1 //===- BBVectorize.cpp - A Basic-Block Vectorizer -------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
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
15 //===----------------------------------------------------------------------===//
17 #define BBV_NAME "bb-vectorize"
18 #define DEBUG_TYPE BBV_NAME
19 #include "llvm/Transforms/Vectorize.h"
20 #include "llvm/ADT/DenseMap.h"
21 #include "llvm/ADT/DenseSet.h"
22 #include "llvm/ADT/STLExtras.h"
23 #include "llvm/ADT/SmallSet.h"
24 #include "llvm/ADT/SmallVector.h"
25 #include "llvm/ADT/Statistic.h"
26 #include "llvm/ADT/StringExtras.h"
27 #include "llvm/Analysis/AliasAnalysis.h"
28 #include "llvm/Analysis/AliasSetTracker.h"
29 #include "llvm/Analysis/Dominators.h"
30 #include "llvm/Analysis/ScalarEvolution.h"
31 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
32 #include "llvm/Analysis/TargetTransformInfo.h"
33 #include "llvm/Analysis/ValueTracking.h"
34 #include "llvm/IR/Constants.h"
35 #include "llvm/IR/DataLayout.h"
36 #include "llvm/IR/DerivedTypes.h"
37 #include "llvm/IR/Function.h"
38 #include "llvm/IR/Instructions.h"
39 #include "llvm/IR/IntrinsicInst.h"
40 #include "llvm/IR/Intrinsics.h"
41 #include "llvm/IR/LLVMContext.h"
42 #include "llvm/IR/Metadata.h"
43 #include "llvm/IR/Type.h"
44 #include "llvm/Pass.h"
45 #include "llvm/Support/CommandLine.h"
46 #include "llvm/Support/Debug.h"
47 #include "llvm/Support/ValueHandle.h"
48 #include "llvm/Support/raw_ostream.h"
49 #include "llvm/Transforms/Utils/Local.h"
54 IgnoreTargetInfo("bb-vectorize-ignore-target-info", cl::init(false),
55 cl::Hidden, cl::desc("Ignore target information"));
57 static cl::opt<unsigned>
58 ReqChainDepth("bb-vectorize-req-chain-depth", cl::init(6), cl::Hidden,
59 cl::desc("The required chain depth for vectorization"));
62 UseChainDepthWithTI("bb-vectorize-use-chain-depth", cl::init(false),
63 cl::Hidden, cl::desc("Use the chain depth requirement with"
64 " target information"));
66 static cl::opt<unsigned>
67 SearchLimit("bb-vectorize-search-limit", cl::init(400), cl::Hidden,
68 cl::desc("The maximum search distance for instruction pairs"));
71 SplatBreaksChain("bb-vectorize-splat-breaks-chain", cl::init(false), cl::Hidden,
72 cl::desc("Replicating one element to a pair breaks the chain"));
74 static cl::opt<unsigned>
75 VectorBits("bb-vectorize-vector-bits", cl::init(128), cl::Hidden,
76 cl::desc("The size of the native vector registers"));
78 static cl::opt<unsigned>
79 MaxIter("bb-vectorize-max-iter", cl::init(0), cl::Hidden,
80 cl::desc("The maximum number of pairing iterations"));
83 Pow2LenOnly("bb-vectorize-pow2-len-only", cl::init(false), cl::Hidden,
84 cl::desc("Don't try to form non-2^n-length vectors"));
86 static cl::opt<unsigned>
87 MaxInsts("bb-vectorize-max-instr-per-group", cl::init(500), cl::Hidden,
88 cl::desc("The maximum number of pairable instructions per group"));
90 static cl::opt<unsigned>
91 MaxCandPairsForCycleCheck("bb-vectorize-max-cycle-check-pairs", cl::init(200),
92 cl::Hidden, cl::desc("The maximum number of candidate pairs with which to use"
93 " a full cycle check"));
96 NoBools("bb-vectorize-no-bools", cl::init(false), cl::Hidden,
97 cl::desc("Don't try to vectorize boolean (i1) values"));
100 NoInts("bb-vectorize-no-ints", cl::init(false), cl::Hidden,
101 cl::desc("Don't try to vectorize integer values"));
104 NoFloats("bb-vectorize-no-floats", cl::init(false), cl::Hidden,
105 cl::desc("Don't try to vectorize floating-point values"));
107 // FIXME: This should default to false once pointer vector support works.
109 NoPointers("bb-vectorize-no-pointers", cl::init(/*false*/ true), cl::Hidden,
110 cl::desc("Don't try to vectorize pointer values"));
113 NoCasts("bb-vectorize-no-casts", cl::init(false), cl::Hidden,
114 cl::desc("Don't try to vectorize casting (conversion) operations"));
117 NoMath("bb-vectorize-no-math", cl::init(false), cl::Hidden,
118 cl::desc("Don't try to vectorize floating-point math intrinsics"));
121 NoFMA("bb-vectorize-no-fma", cl::init(false), cl::Hidden,
122 cl::desc("Don't try to vectorize the fused-multiply-add intrinsic"));
125 NoSelect("bb-vectorize-no-select", cl::init(false), cl::Hidden,
126 cl::desc("Don't try to vectorize select instructions"));
129 NoCmp("bb-vectorize-no-cmp", cl::init(false), cl::Hidden,
130 cl::desc("Don't try to vectorize comparison instructions"));
133 NoGEP("bb-vectorize-no-gep", cl::init(false), cl::Hidden,
134 cl::desc("Don't try to vectorize getelementptr instructions"));
137 NoMemOps("bb-vectorize-no-mem-ops", cl::init(false), cl::Hidden,
138 cl::desc("Don't try to vectorize loads and stores"));
141 AlignedOnly("bb-vectorize-aligned-only", cl::init(false), cl::Hidden,
142 cl::desc("Only generate aligned loads and stores"));
145 NoMemOpBoost("bb-vectorize-no-mem-op-boost",
146 cl::init(false), cl::Hidden,
147 cl::desc("Don't boost the chain-depth contribution of loads and stores"));
150 FastDep("bb-vectorize-fast-dep", cl::init(false), cl::Hidden,
151 cl::desc("Use a fast instruction dependency analysis"));
155 DebugInstructionExamination("bb-vectorize-debug-instruction-examination",
156 cl::init(false), cl::Hidden,
157 cl::desc("When debugging is enabled, output information on the"
158 " instruction-examination process"));
160 DebugCandidateSelection("bb-vectorize-debug-candidate-selection",
161 cl::init(false), cl::Hidden,
162 cl::desc("When debugging is enabled, output information on the"
163 " candidate-selection process"));
165 DebugPairSelection("bb-vectorize-debug-pair-selection",
166 cl::init(false), cl::Hidden,
167 cl::desc("When debugging is enabled, output information on the"
168 " pair-selection process"));
170 DebugCycleCheck("bb-vectorize-debug-cycle-check",
171 cl::init(false), cl::Hidden,
172 cl::desc("When debugging is enabled, output information on the"
173 " cycle-checking process"));
176 PrintAfterEveryPair("bb-vectorize-debug-print-after-every-pair",
177 cl::init(false), cl::Hidden,
178 cl::desc("When debugging is enabled, dump the basic block after"
179 " every pair is fused"));
182 STATISTIC(NumFusedOps, "Number of operations fused by bb-vectorize");
185 struct BBVectorize : public BasicBlockPass {
186 static char ID; // Pass identification, replacement for typeid
188 const VectorizeConfig Config;
190 BBVectorize(const VectorizeConfig &C = VectorizeConfig())
191 : BasicBlockPass(ID), Config(C) {
192 initializeBBVectorizePass(*PassRegistry::getPassRegistry());
195 BBVectorize(Pass *P, const VectorizeConfig &C)
196 : BasicBlockPass(ID), Config(C) {
197 AA = &P->getAnalysis<AliasAnalysis>();
198 DT = &P->getAnalysis<DominatorTree>();
199 SE = &P->getAnalysis<ScalarEvolution>();
200 TD = P->getAnalysisIfAvailable<DataLayout>();
201 TTI = IgnoreTargetInfo ? 0 : &P->getAnalysis<TargetTransformInfo>();
204 typedef std::pair<Value *, Value *> ValuePair;
205 typedef std::pair<ValuePair, int> ValuePairWithCost;
206 typedef std::pair<ValuePair, size_t> ValuePairWithDepth;
207 typedef std::pair<ValuePair, ValuePair> VPPair; // A ValuePair pair
208 typedef std::pair<VPPair, unsigned> VPPairWithType;
214 const TargetTransformInfo *TTI;
216 // FIXME: const correct?
218 bool vectorizePairs(BasicBlock &BB, bool NonPow2Len = false);
220 bool getCandidatePairs(BasicBlock &BB,
221 BasicBlock::iterator &Start,
222 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
223 DenseSet<ValuePair> &FixedOrderPairs,
224 DenseMap<ValuePair, int> &CandidatePairCostSavings,
225 std::vector<Value *> &PairableInsts, bool NonPow2Len);
227 // FIXME: The current implementation does not account for pairs that
228 // are connected in multiple ways. For example:
229 // C1 = A1 / A2; C2 = A2 / A1 (which may be both direct and a swap)
230 enum PairConnectionType {
231 PairConnectionDirect,
236 void computeConnectedPairs(
237 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
238 DenseSet<ValuePair> &CandidatePairsSet,
239 std::vector<Value *> &PairableInsts,
240 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
241 DenseMap<VPPair, unsigned> &PairConnectionTypes);
243 void buildDepMap(BasicBlock &BB,
244 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
245 std::vector<Value *> &PairableInsts,
246 DenseSet<ValuePair> &PairableInstUsers);
248 void choosePairs(DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
249 DenseSet<ValuePair> &CandidatePairsSet,
250 DenseMap<ValuePair, int> &CandidatePairCostSavings,
251 std::vector<Value *> &PairableInsts,
252 DenseSet<ValuePair> &FixedOrderPairs,
253 DenseMap<VPPair, unsigned> &PairConnectionTypes,
254 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
255 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps,
256 DenseSet<ValuePair> &PairableInstUsers,
257 DenseMap<Value *, Value *>& ChosenPairs);
259 void fuseChosenPairs(BasicBlock &BB,
260 std::vector<Value *> &PairableInsts,
261 DenseMap<Value *, Value *>& ChosenPairs,
262 DenseSet<ValuePair> &FixedOrderPairs,
263 DenseMap<VPPair, unsigned> &PairConnectionTypes,
264 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
265 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps);
268 bool isInstVectorizable(Instruction *I, bool &IsSimpleLoadStore);
270 bool areInstsCompatible(Instruction *I, Instruction *J,
271 bool IsSimpleLoadStore, bool NonPow2Len,
272 int &CostSavings, int &FixedOrder);
274 bool trackUsesOfI(DenseSet<Value *> &Users,
275 AliasSetTracker &WriteSet, Instruction *I,
276 Instruction *J, bool UpdateUsers = true,
277 DenseSet<ValuePair> *LoadMoveSetPairs = 0);
279 void computePairsConnectedTo(
280 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
281 DenseSet<ValuePair> &CandidatePairsSet,
282 std::vector<Value *> &PairableInsts,
283 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
284 DenseMap<VPPair, unsigned> &PairConnectionTypes,
287 bool pairsConflict(ValuePair P, ValuePair Q,
288 DenseSet<ValuePair> &PairableInstUsers,
289 DenseMap<ValuePair, std::vector<ValuePair> >
290 *PairableInstUserMap = 0,
291 DenseSet<VPPair> *PairableInstUserPairSet = 0);
293 bool pairWillFormCycle(ValuePair P,
294 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUsers,
295 DenseSet<ValuePair> &CurrentPairs);
298 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
299 std::vector<Value *> &PairableInsts,
300 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
301 DenseSet<ValuePair> &PairableInstUsers,
302 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap,
303 DenseSet<VPPair> &PairableInstUserPairSet,
304 DenseMap<Value *, Value *> &ChosenPairs,
305 DenseMap<ValuePair, size_t> &Tree,
306 DenseSet<ValuePair> &PrunedTree, ValuePair J,
309 void buildInitialTreeFor(
310 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
311 DenseSet<ValuePair> &CandidatePairsSet,
312 std::vector<Value *> &PairableInsts,
313 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
314 DenseSet<ValuePair> &PairableInstUsers,
315 DenseMap<Value *, Value *> &ChosenPairs,
316 DenseMap<ValuePair, size_t> &Tree, ValuePair J);
318 void findBestTreeFor(
319 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
320 DenseSet<ValuePair> &CandidatePairsSet,
321 DenseMap<ValuePair, int> &CandidatePairCostSavings,
322 std::vector<Value *> &PairableInsts,
323 DenseSet<ValuePair> &FixedOrderPairs,
324 DenseMap<VPPair, unsigned> &PairConnectionTypes,
325 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
326 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps,
327 DenseSet<ValuePair> &PairableInstUsers,
328 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap,
329 DenseSet<VPPair> &PairableInstUserPairSet,
330 DenseMap<Value *, Value *> &ChosenPairs,
331 DenseSet<ValuePair> &BestTree, size_t &BestMaxDepth,
332 int &BestEffSize, Value *II, std::vector<Value *>&JJ,
335 Value *getReplacementPointerInput(LLVMContext& Context, Instruction *I,
336 Instruction *J, unsigned o);
338 void fillNewShuffleMask(LLVMContext& Context, Instruction *J,
339 unsigned MaskOffset, unsigned NumInElem,
340 unsigned NumInElem1, unsigned IdxOffset,
341 std::vector<Constant*> &Mask);
343 Value *getReplacementShuffleMask(LLVMContext& Context, Instruction *I,
346 bool expandIEChain(LLVMContext& Context, Instruction *I, Instruction *J,
347 unsigned o, Value *&LOp, unsigned numElemL,
348 Type *ArgTypeL, Type *ArgTypeR, bool IBeforeJ,
349 unsigned IdxOff = 0);
351 Value *getReplacementInput(LLVMContext& Context, Instruction *I,
352 Instruction *J, unsigned o, bool IBeforeJ);
354 void getReplacementInputsForPair(LLVMContext& Context, Instruction *I,
355 Instruction *J, SmallVector<Value *, 3> &ReplacedOperands,
358 void replaceOutputsOfPair(LLVMContext& Context, Instruction *I,
359 Instruction *J, Instruction *K,
360 Instruction *&InsertionPt, Instruction *&K1,
363 void collectPairLoadMoveSet(BasicBlock &BB,
364 DenseMap<Value *, Value *> &ChosenPairs,
365 DenseMap<Value *, std::vector<Value *> > &LoadMoveSet,
366 DenseSet<ValuePair> &LoadMoveSetPairs,
369 void collectLoadMoveSet(BasicBlock &BB,
370 std::vector<Value *> &PairableInsts,
371 DenseMap<Value *, Value *> &ChosenPairs,
372 DenseMap<Value *, std::vector<Value *> > &LoadMoveSet,
373 DenseSet<ValuePair> &LoadMoveSetPairs);
375 bool canMoveUsesOfIAfterJ(BasicBlock &BB,
376 DenseSet<ValuePair> &LoadMoveSetPairs,
377 Instruction *I, Instruction *J);
379 void moveUsesOfIAfterJ(BasicBlock &BB,
380 DenseSet<ValuePair> &LoadMoveSetPairs,
381 Instruction *&InsertionPt,
382 Instruction *I, Instruction *J);
384 void combineMetadata(Instruction *K, const Instruction *J);
386 bool vectorizeBB(BasicBlock &BB) {
387 if (!DT->isReachableFromEntry(&BB)) {
388 DEBUG(dbgs() << "BBV: skipping unreachable " << BB.getName() <<
389 " in " << BB.getParent()->getName() << "\n");
393 DEBUG(if (TTI) dbgs() << "BBV: using target information\n");
395 bool changed = false;
396 // Iterate a sufficient number of times to merge types of size 1 bit,
397 // then 2 bits, then 4, etc. up to half of the target vector width of the
398 // target vector register.
