstrCost(Instruction::ShuffleVector,
2133 VTy, Ty2);
2134 } else if (!Ty2->isVectorTy()) {
2135 // O2 needs to be inserted into a vector of size O1, and then
2136 // both need to be shuffled together.
2137 ESContrib = (int) TTI->getVectorInstrCost(
2138 Instruction::InsertElement, Ty1, 0);
2139 ESContrib += (int) getInstrCost(Instruction::ShuffleVector,
2140 VTy, Ty1);
2141 } else {
2142 Type *TyBig = Ty1, *TySmall = Ty2;
2143 if (Ty2->getVectorNumElements() > Ty1->getVectorNumElements())
2144 std::swap(TyBig, TySmall);
2145
2146 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
2147 VTy, TyBig);
2148 if (TyBig != TySmall)
2149 ESContrib += (int) getInstrCost(Instruction::ShuffleVector,
2150 TyBig, TySmall);
2151 }
2152
2153 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {"
2154 << *O1 << " <-> " << *O2 << "} = " <<
2155 ESContrib << "\n");
2156 EffSize -= ESContrib;
2157 IncomingPairs.insert(VP);
2158 }
2159 }
2160 }
2161
2162 if (!HasNontrivialInsts) {
2163 DEBUG(if (DebugPairSelection) dbgs() <<
2164 "\tNo non-trivial instructions in DAG;"
2165 " override to zero effective size\n");
2166 EffSize = 0;
2167 }
2168 } else {
2169 for (DenseSet<ValuePair>::iterator S = PrunedDAG.begin(),
2170 E = PrunedDAG.end(); S != E; ++S)
2171 EffSize += (int) getDepthFactor(S->first);
2172 }
2173
2174 DEBUG(if (DebugPairSelection)
2175 dbgs() << "BBV: found pruned DAG for pair {"
2176 << *IJ.first << " <-> " << *IJ.second << "} of depth " <<
2177 MaxDepth << " and size " << PrunedDAG.size() <<
2178 " (effective size: " << EffSize << ")\n");
2179 if (((TTI && !UseChainDepthWithTI) ||
2180 MaxDepth >= Config.ReqChainDepth) &&
2181 EffSize > 0 && EffSize > BestEffSize) {
2182 BestMaxDepth = MaxDepth;
2183 BestEffSize = EffSize;
2184 BestDAG = PrunedDAG;
2185 }
2186 }
2187 }
2188
2189 // Given the list of candidate pairs, this function selects those
2190 // that will be fused into vector instructions.
choosePairs(DenseMap<Value *,std::vector<Value * >> & CandidatePairs,DenseSet<ValuePair> & CandidatePairsSet,DenseMap<ValuePair,int> & CandidatePairCostSavings,std::vector<Value * > & PairableInsts,DenseSet<ValuePair> & FixedOrderPairs,DenseMap<VPPair,unsigned> & PairConnectionTypes,DenseMap<ValuePair,std::vector<ValuePair>> & ConnectedPairs,DenseMap<ValuePair,std::vector<ValuePair>> & ConnectedPairDeps,DenseSet<ValuePair> & PairableInstUsers,DenseMap<Value *,Value * > & ChosenPairs)2191 void BBVectorize::choosePairs(
2192 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
2193 DenseSet<ValuePair> &CandidatePairsSet,
2194 DenseMap<ValuePair, int> &CandidatePairCostSavings,
2195 std::vector<Value *> &PairableInsts,
2196 DenseSet<ValuePair> &FixedOrderPairs,
2197 DenseMap<VPPair, unsigned> &PairConnectionTypes,
2198 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
2199 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps,
2200 DenseSet<ValuePair> &PairableInstUsers,
2201 DenseMap<Value *, Value *>& ChosenPairs) {
2202 bool UseCycleCheck =
2203 CandidatePairsSet.size() <= Config.MaxCandPairsForCycleCheck;
2204
2205 DenseMap<Value *, std::vector<Value *> > CandidatePairs2;
2206 for (DenseSet<ValuePair>::iterator I = CandidatePairsSet.begin(),
2207 E = CandidatePairsSet.end(); I != E; ++I) {
2208 std::vector<Value *> &JJ = CandidatePairs2[I->second];
2209 if (JJ.empty()) JJ.reserve(32);
2210 JJ.push_back(I->first);
2211 }
2212
2213 DenseMap<ValuePair, std::vector<ValuePair> > PairableInstUserMap;
2214 DenseSet<VPPair> PairableInstUserPairSet;
2215 for (std::vector<Value *>::iterator I = PairableInsts.begin(),
2216 E = PairableInsts.end(); I != E; ++I) {
2217 // The number of possible pairings for this variable:
2218 size_t NumChoices = CandidatePairs.lookup(*I).size();
2219 if (!NumChoices) continue;
2220
2221 std::vector<Value *> &JJ = CandidatePairs[*I];
2222
2223 // The best pair to choose and its dag:
2224 size_t BestMaxDepth = 0;
2225 int BestEffSize = 0;
2226 DenseSet<ValuePair> BestDAG;
2227 findBestDAGFor(CandidatePairs, CandidatePairsSet,
2228 CandidatePairCostSavings,
2229 PairableInsts, FixedOrderPairs, PairConnectionTypes,
2230 ConnectedPairs, ConnectedPairDeps,
2231 PairableInstUsers, PairableInstUserMap,
2232 PairableInstUserPairSet, ChosenPairs,
2233 BestDAG, BestMaxDepth, BestEffSize, *I, JJ,
2234 UseCycleCheck);
2235
2236 if (BestDAG.empty())
2237 continue;
2238
2239 // A dag has been chosen (or not) at this point. If no dag was
2240 // chosen, then this instruction, I, cannot be paired (and is no longer
2241 // considered).
2242
2243 DEBUG(dbgs() << "BBV: selected pairs in the best DAG for: "
2244 << *cast<Instruction>(*I) << "\n");
2245
2246 for (DenseSet<ValuePair>::iterator S = BestDAG.begin(),
2247 SE2 = BestDAG.end(); S != SE2; ++S) {
2248 // Insert the members of this dag into the list of chosen pairs.
2249 ChosenPairs.insert(ValuePair(S->first, S->second));
2250 DEBUG(dbgs() << "BBV: selected pair: " << *S->first << " <-> " <<
2251 *S->second << "\n");
2252
2253 // Remove all candidate pairs that have values in the chosen dag.