401 (TTI || v <= Config.VectorBits) &&
402 (!Config.MaxIter || n <= Config.MaxIter);
404 DEBUG(dbgs() << "BBV: fusing loop #" << n <<
405 " for " << BB.getName() << " in " <<
406 BB.getParent()->getName() << "...\n");
407 if (vectorizePairs(BB))
413 if (changed && !Pow2LenOnly) {
415 for (; !Config.MaxIter || n <= Config.MaxIter; ++n) {
416 DEBUG(dbgs() << "BBV: fusing for non-2^n-length vectors loop #: " <<
417 n << " for " << BB.getName() << " in " <<
418 BB.getParent()->getName() << "...\n");
419 if (!vectorizePairs(BB, true)) break;
423 DEBUG(dbgs() << "BBV: done!\n");
427 virtual bool runOnBasicBlock(BasicBlock &BB) {
428 AA = &getAnalysis<AliasAnalysis>();
429 DT = &getAnalysis<DominatorTree>();
430 SE = &getAnalysis<ScalarEvolution>();
431 TD = getAnalysisIfAvailable<DataLayout>();
432 TTI = IgnoreTargetInfo ? 0 : &getAnalysis<TargetTransformInfo>();
434 return vectorizeBB(BB);
437 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
438 BasicBlockPass::getAnalysisUsage(AU);
439 AU.addRequired<AliasAnalysis>();
440 AU.addRequired<DominatorTree>();
441 AU.addRequired<ScalarEvolution>();
442 AU.addRequired<TargetTransformInfo>();
443 AU.addPreserved<AliasAnalysis>();
444 AU.addPreserved<DominatorTree>();
445 AU.addPreserved<ScalarEvolution>();
446 AU.setPreservesCFG();
449 static inline VectorType *getVecTypeForPair(Type *ElemTy, Type *Elem2Ty) {
450 assert(ElemTy->getScalarType() == Elem2Ty->getScalarType() &&
451 "Cannot form vector from incompatible scalar types");
452 Type *STy = ElemTy->getScalarType();
455 if (VectorType *VTy = dyn_cast<VectorType>(ElemTy)) {
456 numElem = VTy->getNumElements();
461 if (VectorType *VTy = dyn_cast<VectorType>(Elem2Ty)) {
462 numElem += VTy->getNumElements();
467 return VectorType::get(STy, numElem);
470 static inline void getInstructionTypes(Instruction *I,
471 Type *&T1, Type *&T2) {
472 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
473 // For stores, it is the value type, not the pointer type that matters
474 // because the value is what will come from a vector register.
476 Value *IVal = SI->getValueOperand();
477 T1 = IVal->getType();
482 if (CastInst *CI = dyn_cast<CastInst>(I))
487 if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
488 T2 = SI->getCondition()->getType();
489 } else if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(I)) {
490 T2 = SI->getOperand(0)->getType();
491 } else if (CmpInst *CI = dyn_cast<CmpInst>(I)) {
492 T2 = CI->getOperand(0)->getType();
496 // Returns the weight associated with the provided value. A chain of
497 // candidate pairs has a length given by the sum of the weights of its
498 // members (one weight per pair; the weight of each member of the pair
499 // is assumed to be the same). This length is then compared to the
500 // chain-length threshold to determine if a given chain is significant
501 // enough to be vectorized. The length is also used in comparing
502 // candidate chains where longer chains are considered to be better.
503 // Note: when this function returns 0, the resulting instructions are
504 // not actually fused.
505 inline size_t getDepthFactor(Value *V) {
506 // InsertElement and ExtractElement have a depth factor of zero. This is
507 // for two reasons: First, they cannot be usefully fused. Second, because
508 // the pass generates a lot of these, they can confuse the simple metric
509 // used to compare the trees in the next iteration. Thus, giving them a
510 // weight of zero allows the pass to essentially ignore them in
511 // subsequent iterations when looking for vectorization opportunities
512 // while still tracking dependency chains that flow through those
514 if (isa<InsertElementInst>(V) || isa<ExtractElementInst>(V))
517 // Give a load or store half of the required depth so that load/store
518 // pairs will vectorize.
519 if (!Config.NoMemOpBoost && (isa<LoadInst>(V) || isa<StoreInst>(V)))
520 return Config.ReqChainDepth/2;
525 // Returns the cost of the provided instruction using TTI.
526 // This does not handle loads and stores.
527 unsigned getInstrCost(unsigned Opcode, Type *T1, Type *T2) {
530 case Instruction::GetElementPtr:
531 // We mark this instruction as zero-cost because scalar GEPs are usually
532 // lowered to the intruction addressing mode. At the moment we don't
533 // generate vector GEPs.
535 case Instruction::Br:
536 return TTI->getCFInstrCost(Opcode);
537 case Instruction::PHI:
539 case Instruction::Add:
540 case Instruction::FAdd:
541 case Instruction::Sub:
542 case Instruction::FSub:
543 case Instruction::Mul:
544 case Instruction::FMul:
545 case Instruction::UDiv:
546 case Instruction::SDiv:
547 case Instruction::FDiv:
548 case Instruction::URem:
549 case Instruction::SRem:
550 case Instruction::FRem:
551 case Instruction::Shl:
552 case Instruction::LShr:
553 case Instruction::AShr:
554 case Instruction::And:
555 case Instruction::Or:
556 case Instruction::Xor:
557 return TTI->getArithmeticInstrCost(Opcode, T1);
558 case Instruction::Select:
559 case Instruction::ICmp:
560 case Instruction::FCmp:
561 return TTI->getCmpSelInstrCost(Opcode, T1, T2);
562 case Instruction::ZExt:
563 case Instruction::SExt:
564 case Instruction::FPToUI:
565 case Instruction::FPToSI:
566 case Instruction::FPExt:
567 case Instruction::PtrToInt:
568 case Instruction::IntToPtr:
569 case Instruction::SIToFP:
570 case Instruction::UIToFP:
571 case Instruction::Trunc:
572 case Instruction::FPTrunc:
573 case Instruction::BitCast:
574 case Instruction::ShuffleVector:
575 return TTI->getCastInstrCost(Opcode, T1, T2);
581 // This determines the relative offset of two loads or stores, returning
582 // true if the offset could be determined to be some constant value.
583 // For example, if OffsetInElmts == 1, then J accesses the memory directly
584 // after I; if OffsetInElmts == -1 then I accesses the memory
586 bool getPairPtrInfo(Instruction *I, Instruction *J,
587 Value *&IPtr, Value *&JPtr, unsigned &IAlignment, unsigned &JAlignment,
588 unsigned &IAddressSpace, unsigned &JAddressSpace,
589 int64_t &OffsetInElmts, bool ComputeOffset = true) {
591 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
592 LoadInst *LJ = cast<LoadInst>(J);
593 IPtr = LI->getPointerOperand();
594 JPtr = LJ->getPointerOperand();
595 IAlignment = LI->getAlignment();
596 JAlignment = LJ->getAlignment();
597 IAddressSpace = LI->getPointerAddressSpace();
598 JAddressSpace = LJ->getPointerAddressSpace();
600 StoreInst *SI = cast<StoreInst>(I), *SJ = cast<StoreInst>(J);
601 IPtr = SI->getPointerOperand();
602 JPtr = SJ->getPointerOperand();
603 IAlignment = SI->getAlignment();
604 JAlignment = SJ->getAlignment();
605 IAddressSpace = SI->getPointerAddressSpace();
606 JAddressSpace = SJ->getPointerAddressSpace();
612 const SCEV *IPtrSCEV = SE->getSCEV(IPtr);
613 const SCEV *JPtrSCEV = SE->getSCEV(JPtr);
615 // If this is a trivial offset, then we'll get something like
616 // 1*sizeof(type). With target data, which we need anyway, this will get
617 // constant folded into a number.
618 const SCEV *OffsetSCEV = SE->getMinusSCEV(JPtrSCEV, IPtrSCEV);
619 if (const SCEVConstant *ConstOffSCEV =
620 dyn_cast<SCEVConstant>(OffsetSCEV)) {
621 ConstantInt *IntOff = ConstOffSCEV->getValue();
622 int64_t Offset = IntOff->getSExtValue();
624 Type *VTy = cast<PointerType>(IPtr->getType())->getElementType();
625 int64_t VTyTSS = (int64_t) TD->getTypeStoreSize(VTy);
627 Type *VTy2 = cast<PointerType>(JPtr->getType())->getElementType();
628 if (VTy != VTy2 && Offset < 0) {
629 int64_t VTy2TSS = (int64_t) TD->getTypeStoreSize(VTy2);
630 OffsetInElmts = Offset/VTy2TSS;
631 return (abs64(Offset) % VTy2TSS) == 0;
634 OffsetInElmts = Offset/VTyTSS;
635 return (abs64(Offset) % VTyTSS) == 0;
641 // Returns true if the provided CallInst represents an intrinsic that can
643 bool isVectorizableIntrinsic(CallInst* I) {
644 Function *F = I->getCalledFunction();
645 if (!F) return false;
647 Intrinsic::ID IID = (Intrinsic::ID) F->getIntrinsicID();
648 if (!IID) return false;
653 case Intrinsic::sqrt:
654 case Intrinsic::powi:
658 case Intrinsic::log2:
659 case Intrinsic::log10:
661 case Intrinsic::exp2:
663 return Config.VectorizeMath;
665 case Intrinsic::fmuladd:
666 return Config.VectorizeFMA;
670 bool isPureIEChain(InsertElementInst *IE) {
671 InsertElementInst *IENext = IE;
673 if (!isa<UndefValue>(IENext->getOperand(0)) &&
674 !isa<InsertElementInst>(IENext->getOperand(0))) {
678 dyn_cast<InsertElementInst>(IENext->getOperand(0))));
684 // This function implements one vectorization iteration on the provided
685 // basic block. It returns true if the block is changed.
686 bool BBVectorize::vectorizePairs(BasicBlock &BB, bool NonPow2Len) {
688 BasicBlock::iterator Start = BB.getFirstInsertionPt();
690 std::vector<Value *> AllPairableInsts;
691 DenseMap<Value *, Value *> AllChosenPairs;
692 DenseSet<ValuePair> AllFixedOrderPairs;
693 DenseMap<VPPair, unsigned> AllPairConnectionTypes;
694 DenseMap<ValuePair, std::vector<ValuePair> > AllConnectedPairs,
695 AllConnectedPairDeps;
698 std::vector<Value *> PairableInsts;
699 DenseMap<Value *, std::vector<Value *> > CandidatePairs;
700 DenseSet<ValuePair> FixedOrderPairs;
701 DenseMap<ValuePair, int> CandidatePairCostSavings;
702 ShouldContinue = getCandidatePairs(BB, Start, CandidatePairs,
704 CandidatePairCostSavings,
705 PairableInsts, NonPow2Len);
706 if (PairableInsts.empty()) continue;
708 // Build the candidate pair set for faster lookups.
709 DenseSet<ValuePair> CandidatePairsSet;
710 for (DenseMap<Value *, std::vector<Value *> >::iterator I =
711 CandidatePairs.begin(), E = CandidatePairs.end(); I != E; ++I)
712 for (std::vector<Value *>::iterator J = I->second.begin(),
713 JE = I->second.end(); J != JE; ++J)
714 CandidatePairsSet.insert(ValuePair(I->first, *J));
716 // Now we have a map of all of the pairable instructions and we need to
717 // select the best possible pairing. A good pairing is one such that the
718 // users of the pair are also paired. This defines a (directed) forest
719 // over the pairs such that two pairs are connected iff the second pair
722 // Note that it only matters that both members of the second pair use some
723 // element of the first pair (to allow for splatting).
725 DenseMap<ValuePair, std::vector<ValuePair> > ConnectedPairs,
727 DenseMap<VPPair, unsigned> PairConnectionTypes;
728 computeConnectedPairs(CandidatePairs, CandidatePairsSet,
729 PairableInsts, ConnectedPairs, PairConnectionTypes);
730 if (ConnectedPairs.empty()) continue;
732 for (DenseMap<ValuePair, std::vector<ValuePair> >::iterator
733 I = ConnectedPairs.begin(), IE = ConnectedPairs.end();
735 for (std::vector<ValuePair>::iterator J = I->second.begin(),
736 JE = I->second.end(); J != JE; ++J)
737 ConnectedPairDeps[*J].push_back(I->first);
739 // Build the pairable-instruction dependency map
740 DenseSet<ValuePair> PairableInstUsers;
741 buildDepMap(BB, CandidatePairs, PairableInsts, PairableInstUsers);
743 // There is now a graph of the connected pairs. For each variable, pick
744 // the pairing with the largest tree meeting the depth requirement on at
745 // least one branch. Then select all pairings that are part of that tree
746 // and remove them from the list of available pairings and pairable
749 DenseMap<Value *, Value *> ChosenPairs;
750 choosePairs(CandidatePairs, CandidatePairsSet,
751 CandidatePairCostSavings,
752 PairableInsts, FixedOrderPairs, PairConnectionTypes,
753 ConnectedPairs, ConnectedPairDeps,
754 PairableInstUsers, ChosenPairs);
756 if (ChosenPairs.empty()) continue;
757 AllPairableInsts.insert(AllPairableInsts.end(), PairableInsts.begin(),
758 PairableInsts.end());
759 AllChosenPairs.insert(ChosenPairs.begin(), ChosenPairs.end());
761 // Only for the chosen pairs, propagate information on fixed-order pairs,
762 // pair connections, and their types to the data structures used by the
763 // pair fusion procedures.
764 for (DenseMap<Value *, Value *>::iterator I = ChosenPairs.begin(),
765 IE = ChosenPairs.end(); I != IE; ++I) {
766 if (FixedOrderPairs.count(*I))
767 AllFixedOrderPairs.insert(*I);
768 else if (FixedOrderPairs.count(ValuePair(I->second, I->first)))
769 AllFixedOrderPairs.insert(ValuePair(I->second, I->first));
771 for (DenseMap<Value *, Value *>::iterator J = ChosenPairs.begin();
773 DenseMap<VPPair, unsigned>::iterator K =
774 PairConnectionTypes.find(VPPair(*I, *J));
775 if (K != PairConnectionTypes.end()) {
776 AllPairConnectionTypes.insert(*K);
778 K = PairConnectionTypes.find(VPPair(*J, *I));
779 if (K != PairConnectionTypes.end())
780 AllPairConnectionTypes.insert(*K);
785 for (DenseMap<ValuePair, std::vector<ValuePair> >::iterator
786 I = ConnectedPairs.begin(), IE = ConnectedPairs.end();
788 for (std::vector<ValuePair>::iterator J = I->second.begin(),
789 JE = I->second.end(); J != JE; ++J)
790 if (AllPairConnectionTypes.count(VPPair(I->first, *J))) {
791 AllConnectedPairs[I->first].push_back(*J);
792 AllConnectedPairDeps[*J].push_back(I->first);
794 } while (ShouldContinue);
796 if (AllChosenPairs.empty()) return false;
797 NumFusedOps += AllChosenPairs.size();
799 // A set of pairs has now been selected. It is now necessary to replace the
800 // paired instructions with vector instructions. For this procedure each
801 // operand must be replaced with a vector operand. This vector is formed
802 // by using build_vector on the old operands. The replaced values are then
803 // replaced with a vector_extract on the result. Subsequent optimization
804 // passes should coalesce the build/extract combinations.
806 fuseChosenPairs(BB, AllPairableInsts, AllChosenPairs, AllFixedOrderPairs,
807 AllPairConnectionTypes,
808 AllConnectedPairs, AllConnectedPairDeps);
810 // It is important to cleanup here so that future iterations of this
811 // function have less work to do.
812 (void) SimplifyInstructionsInBlock(&BB, TD, AA->getTargetLibraryInfo());
816 // This function returns true if the provided instruction is capable of being
817 // fused into a vector instruction. This determination is based only on the
818 // type and other attributes of the instruction.
819 bool BBVectorize::isInstVectorizable(Instruction *I,
820 bool &IsSimpleLoadStore) {
821 IsSimpleLoadStore = false;
823 if (CallInst *C = dyn_cast<CallInst>(I)) {
824 if (!isVectorizableIntrinsic(C))
826 } else if (LoadInst *L = dyn_cast<LoadInst>(I)) {
827 // Vectorize simple loads if possbile:
828 IsSimpleLoadStore = L->isSimple();
829 if (!IsSimpleLoadStore || !Config.VectorizeMemOps)
831 } else if (StoreInst *S = dyn_cast<StoreInst>(I)) {
832 // Vectorize simple stores if possbile:
833 IsSimpleLoadStore = S->isSimple();
834 if (!IsSimpleLoadStore || !Config.VectorizeMemOps)
836 } else if (CastInst *C = dyn_cast<CastInst>(I)) {
837 // We can vectorize casts, but not casts of pointer types, etc.
838 if (!Config.VectorizeCasts)
841 Type *SrcTy = C->getSrcTy();
842 if (!SrcTy->isSingleValueType())
845 Type *DestTy = C->getDestTy();
846 if (!DestTy->isSingleValueType())
848 } else if (isa<SelectInst>(I)) {
849 if (!Config.VectorizeSelect)
851 } else if (isa<CmpInst>(I)) {
852 if (!Config.VectorizeCmp)
854 } else if (GetElementPtrInst *G = dyn_cast<GetElementPtrInst>(I)) {
855 if (!Config.VectorizeGEP)
858 // Currently, vector GEPs exist only with one index.
859 if (G->getNumIndices() != 1)
861 } else if (!(I->isBinaryOp() || isa<ShuffleVectorInst>(I) ||
862 isa<ExtractElementInst>(I) || isa<InsertElementInst>(I))) {
866 // We can't vectorize memory operations without target data
867 if (TD == 0 && IsSimpleLoadStore)
871 getInstructionTypes(I, T1, T2);
873 // Not every type can be vectorized...
874 if (!(VectorType::isValidElementType(T1) || T1->isVectorTy()) ||
875 !(VectorType::isValidElementType(T2) || T2->isVectorTy()))
878 if (T1->getScalarSizeInBits() == 1) {
879 if (!Config.VectorizeBools)
882 if (!Config.VectorizeInts && T1->isIntOrIntVectorTy())
886 if (T2->getScalarSizeInBits() == 1) {
887 if (!Config.VectorizeBools)
890 if (!Config.VectorizeInts && T2->isIntOrIntVectorTy())
894 if (!Config.VectorizeFloats
895 && (T1->isFPOrFPVectorTy() || T2->isFPOrFPVectorTy()))
898 // Don't vectorize target-specific types.
899 if (T1->isX86_FP80Ty() || T1->isPPC_FP128Ty() || T1->isX86_MMXTy())
901 if (T2->isX86_FP80Ty() || T2->isPPC_FP128Ty() || T2->isX86_MMXTy())
904 if ((!Config.VectorizePointers || TD == 0) &&
905 (T1->getScalarType()->isPointerTy() ||
906 T2->getScalarType()->isPointerTy()))
909 if (!TTI && (T1->getPrimitiveSizeInBits() >= Config.VectorBits ||
910 T2->getPrimitiveSizeInBits() >= Config.VectorBits))
916 // This function returns true if the two provided instructions are compatible
917 // (meaning that they can be fused into a vector instruction). This assumes
918 // that I has already been determined to be vectorizable and that J is not
919 // in the use tree of I.
920 bool BBVectorize::areInstsCompatible(Instruction *I, Instruction *J,
921 bool IsSimpleLoadStore, bool NonPow2Len,
922 int &CostSavings, int &FixedOrder) {
923 DEBUG(if (DebugInstructionExamination) dbgs() << "BBV: looking at " << *I <<
924 " <-> " << *J << "\n");
929 // Loads and stores can be merged if they have different alignments,
930 // but are otherwise the same.
931 if (!J->isSameOperationAs(I, Instruction::CompareIgnoringAlignment |
932 (NonPow2Len ? Instruction::CompareUsingScalarTypes : 0)))
935 Type *IT1, *IT2, *JT1, *JT2;
936 getInstructionTypes(I, IT1, IT2);
937 getInstructionTypes(J, JT1, JT2);
938 unsigned MaxTypeBits = std::max(
939 IT1->getPrimitiveSizeInBits() + JT1->getPrimitiveSizeInBits(),
940 IT2->getPrimitiveSizeInBits() + JT2->getPrimitiveSizeInBits());
941 if (!TTI && MaxTypeBits > Config.VectorBits)
944 // FIXME: handle addsub-type operations!
946 if (IsSimpleLoadStore) {
948 unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace;
949 int64_t OffsetInElmts = 0;
950 if (getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment,
951 IAddressSpace, JAddressSpace,
952 OffsetInElmts) && abs64(OffsetInElmts) == 1) {
953 FixedOrder = (int) OffsetInElmts;
954 unsigned BottomAlignment = IAlignment;
955 if (OffsetInElmts < 0) BottomAlignment = JAlignment;
957 Type *aTypeI = isa<StoreInst>(I) ?
958 cast<StoreInst>(I)->getValueOperand()->getType() : I->getType();
959 Type *aTypeJ = isa<StoreInst>(J) ?
960 cast<StoreInst>(J)->getValueOperand()->getType() : J->getType();
961 Type *VType = getVecTypeForPair(aTypeI, aTypeJ);
963 if (Config.AlignedOnly) {
964 // An aligned load or store is possible only if the instruction
965 // with the lower offset has an alignment suitable for the
968 unsigned VecAlignment = TD->getPrefTypeAlignment(VType);
969 if (BottomAlignment < VecAlignment)
974 unsigned ICost = TTI->getMemoryOpCost(I->getOpcode(), aTypeI,
975 IAlignment, IAddressSpace);
976 unsigned JCost = TTI->getMemoryOpCost(J->getOpcode(), aTypeJ,
977 JAlignment, JAddressSpace);
978 unsigned VCost = TTI->getMemoryOpCost(I->getOpcode(), VType,
982 ICost += TTI->getAddressComputationCost(aTypeI);
983 JCost += TTI->getAddressComputationCost(aTypeJ);
984 VCost += TTI->getAddressComputationCost(VType);
986 if (VCost > ICost + JCost)
989 // We don't want to fuse to a type that will be split, even
990 // if the two input types will also be split and there is no other
992 unsigned VParts = TTI->getNumberOfParts(VType);
995 else if (!VParts && VCost == ICost + JCost)
998 CostSavings = ICost + JCost - VCost;
1004 unsigned ICost = getInstrCost(I->getOpcode(), IT1, IT2);
1005 unsigned JCost = getInstrCost(J->getOpcode(), JT1, JT2);
1006 Type *VT1 = getVecTypeForPair(IT1, JT1),
1007 *VT2 = getVecTypeForPair(IT2, JT2);
1009 // Note that this procedure is incorrect for insert and extract element
1010 // instructions (because combining these often results in a shuffle),
1011 // but this cost is ignored (because insert and extract element
1012 // instructions are assigned a zero depth factor and are not really
1013 // fused in general).
1014 unsigned VCost = getInstrCost(I->getOpcode(), VT1, VT2);
1016 if (VCost > ICost + JCost)
1019 // We don't want to fuse to a type that will be split, even
1020 // if the two input types will also be split and there is no other
1022 unsigned VParts1 = TTI->getNumberOfParts(VT1),
1023 VParts2 = TTI->getNumberOfParts(VT2);
1024 if (VParts1 > 1 || VParts2 > 1)
1026 else if ((!VParts1 || !VParts2) && VCost == ICost + JCost)
1029 CostSavings = ICost + JCost - VCost;
1032 // The powi intrinsic is special because only the first argument is
1033 // vectorized, the second arguments must be equal.
1034 CallInst *CI = dyn_cast<CallInst>(I);
1036 if (CI && (FI = CI->getCalledFunction())) {
1037 Intrinsic::ID IID = (Intrinsic::ID) FI->getIntrinsicID();
1038 if (IID == Intrinsic::powi) {
1039 Value *A1I = CI->getArgOperand(1),
1040 *A1J = cast<CallInst>(J)->getArgOperand(1);
1041 const SCEV *A1ISCEV = SE->getSCEV(A1I),
1042 *A1JSCEV = SE->getSCEV(A1J);
1043 return (A1ISCEV == A1JSCEV);
1047 SmallVector<Type*, 4> Tys;
1048 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
1049 Tys.push_back(CI->getArgOperand(i)->getType());
1050 unsigned ICost = TTI->getIntrinsicInstrCost(IID, IT1, Tys);
1053 CallInst *CJ = cast<CallInst>(J);
1054 for (unsigned i = 0, ie = CJ->getNumArgOperands(); i != ie; ++i)
1055 Tys.push_back(CJ->getArgOperand(i)->getType());
1056 unsigned JCost = TTI->getIntrinsicInstrCost(IID, JT1, Tys);
1059 assert(CI->getNumArgOperands() == CJ->getNumArgOperands() &&
1060 "Intrinsic argument counts differ");
1061 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
1062 if (IID == Intrinsic::powi && i == 1)
1063 Tys.push_back(CI->getArgOperand(i)->getType());
1065 Tys.push_back(getVecTypeForPair(CI->getArgOperand(i)->getType(),
1066 CJ->getArgOperand(i)->getType()));
1069 Type *RetTy = getVecTypeForPair(IT1, JT1);
1070 unsigned VCost = TTI->getIntrinsicInstrCost(IID, RetTy, Tys);
1072 if (VCost > ICost + JCost)
1075 // We don't want to fuse to a type that will be split, even
1076 // if the two input types will also be split and there is no other
1078 unsigned RetParts = TTI->getNumberOfParts(RetTy);
1081 else if (!RetParts && VCost == ICost + JCost)
1084 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
1085 if (!Tys[i]->isVectorTy())
1088 unsigned NumParts = TTI->getNumberOfParts(Tys[i]);
1091 else if (!NumParts && VCost == ICost + JCost)
1095 CostSavings = ICost + JCost - VCost;
1102 // Figure out whether or not J uses I and update the users and write-set
1103 // structures associated with I. Specifically, Users represents the set of
1104 // instructions that depend on I. WriteSet represents the set
1105 // of memory locations that are dependent on I. If UpdateUsers is true,
1106 // and J uses I, then Users is updated to contain J and WriteSet is updated
1107 // to contain any memory locations to which J writes. The function returns
1108 // true if J uses I. By default, alias analysis is used to determine
1109 // whether J reads from memory that overlaps with a location in WriteSet.
1110 // If LoadMoveSet is not null, then it is a previously-computed map
1111 // where the key is the memory-based user instruction and the value is
1112 // the instruction to be compared with I. So, if LoadMoveSet is provided,
1113 // then the alias analysis is not used. This is necessary because this
1114 // function is called during the process of moving instructions during
1115 // vectorization and the results of the alias analysis are not stable during
1117 bool BBVectorize::trackUsesOfI(DenseSet<Value *> &Users,
1118 AliasSetTracker &WriteSet, Instruction *I,
1119 Instruction *J, bool UpdateUsers,
1120 DenseSet<ValuePair> *LoadMoveSetPairs) {
1123 // This instruction may already be marked as a user due, for example, to
1124 // being a member of a selected pair.
1129 for (User::op_iterator JU = J->op_begin(), JE = J->op_end();
1132 if (I == V || Users.count(V)) {
1137 if (!UsesI && J->mayReadFromMemory()) {
1138 if (LoadMoveSetPairs) {
1139 UsesI = LoadMoveSetPairs->count(ValuePair(J, I));
1141 for (AliasSetTracker::iterator W = WriteSet.begin(),
1142 WE = WriteSet.end(); W != WE; ++W) {
1143 if (W->aliasesUnknownInst(J, *AA)) {
1151 if (UsesI && UpdateUsers) {
1152 if (J->mayWriteToMemory()) WriteSet.add(J);
1159 // This function iterates over all instruction pairs in the provided
1160 // basic block and collects all candidate pairs for vectorization.
1161 bool BBVectorize::getCandidatePairs(BasicBlock &BB,
1162 BasicBlock::iterator &Start,
1163 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1164 DenseSet<ValuePair> &FixedOrderPairs,
1165 DenseMap<ValuePair, int> &CandidatePairCostSavings,
1166 std::vector<Value *> &PairableInsts, bool NonPow2Len) {
1167 BasicBlock::iterator E = BB.end();
1168 if (Start == E) return false;
1170 bool ShouldContinue = false, IAfterStart = false;
1171 for (BasicBlock::iterator I = Start++; I != E; ++I) {
1172 if (I == Start) IAfterStart = true;
1174 bool IsSimpleLoadStore;
1175 if (!isInstVectorizable(I, IsSimpleLoadStore)) continue;
1177 // Look for an instruction with which to pair instruction *I...
1178 DenseSet<Value *> Users;
1179 AliasSetTracker WriteSet(*AA);
1180 bool JAfterStart = IAfterStart;
1181 BasicBlock::iterator J = llvm::next(I);
1182 for (unsigned ss = 0; J != E && ss <= Config.SearchLimit; ++J, ++ss) {
1183 if (J == Start) JAfterStart = true;
1185 // Determine if J uses I, if so, exit the loop.
1186 bool UsesI = trackUsesOfI(Users, WriteSet, I, J, !Config.FastDep);
1187 if (Config.FastDep) {
1188 // Note: For this heuristic to be effective, independent operations
1189 // must tend to be intermixed. This is likely to be true from some
1190 // kinds of grouped loop unrolling (but not the generic LLVM pass),
1191 // but otherwise may require some kind of reordering pass.
1193 // When using fast dependency analysis,
1194 // stop searching after first use:
1197 if (UsesI) continue;
1200 // J does not use I, and comes before the first use of I, so it can be
1201 // merged with I if the instructions are compatible.