2254 std::vector<Value *> &KK = CandidatePairs[S->first];
2255 for (std::vector<Value *>::iterator K = KK.begin(), KE = KK.end();
2256 K != KE; ++K) {
2257 if (*K == S->second)
2258 continue;
2259
2260 CandidatePairsSet.erase(ValuePair(S->first, *K));
2261 }
2262
2263 std::vector<Value *> &LL = CandidatePairs2[S->second];
2264 for (std::vector<Value *>::iterator L = LL.begin(), LE = LL.end();
2265 L != LE; ++L) {
2266 if (*L == S->first)
2267 continue;
2268
2269 CandidatePairsSet.erase(ValuePair(*L, S->second));
2270 }
2271
2272 std::vector<Value *> &MM = CandidatePairs[S->second];
2273 for (std::vector<Value *>::iterator M = MM.begin(), ME = MM.end();
2274 M != ME; ++M) {
2275 assert(*M != S->first && "Flipped pair in candidate list?");
2276 CandidatePairsSet.erase(ValuePair(S->second, *M));
2277 }
2278
2279 std::vector<Value *> &NN = CandidatePairs2[S->first];
2280 for (std::vector<Value *>::iterator N = NN.begin(), NE = NN.end();
2281 N != NE; ++N) {
2282 assert(*N != S->second && "Flipped pair in candidate list?");
2283 CandidatePairsSet.erase(ValuePair(*N, S->first));
2284 }
2285 }
2286 }
2287
2288 DEBUG(dbgs() << "BBV: selected " << ChosenPairs.size() << " pairs.\n");
2289 }
2290
getReplacementName(Instruction * I,bool IsInput,unsigned o,unsigned n=0)2291 std::string getReplacementName(Instruction *I, bool IsInput, unsigned o,
2292 unsigned n = 0) {
2293 if (!I->hasName())
2294 return "";
2295
2296 return (I->getName() + (IsInput ? ".v.i" : ".v.r") + utostr(o) +
2297 (n > 0 ? "." + utostr(n) : "")).str();
2298 }
2299
2300 // Returns the value that is to be used as the pointer input to the vector
2301 // instruction that fuses I with J.
getReplacementPointerInput(LLVMContext & Context,Instruction * I,Instruction * J,unsigned o)2302 Value *BBVectorize::getReplacementPointerInput(LLVMContext& Context,
2303 Instruction *I, Instruction *J, unsigned o) {
2304 Value *IPtr, *JPtr;
2305 unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace;
2306 int64_t OffsetInElmts;
2307
2308 // Note: the analysis might fail here, that is why the pair order has
2309 // been precomputed (OffsetInElmts must be unused here).
2310 (void) getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment,
2311 IAddressSpace, JAddressSpace,
2312 OffsetInElmts, false);
2313
2314 // The pointer value is taken to be the one with the lowest offset.
2315 Value *VPtr = IPtr;
2316
2317 Type *ArgTypeI = IPtr->getType()->getPointerElementType();
2318 Type *ArgTypeJ = JPtr->getType()->getPointerElementType();
2319 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2320 Type *VArgPtrType
2321 = PointerType::get(VArgType,
2322 IPtr->getType()->getPointerAddressSpace());
2323 return new BitCastInst(VPtr, VArgPtrType, getReplacementName(I, true, o),
2324 /* insert before */ I);
2325 }
2326
fillNewShuffleMask(LLVMContext & Context,Instruction * J,unsigned MaskOffset,unsigned NumInElem,unsigned NumInElem1,unsigned IdxOffset,std::vector<Constant * > & Mask)2327 void BBVectorize::fillNewShuffleMask(LLVMContext& Context, Instruction *J,
2328 unsigned MaskOffset, unsigned NumInElem,
2329 unsigned NumInElem1, unsigned IdxOffset,
2330 std::vector<Constant*> &Mask) {
2331 unsigned NumElem1 = J->getType()->getVectorNumElements();
2332 for (unsigned v = 0; v < NumElem1; ++v) {
2333 int m = cast<ShuffleVectorInst>(J)->getMaskValue(v);
2334 if (m < 0) {
2335 Mask[v+MaskOffset] = UndefValue::get(Type::getInt32Ty(Context));
2336 } else {
2337 unsigned mm = m + (int) IdxOffset;
2338 if (m >= (int) NumInElem1)
2339 mm += (int) NumInElem;
2340
2341 Mask[v+MaskOffset] =
2342 ConstantInt::get(Type::getInt32Ty(Context), mm);
2343 }
2344 }
2345 }
2346
2347 // Returns the value that is to be used as the vector-shuffle mask to the
2348 // vector instruction that fuses I with J.
getReplacementShuffleMask(LLVMContext & Context,Instruction * I,Instruction * J)2349 Value *BBVectorize::getReplacementShuffleMask(LLVMContext& Context,
2350 Instruction *I, Instruction *J) {
2351 // This is the shuffle mask. We need to append the second
2352 // mask to the first, and the numbers need to be adjusted.
2353
2354 Type *ArgTypeI = I->getType();
2355 Type *ArgTypeJ = J->getType();
2356 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2357
2358 unsigned NumElemI = ArgTypeI->getVectorNumElements();
2359
2360 // Get the total number of elements in the fused vector type.
2361 // By definition, this must equal the number of elements in
2362 // the final mask.
2363 unsigned NumElem = VArgType->getVectorNumElements();
2364 std::vector<Constant*> Mask(NumElem);
2365
2366 Type *OpTypeI = I->getOperand(0)->getType();
2367 unsigned NumInElemI = OpTypeI->getVectorNumElements();
2368 Type *OpTypeJ = J->getOperand(0)->getType();
2369 unsigned NumInElemJ = OpTypeJ->getVectorNumElements();
2370
2371 // The fused vector will be:
2372 // -----------------------------------------------------
2373 // | NumInElemI | NumInElemJ | NumInElemI | NumInElemJ |
2374 // -----------------------------------------------------
2375 // from which we'll extract NumElem total elements (where the first NumElemI
2376 // of them come from the mask in I and the remainder come from the mask
2377 // in J.
2378
2379 // For the mask from the first pair...
2380 fillNewShuffleMask(Context, I, 0, NumInElemJ, NumInElemI,
2381 0, Mask);
2382
2383 // For the mask from the second pair...
2384 fillNewShuffleMask(Context, J, NumElemI, NumInElemI, NumInElemJ,
2385 NumInElemI, Mask);
2386
2387 return ConstantVector::get(Mask);
2388 }
2389
expandIEChain(LLVMContext & Context,Instruction * I,Instruction * J,unsigned o,Value * & LOp,unsigned numElemL,Type * ArgTypeL,Type * ArgTypeH,bool IBeforeJ,unsigned IdxOff)2390 bool BBVectorize::expandIEChain(LLVMContext& Context, Instruction *I,
2391 Instruction *J, unsigned o, Value *&LOp,
2392 unsigned numElemL,
2393 Type *ArgTypeL, Type *ArgTypeH,
2394 bool IBeforeJ, unsigned IdxOff) {
2395 bool ExpandedIEChain = false;
2396 if (InsertElementInst *LIE = dyn_cast<InsertElementInst>(LOp)) {
2397 // If we have a pure insertelement chain, then this can be rewritten
2398 // into a chain that directly builds the larger type.