1202 int CostSavings, FixedOrder;
1203 if (!areInstsCompatible(I, J, IsSimpleLoadStore, NonPow2Len,
1204 CostSavings, FixedOrder)) continue;
1206 // J is a candidate for merging with I.
1207 if (!PairableInsts.size() ||
1208 PairableInsts[PairableInsts.size()-1] != I) {
1209 PairableInsts.push_back(I);
1212 CandidatePairs[I].push_back(J);
1214 CandidatePairCostSavings.insert(ValuePairWithCost(ValuePair(I, J),
1217 if (FixedOrder == 1)
1218 FixedOrderPairs.insert(ValuePair(I, J));
1219 else if (FixedOrder == -1)
1220 FixedOrderPairs.insert(ValuePair(J, I));
1222 // The next call to this function must start after the last instruction
1223 // selected during this invocation.
1225 Start = llvm::next(J);
1226 IAfterStart = JAfterStart = false;
1229 DEBUG(if (DebugCandidateSelection) dbgs() << "BBV: candidate pair "
1230 << *I << " <-> " << *J << " (cost savings: " <<
1231 CostSavings << ")\n");
1233 // If we have already found too many pairs, break here and this function
1234 // will be called again starting after the last instruction selected
1235 // during this invocation.
1236 if (PairableInsts.size() >= Config.MaxInsts) {
1237 ShouldContinue = true;
1246 DEBUG(dbgs() << "BBV: found " << PairableInsts.size()
1247 << " instructions with candidate pairs\n");
1249 return ShouldContinue;
1252 // Finds candidate pairs connected to the pair P = <PI, PJ>. This means that
1253 // it looks for pairs such that both members have an input which is an
1254 // output of PI or PJ.
1255 void BBVectorize::computePairsConnectedTo(
1256 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1257 DenseSet<ValuePair> &CandidatePairsSet,
1258 std::vector<Value *> &PairableInsts,
1259 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
1260 DenseMap<VPPair, unsigned> &PairConnectionTypes,
1264 // For each possible pairing for this variable, look at the uses of
1265 // the first value...
1266 for (Value::use_iterator I = P.first->use_begin(),
1267 E = P.first->use_end(); I != E; ++I) {
1268 if (isa<LoadInst>(*I)) {
1269 // A pair cannot be connected to a load because the load only takes one
1270 // operand (the address) and it is a scalar even after vectorization.
1272 } else if ((SI = dyn_cast<StoreInst>(*I)) &&
1273 P.first == SI->getPointerOperand()) {
1274 // Similarly, a pair cannot be connected to a store through its
1279 // For each use of the first variable, look for uses of the second
1281 for (Value::use_iterator J = P.second->use_begin(),
1282 E2 = P.second->use_end(); J != E2; ++J) {
1283 if ((SJ = dyn_cast<StoreInst>(*J)) &&
1284 P.second == SJ->getPointerOperand())
1288 if (CandidatePairsSet.count(ValuePair(*I, *J))) {
1289 VPPair VP(P, ValuePair(*I, *J));
1290 ConnectedPairs[VP.first].push_back(VP.second);
1291 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionDirect));
1295 if (CandidatePairsSet.count(ValuePair(*J, *I))) {
1296 VPPair VP(P, ValuePair(*J, *I));
1297 ConnectedPairs[VP.first].push_back(VP.second);
1298 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSwap));
1302 if (Config.SplatBreaksChain) continue;
1303 // Look for cases where just the first value in the pair is used by
1304 // both members of another pair (splatting).
1305 for (Value::use_iterator J = P.first->use_begin(); J != E; ++J) {
1306 if ((SJ = dyn_cast<StoreInst>(*J)) &&
1307 P.first == SJ->getPointerOperand())
1310 if (CandidatePairsSet.count(ValuePair(*I, *J))) {
1311 VPPair VP(P, ValuePair(*I, *J));
1312 ConnectedPairs[VP.first].push_back(VP.second);
1313 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSplat));
1318 if (Config.SplatBreaksChain) return;
1319 // Look for cases where just the second value in the pair is used by
1320 // both members of another pair (splatting).
1321 for (Value::use_iterator I = P.second->use_begin(),
1322 E = P.second->use_end(); I != E; ++I) {
1323 if (isa<LoadInst>(*I))
1325 else if ((SI = dyn_cast<StoreInst>(*I)) &&
1326 P.second == SI->getPointerOperand())
1329 for (Value::use_iterator J = P.second->use_begin(); J != E; ++J) {
1330 if ((SJ = dyn_cast<StoreInst>(*J)) &&
1331 P.second == SJ->getPointerOperand())
1334 if (CandidatePairsSet.count(ValuePair(*I, *J))) {
1335 VPPair VP(P, ValuePair(*I, *J));
1336 ConnectedPairs[VP.first].push_back(VP.second);
1337 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSplat));
1343 // This function figures out which pairs are connected. Two pairs are
1344 // connected if some output of the first pair forms an input to both members
1345 // of the second pair.
1346 void BBVectorize::computeConnectedPairs(
1347 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1348 DenseSet<ValuePair> &CandidatePairsSet,
1349 std::vector<Value *> &PairableInsts,
1350 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
1351 DenseMap<VPPair, unsigned> &PairConnectionTypes) {
1352 for (std::vector<Value *>::iterator PI = PairableInsts.begin(),
1353 PE = PairableInsts.end(); PI != PE; ++PI) {
1354 DenseMap<Value *, std::vector<Value *> >::iterator PP =
1355 CandidatePairs.find(*PI);
1356 if (PP == CandidatePairs.end())
1359 for (std::vector<Value *>::iterator P = PP->second.begin(),
1360 E = PP->second.end(); P != E; ++P)
1361 computePairsConnectedTo(CandidatePairs, CandidatePairsSet,
1362 PairableInsts, ConnectedPairs,
1363 PairConnectionTypes, ValuePair(*PI, *P));
1366 DEBUG(size_t TotalPairs = 0;
1367 for (DenseMap<ValuePair, std::vector<ValuePair> >::iterator I =
1368 ConnectedPairs.begin(), IE = ConnectedPairs.end(); I != IE; ++I)
1369 TotalPairs += I->second.size();
1370 dbgs() << "BBV: found " << TotalPairs
1371 << " pair connections.\n");
1374 // This function builds a set of use tuples such that <A, B> is in the set
1375 // if B is in the use tree of A. If B is in the use tree of A, then B
1376 // depends on the output of A.
1377 void BBVectorize::buildDepMap(
1379 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1380 std::vector<Value *> &PairableInsts,
1381 DenseSet<ValuePair> &PairableInstUsers) {
1382 DenseSet<Value *> IsInPair;
1383 for (DenseMap<Value *, std::vector<Value *> >::iterator C =
1384 CandidatePairs.begin(), E = CandidatePairs.end(); C != E; ++C) {
1385 IsInPair.insert(C->first);
1386 IsInPair.insert(C->second.begin(), C->second.end());
1389 // Iterate through the basic block, recording all users of each
1390 // pairable instruction.
1392 BasicBlock::iterator E = BB.end(), EL =
1393 BasicBlock::iterator(cast<Instruction>(PairableInsts.back()));
1394 for (BasicBlock::iterator I = BB.getFirstInsertionPt(); I != E; ++I) {
1395 if (IsInPair.find(I) == IsInPair.end()) continue;
1397 DenseSet<Value *> Users;
1398 AliasSetTracker WriteSet(*AA);
1399 for (BasicBlock::iterator J = llvm::next(I); J != E; ++J) {
1400 (void) trackUsesOfI(Users, WriteSet, I, J);
1406 for (DenseSet<Value *>::iterator U = Users.begin(), E = Users.end();
1408 if (IsInPair.find(*U) == IsInPair.end()) continue;
1409 PairableInstUsers.insert(ValuePair(I, *U));
1417 // Returns true if an input to pair P is an output of pair Q and also an
1418 // input of pair Q is an output of pair P. If this is the case, then these
1419 // two pairs cannot be simultaneously fused.
1420 bool BBVectorize::pairsConflict(ValuePair P, ValuePair Q,
1421 DenseSet<ValuePair> &PairableInstUsers,
1422 DenseMap<ValuePair, std::vector<ValuePair> > *PairableInstUserMap,
1423 DenseSet<VPPair> *PairableInstUserPairSet) {
1424 // Two pairs are in conflict if they are mutual Users of eachother.
1425 bool QUsesP = PairableInstUsers.count(ValuePair(P.first, Q.first)) ||
1426 PairableInstUsers.count(ValuePair(P.first, Q.second)) ||
1427 PairableInstUsers.count(ValuePair(P.second, Q.first)) ||
1428 PairableInstUsers.count(ValuePair(P.second, Q.second));
1429 bool PUsesQ = PairableInstUsers.count(ValuePair(Q.first, P.first)) ||
1430 PairableInstUsers.count(ValuePair(Q.first, P.second)) ||
1431 PairableInstUsers.count(ValuePair(Q.second, P.first)) ||
1432 PairableInstUsers.count(ValuePair(Q.second, P.second));
1433 if (PairableInstUserMap) {
1434 // FIXME: The expensive part of the cycle check is not so much the cycle
1435 // check itself but this edge insertion procedure. This needs some
1436 // profiling and probably a different data structure.
1438 if (PairableInstUserPairSet->insert(VPPair(Q, P)).second)
1439 (*PairableInstUserMap)[Q].push_back(P);
1442 if (PairableInstUserPairSet->insert(VPPair(P, Q)).second)
1443 (*PairableInstUserMap)[P].push_back(Q);
1447 return (QUsesP && PUsesQ);
1450 // This function walks the use graph of current pairs to see if, starting
1451 // from P, the walk returns to P.
1452 bool BBVectorize::pairWillFormCycle(ValuePair P,
1453 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap,
1454 DenseSet<ValuePair> &CurrentPairs) {
1455 DEBUG(if (DebugCycleCheck)
1456 dbgs() << "BBV: starting cycle check for : " << *P.first << " <-> "
1457 << *P.second << "\n");
1458 // A lookup table of visisted pairs is kept because the PairableInstUserMap
1459 // contains non-direct associations.
1460 DenseSet<ValuePair> Visited;
1461 SmallVector<ValuePair, 32> Q;
1462 // General depth-first post-order traversal:
1465 ValuePair QTop = Q.pop_back_val();
1466 Visited.insert(QTop);
1468 DEBUG(if (DebugCycleCheck)
1469 dbgs() << "BBV: cycle check visiting: " << *QTop.first << " <-> "
1470 << *QTop.second << "\n");
1471 DenseMap<ValuePair, std::vector<ValuePair> >::iterator QQ =
1472 PairableInstUserMap.find(QTop);
1473 if (QQ == PairableInstUserMap.end())
1476 for (std::vector<ValuePair>::iterator C = QQ->second.begin(),
1477 CE = QQ->second.end(); C != CE; ++C) {
1480 << "BBV: rejected to prevent non-trivial cycle formation: "
1481 << QTop.first << " <-> " << C->second << "\n");
1485 if (CurrentPairs.count(*C) && !Visited.count(*C))
1488 } while (!Q.empty());
1493 // This function builds the initial tree of connected pairs with the
1494 // pair J at the root.
1495 void BBVectorize::buildInitialTreeFor(
1496 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1497 DenseSet<ValuePair> &CandidatePairsSet,
1498 std::vector<Value *> &PairableInsts,
1499 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
1500 DenseSet<ValuePair> &PairableInstUsers,
1501 DenseMap<Value *, Value *> &ChosenPairs,
1502 DenseMap<ValuePair, size_t> &Tree, ValuePair J) {
1503 // Each of these pairs is viewed as the root node of a Tree. The Tree
1504 // is then walked (depth-first). As this happens, we keep track of
1505 // the pairs that compose the Tree and the maximum depth of the Tree.
1506 SmallVector<ValuePairWithDepth, 32> Q;
1507 // General depth-first post-order traversal:
1508 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first)));
1510 ValuePairWithDepth QTop = Q.back();
1512 // Push each child onto the queue:
1513 bool MoreChildren = false;
1514 size_t MaxChildDepth = QTop.second;
1515 DenseMap<ValuePair, std::vector<ValuePair> >::iterator QQ =
1516 ConnectedPairs.find(QTop.first);
1517 if (QQ != ConnectedPairs.end())
1518 for (std::vector<ValuePair>::iterator k = QQ->second.begin(),
1519 ke = QQ->second.end(); k != ke; ++k) {
1520 // Make sure that this child pair is still a candidate:
1521 if (CandidatePairsSet.count(*k)) {
1522 DenseMap<ValuePair, size_t>::iterator C = Tree.find(*k);
1523 if (C == Tree.end()) {
1524 size_t d = getDepthFactor(k->first);
1525 Q.push_back(ValuePairWithDepth(*k, QTop.second+d));
1526 MoreChildren = true;
1528 MaxChildDepth = std::max(MaxChildDepth, C->second);
1533 if (!MoreChildren) {
1534 // Record the current pair as part of the Tree:
1535 Tree.insert(ValuePairWithDepth(QTop.first, MaxChildDepth));
1538 } while (!Q.empty());
1541 // Given some initial tree, prune it by removing conflicting pairs (pairs
1542 // that cannot be simultaneously chosen for vectorization).
1543 void BBVectorize::pruneTreeFor(
1544 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1545 std::vector<Value *> &PairableInsts,
1546 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
1547 DenseSet<ValuePair> &PairableInstUsers,
1548 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap,
1549 DenseSet<VPPair> &PairableInstUserPairSet,
1550 DenseMap<Value *, Value *> &ChosenPairs,
1551 DenseMap<ValuePair, size_t> &Tree,
1552 DenseSet<ValuePair> &PrunedTree, ValuePair J,
1553 bool UseCycleCheck) {
1554 SmallVector<ValuePairWithDepth, 32> Q;
1555 // General depth-first post-order traversal:
1556 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first)));
1558 ValuePairWithDepth QTop = Q.pop_back_val();
1559 PrunedTree.insert(QTop.first);
1561 // Visit each child, pruning as necessary...
1562 SmallVector<ValuePairWithDepth, 8> BestChildren;
1563 DenseMap<ValuePair, std::vector<ValuePair> >::iterator QQ =
1564 ConnectedPairs.find(QTop.first);
1565 if (QQ == ConnectedPairs.end())
1568 for (std::vector<ValuePair>::iterator K = QQ->second.begin(),
1569 KE = QQ->second.end(); K != KE; ++K) {
1570 DenseMap<ValuePair, size_t>::iterator C = Tree.find(*K);
1571 if (C == Tree.end()) continue;
1573 // This child is in the Tree, now we need to make sure it is the
1574 // best of any conflicting children. There could be multiple
1575 // conflicting children, so first, determine if we're keeping
1576 // this child, then delete conflicting children as necessary.
1578 // It is also necessary to guard against pairing-induced
1579 // dependencies. Consider instructions a .. x .. y .. b
1580 // such that (a,b) are to be fused and (x,y) are to be fused
1581 // but a is an input to x and b is an output from y. This
1582 // means that y cannot be moved after b but x must be moved
1583 // after b for (a,b) to be fused. In other words, after
1584 // fusing (a,b) we have y .. a/b .. x where y is an input
1585 // to a/b and x is an output to a/b: x and y can no longer
1586 // be legally fused. To prevent this condition, we must
1587 // make sure that a child pair added to the Tree is not
1588 // both an input and output of an already-selected pair.
1590 // Pairing-induced dependencies can also form from more complicated
1591 // cycles. The pair vs. pair conflicts are easy to check, and so
1592 // that is done explicitly for "fast rejection", and because for
1593 // child vs. child conflicts, we may prefer to keep the current
1594 // pair in preference to the already-selected child.
1595 DenseSet<ValuePair> CurrentPairs;
1598 for (SmallVector<ValuePairWithDepth, 8>::iterator C2
1599 = BestChildren.begin(), E2 = BestChildren.end();
1601 if (C2->first.first == C->first.first ||
1602 C2->first.first == C->first.second ||
1603 C2->first.second == C->first.first ||
1604 C2->first.second == C->first.second ||
1605 pairsConflict(C2->first, C->first, PairableInstUsers,
1606 UseCycleCheck ? &PairableInstUserMap : 0,
1607 UseCycleCheck ? &PairableInstUserPairSet : 0)) {
1608 if (C2->second >= C->second) {
1613 CurrentPairs.insert(C2->first);
1616 if (!CanAdd) continue;
1618 // Even worse, this child could conflict with another node already
1619 // selected for the Tree. If that is the case, ignore this child.
1620 for (DenseSet<ValuePair>::iterator T = PrunedTree.begin(),
1621 E2 = PrunedTree.end(); T != E2; ++T) {
1622 if (T->first == C->first.first ||
1623 T->first == C->first.second ||
1624 T->second == C->first.first ||
1625 T->second == C->first.second ||
1626 pairsConflict(*T, C->first, PairableInstUsers,
1627 UseCycleCheck ? &PairableInstUserMap : 0,
1628 UseCycleCheck ? &PairableInstUserPairSet : 0)) {
1633 CurrentPairs.insert(*T);
1635 if (!CanAdd) continue;
1637 // And check the queue too...
1638 for (SmallVector<ValuePairWithDepth, 32>::iterator C2 = Q.begin(),
1639 E2 = Q.end(); C2 != E2; ++C2) {
1640 if (C2->first.first == C->first.first ||
1641 C2->first.first == C->first.second ||
1642 C2->first.second == C->first.first ||
1643 C2->first.second == C->first.second ||
1644 pairsConflict(C2->first, C->first, PairableInstUsers,
1645 UseCycleCheck ? &PairableInstUserMap : 0,
1646 UseCycleCheck ? &PairableInstUserPairSet : 0)) {
1651 CurrentPairs.insert(C2->first);
1653 if (!CanAdd) continue;
1655 // Last but not least, check for a conflict with any of the
1656 // already-chosen pairs.
1657 for (DenseMap<Value *, Value *>::iterator C2 =
1658 ChosenPairs.begin(), E2 = ChosenPairs.end();
1660 if (pairsConflict(*C2, C->first, PairableInstUsers,
1661 UseCycleCheck ? &PairableInstUserMap : 0,
1662 UseCycleCheck ? &PairableInstUserPairSet : 0)) {
1667 CurrentPairs.insert(*C2);
1669 if (!CanAdd) continue;
1671 // To check for non-trivial cycles formed by the addition of the
1672 // current pair we've formed a list of all relevant pairs, now use a
1673 // graph walk to check for a cycle. We start from the current pair and
1674 // walk the use tree to see if we again reach the current pair. If we
1675 // do, then the current pair is rejected.
1677 // FIXME: It may be more efficient to use a topological-ordering
1678 // algorithm to improve the cycle check. This should be investigated.
1679 if (UseCycleCheck &&
1680 pairWillFormCycle(C->first, PairableInstUserMap, CurrentPairs))
1683 // This child can be added, but we may have chosen it in preference
1684 // to an already-selected child. Check for this here, and if a
1685 // conflict is found, then remove the previously-selected child
1686 // before adding this one in its place.
1687 for (SmallVector<ValuePairWithDepth, 8>::iterator C2
1688 = BestChildren.begin(); C2 != BestChildren.end();) {
1689 if (C2->first.first == C->first.first ||
1690 C2->first.first == C->first.second ||
1691 C2->first.second == C->first.first ||
1692 C2->first.second == C->first.second ||
1693 pairsConflict(C2->first, C->first, PairableInstUsers))
1694 C2 = BestChildren.erase(C2);
1699 BestChildren.push_back(ValuePairWithDepth(C->first, C->second));
1702 for (SmallVector<ValuePairWithDepth, 8>::iterator C
1703 = BestChildren.begin(), E2 = BestChildren.end();
1705 size_t DepthF = getDepthFactor(C->first.first);
1706 Q.push_back(ValuePairWithDepth(C->first, QTop.second+DepthF));
1708 } while (!Q.empty());
1711 // This function finds the best tree of mututally-compatible connected
1712 // pairs, given the choice of root pairs as an iterator range.
1713 void BBVectorize::findBestTreeFor(
1714 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1715 DenseSet<ValuePair> &CandidatePairsSet,
1716 DenseMap<ValuePair, int> &CandidatePairCostSavings,
1717 std::vector<Value *> &PairableInsts,
1718 DenseSet<ValuePair> &FixedOrderPairs,
1719 DenseMap<VPPair, unsigned> &PairConnectionTypes,
1720 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
1721 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps,
1722 DenseSet<ValuePair> &PairableInstUsers,
1723 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap,
1724 DenseSet<VPPair> &PairableInstUserPairSet,
1725 DenseMap<Value *, Value *> &ChosenPairs,
1726 DenseSet<ValuePair> &BestTree, size_t &BestMaxDepth,
1727 int &BestEffSize, Value *II, std::vector<Value *>&JJ,
1728 bool UseCycleCheck) {
1729 for (std::vector<Value *>::iterator J = JJ.begin(), JE = JJ.end();
1731 ValuePair IJ(II, *J);
1732 if (!CandidatePairsSet.count(IJ))
1735 // Before going any further, make sure that this pair does not
1736 // conflict with any already-selected pairs (see comment below
1737 // near the Tree pruning for more details).
1738 DenseSet<ValuePair> ChosenPairSet;
1739 bool DoesConflict = false;
1740 for (DenseMap<Value *, Value *>::iterator C = ChosenPairs.begin(),
1741 E = ChosenPairs.end(); C != E; ++C) {
1742 if (pairsConflict(*C, IJ, PairableInstUsers,
1743 UseCycleCheck ? &PairableInstUserMap : 0,
1744 UseCycleCheck ? &PairableInstUserPairSet : 0)) {
1745 DoesConflict = true;
1749 ChosenPairSet.insert(*C);
1751 if (DoesConflict) continue;
1753 if (UseCycleCheck &&
1754 pairWillFormCycle(IJ, PairableInstUserMap, ChosenPairSet))
1757 DenseMap<ValuePair, size_t> Tree;
1758 buildInitialTreeFor(CandidatePairs, CandidatePairsSet,
1759 PairableInsts, ConnectedPairs,
1760 PairableInstUsers, ChosenPairs, Tree, IJ);
1762 // Because we'll keep the child with the largest depth, the largest
1763 // depth is still the same in the unpruned Tree.
1764 size_t MaxDepth = Tree.lookup(IJ);
1766 DEBUG(if (DebugPairSelection) dbgs() << "BBV: found Tree for pair {"
1767 << IJ.first << " <-> " << IJ.second << "} of depth " <<
1768 MaxDepth << " and size " << Tree.size() << "\n");
1770 // At this point the Tree has been constructed, but, may contain
1771 // contradictory children (meaning that different children of
1772 // some tree node may be attempting to fuse the same instruction).
1773 // So now we walk the tree again, in the case of a conflict,
1774 // keep only the child with the largest depth. To break a tie,
1775 // favor the first child.
1777 DenseSet<ValuePair> PrunedTree;
1778 pruneTreeFor(CandidatePairs, PairableInsts, ConnectedPairs,
1779 PairableInstUsers, PairableInstUserMap,
1780 PairableInstUserPairSet,
1781 ChosenPairs, Tree, PrunedTree, IJ, UseCycleCheck);
1785 DenseSet<Value *> PrunedTreeInstrs;
1786 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(),
1787 E = PrunedTree.end(); S != E; ++S) {
1788 PrunedTreeInstrs.insert(S->first);
1789 PrunedTreeInstrs.insert(S->second);
1792 // The set of pairs that have already contributed to the total cost.
1793 DenseSet<ValuePair> IncomingPairs;
1795 // If the cost model were perfect, this might not be necessary; but we
1796 // need to make sure that we don't get stuck vectorizing our own
1798 bool HasNontrivialInsts = false;
1800 // The node weights represent the cost savings associated with
1801 // fusing the pair of instructions.
1802 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(),
1803 E = PrunedTree.end(); S != E; ++S) {
1804 if (!isa<ShuffleVectorInst>(S->first) &&
1805 !isa<InsertElementInst>(S->first) &&
1806 !isa<ExtractElementInst>(S->first))
1807 HasNontrivialInsts = true;
1809 bool FlipOrder = false;
1811 if (getDepthFactor(S->first)) {
1812 int ESContrib = CandidatePairCostSavings.find(*S)->second;
1813 DEBUG(if (DebugPairSelection) dbgs() << "\tweight {"
1814 << *S->first << " <-> " << *S->second << "} = " <<
1816 EffSize += ESContrib;
1819 // The edge weights contribute in a negative sense: they represent
1820 // the cost of shuffles.
1821 DenseMap<ValuePair, std::vector<ValuePair> >::iterator SS =
1822 ConnectedPairDeps.find(*S);
1823 if (SS != ConnectedPairDeps.end()) {
1824 unsigned NumDepsDirect = 0, NumDepsSwap = 0;
1825 for (std::vector<ValuePair>::iterator T = SS->second.begin(),
1826 TE = SS->second.end(); T != TE; ++T) {
1828 if (!PrunedTree.count(Q.second))
1830 DenseMap<VPPair, unsigned>::iterator R =
1831 PairConnectionTypes.find(VPPair(Q.second, Q.first));
1832 assert(R != PairConnectionTypes.end() &&
1833 "Cannot find pair connection type");
1834 if (R->second == PairConnectionDirect)
1836 else if (R->second == PairConnectionSwap)
1840 // If there are more swaps than direct connections, then
1841 // the pair order will be flipped during fusion. So the real
1842 // number of swaps is the minimum number.
1843 FlipOrder = !FixedOrderPairs.count(*S) &&
1844 ((NumDepsSwap > NumDepsDirect) ||
1845 FixedOrderPairs.count(ValuePair(S->second, S->first)));
1847 for (std::vector<ValuePair>::iterator T = SS->second.begin(),
1848 TE = SS->second.end(); T != TE; ++T) {
1850 if (!PrunedTree.count(Q.second))
1852 DenseMap<VPPair, unsigned>::iterator R =
1853 PairConnectionTypes.find(VPPair(Q.second, Q.first));
1854 assert(R != PairConnectionTypes.end() &&
1855 "Cannot find pair connection type");
1856 Type *Ty1 = Q.second.first->getType(),
1857 *Ty2 = Q.second.second->getType();
1858 Type *VTy = getVecTypeForPair(Ty1, Ty2);
1859 if ((R->second == PairConnectionDirect && FlipOrder) ||
1860 (R->second == PairConnectionSwap && !FlipOrder) ||
1861 R->second == PairConnectionSplat) {
1862 int ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
1865 if (VTy->getVectorNumElements() == 2) {
1866 if (R->second == PairConnectionSplat)
1867 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
1868 TargetTransformInfo::SK_Broadcast, VTy));
1870 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
1871 TargetTransformInfo::SK_Reverse, VTy));
1874 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" <<
1875 *Q.second.first << " <-> " << *Q.second.second <<
1877 *S->first << " <-> " << *S->second << "} = " <<
1879 EffSize -= ESContrib;
1884 // Compute the cost of outgoing edges. We assume that edges outgoing
1885 // to shuffles, inserts or extracts can be merged, and so contribute
1886 // no additional cost.
1887 if (!S->first->getType()->isVoidTy()) {
1888 Type *Ty1 = S->first->getType(),
1889 *Ty2 = S->second->getType();
1890 Type *VTy = getVecTypeForPair(Ty1, Ty2);
1892 bool NeedsExtraction = false;
1893 for (Value::use_iterator I = S->first->use_begin(),
1894 IE = S->first->use_end(); I != IE; ++I) {
1895 if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(*I)) {
1896 // Shuffle can be folded if it has no other input
1897 if (isa<UndefValue>(SI->getOperand(1)))
1900 if (isa<ExtractElementInst>(*I))
1902 if (PrunedTreeInstrs.count(*I))
1904 NeedsExtraction = true;
1908 if (NeedsExtraction) {
1910 if (Ty1->isVectorTy()) {
1911 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
1913 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
1914 TargetTransformInfo::SK_ExtractSubvector, VTy, 0, Ty1));
1916 ESContrib = (int) TTI->getVectorInstrCost(
1917 Instruction::ExtractElement, VTy, 0);
1919 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" <<
1920 *S->first << "} = " << ESContrib << "\n");
1921 EffSize -= ESContrib;
1924 NeedsExtraction = false;
1925 for (Value::use_iterator I = S->second->use_begin(),
1926 IE = S->second->use_end(); I != IE; ++I) {
1927 if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(*I)) {
1928 // Shuffle can be folded if it has no other input
1929 if (isa<UndefValue>(SI->getOperand(1)))
1932 if (isa<ExtractElementInst>(*I))
1934 if (PrunedTreeInstrs.count(*I))
1936 NeedsExtraction = true;
1940 if (NeedsExtraction) {
1942 if (Ty2->isVectorTy()) {
1943 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
1945 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
1946 TargetTransformInfo::SK_ExtractSubvector, VTy,
1947 Ty1->isVectorTy() ? Ty1->getVectorNumElements() : 1, Ty2));
1949 ESContrib = (int) TTI->getVectorInstrCost(
1950 Instruction::ExtractElement, VTy, 1);
1951 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" <<
1952 *S->second << "} = " << ESContrib << "\n");
1953 EffSize -= ESContrib;
1957 // Compute the cost of incoming edges.
1958 if (!isa<LoadInst>(S->first) && !isa<StoreInst>(S->first)) {
1959 Instruction *S1 = cast<Instruction>(S->first),
1960 *S2 = cast<Instruction>(S->second);
1961 for (unsigned o = 0; o < S1->getNumOperands(); ++o) {
1962 Value *O1 = S1->getOperand(o), *O2 = S2->getOperand(o);
1964 // Combining constants into vector constants (or small vector
1965 // constants into larger ones are assumed free).
1966 if (isa<Constant>(O1) && isa<Constant>(O2))
1972 ValuePair VP = ValuePair(O1, O2);
1973 ValuePair VPR = ValuePair(O2, O1);
1975 // Internal edges are not handled here.
1976 if (PrunedTree.count(VP) || PrunedTree.count(VPR))
1979 Type *Ty1 = O1->getType(),
1980 *Ty2 = O2->getType();
1981 Type *VTy = getVecTypeForPair(Ty1, Ty2);
1983 // Combining vector operations of the same type is also assumed
1984 // folded with other operations.
1986 // If both are insert elements, then both can be widened.