2399 if (isPureIEChain(LIE)) {
2400 SmallVector<Value *, 8> VectElemts(numElemL,
2401 UndefValue::get(ArgTypeL->getScalarType()));
2402 InsertElementInst *LIENext = LIE;
2403 do {
2404 unsigned Idx =
2405 cast<ConstantInt>(LIENext->getOperand(2))->getSExtValue();
2406 VectElemts[Idx] = LIENext->getOperand(1);
2407 } while ((LIENext =
2408 dyn_cast<InsertElementInst>(LIENext->getOperand(0))));
2409
2410 LIENext = nullptr;
2411 Value *LIEPrev = UndefValue::get(ArgTypeH);
2412 for (unsigned i = 0; i < numElemL; ++i) {
2413 if (isa<UndefValue>(VectElemts[i])) continue;
2414 LIENext = InsertElementInst::Create(LIEPrev, VectElemts[i],
2415 ConstantInt::get(Type::getInt32Ty(Context),
2416 i + IdxOff),
2417 getReplacementName(IBeforeJ ? I : J,
2418 true, o, i+1));
2419 LIENext->insertBefore(IBeforeJ ? J : I);
2420 LIEPrev = LIENext;
2421 }
2422
2423 LOp = LIENext ? (Value*) LIENext : UndefValue::get(ArgTypeH);
2424 ExpandedIEChain = true;
2425 }
2426 }
2427
2428 return ExpandedIEChain;
2429 }
2430
getNumScalarElements(Type * Ty)2431 static unsigned getNumScalarElements(Type *Ty) {
2432 if (VectorType *VecTy = dyn_cast<VectorType>(Ty))
2433 return VecTy->getNumElements();
2434 return 1;
2435 }
2436
2437 // Returns the value to be used as the specified operand of the vector
2438 // instruction that fuses I with J.
getReplacementInput(LLVMContext & Context,Instruction * I,Instruction * J,unsigned o,bool IBeforeJ)2439 Value *BBVectorize::getReplacementInput(LLVMContext& Context, Instruction *I,
2440 Instruction *J, unsigned o, bool IBeforeJ) {
2441 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0);
2442 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), 1);
2443
2444 // Compute the fused vector type for this operand
2445 Type *ArgTypeI = I->getOperand(o)->getType();
2446 Type *ArgTypeJ = J->getOperand(o)->getType();
2447 VectorType *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2448
2449 Instruction *L = I, *H = J;
2450 Type *ArgTypeL = ArgTypeI, *ArgTypeH = ArgTypeJ;
2451
2452 unsigned numElemL = getNumScalarElements(ArgTypeL);
2453 unsigned numElemH = getNumScalarElements(ArgTypeH);
2454
2455 Value *LOp = L->getOperand(o);
2456 Value *HOp = H->getOperand(o);
2457 unsigned numElem = VArgType->getNumElements();
2458
2459 // First, we check if we can reuse the "original" vector outputs (if these
2460 // exist). We might need a shuffle.
2461 ExtractElementInst *LEE = dyn_cast<ExtractElementInst>(LOp);
2462 ExtractElementInst *HEE = dyn_cast<ExtractElementInst>(HOp);
2463 ShuffleVectorInst *LSV = dyn_cast<ShuffleVectorInst>(LOp);
2464 ShuffleVectorInst *HSV = dyn_cast<ShuffleVectorInst>(HOp);
2465
2466 // FIXME: If we're fusing shuffle instructions, then we can't apply this
2467 // optimization. The input vectors to the shuffle might be a different
2468 // length from the shuffle outputs. Unfortunately, the replacement
2469 // shuffle mask has already been formed, and the mask entries are sensitive
2470 // to the sizes of the inputs.
2471 bool IsSizeChangeShuffle =
2472 isa<ShuffleVectorInst>(L) &&
2473 (LOp->getType() != L->getType() || HOp->getType() != H->getType());
2474
2475 if ((LEE || LSV) && (HEE || HSV) && !IsSizeChangeShuffle) {
2476 // We can have at most two unique vector inputs.
2477 bool CanUseInputs = true;
2478 Value *I1, *I2 = nullptr;
2479 if (LEE) {
2480 I1 = LEE->getOperand(0);
2481 } else {
2482 I1 = LSV->getOperand(0);
2483 I2 = LSV->getOperand(1);
2484 if (I2 == I1 || isa<UndefValue>(I2))
2485 I2 = nullptr;
2486 }
2487
2488 if (HEE) {
2489 Value *I3 = HEE->getOperand(0);
2490 if (!I2 && I3 != I1)
2491 I2 = I3;
2492 else if (I3 != I1 && I3 != I2)
2493 CanUseInputs = false;
2494 } else {
2495 Value *I3 = HSV->getOperand(0);
2496 if (!I2 && I3 != I1)
2497 I2 = I3;
2498 else if (I3 != I1 && I3 != I2)
2499 CanUseInputs = false;
2500
2501 if (CanUseInputs) {
2502 Value *I4 = HSV->getOperand(1);
2503 if (!isa<UndefValue>(I4)) {
2504 if (!I2 && I4 != I1)
2505 I2 = I4;
2506 else if (I4 != I1 && I4 != I2)
2507 CanUseInputs = false;
2508 }
2509 }
2510 }
2511
2512 if (CanUseInputs) {
2513 unsigned LOpElem =
2514 cast<Instruction>(LOp)->getOperand(0)->getType()
2515 ->getVectorNumElements();
2516
2517 unsigned HOpElem =
2518 cast<Instruction>(HOp)->getOperand(0)->getType()
2519 ->getVectorNumElements();
2520
2521 // We have one or two input vectors. We need to map each index of the
2522 // operands to the index of the original vector.
2523 SmallVector<std::pair<int, int>, 8> II(numElem);
2524 for (unsigned i = 0; i < numElemL; ++i) {
2525 int Idx, INum;
2526 if (LEE) {
2527 Idx =
2528 cast<ConstantInt>(LEE->getOperand(1))->getSExtValue();
2529 INum = LEE->getOperand(0) == I1 ? 0 : 1;
2530 } else {
2531 Idx = LSV->getMaskValue(i);
2532 if (Idx < (int) LOpElem) {
2533 INum = LSV->getOperand(0) == I1 ? 0 : 1;
2534 } else {
2535 Idx -= LOpElem;
2536 INum = LSV->getOperand(1) == I1 ? 0 : 1;
2537 }
2538 }
2539
2540 II[i] = std::pair<int, int>(Idx, INum);
2541 }
2542 for (unsigned i = 0; i < numElemH; ++i) {
2543 int Idx, INum;
2544 if (HEE) {
2545 Idx =
2546 cast<ConstantInt>(HEE->getOperand(1))->getSExtValue();
2547 INum = HEE->getOperand(0) == I1 ? 0 : 1;
2548 } else {
2549 Idx = HSV->getMaskValue(i);
2550 if (Idx < (int) HOpElem) {
2551 INum = HSV->getOperand(0) == I1 ? 0 : 1;
2552 } else {
2553 Idx -= HOpElem;
2554 INum = HSV->getOperand(1) == I1 ? 0 : 1;
2555 }
2556 }
2557
2558 II[i + numElemL] = std::pair<int, int>(Idx, INum);
2559 }
2560
2561 // We now have an array which tells us from which index of which
2562 // input vector each element of the operand comes.