1987 InsertElementInst *IEO1 = dyn_cast<InsertElementInst>(O1),
1988 *IEO2 = dyn_cast<InsertElementInst>(O2);
1989 if (IEO1 && IEO2 && isPureIEChain(IEO1) && isPureIEChain(IEO2))
1991 // If both are extract elements, and both have the same input
1992 // type, then they can be replaced with a shuffle
1993 ExtractElementInst *EIO1 = dyn_cast<ExtractElementInst>(O1),
1994 *EIO2 = dyn_cast<ExtractElementInst>(O2);
1996 EIO1->getOperand(0)->getType() ==
1997 EIO2->getOperand(0)->getType())
1999 // If both are a shuffle with equal operand types and only two
2000 // unqiue operands, then they can be replaced with a single
2002 ShuffleVectorInst *SIO1 = dyn_cast<ShuffleVectorInst>(O1),
2003 *SIO2 = dyn_cast<ShuffleVectorInst>(O2);
2005 SIO1->getOperand(0)->getType() ==
2006 SIO2->getOperand(0)->getType()) {
2007 SmallSet<Value *, 4> SIOps;
2008 SIOps.insert(SIO1->getOperand(0));
2009 SIOps.insert(SIO1->getOperand(1));
2010 SIOps.insert(SIO2->getOperand(0));
2011 SIOps.insert(SIO2->getOperand(1));
2012 if (SIOps.size() <= 2)
2018 // This pair has already been formed.
2019 if (IncomingPairs.count(VP)) {
2021 } else if (IncomingPairs.count(VPR)) {
2022 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
2025 if (VTy->getVectorNumElements() == 2)
2026 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
2027 TargetTransformInfo::SK_Reverse, VTy));
2028 } else if (!Ty1->isVectorTy() && !Ty2->isVectorTy()) {
2029 ESContrib = (int) TTI->getVectorInstrCost(
2030 Instruction::InsertElement, VTy, 0);
2031 ESContrib += (int) TTI->getVectorInstrCost(
2032 Instruction::InsertElement, VTy, 1);
2033 } else if (!Ty1->isVectorTy()) {
2034 // O1 needs to be inserted into a vector of size O2, and then
2035 // both need to be shuffled together.
2036 ESContrib = (int) TTI->getVectorInstrCost(
2037 Instruction::InsertElement, Ty2, 0);
2038 ESContrib += (int) getInstrCost(Instruction::ShuffleVector,
2040 } else if (!Ty2->isVectorTy()) {
2041 // O2 needs to be inserted into a vector of size O1, and then
2042 // both need to be shuffled together.
2043 ESContrib = (int) TTI->getVectorInstrCost(
2044 Instruction::InsertElement, Ty1, 0);
2045 ESContrib += (int) getInstrCost(Instruction::ShuffleVector,
2048 Type *TyBig = Ty1, *TySmall = Ty2;
2049 if (Ty2->getVectorNumElements() > Ty1->getVectorNumElements())
2050 std::swap(TyBig, TySmall);
2052 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
2054 if (TyBig != TySmall)
2055 ESContrib += (int) getInstrCost(Instruction::ShuffleVector,
2059 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {"
2060 << *O1 << " <-> " << *O2 << "} = " <<
2062 EffSize -= ESContrib;
2063 IncomingPairs.insert(VP);
2068 if (!HasNontrivialInsts) {
2069 DEBUG(if (DebugPairSelection) dbgs() <<
2070 "\tNo non-trivial instructions in tree;"
2071 " override to zero effective size\n");
2075 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(),
2076 E = PrunedTree.end(); S != E; ++S)
2077 EffSize += (int) getDepthFactor(S->first);
2080 DEBUG(if (DebugPairSelection)
2081 dbgs() << "BBV: found pruned Tree for pair {"
2082 << IJ.first << " <-> " << IJ.second << "} of depth " <<
2083 MaxDepth << " and size " << PrunedTree.size() <<
2084 " (effective size: " << EffSize << ")\n");
2085 if (((TTI && !UseChainDepthWithTI) ||
2086 MaxDepth >= Config.ReqChainDepth) &&
2087 EffSize > 0 && EffSize > BestEffSize) {
2088 BestMaxDepth = MaxDepth;
2089 BestEffSize = EffSize;
2090 BestTree = PrunedTree;
2095 // Given the list of candidate pairs, this function selects those
2096 // that will be fused into vector instructions.
2097 void BBVectorize::choosePairs(
2098 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
2099 DenseSet<ValuePair> &CandidatePairsSet,
2100 DenseMap<ValuePair, int> &CandidatePairCostSavings,
2101 std::vector<Value *> &PairableInsts,
2102 DenseSet<ValuePair> &FixedOrderPairs,
2103 DenseMap<VPPair, unsigned> &PairConnectionTypes,
2104 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
2105 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps,
2106 DenseSet<ValuePair> &PairableInstUsers,
2107 DenseMap<Value *, Value *>& ChosenPairs) {
2108 bool UseCycleCheck =
2109 CandidatePairsSet.size() <= Config.MaxCandPairsForCycleCheck;
2111 DenseMap<Value *, std::vector<Value *> > CandidatePairs2;
2112 for (DenseSet<ValuePair>::iterator I = CandidatePairsSet.begin(),
2113 E = CandidatePairsSet.end(); I != E; ++I) {
2114 std::vector<Value *> &JJ = CandidatePairs2[I->second];
2115 if (JJ.empty()) JJ.reserve(32);
2116 JJ.push_back(I->first);
2119 DenseMap<ValuePair, std::vector<ValuePair> > PairableInstUserMap;
2120 DenseSet<VPPair> PairableInstUserPairSet;
2121 for (std::vector<Value *>::iterator I = PairableInsts.begin(),
2122 E = PairableInsts.end(); I != E; ++I) {
2123 // The number of possible pairings for this variable:
2124 size_t NumChoices = CandidatePairs.lookup(*I).size();
2125 if (!NumChoices) continue;
2127 std::vector<Value *> &JJ = CandidatePairs[*I];
2129 // The best pair to choose and its tree:
2130 size_t BestMaxDepth = 0;
2131 int BestEffSize = 0;
2132 DenseSet<ValuePair> BestTree;
2133 findBestTreeFor(CandidatePairs, CandidatePairsSet,
2134 CandidatePairCostSavings,
2135 PairableInsts, FixedOrderPairs, PairConnectionTypes,
2136 ConnectedPairs, ConnectedPairDeps,
2137 PairableInstUsers, PairableInstUserMap,
2138 PairableInstUserPairSet, ChosenPairs,
2139 BestTree, BestMaxDepth, BestEffSize, *I, JJ,
2142 if (BestTree.empty())
2145 // A tree has been chosen (or not) at this point. If no tree was
2146 // chosen, then this instruction, I, cannot be paired (and is no longer
2149 DEBUG(dbgs() << "BBV: selected pairs in the best tree for: "
2150 << *cast<Instruction>(*I) << "\n");
2152 for (DenseSet<ValuePair>::iterator S = BestTree.begin(),
2153 SE2 = BestTree.end(); S != SE2; ++S) {
2154 // Insert the members of this tree into the list of chosen pairs.
2155 ChosenPairs.insert(ValuePair(S->first, S->second));
2156 DEBUG(dbgs() << "BBV: selected pair: " << *S->first << " <-> " <<
2157 *S->second << "\n");
2159 // Remove all candidate pairs that have values in the chosen tree.
2160 std::vector<Value *> &KK = CandidatePairs[S->first],
2161 &LL = CandidatePairs2[S->second],
2162 &MM = CandidatePairs[S->second],
2163 &NN = CandidatePairs2[S->first];
2164 for (std::vector<Value *>::iterator K = KK.begin(), KE = KK.end();
2166 if (*K == S->second)
2169 CandidatePairsSet.erase(ValuePair(S->first, *K));
2171 for (std::vector<Value *>::iterator L = LL.begin(), LE = LL.end();
2176 CandidatePairsSet.erase(ValuePair(*L, S->second));
2178 for (std::vector<Value *>::iterator M = MM.begin(), ME = MM.end();
2180 assert(*M != S->first && "Flipped pair in candidate list?");
2181 CandidatePairsSet.erase(ValuePair(S->second, *M));
2183 for (std::vector<Value *>::iterator N = NN.begin(), NE = NN.end();
2185 assert(*N != S->second && "Flipped pair in candidate list?");
2186 CandidatePairsSet.erase(ValuePair(*N, S->first));
2191 DEBUG(dbgs() << "BBV: selected " << ChosenPairs.size() << " pairs.\n");
2194 std::string getReplacementName(Instruction *I, bool IsInput, unsigned o,
2199 return (I->getName() + (IsInput ? ".v.i" : ".v.r") + utostr(o) +
2200 (n > 0 ? "." + utostr(n) : "")).str();
2203 // Returns the value that is to be used as the pointer input to the vector
2204 // instruction that fuses I with J.
2205 Value *BBVectorize::getReplacementPointerInput(LLVMContext& Context,
2206 Instruction *I, Instruction *J, unsigned o) {
2208 unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace;
2209 int64_t OffsetInElmts;
2211 // Note: the analysis might fail here, that is why the pair order has
2212 // been precomputed (OffsetInElmts must be unused here).
2213 (void) getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment,
2214 IAddressSpace, JAddressSpace,
2215 OffsetInElmts, false);
2217 // The pointer value is taken to be the one with the lowest offset.
2220 Type *ArgTypeI = cast<PointerType>(IPtr->getType())->getElementType();
2221 Type *ArgTypeJ = cast<PointerType>(JPtr->getType())->getElementType();
2222 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2223 Type *VArgPtrType = PointerType::get(VArgType,
2224 cast<PointerType>(IPtr->getType())->getAddressSpace());
2225 return new BitCastInst(VPtr, VArgPtrType, getReplacementName(I, true, o),
2226 /* insert before */ I);
2229 void BBVectorize::fillNewShuffleMask(LLVMContext& Context, Instruction *J,
2230 unsigned MaskOffset, unsigned NumInElem,
2231 unsigned NumInElem1, unsigned IdxOffset,
2232 std::vector<Constant*> &Mask) {
2233 unsigned NumElem1 = cast<VectorType>(J->getType())->getNumElements();
2234 for (unsigned v = 0; v < NumElem1; ++v) {
2235 int m = cast<ShuffleVectorInst>(J)->getMaskValue(v);
2237 Mask[v+MaskOffset] = UndefValue::get(Type::getInt32Ty(Context));
2239 unsigned mm = m + (int) IdxOffset;
2240 if (m >= (int) NumInElem1)
2241 mm += (int) NumInElem;
2243 Mask[v+MaskOffset] =
2244 ConstantInt::get(Type::getInt32Ty(Context), mm);
2249 // Returns the value that is to be used as the vector-shuffle mask to the
2250 // vector instruction that fuses I with J.
2251 Value *BBVectorize::getReplacementShuffleMask(LLVMContext& Context,
2252 Instruction *I, Instruction *J) {
2253 // This is the shuffle mask. We need to append the second
2254 // mask to the first, and the numbers need to be adjusted.
2256 Type *ArgTypeI = I->getType();
2257 Type *ArgTypeJ = J->getType();
2258 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2260 unsigned NumElemI = cast<VectorType>(ArgTypeI)->getNumElements();
2262 // Get the total number of elements in the fused vector type.
2263 // By definition, this must equal the number of elements in
2265 unsigned NumElem = cast<VectorType>(VArgType)->getNumElements();
2266 std::vector<Constant*> Mask(NumElem);
2268 Type *OpTypeI = I->getOperand(0)->getType();
2269 unsigned NumInElemI = cast<VectorType>(OpTypeI)->getNumElements();
2270 Type *OpTypeJ = J->getOperand(0)->getType();
2271 unsigned NumInElemJ = cast<VectorType>(OpTypeJ)->getNumElements();
2273 // The fused vector will be:
2274 // -----------------------------------------------------
2275 // | NumInElemI | NumInElemJ | NumInElemI | NumInElemJ |
2276 // -----------------------------------------------------
2277 // from which we'll extract NumElem total elements (where the first NumElemI
2278 // of them come from the mask in I and the remainder come from the mask
2281 // For the mask from the first pair...
2282 fillNewShuffleMask(Context, I, 0, NumInElemJ, NumInElemI,
2285 // For the mask from the second pair...
2286 fillNewShuffleMask(Context, J, NumElemI, NumInElemI, NumInElemJ,
2289 return ConstantVector::get(Mask);
2292 bool BBVectorize::expandIEChain(LLVMContext& Context, Instruction *I,
2293 Instruction *J, unsigned o, Value *&LOp,
2295 Type *ArgTypeL, Type *ArgTypeH,
2296 bool IBeforeJ, unsigned IdxOff) {
2297 bool ExpandedIEChain = false;
2298 if (InsertElementInst *LIE = dyn_cast<InsertElementInst>(LOp)) {
2299 // If we have a pure insertelement chain, then this can be rewritten
2300 // into a chain that directly builds the larger type.
2301 if (isPureIEChain(LIE)) {
2302 SmallVector<Value *, 8> VectElemts(numElemL,
2303 UndefValue::get(ArgTypeL->getScalarType()));
2304 InsertElementInst *LIENext = LIE;
2307 cast<ConstantInt>(LIENext->getOperand(2))->getSExtValue();
2308 VectElemts[Idx] = LIENext->getOperand(1);
2310 dyn_cast<InsertElementInst>(LIENext->getOperand(0))));
2313 Value *LIEPrev = UndefValue::get(ArgTypeH);
2314 for (unsigned i = 0; i < numElemL; ++i) {
2315 if (isa<UndefValue>(VectElemts[i])) continue;
2316 LIENext = InsertElementInst::Create(LIEPrev, VectElemts[i],
2317 ConstantInt::get(Type::getInt32Ty(Context),
2319 getReplacementName(IBeforeJ ? I : J,
2321 LIENext->insertBefore(IBeforeJ ? J : I);
2325 LOp = LIENext ? (Value*) LIENext : UndefValue::get(ArgTypeH);
2326 ExpandedIEChain = true;
2330 return ExpandedIEChain;
2333 // Returns the value to be used as the specified operand of the vector
2334 // instruction that fuses I with J.
2335 Value *BBVectorize::getReplacementInput(LLVMContext& Context, Instruction *I,
2336 Instruction *J, unsigned o, bool IBeforeJ) {
2337 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0);
2338 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), 1);
2340 // Compute the fused vector type for this operand
2341 Type *ArgTypeI = I->getOperand(o)->getType();
2342 Type *ArgTypeJ = J->getOperand(o)->getType();
2343 VectorType *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2345 Instruction *L = I, *H = J;
2346 Type *ArgTypeL = ArgTypeI, *ArgTypeH = ArgTypeJ;
2349 if (ArgTypeL->isVectorTy())
2350 numElemL = cast<VectorType>(ArgTypeL)->getNumElements();
2355 if (ArgTypeH->isVectorTy())
2356 numElemH = cast<VectorType>(ArgTypeH)->getNumElements();
2360 Value *LOp = L->getOperand(o);
2361 Value *HOp = H->getOperand(o);
2362 unsigned numElem = VArgType->getNumElements();
2364 // First, we check if we can reuse the "original" vector outputs (if these
2365 // exist). We might need a shuffle.