2563 VectorType *I1T = cast<VectorType>(I1->getType());
2564 unsigned I1Elem = I1T->getNumElements();
2565
2566 if (!I2) {
2567 // In this case there is only one underlying vector input. Check for
2568 // the trivial case where we can use the input directly.
2569 if (I1Elem == numElem) {
2570 bool ElemInOrder = true;
2571 for (unsigned i = 0; i < numElem; ++i) {
2572 if (II[i].first != (int) i && II[i].first != -1) {
2573 ElemInOrder = false;
2574 break;
2575 }
2576 }
2577
2578 if (ElemInOrder)
2579 return I1;
2580 }
2581
2582 // A shuffle is needed.
2583 std::vector<Constant *> Mask(numElem);
2584 for (unsigned i = 0; i < numElem; ++i) {
2585 int Idx = II[i].first;
2586 if (Idx == -1)
2587 Mask[i] = UndefValue::get(Type::getInt32Ty(Context));
2588 else
2589 Mask[i] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
2590 }
2591
2592 Instruction *S =
2593 new ShuffleVectorInst(I1, UndefValue::get(I1T),
2594 ConstantVector::get(Mask),
2595 getReplacementName(IBeforeJ ? I : J,
2596 true, o));
2597 S->insertBefore(IBeforeJ ? J : I);
2598 return S;
2599 }
2600
2601 VectorType *I2T = cast<VectorType>(I2->getType());
2602 unsigned I2Elem = I2T->getNumElements();
2603
2604 // This input comes from two distinct vectors. The first step is to
2605 // make sure that both vectors are the same length. If not, the
2606 // smaller one will need to grow before they can be shuffled together.
2607 if (I1Elem < I2Elem) {
2608 std::vector<Constant *> Mask(I2Elem);
2609 unsigned v = 0;
2610 for (; v < I1Elem; ++v)
2611 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2612 for (; v < I2Elem; ++v)
2613 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2614
2615 Instruction *NewI1 =
2616 new ShuffleVectorInst(I1, UndefValue::get(I1T),
2617 ConstantVector::get(Mask),
2618 getReplacementName(IBeforeJ ? I : J,
2619 true, o, 1));
2620 NewI1->insertBefore(IBeforeJ ? J : I);
2621 I1 = NewI1;
2622 I1Elem = I2Elem;
2623 } else if (I1Elem > I2Elem) {
2624 std::vector<Constant *> Mask(I1Elem);
2625 unsigned v = 0;
2626 for (; v < I2Elem; ++v)
2627 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2628 for (; v < I1Elem; ++v)
2629 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2630
2631 Instruction *NewI2 =
2632 new ShuffleVectorInst(I2, UndefValue::get(I2T),
2633 ConstantVector::get(Mask),
2634 getReplacementName(IBeforeJ ? I : J,
2635 true, o, 1));
2636 NewI2->insertBefore(IBeforeJ ? J : I);
2637 I2 = NewI2;
2638 }
2639
2640 // Now that both I1 and I2 are the same length we can shuffle them
2641 // together (and use the result).
2642 std::vector<Constant *> Mask(numElem);
2643 for (unsigned v = 0; v < numElem; ++v) {
2644 if (II[v].first == -1) {
2645 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2646 } else {
2647 int Idx = II[v].first + II[v].second * I1Elem;
2648 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
2649 }
2650 }
2651
2652 Instruction *NewOp =
2653 new ShuffleVectorInst(I1, I2, ConstantVector::get(Mask),
2654 getReplacementName(IBeforeJ ? I : J, true, o));
2655 NewOp->insertBefore(IBeforeJ ? J : I);
2656 return NewOp;
2657 }
2658 }
2659
2660 Type *ArgType = ArgTypeL;
2661 if (numElemL < numElemH) {
2662 if (numElemL == 1 && expandIEChain(Context, I, J, o, HOp, numElemH,
2663 ArgTypeL, VArgType, IBeforeJ, 1)) {
2664 // This is another short-circuit case: we're combining a scalar into
2665 // a vector that is formed by an IE chain. We've just expanded the IE
2666 // chain, now insert the scalar and we're done.
2667
2668 Instruction *S = InsertElementInst::Create(HOp, LOp, CV0,
2669 getReplacementName(IBeforeJ ? I : J, true, o));
2670 S->insertBefore(IBeforeJ ? J : I);
2671 return S;
2672 } else if (!expandIEChain(Context, I, J, o, LOp, numElemL, ArgTypeL,
2673 ArgTypeH, IBeforeJ)) {
2674 // The two vector inputs to the shuffle must be the same length,
2675 // so extend the smaller vector to be the same length as the larger one.
2676 Instruction *NLOp;
2677 if (numElemL > 1) {
2678
2679 std::vector<Constant *> Mask(numElemH);
2680 unsigned v = 0;
2681 for (; v < numElemL; ++v)
2682 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2683 for (; v < numElemH; ++v)
2684 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2685
2686 NLOp = new ShuffleVectorInst(LOp, UndefValue::get(ArgTypeL),
2687 ConstantVector::get(Mask),
2688 getReplacementName(IBeforeJ ? I : J,
2689 true, o, 1));
2690 } else {
2691 NLOp = InsertElementInst::Create(UndefValue::get(ArgTypeH), LOp, CV0,
2692 getReplacementName(IBeforeJ ? I : J,
2693 true, o, 1));
2694 }
2695
2696 NLOp->insertBefore(IBeforeJ ? J : I);
2697 LOp = NLOp;
2698 }
2699
2700 ArgType = ArgTypeH;
2701 } else if (numElemL > numElemH) {
2702 if (numElemH == 1 && expandIEChain(Context, I, J, o, LOp, numElemL,
2703 ArgTypeH, VArgType, IBeforeJ)) {
2704 Instruction *S =
2705 InsertElementInst::Create(LOp, HOp,
2706 ConstantInt::get(Type::getInt32Ty(Context),
2707 numElemL),
2708 getReplacementName(IBeforeJ ? I : J,
2709 true, o));
2710 S->insertBefore(IBeforeJ ? J : I);
2711 return S;
2712 } else if (!expandIEChain(Context, I, J, o, HOp, numElemH, ArgTypeH,
2713 ArgTypeL, IBeforeJ)) {
2714 Instruction *NHOp;
2715 if (numElemH > 1) {
2716 std::vector<Constant *> Mask(numElemL);
2717 unsigned v = 0;
2718 for (; v < numElemH; ++v)
2719 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2720 for (; v < numElemL; ++v)
2721 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2722
2723 NHOp = new ShuffleVectorInst(HOp, UndefValue::get(ArgTypeH),
2724 ConstantVector::get(Mask),
2725 getReplacementName(IBeforeJ ? I : J,
2726 true, o, 1));
2727 } else {
2728 NHOp = InsertElementInst::Create(UndefValue::get(ArgTypeL), HOp, CV0,
2729 getReplacementName(IBeforeJ ? I : J,
2730 true, o, 1));
2731 }
2732
2733 NHOp->insertBefore(IBeforeJ ? J : I);
2734 HOp = NHOp;
2735 }
2736 }
2737
2738 if (ArgType->isVectorTy()) {
2739 unsigned numElem = VArgType->getVectorNumElements();
2740 std::vector<Constant*> Mask(numElem);
2741 for (unsigned v = 0; v < numElem; ++v) {
2742 unsigned Idx = v;
2743 // If the low vector was expanded, we need to skip the extra
2744 // undefined entries.