2366 ExtractElementInst *LEE = dyn_cast<ExtractElementInst>(LOp);
2367 ExtractElementInst *HEE = dyn_cast<ExtractElementInst>(HOp);
2368 ShuffleVectorInst *LSV = dyn_cast<ShuffleVectorInst>(LOp);
2369 ShuffleVectorInst *HSV = dyn_cast<ShuffleVectorInst>(HOp);
2371 // FIXME: If we're fusing shuffle instructions, then we can't apply this
2372 // optimization. The input vectors to the shuffle might be a different
2373 // length from the shuffle outputs. Unfortunately, the replacement
2374 // shuffle mask has already been formed, and the mask entries are sensitive
2375 // to the sizes of the inputs.
2376 bool IsSizeChangeShuffle =
2377 isa<ShuffleVectorInst>(L) &&
2378 (LOp->getType() != L->getType() || HOp->getType() != H->getType());
2380 if ((LEE || LSV) && (HEE || HSV) && !IsSizeChangeShuffle) {
2381 // We can have at most two unique vector inputs.
2382 bool CanUseInputs = true;
2385 I1 = LEE->getOperand(0);
2387 I1 = LSV->getOperand(0);
2388 I2 = LSV->getOperand(1);
2389 if (I2 == I1 || isa<UndefValue>(I2))
2394 Value *I3 = HEE->getOperand(0);
2395 if (!I2 && I3 != I1)
2397 else if (I3 != I1 && I3 != I2)
2398 CanUseInputs = false;
2400 Value *I3 = HSV->getOperand(0);
2401 if (!I2 && I3 != I1)
2403 else if (I3 != I1 && I3 != I2)
2404 CanUseInputs = false;
2407 Value *I4 = HSV->getOperand(1);
2408 if (!isa<UndefValue>(I4)) {
2409 if (!I2 && I4 != I1)
2411 else if (I4 != I1 && I4 != I2)
2412 CanUseInputs = false;
2419 cast<VectorType>(cast<Instruction>(LOp)->getOperand(0)->getType())
2422 cast<VectorType>(cast<Instruction>(HOp)->getOperand(0)->getType())
2425 // We have one or two input vectors. We need to map each index of the
2426 // operands to the index of the original vector.
2427 SmallVector<std::pair<int, int>, 8> II(numElem);
2428 for (unsigned i = 0; i < numElemL; ++i) {
2432 cast<ConstantInt>(LEE->getOperand(1))->getSExtValue();
2433 INum = LEE->getOperand(0) == I1 ? 0 : 1;
2435 Idx = LSV->getMaskValue(i);
2436 if (Idx < (int) LOpElem) {
2437 INum = LSV->getOperand(0) == I1 ? 0 : 1;
2440 INum = LSV->getOperand(1) == I1 ? 0 : 1;
2444 II[i] = std::pair<int, int>(Idx, INum);
2446 for (unsigned i = 0; i < numElemH; ++i) {
2450 cast<ConstantInt>(HEE->getOperand(1))->getSExtValue();
2451 INum = HEE->getOperand(0) == I1 ? 0 : 1;
2453 Idx = HSV->getMaskValue(i);
2454 if (Idx < (int) HOpElem) {
2455 INum = HSV->getOperand(0) == I1 ? 0 : 1;
2458 INum = HSV->getOperand(1) == I1 ? 0 : 1;
2462 II[i + numElemL] = std::pair<int, int>(Idx, INum);
2465 // We now have an array which tells us from which index of which
2466 // input vector each element of the operand comes.
2467 VectorType *I1T = cast<VectorType>(I1->getType());
2468 unsigned I1Elem = I1T->getNumElements();
2471 // In this case there is only one underlying vector input. Check for
2472 // the trivial case where we can use the input directly.
2473 if (I1Elem == numElem) {
2474 bool ElemInOrder = true;
2475 for (unsigned i = 0; i < numElem; ++i) {
2476 if (II[i].first != (int) i && II[i].first != -1) {
2477 ElemInOrder = false;
2486 // A shuffle is needed.
2487 std::vector<Constant *> Mask(numElem);
2488 for (unsigned i = 0; i < numElem; ++i) {
2489 int Idx = II[i].first;
2491 Mask[i] = UndefValue::get(Type::getInt32Ty(Context));
2493 Mask[i] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
2497 new ShuffleVectorInst(I1, UndefValue::get(I1T),
2498 ConstantVector::get(Mask),
2499 getReplacementName(IBeforeJ ? I : J,
2501 S->insertBefore(IBeforeJ ? J : I);
2505 VectorType *I2T = cast<VectorType>(I2->getType());
2506 unsigned I2Elem = I2T->getNumElements();
2508 // This input comes from two distinct vectors. The first step is to
2509 // make sure that both vectors are the same length. If not, the
2510 // smaller one will need to grow before they can be shuffled together.
2511 if (I1Elem < I2Elem) {
2512 std::vector<Constant *> Mask(I2Elem);
2514 for (; v < I1Elem; ++v)
2515 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2516 for (; v < I2Elem; ++v)
2517 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2519 Instruction *NewI1 =
2520 new ShuffleVectorInst(I1, UndefValue::get(I1T),
2521 ConstantVector::get(Mask),
2522 getReplacementName(IBeforeJ ? I : J,
2524 NewI1->insertBefore(IBeforeJ ? J : I);
2528 } else if (I1Elem > I2Elem) {
2529 std::vector<Constant *> Mask(I1Elem);
2531 for (; v < I2Elem; ++v)
2532 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2533 for (; v < I1Elem; ++v)
2534 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2536 Instruction *NewI2 =
2537 new ShuffleVectorInst(I2, UndefValue::get(I2T),
2538 ConstantVector::get(Mask),
2539 getReplacementName(IBeforeJ ? I : J,
2541 NewI2->insertBefore(IBeforeJ ? J : I);
2547 // Now that both I1 and I2 are the same length we can shuffle them
2548 // together (and use the result).
2549 std::vector<Constant *> Mask(numElem);
2550 for (unsigned v = 0; v < numElem; ++v) {
2551 if (II[v].first == -1) {
2552 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2554 int Idx = II[v].first + II[v].second * I1Elem;
2555 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
2559 Instruction *NewOp =
2560 new ShuffleVectorInst(I1, I2, ConstantVector::get(Mask),
2561 getReplacementName(IBeforeJ ? I : J, true, o));
2562 NewOp->insertBefore(IBeforeJ ? J : I);
2567 Type *ArgType = ArgTypeL;
2568 if (numElemL < numElemH) {
2569 if (numElemL == 1 && expandIEChain(Context, I, J, o, HOp, numElemH,
2570 ArgTypeL, VArgType, IBeforeJ, 1)) {
2571 // This is another short-circuit case: we're combining a scalar into
2572 // a vector that is formed by an IE chain. We've just expanded the IE
2573 // chain, now insert the scalar and we're done.
2575 Instruction *S = InsertElementInst::Create(HOp, LOp, CV0,
2576 getReplacementName(IBeforeJ ? I : J, true, o));
2577 S->insertBefore(IBeforeJ ? J : I);
2579 } else if (!expandIEChain(Context, I, J, o, LOp, numElemL, ArgTypeL,
2580 ArgTypeH, IBeforeJ)) {
2581 // The two vector inputs to the shuffle must be the same length,
2582 // so extend the smaller vector to be the same length as the larger one.
2586 std::vector<Constant *> Mask(numElemH);
2588 for (; v < numElemL; ++v)
2589 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2590 for (; v < numElemH; ++v)
2591 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2593 NLOp = new ShuffleVectorInst(LOp, UndefValue::get(ArgTypeL),
2594 ConstantVector::get(Mask),
2595 getReplacementName(IBeforeJ ? I : J,
2598 NLOp = InsertElementInst::Create(UndefValue::get(ArgTypeH), LOp, CV0,
2599 getReplacementName(IBeforeJ ? I : J,
2603 NLOp->insertBefore(IBeforeJ ? J : I);
2608 } else if (numElemL > numElemH) {
2609 if (numElemH == 1 && expandIEChain(Context, I, J, o, LOp, numElemL,
2610 ArgTypeH, VArgType, IBeforeJ)) {
2612 InsertElementInst::Create(LOp, HOp,
2613 ConstantInt::get(Type::getInt32Ty(Context),
2615 getReplacementName(IBeforeJ ? I : J,
2617 S->insertBefore(IBeforeJ ? J : I);
2619 } else if (!expandIEChain(Context, I, J, o, HOp, numElemH, ArgTypeH,
2620 ArgTypeL, IBeforeJ)) {
2623 std::vector<Constant *> Mask(numElemL);
2625 for (; v < numElemH; ++v)
2626 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2627 for (; v < numElemL; ++v)
2628 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2630 NHOp = new ShuffleVectorInst(HOp, UndefValue::get(ArgTypeH),
2631 ConstantVector::get(Mask),
2632 getReplacementName(IBeforeJ ? I : J,
2635 NHOp = InsertElementInst::Create(UndefValue::get(ArgTypeL), HOp, CV0,
2636 getReplacementName(IBeforeJ ? I : J,
2640 NHOp->insertBefore(IBeforeJ ? J : I);
2645 if (ArgType->isVectorTy()) {
2646 unsigned numElem = cast<VectorType>(VArgType)->getNumElements();
2647 std::vector<Constant*> Mask(numElem);
2648 for (unsigned v = 0; v < numElem; ++v) {
2650 // If the low vector was expanded, we need to skip the extra
2651 // undefined entries.
2652 if (v >= numElemL && numElemH > numElemL)
2653 Idx += (numElemH - numElemL);
2654 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
2657 Instruction *BV = new ShuffleVectorInst(LOp, HOp,
2658 ConstantVector::get(Mask),
2659 getReplacementName(IBeforeJ ? I : J, true, o));
2660 BV->insertBefore(IBeforeJ ? J : I);
2664 Instruction *BV1 = InsertElementInst::Create(
2665 UndefValue::get(VArgType), LOp, CV0,
2666 getReplacementName(IBeforeJ ? I : J,
2668 BV1->insertBefore(IBeforeJ ? J : I);
2669 Instruction *BV2 = InsertElementInst::Create(BV1, HOp, CV1,
2670 getReplacementName(IBeforeJ ? I : J,
2672 BV2->insertBefore(IBeforeJ ? J : I);
2676 // This function creates an array of values that will be used as the inputs
2677 // to the vector instruction that fuses I with J.
2678 void BBVectorize::getReplacementInputsForPair(LLVMContext& Context,
2679 Instruction *I, Instruction *J,
2680 SmallVector<Value *, 3> &ReplacedOperands,
2682 unsigned NumOperands = I->getNumOperands();
2684 for (unsigned p = 0, o = NumOperands-1; p < NumOperands; ++p, --o) {
2685 // Iterate backward so that we look at the store pointer
2686 // first and know whether or not we need to flip the inputs.
2688 if (isa<LoadInst>(I) || (o == 1 && isa<StoreInst>(I))) {
2689 // This is the pointer for a load/store instruction.
2690 ReplacedOperands[o] = getReplacementPointerInput(Context, I, J, o);
2692 } else if (isa<CallInst>(I)) {
2693 Function *F = cast<CallInst>(I)->getCalledFunction();
2694 Intrinsic::ID IID = (Intrinsic::ID) F->getIntrinsicID();
2695 if (o == NumOperands-1) {
2696 BasicBlock &BB = *I->getParent();
2698 Module *M = BB.getParent()->getParent();
2699 Type *ArgTypeI = I->getType();
2700 Type *ArgTypeJ = J->getType();
2701 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2703 ReplacedOperands[o] = Intrinsic::getDeclaration(M, IID, VArgType);
2705 } else if (IID == Intrinsic::powi && o == 1) {
2706 // The second argument of powi is a single integer and we've already
2707 // checked that both arguments are equal. As a result, we just keep
2708 // I's second argument.
2709 ReplacedOperands[o] = I->getOperand(o);
2712 } else if (isa<ShuffleVectorInst>(I) && o == NumOperands-1) {
2713 ReplacedOperands[o] = getReplacementShuffleMask(Context, I, J);
2717 ReplacedOperands[o] = getReplacementInput(Context, I, J, o, IBeforeJ);
2721 // This function creates two values that represent the outputs of the
2722 // original I and J instructions. These are generally vector shuffles
2723 // or extracts. In many cases, these will end up being unused and, thus,
2724 // eliminated by later passes.
2725 void BBVectorize::replaceOutputsOfPair(LLVMContext& Context, Instruction *I,
2726 Instruction *J, Instruction *K,
2727 Instruction *&InsertionPt,
2728 Instruction *&K1, Instruction *&K2) {
2729 if (isa<StoreInst>(I)) {
2730 AA->replaceWithNewValue(I, K);
2731 AA->replaceWithNewValue(J, K);
2733 Type *IType = I->getType();
2734 Type *JType = J->getType();
2736 VectorType *VType = getVecTypeForPair(IType, JType);
2737 unsigned numElem = VType->getNumElements();
2739 unsigned numElemI, numElemJ;
2740 if (IType->isVectorTy())
2741 numElemI = cast<VectorType>(IType)->getNumElements();
2745 if (JType->isVectorTy())
2746 numElemJ = cast<VectorType>(JType)->getNumElements();
2750 if (IType->isVectorTy()) {
2751 std::vector<Constant*> Mask1(numElemI), Mask2(numElemI);
2752 for (unsigned v = 0; v < numElemI; ++v) {
2753 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2754 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemJ+v);
2757 K1 = new ShuffleVectorInst(K, UndefValue::get(VType),
2758 ConstantVector::get( Mask1),
2759 getReplacementName(K, false, 1));
2761 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0);
2762 K1 = ExtractElementInst::Create(K, CV0,
2763 getReplacementName(K, false, 1));
2766 if (JType->isVectorTy()) {
2767 std::vector<Constant*> Mask1(numElemJ), Mask2(numElemJ);
2768 for (unsigned v = 0; v < numElemJ; ++v) {
2769 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2770 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemI+v);
2773 K2 = new ShuffleVectorInst(K, UndefValue::get(VType),
2774 ConstantVector::get( Mask2),
2775 getReplacementName(K, false, 2));
2777 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), numElem-1);
2778 K2 = ExtractElementInst::Create(K, CV1,
2779 getReplacementName(K, false, 2));
2783 K2->insertAfter(K1);
2788 // Move all uses of the function I (including pairing-induced uses) after J.
2789 bool BBVectorize::canMoveUsesOfIAfterJ(BasicBlock &BB,
2790 DenseSet<ValuePair> &LoadMoveSetPairs,
2791 Instruction *I, Instruction *J) {
2792 // Skip to the first instruction past I.
2793 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I));
2795 DenseSet<Value *> Users;
2796 AliasSetTracker WriteSet(*AA);
2797 for (; cast<Instruction>(L) != J; ++L)
2798 (void) trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSetPairs);
2800 assert(cast<Instruction>(L) == J &&
2801 "Tracking has not proceeded far enough to check for dependencies");
2802 // If J is now in the use set of I, then trackUsesOfI will return true
2803 // and we have a dependency cycle (and the fusing operation must abort).
2804 return !trackUsesOfI(Users, WriteSet, I, J, true, &LoadMoveSetPairs);
2807 // Move all uses of the function I (including pairing-induced uses) after J.