2745 if (v >= numElemL && numElemH > numElemL)
2746 Idx += (numElemH - numElemL);
2747 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
2748 }
2749
2750 Instruction *BV = new ShuffleVectorInst(LOp, HOp,
2751 ConstantVector::get(Mask),
2752 getReplacementName(IBeforeJ ? I : J, true, o));
2753 BV->insertBefore(IBeforeJ ? J : I);
2754 return BV;
2755 }
2756
2757 Instruction *BV1 = InsertElementInst::Create(
2758 UndefValue::get(VArgType), LOp, CV0,
2759 getReplacementName(IBeforeJ ? I : J,
2760 true, o, 1));
2761 BV1->insertBefore(IBeforeJ ? J : I);
2762 Instruction *BV2 = InsertElementInst::Create(BV1, HOp, CV1,
2763 getReplacementName(IBeforeJ ? I : J,
2764 true, o, 2));
2765 BV2->insertBefore(IBeforeJ ? J : I);
2766 return BV2;
2767 }
2768
2769 // This function creates an array of values that will be used as the inputs
2770 // to the vector instruction that fuses I with J.
getReplacementInputsForPair(LLVMContext & Context,Instruction * I,Instruction * J,SmallVectorImpl<Value * > & ReplacedOperands,bool IBeforeJ)2771 void BBVectorize::getReplacementInputsForPair(LLVMContext& Context,
2772 Instruction *I, Instruction *J,
2773 SmallVectorImpl<Value *> &ReplacedOperands,
2774 bool IBeforeJ) {
2775 unsigned NumOperands = I->getNumOperands();
2776
2777 for (unsigned p = 0, o = NumOperands-1; p < NumOperands; ++p, --o) {
2778 // Iterate backward so that we look at the store pointer
2779 // first and know whether or not we need to flip the inputs.
2780
2781 if (isa<LoadInst>(I) || (o == 1 && isa<StoreInst>(I))) {
2782 // This is the pointer for a load/store instruction.
2783 ReplacedOperands[o] = getReplacementPointerInput(Context, I, J, o);
2784 continue;
2785 } else if (isa<CallInst>(I)) {
2786 Function *F = cast<CallInst>(I)->getCalledFunction();
2787 Intrinsic::ID IID = F->getIntrinsicID();
2788 if (o == NumOperands-1) {
2789 BasicBlock &BB = *I->getParent();
2790
2791 Module *M = BB.getParent()->getParent();
2792 Type *ArgTypeI = I->getType();
2793 Type *ArgTypeJ = J->getType();
2794 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2795
2796 ReplacedOperands[o] = Intrinsic::getDeclaration(M, IID, VArgType);
2797 continue;
2798 } else if ((IID == Intrinsic::powi || IID == Intrinsic::ctlz ||
2799 IID == Intrinsic::cttz) && o == 1) {
2800 // The second argument of powi/ctlz/cttz is a single integer/constant
2801 // and we've already checked that both arguments are equal.
2802 // As a result, we just keep I's second argument.
2803 ReplacedOperands[o] = I->getOperand(o);
2804 continue;
2805 }
2806 } else if (isa<ShuffleVectorInst>(I) && o == NumOperands-1) {
2807 ReplacedOperands[o] = getReplacementShuffleMask(Context, I, J);
2808 continue;
2809 }
2810
2811 ReplacedOperands[o] = getReplacementInput(Context, I, J, o, IBeforeJ);
2812 }
2813 }
2814
2815 // This function creates two values that represent the outputs of the
2816 // original I and J instructions. These are generally vector shuffles
2817 // or extracts. In many cases, these will end up being unused and, thus,
2818 // eliminated by later passes.
replaceOutputsOfPair(LLVMContext & Context,Instruction * I,Instruction * J,Instruction * K,Instruction * & InsertionPt,Instruction * & K1,Instruction * & K2)2819 void BBVectorize::replaceOutputsOfPair(LLVMContext& Context, Instruction *I,
2820 Instruction *J, Instruction *K,
2821 Instruction *&InsertionPt,
2822 Instruction *&K1, Instruction *&K2) {
2823 if (isa<StoreInst>(I))
2824 return;
2825
2826 Type *IType = I->getType();
2827 Type *JType = J->getType();
2828
2829 VectorType *VType = getVecTypeForPair(IType, JType);
2830 unsigned numElem = VType->getNumElements();
2831
2832 unsigned numElemI = getNumScalarElements(IType);
2833 unsigned numElemJ = getNumScalarElements(JType);
2834
2835 if (IType->isVectorTy()) {
2836 std::vector<Constant *> Mask1(numElemI), Mask2(numElemI);
2837 for (unsigned v = 0; v < numElemI; ++v) {
2838 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2839 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemJ + v);
2840 }
2841
2842 K1 = new ShuffleVectorInst(K, UndefValue::get(VType),
2843 ConstantVector::get(Mask1),
2844 getReplacementName(K, false, 1));
2845 } else {
2846 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0);
2847 K1 = ExtractElementInst::Create(K, CV0, getReplacementName(K, false, 1));
2848 }
2849
2850 if (JType->isVectorTy()) {
2851 std::vector<Constant *> Mask1(numElemJ), Mask2(numElemJ);
2852 for (unsigned v = 0; v < numElemJ; ++v) {
2853 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2854 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemI + v);
2855 }
2856
2857 K2 = new ShuffleVectorInst(K, UndefValue::get(VType),
2858 ConstantVector::get(Mask2),
2859 getReplacementName(K, false, 2));
2860 } else {
2861 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), numElem - 1);
2862 K2 = ExtractElementInst::Create(K, CV1, getReplacementName(K, false, 2));
2863 }
2864
2865 K1->insertAfter(K);
2866 K2->insertAfter(K1);
2867 InsertionPt = K2;
2868 }
2869
2870 // Move all uses of the function I (including pairing-induced uses) after J.
canMoveUsesOfIAfterJ(BasicBlock & BB,DenseSet<ValuePair> & LoadMoveSetPairs,Instruction * I,Instruction * J)2871 bool BBVectorize::canMoveUsesOfIAfterJ(BasicBlock &BB,
2872 DenseSet<ValuePair> &LoadMoveSetPairs,
2873 Instruction *I, Instruction *J) {
2874 // Skip to the first instruction past I.
2875 BasicBlock::iterator L = std::next(BasicBlock::iterator(I));
2876
2877 DenseSet<Value *> Users;
2878 AliasSetTracker WriteSet(*AA);
2879 if (I->mayWriteToMemory()) WriteSet.add(I);
2880
2881 for (; cast<Instruction>(L) != J; ++L)
2882 (void)trackUsesOfI(Users, WriteSet, I, &*L, true, &LoadMoveSetPairs);
2883
2884 assert(cast<Instruction>(L) == J &&
2885 "Tracking has not proceeded far enough to check for dependencies");
2886 // If J is now in the use set of I, then trackUsesOfI will return true
2887 // and we have a dependency cycle (and the fusing operation must abort).
2888 return !trackUsesOfI(Users, WriteSet, I, J, true, &LoadMoveSetPairs);
2889 }
2890
2891 // Move all uses of the function I (including pairing-induced uses) after J.
moveUsesOfIAfterJ(BasicBlock & BB,DenseSet<ValuePair> & LoadMoveSetPairs,Instruction * & InsertionPt,Instruction * I,Instruction * J)2892 void BBVectorize::moveUsesOfIAfterJ(BasicBlock &BB,
2893 DenseSet<ValuePair> &LoadMoveSetPairs,
2894 Instruction *&InsertionPt,
2895 Instruction *I, Instruction *J) {
2896 // Skip to the first instruction past I.
2897 BasicBlock::iterator L = std::next(BasicBlock::iterator(I));
2898
2899 DenseSet<Value *> Users;
2900 AliasSetTracker WriteSet(*AA);
2901 if (I->mayWriteToMemory()) WriteSet.add(I);
2902
2903 for (; cast<Instruction>(L) != J;) {
2904 if (trackUsesOfI(Users, WriteSet, I, &*L, true, &LoadMoveSetPairs)) {
2905 // Move this instruction
2906 Instruction *InstToMove = &*L++;
2907
2908 DEBUG(dbgs() << "BBV: moving: " << *InstToMove <<
2909 " to after " << *InsertionPt << "\n");
2910 InstToMove->removeFromParent();
2911 InstToMove->insertAfter(InsertionPt);
2912 InsertionPt = InstToMove;
2913 } else {
2914 ++L;
2915 }
2916 }
2917 }
2918
2919 // Collect all load instruction that are in the move set of a given first
2920 // pair member. These loads depend on the first instruction, I, and so need
2921 // to be moved after J (the second instruction) when the pair is fused.
collectPairLoadMoveSet(BasicBlock & BB,DenseMap<Value *,Value * > & ChosenPairs,DenseMap<Value *,std::vector<Value * >> & LoadMoveSet,DenseSet<ValuePair> & LoadMoveSetPairs,Instruction * I)2922 void BBVectorize::collectPairLoadMoveSet(BasicBlock &BB,
2923 DenseMap<Value *, Value *> &ChosenPairs,
2924 DenseMap<Value *, std::vector<Value *> > &LoadMoveSet,
2925 DenseSet<ValuePair> &LoadMoveSetPairs,
2926 Instruction *I) {
2927 // Skip to the first instruction past I.
2928 BasicBlock::iterator L = std::next(BasicBlock::iterator(I));
2929
2930 DenseSet<Value *> Users;
2931 AliasSetTracker WriteSet(*AA);
2932 if (I->mayWriteToMemory()) WriteSet.add(I);
2933
2934 // Note: We cannot end the loop when we reach J because J could be moved
2935 // farther down the use chain by another instruction pairing. Also, J
2936 // could be before I if this is an inverted input.
2937 for (BasicBlock::iterator E = BB.end(); L != E; ++L) {
2938 if (trackUsesOfI(Users, WriteSet, I, &*L)) {
2939 if (L->mayReadFromMemory()) {
2940 LoadMoveSet[&*L].push_back(I);
2941 LoadMoveSetPairs.insert(ValuePair(&*L, I));
2942 }
2943 }
2944 }
2945 }
2946
2947 // In cases where both load/stores and the computation of their pointers
2948 // are chosen for vectorization, we can end up in a situation where the
2949 // aliasing analysis starts returning different query results as the
2950 // process of fusing instruction pairs continues. Because the algorithm
2951 // relies on finding the same use dags here as were found earlier, we'll
2952 // need to precompute the necessary aliasing information here and then
2953 // manually update it during the fusion process.
collectLoadMoveSet(BasicBlock & BB,std::vector<Value * > & PairableInsts,DenseMap<Value *,Value * > & ChosenPairs,DenseMap<Value *,std::vector<Value * >> & LoadMoveSet,DenseSet<ValuePair> & LoadMoveSetPairs)2954 void BBVectorize::collectLoadMoveSet(BasicBlock &BB,
2955 std::vector<Value *> &PairableInsts,
2956 DenseMap<Value *, Value *> &ChosenPairs,
2957 DenseMap<Value *, std::vector<Value *> > &LoadMoveSet,
2958 DenseSet<ValuePair> &LoadMoveSetPairs) {
2959 for (std::vector<Value *>::iterator PI = PairableInsts.begin(),
2960 PIE = PairableInsts.end(); PI != PIE; ++PI) {
2961 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(*PI);
2962 if (P == ChosenPairs.end()) continue;
2963
2964 Instruction *I = cast<Instruction>(P->first);
2965 collectPairLoadMoveSet(BB, ChosenPairs, LoadMoveSet,
2966 LoadMoveSetPairs, I);
2967 }
2968 }
2969
2970 // This function fuses the chosen instruction pairs into vector instructions,
2971 // taking care preserve any needed scalar outputs and, then, it reorders the
2972 // remaining instructions as needed (users of the first member of the pair
2973 // need to be moved to after the location of the second member of the pair
2974 // because the vector instruction is inserted in the location of the pair's
2975 // second member).
fuseChosenPairs(BasicBlock & BB,std::vector<Value * > & PairableInsts,DenseMap<Value *,Value * > & ChosenPairs,DenseSet<ValuePair> & FixedOrderPairs,DenseMap<VPPair,unsigned> & PairConnectionTypes,DenseMap<ValuePair,std::vector<ValuePair>> & ConnectedPairs,DenseMap<ValuePair,std::vector<ValuePair>> & ConnectedPairDeps)2976 void BBVectorize::fuseChosenPairs(BasicBlock &BB,
2977 std::vector<Value *> &PairableInsts,
2978 DenseMap<Value *, Value *> &ChosenPairs,
2979 DenseSet<ValuePair> &FixedOrderPairs,
2980 DenseMap<VPPair, unsigned> &PairConnectionTypes,
2981 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
2982 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps) {
2983 LLVMContext& Context = BB.getContext();
2984
2985 // During the vectorization process, the order of the pairs to be fused
2986 // could be flipped. So we'll add each pair, flipped, into the ChosenPairs
2987 // list. After a pair is fused, the flipped pair is removed from the list.