2808 void BBVectorize::moveUsesOfIAfterJ(BasicBlock &BB,
2809 DenseSet<ValuePair> &LoadMoveSetPairs,
2810 Instruction *&InsertionPt,
2811 Instruction *I, Instruction *J) {
2812 // Skip to the first instruction past I.
2813 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I));
2815 DenseSet<Value *> Users;
2816 AliasSetTracker WriteSet(*AA);
2817 for (; cast<Instruction>(L) != J;) {
2818 if (trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSetPairs)) {
2819 // Move this instruction
2820 Instruction *InstToMove = L; ++L;
2822 DEBUG(dbgs() << "BBV: moving: " << *InstToMove <<
2823 " to after " << *InsertionPt << "\n");
2824 InstToMove->removeFromParent();
2825 InstToMove->insertAfter(InsertionPt);
2826 InsertionPt = InstToMove;
2833 // Collect all load instruction that are in the move set of a given first
2834 // pair member. These loads depend on the first instruction, I, and so need
2835 // to be moved after J (the second instruction) when the pair is fused.
2836 void BBVectorize::collectPairLoadMoveSet(BasicBlock &BB,
2837 DenseMap<Value *, Value *> &ChosenPairs,
2838 DenseMap<Value *, std::vector<Value *> > &LoadMoveSet,
2839 DenseSet<ValuePair> &LoadMoveSetPairs,
2841 // Skip to the first instruction past I.
2842 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I));
2844 DenseSet<Value *> Users;
2845 AliasSetTracker WriteSet(*AA);
2847 // Note: We cannot end the loop when we reach J because J could be moved
2848 // farther down the use chain by another instruction pairing. Also, J
2849 // could be before I if this is an inverted input.
2850 for (BasicBlock::iterator E = BB.end(); cast<Instruction>(L) != E; ++L) {
2851 if (trackUsesOfI(Users, WriteSet, I, L)) {
2852 if (L->mayReadFromMemory()) {
2853 LoadMoveSet[L].push_back(I);
2854 LoadMoveSetPairs.insert(ValuePair(L, I));
2860 // In cases where both load/stores and the computation of their pointers
2861 // are chosen for vectorization, we can end up in a situation where the
2862 // aliasing analysis starts returning different query results as the
2863 // process of fusing instruction pairs continues. Because the algorithm
2864 // relies on finding the same use trees here as were found earlier, we'll
2865 // need to precompute the necessary aliasing information here and then
2866 // manually update it during the fusion process.
2867 void BBVectorize::collectLoadMoveSet(BasicBlock &BB,
2868 std::vector<Value *> &PairableInsts,
2869 DenseMap<Value *, Value *> &ChosenPairs,
2870 DenseMap<Value *, std::vector<Value *> > &LoadMoveSet,
2871 DenseSet<ValuePair> &LoadMoveSetPairs) {
2872 for (std::vector<Value *>::iterator PI = PairableInsts.begin(),
2873 PIE = PairableInsts.end(); PI != PIE; ++PI) {
2874 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(*PI);
2875 if (P == ChosenPairs.end()) continue;
2877 Instruction *I = cast<Instruction>(P->first);
2878 collectPairLoadMoveSet(BB, ChosenPairs, LoadMoveSet,
2879 LoadMoveSetPairs, I);
2883 // When the first instruction in each pair is cloned, it will inherit its
2884 // parent's metadata. This metadata must be combined with that of the other
2885 // instruction in a safe way.
2886 void BBVectorize::combineMetadata(Instruction *K, const Instruction *J) {
2887 SmallVector<std::pair<unsigned, MDNode*>, 4> Metadata;
2888 K->getAllMetadataOtherThanDebugLoc(Metadata);
2889 for (unsigned i = 0, n = Metadata.size(); i < n; ++i) {
2890 unsigned Kind = Metadata[i].first;
2891 MDNode *JMD = J->getMetadata(Kind);
2892 MDNode *KMD = Metadata[i].second;
2896 K->setMetadata(Kind, 0); // Remove unknown metadata
2898 case LLVMContext::MD_tbaa:
2899 K->setMetadata(Kind, MDNode::getMostGenericTBAA(JMD, KMD));
2901 case LLVMContext::MD_fpmath:
2902 K->setMetadata(Kind, MDNode::getMostGenericFPMath(JMD, KMD));
2908 // This function fuses the chosen instruction pairs into vector instructions,
2909 // taking care preserve any needed scalar outputs and, then, it reorders the
2910 // remaining instructions as needed (users of the first member of the pair
2911 // need to be moved to after the location of the second member of the pair
2912 // because the vector instruction is inserted in the location of the pair's
2914 void BBVectorize::fuseChosenPairs(BasicBlock &BB,
2915 std::vector<Value *> &PairableInsts,
2916 DenseMap<Value *, Value *> &ChosenPairs,
2917 DenseSet<ValuePair> &FixedOrderPairs,
2918 DenseMap<VPPair, unsigned> &PairConnectionTypes,
2919 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
2920 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps) {
2921 LLVMContext& Context = BB.getContext();
2923 // During the vectorization process, the order of the pairs to be fused
2924 // could be flipped. So we'll add each pair, flipped, into the ChosenPairs
2925 // list. After a pair is fused, the flipped pair is removed from the list.
2926 DenseSet<ValuePair> FlippedPairs;
2927 for (DenseMap<Value *, Value *>::iterator P = ChosenPairs.begin(),
2928 E = ChosenPairs.end(); P != E; ++P)
2929 FlippedPairs.insert(ValuePair(P->second, P->first));
2930 for (DenseSet<ValuePair>::iterator P = FlippedPairs.begin(),
2931 E = FlippedPairs.end(); P != E; ++P)
2932 ChosenPairs.insert(*P);
2934 DenseMap<Value *, std::vector<Value *> > LoadMoveSet;
2935 DenseSet<ValuePair> LoadMoveSetPairs;
2936 collectLoadMoveSet(BB, PairableInsts, ChosenPairs,
2937 LoadMoveSet, LoadMoveSetPairs);
2939 DEBUG(dbgs() << "BBV: initial: \n" << BB << "\n");
2941 for (BasicBlock::iterator PI = BB.getFirstInsertionPt(); PI != BB.end();) {
2942 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(PI);
2943 if (P == ChosenPairs.end()) {
2948 if (getDepthFactor(P->first) == 0) {
2949 // These instructions are not really fused, but are tracked as though
2950 // they are. Any case in which it would be interesting to fuse them
2951 // will be taken care of by InstCombine.
2957 Instruction *I = cast<Instruction>(P->first),
2958 *J = cast<Instruction>(P->second);
2960 DEBUG(dbgs() << "BBV: fusing: " << *I <<
2961 " <-> " << *J << "\n");
2963 // Remove the pair and flipped pair from the list.
2964 DenseMap<Value *, Value *>::iterator FP = ChosenPairs.find(P->second);
2965 assert(FP != ChosenPairs.end() && "Flipped pair not found in list");
2966 ChosenPairs.erase(FP);
2967 ChosenPairs.erase(P);
2969 if (!canMoveUsesOfIAfterJ(BB, LoadMoveSetPairs, I, J)) {
2970 DEBUG(dbgs() << "BBV: fusion of: " << *I <<
2972 " aborted because of non-trivial dependency cycle\n");
2978 // If the pair must have the other order, then flip it.
2979 bool FlipPairOrder = FixedOrderPairs.count(ValuePair(J, I));
2980 if (!FlipPairOrder && !FixedOrderPairs.count(ValuePair(I, J))) {
2981 // This pair does not have a fixed order, and so we might want to
2982 // flip it if that will yield fewer shuffles. We count the number
2983 // of dependencies connected via swaps, and those directly connected,
2984 // and flip the order if the number of swaps is greater.
2985 bool OrigOrder = true;
2986 DenseMap<ValuePair, std::vector<ValuePair> >::iterator IJ =
2987 ConnectedPairDeps.find(ValuePair(I, J));
2988 if (IJ == ConnectedPairDeps.end()) {
2989 IJ = ConnectedPairDeps.find(ValuePair(J, I));
2993 if (IJ != ConnectedPairDeps.end()) {
2994 unsigned NumDepsDirect = 0, NumDepsSwap = 0;
2995 for (std::vector<ValuePair>::iterator T = IJ->second.begin(),
2996 TE = IJ->second.end(); T != TE; ++T) {
2997 VPPair Q(IJ->first, *T);
2998 DenseMap<VPPair, unsigned>::iterator R =
2999 PairConnectionTypes.find(VPPair(Q.second, Q.first));
3000 assert(R != PairConnectionTypes.end() &&
3001 "Cannot find pair connection type");
3002 if (R->second == PairConnectionDirect)
3004 else if (R->second == PairConnectionSwap)
3009 std::swap(NumDepsDirect, NumDepsSwap);
3011 if (NumDepsSwap > NumDepsDirect) {
3012 FlipPairOrder = true;
3013 DEBUG(dbgs() << "BBV: reordering pair: " << *I <<
3014 " <-> " << *J << "\n");
3019 Instruction *L = I, *H = J;
3023 // If the pair being fused uses the opposite order from that in the pair
3024 // connection map, then we need to flip the types.
3025 DenseMap<ValuePair, std::vector<ValuePair> >::iterator HL =
3026 ConnectedPairs.find(ValuePair(H, L));
3027 if (HL != ConnectedPairs.end())
3028 for (std::vector<ValuePair>::iterator T = HL->second.begin(),
3029 TE = HL->second.end(); T != TE; ++T) {
3030 VPPair Q(HL->first, *T);
3031 DenseMap<VPPair, unsigned>::iterator R = PairConnectionTypes.find(Q);
3032 assert(R != PairConnectionTypes.end() &&
3033 "Cannot find pair connection type");
3034 if (R->second == PairConnectionDirect)
3035 R->second = PairConnectionSwap;
3036 else if (R->second == PairConnectionSwap)
3037 R->second = PairConnectionDirect;
3040 bool LBeforeH = !FlipPairOrder;
3041 unsigned NumOperands = I->getNumOperands();
3042 SmallVector<Value *, 3> ReplacedOperands(NumOperands);
3043 getReplacementInputsForPair(Context, L, H, ReplacedOperands,
3046 // Make a copy of the original operation, change its type to the vector
3047 // type and replace its operands with the vector operands.
3048 Instruction *K = L->clone();
3051 else if (H->hasName())
3054 if (!isa<StoreInst>(K))
3055 K->mutateType(getVecTypeForPair(L->getType(), H->getType()));
3057 combineMetadata(K, H);
3058 K->intersectOptionalDataWith(H);
3060 for (unsigned o = 0; o < NumOperands; ++o)
3061 K->setOperand(o, ReplacedOperands[o]);
3065 // Instruction insertion point:
3066 Instruction *InsertionPt = K;
3067 Instruction *K1 = 0, *K2 = 0;
3068 replaceOutputsOfPair(Context, L, H, K, InsertionPt, K1, K2);
3070 // The use tree of the first original instruction must be moved to after
3071 // the location of the second instruction. The entire use tree of the
3072 // first instruction is disjoint from the input tree of the second
3073 // (by definition), and so commutes with it.
3075 moveUsesOfIAfterJ(BB, LoadMoveSetPairs, InsertionPt, I, J);
3077 if (!isa<StoreInst>(I)) {
3078 L->replaceAllUsesWith(K1);
3079 H->replaceAllUsesWith(K2);
3080 AA->replaceWithNewValue(L, K1);
3081 AA->replaceWithNewValue(H, K2);
3084 // Instructions that may read from memory may be in the load move set.
3085 // Once an instruction is fused, we no longer need its move set, and so
3086 // the values of the map never need to be updated. However, when a load
3087 // is fused, we need to merge the entries from both instructions in the
3088 // pair in case those instructions were in the move set of some other
3089 // yet-to-be-fused pair. The loads in question are the keys of the map.
3090 if (I->mayReadFromMemory()) {
3091 std::vector<ValuePair> NewSetMembers;
3092 DenseMap<Value *, std::vector<Value *> >::iterator II =
3093 LoadMoveSet.find(I);
3094 if (II != LoadMoveSet.end())
3095 for (std::vector<Value *>::iterator N = II->second.begin(),
3096 NE = II->second.end(); N != NE; ++N)
3097 NewSetMembers.push_back(ValuePair(K, *N));
3098 DenseMap<Value *, std::vector<Value *> >::iterator JJ =
3099 LoadMoveSet.find(J);
3100 if (JJ != LoadMoveSet.end())
3101 for (std::vector<Value *>::iterator N = JJ->second.begin(),
3102 NE = JJ->second.end(); N != NE; ++N)
3103 NewSetMembers.push_back(ValuePair(K, *N));
3104 for (std::vector<ValuePair>::iterator A = NewSetMembers.begin(),
3105 AE = NewSetMembers.end(); A != AE; ++A) {
3106 LoadMoveSet[A->first].push_back(A->second);
3107 LoadMoveSetPairs.insert(*A);
3111 // Before removing I, set the iterator to the next instruction.
3112 PI = llvm::next(BasicBlock::iterator(I));
3113 if (cast<Instruction>(PI) == J)
3118 I->eraseFromParent();
3119 J->eraseFromParent();
3121 DEBUG(if (PrintAfterEveryPair) dbgs() << "BBV: block is now: \n" <<
3125 DEBUG(dbgs() << "BBV: final: \n" << BB << "\n");
3129 char BBVectorize::ID = 0;
3130 static const char bb_vectorize_name[] = "Basic-Block Vectorization";
3131 INITIALIZE_PASS_BEGIN(BBVectorize, BBV_NAME, bb_vectorize_name, false, false)
3132 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
3133 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
3134 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
3135 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
3136 INITIALIZE_PASS_END(BBVectorize, BBV_NAME, bb_vectorize_name, false, false)
3138 BasicBlockPass *llvm::createBBVectorizePass(const VectorizeConfig &C) {
3139 return new BBVectorize(C);
3143 llvm::vectorizeBasicBlock(Pass *P, BasicBlock &BB, const VectorizeConfig &C) {
3144 BBVectorize BBVectorizer(P, C);
3145 return BBVectorizer.vectorizeBB(BB);
3148 //===----------------------------------------------------------------------===//
3149 VectorizeConfig::VectorizeConfig() {
3150 VectorBits = ::VectorBits;
3151 VectorizeBools = !::NoBools;
3152 VectorizeInts = !::NoInts;
3153 VectorizeFloats = !::NoFloats;
3154 VectorizePointers = !::NoPointers;
3155 VectorizeCasts = !::NoCasts;
3156 VectorizeMath = !::NoMath;
3157 VectorizeFMA = !::NoFMA;
3158 VectorizeSelect = !::NoSelect;
3159 VectorizeCmp = !::NoCmp;
3160 VectorizeGEP = !::NoGEP;
3161 VectorizeMemOps = !::NoMemOps;
3162 AlignedOnly = ::AlignedOnly;
3163 ReqChainDepth= ::ReqChainDepth;
3164 SearchLimit = ::SearchLimit;
3165 MaxCandPairsForCycleCheck = ::MaxCandPairsForCycleCheck;
3166 SplatBreaksChain = ::SplatBreaksChain;
3167 MaxInsts = ::MaxInsts;
3168 MaxIter = ::MaxIter;
3169 Pow2LenOnly = ::Pow2LenOnly;
3170 NoMemOpBoost = ::NoMemOpBoost;
3171 FastDep = ::FastDep;