2988 DenseSet<ValuePair> FlippedPairs;
2989 for (DenseMap<Value *, Value *>::iterator P = ChosenPairs.begin(),
2990 E = ChosenPairs.end(); P != E; ++P)
2991 FlippedPairs.insert(ValuePair(P->second, P->first));
2992 for (DenseSet<ValuePair>::iterator P = FlippedPairs.begin(),
2993 E = FlippedPairs.end(); P != E; ++P)
2994 ChosenPairs.insert(*P);
2995
2996 DenseMap<Value *, std::vector<Value *> > LoadMoveSet;
2997 DenseSet<ValuePair> LoadMoveSetPairs;
2998 collectLoadMoveSet(BB, PairableInsts, ChosenPairs,
2999 LoadMoveSet, LoadMoveSetPairs);
3000
3001 DEBUG(dbgs() << "BBV: initial: \n" << BB << "\n");
3002
3003 for (BasicBlock::iterator PI = BB.getFirstInsertionPt(); PI != BB.end();) {
3004 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(&*PI);
3005 if (P == ChosenPairs.end()) {
3006 ++PI;
3007 continue;
3008 }
3009
3010 if (getDepthFactor(P->first) == 0) {
3011 // These instructions are not really fused, but are tracked as though
3012 // they are. Any case in which it would be interesting to fuse them
3013 // will be taken care of by InstCombine.
3014 --NumFusedOps;
3015 ++PI;
3016 continue;
3017 }
3018
3019 Instruction *I = cast<Instruction>(P->first),
3020 *J = cast<Instruction>(P->second);
3021
3022 DEBUG(dbgs() << "BBV: fusing: " << *I <<
3023 " <-> " << *J << "\n");
3024
3025 // Remove the pair and flipped pair from the list.
3026 DenseMap<Value *, Value *>::iterator FP = ChosenPairs.find(P->second);
3027 assert(FP != ChosenPairs.end() && "Flipped pair not found in list");
3028 ChosenPairs.erase(FP);
3029 ChosenPairs.erase(P);
3030
3031 if (!canMoveUsesOfIAfterJ(BB, LoadMoveSetPairs, I, J)) {
3032 DEBUG(dbgs() << "BBV: fusion of: " << *I <<
3033 " <-> " << *J <<
3034 " aborted because of non-trivial dependency cycle\n");
3035 --NumFusedOps;
3036 ++PI;
3037 continue;
3038 }
3039
3040 // If the pair must have the other order, then flip it.
3041 bool FlipPairOrder = FixedOrderPairs.count(ValuePair(J, I));
3042 if (!FlipPairOrder && !FixedOrderPairs.count(ValuePair(I, J))) {
3043 // This pair does not have a fixed order, and so we might want to
3044 // flip it if that will yield fewer shuffles. We count the number
3045 // of dependencies connected via swaps, and those directly connected,
3046 // and flip the order if the number of swaps is greater.
3047 bool OrigOrder = true;
3048 DenseMap<ValuePair, std::vector<ValuePair> >::iterator IJ =
3049 ConnectedPairDeps.find(ValuePair(I, J));
3050 if (IJ == ConnectedPairDeps.end()) {
3051 IJ = ConnectedPairDeps.find(ValuePair(J, I));
3052 OrigOrder = false;
3053 }
3054
3055 if (IJ != ConnectedPairDeps.end()) {
3056 unsigned NumDepsDirect = 0, NumDepsSwap = 0;
3057 for (std::vector<ValuePair>::iterator T = IJ->second.begin(),
3058 TE = IJ->second.end(); T != TE; ++T) {
3059 VPPair Q(IJ->first, *T);
3060 DenseMap<VPPair, unsigned>::iterator R =
3061 PairConnectionTypes.find(VPPair(Q.second, Q.first));
3062 assert(R != PairConnectionTypes.end() &&
3063 "Cannot find pair connection type");
3064 if (R->second == PairConnectionDirect)
3065 ++NumDepsDirect;
3066 else if (R->second == PairConnectionSwap)
3067 ++NumDepsSwap;
3068 }
3069
3070 if (!OrigOrder)
3071 std::swap(NumDepsDirect, NumDepsSwap);
3072
3073 if (NumDepsSwap > NumDepsDirect) {
3074 FlipPairOrder = true;
3075 DEBUG(dbgs() << "BBV: reordering pair: " << *I <<
3076 " <-> " << *J << "\n");
3077 }
3078 }
3079 }
3080
3081 Instruction *L = I, *H = J;
3082 if (FlipPairOrder)
3083 std::swap(H, L);
3084
3085 // If the pair being fused uses the opposite order from that in the pair
3086 // connection map, then we need to flip the types.
3087 DenseMap<ValuePair, std::vector<ValuePair> >::iterator HL =
3088 ConnectedPairs.find(ValuePair(H, L));
3089 if (HL != ConnectedPairs.end())
3090 for (std::vector<ValuePair>::iterator T = HL->second.begin(),
3091 TE = HL->second.end(); T != TE; ++T) {
3092 VPPair Q(HL->first, *T);
3093 DenseMap<VPPair, unsigned>::iterator R = PairConnectionTypes.find(Q);
3094 assert(R != PairConnectionTypes.end() &&
3095 "Cannot find pair connection type");
3096 if (R->second == PairConnectionDirect)
3097 R->second = PairConnectionSwap;
3098 else if (R->second == PairConnectionSwap)
3099 R->second = PairConnectionDirect;
3100 }
3101
3102 bool LBeforeH = !FlipPairOrder;
3103 unsigned NumOperands = I->getNumOperands();
3104 SmallVector<Value *, 3> ReplacedOperands(NumOperands);
3105 getReplacementInputsForPair(Context, L, H, ReplacedOperands,
3106 LBeforeH);
3107
3108 // Make a copy of the original operation, change its type to the vector
3109 // type and replace its operands with the vector operands.
3110 Instruction *K = L->clone();
3111 if (L->hasName())
3112 K->takeName(L);
3113 else if (H->hasName())
3114 K->takeName(H);
3115
3116 if (auto CS = CallSite(K)) {
3117 SmallVector<Type *, 3> Tys;
3118 FunctionType *Old = CS.getFunctionType();
3119 unsigned NumOld = Old->getNumParams();
3120 assert(NumOld <= ReplacedOperands.size());
3121 for (unsigned i = 0; i != NumOld; ++i)
3122 Tys.push_back(ReplacedOperands[i]->getType());
3123 CS.mutateFunctionType(
3124 FunctionType::get(getVecTypeForPair(L->getType(), H->getType()),
3125 Tys, Old->isVarArg()));
3126 } else if (!isa<StoreInst>(K))
3127 K->mutateType(getVecTypeForPair(L->getType(), H->getType()));
3128
3129 unsigned KnownIDs[] = {LLVMContext::MD_tbaa, LLVMContext::MD_alias_scope,
3130 LLVMContext::MD_noalias, LLVMContext::MD_fpmath,
3131 LLVMContext::MD_invariant_group};
3132 combineMetadata(K, H, KnownIDs);
3133 K->intersectOptionalDataWith(H);
3134
3135 for (unsigned o = 0; o < NumOperands; ++o)
3136 K->setOperand(o, ReplacedOperands[o]);
3137
3138 K->insertAfter(J);
3139
3140 // Instruction insertion point:
3141 Instruction *InsertionPt = K;
3142 Instruction *K1 = nullptr, *K2 = nullptr;
3143 replaceOutputsOfPair(Context, L, H, K, InsertionPt, K1, K2);
3144
3145 // The use dag of the first original instruction must be moved to after
3146 // the location of the second instruction. The entire use dag of the
3147 // first instruction is disjoint from the input dag of the second
3148 // (by definition), and so commutes with it.
3149
3150 moveUsesOfIAfterJ(BB, LoadMoveSetPairs, InsertionPt, I, J);
3151
3152 if (!isa<StoreInst>(I)) {
3153 L->replaceAllUsesWith(K1);
3154 H->replaceAllUsesWith(K2);
3155 }
3156
3157 // Instructions that may read from memory may be in the load move set.
3158 // Once an instruction is fused, we no longer need its move set, and so
3159 // the values of the map never need to be updated. However, when a load
3160 // is fused, we need to merge the entries from both instructions in the
3161 // pair in case those instructions were in the move set of some other
3162 // yet-to-be-fused pair. The loads in question are the keys of the map.
3163 if (I->mayReadFromMemory()) {
3164 std::vector<ValuePair> NewSetMembers;
3165 DenseMap<Value *, std::vector<Value *> >::iterator II =
3166 LoadMoveSet.find(I);
3167 if (II != LoadMoveSet.end())
3168 for (std::vector<Value *>::iterator N = II->second.begin(),
3169 NE = II->second.end(); N != NE; ++N)
3170 NewSetMembers.push_back(ValuePair(K, *N));
3171 DenseMap<Value *, std::vector<Value *> >::iterator JJ =
3172 LoadMoveSet.find(J);
3173 if (JJ != LoadMoveSet.end())
3174 for (std::vector<Value *>::iterator N = JJ->second.begin(),
3175 NE = JJ->second.end(); N != NE; ++N)
3176 NewSetMembers.push_back(ValuePair(K, *N));
3177 for (std::vector<ValuePair>::iterator A = NewSetMembers.begin(),
3178 AE = NewSetMembers.end(); A != AE; ++A) {
3179 LoadMoveSet[A->first].push_back(A->second);
3180 LoadMoveSetPairs.insert(*A);
3181 }
3182 }
3183
3184 // Before removing I, set the iterator to the next instruction.
3185 PI = std::next(BasicBlock::iterator(I));
3186 if (cast<Instruction>(PI) == J)
3187 ++PI;
3188
3189 SE->forgetValue(I);
3190 SE->forgetValue(J);
3191 I->eraseFromParent();
3192 J->eraseFromParent();
3193
3194 DEBUG(if (PrintAfterEveryPair) dbgs() << "BBV: block is now: \n" <<
3195 BB << "\n");
3196 }
3197
3198 DEBUG(dbgs() << "BBV: final: \n" << BB << "\n");
3199 }
3200 }
3201
3202 char BBVectorize::ID = 0;
3203 static const char bb_vectorize_name[] = "Basic-Block Vectorization";
INITIALIZE_PASS_BEGIN(BBVectorize,BBV_NAME,bb_vectorize_name,false,false)3204 INITIALIZE_PASS_BEGIN(BBVectorize, BBV_NAME, bb_vectorize_name, false, false)
3205 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
3206 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
3207 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
3208 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
3209 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
3210 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
3211 INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)
3212 INITIALIZE_PASS_DEPENDENCY(SCEVAAWrapperPass)
3213 INITIALIZE_PASS_END(BBVectorize, BBV_NAME, bb_vectorize_name, false, false)
3214
3215 BasicBlockPass *llvm::createBBVectorizePass(const VectorizeConfig &C) {
3216 return new BBVectorize(C);
3217 }
3218
3219 bool
vectorizeBasicBlock(Pass * P,BasicBlock & BB,const VectorizeConfig & C)3220 llvm::vectorizeBasicBlock(Pass *P, BasicBlock &BB, const VectorizeConfig &C) {
3221 BBVectorize BBVectorizer(P, *BB.getParent(), C);
3222 return BBVectorizer.vectorizeBB(BB);
3223 }
3224
3225 //===----------------------------------------------------------------------===//
VectorizeConfig()3226 VectorizeConfig::VectorizeConfig() {
3227 VectorBits = ::VectorBits;
3228 VectorizeBools = !::NoBools;
3229 VectorizeInts = !::NoInts;
3230 VectorizeFloats = !::NoFloats;
3231 VectorizePointers = !::NoPointers;
3232 VectorizeCasts = !::NoCasts;
3233 VectorizeMath = !::NoMath;
3234 VectorizeBitManipulations = !::NoBitManipulation;
3235 VectorizeFMA = !::NoFMA;
3236 VectorizeSelect = !::NoSelect;
3237 VectorizeCmp = !::NoCmp;
3238 VectorizeGEP = !::NoGEP;
3239 VectorizeMemOps = !::NoMemOps;
3240 AlignedOnly = ::AlignedOnly;
3241 ReqChainDepth= ::ReqChainDepth;
3242 SearchLimit = ::SearchLimit;
3243 MaxCandPairsForCycleCheck = ::MaxCandPairsForCycleCheck;
3244 SplatBreaksChain = ::SplatBreaksChain;
3245 MaxInsts = ::MaxInsts;
3246 MaxPairs = ::MaxPairs;
3247 MaxIter = ::MaxIter;
3248 Pow2LenOnly = ::Pow2LenOnly;
3249 NoMemOpBoost = ::NoMemOpBoost;
3250 FastDep = ::FastDep;
3251 }
3252